test: add dataset guard + fix imports

update docs + docstrings + examples + add minimal test
add context guards
2026-05-30 18:31:25 +00:00 · 2026-04-20 00:36:02 +02:00 · 2026-04-19 23:53:53 +02:00 · 2026-04-19 23:21:14 +02:00 · 2026-04-19 22:48:08 +02:00 · 2026-04-19 16:50:19 +02:00
310 changed files with 6584 additions and 26697 deletions
--- a/.github/workflows/benchmark_tests.yml
+++ b/.github/workflows/benchmark_tests.yml
@@ -83,13 +83,10 @@ jobs:
          cache-binary: false

      - name: Login to Docker Hub
-        if: ${{ env.DOCKERHUB_USERNAME != '' }}
        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}

      # Build the benchmark-specific image. The Dockerfile separates dep-install
      # from source-copy, so code-only changes skip the slow uv-sync layer
@@ -118,7 +115,7 @@ jobs:
            bash -c "
              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
              lerobot-eval \
-                --policy.path=lerobot/smolvla_libero \
+                --policy.path=pepijn223/smolvla_libero \
                --env.type=libero \
                --env.task=libero_spatial \
                --eval.batch_size=1 \
@@ -147,7 +144,7 @@ jobs:
            --artifacts-dir /tmp/libero-artifacts \
            --env libero \
            --task libero_spatial \
-            --policy lerobot/smolvla_libero
+            --policy pepijn223/smolvla_libero

      - name: Upload Libero rollout video
        if: always()
@@ -241,13 +238,10 @@ jobs:
          cache-binary: false

      - name: Login to Docker Hub
-        if: ${{ env.DOCKERHUB_USERNAME != '' }}
        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}

      - name: Build MetaWorld benchmark image
        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
@@ -270,7 +264,7 @@ jobs:
            bash -c "
              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
              lerobot-eval \
-                --policy.path=lerobot/smolvla_metaworld \
+                --policy.path=pepijn223/smolvla_metaworld \
                --env.type=metaworld \
                --env.task=metaworld-push-v3 \
                --eval.batch_size=1 \
@@ -299,7 +293,7 @@ jobs:
            --artifacts-dir /tmp/metaworld-artifacts \
            --env metaworld \
            --task metaworld-push-v3 \
-            --policy lerobot/smolvla_metaworld
+            --policy pepijn223/smolvla_metaworld

      - name: Upload MetaWorld rollout video
        if: always()
@@ -316,636 +310,3 @@ jobs:
          name: metaworld-metrics
          path: /tmp/metaworld-artifacts/metrics.json
          if-no-files-found: warn
-
-  # ── ROBOTWIN 2.0 ──────────────────────────────────────────────────────────
-  # Isolated image: full RoboTwin 2.0 stack — SAPIEN, mplib, CuRobo,
-  # pytorch3d, + simulation assets (~4 GB).
-  # Build takes ~20 min on first run; subsequent runs hit the layer cache.
-  # Requires an NVIDIA GPU runner with CUDA 12.1 drivers.
-  robotwin-integration-test:
-    name: RoboTwin 2.0 — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-      ROBOTWIN_POLICY: lerobot/smolvla_robotwin
-      ROBOTWIN_TASKS: beat_block_hammer,click_bell,handover_block,stack_blocks_two,click_alarmclock,open_microwave,adjust_bottle,lift_pot,stamp_seal,turn_switch
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          cache-binary: false
-
-      - name: Login to Docker Hub
-        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      # Build the full-install image: SAPIEN, mplib, CuRobo, pytorch3d +
-      # simulation assets (~4 GB). Layer cache lives in the runner's local
-      # Docker daemon — reused across re-runs on the same machine.
-      - name: Build RoboTwin 2.0 benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.robotwin
-          push: false
-          load: true
-          tags: lerobot-benchmark-robotwin:ci
-          cache-from: type=local,src=/tmp/.buildx-cache-robotwin
-          cache-to: type=local,dest=/tmp/.buildx-cache-robotwin,mode=max
-
-      - name: Run RoboTwin 2.0 smoke eval (10 tasks, 1 episode each)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          # Named container (no --rm) so we can docker cp artifacts out.
-          docker run --name robotwin-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e ROBOTWIN_POLICY="${ROBOTWIN_POLICY}" \
-            -e ROBOTWIN_TASKS="${ROBOTWIN_TASKS}" \
-            lerobot-benchmark-robotwin:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              cd /opt/robotwin && lerobot-eval \
-                --policy.path=\"\$ROBOTWIN_POLICY\" \
-                --env.type=robotwin \
-                --env.task=\"\$ROBOTWIN_TASKS\" \
-                --env.max_parallel_tasks=5 \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--rename_map={\"observation.images.head_camera\": \"observation.images.camera1\", \"observation.images.left_camera\": \"observation.images.camera2\", \"observation.images.right_camera\": \"observation.images.camera3\"}' \
-                --output_dir=/tmp/eval-artifacts
-              python /lerobot/scripts/ci/extract_task_descriptions.py \
-                --env robotwin \
-                --task \"\$ROBOTWIN_TASKS\" \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy RoboTwin artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/robotwin-artifacts
-          docker cp robotwin-eval:/tmp/eval-artifacts/. /tmp/robotwin-artifacts/ 2>/dev/null || true
-          docker rm -f robotwin-eval || true
-
-      - name: Parse RoboTwin eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/robotwin-artifacts \
-            --env robotwin \
-            --task "${ROBOTWIN_TASKS}" \
-            --policy "${ROBOTWIN_POLICY}"
-
-      - name: Upload RoboTwin rollout video
-        if: always()
-        uses: actions/upload-artifact@v4
-        with:
-          name: robotwin-rollout-video
-          path: /tmp/robotwin-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload RoboTwin eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4
-        with:
-          name: robotwin-metrics
-          path: /tmp/robotwin-artifacts/metrics.json
-          if-no-files-found: warn
-
-  # ── ROBOCASA365 ──────────────────────────────────────────────────────────
-  # Isolated image: robocasa + robosuite installed manually as editable
-  # clones (no `lerobot[robocasa]` extra — robocasa's setup.py pins
-  # `lerobot==0.3.3`, which would shadow this repo's lerobot).
-  robocasa-integration-test:
-    name: RoboCasa365 — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          cache-binary: false
-
-      - name: Login to Docker Hub
-        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      - name: Build RoboCasa365 benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.robocasa
-          push: false
-          load: true
-          tags: lerobot-benchmark-robocasa:ci
-
-      - name: Run RoboCasa365 smoke eval (10 atomic tasks, 1 episode each)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          docker run --name robocasa-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
-            -e MUJOCO_GL=egl \
-            lerobot-benchmark-robocasa:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              lerobot-eval \
-                --policy.path=lerobot/smolvla_robocasa \
-                --env.type=robocasa \
-                --env.task=CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove,CloseToasterOvenDoor,SlideDishwasherRack,TurnOnSinkFaucet,NavigateKitchen,TurnOnElectricKettle \
-                --env.max_parallel_tasks=5 \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--rename_map={\"observation.images.robot0_agentview_left\": \"observation.images.camera1\", \"observation.images.robot0_eye_in_hand\": \"observation.images.camera2\", \"observation.images.robot0_agentview_right\": \"observation.images.camera3\"}' \
-                --output_dir=/tmp/eval-artifacts
-              python scripts/ci/extract_task_descriptions.py \
-                --env robocasa \
-                --task CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove,CloseToasterOvenDoor,SlideDishwasherRack,TurnOnSinkFaucet,NavigateKitchen,TurnOnElectricKettle \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy RoboCasa365 artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/robocasa-artifacts
-          docker cp robocasa-eval:/tmp/eval-artifacts/. /tmp/robocasa-artifacts/ 2>/dev/null || true
-          docker rm -f robocasa-eval || true
-
-      - name: Parse RoboCasa365 eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/robocasa-artifacts \
-            --env robocasa \
-            --task atomic_smoke_10 \
-            --policy lerobot/smolvla_robocasa
-
-      - name: Upload RoboCasa365 rollout video
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robocasa-rollout-video
-          path: /tmp/robocasa-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload RoboCasa365 eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robocasa-metrics
-          path: /tmp/robocasa-artifacts/metrics.json
-          if-no-files-found: warn
-
-  # ── ROBOCEREBRA ───────────────────────────────────────────────────────────
-  # Reuses the LIBERO simulator (libero_10 suite) with RoboCerebra camera
-  # defaults (image/wrist_image). The image is layered on
-  # huggingface/lerobot-gpu, which already ships [libero] as part of [all].
-  robocerebra-integration-test:
-    name: RoboCerebra — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          cache-binary: false
-
-      - name: Login to Docker Hub
-        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      - name: Build RoboCerebra benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.robocerebra
-          push: false
-          load: true
-          tags: lerobot-benchmark-robocerebra:ci
-          cache-from: type=local,src=/tmp/.buildx-cache-robocerebra
-          cache-to: type=local,dest=/tmp/.buildx-cache-robocerebra,mode=max
-
-      - name: Run RoboCerebra smoke eval (1 episode)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          docker run --name robocerebra-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
-            -e LIBERO_DATA_FOLDER=/tmp/libero_data \
-            lerobot-benchmark-robocerebra:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              lerobot-eval \
-                --policy.path=lerobot/smolvla_robocerebra \
-                --env.type=libero \
-                --env.task=libero_10 \
-                --env.fps=20 \
-                --env.obs_type=pixels_agent_pos \
-                --env.observation_height=256 \
-                --env.observation_width=256 \
-                '--env.camera_name_mapping={\"agentview_image\": \"image\", \"robot0_eye_in_hand_image\": \"wrist_image\"}' \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--rename_map={\"observation.images.image\": \"observation.images.camera1\", \"observation.images.wrist_image\": \"observation.images.camera2\"}' \
-                --policy.empty_cameras=1 \
-                --output_dir=/tmp/eval-artifacts
-              python scripts/ci/extract_task_descriptions.py \
-                --env libero --task libero_10 \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy RoboCerebra artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/robocerebra-artifacts
-          docker cp robocerebra-eval:/tmp/eval-artifacts/. /tmp/robocerebra-artifacts/ 2>/dev/null || true
-          docker rm -f robocerebra-eval || true
-
-      - name: Parse RoboCerebra eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/robocerebra-artifacts \
-            --env robocerebra \
-            --task libero_10 \
-            --policy lerobot/smolvla_robocerebra
-
-      - name: Upload RoboCerebra rollout video
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robocerebra-rollout-video
-          path: /tmp/robocerebra-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload RoboCerebra eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robocerebra-metrics
-          path: /tmp/robocerebra-artifacts/metrics.json
-          if-no-files-found: warn
-
-  # ── ROBOMME ───────────────────────────────────────────────────────────────
-  # Isolated image: mani-skill/SAPIEN/Vulkan chain with gymnasium and numpy
-  # overrides (robomme can't be a pyproject extra due to numpy<2 pin).
-  robomme-integration-test:
-    name: RoboMME — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-      ROBOMME_POLICY: lerobot/smolvla_robomme
-      ROBOMME_TASKS: PickXtimes,BinFill,StopCube,MoveCube,InsertPeg,SwingXtimes,VideoUnmask,ButtonUnmask,PickHighlight,PatternLock
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          cache-binary: false
-
-      - name: Login to Docker Hub
-        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      - name: Build RoboMME benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.robomme
-          push: false
-          load: true
-          tags: lerobot-benchmark-robomme:ci
-
-      - name: Run RoboMME smoke eval (10 tasks, 1 episode each)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          docker run --name robomme-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
-            -e ROBOMME_POLICY="${ROBOMME_POLICY}" \
-            -e ROBOMME_TASKS="${ROBOMME_TASKS}" \
-            lerobot-benchmark-robomme:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              lerobot-eval \
-                --policy.path=\"\$ROBOMME_POLICY\" \
-                --env.type=robomme \
-                --env.task=\"\$ROBOMME_TASKS\" \
-                --env.dataset_split=test \
-                --env.task_ids=[0] \
-                --env.max_parallel_tasks=5 \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--rename_map={\"observation.images.image\": \"observation.images.camera1\", \"observation.images.wrist_image\": \"observation.images.camera2\"}' \
-                --policy.empty_cameras=3 \
-                --output_dir=/tmp/eval-artifacts
-              python scripts/ci/extract_task_descriptions.py \
-                --env robomme --task \"\$ROBOMME_TASKS\" \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy RoboMME artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/robomme-artifacts
-          docker cp robomme-eval:/tmp/eval-artifacts/. /tmp/robomme-artifacts/ 2>/dev/null || true
-          docker rm -f robomme-eval || true
-
-      - name: Parse RoboMME eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/robomme-artifacts \
-            --env robomme \
-            --task "${ROBOMME_TASKS}" \
-            --policy "${ROBOMME_POLICY}"
-
-      - name: Upload RoboMME rollout video
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robomme-rollout-video
-          path: /tmp/robomme-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload RoboMME eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robomme-metrics
-          path: /tmp/robomme-artifacts/metrics.json
-          if-no-files-found: warn
-
-  # ── LIBERO-plus ───────────────────────────────────────────────────────────
-  # Isolated image: LIBERO-plus fork cloned into /home/user_lerobot on top of
-  # huggingface/lerobot-gpu (see docker/Dockerfile.benchmark.libero_plus).
-  libero-plus-integration-test:
-    name: LIBERO-plus — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-      LIBERO_PLUS_SUITE: libero_spatial
-      LIBERO_PLUS_POLICY: lerobot/smolvla_libero_plus
-      LIBERO_PLUS_TASK_IDS: "[0,100,260,500,1000,1500,2000,2400]"
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          cache-binary: false
-
-      - name: Login to Docker Hub
-        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      - name: Build LIBERO-plus benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.libero_plus
-          push: false
-          load: true
-          tags: lerobot-benchmark-libero-plus:ci
-          cache-from: type=local,src=/tmp/.buildx-cache-libero-plus
-          cache-to: type=local,dest=/tmp/.buildx-cache-libero-plus,mode=max
-
-      - name: Run LIBERO-plus smoke eval (1 episode)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          docker run --name libero-plus-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
-            -e LIBERO_PLUS_SUITE="${LIBERO_PLUS_SUITE}" \
-            -e LIBERO_PLUS_POLICY="${LIBERO_PLUS_POLICY}" \
-            -e LIBERO_PLUS_TASK_IDS="${LIBERO_PLUS_TASK_IDS}" \
-            lerobot-benchmark-libero-plus:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              lerobot-eval \
-                --policy.path=\"\$LIBERO_PLUS_POLICY\" \
-                --env.type=libero_plus \
-                --env.task=\"\$LIBERO_PLUS_SUITE\" \
-                --env.task_ids=\"\$LIBERO_PLUS_TASK_IDS\" \
-                --env.max_parallel_tasks=5 \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--env.camera_name_mapping={\"agentview_image\": \"camera1\", \"robot0_eye_in_hand_image\": \"camera2\"}' \
-                --policy.empty_cameras=1 \
-                --output_dir=/tmp/eval-artifacts
-              python scripts/ci/extract_task_descriptions.py \
-                --env libero_plus --task \"\$LIBERO_PLUS_SUITE\" \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy LIBERO-plus artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/libero-plus-artifacts
-          docker cp libero-plus-eval:/tmp/eval-artifacts/. /tmp/libero-plus-artifacts/ 2>/dev/null || true
-          docker rm -f libero-plus-eval || true
-
-      - name: Parse LIBERO-plus eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/libero-plus-artifacts \
-            --env libero_plus \
-            --task "${LIBERO_PLUS_SUITE}" \
-            --policy "${LIBERO_PLUS_POLICY}"
-
-      - name: Upload LIBERO-plus rollout video
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: libero-plus-rollout-video
-          path: /tmp/libero-plus-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload LIBERO-plus eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: libero-plus-metrics
-          path: /tmp/libero-plus-artifacts/metrics.json
-          if-no-files-found: warn
-
-  # ── VLABENCH ─────────────────────────────────────────────────────────────
-  # Isolated image: lerobot[vlabench] only (VLABench, mujoco==3.2.2, dm-control chain)
-  vlabench-integration-test:
-    name: VLABench — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          cache-binary: false
-
-      - name: Login to Docker Hub
-        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      - name: Build VLABench benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.vlabench
-          push: false
-          load: true
-          tags: lerobot-benchmark-vlabench:ci
-          build-args: |
-            VLABENCH_ASSETS_REPO=lerobot/vlabench-assets
-
-      - name: Run VLABench smoke eval (10 tasks, 1 episode each)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          docker run --name vlabench-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
-            -e MUJOCO_GL=egl \
-            lerobot-benchmark-vlabench:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              lerobot-eval \
-                --policy.path=lerobot/smolvla_vlabench \
-                --env.type=vlabench \
-                --env.task=select_fruit,select_toy,select_book,select_painting,select_drink,select_ingredient,select_billiards,select_poker,add_condiment,insert_flower \
-                --env.episode_length=50 \
-                --env.max_parallel_tasks=5 \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--rename_map={\"observation.images.image\": \"observation.images.camera1\", \"observation.images.second_image\": \"observation.images.camera2\", \"observation.images.wrist_image\": \"observation.images.camera3\"}' \
-                --output_dir=/tmp/eval-artifacts
-              python scripts/ci/extract_task_descriptions.py \
-                --env vlabench \
-                --task select_fruit,select_toy,select_book,select_painting,select_drink,select_ingredient,select_billiards,select_poker,add_condiment,insert_flower \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy VLABench artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/vlabench-artifacts
-          docker cp vlabench-eval:/tmp/eval-artifacts/. /tmp/vlabench-artifacts/ 2>/dev/null || true
-          docker rm -f vlabench-eval || true
-
-      - name: Parse VLABench eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/vlabench-artifacts \
-            --env vlabench \
-            --task select_fruit,select_toy,select_book,select_painting,select_drink,select_ingredient,select_billiards,select_poker,add_condiment,insert_flower \
-            --policy lerobot/smolvla_vlabench
-
-      - name: Upload VLABench rollout video
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: vlabench-rollout-video
-          path: /tmp/vlabench-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload VLABench eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: vlabench-metrics
-          path: /tmp/vlabench-artifacts/metrics.json
-          if-no-files-found: warn
--- a/.github/workflows/documentation-upload-pr.yml
+++ b/.github/workflows/documentation-upload-pr.yml
@@ -33,7 +33,7 @@ jobs:
      github.event.workflow_run.event == 'pull_request' &&
      github.event.workflow_run.conclusion == 'success' &&
      github.repository == 'huggingface/lerobot'
-    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@9ad2de8582b56c017cb530c1165116d40433f1c6  # main
    with:
      package_name: lerobot
    secrets:
--- a/.github/workflows/documentation.yml
+++ b/.github/workflows/documentation.yml
@@ -55,7 +55,7 @@ jobs:
      github.repository == 'huggingface/lerobot'
    permissions:
      contents: read
-    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
+    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
    with:
      commit_sha: ${{ github.sha }}
      package: lerobot
@@ -78,7 +78,7 @@ jobs:
    permissions:
      contents: read
      pull-requests: write
-    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
+    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
    with:
      commit_sha: ${{ github.event.pull_request.head.sha }}
      pr_number: ${{ github.event.number }}
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -152,14 +152,13 @@ jobs:
            BASE_VERSION="${VERSION%%-*}"
            echo "Installing pre-release version $BASE_VERSION from TestPyPI..."
            uv pip install \
-              --torch-backend cpu \
              --index-url https://test.pypi.org/simple/ \
              --extra-index-url https://pypi.org/simple \
              --index-strategy unsafe-best-match \
               "lerobot[all]==$BASE_VERSION"
          else
            echo "Installing release version $VERSION from PyPI..."
-            uv pip install --torch-backend cpu "lerobot[all]==$VERSION"
+            uv pip install "lerobot[all]==$VERSION"
          fi
      - name: Check lerobot version
        run: uv run python -c "import lerobot; print(lerobot.__version__)"
--- a/.github/workflows/stale.yml
+++ b/.github/workflows/stale.yml
@@ -19,19 +19,19 @@ on:
  workflow_dispatch:

  # Runs at 02:00
-  # schedule:
-  #   - cron: "0 2 * * *"
+  schedule:
+    - cron: "0 2 * * *"

 env:
  CLOSE_ISSUE_MESSAGE: >
-    This issue was closed because it has been stalled for 30 days with no activity.
+    This issue was closed because it has been stalled for 14 days with no activity.
    Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
  CLOSE_PR_MESSAGE: >
-    This PR was closed because it has been stalled for 30 days with no activity.
+    This PR was closed because it has been stalled for 21 days with no activity.
    Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
  WARN_ISSUE_MESSAGE: >
    This issue has been automatically marked as stale because it has not had
-    recent activity (1 year). It will be closed if no further activity occurs.
+    recent activity (6 months). It will be closed if no further activity occurs.
    Any change, comment or update to this issue will reset this count.
    Thank you for your contributions.
  WARN_PR_MESSAGE: >
@@ -59,10 +59,10 @@ jobs:
          stale-pr-label: stale
          exempt-issue-labels: never-stale
          exempt-pr-labels: never-stale
-          days-before-issue-stale: 365
-          days-before-issue-close: 30
+          days-before-issue-stale: 180
+          days-before-issue-close: 14
          days-before-pr-stale: 365
-          days-before-pr-close: 30
+          days-before-pr-close: 21
          delete-branch: true
          close-issue-message: ${{ env.CLOSE_ISSUE_MESSAGE }}
          close-pr-message: ${{ env.CLOSE_PR_MESSAGE }}
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -1,7 +1,5 @@
 This file provides guidance to AI agents when working with code in this repository.

-> **User-facing help → [`AGENT_GUIDE.md`](./AGENT_GUIDE.md)** (SO-101 setup, recording, picking a policy, training duration, eval — with copy-pasteable commands).
-
 ## Project Overview

 LeRobot is a PyTorch-based library for real-world robotics, providing datasets, pretrained policies, and tools for training, evaluation, data collection, and robot control. It integrates with Hugging Face Hub for model/dataset sharing.
--- a/AGENT_GUIDE.md
+++ b/AGENT_GUIDE.md
@@ -1,412 +0,0 @@
-# AGENT_GUIDE.md — LeRobot Helper for AI Agents & Users
-
-This file is a practical, copy-paste-friendly companion for any AI agent (Cursor, Claude, ChatGPT, Codex, etc.) helping a user work with LeRobot. It complements [`AGENTS.md`](./AGENTS.md) (dev/contributor context) with **user-facing guidance**: how to start, what to train, how long, how to record, and how to calibrate an SO-101.
-
---
-
-## 1. Start here — ask the user first (MANDATORY)
-
-Before suggesting any command, an agent MUST ask the user at least these questions and wait for answers:
-
-1. **What's your goal?** (e.g. "teach my SO-101 to fold a cloth", "train a policy on an existing HF dataset", "contribute a PR", "understand the codebase")
-2. **What hardware do you have?**
-   - Robot: none / SO-100 / SO-101 / Koch / LeKiwi / Reachy / other
-   - Teleop: leader arm / phone / keyboard / gamepad / none
-   - Cameras: how many, resolution, fixed or moving?
-3. **What machine will you train on?**
-   - GPU model + VRAM (e.g. "laptop 3060 6 GB", "RTX 4090 24 GB", "A100 80 GB", "CPU only")
-   - OS: macOS / Linux / Windows
-4. **Skill level & time budget?** First time, some ML, experienced? Hours, days, a weekend?
-5. **Do you already have a dataset?** Yes (HF repo id?) / no / want to record one
-6. **How can I help right now?** (pick one concrete next step)
-
-Only after you have answers, propose a concrete path. If something is ambiguous, ask again rather than guessing. Bias toward **the simplest thing that works** for the user's hardware and goal.
-
---
-
-## 2. LeRobot in 60 seconds
-
-LeRobot = **datasets + policies + envs + robot control**, unified by a small set of strong abstractions.
-
- **`LeRobotDataset`** — episode-aware dataset (video or images + actions + state), loadable from the Hub or disk.
- **Policies** (`ACT`, `Diffusion`, `SmolVLA`, `π0`, `π0.5`, `Wall-X`, `X-VLA`, `VQ-BeT`, `TD-MPC`, …) — all inherit `PreTrainedPolicy` and can be pushed/pulled from the Hub.
- **Processors** — small composable transforms between dataset → policy → robot.
- **Envs** (sim) and **Robots** (real) — same action/observation contract so code swaps cleanly.
- **CLI** — `lerobot-record`, `lerobot-train`, `lerobot-eval`, `lerobot-teleoperate`, `lerobot-calibrate`, `lerobot-find-port`, `lerobot-setup-motors`, `lerobot-replay`.
-
-See [`AGENTS.md`](./AGENTS.md) for repo architecture.
-
---
-
-## 3. Quickstart paths (pick one)
-
-### Path A — "I have an SO-101 and want my first trained policy"
-
-Go to §4 (SO-101 end-to-end), then §5 (data tips), then §6 (pick a policy — likely **ACT**), then §7 (how long), then §8 (eval).
-
-### Path B — "No hardware, I want to train on an existing dataset"
-
-Skip §4. Pick a policy in §6, pick a duration in §7, then run `lerobot-train` per §4.9 with a Hub `--dataset.repo_id` and an `--env.type` for eval. Finish with §8.
-
-### Path C — "I just want to understand the codebase"
-
-Read §2 above, then `AGENTS.md` "Architecture", then open `src/lerobot/policies/act/` and `src/lerobot/datasets/lerobot_dataset.py` as canonical examples.
-
---
-
-## 4. SO-101 end-to-end cheat-sheet
-
-Full details in [`docs/source/so101.mdx`](./docs/source/so101.mdx) and [`docs/source/il_robots.mdx`](./docs/source/il_robots.mdx). Minimum commands in order. Confirm arms are assembled + powered before issuing.
-
-**4.1 Install**
-
-```bash
-pip install 'lerobot[feetech]'              # SO-100/SO-101 motor stack
-# pip install 'lerobot[all]'                # everything
-# pip install 'lerobot[aloha,pusht]'        # specific features
-# pip install 'lerobot[smolvla]'            # add SmolVLA deps
-git lfs install && git lfs pull
-hf auth login                               # required to push datasets/policies
-```
-
-Contributors can alternatively use `uv sync --locked --extra feetech` (see `AGENTS.md`).
-
-**4.2 Find USB ports** — run once per arm, unplug when prompted.
-
-```bash
-lerobot-find-port
-```
-
-macOS: `/dev/tty.usbmodem...`; Linux: `/dev/ttyACM0` (may need `sudo chmod 666 /dev/ttyACM0`).
-
-**4.3 Setup motor IDs & baudrate** (one-time, per arm)
-
-```bash
-lerobot-setup-motors --robot.type=so101_follower --robot.port=<FOLLOWER_PORT>
-lerobot-setup-motors --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>
-```
-
-**4.4 Calibrate** — center all joints, press Enter, sweep each joint through its full range. The `id` is the calibration key — reuse it everywhere.
-
-```bash
-lerobot-calibrate --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower
-lerobot-calibrate --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>   --teleop.id=my_leader
-```
-
-**4.5 Teleoperate** (sanity check, no recording)
-
-```bash
-lerobot-teleoperate \
-  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
-  --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>  --teleop.id=my_leader \
-  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-  --display_data=true
-```
-
-> **Feetech timeout / comms error on SO-100 / SO-101?** Before touching software, check the **red motor LEDs** on the daisy chain.
->
-> - **All steady red, gripper → base chain** → wiring OK.
-> - **One or more motors dark / chain stops mid-way** → wiring issue: reseat the 3-pin cables, check the controller-board power supply, and make sure each motor is fully clicked in.
-> - **LEDs blinking** → the motor is in an **error state**: usually overload (forcing a joint past its limit) **or wrong power supply voltage**. SO-100 / SO-101 ship in two variants — a **5 V / 7.4 V** build and a **12 V** build — they are NOT interchangeable. Using a 12 V PSU on a 5 V / 7.4 V arm (or vice-versa) will trip this error; confirm your motor variant before powering up.
->
-> Most "timeout" errors are physical, not code.
-
-**4.6 Record a dataset** — keys: **→** next, **←** redo, **ESC** finish & upload.
-
-```bash
-HF_USER=$(NO_COLOR=1 hf auth whoami | awk -F': *' 'NR==1 {print $2}')
-
-lerobot-record \
-  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
-  --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>  --teleop.id=my_leader \
-  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-  --dataset.repo_id=${HF_USER}/my_task \
-  --dataset.single_task="<describe the task in one sentence>" \
-  --dataset.num_episodes=50 \
-  --dataset.episode_time_s=30 \
-  --dataset.reset_time_s=10 \
-  --display_data=true
-```
-
-**4.7 Visualize** — **always** do this before training. Look for missing frames, camera blur, unreachable targets, inconsistent object positions.
-After upload: https://huggingface.co/spaces/lerobot/visualize_dataset → paste `${HF_USER}/my_task`. Works for **any LeRobot-formatted Hub dataset** — use it to scout other datasets, inspect episode quality, or debug your own data before retraining.
-
-**4.8 Replay an episode** (sanity check)
-
-```bash
-lerobot-replay --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
-  --dataset.repo_id=${HF_USER}/my_task --dataset.episode=0
-```
-
-**4.9 Train** (default: ACT — fastest, lowest memory). Apple silicon: `--policy.device=mps`. See §6/§7 for policy and duration.
-
-```bash
-lerobot-train \
-  --dataset.repo_id=${HF_USER}/my_task \
-  --policy.type=act \
-  --policy.device=cuda \
-  --output_dir=outputs/train/act_my_task \
-  --job_name=act_my_task \
-  --batch_size=8 \
-  --wandb.enable=true \
-  --policy.repo_id=${HF_USER}/act_my_task
-```
-
-**4.10 Evaluate on the real robot** — compare success rate to a teleoperated baseline.
-
-```bash
-lerobot-record \
-  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
-  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-  --dataset.repo_id=${HF_USER}/eval_my_task \
-  --dataset.single_task="<same task description as training>" \
-  --dataset.num_episodes=10 \
-  --policy.path=${HF_USER}/act_my_task
-```
-
---
-
-## 5. Data collection tips (beginner → reliable policy)
-
-Good data beats clever models. Adopt these defaults and deviate only with evidence.
-
-### 5.1 Setup & ergonomics
-
- **Fix the rig and cameras** before touching the software. If the rig vibrates or the operator gets frustrated, fix that first — more bad data won't help.
- **Lighting matters more than resolution.** Diffuse, consistent light. Avoid moving shadows.
- **"Can you do the task from the camera view alone?"** If no, your cameras are wrong. Fix before recording.
- Enable **action interpolation** for rollouts when available for smoother trajectories.
-
-### 5.2 Practice before you record
-
- Do 5–10 demos without recording. Build a deliberate, repeatable strategy.
- Hesitant or inconsistent demos teach the model hesitation.
-
-### 5.3 Quality over speed
-
-Deliberate, high-quality execution beats fast sloppy runs. Optimize for speed only **after** strategy is dialed in — never trade quality for it.
-
-### 5.4 Consistency within and across episodes
-
-Same grasp, approach vector, and timing. Coherent strategies are much easier to learn than wildly varying movements.
-
-### 5.5 Start small, then extend (the golden rule)
-
- **First 50 episodes = constrained version** of the task: one object, fixed position, fixed camera setup, one operator.
- Train a quick ACT model. See what fails.
- **Then add diversity** along one axis at a time: more positions → more lighting → more objects → more operators.
- Don't try to collect the "perfect dataset" on day one. Iterate.
-
-### 5.6 Policy choice for beginners
-
- **Laptop / first time / want results fast → ACT.** Works surprisingly well, trains fast even on a laptop GPU.
- **Bigger GPU / language-conditioned / multi-task → SmolVLA.** Unfreezing the vision encoder (see §7) is a big win here.
- Defer π0 / π0.5 / Wall-X / X-VLA until you have a proven ACT baseline and a 20+ GB GPU.
-
-### 5.7 Recommended defaults for your first task
-
-| Setting          | Value                                                                                                                                                 |
-| ---------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------- |
-| Episodes         | **50** to start, scale to 100–300 after first training                                                                                                |
-| Episode length   | 20–45 s (shorter is fine for grasp/place)                                                                                                             |
-| Reset time       | 10 s                                                                                                                                                  |
-| FPS              | 30                                                                                                                                                    |
-| Cameras          | **2 cameras recommended**: 1 fixed front + 1 wrist. Multi-view often outperforms single-view. A single fixed camera also works to keep things simple. |
-| Task description | Short, specific, action-phrased sentence                                                                                                              |
-
-### 5.8 Troubleshooting signal
-
- Policy fails at one specific stage → record 10–20 more episodes **targeting that stage**.
- Policy flaps / oscillates → likely inconsistent demos, or need more training; re-record worst episodes (use **←** to redo).
- Policy ignores the object → camera framing or lighting issue, not a model issue.
-
-See also: [What makes a good dataset](https://huggingface.co/blog/lerobot-datasets#what-makes-a-good-dataset).
-
---
-
-## 6. Which policy should I train?
-
-Match the policy to the user's **GPU memory** and **time budget**. Numbers below come from an internal profiling run (one training update per policy). They are **indicative only** — see caveats.
-
-### 6.1 Profiling snapshot (indicative)
-
-All policies typically train for **5–10 epochs** (see §7).
-
-> **Human-facing version:** the [Compute Hardware Guide](./docs/source/hardware_guide.mdx) reuses the table below and adds a cloud-GPU tier guide and a Hugging Face Jobs pointer.
-
-| Policy      | Batch | Update (ms) | Peak GPU mem (GB) | Best for                                                                                         |
-| ----------- | ----: | ----------: | ----------------: | ------------------------------------------------------------------------------------------------ |
-| `act`       |     4 |    **83.9** |          **0.94** | First-time users, laptops, single-task. Fast and reliable.                                       |
-| `diffusion` |     4 |       168.6 |              4.94 | Multi-modal action distributions; needs mid-range GPU.                                           |
-| `smolvla`   |     1 |       357.8 |              3.93 | Language-conditioned, multi-task, small VLA. **Unfreeze vision encoder for big gains** (see §7). |
-| `xvla`      |     1 |       731.6 |             15.52 | Large VLA, multi-task.                                                                           |
-| `wall_x`    |     1 |       716.5 |             15.95 | Large VLA with world-model objective.                                                            |
-| `pi0`       |     1 |       940.3 |             15.50 | Strong large VLA baseline (Physical Intelligence).                                               |
-| `pi05`      |     1 |      1055.8 |             16.35 | Newer π policy; similar footprint to `pi0`.                                                      |
-
-**Critical caveats:**
-
- **Optimizer:** measured with **SGD**. LeRobot's default is **AdamW**, which keeps extra optimizer state → **peak memory will be noticeably higher** with the default, especially for `pi0`, `pi05`, `wall_x`, `xvla`.
- **Batch size:** the large policies were profiled at batch 1. In practice use a **larger batch** for stable training (see §7.4). Memory scales roughly linearly with batch.
-
-### 6.2 Decision rules
-
- **< 8 GB VRAM (laptop, 3060, M-series Mac):** → `act`. Maybe `diffusion` if you have ~6–8 GB free.
- **12–16 GB VRAM (4070/4080, A4000):** → `smolvla` with defaults, or `act`/`diffusion` with larger batch. `pi0`/`pi05`/`wall_x`/`xvla` feasible only with small batch + gradient accumulation.
- **24+ GB VRAM (3090/4090/A5000):** → any policy. Prefer `smolvla` (unfrozen) for multi-task; `act` for single-task grasp-and-place (still often the best ROI). Could experiment with `pi0` or `pi05` or `xvla`
- **80 GB (A100/H100):** → any, with healthy batch. `pi05`, `xvla`, `wall_x` become comfortable.
- **CPU only:** → don't train here. Use Google Colab (see [`docs/source/notebooks.mdx`](./docs/source/notebooks.mdx)) or a rented GPU.
-
---
-
-## 7. How long should I train?
-
-Robotics imitation learning usually converges in a **few epochs over the dataset**, not hundreds of thousands of raw steps. Think **epochs first**, then translate to steps.
-
-### 7.1 Rule of thumb
-
- **Typical total: 5–10 epochs.** Start at 5, eval, then decide if more helps.
- Very small datasets (< 30 episodes) may want slightly more epochs — but first, **collect more data**.
- VLAs with a pretrained vision backbone typically need **fewer** epochs than training from scratch.
-
-### 7.2 Steps ↔ epochs conversion
-
-```
-total_frames     = sum of frames over all episodes      # e.g. 50 eps × 30 fps × 30 s ≈ 45,000
-steps_per_epoch  = ceil(total_frames / batch_size)
-total_steps      = epochs × steps_per_epoch
-```
-
-Examples for `--batch_size=8`:
-
-| Dataset size            |  Frames | Steps / epoch | 5 epochs | 10 epochs |
-| ----------------------- | ------: | ------------: | -------: | --------: |
-| 50 eps × 30 s @ 30 fps  |  45,000 |        ~5,625 |      28k |       56k |
-| 100 eps × 30 s @ 30 fps |  90,000 |       ~11,250 |      56k |      113k |
-| 300 eps × 30 s @ 30 fps | 270,000 |       ~33,750 |     169k |      338k |
-
-Pass the resulting total with `--steps=<N>`; eval at intermediate checkpoints (`outputs/train/.../checkpoints/`).
-
-### 7.3 Per-policy starting points (single-task, ~50 episodes)
-
-| Policy         | Batch | Steps (first run) | Notes                                                             |
-| -------------- | ----: | ----------------: | ----------------------------------------------------------------- |
-| `act`          |  8–16 |           30k–80k | Usually converges under 50k for single-task.                      |
-| `diffusion`    |  8–16 |          80k–150k | Benefits from longer training than ACT.                           |
-| `smolvla`      |   4–8 |           30k–80k | Pretrained VLM → converges fast.                                  |
-| `pi0` / `pi05` |   1–4 |           30k–80k | Memory-bound; use gradient accumulation for effective batch ≥ 16! |
-
-### 7.4 Batch size guidance
-
- **Bigger batch is preferable** for stable gradients on teleop data.
- If GPU memory is the bottleneck, use **gradient accumulation** to raise _effective_ batch without raising peak memory.
- Scale **learning rate** gently with batch; most LeRobot defaults work fine for a 2–4× batch change.
-
-### 7.5 Scale LR schedule & checkpoints with `--steps`
-
-LeRobot's default schedulers (e.g. SmolVLA's cosine decay) use `scheduler_decay_steps=30_000`, which is sized for long training runs. When you shorten training (e.g. 5k–10k steps on a small dataset), **scale the scheduler down to match** — otherwise the LR stays near the peak and never decays. Same for checkpoint frequency.
-
-```bash
-lerobot-train ... \
-  --steps=5000 \
-  --policy.scheduler_decay_steps=5000 \
-  --save_freq=5000
-```
-
-Rule of thumb: set `scheduler_decay_steps ≈ steps`, and `save_freq` to whatever granularity you want for eval (e.g. every 1k–5k steps). Match `scheduler_warmup_steps` proportionally if your run is very short.
-
-### 7.6 SmolVLA: unfreeze the vision encoder for real gains
-
-SmolVLA ships with `freeze_vision_encoder=True`. Unfreezing usually **improves performance substantially** on specialized tasks, at the cost of more VRAM and slower steps. Enable with:
-
-```bash
-lerobot-train ... --policy.type=smolvla \
-  --policy.freeze_vision_encoder=false \
-  --policy.train_expert_only=false
-```
-
-### 7.7 Signals to stop / keep going
-
- Train loss plateaus → stop, save a Hub checkpoint.
- Train loss still dropping and you're under 10 epochs → keep going.
-
---
-
-## 8. Evaluation & benchmarks
-
-Two flavors of evaluation:
-
-### 8.1 Real-robot eval (SO-101, etc.)
-
-Reuse `lerobot-record` with `--policy.path` to run the trained policy on-robot and save the run as an eval dataset. Convention: prefix the dataset with `eval_`.
-
-```bash
-lerobot-record \
-  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
-  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-  --dataset.repo_id=${HF_USER}/eval_my_task \
-  --dataset.single_task="<same task description used during training>" \
-  --dataset.num_episodes=10 \
-  --policy.path=${HF_USER}/act_my_task
-```
-
-Report success rate across episodes. Compare to a teleoperated baseline and to an earlier checkpoint to catch regressions.
-
-### 8.2 Sim-benchmark eval
-
-For policies trained on sim datasets (PushT, Aloha, LIBERO, MetaWorld, RoboCasa, …) use `lerobot-eval` against the matching `env.type`:
-
-```bash
-lerobot-eval \
-  --policy.path=${HF_USER}/diffusion_pusht \
-  --env.type=pusht \
-  --eval.n_episodes=50 \
-  --eval.batch_size=10 \
-  --policy.device=cuda
-```
-
- Use `--policy.path=outputs/train/.../checkpoints/<step>/pretrained_model` for local checkpoints.
- `--eval.n_episodes` should be ≥ 50 for a stable success-rate estimate.
- Available envs live in `src/lerobot/envs/`. See [`docs/source/libero.mdx`](./docs/source/libero.mdx), [`metaworld.mdx`](./docs/source/metaworld.mdx), [`robocasa.mdx`](./docs/source/robocasa.mdx), [`vlabench.mdx`](./docs/source/vlabench.mdx) for specific benchmarks.
- To add a new benchmark, see [`docs/source/adding_benchmarks.mdx`](./docs/source/adding_benchmarks.mdx) and [`envhub.mdx`](./docs/source/envhub.mdx).
-
-### 8.2b Dockerfiles for benchmark eval
-
-Benchmark envs have native dependencies that are painful to install locally. The repo ships **pre-baked Dockerfiles** for each supported benchmark — use these to run `lerobot-eval` in a reproducible environment:
-
-| Benchmark   | Dockerfile                                                                             |
-| ----------- | -------------------------------------------------------------------------------------- |
-| LIBERO      | [`docker/Dockerfile.benchmark.libero`](./docker/Dockerfile.benchmark.libero)           |
-| LIBERO+     | [`docker/Dockerfile.benchmark.libero_plus`](./docker/Dockerfile.benchmark.libero_plus) |
-| MetaWorld   | [`docker/Dockerfile.benchmark.metaworld`](./docker/Dockerfile.benchmark.metaworld)     |
-| RoboCasa    | [`docker/Dockerfile.benchmark.robocasa`](./docker/Dockerfile.benchmark.robocasa)       |
-| RoboCerebra | [`docker/Dockerfile.benchmark.robocerebra`](./docker/Dockerfile.benchmark.robocerebra) |
-| RoboMME     | [`docker/Dockerfile.benchmark.robomme`](./docker/Dockerfile.benchmark.robomme)         |
-| RoboTwin    | [`docker/Dockerfile.benchmark.robotwin`](./docker/Dockerfile.benchmark.robotwin)       |
-| VLABench    | [`docker/Dockerfile.benchmark.vlabench`](./docker/Dockerfile.benchmark.vlabench)       |
-
-Build and run (adapt to your benchmark):
-
-```bash
-docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-bench-robomme .
-docker run --gpus all --rm -it \
-  -v $HOME/.cache/huggingface:/root/.cache/huggingface \
-  lerobot-bench-robomme \
-  lerobot-eval --policy.path=<your_policy> --env.type=<env> --eval.n_episodes=50
-```
-
-See [`docker/README.md`](./docker/README.md) for base-image details.
-
-### 8.3 Target success rates
-
-Single-task grasp-and-place with 50 clean episodes: ACT should reach **> 70% success** on the training configuration. Less → data problem (see §5), not model problem. Expect a drop when generalizing to new positions — scale episodes or diversity to recover.
-
---
-
-## 9. Further reading & resources
-
- **Getting started:** [`installation.mdx`](./docs/source/installation.mdx) · [`il_robots.mdx`](./docs/source/il_robots.mdx) · [What makes a good dataset](https://huggingface.co/blog/lerobot-datasets)
- **Per-policy docs:** browse [`docs/source/*.mdx`](./docs/source/) (policies, hardware, benchmarks, advanced training).
- **Community:** [Discord](https://discord.com/invite/s3KuuzsPFb) · [Hub `LeRobot` tag](https://huggingface.co/datasets?other=LeRobot) · [Dataset visualizer](https://huggingface.co/spaces/lerobot/visualize_dataset)
-
-> Keep this file current. If you learn a rule that would prevent a class of user mistakes, add it here and in [`AGENTS.md`](./AGENTS.md).
--- a/MANIFEST.in
+++ b/MANIFEST.in
@@ -1,4 +1,3 @@
 include src/lerobot/templates/lerobot_modelcard_template.md
-include src/lerobot/templates/lerobot_rewardmodel_modelcard_template.md
 include src/lerobot/datasets/card_template.md
 include src/lerobot/envs/metaworld_config.json
--- a/README.md
+++ b/README.md
@@ -109,7 +109,7 @@ lerobot-train \

 Similarly to the hardware, you can easily implement your own policy & leverage LeRobot's data collection, training, and visualization tools, and share your model to the HF Hub

-For detailed policy setup guides, see the [Policy Documentation](https://huggingface.co/docs/lerobot/bring_your_own_policies). For GPU/RAM requirements and expected training time per policy, see the [Compute Hardware Guide](https://huggingface.co/docs/lerobot/hardware_guide).
+For detailed policy setup guides, see the [Policy Documentation](https://huggingface.co/docs/lerobot/bring_your_own_policies).

 ## Inference & Evaluation

--- a/benchmarks/video/README.md
+++ b/benchmarks/video/README.md
@@ -0,0 +1,288 @@
+# Video benchmark
+
+## Questions
+
+What is the optimal trade-off between:
+
+- maximizing loading time with random access,
+- minimizing memory space on disk,
+- maximizing success rate of policies,
+- compatibility across devices/platforms for decoding videos (e.g. video players, web browsers).
+
+How to encode videos?
+
+- Which video codec (`-vcodec`) to use? h264, h265, AV1?
+- What pixel format to use (`-pix_fmt`)? `yuv444p` or `yuv420p`?
+- How much compression (`-crf`)? No compression with `0`, intermediate compression with `25` or extreme with `50+`?
+- Which frequency to chose for key frames (`-g`)? A key frame every `10` frames?
+
+How to decode videos?
+
+- Which `decoder`? `torchvision`, `torchaudio`, `ffmpegio`, `decord`, or `nvc`?
+- What scenarios to use for the requesting timestamps during benchmark? (`timestamps_mode`)
+
+## Variables
+
+**Image content & size**
+We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an apartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
+For these reasons, we run this benchmark on four representative datasets:
+
+- `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
+- `lerobot/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
+- `lerobot/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
+- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+
+Note: The datasets used for this benchmark need to be image datasets, not video datasets.
+
+**Data augmentations**
+We might revisit this benchmark and find better settings if we train our policies with various data augmentations to make them more robust (e.g. robust to color changes, compression, etc.).
+
+### Encoding parameters
+
+| parameter   | values                                                       |
+| ----------- | ------------------------------------------------------------ |
+| **vcodec**  | `libx264`, `libx265`, `libsvtav1`                            |
+| **pix_fmt** | `yuv444p`, `yuv420p`                                         |
+| **g**       | `1`, `2`, `3`, `4`, `5`, `6`, `10`, `15`, `20`, `40`, `None` |
+| **crf**     | `0`, `5`, `10`, `15`, `20`, `25`, `30`, `40`, `50`, `None`   |
+
+Note that `crf` value might be interpreted differently by various video codecs. In other words, the same value used with one codec doesn't necessarily translate into the same compression level with another codec. In fact, the default value (`None`) isn't the same amongst the different video codecs. Importantly, it is also the case for many other ffmpeg arguments like `g` which specifies the frequency of the key frames.
+
+For a comprehensive list and documentation of these parameters, see the ffmpeg documentation depending on the video codec used:
+
+- h264: https://trac.ffmpeg.org/wiki/Encode/H.264
+- h265: https://trac.ffmpeg.org/wiki/Encode/H.265
+- AV1: https://trac.ffmpeg.org/wiki/Encode/AV1
+
+### Decoding parameters
+
+**Decoder**
+We tested two video decoding backends from torchvision:
+
+- `pyav`
+- `video_reader` (requires to build torchvision from source)
+
+**Requested timestamps**
+Given the way video decoding works, once a keyframe has been loaded, the decoding of subsequent frames is fast.
+This of course is affected by the `-g` parameter during encoding, which specifies the frequency of the keyframes. Given our typical use cases in robotics policies which might request a few timestamps in different random places, we want to replicate these use cases with the following scenarios:
+
+- `1_frame`: 1 frame,
+- `2_frames`: 2 consecutive frames (e.g. `[t, t + 1 / fps]`),
+- `6_frames`: 6 consecutive frames (e.g. `[t + i / fps for i in range(6)]`)
+
+Note that this differs significantly from a typical use case like watching a movie, in which every frame is loaded sequentially from the beginning to the end and it's acceptable to have big values for `-g`.
+
+Additionally, because some policies might request single timestamps that are a few frames apart, we also have the following scenario:
+
+- `2_frames_4_space`: 2 frames with 4 consecutive frames of spacing in between (e.g `[t, t + 5 / fps]`),
+
+However, due to how video decoding is implemented with `pyav`, we don't have access to an accurate seek so in practice this scenario is essentially the same as `6_frames` since all 6 frames between `t` and `t + 5 / fps` will be decoded.
+
+## Metrics
+
+**Data compression ratio (lower is better)**
+`video_images_size_ratio` is the ratio of the memory space on disk taken by the encoded video over the memory space taken by the original images. For instance, `video_images_size_ratio=25%` means that the video takes 4 times less memory space on disk compared to the original images.
+
+**Loading time ratio (lower is better)**
+`video_images_load_time_ratio` is the ratio of the time it takes to decode frames from the video at a given timestamps over the time it takes to load the exact same original images. Lower is better. For instance, `video_images_load_time_ratio=200%` means that decoding from video is 2 times slower than loading the original images.
+
+**Average Mean Square Error (lower is better)**
+`avg_mse` is the average mean square error between each decoded frame and its corresponding original image over all requested timestamps, and also divided by the number of pixels in the image to be comparable when switching to different image sizes.
+
+**Average Peak Signal to Noise Ratio (higher is better)**
+`avg_psnr` measures the ratio between the maximum possible power of a signal and the power of corrupting noise that affects the fidelity of its representation. Higher PSNR indicates better quality.
+
+**Average Structural Similarity Index Measure (higher is better)**
+`avg_ssim` evaluates the perceived quality of images by comparing luminance, contrast, and structure. SSIM values range from -1 to 1, where 1 indicates perfect similarity.
+
+One aspect that can't be measured here with those metrics is the compatibility of the encoding across platforms, in particular on web browser, for visualization purposes.
+h264, h265 and AV1 are all commonly used codecs and should not pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
+
+- `yuv420p` is more widely supported across various platforms, including web browsers.
+- `yuv444p` offers higher color fidelity but might not be supported as broadly.
+
+<!-- **Loss of a pretrained policy (higher is better)** (not available)
+`loss_pretrained` is the result of evaluating with the selected encoding/decoding settings a policy pretrained on original images. It is easier to understand than `avg_l2_error`.
+
+**Success rate after retraining (higher is better)** (not available)
+`success_rate` is the result of training and evaluating a policy with the selected encoding/decoding settings. It is the most difficult metric to get but also the very best. -->
+
+## How the benchmark works
+
+The benchmark evaluates both encoding and decoding of video frames on the first episode of each dataset.
+
+**Encoding:** for each `vcodec` and `pix_fmt` pair, we use a default value for `g` and `crf` upon which we change a single value (either `g` or `crf`) to one of the specified values (we don't test every combination of those as this would be computationally too heavy).
+This gives a unique set of encoding parameters which is used to encode the episode.
+
+**Decoding:** Then, for each of those unique encodings, we iterate through every combination of the decoding parameters `backend` and `timestamps_mode`. For each of them, we record the metrics of a number of samples (given by `--num-samples`). This is parallelized for efficiency and the number of processes can be controlled with `--num-workers`. Ideally, it's best to have a `--num-samples` that is divisible by `--num-workers`.
+
+Intermediate results saved for each `vcodec` and `pix_fmt` combination in csv tables.
+These are then all concatenated to a single table ready for analysis.
+
+## Caveats
+
+We tried to measure the most impactful parameters for both encoding and decoding. However, for computational reasons we can't test out every combination.
+
+Additional encoding parameters exist that are not included in this benchmark. In particular:
+
+- `-preset` which allows for selecting encoding presets. This represents a collection of options that will provide a certain encoding speed to compression ratio. By leaving this parameter unspecified, it is considered to be `medium` for libx264 and libx265 and `8` for libsvtav1.
+- `-tune` which allows to optimize the encoding for certain aspects (e.g. film quality, fast decoding, etc.).
+
+See the documentation mentioned above for more detailed info on these settings and for a more comprehensive list of other parameters.
+
+Similarly on the decoding side, other decoders exist but are not implemented in our current benchmark. To name a few:
+
+- `torchaudio`
+- `ffmpegio`
+- `decord`
+- `nvc`
+
+Note as well that since we are mostly interested in the performance at decoding time (also because encoding is done only once before uploading a dataset), we did not measure encoding times nor have any metrics regarding encoding.
+However, besides the necessity to build ffmpeg from source, encoding did not pose any issue and it didn't take a significant amount of time during this benchmark.
+
+## Install
+
+Building ffmpeg from source is required to include libx265 and libaom/libsvtav1 (av1) video codecs ([compilation guide](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu)).
+
+**Note:** While you still need to build torchvision with a conda-installed `ffmpeg<4.3` to use the `video_reader` decoder (as described in [#220](https://github.com/huggingface/lerobot/pull/220)), you also need another version which is custom-built with all the video codecs for encoding. For the script to then use that version, you can prepend the command above with `PATH="$HOME/bin:$PATH"`, which is where ffmpeg should be built.
+
+## Adding a video decoder
+
+Right now, we're only benchmarking the two video decoder available with torchvision: `pyav` and `video_reader`.
+You can easily add a new decoder to benchmark by adding it to this function in the script:
+
+```diff
+def decode_video_frames(
+    video_path: str,
+    timestamps: list[float],
+    tolerance_s: float,
+    backend: str,
+) -> torch.Tensor:
+    if backend in ["pyav", "video_reader"]:
+        return decode_video_frames_torchvision(
+            video_path, timestamps, tolerance_s, backend
+        )
+    elif backend == ["your_decoder"]:
+        return your_decoder_function(
+            video_path, timestamps, tolerance_s, backend
+        )
+    else:
+        raise NotImplementedError(backend)
+```
+
+## Example
+
+For a quick run, you can try these parameters:
+
+```bash
+python benchmark/video/run_video_benchmark.py \
+    --output-dir outputs/video_benchmark \
+    --repo-ids \
+        lerobot/pusht_image \
+        lerobot/aloha_mobile_shrimp_image \
+    --vcodec libx264 libx265 \
+    --pix-fmt yuv444p yuv420p \
+    --g 2 20 None \
+    --crf 10 40 None \
+    --timestamps-modes 1_frame 2_frames \
+    --backends pyav video_reader \
+    --num-samples 5 \
+    --num-workers 5 \
+    --save-frames 0
+```
+
+## Results
+
+### Reproduce
+
+We ran the benchmark with the following parameters:
+
+```bash
+# h264 and h265 encodings
+python benchmark/video/run_video_benchmark.py \
+    --output-dir outputs/video_benchmark \
+    --repo-ids \
+        lerobot/pusht_image \
+        lerobot/aloha_mobile_shrimp_image \
+        lerobot/paris_street \
+        lerobot/kitchen \
+    --vcodec libx264 libx265 \
+    --pix-fmt yuv444p yuv420p \
+    --g 1 2 3 4 5 6 10 15 20 40 None \
+    --crf 0 5 10 15 20 25 30 40 50 None \
+    --timestamps-modes 1_frame 2_frames 6_frames \
+    --backends pyav video_reader \
+    --num-samples 50 \
+    --num-workers 5 \
+    --save-frames 1
+
+# av1 encoding (only compatible with yuv420p and pyav decoder)
+python benchmark/video/run_video_benchmark.py \
+    --output-dir outputs/video_benchmark \
+    --repo-ids \
+        lerobot/pusht_image \
+        lerobot/aloha_mobile_shrimp_image \
+        lerobot/paris_street \
+        lerobot/kitchen \
+    --vcodec libsvtav1 \
+    --pix-fmt yuv420p \
+    --g 1 2 3 4 5 6 10 15 20 40 None \
+    --crf 0 5 10 15 20 25 30 40 50 None \
+    --timestamps-modes 1_frame 2_frames 6_frames \
+    --backends pyav \
+    --num-samples 50 \
+    --num-workers 5 \
+    --save-frames 1
+```
+
+The full results are available [here](https://docs.google.com/spreadsheets/d/1OYJB43Qu8fC26k_OyoMFgGBBKfQRCi4BIuYitQnq3sw/edit?usp=sharing)
+
+### Parameters selected for LeRobotDataset
+
+Considering these results, we chose what we think is the best set of encoding parameter:
+
+- vcodec: `libsvtav1`
+- pix-fmt: `yuv420p`
+- g: `2`
+- crf: `30`
+
+Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_reader` does not support it (and `pyav` doesn't require a custom build of `torchvision`).
+
+### Summary
+
+These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`
+
+| video_images_size_ratio           | vcodec     | pix_fmt |           |           |           |
+| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
+|                                   | libx264    |         | libx265   |           | libsvtav1 |
+| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| lerobot/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
+| lerobot/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
+| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+
+| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
+| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
+|                                   | libx264 |         | libx265  |         | libsvtav1 |
+| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| lerobot/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
+| lerobot/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
+| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+
+|                                   |          | vcodec   | pix_fmt      |          |           |              |
+| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
+|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
+| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| lerobot/pusht_image               | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
+|                                   | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
+|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+| lerobot/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
+|                                   | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
+|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+| lerobot/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
+|                                   | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
+|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+| lerobot/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
+|                                   | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
+|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
--- a/benchmarks/video/run_video_benchmark.py
+++ b/benchmarks/video/run_video_benchmark.py
@@ -0,0 +1,488 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Assess the performance of video decoding in various configurations.
+
+This script will benchmark different video encoding and decoding parameters.
+See the provided README.md or run `python benchmark/video/run_video_benchmark.py --help` for usage info.
+"""
+
+import argparse
+import datetime as dt
+import itertools
+import random
+import shutil
+from collections import OrderedDict
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+from threading import Lock
+
+import einops
+import numpy as np
+import pandas as pd
+import PIL
+import torch
+from skimage.metrics import mean_squared_error, peak_signal_noise_ratio, structural_similarity
+from tqdm import tqdm
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.video_utils import (
+    decode_video_frames,
+    encode_video_frames,
+)
+from lerobot.utils.constants import OBS_IMAGE
+from lerobot.utils.utils import TimerManager
+
+BASE_ENCODING = OrderedDict(
+    [
+        ("vcodec", "libx264"),
+        ("pix_fmt", "yuv444p"),
+        ("g", 2),
+        ("crf", None),
+        # TODO(aliberts): Add fastdecode
+        # ("fastdecode", 0),
+    ]
+)
+
+
+# TODO(rcadene, aliberts): move to `utils.py` folder when we want to refactor
+def parse_int_or_none(value) -> int | None:
+    if value.lower() == "none":
+        return None
+    try:
+        return int(value)
+    except ValueError as e:
+        raise argparse.ArgumentTypeError(f"Invalid int or None: {value}") from e
+
+
+def check_datasets_formats(repo_ids: list) -> None:
+    for repo_id in repo_ids:
+        dataset = LeRobotDataset(repo_id)
+        if len(dataset.meta.video_keys) > 0:
+            raise ValueError(
+                f"Use only image dataset for running this benchmark. Video dataset provided: {repo_id}"
+            )
+
+
+def get_directory_size(directory: Path) -> int:
+    total_size = 0
+    for item in directory.rglob("*"):
+        if item.is_file():
+            total_size += item.stat().st_size
+    return total_size
+
+
+def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> torch.Tensor:
+    frames = []
+    for ts in timestamps:
+        idx = int(ts * fps)
+        frame = PIL.Image.open(imgs_dir / f"frame-{idx:06d}.png")
+        frame = torch.from_numpy(np.array(frame))
+        frame = frame.type(torch.float32) / 255
+        frame = einops.rearrange(frame, "h w c -> c h w")
+        frames.append(frame)
+    return torch.stack(frames)
+
+
+def save_decoded_frames(
+    imgs_dir: Path, save_dir: Path, frames: torch.Tensor, timestamps: list[float], fps: int
+) -> None:
+    if save_dir.exists() and len(list(save_dir.glob("frame-*.png"))) == len(timestamps):
+        return
+
+    save_dir.mkdir(parents=True, exist_ok=True)
+    for i, ts in enumerate(timestamps):
+        idx = int(ts * fps)
+        frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
+        PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame-{idx:06d}_decoded.png")
+        shutil.copyfile(imgs_dir / f"frame-{idx:06d}.png", save_dir / f"frame-{idx:06d}_original.png")
+
+
+def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
+    episode_index = 0
+    ep_num_images = dataset.meta.episodes["length"][episode_index]
+    if imgs_dir.exists() and len(list(imgs_dir.glob("frame-*.png"))) == ep_num_images:
+        return
+
+    imgs_dir.mkdir(parents=True, exist_ok=True)
+    hf_dataset = dataset.hf_dataset.with_format(None)
+
+    # We only save images from the first camera
+    img_keys = [key for key in hf_dataset.features if key.startswith(OBS_IMAGE)]
+    imgs_dataset = hf_dataset.select_columns(img_keys[0])
+
+    for i, item in enumerate(
+        tqdm(imgs_dataset, desc=f"saving {dataset.repo_id} first episode images", leave=False)
+    ):
+        img = item[img_keys[0]]
+        img.save(str(imgs_dir / f"frame-{i:06d}.png"), quality=100)
+
+        if i >= ep_num_images - 1:
+            break
+
+
+def sample_timestamps(timestamps_mode: str, ep_num_images: int, fps: int) -> list[float]:
+    # Start at 5 to allow for 2_frames_4_space and 6_frames
+    idx = random.randint(5, ep_num_images - 1)
+    match timestamps_mode:
+        case "1_frame":
+            frame_indexes = [idx]
+        case "2_frames":
+            frame_indexes = [idx - 1, idx]
+        case "2_frames_4_space":
+            frame_indexes = [idx - 5, idx]
+        case "6_frames":
+            frame_indexes = [idx - i for i in range(6)][::-1]
+        case _:
+            raise ValueError(timestamps_mode)
+
+    return [idx / fps for idx in frame_indexes]
+
+
+def benchmark_decoding(
+    imgs_dir: Path,
+    video_path: Path,
+    timestamps_mode: str,
+    backend: str,
+    ep_num_images: int,
+    fps: int,
+    num_samples: int = 50,
+    num_workers: int = 4,
+    save_frames: bool = False,
+) -> dict:
+    def process_sample(sample: int, lock: Lock):
+        time_benchmark = TimerManager(log=False)
+        timestamps = sample_timestamps(timestamps_mode, ep_num_images, fps)
+        num_frames = len(timestamps)
+        result = {
+            "psnr_values": [],
+            "ssim_values": [],
+            "mse_values": [],
+        }
+
+        with time_benchmark, lock:
+            frames = decode_video_frames(video_path, timestamps=timestamps, tolerance_s=5e-1, backend=backend)
+        result["load_time_video_ms"] = (time_benchmark.last * 1000) / num_frames
+
+        with time_benchmark:
+            original_frames = load_original_frames(imgs_dir, timestamps, fps)
+        result["load_time_images_ms"] = (time_benchmark.last * 1000) / num_frames
+
+        frames_np, original_frames_np = frames.numpy(), original_frames.numpy()
+        for i in range(num_frames):
+            result["mse_values"].append(mean_squared_error(original_frames_np[i], frames_np[i]))
+            result["psnr_values"].append(
+                peak_signal_noise_ratio(original_frames_np[i], frames_np[i], data_range=1.0)
+            )
+            result["ssim_values"].append(
+                structural_similarity(original_frames_np[i], frames_np[i], data_range=1.0, channel_axis=0)
+            )
+
+        if save_frames and sample == 0:
+            save_dir = video_path.with_suffix("") / f"{timestamps_mode}_{backend}"
+            save_decoded_frames(imgs_dir, save_dir, frames, timestamps, fps)
+
+        return result
+
+    load_times_video_ms = []
+    load_times_images_ms = []
+    mse_values = []
+    psnr_values = []
+    ssim_values = []
+
+    # A sample is a single set of decoded frames specified by timestamps_mode (e.g. a single frame, 2 frames, etc.).
+    # For each sample, we record metrics (loading time and quality metrics) which are then averaged over all samples.
+    # As these samples are independent, we run them in parallel threads to speed up the benchmark.
+    # Use a single shared lock for all worker threads
+    shared_lock = Lock()
+    with ThreadPoolExecutor(max_workers=num_workers) as executor:
+        futures = [executor.submit(process_sample, i, shared_lock) for i in range(num_samples)]
+        for future in tqdm(as_completed(futures), total=num_samples, desc="samples", leave=False):
+            result = future.result()
+            load_times_video_ms.append(result["load_time_video_ms"])
+            load_times_images_ms.append(result["load_time_images_ms"])
+            psnr_values.extend(result["psnr_values"])
+            ssim_values.extend(result["ssim_values"])
+            mse_values.extend(result["mse_values"])
+
+    avg_load_time_video_ms = float(np.array(load_times_video_ms).mean())
+    avg_load_time_images_ms = float(np.array(load_times_images_ms).mean())
+    video_images_load_time_ratio = avg_load_time_video_ms / avg_load_time_images_ms
+
+    return {
+        "avg_load_time_video_ms": avg_load_time_video_ms,
+        "avg_load_time_images_ms": avg_load_time_images_ms,
+        "video_images_load_time_ratio": video_images_load_time_ratio,
+        "avg_mse": float(np.mean(mse_values)),
+        "avg_psnr": float(np.mean(psnr_values)),
+        "avg_ssim": float(np.mean(ssim_values)),
+    }
+
+
+def benchmark_encoding_decoding(
+    dataset: LeRobotDataset,
+    video_path: Path,
+    imgs_dir: Path,
+    encoding_cfg: dict,
+    decoding_cfg: dict,
+    num_samples: int,
+    num_workers: int,
+    save_frames: bool,
+    overwrite: bool = False,
+    seed: int = 1337,
+) -> list[dict]:
+    fps = dataset.fps
+
+    if overwrite or not video_path.is_file():
+        tqdm.write(f"encoding {video_path}")
+        encode_video_frames(
+            imgs_dir=imgs_dir,
+            video_path=video_path,
+            fps=fps,
+            vcodec=encoding_cfg["vcodec"],
+            pix_fmt=encoding_cfg["pix_fmt"],
+            g=encoding_cfg.get("g"),
+            crf=encoding_cfg.get("crf"),
+            # fast_decode=encoding_cfg.get("fastdecode"),
+            overwrite=True,
+        )
+
+    episode_index = 0
+    ep_num_images = dataset.meta.episodes["length"][episode_index]
+    width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
+    num_pixels = width * height
+    video_size_bytes = video_path.stat().st_size
+    images_size_bytes = get_directory_size(imgs_dir)
+    video_images_size_ratio = video_size_bytes / images_size_bytes
+
+    random.seed(seed)
+    benchmark_table = []
+    for timestamps_mode in tqdm(
+        decoding_cfg["timestamps_modes"], desc="decodings (timestamps_modes)", leave=False
+    ):
+        for backend in tqdm(decoding_cfg["backends"], desc="decodings (backends)", leave=False):
+            benchmark_row = benchmark_decoding(
+                imgs_dir,
+                video_path,
+                timestamps_mode,
+                backend,
+                ep_num_images,
+                fps,
+                num_samples,
+                num_workers,
+                save_frames,
+            )
+            benchmark_row.update(
+                **{
+                    "repo_id": dataset.repo_id,
+                    "resolution": f"{width} x {height}",
+                    "num_pixels": num_pixels,
+                    "video_size_bytes": video_size_bytes,
+                    "images_size_bytes": images_size_bytes,
+                    "video_images_size_ratio": video_images_size_ratio,
+                    "timestamps_mode": timestamps_mode,
+                    "backend": backend,
+                },
+                **encoding_cfg,
+            )
+            benchmark_table.append(benchmark_row)
+
+    return benchmark_table
+
+
+def main(
+    output_dir: Path,
+    repo_ids: list[str],
+    vcodec: list[str],
+    pix_fmt: list[str],
+    g: list[int],
+    crf: list[int],
+    # fastdecode: list[int],
+    timestamps_modes: list[str],
+    backends: list[str],
+    num_samples: int,
+    num_workers: int,
+    save_frames: bool,
+):
+    check_datasets_formats(repo_ids)
+    encoding_benchmarks = {
+        "g": g,
+        "crf": crf,
+        # "fastdecode": fastdecode,
+    }
+    decoding_benchmarks = {
+        "timestamps_modes": timestamps_modes,
+        "backends": backends,
+    }
+    headers = ["repo_id", "resolution", "num_pixels"]
+    headers += list(BASE_ENCODING.keys())
+    headers += [
+        "timestamps_mode",
+        "backend",
+        "video_size_bytes",
+        "images_size_bytes",
+        "video_images_size_ratio",
+        "avg_load_time_video_ms",
+        "avg_load_time_images_ms",
+        "video_images_load_time_ratio",
+        "avg_mse",
+        "avg_psnr",
+        "avg_ssim",
+    ]
+    file_paths = []
+    for video_codec in tqdm(vcodec, desc="encodings (vcodec)"):
+        for pixel_format in tqdm(pix_fmt, desc="encodings (pix_fmt)", leave=False):
+            benchmark_table = []
+            for repo_id in tqdm(repo_ids, desc="encodings (datasets)", leave=False):
+                dataset = LeRobotDataset(repo_id)
+                imgs_dir = output_dir / "images" / dataset.repo_id.replace("/", "_")
+                # We only use the first episode
+                save_first_episode(imgs_dir, dataset)
+                for duet in [
+                    dict(zip(encoding_benchmarks.keys(), unique_combination, strict=False))
+                    for unique_combination in itertools.product(*encoding_benchmarks.values())
+                ]:
+                    encoding_cfg = BASE_ENCODING.copy()
+                    encoding_cfg["vcodec"] = video_codec
+                    encoding_cfg["pix_fmt"] = pixel_format
+                    for key, value in duet.items():
+                        encoding_cfg[key] = value
+                    args_path = Path("_".join(str(value) for value in encoding_cfg.values()))
+                    video_path = output_dir / "videos" / args_path / f"{repo_id.replace('/', '_')}.mp4"
+                    benchmark_table += benchmark_encoding_decoding(
+                        dataset,
+                        video_path,
+                        imgs_dir,
+                        encoding_cfg,
+                        decoding_benchmarks,
+                        num_samples,
+                        num_workers,
+                        save_frames,
+                    )
+
+            # Save intermediate results
+            benchmark_df = pd.DataFrame(benchmark_table, columns=headers)
+            now = dt.datetime.now()
+            csv_path = (
+                output_dir
+                / f"{now:%Y-%m-%d}_{now:%H-%M-%S}_{video_codec}_{pixel_format}_{num_samples}-samples.csv"
+            )
+            benchmark_df.to_csv(csv_path, header=True, index=False)
+            file_paths.append(csv_path)
+            del benchmark_df
+
+    # Concatenate all results
+    df_list = [pd.read_csv(csv_path) for csv_path in file_paths]
+    concatenated_df = pd.concat(df_list, ignore_index=True)
+    concatenated_path = output_dir / f"{now:%Y-%m-%d}_{now:%H-%M-%S}_all_{num_samples}-samples.csv"
+    concatenated_df.to_csv(concatenated_path, header=True, index=False)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=Path("outputs/video_benchmark"),
+        help="Directory where the video benchmark outputs are written.",
+    )
+    parser.add_argument(
+        "--repo-ids",
+        type=str,
+        nargs="*",
+        default=[
+            "lerobot/pusht_image",
+            "lerobot/aloha_mobile_shrimp_image",
+            "lerobot/paris_street",
+            "lerobot/kitchen",
+        ],
+        help="Datasets repo-ids to test against. First episodes only are used. Must be images.",
+    )
+    parser.add_argument(
+        "--vcodec",
+        type=str,
+        nargs="*",
+        default=["h264", "hevc", "libsvtav1"],
+        help="Video codecs to be tested",
+    )
+    parser.add_argument(
+        "--pix-fmt",
+        type=str,
+        nargs="*",
+        default=["yuv444p", "yuv420p"],
+        help="Pixel formats (chroma subsampling) to be tested",
+    )
+    parser.add_argument(
+        "--g",
+        type=parse_int_or_none,
+        nargs="*",
+        default=[1, 2, 3, 4, 5, 6, 10, 15, 20, 40, 100, None],
+        help="Group of pictures sizes to be tested.",
+    )
+    parser.add_argument(
+        "--crf",
+        type=parse_int_or_none,
+        nargs="*",
+        default=[0, 5, 10, 15, 20, 25, 30, 40, 50, None],
+        help="Constant rate factors to be tested.",
+    )
+    # parser.add_argument(
+    #     "--fastdecode",
+    #     type=int,
+    #     nargs="*",
+    #     default=[0, 1],
+    #     help="Use the fastdecode tuning option. 0 disables it. "
+    #         "For libx264 and libx265/hevc, only 1 is possible. "
+    #         "For libsvtav1, 1, 2 or 3 are possible values with a higher number meaning a faster decoding optimization",
+    # )
+    parser.add_argument(
+        "--timestamps-modes",
+        type=str,
+        nargs="*",
+        default=[
+            "1_frame",
+            "2_frames",
+            "2_frames_4_space",
+            "6_frames",
+        ],
+        help="Timestamps scenarios to be tested.",
+    )
+    parser.add_argument(
+        "--backends",
+        type=str,
+        nargs="*",
+        default=["torchcodec", "pyav"],
+        help="Torchvision decoding backend to be tested.",
+    )
+    parser.add_argument(
+        "--num-samples",
+        type=int,
+        default=50,
+        help="Number of samples for each encoding x decoding config.",
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=10,
+        help="Number of processes for parallelized sample processing.",
+    )
+    parser.add_argument(
+        "--save-frames",
+        type=int,
+        default=0,
+        help="Whether to save decoded frames or not. Enter a non-zero number for true.",
+    )
+    args = parser.parse_args()
+    main(**vars(args))
--- a/docker/Dockerfile.benchmark.libero_plus
+++ b/docker/Dockerfile.benchmark.libero_plus
@@ -1,84 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for LIBERO-plus integration tests.
-# Extends the nightly GPU image (which has lerobot[all]) with the LIBERO-plus
-# fork source + its 6.4 GB perturbation assets.
-#
-# Build:  docker build -f docker/Dockerfile.benchmark.libero_plus -t lerobot-benchmark-libero-plus .
-# Run:    docker run --gpus all --rm lerobot-benchmark-libero-plus lerobot-eval ...
-
-FROM huggingface/lerobot-gpu:latest
-ENV MUJOCO_GL=egl
-
-# unzip for the 6.4 GB assets.zip; the rest are LIBERO-plus build-time extras
-# (wand / ImageMagick / fontconfig) not in the nightly base.
-USER root
-RUN apt-get update \
-    && apt-get install -y --no-install-recommends \
-         unzip libexpat1 libfontconfig1-dev libmagickwand-dev \
-    && apt-get clean && rm -rf /var/lib/apt/lists/*
-USER user_lerobot
-
-# robosuite==1.4.1 is mandatory (the fork uses `single_arm_env` removed in
-# v1.5+). The rest are LIBERO-plus runtime deps pulled from its setup.py.
-# We install these explicitly instead of via the [libero_plus] extra because
-# the extra's `libero @ git+...` dep installs as a namespace package and then
-# clone and PYTHONPATH-override it below.
-RUN uv pip install --no-cache \
-        "robosuite==1.4.1" \
-        "bddl==1.0.1" \
-        "easydict==1.13" \
-        "mujoco==3.7.0" \
-        "matplotlib==3.10.8" \
-        "Wand==0.6.13" \
-        "scikit-image==0.25.2" \
-        "gym==0.26.2"
-
-# Clone LIBERO-plus and make it importable as `libero`. The nightly base has
-# hf-libero (10 tasks) preinstalled via lerobot[libero]; uninstall it so
-# Python resolves `import libero` to the 2402-task LIBERO-plus module instead.
-# Pinned to the current upstream main SHA so benchmark builds stay reproducible.
-ARG LIBERO_PLUS_SHA=4976dc3
-ENV LIBERO_PLUS_ROOT=/home/user_lerobot/libero-plus/libero/libero
-RUN git clone https://github.com/sylvestf/LIBERO-plus.git /home/user_lerobot/libero-plus \
-    && git -C /home/user_lerobot/libero-plus checkout ${LIBERO_PLUS_SHA} \
-    && cd /home/user_lerobot/libero-plus && uv pip install --no-cache --no-deps -e "." \
-    && (uv pip uninstall hf-libero 2>/dev/null || true)
-ENV PYTHONPATH="/home/user_lerobot/libero-plus:${PYTHONPATH}"
-
-# Perturbation textures/scenes: bddl_base_domain.py resolves XMLs via
-# DIR_PATH/../assets (package-relative, ignoring ~/.libero/config.yaml). All
-# 2402 tasks reference files that ship only in Sylvest/LIBERO-plus's
-# assets.zip (6.4 GB) under a deep author-internal prefix — extract and
-# flatten it under ${LIBERO_PLUS_ROOT}/assets.
-RUN python -c "\
-from huggingface_hub import hf_hub_download; \
-hf_hub_download(repo_id='Sylvest/LIBERO-plus', repo_type='dataset', \
-                filename='assets.zip', local_dir='/tmp/libero-plus-dl')" \
-    && unzip -q /tmp/libero-plus-dl/assets.zip -d /tmp/libero-plus-dl/extract \
-    && ASSETS_DIR=$(find /tmp/libero-plus-dl/extract -type d -name assets | head -1) \
-    && mv "${ASSETS_DIR}" ${LIBERO_PLUS_ROOT}/assets \
-    && rm -rf /tmp/libero-plus-dl
-
-# Point ~/.libero/config.yaml at the clone so LIBERO-plus's imports are
-# non-interactive (it calls input() when the config is missing).
-RUN mkdir -p /home/user_lerobot/.libero \
-    && printf "assets: ${LIBERO_PLUS_ROOT}/assets\nbddl_files: ${LIBERO_PLUS_ROOT}/bddl_files\ndatasets: ${LIBERO_PLUS_ROOT}/../datasets\ninit_states: ${LIBERO_PLUS_ROOT}/init_files\n" \
-       > /home/user_lerobot/.libero/config.yaml
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-CMD ["/bin/bash"]
--- a/docker/Dockerfile.benchmark.robocasa
+++ b/docker/Dockerfile.benchmark.robocasa
@@ -1,71 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for RoboCasa365 integration tests.
-# Extends the nightly GPU image (which already has all extras installed)
-# with the PR's source code and RoboCasa-specific asset setup.
-#
-# Build:  docker build -f docker/Dockerfile.benchmark.robocasa -t lerobot-benchmark-robocasa .
-# Run:    docker run --gpus all --rm lerobot-benchmark-robocasa lerobot-eval ...
-
-FROM huggingface/lerobot-gpu:latest
-
-# Install robocasa + robosuite as editable clones. pip-installing from git
-# omits data files like robocasa/models/assets/box_links/box_links_assets.json
-# (not declared in package_data), which download_kitchen_assets needs at import.
-#
-# `--no-deps` on robocasa is deliberate: its setup.py pins `lerobot==0.3.3`
-# in install_requires, which would shadow the editable lerobot baked into
-# this image. We install robocasa's actual runtime deps explicitly instead.
-# Pinned SHAs for reproducible benchmark runs. Bump when you need an
-# upstream fix; don't rely on `main`/`master` drift.
-ARG ROBOCASA_SHA=56e355ccc64389dfc1b8a61a33b9127b975ba681
-ARG ROBOSUITE_SHA=aaa8b9b214ce8e77e82926d677b4d61d55e577ab
-RUN git clone https://github.com/robocasa/robocasa.git ~/robocasa && \
-    git -C ~/robocasa checkout ${ROBOCASA_SHA} && \
-    git clone https://github.com/ARISE-Initiative/robosuite.git ~/robosuite && \
-    git -C ~/robosuite checkout ${ROBOSUITE_SHA} && \
-    uv pip install --no-cache -e ~/robocasa --no-deps && \
-    uv pip install --no-cache -e ~/robosuite && \
-    uv pip install --no-cache \
-      "numpy==2.2.5" "numba==0.61.2" "scipy==1.15.3" "mujoco==3.3.1" \
-      "pygame==2.6.1" "Pillow==12.2.0" "opencv-python==4.13.0.92" \
-      "pyyaml==6.0.3" "pynput==1.8.1" "tqdm==4.67.3" "termcolor==3.3.0" \
-      "imageio==2.37.3" "h5py==3.16.0" "lxml==6.0.4" "hidapi==0.14.0.post4" \
-      "tianshou==0.4.10" "gymnasium==1.2.3"
-
-# Set up robocasa macros and download kitchen assets. We need:
-#   - tex              : base environment textures
-#   - tex_generative   : AI-generated textures; kitchen fixture XMLs embed
-#                        refs to generative_textures/wall/tex*.png
-#                        unconditionally, so MjModel.from_xml_string fails
-#                        at reset time without them (even if the env is
-#                        constructed with generative_textures=None).
-#   - fixtures_lw      : lightwheel kitchen fixtures (fridge, counters...)
-#   - objs_lw          : lightwheel object meshes (stools, misc props)
-# We skip the objaverse/aigen object packs (~30GB combined) by pairing
-# this with --env.obj_registries=["lightwheel"] on the lerobot side.
-# The download script prompts interactively, so pipe 'y' to auto-accept.
-RUN python -m robocasa.scripts.setup_macros && \
-    yes y | python -m robocasa.scripts.download_kitchen_assets \
-      --type tex tex_generative fixtures_lw objs_lw
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-# Re-install lerobot editably so the new source (with RoboCasaEnv registration)
-# replaces the stale package baked into the nightly image.
-RUN uv pip install --no-cache --no-deps -e .
-
-CMD ["/bin/bash"]
--- a/docker/Dockerfile.benchmark.robocerebra
+++ b/docker/Dockerfile.benchmark.robocerebra
@@ -1,43 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for RoboCerebra integration tests.
-# RoboCerebra reuses LIBERO's simulator (libero_10 suite) with a different
-# rename_map, so this image is identical to the LIBERO benchmark image —
-# extends the nightly GPU base with LIBERO assets + the PR's source code.
-#
-# Build:  docker build -f docker/Dockerfile.benchmark.robocerebra -t lerobot-benchmark-robocerebra .
-# Run:    docker run --gpus all --rm lerobot-benchmark-robocerebra lerobot-eval ...
-
-FROM huggingface/lerobot-gpu:latest
-
-# Pre-download lerobot/libero-assets from HF Hub so nothing is fetched at
-# runtime (which times out on CI). Point the libero config at the cached path.
-# libero/libero/__init__.py calls input() when ~/.libero/config.yaml is missing,
-# so we write the config before any libero import can happen.
-RUN LIBERO_DIR=$(python -c \
-      "import importlib.util, os; s=importlib.util.find_spec('libero'); \
-       print(os.path.join(os.path.dirname(s.origin), 'libero'))") && \
-    mkdir -p /home/user_lerobot/.libero && \
-    python -c "\
-from huggingface_hub import snapshot_download; \
-snapshot_download(repo_id='lerobot/libero-assets', repo_type='dataset', \
-                  local_dir='/home/user_lerobot/.libero/assets')" && \
-    printf "assets: /home/user_lerobot/.libero/assets\nbddl_files: ${LIBERO_DIR}/bddl_files\ndatasets: ${LIBERO_DIR}/../datasets\ninit_states: ${LIBERO_DIR}/init_files\n" \
-    > /home/user_lerobot/.libero/config.yaml
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-CMD ["/bin/bash"]
--- a/docker/Dockerfile.benchmark.robomme
+++ b/docker/Dockerfile.benchmark.robomme
@@ -1,56 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for RoboMME integration tests.
-# Extends the nightly GPU image (which has lerobot[all]) with Vulkan system
-# libs for ManiSkill/SAPIEN and the robomme extra. robomme isn't in [all]
-# because mani-skill hard-pins gymnasium==0.29.1 and numpy<2.0.0 which
-# conflict with lerobot's defaults; both are safe at runtime:
-#   - gymnasium 0.29.x has the same 5-tuple step() API as 1.x (since 0.26)
-#   - numpy 1.26.4 is API-compatible with lerobot's actual usage.
-#
-# Build:  docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-benchmark-robomme .
-# Run:    docker run --gpus all --rm lerobot-benchmark-robomme lerobot-eval ...
-
-FROM huggingface/lerobot-gpu:latest
-
-# NVIDIA Container Toolkit: expose Vulkan driver capability for headless rendering.
-ENV NVIDIA_DRIVER_CAPABILITIES=all \
-    VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json
-
-# ManiSkill/SAPIEN's renderer needs Vulkan, which isn't in the base image.
-USER root
-RUN apt-get update \
-    && apt-get install -y --no-install-recommends \
-         libvulkan1 libvulkan-dev mesa-vulkan-drivers \
-    && mkdir -p /usr/share/vulkan/icd.d \
-    && echo '{"file_format_version":"1.0.0","ICD":{"library_path":"libGLX_nvidia.so.0","api_version":"1.3.0"}}' \
-       > /usr/share/vulkan/icd.d/nvidia_icd.json \
-    && apt-get clean && rm -rf /var/lib/apt/lists/*
-USER user_lerobot
-
-# Install smolvla + av-dep via the PR's pyproject, then layer robomme on top
-# with gymnasium/numpy overrides. robomme isn't a pyproject extra because its
-# mani-skill pin conflicts with lerobot's base numpy>=2 (see pyproject.toml).
-COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
-RUN printf 'gymnasium==0.29.1\nnumpy==1.26.4\n' > /tmp/robomme_override.txt \
-    && uv pip install --no-cache --override /tmp/robomme_override.txt \
-         -e ".[smolvla,av-dep]" \
-         "robomme @ git+https://github.com/RoboMME/robomme_benchmark.git@main" \
-    && python -c "import robomme; print('robomme import OK')"
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-CMD ["/bin/bash"]
--- a/docker/Dockerfile.benchmark.robotwin
+++ b/docker/Dockerfile.benchmark.robotwin
@@ -1,138 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for RoboTwin 2.0 integration tests.
-# Extends the nightly GPU image with the RoboTwin simulator stack:
-#   sapien/mplib/pytorch3d + NVlabs CuRobo + embodiments.zip + objects.zip
-# (~3.96 GB of assets; background_texture.zip ~11 GB skipped for smoke eval).
-#
-# Build: docker build -f docker/Dockerfile.benchmark.robotwin -t lerobot-benchmark-robotwin .
-# Run:   docker run --gpus all --rm lerobot-benchmark-robotwin \
-#            lerobot-eval --env.type=robotwin --env.task=beat_block_hammer ...
-
-FROM huggingface/lerobot-gpu:latest
-
-ENV NVIDIA_DRIVER_CAPABILITIES=all \
-    VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json \
-    ROBOTWIN_ROOT=/opt/robotwin
-
-# The nightly base is CUDA -base (no compiler, no Vulkan loader). CuRobo's
-# `pip install -e .` runs nvcc, and SAPIEN renders via Vulkan — add both.
-USER root
-# Pinned upstream SHA for reproducible benchmark runs. Bump when we need
-# an upstream fix; don't rely on `main` drift.
-ARG ROBOTWIN_SHA=0aeea2d669c0f8516f4d5785f0aa33ba812c14b4
-RUN apt-get update \
-    && apt-get install -y --no-install-recommends \
-         cuda-nvcc-12-8 cuda-cudart-dev-12-8 \
-         libvulkan1 vulkan-tools \
-    && mkdir -p /usr/share/vulkan/icd.d \
-    && echo '{"file_format_version":"1.0.0","ICD":{"library_path":"libGLX_nvidia.so.0","api_version":"1.3.0"}}' \
-       > /usr/share/vulkan/icd.d/nvidia_icd.json \
-    && git clone https://github.com/RoboTwin-Platform/RoboTwin.git ${ROBOTWIN_ROOT} \
-    && git -C ${ROBOTWIN_ROOT} checkout ${ROBOTWIN_SHA} \
-    && chown -R user_lerobot:user_lerobot ${ROBOTWIN_ROOT} \
-    && apt-get clean && rm -rf /var/lib/apt/lists/*
-USER user_lerobot
-
-# RoboTwin runtime deps (av is already in the base via [av-dep]).
-RUN uv pip install --no-cache \
-        "sapien==3.0.0b1" "mplib==0.2.1" "transforms3d==0.4.2" "trimesh==4.4.3" \
-        "open3d==0.19.0" "imageio==2.34.2" termcolor zarr pydantic h5py
-
-# pytorch3d has no universal wheel; must be built from source (~10 min, cached).
-RUN uv pip install --no-cache --no-build-isolation \
-        "git+https://github.com/facebookresearch/pytorch3d.git@stable"
-
-# CuRobo — NVlabs motion generator; TORCH_CUDA_ARCH_LIST must be set or the
-# build aborts on an empty arch list. RoboTwin's own installer pins v0.7.8,
-# which still exposes the v1 API (`curobo.types.math`) that RoboTwin imports.
-ARG CUROBO_REF=v0.7.8
-RUN cd ${ROBOTWIN_ROOT}/envs \
-    && git clone --branch ${CUROBO_REF} --depth 1 https://github.com/NVlabs/curobo.git \
-    && cd curobo \
-    && TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;8.9;9.0" \
-       uv pip install -e . --no-build-isolation --no-cache
-
-# Upstream patches (mirror RoboTwin's script/_install.sh).
-# These patches target the exact versions pinned above; re-check when upgrading.
-# mplib==0.2.1: drop a broken `or collide` clause in planner.py.
-#   Safe to remove once mplib > 0.2.1 ships with the fix upstream.
-# sapien==3.0.0b1: fix URDF loader encoding + .srdf extension check.
-#   Safe to remove once sapien > 3.0.0b1 ships with the fix upstream.
-RUN python - <<'EOF'
-import pathlib, re, site
-for d in site.getsitepackages():
-    p = pathlib.Path(d) / "mplib" / "planner.py"
-    if p.exists():
-        p.write_text(re.sub(r"\bor collide\b", "", p.read_text(), count=1))
-        print(f"mplib patch applied: {p}")
-    p = pathlib.Path(d) / "sapien" / "wrapper" / "urdf_loader.py"
-    if p.exists():
-        src = p.read_text().replace(
-            "with open(srdf_path) as f:", 'with open(srdf_path, encoding="utf-8") as f:'
-        ).replace('"srdf"', '".srdf"')
-        p.write_text(src)
-        print(f"sapien patch applied: {p}")
-EOF
-
-# Simulation assets from TianxingChen/RoboTwin2.0: embodiments (~220 MB) +
-# objects (~3.74 GB). background_texture (~11 GB) is intentionally skipped.
-# The dataset is public — no auth token needed.
-RUN python - <<'EOF'
-import os, pathlib, zipfile
-from huggingface_hub import hf_hub_download
-
-assets_dir = pathlib.Path(os.environ["ROBOTWIN_ROOT"]) / "assets"
-assets_dir.mkdir(parents=True, exist_ok=True)
-for fname in ("embodiments.zip", "objects.zip"):
-    local = hf_hub_download(
-        repo_id="TianxingChen/RoboTwin2.0",
-        repo_type="dataset",
-        filename=fname,
-        local_dir=str(assets_dir),
-    )
-    with zipfile.ZipFile(local, "r") as z:
-        z.extractall(str(assets_dir))
-    pathlib.Path(local).unlink()
-EOF
-
-WORKDIR ${ROBOTWIN_ROOT}
-RUN python script/update_embodiment_config_path.py
-
-ENV PYTHONPATH="${ROBOTWIN_ROOT}"
-
-# Fail the image build early if the CuRobo package layout regresses. Importing
-# RoboTwin's planner here is too eager because CuRobo constructs CUDA-backed
-# defaults at import time, while Docker builds don't have access to an NVIDIA
-# driver.
-RUN python - <<'EOF'
-from pathlib import Path
-
-from curobo.types.math import Pose
-
-planner_src = (Path("/opt/robotwin/envs/robot/planner.py")).read_text()
-assert "from curobo.types.math import Pose as CuroboPose" in planner_src
-
-print("CuRobo import OK:", Pose.__name__)
-print("RoboTwin planner import references curobo.types.math")
-EOF
-
-# Return to the lerobot source directory (set by base image) before overlaying.
-WORKDIR /lerobot
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-CMD ["/bin/bash"]
--- a/docker/Dockerfile.benchmark.vlabench
+++ b/docker/Dockerfile.benchmark.vlabench
@@ -1,99 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for VLABench integration tests.
-# Extends the nightly GPU image with the PR's source code and VLABench setup.
-#
-# Build:  docker build -f docker/Dockerfile.benchmark.vlabench -t lerobot-benchmark-vlabench .
-# Run:    docker run --gpus all --rm lerobot-benchmark-vlabench lerobot-eval ...
-
-FROM huggingface/lerobot-gpu:latest
-
-# Install VLABench from GitHub (not on PyPI) and pin MuJoCo/dm-control.
-# Shallow-clone without submodule recursion (nested SSH-only submodules fail in CI).
-# Editable install (-e) because VLABench/utils/ has no __init__.py, so
-# find_packages() omits it from wheels; editable mode uses the source tree directly.
-# rrt-algorithms has the same packaging issue (rrt/ dir missing __init__.py).
-# Patch: constant.py calls os.listdir on ~100 asset/obj/meshes/* dirs at import
-# time. Guard the call so missing dirs return [] instead of crashing (in case
-# the asset download is partial).
-#
-# Pinned upstream SHAs for reproducible benchmark runs. Bump when you need
-# an upstream fix; don't rely on `main`/`develop` drift.
-ARG VLABENCH_SHA=cf588fe60c0c7282174fe979f5913170cfe69017
-ARG RRT_ALGORITHMS_SHA=e51d95ee489a225220d6ae2a764c4111f6ba7d85
-RUN git clone https://github.com/OpenMOSS/VLABench.git ~/VLABench && \
-    git -C ~/VLABench checkout ${VLABENCH_SHA} && \
-    git clone https://github.com/motion-planning/rrt-algorithms.git ~/rrt-algorithms && \
-    git -C ~/rrt-algorithms checkout ${RRT_ALGORITHMS_SHA} && \
-    python3 -c "\
-import pathlib; \
-p = pathlib.Path.home() / 'VLABench/VLABench/configs/constant.py'; \
-t = p.read_text(); \
-p.write_text(t.replace( \
-    'subdirs = os.listdir(xml_dir)', \
-    'if not os.path.isdir(xml_dir): return []\n    subdirs = os.listdir(xml_dir)'))" && \
-    uv pip install --no-cache -e ~/VLABench -e ~/rrt-algorithms \
-      mujoco==3.2.2 dm-control==1.0.22 \
-      open3d colorlog scikit-learn openai gdown
-
-# Download VLABench mesh assets. Task configs reference object meshes
-# (obj/meshes/fruit/, containers/basket/, tablewares/plates/, etc.); without
-# them the task builder picks from an empty mesh list and crashes with
-# IndexError at task-build time (random.choice([]) in config_manager.py).
-#
-# Preferred source: an HF Hub mirror. Set VLABENCH_ASSETS_REPO at build time
-# (e.g. --build-arg VLABENCH_ASSETS_REPO=lerobot/vlabench-assets) and we'll
-# snapshot_download the repo into VLABench's assets dir. This is the reliable
-# path for CI — Google Drive frequently returns HTTP 429 ("Too many users have
-# viewed or downloaded this file recently") on shared academic files.
-#
-# After download we *validate* that at least one XML exists under each
-# task-critical subtree and fail the build loudly if not. Silent-empty asset
-# dirs are the #1 cause of VLABench runtime crashes in CI, so we surface them
-# here rather than after a 10-minute eval build.
-#
-# Fallback: VLABench's own gdown-based script. Best-effort only.
-ARG VLABENCH_ASSETS_REPO=""
-RUN ASSETS_DIR="$HOME/VLABench/VLABench/assets" && \
-    if [ -n "${VLABENCH_ASSETS_REPO}" ]; then \
-        echo "Downloading VLABench assets from HF Hub: ${VLABENCH_ASSETS_REPO}" && \
-        uv pip install --no-cache "huggingface_hub[hf_xet]>=0.26" && \
-        python -c "from huggingface_hub import snapshot_download; \
-p = snapshot_download(repo_id='${VLABENCH_ASSETS_REPO}', repo_type='dataset', \
-    local_dir='${ASSETS_DIR}', allow_patterns=['obj/**', 'scenes/**']); \
-print('snapshot_download returned:', p)"; \
-    else \
-        echo "No VLABENCH_ASSETS_REPO set — falling back to gdown" && \
-        python ~/VLABench/scripts/download_assets.py --choice all; \
-    fi && \
-    python -c "\
-from pathlib import Path; \
-import sys; \
-root = Path('${ASSETS_DIR}'); \
-checks = ['obj/meshes/tablewares/plates', 'obj/meshes/containers/basket', 'obj/meshes/fruit', 'obj/meshes/containers/tray']; \
-failed = []; \
-print(f'Validating VLABench assets under {root}'); \
-[print(f'  {c}: {len(list((root/c).rglob(\"*.xml\")))} XMLs') for c in checks]; \
-[failed.append(c) for c in checks if not any((root/c).rglob('*.xml'))]; \
-sys.exit(f'Empty asset dirs (no *.xml): {failed}') if failed else print('All asset dirs populated.')"
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-# Re-install lerobot editably so the new source (with VLABenchEnv registration
-# and updated obs handling) replaces the stale package baked into the nightly image.
-RUN uv pip install --no-cache --no-deps -e .
-
-CMD ["/bin/bash"]
--- a/docker/Dockerfile.internal
+++ b/docker/Dockerfile.internal
@@ -18,8 +18,9 @@
 # docker build -f docker/Dockerfile.internal -t lerobot-internal .

 # Configure the base image for CI with GPU access
-ARG CUDA_VERSION=12.8.1
-ARG OS_VERSION=24.04
+# TODO(Steven): Bump these versions
+ARG CUDA_VERSION=12.4.1
+ARG OS_VERSION=22.04
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}

 # Define Python version argument
@@ -35,13 +36,16 @@ ENV DEBIAN_FRONTEND=noninteractive \

 # Install Python, system dependencies, and uv (as root)
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential git curl \
-    libglib2.0-0 libgl1 libegl1 ffmpeg \
+    software-properties-common build-essential git curl \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
    libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
    cmake pkg-config ninja-build \
-    python${PYTHON_VERSION} \
-    python${PYTHON_VERSION}-venv \
-    python${PYTHON_VERSION}-dev \
+    && add-apt-repository -y ppa:deadsnakes/ppa \
+    && apt-get update \
+    && apt-get install -y --no-install-recommends \
+       python${PYTHON_VERSION} \
+       python${PYTHON_VERSION}-venv \
+       python${PYTHON_VERSION}-dev \
    && curl -LsSf https://astral.sh/uv/install.sh | sh \
    && mv /root/.local/bin/uv /usr/local/bin/uv \
    && useradd --create-home --shell /bin/bash user_lerobot \
--- a/docs/source/_toctree.yml
+++ b/docs/source/_toctree.yml
@@ -3,14 +3,12 @@
    title: LeRobot
  - local: installation
    title: Installation
-  - local: cheat-sheet
-    title: Cheat sheet
  title: Get started
 - sections:
  - local: il_robots
    title: Imitation Learning for Robots
  - local: bring_your_own_policies
-    title: Adding a Policy
+    title: Bring Your Own Policies
  - local: integrate_hardware
    title: Bring Your Own Hardware
  - local: hilserl
@@ -26,12 +24,6 @@
  - local: rename_map
    title: Using Rename Map and Empty Cameras
  title: "Tutorials"
- sections:
-  - local: hardware_guide
-    title: Compute Hardware Guide
-  - local: torch_accelerators
-    title: PyTorch accelerators
-  title: "Compute & Hardware"
 - sections:
  - local: lerobot-dataset-v3
    title: Using LeRobotDataset
@@ -39,12 +31,8 @@
    title: Porting Large Datasets
  - local: using_dataset_tools
    title: Using the Dataset Tools
-  - local: language_and_recipes
-    title: Language Columns and Recipes
-  - local: tools
-    title: Tools
-  - local: video_encoding_parameters
-    title: Video encoding parameters
+  - local: dataset_subtask
+    title: Using Subtasks in the Dataset
  - local: streaming_video_encoding
    title: Streaming Video Encoding
  title: "Datasets"
@@ -59,8 +47,6 @@
    title: π₀-FAST (Pi0Fast)
  - local: pi05
    title: π₀.₅ (Pi05)
-  - local: eo1
-    title: EO-1
  - local: groot
    title: NVIDIA GR00T N1.5
  - local: xvla
@@ -93,22 +79,10 @@
    title: Adding a New Benchmark
  - local: libero
    title: LIBERO
-  - local: libero_plus
-    title: LIBERO-plus
  - local: metaworld
    title: Meta-World
-  - local: robotwin
-    title: RoboTwin 2.0
-  - local: robocasa
-    title: RoboCasa365
-  - local: robocerebra
-    title: RoboCerebra
-  - local: robomme
-    title: RoboMME
  - local: envhub_isaaclab_arena
    title: NVIDIA IsaacLab Arena Environments
-  - local: vlabench
-    title: VLABench
  title: "Benchmarks"
 - sections:
  - local: introduction_processors
@@ -145,8 +119,6 @@
    title: OMX
  - local: openarm
    title: OpenArm
-  - local: rebot_b601
-    title: reBot B601-DM
  title: "Robots"
 - sections:
  - local: phone_teleop
@@ -156,6 +128,10 @@
  - local: cameras
    title: Cameras
  title: "Sensors"
+- sections:
+  - local: torch_accelerators
+    title: PyTorch accelerators
+  title: "Supported Hardware"
 - sections:
  - local: notebooks
    title: Notebooks
--- a/docs/source/act.mdx
+++ b/docs/source/act.mdx
@@ -79,13 +79,17 @@ If your local computer doesn't have a powerful GPU, you can utilize Google Colab
 Once training is complete, you can evaluate your ACT policy using the `lerobot-record` command with your trained policy. This will run inference and record evaluation episodes:

 ```bash
-lerobot-rollout \
-  --strategy.type=base \
-  --policy.path=${HF_USER}/act_policy \
-  --robot.type=so101_follower \
+lerobot-record \
+  --robot.type=so100_follower \
  --robot.port=/dev/ttyACM0 \
+  --robot.id=my_robot \
  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
  --display_data=true \
-  --task="Your task description" \ # can be skipped for ACT
-  --duration=60
+  --dataset.repo_id=${HF_USER}/eval_act_your_dataset \
+  --dataset.num_episodes=10 \
+  --dataset.single_task="Your task description" \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
+  --policy.path=${HF_USER}/act_policy
 ```
--- a/docs/source/bring_your_own_policies.mdx
+++ b/docs/source/bring_your_own_policies.mdx
@@ -1,37 +1,60 @@
-# Adding a Policy
+# Bring Your Own Policies

-This guide walks you through implementing a custom policy and getting it to work with LeRobot's training, evaluation, and deployment tools. There are two paths:
+This tutorial explains how to integrate your own custom policy implementations into the LeRobot ecosystem, allowing you to leverage all LeRobot tools for training, evaluation, and deployment while using your own algorithms.

- **Plugin (out-of-tree)** — ship your policy as a standalone `lerobot_policy_*` package. Faster, no PR required, easy to iterate. Right for experimentation, internal use, or when you want to publish independently.
- **In-tree (contributed to LeRobot)** — land your policy directly in `src/lerobot/policies/`. Requires a PR, but makes your policy a first-class citizen of the library.
+## Step 1: Create a Policy Package

-The plugin route is usually the right starting point — promote to in-tree once the policy has stabilized and there's clear value in shipping it with the library.
+Your custom policy should be organized as an installable Python package following LeRobot's plugin conventions.

-Either way, the building blocks are the same: a configuration class, a policy class, and a processor factory. The first half of this guide covers those shared pieces; the second half covers the path-specific scaffolding ([Path A](#path-a-out-of-tree-plugin), [Path B](#path-b-contributing-in-tree)).
+### Package Structure

-A note on tone: robot-learning is an actively evolving field, and "what a policy looks like" can shift with each new architecture. The conventions described here exist because they let `lerobot-train` and `lerobot-eval` work uniformly across very different models. When a new policy genuinely doesn't fit them, raise it (in your PR, or an issue) — the conventions are not sacred.
+Create a package with the prefix `lerobot_policy_` (IMPORTANT!) followed by your policy name:

---
+```bash
+lerobot_policy_my_custom_policy/
+├── pyproject.toml
+└── src/
+    └── lerobot_policy_my_custom_policy/
+        ├── __init__.py
+        ├── configuration_my_custom_policy.py
+        ├── modeling_my_custom_policy.py
+        └── processor_my_custom_policy.py
+```

-## Anatomy of a policy
+### Package Configuration

-Three building blocks make up every policy. The names below use `my_policy` as a placeholder — replace with your policy's name. That name is load-bearing: it must match the string you pass to `@PreTrainedConfig.register_subclass`, the `MyPolicy.name` class attribute, and the `make_<name>_pre_post_processors` factory function (more on each below).
+Set up your `pyproject.toml`:

-### Configuration class
+```toml
+[project]
+name = "lerobot_policy_my_custom_policy"
+version = "0.1.0"
+dependencies = [
+    # your policy-specific dependencies
+]
+requires-python = ">= 3.12"

-Inherit from [`PreTrainedConfig`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/configs/policies.py) and register your policy type. Here is a template — customize the parameters and methods as needed for your policy's architecture and training requirements.
+[build-system]
+build-backend = # your-build-backend
+requires = # your-build-system
+```
+
+## Step 2: Define the Policy Configuration
+
+Create a configuration class that inherits from [`PreTrainedConfig`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/configs/policies.py) and registers your policy type:
+Here is a template to get you started, customize the parameters and methods as needed for your policy's architecture and training requirements.

 ```python
-# configuration_my_policy.py
+# configuration_my_custom_policy.py
 from dataclasses import dataclass, field
 from lerobot.configs import PreTrainedConfig
 from lerobot.optim import AdamWConfig
 from lerobot.optim import CosineDecayWithWarmupSchedulerConfig

-@PreTrainedConfig.register_subclass("my_policy")
+@PreTrainedConfig.register_subclass("my_custom_policy")
@dataclass
-class MyPolicyConfig(PreTrainedConfig):
-    """Configuration class for MyPolicy.
+class MyCustomPolicyConfig(PreTrainedConfig):
+    """Configuration class for MyCustomPolicy.

    Args:
        n_obs_steps: Number of observation steps to use as input
@@ -54,20 +77,16 @@ class MyPolicyConfig(PreTrainedConfig):
            raise ValueError("n_action_steps cannot exceed horizon")

    def validate_features(self) -> None:
-        """Validate input/output feature compatibility.
-
-        Call this explicitly from your policy's __init__ — the base class does not.
-        """
+        """Validate input/output feature compatibility."""
        if not self.image_features:
-            raise ValueError("MyPolicy requires at least one image feature.")
+            raise ValueError("MyCustomPolicy requires at least one image feature.")
        if self.action_feature is None:
-            raise ValueError("MyPolicy requires 'action' in output_features.")
+            raise ValueError("MyCustomPolicy requires 'action' in output_features.")

    def get_optimizer_preset(self) -> AdamWConfig:
        return AdamWConfig(lr=self.optimizer_lr, weight_decay=self.optimizer_weight_decay)

    def get_scheduler_preset(self):
-        """Return a LRSchedulerConfig from lerobot.optim, or None."""
        return None

    @property
@@ -82,7 +101,8 @@ class MyPolicyConfig(PreTrainedConfig):

    @property
    def action_delta_indices(self) -> list[int]:
-        """Relative timestep offsets for the action chunk the dataset loader returns."""
+        """Relative timestep offsets for the action chunk the dataset loader returns.
+        """
        return list(range(self.horizon))

    @property
@@ -90,34 +110,32 @@ class MyPolicyConfig(PreTrainedConfig):
        return None
 ```

-The string you pass to `@register_subclass` must match `MyPolicy.name` (next section) and is what users supply as `--policy.type` on the CLI. Default to `AdamW` from `lerobot.optim` for `get_optimizer_preset` unless you genuinely need otherwise.
+## Step 3: Implement the Policy Class

-### Policy class
-
-Inherit from [`PreTrainedPolicy`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pretrained.py) and set two class attributes — both are checked by `__init_subclass__`:
+Create your policy implementation by inheriting from [`PreTrainedPolicy`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pretrained.py):

 ```python
-# modeling_my_policy.py
+# modeling_my_custom_policy.py
 import torch
 import torch.nn as nn
 from typing import Any

 from lerobot.policies import PreTrainedPolicy
 from lerobot.utils.constants import ACTION
-from .configuration_my_policy import MyPolicyConfig
+from .configuration_my_custom_policy import MyCustomPolicyConfig

-class MyPolicy(PreTrainedPolicy):
-    config_class = MyPolicyConfig  # must match the string in @register_subclass
-    name = "my_policy"
+class MyCustomPolicy(PreTrainedPolicy):
+    config_class = MyCustomPolicyConfig  # must match the string in @register_subclass
+    name = "my_custom_policy"

-    def __init__(self, config: MyPolicyConfig, dataset_stats: dict[str, Any] = None):
+    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: dict[str, Any] = None):
        super().__init__(config, dataset_stats)
        config.validate_features()  # not called automatically by the base class
        self.config = config
        self.model = ...  # your nn.Module here

    def reset(self):
-        """Reset per-episode state. Called by lerobot-eval at the start of each episode."""
+        """Reset episode state."""
        ...

    def get_optim_params(self) -> dict:
@@ -129,51 +147,35 @@ class MyPolicy(PreTrainedPolicy):
        ...

    def select_action(self, batch: dict[str, torch.Tensor], **kwargs) -> torch.Tensor:
-        """Return a single action for the current timestep (called every step at inference)."""
+        """Return a single action for the current timestep (called at inference)."""
        ...

-    def forward(self, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, dict | None]:
+    def forward(self, batch: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
        """Compute the training loss.

-        Returns `(loss, output_dict)`. `output_dict` may be `None`; everything in it must be
-        logging-friendly Python natives (no tensors with gradients).
-
        `batch["action_is_pad"]` is a bool mask of shape (B, horizon) that marks
-        timesteps padded because the episode ended before `horizon` steps; you
+        timesteps padded because the episode ended before `horizon` steps, you
        can exclude those from your loss.
        """
        actions = batch[ACTION]
        action_is_pad = batch.get("action_is_pad")
        ...
-        return loss, {"some_loss_component": some_loss_component.item()}
+        return {"loss": ...}
 ```

-The methods called by the train/eval loops:
+## Step 4: Add Data Processors

-| Method                                                            | Used by           | What it does                                                                                                                                                                                                                                         |
-| ----------------------------------------------------------------- | ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `reset() -> None`                                                 | `lerobot-eval`    | Clear per-episode state at the start of each episode.                                                                                                                                                                                                |
-| `select_action(batch, **kwargs) -> Tensor`                        | `lerobot-eval`    | Return the next action `(B, action_dim)`. Called every step.                                                                                                                                                                                         |
-| `predict_action_chunk(batch, **kwargs) -> Tensor`                 | the policy itself | Return an action chunk `(B, chunk_size, action_dim)`. Currently abstract on the base class — raise `NotImplementedError` if your policy doesn't chunk.                                                                                               |
-| `forward(batch, reduction="mean") -> tuple[Tensor, dict \| None]` | `lerobot-train`   | Return `(loss, output_dict)`. Accept `reduction="none"` if you want to support per-sample weighting.                                                                                                                                                 |
-| `get_optim_params() -> dict`                                      | the optimizer     | Return `self.parameters()` for simple policies; return a named parameter dict for [multi-optimizer policies](https://github.com/huggingface/lerobot/blob/ecd38c50d7d15b4184cf42649ff1185ee2e11eeb/src/lerobot/policies/sac/modeling_sac.py#L61-L73). |
-| `update() -> None` _(optional)_                                   | `lerobot-train`   | Called after each optimizer step _if defined_. Use for EMA, target nets, replay buffers (TDMPC uses this).                                                                                                                                           |
-
-Batches are flat dictionaries keyed by the constants in [`lerobot.utils.constants`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/utils/constants.py): `OBS_STATE` (`observation.state.<motor>`), `OBS_IMAGES` (`observation.images.<camera>`), `OBS_LANGUAGE`, `ACTION`, etc. Reuse the constants — don't invent new prefixes.
-
-### Processor functions
-
-LeRobot uses `PolicyProcessorPipeline`s to normalize inputs and de-normalize outputs around your policy. For a concrete reference, see [`processor_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/processor_act.py) or [`processor_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/processor_diffusion.py).
+Create processor functions. For a concrete reference, see [processor_act.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/processor_act.py) or [processor_diffusion.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/processor_diffusion.py).

 ```python
-# processor_my_policy.py
+# processor_my_custom_policy.py
 from typing import Any
 import torch

 from lerobot.processor import PolicyAction, PolicyProcessorPipeline


-def make_my_policy_pre_post_processors(
+def make_my_custom_policy_pre_post_processors(
    config,
    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
 ) -> tuple[
@@ -185,48 +187,11 @@ def make_my_policy_pre_post_processors(
    return preprocessor, postprocessor
 ```

-**Important — function naming:** LeRobot discovers your processor by name. The function **must** be called `make_{policy_name}_pre_post_processors` (matching the string you passed to `@PreTrainedConfig.register_subclass`).
+**Important - function naming:** LeRobot discovers your processor by name. The function **must** be called `make_{policy_name}_pre_post_processors` (matching the string you passed to `@PreTrainedConfig.register_subclass`).

---
+## Step 5: Package Initialization

-## Path A: Out-of-tree plugin
-
-The fastest way to ship a policy: package it as a standalone Python distribution and install it alongside LeRobot. No PR required, you own the release cycle, and you can publish to PyPI under your own namespace.
-
-### Package structure
-
-Create a package with the prefix `lerobot_policy_` (IMPORTANT!) followed by your policy name:
-
-```bash
-lerobot_policy_my_policy/
-├── pyproject.toml
-└── src/
-    └── lerobot_policy_my_policy/
-        ├── __init__.py
-        ├── configuration_my_policy.py
-        ├── modeling_my_policy.py
-        └── processor_my_policy.py
-```
-
-### `pyproject.toml`
-
-```toml
-[project]
-name = "lerobot_policy_my_policy"
-version = "0.1.0"
-dependencies = [
-    # your policy-specific dependencies
-]
-requires-python = ">= 3.12"
-
-[build-system]
-build-backend = # your-build-backend
-requires = # your-build-system
-```
-
-### Package `__init__.py`
-
-Expose your classes in the package's `__init__.py` and guard against missing `lerobot`:
+Expose your classes in the package's `__init__.py`:

 ```python
 # __init__.py
@@ -239,148 +204,44 @@ except ImportError:
        "lerobot is not installed. Please install lerobot to use this policy package."
    )

-from .configuration_my_policy import MyPolicyConfig
-from .modeling_my_policy import MyPolicy
-from .processor_my_policy import make_my_policy_pre_post_processors
+from .configuration_my_custom_policy import MyCustomPolicyConfig
+from .modeling_my_custom_policy import MyCustomPolicy
+from .processor_my_custom_policy import make_my_custom_policy_pre_post_processors

 __all__ = [
-    "MyPolicyConfig",
-    "MyPolicy",
-    "make_my_policy_pre_post_processors",
+    "MyCustomPolicyConfig",
+    "MyCustomPolicy",
+    "make_my_custom_policy_pre_post_processors",
 ]
 ```

-### Install and use
+## Step 6: Installation and Usage
+
+### Install Your Policy Package

 ```bash
-cd lerobot_policy_my_policy
+cd lerobot_policy_my_custom_policy
 pip install -e .

 # Or install from PyPI if published
-pip install lerobot_policy_my_policy
+pip install lerobot_policy_my_custom_policy
 ```

+### Use Your Policy
+
 Once installed, your policy automatically integrates with LeRobot's training and evaluation tools:

 ```bash
 lerobot-train \
-    --policy.type my_policy \
+    --policy.type my_custom_policy \
    --env.type pusht \
    --steps 200000
 ```

---
-
-## Path B: Contributing in-tree
-
-When your policy has stabilized and there's clear value in shipping it with the library, you can land it directly in LeRobot. Read the general [contribution guide](./contributing) and the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md) first — that's where you'll find the testing/quality expectations every PR has to meet (`pre-commit run -a`, `pytest`, the community-review rule, etc.). What's below is the policy-specific layer on top of that.
-
-### In-tree layout
-
-```
-src/lerobot/policies/my_policy/
-├── __init__.py                    # re-exports config + modeling + processor factory
-├── configuration_my_policy.py     # MyPolicyConfig + @register_subclass
-├── modeling_my_policy.py          # MyPolicy(PreTrainedPolicy)
-├── processor_my_policy.py         # make_my_policy_pre_post_processors
-└── README.md                      # symlink → ../../../../docs/source/policy_my_policy_README.md
-```
-
-Two notes:
-
- The `README.md` next to the source is a **symlink** into `docs/source/policy_<name>_README.md` — the actual file lives under `docs/`. Existing policies (act, smolvla, diffusion, …) all do this; copy one of those symlinks. The policy README is conventionally minimal: paper link + BibTeX citation.
- The user-facing tutorial — what to install, how to train, hyperparameters, benchmark numbers — lives separately at `docs/source/<my_policy>.mdx` and is registered in `_toctree.yml` under "Policies".
-
-The file names are load-bearing: the factory does lazy imports by name, and the processor is discovered by the `make_<policy_name>_pre_post_processors` convention.
-
-### Wiring
-
-Three places need to know about your policy. All by name.
-
-1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast).
-2. **`factory.py:get_policy_class`** — add a branch returning `MyPolicy` from a lazy import.
-3. **`factory.py:make_policy_config`** and **`factory.py:make_pre_post_processors`** — same idea, two more branches.
-
-Mirror an existing policy that's structurally similar to yours; the diff is small.
-
-### Heavy / optional dependencies
-
-Most policies need a heavy backbone (transformers, diffusers, a specific VLM SDK). The convention is **two-step gating**: a `TYPE_CHECKING`-guarded import at module top, and a `require_package` runtime check in the constructor. [`modeling_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/modeling_diffusion.py) is the canonical reference:
-
-```python
-from typing import TYPE_CHECKING
-from lerobot.utils.import_utils import _diffusers_available, require_package
-
-if TYPE_CHECKING or _diffusers_available:
-    from diffusers.schedulers.scheduling_ddim import DDIMScheduler
-else:
-    DDIMScheduler = None  # keeps the symbol bindable at import time
-
-class DiffusionPolicy(PreTrainedPolicy):
-    def __init__(self, config):
-        require_package("diffusers", extra="diffusion")
-        super().__init__(config)
-        ...
-```
-
-This way:
-
- `import lerobot.policies` keeps working without the extra installed (the symbol is just bound to `None`).
- Type checkers see the real symbol.
- Instantiating the policy without the extra raises a clear `ImportError` pointing at `pip install 'lerobot[diffusion]'`.
-
-Add a matching extra to [`pyproject.toml`](https://github.com/huggingface/lerobot/blob/main/pyproject.toml) `[project.optional-dependencies]` and include it in the `all` extra so `pip install 'lerobot[all]'` keeps installing everything.
-
-### Benchmarks and a published checkpoint
-
-A new policy is much easier to review — and far more useful — when it ships with a working checkpoint and at least one number you can reproduce.
-
-**Pick at least one in-tree benchmark.** LeRobot ships sim benchmarks with per-benchmark Docker images (LIBERO, LIBERO-plus, Meta-World, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, VLABench and more). Pick the one that matches your policy's modality — VLAs usually go to LIBERO or VLABench; image-only BC to LIBERO or Meta-World. The full list lives under [Benchmarks](./libero) in the docs sidebar.
-
-**Push the checkpoint & processors** to the Hub under `lerobot/<policy>_<benchmark>` (or your namespace if you don't have write access; a maintainer can mirror it). Use `PreTrainedPolicy.push_model_to_hub` so the repo gets `config.json`, `model.safetensors`, and a model card.
-
-**Report results in your policy's MDX**, with the exact `lerobot-eval` command and hardware so anyone can re-run:
-
-```markdown
-## Results
-
-Evaluated on LIBERO with `lerobot/<policy>_libero`:
-
-| Suite          | Success rate | n_episodes |
-| -------------- | -----------: | ---------: |
-| libero_spatial |        87.5% |         50 |
-| libero_object  |        93.0% |         50 |
-| libero_goal    |        81.5% |         50 |
-| libero_10      |        62.0% |         50 |
-| **average**    |    **81.0%** |        200 |
-
-Reproduce: `lerobot-eval --policy.path=lerobot/<policy>_libero --env.type=libero --env.task=libero_spatial --eval.n_episodes=50` (1× A100 40 GB).
-```
-
-Use `n_episodes ≥ 50` per suite for stable success-rate estimates.
-
-If your policy is real-robot-only and no sim benchmark applies, swap the sim eval for: a public training dataset on the Hub, the `lerobot-train` command, the checkpoint, and a real-robot success rate over ≥10 episodes via `lerobot-rollout --policy.path=...`.
-
-### PR checklist
-
-The general expectations are in [`CONTRIBUTING.md`](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md) and the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md). On top of those, reviewers will look for:
-
- [ ] `MyPolicy` and `MyPolicyConfig` cover the surface above; `__init_subclass__` accepts the class.
- [ ] `factory.py` and `policies/__init__.py` are wired (lazy imports for modeling).
- [ ] `make_my_policy_pre_post_processors` follows the naming convention.
- [ ] Optional deps live behind a `[project.optional-dependencies]` extra and the `TYPE_CHECKING + require_package` guard.
- [ ] `tests/policies/` updated; backward-compat artifact committed & policy-specific tests.
- [ ] `src/lerobot/policies/<name>/README.md` symlinked into `docs/source/policy_<name>_README.md`; user-facing `docs/source/<name>.mdx` written and added to `_toctree.yml`.
- [ ] At least one reproducible benchmark eval in the policy MDX with a published checkpoint (sim benchmark, or real-robot dataset + checkpoint).
-
-The fastest way to get a clean PR is to copy the directory of the existing policy closest to yours, rename, and replace contents method by method. Don't wait until everything is polished — open a draft PR early and iterate with us; reviewers would much rather give feedback on a half-finished branch than a fully-merged one.
-
---
-
-## Examples and community contributions
+## Examples and Community Contributions

 Check out these example policy implementations:

- [DiTFlow Policy](https://github.com/danielsanjosepro/lerobot_policy_ditflow) — Diffusion Transformer policy with flow-matching objective. Try it out in this example: [DiTFlow Example](https://github.com/danielsanjosepro/test_lerobot_policy_ditflow)
+- [DiTFlow Policy](https://github.com/danielsanjosepro/lerobot_policy_ditflow) - Diffusion Transformer policy with flow-matching objective. Try it out in this example: [DiTFlow Example](https://github.com/danielsanjosepro/test_lerobot_policy_ditflow)

-Thanks for taking the time to bring a new policy into LeRobot. Every architecture that lands in `main` — and every plugin published by the community — makes the library a little more useful for the next person, and a little more representative of where robot learning is going. We're looking forward to seeing what you ship. 🤗
+Share your policy implementations with the community! 🤗
--- a/docs/source/cheat-sheet.mdx
+++ b/docs/source/cheat-sheet.mdx
@@ -1,139 +0,0 @@
-# Cheat sheet
-
-All of the LeRobot commands in one place. If you forgot how to use a specific command or want to learn about a new one you can do it here.
-
-> [!WARNING]
-> For all of the commands listed below remember to change the ports/names/ids to your own values!
-
-> [!TIP]
-> Another great way to look at all the commands and get them configured for your specific setup is to use this [Jupyter Notebook](https://github.com/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb).
-
-### Setup and installation
-
-For installation please look at [LeRobot Installation](https://huggingface.co/docs/lerobot/main/en/installation).
-
-### Useful tools
-
-###### Find port
-
-Use this to identify which serial ports your robots are connected to. Follow the instructions in your terminal: you will be asked to unplug the USB cable and press Enter. The script will then detect and print the correct serial port for that robot.
-
-```bash
-lerobot-find-port
-```
-
-###### Find cameras
-
-Quickly find camera indices and verify their output. This command prints camera information to the terminal and saves test frames from each detected camera to `lerobot/outputs/captured_images`
-
-```bash
-lerobot-find-cameras
-```
-
-### Calibration
-
-In most cases you will need to perform calibration just once for each robot and teleoperation device. Before performing the calibration make sure that all the joints are roughly in the middle position.
-
-```bash
-lerobot-calibrate \
-    --robot.type=so101_follower \
-    --robot.port=/dev/ttyACM0 \
-    --robot.id=my_follower_arm
-```
-
-Make sure that you use the same IDs used during calibration later for the other scripts. That's how LeRobot finds the calibration files.
-
-### Teleoperation
-
-Teleoperating with two cameras and displaying the data with Rerun.
-
-```bash
-lerobot-teleoperate \
-    --robot.type=so101_follower \
-    --robot.port=/dev/ttyACM0 \
-    --robot.id=my_follower_arm \
-    --robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
-    --teleop.type=so101_leader \
-    --teleop.port=/dev/ttyACM1 \
-    --teleop.id=my_leader_arm \
-    --display_data=true
-```
-
-### Recording a dataset
-
-The dataset is automatically uploaded to the server and saved under repo_id, make sure you are logged in to your HF account with CLI:
-`hf auth login`
-
-You can get the token from: [https://huggingface.co/settings/tokens](https://huggingface.co/settings/tokens)
-
-```bash
-lerobot-record \
-    --robot.type=so101_follower \
-    --robot.port=/dev/ttyACM0 \
-    --robot.id=my_follower_arm \
-    --robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
-    --teleop.type=so101_leader \
-    --teleop.port=/dev/ttyACM1 \
-    --teleop.id=my_leader_arm \
-    --dataset.repo_id=${HF_USER}/so101_dataset_test \
-    --dataset.num_episodes=30 \
-    --dataset.single_task="put the red brick in a bowl" \
-    --dataset.streaming_encoding=true \
-    --display_data=true
-```
-
-While collecting the dataset you can control the process with your keyboard:
-Control the data recording flow using keyboard shortcuts:
-
- Press **Right Arrow (`→`)**: Save episode and move to the next.
- Press **Left Arrow (`←`)**: Delete current episode and retry.
- Press **Escape (`ESC`)**: Stop, encode videos, and upload.
-
-### Training
-
-Depending on your hardware training the policy might take a few hours. That's how you train simple `ACT` policy:
-
-```bash
-lerobot-train \
-    --dataset.repo_id=${HF_USER}/so101_dataset_test \
-    --policy.type=act \
-    --output_dir=outputs/train/act_so101_test \
-    --job_name=act_so101_test \
-    --policy.device=cuda \
-    --wandb.enable=true \
-    --policy.repo_id=${HF_USER}/policy_test \
-    --steps=20000
-```
-
- Policy Types: `act`, `diffusion`, `smolvla`, `pi05`
- Devices: `cuda` (NVIDIA), `mps` (Apple Silicon), `cpu`
-
-If you want to fine-tune a specific model you can provide the path to the model. In this case path is enough and type can be skipped.
-
-```bash
-lerobot-train \
-    --dataset.repo_id=${HF_USER}/so101_dataset_test \
-    --policy.path=username/the_policy_to_finetune \
-    --policy.device=cuda \
-    --policy.repo_id=${HF_USER}/policy_test \
-    --output_dir=outputs/train/act_so101_test \
-    --steps=20000
-```
-
-### Inference
-
-Inference means running the trained policy/model on a robot. For that we use `lerobot-rollout`. You will need to provide a path to your policy. It can be a local path or a path to Hugging Face for example "lerobot/folding_latest". Your cameras configuration needs to match what was used when collecting the dataset. Duration is in seconds if unspecified, it will run forever.
-
-> [!TIP]
-> If you are using the previous release V0.5.1 instead of `lerobot-rollout` you need to use `lerobot-record`. More information [here](https://huggingface.co/docs/lerobot/v0.5.1/en/il_robots#run-inference-and-evaluate-your-policy).
-
-```bash
-lerobot-rollout \
-    --strategy.type=base \
-    --policy.path=${HF_USER}/my_policy \
-    --robot.type=so101_follower \
-    --robot.port=/dev/ttyACM1 \
-    --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video1, width: 640, height: 480, fps: 30}, side: {type: opencv, index_or_path: /dev/video5, width: 640, height: 480, fps: 30}}" \
-    --task="Put lego brick into the transparent box" \
-    --duration=60
-```
--- a/docs/source/dataset_subtask.mdx
+++ b/docs/source/dataset_subtask.mdx
@@ -0,0 +1,277 @@
+# Using Subtasks in LeRobot Datasets
+
+Subtask support in robotics datasets has proven effective in improving robot reasoning and understanding. Subtasks are particularly useful for:
+
+- **Hierarchical policies**: Building policies that include subtask predictions to visualize robot reasoning in real time
+- **Reward modeling**: Helping reward models understand task progression (e.g., SARM-style stage-aware reward models)
+- **Task decomposition**: Breaking down complex manipulation tasks into atomic, interpretable steps
+
+LeRobotDataset now supports subtasks as part of its dataset structure, alongside tasks.
+
+## What are Subtasks?
+
+While a **task** describes the overall goal (e.g., "Pick up the apple and place it in the basket"), **subtasks** break down the execution into finer-grained steps:
+
+1. "Approach the apple"
+2. "Grasp the apple"
+3. "Lift the apple"
+4. "Move to basket"
+5. "Release the apple"
+
+Each frame in the dataset can be annotated with its corresponding subtask, enabling models to learn and predict these intermediate stages.
+
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/subtask-asset.png"
+  alt="An overview of subtask annotation showing how frames are labeled with intermediate subtask stages"
+  width="80%"
+/>
+
+<p>
+  <em>Figure: Overview of subtask annotation.</em>
+</p>
+
+**Reference:** _Subtask-learning based for robot self-assembly in flexible collaborative assembly in manufacturing_, Original Article, Published: 19 April 2022.
+
+## Dataset Structure
+
+Subtask information is stored in the dataset metadata:
+
+```
+my-dataset/
+├── data/
+│   └── ...
+├── meta/
+│   ├── info.json
+│   ├── stats.json
+│   ├── tasks.parquet
+│   ├── subtasks.parquet      # Subtask index → subtask string mapping
+│   └── episodes/
+│       └── ...
+└── videos/
+    └── ...
+```
+
+### Subtasks Parquet File
+
+The `meta/subtasks.parquet` file maps subtask indices to their natural language descriptions:
+
+| subtask_index | subtask (index column) |
+| ------------- | ---------------------- |
+| 0             | "Approach the apple"   |
+| 1             | "Grasp the apple"      |
+| 2             | "Lift the apple"       |
+| ...           | ...                    |
+
+### Frame-Level Annotations
+
+Each frame in the dataset can include a `subtask_index` field that references the subtasks parquet file:
+
+```python
+# Example frame data in the parquet file
+{
+    "index": 42,
+    "timestamp": 1.4,
+    "episode_index": 0,
+    "task_index": 0,
+    "subtask_index": 2,  # References "Lift the apple"
+    "observation.state": [...],
+    "action": [...],
+}
+```
+
+## Annotating Datasets with Subtasks
+
+We provide a HuggingFace Space for easily annotating any LeRobotDataset with subtasks:
+
+**[https://huggingface.co/spaces/lerobot/annotate](https://huggingface.co/spaces/lerobot/annotate)**
+
+After completing your annotation:
+
+1. Click "Push to Hub" to upload your annotated dataset
+2. You can also run the annotation space locally by following the instructions at [github.com/huggingface/lerobot-annotate](https://github.com/huggingface/lerobot-annotate)
+
+## Loading Datasets with Subtasks
+
+When you load a dataset with subtask annotations, the subtask information is automatically available:
+
+```python
+from lerobot.datasets import LeRobotDataset
+
+# Load a dataset with subtask annotations
+dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
+
+# Access a sample
+sample = dataset[100]
+
+# The sample includes both task and subtask information
+print(sample["task"])        # "Collect the fruit"
+print(sample["subtask"])     # "Grasp the apple"
+print(sample["task_index"])  # tensor(0)
+print(sample["subtask_index"])  # tensor(2)
+```
+
+### Checking for Subtask Support
+
+You can check if a dataset has subtask annotations:
+
+```python
+# Check if subtasks are available
+has_subtasks = (
+    "subtask_index" in dataset.features
+    and dataset.meta.subtasks is not None
+)
+
+if has_subtasks:
+    print(f"Dataset has {len(dataset.meta.subtasks)} unique subtasks")
+    print("Subtasks:", list(dataset.meta.subtasks.index))
+```
+
+## Using Subtasks for Training
+
+### With the Tokenizer Processor
+
+The `TokenizerProcessor` automatically handles subtask tokenization for Vision-Language Action (VLA) models:
+
+```python
+from lerobot.processor import TokenizerProcessorStep
+
+# Create a tokenizer processor step
+tokenizer_processor = TokenizerProcessorStep(
+    tokenizer_name_or_path="google/paligemma-3b-pt-224",
+    padding="max_length",
+    max_length=64,
+)
+
+# The processor will automatically tokenize subtasks if present in the batch
+# and add them to the observation under:
+# - "observation.subtask.tokens"
+# - "observation.subtask.attention_mask"
+```
+
+When subtasks are available in the batch, the tokenizer processor adds:
+
+- `observation.subtask.tokens`: Tokenized subtask text
+- `observation.subtask.attention_mask`: Attention mask for the subtask tokens
+
+### DataLoader with Subtasks
+
+```python
+import torch
+from lerobot.datasets import LeRobotDataset
+
+dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
+
+dataloader = torch.utils.data.DataLoader(
+    dataset,
+    batch_size=16,
+    shuffle=True,
+)
+
+for batch in dataloader:
+    # Access subtask information in the batch
+    subtasks = batch["subtask"]  # List of subtask strings
+    subtask_indices = batch["subtask_index"]  # Tensor of subtask indices
+
+    # Use for training hierarchical policies or reward models
+    print(f"Batch subtasks: {set(subtasks)}")
+```
+
+## Example Datasets with Subtask Annotations
+
+Try loading a dataset with subtask annotations:
+
+```python
+from lerobot.datasets import LeRobotDataset
+
+# Example dataset with subtask annotations
+dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
+
+# Explore the subtasks
+print("Available subtasks:")
+for subtask_name in dataset.meta.subtasks.index:
+    print(f"  - {subtask_name}")
+
+# Get subtask distribution
+subtask_counts = {}
+for i in range(len(dataset)):
+    sample = dataset[i]
+    subtask = sample["subtask"]
+    subtask_counts[subtask] = subtask_counts.get(subtask, 0) + 1
+
+print("\nSubtask distribution:")
+for subtask, count in sorted(subtask_counts.items(), key=lambda x: -x[1]):
+    print(f"  {subtask}: {count} frames")
+```
+
+## Use Cases
+
+### 1. Hierarchical Policy Training
+
+Train policies that predict both actions and current subtask:
+
+```python
+class HierarchicalPolicy(nn.Module):
+    def __init__(self, num_subtasks):
+        super().__init__()
+        self.action_head = nn.Linear(hidden_dim, action_dim)
+        self.subtask_head = nn.Linear(hidden_dim, num_subtasks)
+
+    def forward(self, observations):
+        features = self.encoder(observations)
+        actions = self.action_head(features)
+        subtask_logits = self.subtask_head(features)
+        return actions, subtask_logits
+```
+
+### 2. Stage-Aware Reward Modeling (SARM)
+
+Build reward models that understand task progression:
+
+```python
+# SARM predicts:
+# - Stage: Which subtask is being executed (discrete)
+# - Progress: How far along the subtask (continuous 0-1)
+
+class SARMRewardModel(nn.Module):
+    def forward(self, observations):
+        features = self.encoder(observations)
+        stage_logits = self.stage_classifier(features)
+        progress = self.progress_regressor(features)
+        return stage_logits, progress
+```
+
+### 3. Progress Visualization
+
+Monitor robot execution by tracking subtask progression:
+
+```python
+def visualize_execution(model, observations):
+    for t, obs in enumerate(observations):
+        action, subtask_logits = model(obs)
+        predicted_subtask = subtask_names[subtask_logits.argmax()]
+        print(f"t={t}: Executing '{predicted_subtask}'")
+```
+
+## API Reference
+
+### LeRobotDataset Properties
+
+| Property                    | Type                   | Description                                |
+| --------------------------- | ---------------------- | ------------------------------------------ |
+| `meta.subtasks`             | `pd.DataFrame \| None` | DataFrame mapping subtask names to indices |
+| `features["subtask_index"]` | `dict`                 | Feature spec for subtask_index if present  |
+
+### Sample Keys
+
+When subtasks are available, each sample includes:
+
+| Key             | Type           | Description                          |
+| --------------- | -------------- | ------------------------------------ |
+| `subtask_index` | `torch.Tensor` | Integer index of the current subtask |
+| `subtask`       | `str`          | Natural language subtask description |
+
+## Related Resources
+
+- [SARM Paper](https://arxiv.org/pdf/2509.25358) - Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
+- [LeRobot Annotate Space](https://huggingface.co/spaces/lerobot/annotate) - Interactive annotation tool
+- [LeRobotDataset v3.0](./lerobot-dataset-v3) - Dataset format documentation
--- a/docs/source/earthrover_mini_plus.mdx
+++ b/docs/source/earthrover_mini_plus.mdx
@@ -194,7 +194,7 @@ lerobot-record \
    --dataset.single_task="Navigate around obstacles" \
    --dataset.streaming_encoding=true \
    --dataset.encoder_threads=2 \
-    # --dataset.camera_encoder.vcodec=auto \
+    # --dataset.vcodec=auto \
    --display_data=true
 ```

--- a/docs/source/eo1.mdx
+++ b/docs/source/eo1.mdx
@@ -1,168 +0,0 @@
-# EO-1
-
-EO-1 is a **Vision-Language-Action policy for robot control**. The LeRobot implementation integrates EO-1 with the standard LeRobot training, evaluation, processor interface.
-
-## Model Overview
-
-EO-1 uses a Qwen2.5-VL backbone for vision-language understanding and adds a continuous flow-matching action head for robot control. The policy formats each robot-control sample as a multimodal conversation: camera images are passed to Qwen2.5-VL, the robot state is represented with EO-1 state tokens, and the future action chunk is represented with EO-1 action tokens.
-
-<img
-  src="https://huggingface.co/datasets/HaomingSong/lerobot-documentation-images/resolve/main/lerobot/eo_pipeline.png"
-  alt="An overview of EO-1"
-  width="85%"
-/>
-
-During training, EO-1 learns to denoise continuous action chunks at the action-token positions. During inference, it samples an action chunk, returns continuous actions, and executes `n_action_steps` from the chunk before sampling again.
-
-### What the LeRobot Integration Covers
-
- Standard `policy.type=eo1` configuration through LeRobot
- Qwen2.5-VL image and text preprocessing through policy processors
- Continuous flow-matching action prediction
- Checkpoint save/load through LeRobot policy APIs
- Training with `lerobot-train` and evaluation with `lerobot-eval`
-
-The broader EO-1 project also includes interleaved vision-text-action pretraining and multimodal reasoning workflows. This page focuses on the LeRobot robot-control policy path.
-
-## Installation Requirements
-
-1. Install LeRobot by following the [Installation Guide](./installation).
-2. Install EO-1 dependencies by running:
-
-   ```bash
-   pip install -e ".[eo1]"
-   ```
-
-3. If you want to train or evaluate on LIBERO, install the LIBERO dependencies too:
-
-   ```bash
-   pip install -e ".[eo1,libero]"
-   ```
-
-EO-1 can use the standard PyTorch scaled-dot-product attention backend through `policy.attn_implementation=sdpa`. If your environment has a compatible `flash_attn` installation, you can request `policy.attn_implementation=flash_attention_2`.
-
-## Data Requirements
-
-EO-1 expects a LeRobot dataset with:
-
- At least one visual observation, for example `observation.images.image`
- `observation.state`
- `action`
- A language task instruction through the dataset `task` field
-
-If your dataset uses different observation names, use `rename_map` to align them with the names expected by your training or evaluation setup.
-
-## Usage
-
-To use EO-1 in a LeRobot configuration, specify the policy type as:
-
-```python
-policy.type=eo1
-```
-
-By default, a new EO-1 policy initializes its backbone from:
-
-```python
-policy.vlm_base=Qwen/Qwen2.5-VL-3B-Instruct
-```
-
-Once a LeRobot-format EO-1 checkpoint is available, load it with:
-
-```python
-policy.path=your-org/your-eo1-checkpoint
-```
-
-## Training
-
-### Training Command Example
-
-```bash
-lerobot-train \
-  --dataset.repo_id=your_org/your_dataset \
-  --policy.type=eo1 \
-  --policy.vlm_base=Qwen/Qwen2.5-VL-3B-Instruct \
-  --policy.dtype=bfloat16 \
-  --policy.attn_implementation=sdpa \
-  --policy.gradient_checkpointing=false \
-  --output_dir=./outputs/eo1_training \
-  --job_name=eo1_training \
-  --steps=300000 \
-  --batch_size=16 \
-  --policy.device=cuda
-```
-
-### Key Training Parameters
-
-| Parameter                              | Default                       | Description                                                             |
-| -------------------------------------- | ----------------------------- | ----------------------------------------------------------------------- |
-| `policy.vlm_base`                      | `Qwen/Qwen2.5-VL-3B-Instruct` | Qwen2.5-VL checkpoint used to initialize a new policy                   |
-| `policy.dtype`                         | `auto`                        | Backbone dtype request: `auto`, `bfloat16`, or `float32`                |
-| `policy.attn_implementation`           | `None`                        | Optional Qwen attention backend, such as `sdpa`                         |
-| `policy.gradient_checkpointing`        | `false`                       | Reduces memory usage during training                                    |
-| `policy.chunk_size`                    | `8`                           | Number of future actions predicted per chunk                            |
-| `policy.n_action_steps`                | `8`                           | Number of actions consumed from a sampled chunk                         |
-| `policy.num_denoise_steps`             | `10`                          | Number of flow-matching denoising steps used during sampling            |
-| `policy.max_state_dim`                 | `32`                          | State padding dimension                                                 |
-| `policy.max_action_dim`                | `32`                          | Action padding dimension                                                |
-| `policy.force_fp32_autocast`           | `true`                        | Keeps the flow head in fp32 even when the backbone uses mixed precision |
-| `policy.supervise_padding_action_dims` | `true`                        | Controls whether padded action dimensions are supervised                |
-| `policy.supervise_padding_actions`     | `true`                        | Controls whether padded future action rows are supervised               |
-
-## Evaluation
-
-EO-1 can be evaluated through `lerobot-eval` once you have a LeRobot-format checkpoint:
-
-```bash
-lerobot-eval \
-  --policy.path=your-org/your-eo1-checkpoint \
-  --env.type=libero \
-  --env.task=libero_object \
-  --eval.batch_size=1 \
-  --eval.n_episodes=20
-```
-
-For datasets or environments whose camera names differ from the checkpoint configuration, pass a `rename_map`:
-
-```bash
-lerobot-eval \
-  --policy.path=your-org/your-eo1-checkpoint \
-  --env.type=libero \
-  --env.task=libero_object \
-  --rename_map='{"observation.images.image2":"observation.images.wrist_image"}'
-```
-
-## Configuration Notes
-
-### Image Processing
-
-EO-1 uses the Qwen2.5-VL processor. The `policy.image_min_pixels` and `policy.image_max_pixels` settings control the image resizing bounds before the visual tokens are passed into the backbone.
-
-### State and Action Dimensions
-
-The policy pads state and action vectors to `policy.max_state_dim` and `policy.max_action_dim` before the EO-1 flow head. Predictions are cropped back to the original action dimension before being returned by the policy.
-
-### Attention Backend
-
-Use `policy.attn_implementation=sdpa` for a portable setup. Use `flash_attention_2` only when `flash_attn` is installed and compatible with your environment.
-
-## References
-
- [EO-1 project](https://github.com/EO-Robotics/EO1)
- [EO-1 paper](https://arxiv.org/abs/2508.21112)
- [Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct)
-
-## Citation
-
-```bibtex
-@article{eo1,
-  title={EO-1: Interleaved Vision-Text-Action Pretraining for General Robot Control},
-  author={Delin Qu and Haoming Song and Qizhi Chen and Zhaoqing Chen and Xianqiang Gao and Xinyi Ye and Qi Lv and Modi Shi and Guanghui Ren and Cheng Ruan and Maoqing Yao and Haoran Yang and Jiacheng Bao and Bin Zhao and Dong Wang},
-  journal={arXiv preprint},
-  year={2025},
-  url={https://arxiv.org/abs/2508.21112}
-}
-```
-
-## License
-
-This LeRobot integration follows the **Apache 2.0 License** used by LeRobot. Check the upstream EO-1 model and dataset pages for the licenses of released EO-1 checkpoints and data.
--- a/docs/source/groot.mdx
+++ b/docs/source/groot.mdx
@@ -105,12 +105,10 @@ These results demonstrate GR00T's strong generalization capabilities across dive

 ### Evaluate in your hardware setup

-Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Policy Deployment (lerobot-rollout)](./inference). For example:
+Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Imitation Learning for Robots](./il_robots). For example:

 ```bash
-lerobot-rollout\
-  --strategy.type=sentry \
-  --strategy.upload_every_n_episodes=5 \
+lerobot-record \
  --robot.type=bi_so_follower \
  --robot.left_arm_port=/dev/ttyACM1 \
  --robot.right_arm_port=/dev/ttyACM0 \
@@ -121,12 +119,14 @@ lerobot-rollout\
  }' \
  --display_data=true \
  --dataset.repo_id=<user>/eval_groot-bimanual  \
+  --dataset.num_episodes=10 \
  --dataset.single_task="Grab and handover the red cube to the other arm" \
  --dataset.streaming_encoding=true \
  --dataset.encoder_threads=2 \
-  # --dataset.camera_encoder.vcodec=auto \
+  # --dataset.vcodec=auto \
  --policy.path=<user>/groot-bimanual \ # your trained model
-  --duration=600
+  --dataset.episode_time_s=30 \
+  --dataset.reset_time_s=10
 ```

 ## License
--- a/docs/source/hardware_guide.mdx
+++ b/docs/source/hardware_guide.mdx
@@ -1,98 +0,0 @@
-# Compute HW Guide for LeRobot Training
-
-Rough sizing for training a LeRobot policy: how much VRAM each policy needs, what training time looks like, and where to run when local hardware isn't enough.
-
-The numbers below are **indicative** — order-of-magnitude figures for picking hardware, not exact predictions. Throughput depends heavily on dataset I/O, image resolution, batch size, and number of GPUs.
-
-## Memory by policy group
-
-Policies cluster by backbone size; the groupings below give a single VRAM envelope per group instead of repeating numbers per policy. Memory scales roughly linearly with batch size; AdamW (the LeRobot default) carries optimizer state that adds ~30–100% over a forward+backward pass alone.
-
-| Group      | Policies                                    | Peak VRAM (BS 8, AdamW) | Suitable starter GPUs             |
-| ---------- | ------------------------------------------- | ----------------------: | --------------------------------- |
-| Light BC   | `act`, `vqbet`, `tdmpc`                     |                  ~2–6GB | Laptop GPU (RTX 3060), L4, A10G   |
-| Diffusion  | `diffusion`, `multi_task_dit`               |                 ~8–14GB | RTX 4070+ / L4 / A10G             |
-| Small VLA  | `smolvla`                                   |                ~10–16GB | RTX 4080+ / L4 / A10G             |
-| Large VLA  | `pi0`, `pi0_fast`, `pi05`, `xvla`, `wall_x` |                ~24–40GB | A100 40 GB+ (24 GB tight at BS 1) |
-| Multimodal | `groot`, `eo1`                              |                ~24–40GB | A100 40 GB+                       |
-| RL         | `sac`                                       |             config-dep. | See [HIL-SERL guide](./hilserl)   |
-
-Memory-bound? Drop the batch size (~linear), use gradient accumulation to recover effective batch, or for SmolVLA leave `freeze_vision_encoder=True`.
-
-## Training time
-
-Robotics imitation learning typically converges in **5–10 epochs over the dataset**, not hundreds of thousands of raw steps. Once you know your epoch count, wall-clock is essentially:
-
-```text
-total_frames    = sum of frames over all episodes      # 50 ep × 30 fps × 30 s ≈ 45,000
-steps_per_epoch = ceil(total_frames / (num_gpus × batch_size))
-total_steps     = epochs × steps_per_epoch
-wall_clock      ≈ total_steps × per_step_time
-```
-
-Per-step time depends on the policy and the GPU. The numbers in the table below are anchors — pick the row closest to your setup and scale linearly with `total_steps` if you train longer or shorter.
-
-### Common scenarios
-
-Indicative wall-clock for **5 epochs on a ~50-episode dataset (~45k frames at 30 fps × 30 s)**, default optimizer (AdamW), 640×480 images:
-
-| Setup                                | Policy         | Batch | Wall-clock |
-| ------------------------------------ | -------------- | ----- | ---------: |
-| Single RTX 4090 / RTX 3090 (24 GB)   | `act`          | 8     |  ~30–60min |
-| Single RTX 4090 / RTX 3090 (24 GB)   | `diffusion`    | 8     |      ~2–4h |
-| Single L4 / A10G (24 GB)             | `act`          | 8     |      ~1–2h |
-| Single L4 / A10G (24 GB)             | `smolvla`      | 4     |      ~3–6h |
-| Single A100 40 GB                    | `smolvla`      | 16    |      ~1–2h |
-| Single A100 40 GB                    | `pi0` / `pi05` | 4     |      ~4–8h |
-| 4× H100 80 GB cluster (`accelerate`) | `diffusion`    | 32    |  ~30–60min |
-| 4× H100 80 GB cluster (`accelerate`) | `smolvla`      | 32    |      ~1–2h |
-| Apple Silicon M1/M2/M3 Max (MPS)     | `act`          | 4     |     ~6–14h |
-
-These are order-of-magnitude figures. Real runs deviate by ±50% depending on image resolution, dataset I/O, dataloader threading, and exact GPU SKU. They are useful as "is this run going to take an hour or a day?" intuition, not as SLAs.
-
-### Multi-GPU matters a lot
-
-`accelerate launch --num_processes=N` is the easiest way to cut training time. Each optimizer step processes `N × batch_size` samples in roughly the same wall-clock as a single-GPU step, so 4 GPUs ≈ 4× speedup for compute-bound runs. See the [Multi GPU training](./multi_gpu_training) guide for the full setup.
-
-Reference data points on a 4×H100 80 GB cluster (`accelerate launch --num_processes=4`), 5000 steps, batch 32, AdamW, dataset [`imstevenpmwork/super_poulain_draft`](https://huggingface.co/datasets/imstevenpmwork/super_poulain_draft) (~50 episodes, ~640×480 images):
-
-| Policy      | Wall-clock | `update_s` | `dataloading_s` | GPU util | Notable flags                                                                                                                  |
-| ----------- | ---------- | ---------: | --------------: | -------- | ------------------------------------------------------------------------------------------------------------------------------ |
-| `diffusion` | 16m 17s    |      0.167 |           0.015 | ~90%     | defaults (training from scratch)                                                                                               |
-| `smolvla`   | 27m 49s    |      0.312 |           0.011 | ~80%     | `--policy.path=lerobot/smolvla_base`, `freeze_vision_encoder=false`, `train_expert_only=false`                                 |
-| `pi05`      | 3h 41m     |      2.548 |           0.014 | ~95%     | `--policy.pretrained_path=lerobot/pi05_base`, `gradient_checkpointing=true`, `dtype=bfloat16`, vision encoder + expert trained |
-
-The `dataloading_s` vs. `update_s` ratio is the diagnostic that matters: when `dataloading_s` approaches `update_s`, more GPUs stop helping — your dataloader is the bottleneck and you should look at `--num_workers`, image resolution, and disk speed before adding compute.
-
-### Schedule and checkpoints
-
-If you shorten training (e.g. 5k–10k steps on a small dataset), also shorten the LR schedule with `--policy.scheduler_decay_steps≈--steps`. Otherwise the LR stays near its peak and never decays. Same for `--save_freq`.
-
-## Where to run
-
-VRAM is the first filter. Within a tier, pick by budget and availability — the `$`–`$$$$` columns are relative; check current pricing on the provider you actually use.
-
-| Class                      | VRAM  | Tier   | Comfortable for                                             |
-| -------------------------- | ----- | ------ | ----------------------------------------------------------- |
-| RTX 3090 / 4090 (consumer) | 24 GB | `$`    | Light BC, Diffusion, SmolVLA. Tight for VLAs at batch 1.    |
-| L4 / A10G (cloud)          | 24 GB | `$–$$` | Same envelope; common on Google Cloud, RunPod, AWS `g5/g6`. |
-| A100 40 GB                 | 40 GB | `$$$`  | Any policy at reasonable batch sizes.                       |
-| A100 80 GB / H100 80 GB    | 80 GB | `$$$$` | Multi-GPU clusters; large batches for VLAs.                 |
-| **CPU only**               | —     | —      | Don't train. Use Colab or rent a GPU.                       |
-
-### Hugging Face Jobs
-
-[Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) lets you run training on managed HF infrastructure, billed by the second. The repo publishes a ready-to-use image: **`huggingface/lerobot-gpu:latest`**, rebuilt **every night at 02:00 UTC from `main`** ([`docker_publish.yml`](https://github.com/huggingface/lerobot/blob/main/.github/workflows/docker_publish.yml)) — so it tracks the current state of the repo, not a tagged release.
-
-```bash
-hf jobs run --flavor a10g-large huggingface/lerobot-gpu:latest \
-  bash -c "nvidia-smi && lerobot-train \
-    --policy.type=act --dataset.repo_id=<USER>/<DATASET> \
-    --policy.repo_id=<USER>/act_<task> --batch_size=8 --steps=50000"
-```
-
-Notes:
-
- The leading `nvidia-smi` is a quick sanity check that CUDA is visible inside the container — useful to fail fast if the flavor or driver mismatched.
- The default Job timeout is 30 minutes; pass `--timeout 4h` (or longer) for real training.
- `--flavor` maps onto the table above: `t4-small`/`t4-medium` (T4, ACT only), `l4x1`/`l4x4` (L4 24 GB), `a10g-small/large/largex2/largex4` (A10G 24 GB scaled out), `a100-large` (A100). For the current full catalogue + pricing see [https://huggingface.co/docs/hub/jobs](https://huggingface.co/docs/hub/jobs).
--- a/docs/source/hil_data_collection.mdx
+++ b/docs/source/hil_data_collection.mdx
@@ -108,7 +108,7 @@ lerobot-rollout --strategy.type=dagger \
    --teleop.port_left=/dev/ttyACM0 \
    --teleop.port_right=/dev/ttyACM1 \
    --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
-    --dataset.repo_id=your-username/rollout_hil_dataset \
+    --dataset.repo_id=your-username/hil-dataset \
    --dataset.single_task="Fold the T-shirt properly" \
    --dataset.fps=30 \
    --strategy.num_episodes=50 \
@@ -135,7 +135,7 @@ lerobot-rollout --strategy.type=dagger \
    --teleop.port_left=/dev/ttyACM0 \
    --teleop.port_right=/dev/ttyACM1 \
    --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
-    --dataset.repo_id=your-username/rollout_hil_rtc_dataset \
+    --dataset.repo_id=your-username/hil-rtc-dataset \
    --dataset.single_task="Fold the T-shirt properly" \
    --dataset.fps=30 \
    --strategy.num_episodes=50 \
--- a/docs/source/hilserl.mdx
+++ b/docs/source/hilserl.mdx
@@ -62,7 +62,7 @@ pip install -e ".[hilserl]"

 ### Understanding Configuration

-The training process begins with proper configuration for the HILSERl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` (defined in `lerobot/envs/configs.py`) and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
+The training process begins with proper configuration for the HILSerl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:

 <!-- prettier-ignore-start -->
 ```python
@@ -95,7 +95,6 @@ class HILSerlProcessorConfig:
 class ObservationConfig:
    add_joint_velocity_to_observation: bool = False    # Add joint velocities to state
    add_current_to_observation: bool = False    # Add motor currents to state
-    add_ee_pose_to_observation: bool = False    # Add end-effector pose to state
    display_cameras: bool = False    # Display camera feeds during execution

 class ImagePreprocessingConfig:
@@ -327,22 +326,14 @@ lerobot-find-joint-limits \
   Max joint positions [-20.0, -20.0, -20.0, -20.0, -20.0, -20.0]
   Min joint positions [50.0, 50.0, 50.0, 50.0, 50.0, 50.0]
   ```
-3. Use these values in your environment configuration under `env.processor.inverse_kinematics.end_effector_bounds` (see `InverseKinematicsConfig` in `lerobot/envs/configs.py`)
+3. Use these values in the configuration of your teleoperation device (TeleoperatorConfig) under the `end_effector_bounds` field

 **Example Configuration**

 ```json
-{
-  "env": {
-    "processor": {
-      "inverse_kinematics": {
-        "end_effector_bounds": {
-          "max": [0.24, 0.2, 0.1],
-          "min": [0.16, -0.08, 0.03]
-        }
-      }
-    }
-  }
+"end_effector_bounds": {
+    "max": [0.24, 0.20, 0.10],
+    "min": [0.16, -0.08, 0.03]
 }
 ```

@@ -413,24 +404,30 @@ We support using a gamepad or a keyboard or the leader arm of the robot.

 HIL-Serl learns actions in the end-effector space of the robot. Therefore, the teleoperation will control the end-effector's x,y,z displacements.

-The end-effector transformation is applied by the processor pipeline (`InverseKinematicsRLStep`, `EEBoundsAndSafety`, `EEReferenceAndDelta`, `GripperVelocityToJoint`) configured under `env.processor.inverse_kinematics` (`InverseKinematicsConfig`) and `env.processor.gripper` / `env.processor.max_gripper_pos`. The defaults related to the end-effector space are:
+For that we need to define a version of the robot that takes actions in the end-effector space. Check the robot class `SO100FollowerEndEffector` and its configuration `SO100FollowerEndEffectorConfig` for the default parameters related to the end-effector space.

 <!-- prettier-ignore-start -->
 ```python
-class InverseKinematicsConfig:
-    """Configuration for inverse kinematics processing."""
+class SO100FollowerEndEffectorConfig(SO100FollowerConfig):
+    """Configuration for the SO100FollowerEndEffector robot."""

-    urdf_path: str | None = None
-    target_frame_name: str | None = None
-    # bounds for the end-effector in x,y,z direction
-    end_effector_bounds: dict[str, list[float]] | None = None
-    # maximum step size for the end-effector in x,y,z direction
-    end_effector_step_sizes: dict[str, float] | None = None
+    # Default bounds for the end-effector position (in meters)
+    end_effector_bounds: dict[str, list[float]] = field( # bounds for the end-effector in x,y,z direction
+        default_factory=lambda: {
+            "min": [-1.0, -1.0, -1.0],  # min x, y, z
+            "max": [1.0, 1.0, 1.0],  # max x, y, z
+        }
+    )

-class HILSerlProcessorConfig:
-    ...
-    # maximum gripper position that the gripper will be open at
-    max_gripper_pos: float | None = 100.0
+    max_gripper_pos: float = 50 # maximum gripper position that the gripper will be open at
+
+    end_effector_step_sizes: dict[str, float] = field( # maximum step size for the end-effector in x,y,z direction
+        default_factory=lambda: {
+            "x": 0.02,
+            "y": 0.02,
+            "z": 0.02,
+        }
+    )
 ```
 <!-- prettier-ignore-end -->

@@ -609,11 +606,11 @@ This guide explains how to train a reward classifier for human-in-the-loop reinf

 **Note**: Training a reward classifier is optional. You can start the first round of RL experiments by annotating the success manually with your gamepad or keyboard device.

-The reward classifier implementation in `lerobot/rewards/classifier/modeling_classifier.py` uses a pretrained vision model to process the images. It can output either a single value for binary rewards to predict success/fail cases or multiple values for multi-class settings.
+The reward classifier implementation in `modeling_classifier.py` uses a pretrained vision model to process the images. It can output either a single value for binary rewards to predict success/fail cases or multiple values for multi-class settings.

 **Collecting a Dataset for the reward classifier**

-Before training, you need to collect a dataset with labeled examples. Setting `mode: "record"` in your config and running `gym_manipulator.py` enables the process of collecting a dataset of observations, actions, and rewards.
+Before training, you need to collect a dataset with labeled examples. The `record_dataset` function in `gym_manipulator.py` enables the process of collecting a dataset of observations, actions, and rewards.

 To collect a dataset, you need to modify some parameters in the environment configuration based on HILSerlRobotEnvConfig.

@@ -661,7 +658,7 @@ Example configuration section for data collection:
  },
  "dataset": {
    "repo_id": "hf_username/dataset_name",
-    "root": "data/your_dataset",
+    "dataset_root": "data/your_dataset",
    "task": "reward_classifier_task",
    "num_episodes_to_record": 20,
    "replay_episode": null,
@@ -674,7 +671,7 @@ Example configuration section for data collection:

 **Reward Classifier Configuration**

-The reward classifier is configured using `lerobot/rewards/classifier/configuration_classifier.py`. Here are the key parameters:
+The reward classifier is configured using `configuration_classifier.py`. Here are the key parameters:

 - **model_name**: Base model architecture (e.g., we mainly use `"helper2424/resnet10"`)
 - **model_type**: `"cnn"` or `"transformer"`
@@ -692,7 +689,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
    "repo_id": "hf_username/dataset_name",
    "root": null
  },
-  "reward_model": {
+  "policy": {
    "type": "reward_classifier",
    "model_name": "helper2424/resnet10",
    "model_type": "cnn",
@@ -702,6 +699,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
    "dropout_rate": 0.1,
    "learning_rate": 1e-4,
    "device": "cuda",
+    "use_amp": true,
    "input_features": {
      "observation.images.front": {
        "type": "VISUAL",
@@ -820,14 +818,13 @@ The LeRobot system uses a distributed actor-learner architecture for training. T

 **Configuration Setup**

-Create a training configuration file (example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/train_config.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/rl/train_rl.py`.
+Create a training configuration file (example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/train_config.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/configs/train.py`.

-1. Configure the policy settings (`type="gaussian_actor"`, `device`, etc.)
-2. Configure the algorithm settings under the top-level `algorithm` block (`type="sac"`, learning rates, discount, etc., defined in `lerobot/rl/algorithms/sac/configuration_sac.py`).
-3. Set `dataset` to your cropped dataset
-4. Configure environment settings with crop parameters
-5. Check the other parameters related to the Gaussian Actor in [configuration_gaussian_actor.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/gaussian_actor/configuration_gaussian_actor.py#L79).
-6. Verify that the `policy` config is correct with the right `input_features` and `output_features` for your task.
+1. Configure the policy settings (`type="sac"`, `device`, etc.)
+2. Set `dataset` to your cropped dataset
+3. Configure environment settings with crop parameters
+4. Check the other parameters related to SAC in [configuration_sac.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/sac/configuration_sac.py#L79).
+5. Verify that the `policy` config is correct with the right `input_features` and `output_features` for your task.

 **Starting the Learner**

@@ -929,7 +926,7 @@ The ideal behaviour is that your intervention rate should drop gradually during

 Some configuration values have a disproportionate impact on training stability and speed:

- **`temperature_init`** (`algorithm.temperature_init`) – initial entropy temperature in SAC. Higher values encourage more exploration; lower values make the policy more deterministic early on. A good starting point is `1e-2`. We observed that setting it too high can make human interventions ineffective and slow down learning.
+- **`temperature_init`** (`policy.temperature_init`) – initial entropy temperature in SAC. Higher values encourage more exploration; lower values make the policy more deterministic early on. A good starting point is `1e-2`. We observed that setting it too high can make human interventions ineffective and slow down learning.
 - **`policy_parameters_push_frequency`** (`policy.actor_learner_config.policy_parameters_push_frequency`) – interval in _seconds_ between two weight pushes from the learner to the actor. The default is `4 s`. Decrease to **1-2 s** to provide fresher weights (at the cost of more network traffic); increase only if your connection is slow, as this will reduce sample efficiency.
 - **`storage_device`** (`policy.storage_device`) – device on which the learner keeps the policy parameters. If you have spare GPU memory, set this to `"cuda"` (instead of the default `"cpu"`). Keeping the weights on-GPU removes CPU→GPU transfer overhead and can significantly increase the number of learner updates per second.

--- a/docs/source/hope_jr.mdx
+++ b/docs/source/hope_jr.mdx
@@ -232,7 +232,7 @@ lerobot-record \
    --dataset.private=true \
    --dataset.streaming_encoding=true \
    --dataset.encoder_threads=2 \
-    # --dataset.camera_encoder.vcodec=auto \
+    # --dataset.vcodec=auto \
    --display_data=true
 ```

@@ -278,6 +278,6 @@ lerobot-record \
  --dataset.num_episodes=10 \
  --dataset.streaming_encoding=true \
  --dataset.encoder_threads=2 \
-  # --dataset.camera_encoder.vcodec=auto \
+  # --dataset.vcodec=auto \
  --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```
--- a/docs/source/il_robots.mdx
+++ b/docs/source/il_robots.mdx
@@ -68,13 +68,13 @@ from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
 from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig

 robot_config = SO101FollowerConfig(
-    port="/dev/tty.usbmodem5AB90687491",
-    id="my_follower_arm",
+    port="/dev/tty.usbmodem58760431541",
+    id="my_red_robot_arm",
 )

 teleop_config = SO101LeaderConfig(
-    port="/dev/tty.usbmodem5AB90689011",
-    id="my_leader_arm",
+    port="/dev/tty.usbmodem58760431551",
+    id="my_blue_leader_arm",
 )

 robot = SO101Follower(robot_config)
@@ -108,13 +108,13 @@ With `rerun`, you can teleoperate again while simultaneously visualizing the cam
 <hfoption id="Command">
 ```bash
 lerobot-teleoperate \
-    --robot.type=so101_follower \
-    --robot.port=/dev/tty.usbmodem5AB90687491 \
-    --robot.id=my_follower_arm \
-    --robot.cameras="{front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-    --teleop.type=so101_leader \
-    --teleop.port=/dev/tty.usbmodem5AB90689011 \
-    --teleop.id=my_leader_arm \
+    --robot.type=koch_follower \
+    --robot.port=/dev/tty.usbmodem58760431541 \
+    --robot.id=my_awesome_follower_arm \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
+    --teleop.type=koch_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --teleop.id=my_awesome_leader_arm \
    --display_data=true
 ```
 </hfoption>
@@ -122,48 +122,34 @@ lerobot-teleoperate \

 <!-- prettier-ignore-start -->
 ```python
-import time
-from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
-from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.utils.visualization_utils import init_rerun, log_rerun_data, shutdown_rerun
+from lerobot.teleoperators.koch_leader import KochLeader, KochLeaderConfig
+from lerobot.robots.koch_follower import KochFollower, KochFollowerConfig

-robot_config = SO101FollowerConfig(
-    port="/dev/tty.usbmodem5AB90687491",
-    id="my_follower_arm",
-    cameras={
-        "wrist": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-        "top": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30)
-    }
+camera_config = {
+    "front": OpenCVCameraConfig(index_or_path=0, width=1920, height=1080, fps=30)
+}
+
+robot_config = KochFollowerConfig(
+    port="/dev/tty.usbmodem585A0076841",
+    id="my_red_robot_arm",
+    cameras=camera_config
 )

-teleop_config = SO101LeaderConfig(
-    port="/dev/tty.usbmodem5AB90689011",
-    id="my_leader_arm",
+teleop_config = KochLeaderConfig(
+    port="/dev/tty.usbmodem58760431551",
+    id="my_blue_leader_arm",
 )

-init_rerun(session_name="teleoperation")
-
-robot = SO101Follower(robot_config)
-teleop_device = SO101Leader(teleop_config)
+robot = KochFollower(robot_config)
+teleop_device = KochLeader(teleop_config)
 robot.connect()
 teleop_device.connect()

-TARGET_HZ = 30
-TIME_PER_FRAME = 1.0 / TARGET_HZ
-
 while True:
-    start_time = time.perf_counter()
-
    observation = robot.get_observation()
    action = teleop_device.get_action()
    robot.send_action(action)
-    log_rerun_data(observation=observation, action=action)
-
-    elapsed_time = time.perf_counter() - start_time
-    sleep_time = TIME_PER_FRAME - elapsed_time
-    if sleep_time > 0:
-        time.sleep(sleep_time)
 ```
 <!-- prettier-ignore-end -->

@@ -207,7 +193,7 @@ lerobot-record \
    --dataset.num_episodes=5 \
    --dataset.single_task="Grab the black cube" \
    --dataset.streaming_encoding=true \
-    # --dataset.camera_encoder.vcodec=auto \
+    # --dataset.vcodec=auto \
    --dataset.encoder_threads=2
 ```
 </hfoption>
@@ -216,11 +202,10 @@ lerobot-record \
 <!-- prettier-ignore-start -->
 ```python
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 from lerobot.utils.feature_utils import hw_to_dataset_features
-from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
-from lerobot.teleoperators.so_leader.config_so_leader import SO101LeaderConfig
-from lerobot.teleoperators.so_leader.so_leader import SO101Leader
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
@@ -233,56 +218,71 @@ EPISODE_TIME_SEC = 60
 RESET_TIME_SEC = 10
 TASK_DESCRIPTION = "My task description"

-def main():
-    # Create robot configuration
-    robot_config = SO101FollowerConfig(
-        port="/dev/tty.usbmodem5AB90687491",
-        id="my_follower_arm",
-        cameras={
-            "wrist": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-            "top": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30)
-        }
-    )
+# Create robot configuration
+robot_config = SO100FollowerConfig(
+    id="my_awesome_follower_arm",
+    cameras={
+        "front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS) # Optional: fourcc="MJPG" for troubleshooting OpenCV async error.
+    },
+    port="/dev/tty.usbmodem58760434471",
+)

-    teleop_config = SO101LeaderConfig(
-        port="/dev/tty.usbmodem5AB90689011",
-        id="my_leader_arm",
-    )
+teleop_config = SO100LeaderConfig(
+    id="my_awesome_leader_arm",
+    port="/dev/tty.usbmodem585A0077581",
+)

-    # Initialize the robot and teleoperator
-    robot = SO101Follower(robot_config)
-    teleop = SO101Leader(teleop_config)
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+teleop = SO100Leader(teleop_config)

-    # Configure the dataset features
-    action_features = hw_to_dataset_features(robot.action_features, "action")
-    obs_features = hw_to_dataset_features(robot.observation_features, "observation")
-    dataset_features = {**action_features, **obs_features}
+# Configure the dataset features
+action_features = hw_to_dataset_features(robot.action_features, "action")
+obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+dataset_features = {**action_features, **obs_features}

-    # Create the dataset
-    dataset = LeRobotDataset.create(
-        repo_id="<hf_username>/<dataset_repo_id>",
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id="<hf_username>/<dataset_repo_id>",
+    fps=FPS,
+    features=dataset_features,
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Initialize the keyboard listener and rerun visualization
+_, events = init_keyboard_listener()
+init_rerun(session_name="recording")
+
+# Connect the robot and teleoperator
+robot.connect()
+teleop.connect()
+
+# Create the required processors
+teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+
+episode_idx = 0
+while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    record_loop(
+        robot=robot,
+        events=events,
        fps=FPS,
-        features=dataset_features,
-        robot_type=robot.name,
-        use_videos=True,
-        image_writer_threads=4,
+        teleop_action_processor=teleop_action_processor,
+        robot_action_processor=robot_action_processor,
+        robot_observation_processor=robot_observation_processor,
+        teleop=teleop,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
    )

-    # Initialize the keyboard listener and rerun visualization
-    _, events = init_keyboard_listener()
-    init_rerun(session_name="recording")
-
-    # Connect the robot and teleoperator
-    robot.connect()
-    teleop.connect()
-
-    # Create the required processors
-    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
-
-    episode_idx = 0
-    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
-
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+        log_say("Reset the environment")
        record_loop(
            robot=robot,
            events=events,
@@ -291,50 +291,26 @@ def main():
            robot_action_processor=robot_action_processor,
            robot_observation_processor=robot_observation_processor,
            teleop=teleop,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
+            control_time_s=RESET_TIME_SEC,
            single_task=TASK_DESCRIPTION,
            display_data=True,
        )

-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-            log_say("Reset the environment")
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                teleop_action_processor=teleop_action_processor,
-                robot_action_processor=robot_action_processor,
-                robot_observation_processor=robot_observation_processor,
-                teleop=teleop,
-                control_time_s=RESET_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-            )
+    if events["rerecord_episode"]:
+        log_say("Re-recording episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue

-        if events["rerecord_episode"]:
-            log_say("Re-recording episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+    dataset.save_episode()
+    episode_idx += 1

-        dataset.save_episode()
-        episode_idx += 1
-
-    # finalize dataset
-    log_say("Finalizing dataset...")
-    dataset.finalize()
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    teleop.disconnect()
-    dataset.push_to_hub()
-
-
-if __name__ == "__main__":
-    main()
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+teleop.disconnect()
+dataset.push_to_hub()
 ```
 <!-- prettier-ignore-end -->

@@ -372,7 +348,7 @@ The `record` function provides a suite of tools for capturing and managing data
 ##### 2. Checkpointing and Resuming

 - Checkpoints are automatically created during recording.
- If an issue occurs or you want to record additional episodes in the same dataset, you can resume by re-running the same command with `--resume=true`. When resuming a recording, `--dataset.num_episodes` must be set to the **number of additional episodes to be recorded**, and not to the targeted total number of episodes in the dataset! Make sure that you also set `--dataset.root="local_path"`, it's a local path to save the new part of the dataset and is required to resume.
+- If an issue occurs, you can resume by re-running the same command with `--resume=true`. When resuming a recording, `--dataset.num_episodes` must be set to the **number of additional episodes to be recorded**, and not to the targeted total number of episodes in the dataset !
 - To start recording from scratch, **manually delete** the dataset directory.

 ##### 3. Recording Parameters
@@ -446,7 +422,7 @@ from lerobot.utils.utils import log_say

 episode_idx = 0

-robot_config = SO100FollowerConfig(port="/dev/tty.usbmodem5AB90687491", id="my_follower_arm")
+robot_config = SO100FollowerConfig(port="/dev/tty.usbmodem58760434471", id="my_awesome_follower_arm")

 robot = SO100Follower(robot_config)
 robot.connect()
@@ -514,83 +490,6 @@ Additionally you can provide extra `tags` or specify a `license` for your model

 If your local computer doesn't have a powerful GPU you could utilize Google Colab to train your model by following the [ACT training notebook](./notebooks#training-act).

-#### Train using Hugging Face Jobs
-
-Hugging Face jobs let's you easily select hardware and run the training in the cloud. So if you don't have a powerful GPU or you need more VRAM or just want to train a model much faster use HF Jobs! It's pay as you go and you simply pay for each second of use, you can see the pricing and additional information [here](https://huggingface.co/docs/hub/jobs).
-
-To run the training use this command:
-
-<hfoptions id="train_with_hf_jobs">
-<hfoption id="Command">
-```bash
-hf jobs run \
-  --flavor a10g-small \
-  --timeout 4h \
-  --secrets HF_TOKEN \
-  huggingface/lerobot-gpu:latest \
-  -- \
-  python -m lerobot.scripts.lerobot_train \
-    --dataset.repo_id=username/dataset \
-    --policy.type=act \
-    --steps=5000 \
-    --batch_size=16 \
-    --policy.device=cuda \
-    --policy.repo_id=username/your_policy \
-    --log_freq=100
-```
-</hfoption>
-<hfoption id="API example">
-
-<!-- prettier-ignore-start -->
-```python
-from huggingface_hub import run_job, get_token
-
-run_name = "act_so101_hf_jobs"
-dataset_id = "username/dataset"
-user_hub_id = "username"
-
-command_args = [
-    "python", "-m", "lerobot.scripts.lerobot_train",
-    "--dataset.repo_id", dataset_id,
-    "--policy.type", "act",
-    "--steps", "5000",
-    "--batch_size", "16",
-    "--num_workers", "4",
-    "--policy.device", "cuda",
-    "--log_freq", "100",
-    "--save_freq", "1000",
-    "--save_checkpoint", "true",
-    "--wandb.enable", "false",
-    "--policy.repo_id", f"{user_hub_id}/{run_name}"
-]
-
-print(f"Submitting job '{run_name}' to Hugging Face Infrastructure...")
-
-job_info = run_job(
-    image="huggingface/lerobot-gpu:latest",
-    command=command_args,
-    flavor="a10g-small",
-    timeout="4h",
-    secrets={"HF_TOKEN": get_token()}
-)
-
-print("\n🚀 Job successfully launched!")
-print(f"🔹 Job ID: {job_info.id}")
-print(f"🔗 Live UI Dashboard & Logs: {job_info.url}")
-```
-<!-- prettier-ignore-end -->
-
-</hfoption>
-</hfoptions>
-
-You can modify the `--flavor` to use different hardware, for example: `t4-small`, `a100-large`, `h200`. Use `hf jobs hardware` to see the full list with pricing.
-Depending on the model you want to train and the hardware you selected you can also modify the `--batch_size` and `--number_of_workers`.
-For longer training sessions increase the timeout.
-
-Once the training is started you can go to [Jobs](https://huggingface.co/settings/jobs) and see if your jobs is running as well as all the outputs. Sometimes it takes a few minutes to schedule your job so be patient.
-
-After training the model will be pushed to hub and you can use it as any other model with LeRobot.
-
 #### Upload policy checkpoints

 Once training is done, upload the latest checkpoint with:
--- a/docs/source/inference.mdx
+++ b/docs/source/inference.mdx
@@ -59,7 +59,7 @@ lerobot-rollout \
    --robot.type=so100_follower \
    --robot.port=/dev/ttyACM0 \
    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-    --dataset.repo_id=${HF_USER}/rollout_eval_data \
+    --dataset.repo_id=${HF_USER}/eval_data \
    --dataset.single_task="Put lego brick into the box" \
    --duration=3600
 ```
@@ -84,7 +84,7 @@ lerobot-rollout \
    --policy.path=${HF_USER}/my_policy \
    --robot.type=koch_follower \
    --robot.port=/dev/ttyACM0 \
-    --dataset.repo_id=${HF_USER}/rollout_highlight_data \
+    --dataset.repo_id=${HF_USER}/highlight_data \
    --dataset.single_task="Pick up the red cube"
 ```

@@ -118,7 +118,7 @@ lerobot-rollout \
    --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
    --robot.type=bi_openarm_follower \
    --teleop.type=openarm_mini \
-    --dataset.repo_id=${HF_USER}/rollout_hil_data \
+    --dataset.repo_id=${HF_USER}/hil_data \
    --dataset.single_task="Fold the T-shirt"
 ```

@@ -134,7 +134,7 @@ lerobot-rollout \
    --robot.port=/dev/ttyACM0 \
    --teleop.type=so101_leader \
    --teleop.port=/dev/ttyACM1 \
-    --dataset.repo_id=${HF_USER}/rollout_dagger_data \
+    --dataset.repo_id=${HF_USER}/dagger_data \
    --dataset.single_task="Grasp the block"
 ```

--- a/docs/source/installation.mdx
+++ b/docs/source/installation.mdx
@@ -207,56 +207,6 @@ pip install 'lerobot[feetech]'        # Feetech motor support

 _Multiple extras can be combined (e.g., `.[core_scripts,pi,pusht]`). For a full list of available extras, refer to `pyproject.toml`._

-### PyTorch CUDA variant (Linux only)
-
-On Linux, the install path determines which CUDA wheel you get. macOS and Windows installs use the PyPI default (MPS / CPU / CUDA-Windows wheel respectively) and can skip this section.
-
-<!-- prettier-ignore-start -->
-
-<hfoptions id="cuda_variant">
-<hfoption id="uv-source">
-
-**Source install via `uv` (`uv sync` or `uv pip install -e .`)**
-
-`torch` and `torchvision` are pinned by the project to the **CUDA 12.8** PyTorch index (`https://download.pytorch.org/whl/cu128`, driver floor **570.86**) — covers Ampere/Ada/Hopper/Blackwell GPUs. No action needed for typical NVIDIA setups.
-
-To override for a different CUDA variant:
-
-```bash
-uv pip install --force-reinstall torch torchvision \
-    --index-url https://download.pytorch.org/whl/cu126   # older drivers; or cu130 for Blackwell on driver ≥ 580
-```
-
-</hfoption>
-<hfoption id="pip-conda">
-
-**Source install via `pip`/`conda`, or `pip install lerobot` from PyPI**
-
-PyPI default torch wheel is currently a cu130-bundled Linux wheel, driver floor **580.65**.
-
-To pick a specific CUDA variant:
-
-**Using `pip` or `conda`** — install torch first with an explicit index, then lerobot:
-
-```bash
-pip install --index-url https://download.pytorch.org/whl/cu128 torch torchvision
-pip install -e ".[all]"          # source
-# — or —
-pip install lerobot              # from PyPI
-```
-
-**Using `uv` to install from PyPI** — one-liner via `--torch-backend` (uv ≥ 0.6):
-
-```bash
-uv pip install --torch-backend cu128 lerobot
-```
-
-Supported values include `auto`, `cpu`, `cu126`, `cu128`, `cu129`, `cu130`, plus various `rocm*` and `xpu`. Swap as needed for your driver.
-
-</hfoption>
-</hfoptions>
-<!-- prettier-ignore-end -->
-
 ### Troubleshooting

 If you encounter build errors, you may need to install additional system dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
--- a/docs/source/language_and_recipes.mdx
+++ b/docs/source/language_and_recipes.mdx
@@ -1,147 +0,0 @@
-# Language columns and recipes
-
-Most LeRobot datasets ship with a single `task` string per episode — fine for
-short, single-instruction skills, but not enough for the longer-horizon,
-multi-modal robot policies the field is moving toward (high-level planning,
-memory, interjections, VQA, tool use). To support those policies without
-forking the dataset format, LeRobot extends `LeRobotDataset` with two optional
-language columns and a small recipe layer that turns those rows into
-chat-style training samples on the fly.
-
-The design splits cleanly into three layers:
-
-1. **Data in the dataset** — language annotations stored next to frames in
-   `data/chunk-*/file-*.parquet` as two optional columns (`language_persistent`
-   and `language_events`). Datasets without these columns keep their existing
-   behavior.
-2. **Recipe** — a YAML file that declares which annotation rows to bind and
-   how to lay them out as chat turns (`role`, `content`, optional images,
-   optional tool calls). Recipes are pure config; no Python required to add a
-   new one.
-3. **Training format** — at sample time, `RenderMessagesStep` resolves the
-   recipe against the per-frame annotations and emits HF-style `messages` plus
-   LeRobot-specific sidecars (`message_streams`, `target_message_indices`)
-   that policy processors consume.
-
-This page describes each layer in turn.
-
-## Layer 1 — language columns in the dataset
-
-The two optional columns live next to frame data in
-`data/chunk-*/file-*.parquet`:
-
- `language_persistent`: a list of rows broadcast across every frame in an episode for state that remains active, such as `subtask`, `plan`, and `memory`.
- `language_events`: a list of rows only on the exact frame where an event was emitted, such as `interjection`, `vqa`, and speech tool calls.
-
-Both columns share the same row shape (event rows omit `timestamp` because the
-frame the row sits on already provides it):
-
-```text
-role: string
-content: string | null
-style: string | null
-timestamp: float32        # persistent rows only
-camera: string | null     # observation.images.* feature key, view-dependent rows only
-tool_calls: list[Json] | null
-```
-
-The `camera` field tags rows whose `content` is grounded in a specific camera
-view. Rows of view-dependent styles (`vqa` and `trace`) MUST set `camera` to
-the matching `observation.images.*` feature key. Rows of every other style —
-including `motion`, which describes robot-frame primitives in joint / Cartesian
-terms — MUST leave `camera` as `null`. Pipeline writers and the validator
-enforce this via `validate_camera_field(style, camera)`.
-
-`meta/tasks.parquet` remains the canonical source for the task. The special `${task}` recipe binding always reads that task string and does not depend on language annotations.
-
-### Architecture
-
-The language stack itself has three internal modules backing layer 1:
-
-1. `lerobot.datasets.language` defines the schema, style registry, and `column_for_style`.
-2. `lerobot.datasets.language_render` resolves rows and renders messages.
-3. `RenderMessagesStep` turns dataset samples into `messages`, `message_streams`, and `target_message_indices`.
-
-`LeRobotDataset` stays recipe-agnostic. It passes `language_persistent` and `language_events` through when present, and unannotated datasets keep their existing behavior.
-
-## Layer 2 — recipe anatomy
-
-Recipes are YAML files backed by `TrainingRecipe` and `MessageTurn`. They
-declare which annotation rows to pull (via `bindings`) and how to compose them
-into chat turns (`messages`).
-
-```yaml
-messages:
-  - { role: user, content: "${task}", stream: high_level }
-  - { role: assistant, content: "${subtask}", stream: low_level, target: true }
-```
-
-A recipe can also branch into a weighted **blend** of sub-recipes. At sample
-time, exactly one branch is selected deterministically from the sample index,
-so different frames train different objectives (e.g. memory updates vs.
-low-level execution vs. VQA) without any Python wiring.
-
-### Temporal semantics
-
-Persistent styles are active after emission until replaced:
-
- `active_at(t, style=subtask)`
- `nth_prev(style=memory, offset=1)`
- `nth_next(style=subtask, offset=1)`
-
-Event styles only exist on their exact timestamp:
-
- `emitted_at(t, style=interjection)`
- `emitted_at(t, style=vqa, role=user, camera=observation.images.top)`
- `emitted_at(t, role=assistant, tool_name=say)`
-
-Exact event matching has no tolerance window, so writers must stamp event rows with frame timestamps from the parquet data.
-
-### View-dependent resolution
-
-For view-dependent styles (`vqa` and `trace`), the resolver gains a
-`camera=` filter parallel to `role=` and `tool_name=`. Datasets with multiple
-cameras typically emit one (`vqa`, `user`) + (`vqa`, `assistant`) pair per
-camera at the same timestamp; without `camera=`, those resolvers see two
-matches and raise an ambiguity error. Recipes consume each camera through its
-own binding plus a matching image block, e.g.
-
-```yaml
-ask_vqa_top:
-  bindings:
-    vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
-    vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
-  messages:
-    - role: user
-      stream: high_level
-      if_present: vqa_query
-      content:
-        - { type: image, feature: observation.images.top }
-        - { type: text, text: "${vqa_query}" }
-    - {
-        role: assistant,
-        content: "${vqa}",
-        stream: high_level,
-        target: true,
-        if_present: vqa,
-      }
-```
-
-Add one such sub-recipe per camera the dataset records.
-
-## Layer 3 — training format
-
-Rendered samples use HF-style chat messages plus LeRobot sidecars:
-
-```python
-sample["messages"]
-sample["message_streams"]
-sample["target_message_indices"]
-```
-
-The renderer does not apply a tokenizer chat template. Policy processors decide how to serialize the messages for their backbone, which keeps the same dataset usable across SmolVLA, Pi0.5, and any future VLM that expects OpenAI-style chat messages.
-
-## Graceful absence
-
-If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.
-If an event-scoped branch is selected on a frame without the required event row, rendering returns `None`, allowing a loader to retry another sample.
--- a/docs/source/lerobot-dataset-v3.mdx
+++ b/docs/source/lerobot-dataset-v3.mdx
@@ -10,7 +10,6 @@ This docs will guide you to:
 - Stream datasets without downloading using `StreamingLeRobotDataset`
 - Apply image transforms for data augmentation during training
 - Migrate existing `v2.1` datasets to `v3.0`
- Experiment with other `LeRobotDataset` formats and implementations like Lance

 ## What’s new in `v3`

@@ -44,7 +43,7 @@ lerobot-record \
  --dataset.num_episodes=5 \
  --dataset.single_task="Grab the black cube" \
  --dataset.streaming_encoding=true \
-  # --dataset.camera_encoder.vcodec=auto \
+  # --dataset.vcodec=auto \
  --dataset.encoder_threads=2
 ```

@@ -316,39 +315,3 @@ Dataset v3.0 uses incremental parquet writing with buffered metadata for efficie
 - Ensures the dataset is valid for loading

 Without calling `finalize()`, your parquet files will be incomplete and the dataset won't load properly.
-
-## Other formats and implementations
-
-### Lance
-
-Lance is a useful format for multimodal AI datasets, especially for large-scale training requiring high performance IO and random access.
-
-The `lerobot-lancedb` package implements `LeRobotLanceDataset` (for JPEG images) and `LeRobotLanceVideoDataset` (for mp4 videos).
-Those two storage layouts both subclass LeRobotDataset and can provide data loading speed ups.
-
-`LeRobotLanceDataset` is a drop-in replacement for `LeRobotDataset`:
-
-```python
-from lerobot.datasets import LeRobotDatasetMetadata
-from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
-from lerobot_lancedb import LeRobotLanceDataset, LeRobotLanceVideoDataset
-
-cfg = DiffusionConfig(...)
-meta = LeRobotDatasetMetadata(root=local_dataset_path)  # or use repo_id=... to load metadata from the Hub
-delta_timestamps = {...}
-
-# Use LeRobotLanceDataset for image datasets
-dataset = LeRobotLanceDataset(
-    root=local_dataset_path,                            # or use repo_id=... to stream from the Hub
-    delta_timestamps=delta_timestamps,
-    return_uint8=True,
-)
-# Or use LeRobotLanceVideoDataset for video datasets:
-dataset = LeRobotLanceVideoDataset(
-    root=local_dataset_path,                            # or use repo_id=... to stream from the Hub
-    delta_timestamps=delta_timestamps,
-    return_uint8=True,
-)
-```
-
-Join the discussion on [Github](https://github.com/huggingface/lerobot/issues/3608) and explore the `lerobot-lancedb` documentation [here](https://lancedb.github.io/lerobot-lancedb/).
--- a/docs/source/libero_plus.mdx
+++ b/docs/source/libero_plus.mdx
@@ -1,188 +0,0 @@
-# LIBERO-plus
-
-LIBERO-plus is a **robustness benchmark** for Vision-Language-Action (VLA) models built on top of [LIBERO](./libero). It systematically stress-tests policies by applying **seven independent perturbation dimensions** to the original LIBERO task set, exposing failure modes that standard benchmarks miss.
-
- Paper: [In-depth Robustness Analysis of Vision-Language-Action Models](https://arxiv.org/abs/2510.13626)
- GitHub: [sylvestf/LIBERO-plus](https://github.com/sylvestf/LIBERO-plus)
- Dataset: [lerobot/libero_plus](https://huggingface.co/datasets/lerobot/libero_plus)
-
-![An overview of the LIBERO-plus benchmark perturbation dimensions](https://github.com/sylvestf/LIBERO-plus/raw/main/static/images/libero-plus.jpg)
-
-## Perturbation dimensions
-
-LIBERO-plus creates ~10 000 task variants by perturbing each original LIBERO task along these axes:
-
-| Dimension             | What changes                                          |
-| --------------------- | ----------------------------------------------------- |
-| Objects layout        | Target position, presence of confounding objects      |
-| Camera viewpoints     | Camera position, orientation, field-of-view           |
-| Robot initial states  | Manipulator start pose                                |
-| Language instructions | LLM-rewritten task description (paraphrase / synonym) |
-| Light conditions      | Intensity, direction, color, shadow                   |
-| Background textures   | Scene surface and object appearance                   |
-| Sensor noise          | Photometric distortions and image degradation         |
-
-## Available task suites
-
-LIBERO-plus covers the same five suites as LIBERO:
-
-| Suite          | CLI name         | Tasks | Max steps | Description                                        |
-| -------------- | ---------------- | ----- | --------- | -------------------------------------------------- |
-| LIBERO-Spatial | `libero_spatial` | 10    | 280       | Tasks requiring reasoning about spatial relations  |
-| LIBERO-Object  | `libero_object`  | 10    | 280       | Tasks centered on manipulating different objects   |
-| LIBERO-Goal    | `libero_goal`    | 10    | 300       | Goal-conditioned tasks with changing targets       |
-| LIBERO-90      | `libero_90`      | 90    | 400       | Short-horizon tasks from the LIBERO-100 collection |
-| LIBERO-Long    | `libero_10`      | 10    | 520       | Long-horizon tasks from the LIBERO-100 collection  |
-
-<Tip warning={true}>
-  Installing LIBERO-plus **replaces** vanilla LIBERO — it uninstalls `hf-libero`
-  so that `import libero` resolves to the LIBERO-plus fork. You cannot have both
-  installed at the same time. To switch back to vanilla LIBERO, uninstall the
-  fork and reinstall with `pip install -e ".[libero]"`.
-</Tip>
-
-## Installation
-
-### System dependencies (Linux only)
-
-```bash
-sudo apt install libexpat1 libfontconfig1-dev libmagickwand-dev
-```
-
-### Python package
-
-```bash
-pip install -e ".[libero]" "robosuite==1.4.1" bddl easydict mujoco wand scikit-image gym
-git clone https://github.com/sylvestf/LIBERO-plus.git
-cd LIBERO-plus && pip install --no-deps -e .
-pip uninstall -y hf-libero  # so `import libero` resolves to the fork
-```
-
-LIBERO-plus is installed from its GitHub fork rather than a pyproject extra — the fork ships as a namespace package that pip can't handle, so it must be cloned and added to `PYTHONPATH`. See `docker/Dockerfile.benchmark.libero_plus` for the canonical install. MuJoCo is required, so only Linux is supported.
-
-<Tip>
-Set the MuJoCo rendering backend before running evaluation:
-
-```bash
-export MUJOCO_GL=egl   # headless / HPC / cloud
-```
-
-</Tip>
-
-### Download LIBERO-plus assets
-
-LIBERO-plus ships its extended asset pack separately. Download `assets.zip` from the [Hugging Face dataset](https://huggingface.co/datasets/Sylvest/LIBERO-plus/tree/main) and extract it into the LIBERO-plus package directory:
-
-```bash
-# After installing the package, find where it was installed:
-python -c "import libero; print(libero.__file__)"
-# Then extract assets.zip into <package_root>/libero/assets/
-```
-
-## Evaluation
-
-### Default evaluation (recommended)
-
-Evaluate across the four standard suites (10 episodes per task):
-
-```bash
-lerobot-eval \
-  --policy.path="your-policy-id" \
-  --env.type=libero_plus \
-  --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --env.max_parallel_tasks=1
-```
-
-### Single-suite evaluation
-
-Evaluate on one LIBERO-plus suite:
-
-```bash
-lerobot-eval \
-  --policy.path="your-policy-id" \
-  --env.type=libero_plus \
-  --env.task=libero_spatial \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10
-```
-
- `--env.task` picks the suite (`libero_spatial`, `libero_object`, etc.).
- `--env.task_ids` restricts to specific task indices (`[0]`, `[1,2,3]`, etc.). Omit to run all tasks in the suite.
- `--eval.batch_size` controls how many environments run in parallel.
- `--eval.n_episodes` sets how many episodes to run per task.
-
-### Multi-suite evaluation
-
-Benchmark a policy across multiple suites at once by passing a comma-separated list:
-
-```bash
-lerobot-eval \
-  --policy.path="your-policy-id" \
-  --env.type=libero_plus \
-  --env.task=libero_spatial,libero_object \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10
-```
-
-### Control mode
-
-LIBERO-plus supports two control modes — `relative` (default) and `absolute`. Different VLA checkpoints are trained with different action parameterizations, so make sure the mode matches your policy:
-
-```bash
--env.control_mode=relative   # or "absolute"
-```
-
-### Policy inputs and outputs
-
-**Observations:**
-
- `observation.state` — 8-dim proprioceptive features (eef position, axis-angle orientation, gripper qpos)
- `observation.images.image` — main camera view (`agentview_image`), HWC uint8
- `observation.images.image2` — wrist camera view (`robot0_eye_in_hand_image`), HWC uint8
-
-**Actions:**
-
- Continuous control in `Box(-1, 1, shape=(7,))` — 6D end-effector delta + 1D gripper
-
-### Recommended evaluation episodes
-
-For reproducible benchmarking, use **10 episodes per task** across all four standard suites (Spatial, Object, Goal, Long). This gives 400 total episodes and matches the protocol used for published results.
-
-## Training
-
-### Dataset
-
-A LeRobot-format training dataset for LIBERO-plus is available at:
-
- [lerobot/libero_plus](https://huggingface.co/datasets/lerobot/libero_plus)
-
-### Example training command
-
-```bash
-lerobot-train \
-    --policy.type=smolvla \
-    --policy.repo_id=${HF_USER}/smolvla_libero_plus \
-    --policy.load_vlm_weights=true \
-    --dataset.repo_id=lerobot/libero_plus \
-    --env.type=libero_plus \
-    --env.task=libero_spatial \
-    --output_dir=./outputs/ \
-    --steps=100000 \
-    --batch_size=4 \
-    --eval.batch_size=1 \
-    --eval.n_episodes=1 \
-    --eval_freq=1000
-```
-
-## Relationship to LIBERO
-
-LIBERO-plus is a drop-in extension of LIBERO:
-
- Same Python gym interface (`LiberoEnv`, `LiberoProcessorStep`)
- Same camera names and observation/action format
- Same task suite names
- Installs under the same `libero` Python package name (different GitHub repo)
-
-To use the original LIBERO benchmark, see [LIBERO](./libero) and use `--env.type=libero`.
--- a/docs/source/peft_training.mdx
+++ b/docs/source/peft_training.mdx
@@ -28,15 +28,13 @@ lerobot-train \
 --steps=100000 \
 --batch_size=32 \
 --peft.method_type=LORA \
- --peft.r=64 \
- --peft.lora_alpha=64
+ --peft.r=64
 ```

 Note the `--peft.method_type` parameter that let's you select which PEFT method to use. Here we use
 [LoRA](https://huggingface.co/docs/peft/main/en/package_reference/lora) (Low-Rank Adapter) which is probably the most
 popular fine-tuning method to date. Low-rank adaption means that we only fine-tune a matrix with comparably low rank
-instead of the full weight matrix. This rank can be specified using the `--peft.r` parameter, and the LoRA scaling factor with
-`--peft.lora_alpha` (where `scaling = lora_alpha / r`). The higher the rank
+instead of the full weight matrix. This rank can be specified using the `--peft.r` parameter. The higher the rank
 the closer you get to full fine-tuning

 There are more complex methods that have more parameters. These are not yet supported, feel free to raise an issue
--- a/docs/source/reachy2.mdx
+++ b/docs/source/reachy2.mdx
@@ -161,7 +161,7 @@ lerobot-record \
    --dataset.private=true \
    --dataset.streaming_encoding=true \
    --dataset.encoder_threads=2 \
-    # --dataset.camera_encoder.vcodec=auto \
+    # --dataset.vcodec=auto \
    --display_data=true
 ```

@@ -203,7 +203,7 @@ lerobot-record \
    --dataset.private=true \
    --dataset.streaming_encoding=true \
    --dataset.encoder_threads=2 \
-    # --dataset.camera_encoder.vcodec=auto \
+    # --dataset.vcodec=auto \
    --display_data=true
 ```

--- a/docs/source/rebot_b601.mdx
+++ b/docs/source/rebot_b601.mdx
@@ -1,186 +0,0 @@
-# reBot B601-DM
-
-[reBot B601-DM](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/) is an open-source, low-cost robot arm from Seeed Studio for embodied-AI and imitation learning. It comes as a **follower** arm (the `B601-DM`, a 6-DOF arm plus gripper driven by Damiao CAN motors) and a **leader** arm (the `StarArm102` / `reBot Arm 102`, driven by FashionStar UART smart servos) used to teleoperate it.
-
-This page covers **calibration** and **teleoperation** for both single-arm and bimanual (dual-arm) setups.
-
-<div style="display: flex; align-items: center; gap: 10px;">
-  <img
-    src="https://files.seeedstudio.com/wiki/robotics/projects/lerobot/b601dm_zeroposition.jpg"
-    alt="reBot B601-DM follower arm at its zero position"
-    width="48%"
-  />
-  <img
-    src="https://files.seeedstudio.com/wiki/robotics/projects/lerobot/102_zeroposition.jpg"
-    alt="reBot Arm 102 leader arm at its zero position"
-    width="48%"
-  />
-</div>
-
-_Left: the B601-DM follower at its zero position. Right: the reBot Arm 102 leader at its zero position. Images courtesy of [Seeed Studio](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/)._
-
-## Install LeRobot 🤗
-
-Follow our [Installation Guide](./installation), then install the reBot support:
-
-```bash
-pip install -e ".[rebot]"
-```
-
-This pulls in `motorbridge` (CAN motor control for the B601-DM follower) and `motorbridge-smart-servo` (FashionStar UART servos for the reBot Arm 102 leader).
-
-## Registered device types
-
-| Type                     | Kind                                         |
-| ------------------------ | -------------------------------------------- |
-| `rebot_b601_follower`    | single-arm B601-DM follower robot            |
-| `bi_rebot_b601_follower` | bimanual (dual-arm) follower robot           |
-| `rebot_102_leader`       | single-arm reBot Arm 102 leader teleoperator |
-| `bi_rebot_102_leader`    | bimanual (dual-arm) leader teleoperator      |
-
-The bimanual types compose two single-arm instances and namespace each arm's
-observation/action keys with a `left_` / `right_` prefix. Per-arm settings are
-passed through nested `left_arm_config.*` / `right_arm_config.*` arguments.
-
-## Find the USB ports
-
-For each device, find the USB port associated with its motor bus using:
-
-```bash
-lerobot-find-port
-```
-
-<Tip warning={true}>
-  On Linux, remove `brltty` (`sudo apt remove brltty`) so it does not hold the
-  leader's USB serial port. You may also need to grant access to the serial
-  devices: `sudo chmod 666 /dev/ttyACM* /dev/ttyUSB*`.
-</Tip>
-
-## Calibration
-
-Neither arm stores a persistent hardware calibration: every time it connects, the motors are re-zeroed against the pose the arm is physically holding. Calibration simply records that zero pose. When prompted, **manually move the arm to its zero position** (the default sit-down pose shown above, gripper fully closed) and press <kbd>ENTER</kbd>.
-
-### Follower (B601-DM)
-
-<hfoptions id="calibrate-follower">
-<hfoption id="Single arm">
-
-```bash
-lerobot-calibrate \
-    --robot.type=rebot_b601_follower \
-    --robot.port=/dev/ttyACM0 \
-    --robot.id=follower \
-    --robot.can_adapter=damiao
-```
-
-</hfoption>
-<hfoption id="Dual arm">
-
-Connect the bimanual follower; calibration runs for the left arm, then the right arm.
-
-```bash
-lerobot-calibrate \
-    --robot.type=bi_rebot_b601_follower \
-    --robot.id=bi_follower \
-    --robot.left_arm_config.port=/dev/ttyACM0 \
-    --robot.left_arm_config.can_adapter=damiao \
-    --robot.right_arm_config.port=/dev/ttyACM1 \
-    --robot.right_arm_config.can_adapter=damiao
-```
-
-Per-arm calibration files are saved with `_left` / `_right` suffixes on the id.
-
-</hfoption>
-</hfoptions>
-
-### Leader (reBot Arm 102)
-
-<hfoptions id="calibrate-leader">
-<hfoption id="Single arm">
-
-```bash
-lerobot-calibrate \
-    --teleop.type=rebot_102_leader \
-    --teleop.port=/dev/ttyUSB0 \
-    --teleop.id=leader
-```
-
-</hfoption>
-<hfoption id="Dual arm">
-
-```bash
-lerobot-calibrate \
-    --teleop.type=bi_rebot_102_leader \
-    --teleop.id=bi_leader \
-    --teleop.left_arm_config.port=/dev/ttyUSB0 \
-    --teleop.right_arm_config.port=/dev/ttyUSB1
-```
-
-</hfoption>
-</hfoptions>
-
-## Teleoperation
-
-Once both arms are calibrated, drive the follower with the leader. The follower talks to its CAN bus through a Damiao serial bridge (`can_adapter=damiao`, the default) or a SocketCAN adapter (`can_adapter=socketcan`). See the [OpenArm page](./openarm) for more details on the SocketCAN adapter configuration.
-
-<hfoptions id="teleoperate">
-<hfoption id="Single arm">
-
-```bash
-lerobot-teleoperate \
-    --robot.type=rebot_b601_follower \
-    --robot.port=/dev/ttyACM0 \
-    --robot.id=follower \
-    --robot.can_adapter=damiao \
-    --teleop.type=rebot_102_leader \
-    --teleop.port=/dev/ttyUSB0 \
-    --teleop.id=leader
-```
-
-</hfoption>
-<hfoption id="Dual arm">
-
-The bimanual leader and follower reuse the single-arm classes; each arm is
-configured through nested `left_arm_config.*` / `right_arm_config.*` arguments,
-so a bimanual reBot Arm 102 leader drives a bimanual B601-DM follower.
-
-```bash
-lerobot-teleoperate \
-    --robot.type=bi_rebot_b601_follower \
-    --robot.id=bi_follower \
-    --robot.left_arm_config.port=/dev/ttyACM0 \
-    --robot.left_arm_config.can_adapter=damiao \
-    --robot.right_arm_config.port=/dev/ttyACM1 \
-    --robot.right_arm_config.can_adapter=damiao \
-    --teleop.type=bi_rebot_102_leader \
-    --teleop.id=bi_leader \
-    --teleop.left_arm_config.port=/dev/ttyUSB0 \
-    --teleop.right_arm_config.port=/dev/ttyUSB1
-```
-
-</hfoption>
-</hfoptions>
-
-<Tip>
-  The leader and follower share the same joint names (`shoulder_pan,
-  shoulder_lift, elbow_flex, wrist_flex, wrist_yaw, wrist_roll, gripper`), so
-  leader actions map directly onto the follower.
-</Tip>
-
-If the motion of a joint is reversed, flip its sign in the leader's `joint_directions` (the gripper also carries a scale to widen its range to the follower):
-
-```bash
-lerobot-teleoperate \
-    --robot.type=rebot_b601_follower \
-    --robot.port=/dev/ttyACM0 \
-    --robot.can_adapter=damiao \
-    --teleop.type=rebot_102_leader \
-    --teleop.port=/dev/ttyUSB0 \
-    --teleop.joint_directions='{"shoulder_pan":-1,"shoulder_lift":-1,"elbow_flex":1,"wrist_flex":1,"wrist_yaw":1,"wrist_roll":-1,"gripper":-6}'
-```
-
-## Recording datasets
-
-Swap `lerobot-teleoperate` for `lerobot-record` (with the same `--robot.*` / `--teleop.*` arguments, plus `--dataset.*`) to record demonstrations for training. See [Imitation Learning for Robots](./il_robots) for the full workflow.
-
-For hardware assembly and wiring, see the [Seeed Studio reBot wiki](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/).
--- a/docs/source/rename_map.mdx
+++ b/docs/source/rename_map.mdx
@@ -61,6 +61,17 @@ lerobot-eval \
  --rename_map='{"observation.images.image": "observation.images.base_0_rgb", "observation.images.image2": "observation.images.left_wrist_0_rgb"}'
 ```

+### Recording
+
+`lerobot-record` also supports rename maps, nested under the dataset config:
+
+```bash
+lerobot-record \ # When running inference
+  --policy.path="<user>/smolVLA_finetuned" \
+  ... \
+  --dataset.rename_map='{"observation.images.glove2": "observation.images.image"}'
+```
+
 ## Alternative: edit the policy config directly

 If you always use the same dataset or environment, you can **edit the policy's `config.json`** so its observation keys match your data source. Then no rename map is needed.
@@ -94,10 +105,10 @@ XVLA-base has three visual inputs and `empty_cameras=0` by default. Your dataset

 ## Quick reference

-| Goal                                    | What to do                                                                  |
-| --------------------------------------- | --------------------------------------------------------------------------- |
-| Dataset keys ≠ policy keys              | `--rename_map='{"dataset_key": "policy_key", ...}'`                         |
-| Env keys ≠ policy keys (eval)           | `--rename_map='{"env_key": "policy_key", ...}'`                             |
-| Rollout with different keys (inference) | `--rename_map='{"source_key": "policy_key", ...}'`.                         |
-| Fewer cameras than policy expects       | `--policy.empty_cameras=N` (supported by PI0, PI05, PI0Fast, SmolVLA, XVLA) |
-| Avoid passing a rename map              | Edit the policy's `config.json` so its keys match your data source          |
+| Goal                                      | What to do                                                                  |
+| ----------------------------------------- | --------------------------------------------------------------------------- |
+| Dataset keys ≠ policy keys                | `--rename_map='{"dataset_key": "policy_key", ...}'`                         |
+| Env keys ≠ policy keys (eval)             | `--rename_map='{"env_key": "policy_key", ...}'`                             |
+| Recording with different keys (inference) | `--dataset.rename_map='{"source_key": "policy_key", ...}'`.                 |
+| Fewer cameras than policy expects         | `--policy.empty_cameras=N` (supported by PI0, PI05, PI0Fast, SmolVLA, XVLA) |
+| Avoid passing a rename map                | Edit the policy's `config.json` so its keys match your data source          |
--- a/docs/source/robocasa.mdx
+++ b/docs/source/robocasa.mdx
@@ -1,188 +0,0 @@
-# RoboCasa365
-
-[RoboCasa365](https://robocasa.ai) is a large-scale simulation framework for training and benchmarking **generalist robots** in everyday kitchen tasks. It ships 365 diverse manipulation tasks across 2,500 kitchen environments, 3,200+ object assets and 600+ hours of human demonstration data, on a PandaOmron 12-DOF mobile manipulator (Franka arm on a holonomic base).
-
- Paper: [RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots](https://arxiv.org/abs/2406.02523)
- GitHub: [robocasa/robocasa](https://github.com/robocasa/robocasa)
- Project website: [robocasa.ai](https://robocasa.ai)
- Pretrained policy: [`lerobot/smolvla_robocasa`](https://huggingface.co/lerobot/smolvla_robocasa)
- Single-task dataset (CloseFridge): [`pepijn223/robocasa_CloseFridge`](https://huggingface.co/datasets/pepijn223/robocasa_CloseFridge)
-
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/robocasa-banner.webp"
-  alt="RoboCasa365 benchmark overview"
-  width="85%"
-/>
-
-## Available tasks
-
-RoboCasa365 organizes its 365 tasks into two families and three upstream benchmark groups that LeRobot exposes as first-class `--env.task` shortcuts:
-
-| Family    | Tasks | Description                                                                     |
-| --------- | ----- | ------------------------------------------------------------------------------- |
-| Atomic    | ~65   | Single-skill tasks: pick-and-place, door/drawer manipulation, appliance control |
-| Composite | ~300  | Multi-step tasks across 60+ categories: cooking, cleaning, organizing, etc.     |
-
-**Atomic task examples:** `CloseFridge`, `OpenDrawer`, `OpenCabinet`, `TurnOnMicrowave`, `TurnOffStove`, `NavigateKitchen`, `PickPlaceCounterToStove`.
-
-**Composite task categories:** baking, boiling, brewing, chopping, clearing table, defrosting food, loading dishwasher, making tea, microwaving food, washing dishes, and more.
-
-`--env.task` accepts three forms:
-
- a single task name (`CloseFridge`)
- a comma-separated list (`CloseFridge,OpenBlenderLid,PickPlaceCoffee`)
- a benchmark-group shortcut — `atomic_seen`, `composite_seen`, `composite_unseen`, `pretrain50`, `pretrain100`, `pretrain200`, `pretrain300` — which auto-expands to the upstream task list and auto-sets the dataset `split` (`target` or `pretrain`).
-
-## Installation
-
-RoboCasa and its dependency `robosuite` are not published on PyPI, and RoboCasa's own `setup.py` hardcodes `lerobot==0.3.3`, which conflicts with this repo's `lerobot`. LeRobot therefore does **not** expose a `robocasa` extra — install the two packages manually as editable clones (using `--no-deps` on `robocasa` to skip its shadowed `lerobot` pin):
-
-```bash
-# After following the standard LeRobot installation instructions.
-
-git clone https://github.com/robocasa/robocasa.git ~/robocasa
-git clone https://github.com/ARISE-Initiative/robosuite.git ~/robosuite
-pip install -e ~/robocasa --no-deps
-pip install -e ~/robosuite
-
-# Robocasa's runtime deps (the ones its setup.py would have pulled, minus
-# the bad lerobot pin).
-pip install numpy numba scipy mujoco pygame Pillow opencv-python \
-            pyyaml pynput tqdm termcolor imageio h5py lxml hidapi \
-            tianshou gymnasium
-
-python -m robocasa.scripts.setup_macros
-# Lightweight assets (lightwheel object meshes + textures). Enough for
-# the default env out of the box.
-python -m robocasa.scripts.download_kitchen_assets \
-  --type tex tex_generative fixtures_lw objs_lw
-# Optional: full objaverse/aigen registries (~30GB) for richer object
-# variety. Enable at eval time via --env.obj_registries (see below).
-# python -m robocasa.scripts.download_kitchen_assets --type objs_objaverse
-```
-
-<Tip>
-RoboCasa requires MuJoCo. Set the rendering backend before training or evaluation:
-
-```bash
-export MUJOCO_GL=egl  # for headless servers (HPC, cloud)
-```
-
-</Tip>
-
-### Object registries
-
-By default the env samples objects only from the `lightwheel` registry (what `--type objs_lw` ships), which avoids a `Probabilities contain NaN` crash when the objaverse / aigen packs aren't on disk. If you've downloaded the full asset set, enable the full registry at runtime:
-
-```bash
--env.obj_registries='[objaverse,lightwheel]'
-```
-
-## Evaluation
-
-All eval snippets below mirror the CI command (see `.github/workflows/benchmark_tests.yml`). The `--rename_map` argument maps RoboCasa's native camera keys (`robot0_agentview_left` / `robot0_eye_in_hand` / `robot0_agentview_right`) onto the three-camera (`camera1` / `camera2` / `camera3`) input layout the released `smolvla_robocasa` policy was trained on.
-
-### Single-task evaluation (recommended for quick iteration)
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_robocasa \
-  --env.type=robocasa \
-  --env.task=CloseFridge \
-  --eval.batch_size=1 \
-  --eval.n_episodes=20 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
-```
-
-### Multi-task evaluation
-
-Pass a comma-separated list of tasks:
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_robocasa \
-  --env.type=robocasa \
-  --env.task=CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove \
-  --eval.batch_size=1 \
-  --eval.n_episodes=20 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
-```
-
-### Benchmark-group evaluation
-
-Run an entire upstream group (e.g. all 18 `atomic_seen` tasks with `split=target`):
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_robocasa \
-  --env.type=robocasa \
-  --env.task=atomic_seen \
-  --eval.batch_size=1 \
-  --eval.n_episodes=20 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
-```
-
-### Recommended evaluation episodes
-
-**20 episodes per task** for reproducible benchmarking. Matches the protocol used in published results.
-
-## Policy inputs and outputs
-
-**Observations** (raw RoboCasa camera names are preserved verbatim):
-
- `observation.state` — 16-dim proprioceptive state (base position, base quaternion, relative end-effector position, relative end-effector quaternion, gripper qpos)
- `observation.images.robot0_agentview_left` — left agent view, 256×256 HWC uint8
- `observation.images.robot0_eye_in_hand` — wrist camera view, 256×256 HWC uint8
- `observation.images.robot0_agentview_right` — right agent view, 256×256 HWC uint8
-
-**Actions:**
-
- Continuous control in `Box(-1, 1, shape=(12,))` — base motion (4D) + control mode (1D) + end-effector position (3D) + end-effector rotation (3D) + gripper (1D).
-
-## Training
-
-### Single-task example
-
-A ready-to-use single-task dataset is on the Hub:
-[`pepijn223/robocasa_CloseFridge`](https://huggingface.co/datasets/pepijn223/robocasa_CloseFridge).
-
-Fine-tune a SmolVLA base on `CloseFridge`:
-
-```bash
-lerobot-train \
-  --policy.type=smolvla \
-  --policy.repo_id=${HF_USER}/smolvla_robocasa_CloseFridge \
-  --policy.load_vlm_weights=true \
-  --policy.push_to_hub=true \
-  --dataset.repo_id=pepijn223/robocasa_CloseFridge \
-  --env.type=robocasa \
-  --env.task=CloseFridge \
-  --output_dir=./outputs/smolvla_robocasa_CloseFridge \
-  --steps=100000 \
-  --batch_size=4 \
-  --eval_freq=5000 \
-  --eval.batch_size=1 \
-  --eval.n_episodes=5 \
-  --save_freq=10000
-```
-
-Evaluate the resulting checkpoint:
-
-```bash
-lerobot-eval \
-  --policy.path=${HF_USER}/smolvla_robocasa_CloseFridge \
-  --env.type=robocasa \
-  --env.task=CloseFridge \
-  --eval.batch_size=1 \
-  --eval.n_episodes=20
-```
-
-## Reproducing published results
-
-The released checkpoint [`lerobot/smolvla_robocasa`](https://huggingface.co/lerobot/smolvla_robocasa) is evaluated with the commands in the [Evaluation](#evaluation) section. CI runs a 10-atomic-task smoke eval (one episode each) on every PR touching the benchmark, picking fixture-centric tasks that don't require the objaverse asset pack.
--- a/docs/source/robocerebra.mdx
+++ b/docs/source/robocerebra.mdx
@@ -1,99 +0,0 @@
-# RoboCerebra
-
-[RoboCerebra](https://robocerebra-project.github.io/) is a long-horizon manipulation benchmark that evaluates **high-level reasoning, planning, and memory** in VLAs. Episodes chain multiple sub-goals with language-grounded intermediate instructions, built on top of LIBERO's simulator stack (MuJoCo + robosuite, Franka Panda 7-DOF).
-
- Paper: [RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation](https://arxiv.org/abs/2506.06677)
- Project website: [robocerebra-project.github.io](https://robocerebra-project.github.io/)
- Dataset: [`lerobot/robocerebra_unified`](https://huggingface.co/datasets/lerobot/robocerebra_unified) — LeRobot v3.0, 6,660 episodes / 571,116 frames at 20 fps, 1,728 language-grounded sub-tasks.
- Pretrained policy: [`lerobot/smolvla_robocerebra`](https://huggingface.co/lerobot/smolvla_robocerebra)
-
-## Available tasks
-
-RoboCerebra reuses LIBERO's simulator, so evaluation runs against the LIBERO `libero_10` long-horizon suite:
-
-| Suite     | CLI name    | Tasks | Description                                                   |
-| --------- | ----------- | ----- | ------------------------------------------------------------- |
-| LIBERO-10 | `libero_10` | 10    | Long-horizon kitchen/living room tasks chaining 3–6 sub-goals |
-
-Each RoboCerebra episode in the dataset is segmented into multiple sub-tasks with natural-language instructions, which the unified dataset exposes as independent supervision signals.
-
-## Installation
-
-RoboCerebra piggybacks on LIBERO, so the `libero` extra is all you need:
-
-```bash
-pip install -e ".[libero]"
-```
-
-<Tip>
-RoboCerebra requires Linux (MuJoCo / robosuite). Set the rendering backend before training or evaluation:
-
-```bash
-export MUJOCO_GL=egl  # for headless servers (HPC, cloud)
-```
-
-</Tip>
-
-## Evaluation
-
-RoboCerebra eval runs against LIBERO's `libero_10` suite with RoboCerebra's camera naming (`image` + `wrist_image`) and an extra empty-camera slot so a three-view-trained policy receives the expected input layout:
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_robocerebra \
-  --env.type=libero \
-  --env.task=libero_10 \
-  --env.fps=20 \
-  --env.obs_type=pixels_agent_pos \
-  --env.observation_height=256 \
-  --env.observation_width=256 \
-  '--env.camera_name_mapping={"agentview_image": "image", "robot0_eye_in_hand_image": "wrist_image"}' \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.wrist_image": "observation.images.camera2"}' \
-  --policy.empty_cameras=1
-```
-
-### Recommended evaluation episodes
-
-**10 episodes per task** across the `libero_10` suite (100 total) for reproducible benchmarking. Matches the protocol used in the RoboCerebra paper.
-
-## Policy inputs and outputs
-
-**Observations:**
-
- `observation.state` — 8-dim proprioceptive state (7 joint positions + gripper)
- `observation.images.image` — third-person view, 256×256 HWC uint8
- `observation.images.wrist_image` — wrist-mounted camera view, 256×256 HWC uint8
-
-**Actions:**
-
- Continuous control in `Box(-1, 1, shape=(7,))` — end-effector delta (6D) + gripper (1D)
-
-## Training
-
-The unified dataset at [`lerobot/robocerebra_unified`](https://huggingface.co/datasets/lerobot/robocerebra_unified) exposes two RGB streams and language-grounded sub-task annotations:
-
-| Feature                          | Shape         | Description          |
-| -------------------------------- | ------------- | -------------------- |
-| `observation.images.image`       | (256, 256, 3) | Third-person view    |
-| `observation.images.wrist_image` | (256, 256, 3) | Wrist-mounted camera |
-| `observation.state`              | (8,)          | Joint pos + gripper  |
-| `action`                         | (7,)          | EEF delta + gripper  |
-
-Fine-tune a SmolVLA base on it:
-
-```bash
-lerobot-train \
-  --policy.path=lerobot/smolvla_base \
-  --dataset.repo_id=lerobot/robocerebra_unified \
-  --env.type=libero \
-  --env.task=libero_10 \
-  --output_dir=outputs/smolvla_robocerebra
-```
-
-## Reproducing published results
-
-The released checkpoint [`lerobot/smolvla_robocerebra`](https://huggingface.co/lerobot/smolvla_robocerebra) was trained on `lerobot/robocerebra_unified` and evaluated with the command in the [Evaluation](#evaluation) section. CI runs the same command with `--eval.n_episodes=1` as a smoke test on every PR touching the benchmark.
--- a/docs/source/robomme.mdx
+++ b/docs/source/robomme.mdx
@@ -1,130 +0,0 @@
-# RoboMME
-
-[RoboMME](https://robomme.github.io) is a memory-augmented manipulation benchmark built on ManiSkill (SAPIEN). It evaluates a robot's ability to retain and use information across an episode — counting, object permanence, reference, and imitation.
-
- **16 tasks** across 4 memory-skill suites
- **1,600 training demos** (100 per task, 50 val, 50 test)
- **Dataset**: [`lerobot/robomme`](https://huggingface.co/datasets/lerobot/robomme) — LeRobot v3.0, 768K frames at 10 fps
- **Simulator**: ManiSkill / SAPIEN, Panda arm, Linux only
-
-![RoboMME benchmark tasks overview](https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04639/gradient.png)
-
-## Tasks
-
-| Suite                             | Tasks                                                         |
-| --------------------------------- | ------------------------------------------------------------- |
-| **Counting** (temporal memory)    | BinFill, PickXtimes, SwingXtimes, StopCube                    |
-| **Permanence** (spatial memory)   | VideoUnmask, VideoUnmaskSwap, ButtonUnmask, ButtonUnmaskSwap  |
-| **Reference** (object memory)     | PickHighlight, VideoRepick, VideoPlaceButton, VideoPlaceOrder |
-| **Imitation** (procedural memory) | MoveCube, InsertPeg, PatternLock, RouteStick                  |
-
-## Installation
-
-> RoboMME requires **Linux** (ManiSkill/SAPIEN uses Vulkan rendering). Docker is recommended to isolate dependency conflicts.
-
-### Native (Linux)
-
-```bash
-pip install --override <(printf 'gymnasium==0.29.1\nnumpy==1.26.4\n') \
-  -e '.[smolvla,av-dep]' \
-  'robomme @ git+https://github.com/RoboMME/robomme_benchmark.git@main'
-```
-
-> **Dependency note**: `mani-skill` (pulled by `robomme`) pins `gymnasium==0.29.1` and `numpy<2.0.0`, which conflict with lerobot's base `numpy>=2.0.0`. That's why `robomme` is not a pyproject extra — use the override install above, or the Docker approach below to avoid conflicts entirely.
-
-### Docker (recommended)
-
-```bash
-# Build base image first (from repo root)
-docker build -f docker/Dockerfile.eval-base -t lerobot-eval-base .
-
-# Build RoboMME eval image (applies gymnasium + numpy pin overrides)
-docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-robomme .
-```
-
-The `docker/Dockerfile.benchmark.robomme` image overrides `gymnasium==0.29.1` and `numpy==1.26.4` after lerobot's install. Both versions are runtime-safe for lerobot's actual API usage.
-
-## Running Evaluation
-
-### Default (single task, single episode)
-
-```bash
-lerobot-eval \
-    --policy.path=<your_policy_repo> \
-    --env.type=robomme \
-    --env.task=PickXtimes \
-    --env.dataset_split=test \
-    --env.task_ids=[0] \
-    --eval.batch_size=1 \
-    --eval.n_episodes=1
-```
-
-### Multi-task evaluation
-
-Evaluate multiple tasks in one run by comma-separating task names. Use `task_ids` to control which episodes are evaluated per task. Recommended: 50 episodes per task for the test split.
-
-```bash
-lerobot-eval \
-    --policy.path=<your_policy_repo> \
-    --env.type=robomme \
-    --env.task=PickXtimes,BinFill,StopCube,MoveCube,InsertPeg \
-    --env.dataset_split=test \
-    --env.task_ids=[0,1,2,3,4,5,6,7,8,9] \
-    --eval.batch_size=1 \
-    --eval.n_episodes=50
-```
-
-### Key CLI options for `env.type=robomme`
-
-| Option               | Default       | Description                                        |
-| -------------------- | ------------- | -------------------------------------------------- |
-| `env.task`           | `PickXtimes`  | Any of the 16 task names above (comma-separated)   |
-| `env.dataset_split`  | `test`        | `train`, `val`, or `test`                          |
-| `env.action_space`   | `joint_angle` | `joint_angle` (8-D) or `ee_pose` (7-D)             |
-| `env.episode_length` | `300`         | Max steps per episode                              |
-| `env.task_ids`       | `null`        | List of episode indices to evaluate (null = `[0]`) |
-
-## Dataset
-
-The dataset [`lerobot/robomme`](https://huggingface.co/datasets/lerobot/robomme) is in **LeRobot v3.0 format** and can be loaded directly:
-
-```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-dataset = LeRobotDataset("lerobot/robomme")
-```
-
-### Dataset features
-
-| Feature            | Shape         | Description                     |
-| ------------------ | ------------- | ------------------------------- |
-| `image`            | (256, 256, 3) | Front camera RGB                |
-| `wrist_image`      | (256, 256, 3) | Wrist camera RGB                |
-| `actions`          | (8,)          | Joint angles + gripper          |
-| `state`            | (8,)          | Joint positions + gripper state |
-| `simple_subgoal`   | str           | High-level language annotation  |
-| `grounded_subgoal` | str           | Grounded language annotation    |
-| `episode_index`    | int           | Episode ID                      |
-| `frame_index`      | int           | Frame within episode            |
-
-### Feature key alignment (training)
-
-The env wrapper exposes `pixels/image` and `pixels/wrist_image` as observation keys. The `features_map` in `RoboMMEEnv` maps these to `observation.images.image` and `observation.images.wrist_image` for the policy. State is exposed as `agent_pos` and maps to `observation.state`.
-
-The dataset's `image` and `wrist_image` columns already align with the policy input keys, so no renaming is needed when fine-tuning.
-
-## Action Spaces
-
-| Type          | Dim | Description                                               |
-| ------------- | --- | --------------------------------------------------------- |
-| `joint_angle` | 8   | 7 joint angles + 1 gripper (−1 closed, +1 open, absolute) |
-| `ee_pose`     | 7   | xyz + roll/pitch/yaw + gripper                            |
-
-Set via `--env.action_space=joint_angle` (default) or `--env.action_space=ee_pose`.
-
-## Platform Notes
-
- **Linux only**: ManiSkill requires SAPIEN/Vulkan. macOS and Windows are not supported.
- **GPU recommended**: Rendering is CPU-capable but slow; CUDA + Vulkan gives full speed.
- **gymnasium / numpy conflict**: See installation note above. Docker image handles this automatically.
- **ManiSkill fork**: `robomme` depends on a specific ManiSkill fork (`YinpeiDai/ManiSkill`), pulled in automatically via the `robomme` package.
--- a/docs/source/robotwin.mdx
+++ b/docs/source/robotwin.mdx
@@ -1,223 +0,0 @@
-# RoboTwin 2.0
-
-RoboTwin 2.0 is a **large-scale dual-arm manipulation benchmark** built on the SAPIEN physics engine. It provides a standardized evaluation protocol for bimanual robotic policies across 50 tasks (as of upstream `main`) with strong domain randomization (clutter, lighting, background, tabletop height, and language instructions).
-
- Paper: [RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation](https://arxiv.org/abs/2506.18088)
- GitHub: [RoboTwin-Platform/RoboTwin](https://github.com/RoboTwin-Platform/RoboTwin)
- Leaderboard: [robotwin-platform.github.io/leaderboard](https://robotwin-platform.github.io/leaderboard)
- Dataset: [lerobot/robotwin_unified](https://huggingface.co/datasets/lerobot/robotwin_unified)
-
-![RoboTwin 2.0 benchmark overview](https://www.aitntnews.com/pictures/2025/7/8/9a7f79cb-5ba9-11f0-8581-fa163e47d677.png)
-
-## Overview
-
-| Property      | Value                                                    |
-| ------------- | -------------------------------------------------------- |
-| Tasks         | 50 dual-arm manipulation tasks                           |
-| Robot         | Aloha-AgileX bimanual (14 DOF, 7 per arm)                |
-| Action space  | 14-dim joint-space, continuous in `[-1, 1]`              |
-| Cameras       | `head_camera`, `left_camera`, `right_camera`             |
-| Simulator     | SAPIEN (not MuJoCo)                                      |
-| Eval protocol | 100 episodes/task, 50 demo_clean demonstrations          |
-| Eval settings | **Easy** (`demo_clean`) and **Hard** (`demo_randomized`) |
-
-## Available tasks
-
-RoboTwin 2.0 ships 50 dual-arm manipulation tasks in its upstream `envs/` directory. The canonical list is the `ROBOTWIN_TASKS` tuple in `src/lerobot/envs/robotwin.py`, mirrored verbatim from the upstream repo. Example tasks:
-
-| Task                     | CLI name                 | Category          |
-| ------------------------ | ------------------------ | ----------------- |
-| Beat block with hammer   | `beat_block_hammer`      | Tool use          |
-| Click bell / alarm clock | `click_bell`             | Precision press   |
-| Stack blocks (2 / 3)     | `stack_blocks_two/three` | Stacking          |
-| Stack bowls (2 / 3)      | `stack_bowls_two/three`  | Stacking          |
-| Handover block / mic     | `handover_block`         | Bimanual coord.   |
-| Lift pot                 | `lift_pot`               | Bimanual lift     |
-| Shake bottle             | `shake_bottle`           | Continuous motion |
-| Turn switch              | `turn_switch`            | Articulated obj   |
-| Stamp seal               | `stamp_seal`             | Precision place   |
-| Scan object              | `scan_object`            | Mobile manip.     |
-
-Pass a comma-separated list to `--env.task` to run multiple tasks in a single eval sweep.
-
-<Tip warning={true}>
-  `open_laptop` is currently broken upstream (its `check_success()` uses
-  `self.arm_tag`, which is only set inside the scripted-expert `play_once()`
-  path and therefore unavailable during normal policy eval). Avoid it until the
-  upstream bug is fixed, or patch the task to default `self.arm_tag = "left"` in
-  `load_actors()`.
-</Tip>
-
-## Dataset
-
-The RoboTwin 2.0 dataset is available in **LeRobot v3.0 format** on the Hugging Face Hub:
-
-```
-lerobot/robotwin_unified
-```
-
-It contains over 100,000 pre-collected trajectories across all 50 tasks (79.6 GB, Apache 2.0 license). No format conversion is needed — it is already in the correct LeRobot v3.0 schema with video observations and action labels.
-
-You can load it directly with the HF Datasets library:
-
-```python
-from datasets import load_dataset
-
-ds = load_dataset("lerobot/robotwin_unified", split="train")
-```
-
-## Installation
-
-RoboTwin 2.0 requires **Linux** with an NVIDIA GPU (CUDA 12.1 recommended). Installation takes approximately 20 minutes.
-
-### 1. Create a conda environment
-
-```bash
-conda create -n robotwin python=3.10 -y
-conda activate robotwin
-```
-
-### 2. Install LeRobot
-
-```bash
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-pip install -e "."
-```
-
-### 3. Install RoboTwin 2.0
-
-```bash
-git clone https://github.com/RoboTwin-Platform/RoboTwin.git
-cd RoboTwin
-bash script/_install.sh
-bash script/_download_assets.sh
-```
-
-The install script handles all Python dependencies including SAPIEN, CuRobo, mplib, and pytorch3d.
-
-<Tip warning={true}>
-If the automated install fails, install manually:
-
-```bash
-pip install -r requirements.txt
-pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"
-cd envs && git clone https://github.com/NVlabs/curobo.git && cd curobo
-pip install -e . --no-build-isolation
-```
-
-Then apply the required mplib fix: in `mplib/planner.py` line 807, remove `or collide` from the conditional.
-
-</Tip>
-
-### 4. Add RoboTwin to PYTHONPATH
-
-The RoboTwin task modules must be importable by LeRobot. From within the `RoboTwin/` directory:
-
-```bash
-export PYTHONPATH="${PYTHONPATH}:$(pwd)"
-```
-
-Add this to your shell profile to make it permanent.
-
-## Evaluation
-
-### Standard evaluation (recommended)
-
-Evaluate a policy on a single task with the official protocol (100 episodes):
-
-```bash
-lerobot-eval \
-  --policy.path="your-hf-policy-id" \
-  --env.type=robotwin \
-  --env.task=beat_block_hammer \
-  --eval.batch_size=1 \
-  --eval.n_episodes=100
-```
-
-### Single-task quick check
-
-```bash
-lerobot-eval \
-  --policy.path="your-hf-policy-id" \
-  --env.type=robotwin \
-  --env.task=beat_block_hammer \
-  --eval.batch_size=1 \
-  --eval.n_episodes=5
-```
-
-### Multi-task sweep
-
-Evaluate on several tasks in one run:
-
-```bash
-lerobot-eval \
-  --policy.path="your-hf-policy-id" \
-  --env.type=robotwin \
-  --env.task=beat_block_hammer,click_bell,handover_block,stack_blocks_two \
-  --eval.batch_size=1 \
-  --eval.n_episodes=100
-```
-
-### Full benchmark (all 50 tasks)
-
-```bash
-lerobot-eval \
-  --policy.path="your-hf-policy-id" \
-  --env.type=robotwin \
-  --env.task=adjust_bottle,beat_block_hammer,blocks_ranking_rgb,blocks_ranking_size,click_alarmclock,click_bell,dump_bin_bigbin,grab_roller,handover_block,handover_mic,hanging_mug,lift_pot,move_can_pot,move_pillbottle_pad,move_playingcard_away,move_stapler_pad,open_microwave,pick_diverse_bottles,pick_dual_bottles,place_a2b_left,place_a2b_right,place_bread_basket,place_bread_skillet,place_burger_fries,place_can_basket,place_cans_plasticbox,place_container_plate,place_dual_shoes,place_empty_cup,place_fan,place_mouse_pad,place_object_basket,place_object_scale,place_object_stand,place_phone_stand,place_shoe,press_stapler,put_bottles_dustbin,put_object_cabinet,rotate_qrcode,scan_object,shake_bottle,shake_bottle_horizontally,stack_blocks_three,stack_blocks_two,stack_bowls_three,stack_bowls_two,stamp_seal,turn_switch \
-  --eval.batch_size=1 \
-  --eval.n_episodes=100
-```
-
-<Tip>
-  `open_laptop` is intentionally omitted above because of the upstream
-  `self.arm_tag` bug (see the **Available tasks** section). Re-add it once the
-  upstream fix lands.
-</Tip>
-
-## Camera configuration
-
-By default, all three cameras are included:
-
-| Camera key     | Description                    |
-| -------------- | ------------------------------ |
-| `head_camera`  | Torso-mounted overhead view    |
-| `left_camera`  | Left arm wrist-mounted camera  |
-| `right_camera` | Right arm wrist-mounted camera |
-
-To use a subset of cameras, override `--env.camera_names`:
-
-```bash
-lerobot-eval \
-  --policy.path="your-hf-policy-id" \
-  --env.type=robotwin \
-  --env.task=beat_block_hammer \
-  --env.camera_names="head_camera,left_camera" \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10
-```
-
-## Environment config reference
-
-Key parameters for `RoboTwinEnvConfig`:
-
-| Parameter            | Default                                  | Description                        |
-| -------------------- | ---------------------------------------- | ---------------------------------- |
-| `task`               | `"beat_block_hammer"`                    | Comma-separated task name(s)       |
-| `fps`                | `25`                                     | Simulation FPS                     |
-| `episode_length`     | `300`                                    | Max steps per episode              |
-| `obs_type`           | `"pixels_agent_pos"`                     | `"pixels"` or `"pixels_agent_pos"` |
-| `camera_names`       | `"head_camera,left_camera,right_camera"` | Comma-separated active cameras     |
-| `observation_height` | `240`                                    | Camera pixel height                |
-| `observation_width`  | `320`                                    | Camera pixel width                 |
-
-## Leaderboard submission
-
-Results can be submitted to the [RoboTwin 2.0 leaderboard](https://robotwin-platform.github.io/leaderboard). The official protocol requires:
-
- Training on 50 `demo_clean` demonstrations per task
- Evaluating 100 episodes per task
- Reporting success rate separately for **Easy** (`demo_clean`) and **Hard** (`demo_randomized`) settings
-
-For submission instructions, refer to the [RoboTwin 2.0 documentation](https://robotwin-platform.github.io/doc/).
--- a/docs/source/sarm.mdx
+++ b/docs/source/sarm.mdx
@@ -46,7 +46,7 @@ This ensures identical task states map to consistent progress values, even acros

 ## Inputs and Targets (What the new code expects)

-SARM is trained through its processor (`src/lerobot/rewards/sarm/processor_sarm.py`), which:
+SARM is trained through its processor (`src/lerobot/policies/sarm/processor_sarm.py`), which:

 - **Encodes** images and task text with CLIP (ViT-B/32) into `video_features` and `text_features`
 - **Pads/truncates** robot state into `state_features` (up to `max_state_dim`)
@@ -347,7 +347,7 @@ Use `compute_rabc_weights.py` with `--visualize-only` to visualize model predict
 <hfoption id="single_stage">

 ```bash
-python -m lerobot.rewards.sarm.compute_rabc_weights \
+python src/lerobot/policies/sarm/compute_rabc_weights.py \
  --dataset-repo-id your-username/your-dataset \
  --reward-model-path your-username/sarm-model \
  --visualize-only \
@@ -360,7 +360,7 @@ python -m lerobot.rewards.sarm.compute_rabc_weights \
 <hfoption id="dense_only">

 ```bash
-python -m lerobot.rewards.sarm.compute_rabc_weights \
+python src/lerobot/policies/sarm/compute_rabc_weights.py \
  --dataset-repo-id your-username/your-dataset \
  --reward-model-path your-username/sarm-model \
  --visualize-only \
@@ -373,7 +373,7 @@ python -m lerobot.rewards.sarm.compute_rabc_weights \
 <hfoption id="dual">

 ```bash
-python -m lerobot.rewards.sarm.compute_rabc_weights \
+python src/lerobot/policies/sarm/compute_rabc_weights.py \
  --dataset-repo-id your-username/your-dataset \
  --reward-model-path your-username/sarm-model \
  --visualize-only \
@@ -429,7 +429,7 @@ The weighting follows **Equations 8-9** from the paper:
 First, run the SARM model on all frames in your dataset to compute progress values:

 ```bash
-python -m lerobot.rewards.sarm.compute_rabc_weights \
+python src/lerobot/policies/sarm/compute_rabc_weights.py \
  --dataset-repo-id your-username/your-dataset \
  --reward-model-path your-username/sarm-model \
  --head-mode sparse \
@@ -465,15 +465,15 @@ This script:

 ### Step 5b: Train Policy with RA-BC

-Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`) if not explicitly provided. Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
+Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`). Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:

 ```bash
 lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=pi0 \
-  --sample_weighting.type=rabc \
-  --sample_weighting.head_mode=sparse \
-  --sample_weighting.kappa=0.01 \
+  --use_rabc=true \
+  --rabc_head_mode=sparse \
+  --rabc_kappa=0.01 \
  --output_dir=outputs/train/policy_rabc \
  --batch_size=32 \
  --steps=40000
@@ -488,13 +488,12 @@ The training script automatically:

 **RA-BC Arguments:**

-| Argument                           | Description                                            | Default                 |
-| ---------------------------------- | ------------------------------------------------------ | ----------------------- |
-| `--sample_weighting.type`          | Weighting strategy type (`rabc` or `uniform`)          | `rabc`                  |
-| `--sample_weighting.progress_path` | Path to progress parquet file                          | `sarm_progress.parquet` |
-| `--sample_weighting.head_mode`     | Which SARM head's progress to use: `sparse` or `dense` | `sparse`                |
-| `--sample_weighting.kappa`         | Threshold κ for high-quality samples                   | `0.01`                  |
-| `--sample_weighting.epsilon`       | Small constant for numerical stability                 | `1e-6`                  |
+| Argument               | Description                                                | Default                            |
+| ---------------------- | ---------------------------------------------------------- | ---------------------------------- |
+| `--use_rabc`           | Enable RA-BC sample weighting                              | `false`                            |
+| `--rabc_progress_path` | Path to progress parquet file (auto-detected from dataset) | `sarm_progress.parquet` in dataset |
+| `--rabc_head_mode`     | Which SARM head's progress to use: `sparse` or `dense`     | `sparse`                           |
+| `--rabc_kappa`         | Threshold κ for high-quality samples                       | `0.01`                             |

 ### Tuning RA-BC Kappa

@@ -512,30 +511,30 @@ The `kappa` parameter is the threshold that determines which samples get full we

 Monitor these WandB metrics during training:

-| Metric                        | Healthy Range | Problem Indicator         |
-| ----------------------------- | ------------- | ------------------------- |
-| `sample_weight_mean_weight`   | 0.3 - 0.8     | ≈ 1.0 means kappa too low |
-| `sample_weighting/delta_mean` | > 0           | Should be positive        |
-| `sample_weighting/delta_std`  | > 0           | Variance in data quality  |
+| Metric             | Healthy Range | Problem Indicator         |
+| ------------------ | ------------- | ------------------------- |
+| `rabc_mean_weight` | 0.3 - 0.8     | ≈ 1.0 means kappa too low |
+| `rabc_delta_mean`  | > 0           | Should be positive        |
+| `rabc_delta_std`   | > 0           | Variance in data quality  |

-**If `sample_weight_mean_weight ≈ 1.0`:** Your kappa is too low. Most samples have `delta > kappa` and bypass the soft-weighting entirely. RA-BC becomes equivalent to vanilla BC.
+**If `rabc_mean_weight ≈ 1.0`:** Your kappa is too low. Most samples have `delta > kappa` and bypass the soft-weighting entirely. RA-BC becomes equivalent to vanilla BC.

 **Setting kappa based on your data:**

-The default `kappa=0.01` was tuned for the paper's T-shirt folding task (~90s episodes at 30fps). For your dataset, check the logged `sample_weighting/delta_mean` and `sample_weighting/delta_std`:
+The default `kappa=0.01` was tuned for the paper's T-shirt folding task (~90s episodes at 30fps). For your dataset, check the logged `rabc_delta_mean` and `rabc_delta_std`:

 ```
 # If delta_mean ≈ 0.03 and delta_std ≈ 0.02:
 # Most deltas fall in range [0.01, 0.05]

 # Option 1: Set kappa = delta_mean (medium selectivity)
--sample_weighting.kappa=0.03
+--rabc_kappa=0.03

 # Option 2: Set kappa = delta_mean + delta_std (high selectivity)
--sample_weighting.kappa=0.05
+--rabc_kappa=0.05

 # Option 3: Set kappa = delta_mean + 2*delta_std (very selective)
--sample_weighting.kappa=0.07
+--rabc_kappa=0.07
 ```

 **When RA-BC may not help:**
@@ -551,8 +550,8 @@ accelerate launch \
  src/lerobot/scripts/lerobot_train.py \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=pi0 \
-  --sample_weighting.type=rabc \
-  --sample_weighting.kappa=0.01 \
+  --use_rabc=true \
+  --rabc_kappa=0.01 \
  --output_dir=outputs/train/policy_rabc \
  --batch_size=32 \
  --steps=40000
@@ -577,7 +576,7 @@ accelerate launch \
 ### RA-BC

 1. **Train SARM first**: RA-BC quality depends entirely on SARM quality
-2. **Monitor `sample_weight_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))
+2. **Monitor `rabc_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))

 ---

--- a/docs/source/smolvla.mdx
+++ b/docs/source/smolvla.mdx
@@ -97,22 +97,22 @@ Similarly for when recording an episode, it is recommended that you are logged i
 Once you are logged in, you can run inference in your setup by doing:

 ```bash
-lerobot-rollout \
-  --strategy.type=base \
+lerobot-record \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM0 \ # <- Use your port
  --robot.id=my_blue_follower_arm \ # <- Use your robot id
  --robot.cameras="{ front: {type: opencv, index_or_path: 8, width: 640, height: 480, fps: 30}}" \ # <- Use your cameras
-  --task="Grasp a lego block and put it in the bin." \ # <- Use the same task description you used in your dataset recording
-  # <- RTC optional, use when running on low power hardware \
-  # --inference.type=rtc \
-  # --inference.rtc.execution_horizon=10 \
-  # --inference.rtc.max_guidance_weight=10.0 \
+  --dataset.single_task="Grasp a lego block and put it in the bin." \ # <- Use the same task description you used in your dataset recording
+  --dataset.repo_id=${HF_USER}/eval_DATASET_NAME_test \  # <- This will be the dataset name on HF Hub
+  --dataset.episode_time_s=50 \
+  --dataset.num_episodes=10 \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
  # <- Teleop optional if you want to teleoperate in between episodes \
  # --teleop.type=so100_leader \
  # --teleop.port=/dev/ttyACM0 \
  # --teleop.id=my_red_leader_arm \
-  # --display_data=true #optional use if you want to see the camera stream \
  --policy.path=HF_USER/FINETUNE_MODEL_NAME # <- Use your fine-tuned model
 ```

--- a/docs/source/streaming_video_encoding.mdx
+++ b/docs/source/streaming_video_encoding.mdx
@@ -17,9 +17,9 @@ This makes `save_episode()` near-instant (the video is already encoded by the ti
 | Parameter               | CLI Flag                          | Type          | Default       | Description                                                       |
 | ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
 | `streaming_encoding`    | `--dataset.streaming_encoding`    | `bool`        | `True`        | Enable real-time encoding during capture                          |
-| `vcodec`                | `--dataset.camera_encoder.vcodec` | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
+| `vcodec`                | `--dataset.vcodec`                | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
 | `encoder_threads`       | `--dataset.encoder_threads`       | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
-| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `30`          | Max buffered frames per camera (~1s at 30fps). Consumes RAM       |
+| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `60`          | Max buffered frames per camera (~2s at 30fps). Consumes RAM       |

 ## 3. Performance Considerations

@@ -48,7 +48,7 @@ This parameter controls how many threads each encoder instance uses internally:

 ### Backpressure and Frame Dropping

-Each camera has a bounded queue (`encoder_queue_maxsize`, default 30 frames). When the encoder can't keep up:
+Each camera has a bounded queue (`encoder_queue_maxsize`, default 60 frames). When the encoder can't keep up:

 1. The queue fills up (consuming RAM)
 2. New frames are **dropped** (not blocked) — the capture loop continues uninterrupted
@@ -82,15 +82,15 @@ Use HW encoding when:

 ### Available HW Encoders

-| Encoder             | Platform      | Hardware                                                                                         | CLI Value                                           |
-| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | --------------------------------------------------- |
-| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.camera_encoder.vcodec=h264_videotoolbox` |
-| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.camera_encoder.vcodec=hevc_videotoolbox` |
-| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.camera_encoder.vcodec=h264_nvenc`        |
-| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.camera_encoder.vcodec=hevc_nvenc`        |
-| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.camera_encoder.vcodec=h264_vaapi`        |
-| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.camera_encoder.vcodec=h264_qsv`          |
-| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.camera_encoder.vcodec=auto`              |
+| Encoder             | Platform      | Hardware                                                                                         | CLI Value                            |
+| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------ |
+| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=h264_videotoolbox` |
+| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=hevc_videotoolbox` |
+| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=h264_nvenc`        |
+| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=hevc_nvenc`        |
+| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.vcodec=h264_vaapi`        |
+| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.vcodec=h264_qsv`          |
+| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.vcodec=auto`              |

 > [!NOTE]
 > In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
@@ -100,15 +100,15 @@ Use HW encoding when:

 ## 5. Troubleshooting

-| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                                 |
-| ------------------------------------------------------------------ | -------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.camera_encoder.vcodec=auto`) |
-| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.camera_encoder.vcodec=auto`).                                                                                                                                                    |
-| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                                    |
-| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                                             |
-| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                                             |
-| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.camera_encoder.vcodec=auto`                                                                                                                                                                                                                             |
-| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                                  |
+| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                  |
+| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.vcodec=auto`) |
+| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.vcodec=auto`).                                                                                                                                                    |
+| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                     |
+| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                              |
+| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                              |
+| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.vcodec=auto`                                                                                                                                                                                                                             |
+| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                   |

 ## 6. Recommended Configurations

@@ -146,7 +146,7 @@ On very constrained systems, streaming encoding may compete too heavily with the
 # 2camsx 640x480x3 @30fps: Requires some tuning.

 # Use H.264, disable streaming, consider batching encoding
-lerobot-record --dataset.camera_encoder.vcodec=h264 --dataset.streaming_encoding=false ...
+lerobot-record --dataset.vcodec=h264 --dataset.streaming_encoding=false ...
 ```

 ## 7. Closing note
--- a/docs/source/tools.mdx
+++ b/docs/source/tools.mdx
@@ -1,210 +0,0 @@
-# Tools
-
-LeRobot v3.1 supports **tool calls** in policies — assistant messages can
-emit structured invocations like `say(text="OK, starting now")` that the
-runtime dispatches to a real implementation (TTS, controller, logger, …).
-
-This page covers:
-
-1. Where the tool catalog lives.
-2. How the annotation pipeline produces tool-call atoms.
-3. How to add your own tool.
-
-## Where tools are declared
-
-Two layers.
-
-**The catalog** — a list of OpenAI-style function schemas — lives at
-`meta/info.json["tools"]` on each dataset. Example:
-
-```json
-{
-  "features": { "...": "..." },
-  "tools": [
-    {
-      "type": "function",
-      "function": {
-        "name": "say",
-        "description": "Speak a short utterance to the user via the TTS executor.",
-        "parameters": {
-          "type": "object",
-          "properties": {
-            "text": {
-              "type": "string",
-              "description": "The verbatim text to speak."
-            }
-          },
-          "required": ["text"]
-        }
-      }
-    }
-  ]
-}
-```
-
-Read it via the dataset metadata accessor:
-
-```python
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-
-meta = LeRobotDatasetMetadata(repo_id="pepijn/super_poulain_final_annotations")
-tools = meta.tools     # list[dict] — OpenAI tool schemas
-```
-
-If the dataset's `info.json` doesn't declare any tools, `meta.tools`
-returns `DEFAULT_TOOLS` from `lerobot.datasets.language` — currently a
-single-entry list with the canonical `say` schema. So unannotated
-datasets and chat-template consumers keep working without any
-configuration:
-
-```python
-prompt_str = tokenizer.apply_chat_template(
-    sample["messages"],
-    tools=meta.tools,                 # works either way
-    add_generation_prompt=False,
-    tokenize=False,
-)
-```
-
-**The implementations** — runnable Python — will live under
-`src/lerobot/tools/`, one file per tool. The runtime dispatcher and
-the canonical `say` implementation (wrapping Kyutai's pocket-tts) are
-not part of the catalog layer described here; today this layer ships
-only the schema storage and the `DEFAULT_TOOLS` fallback constant.
-
-## Per-row tool _invocations_
-
-The catalog above describes _what can be called_. The actual _call_ — the
-function name plus the argument values — is stored per-row, on the
-assistant atoms in `language_events`:
-
-```python
-{
-  "role": "assistant",
-  "content": null,
-  "style": null,
-  "timestamp": 12.4,
-  "camera": null,
-  "tool_calls": [
-    { "type": "function",
-      "function": { "name": "say", "arguments": { "text": "On it." } } }
-  ]
-}
-```
-
-Recipes splice these into rendered messages via `tool_calls_from`:
-
-```yaml
-user_interjection_response:
-  bindings:
-    speech: "emitted_at(t, role=assistant, tool_name=say)"
-  messages:
-    - { role: user, content: "${task}", stream: high_level }
-    - {
-        role: assistant,
-        content: "${current_plan}",
-        stream: high_level,
-        target: true,
-        tool_calls_from: speech,
-      }
-```
-
-The model's training target is one assistant turn that carries both the
-plan text _and_ the `say` tool call. At inference, the runtime parses
-the generated text back into structured `tool_calls` and dispatches to
-the matching implementation.
-
-## How to add your own tool
-
-> **Note:** Steps 2 and 3 below describe the runtime layer
-> (`src/lerobot/tools/`, the `Tool` protocol, `TOOL_REGISTRY`,
-> `get_tools(meta)`) which is not part of the catalog layer shipped
-> today — those modules don't yet exist in the tree. Step 1 alone is
-> enough to make the tool visible to the chat template via
-> `meta.tools` so the model can learn to _generate_ the call;
-> executing the call at inference requires the runtime layer.
-
-Three steps. Concrete example: a `record_observation` tool the policy
-can call to capture an extra observation outside the regular control
-loop.
-
-### Step 1 — declare the schema
-
-Add an entry under `meta/info.json["tools"]`. Either edit the file
-directly on disk _before_ running the annotation pipeline (it'll be
-preserved) or hand it to `lerobot-annotate` via a config flag.
-
-```json
-{
-  "tools": [
-    { "type": "function", "function": { "name": "say", "...": "..." } },
-    {
-      "type": "function",
-      "function": {
-        "name": "record_observation",
-        "description": "Capture a high-resolution still image for the user.",
-        "parameters": {
-          "type": "object",
-          "properties": {
-            "label": {
-              "type": "string",
-              "description": "Short label for the saved image."
-            }
-          },
-          "required": ["label"]
-        }
-      }
-    }
-  ]
-}
-```
-
-The schema follows OpenAI's function-calling convention exactly, so the
-chat template can render it natively.
-
-### Step 2 — implement the call
-
-Create `src/lerobot/tools/record_observation.py`:
-
-```python
-from .base import Tool
-from typing import Any
-
-RECORD_OBSERVATION_SCHEMA: dict[str, Any] = { "...": "..." }   # mirrors the JSON above
-
-
-class RecordObservationTool:
-    name = "record_observation"
-    schema = RECORD_OBSERVATION_SCHEMA
-
-    def __init__(self, schema: dict | None = None, output_dir: str = "."):
-        self.output_dir = output_dir
-
-    def call(self, arguments: dict) -> str:
-        label = arguments["label"]
-        # ... save the latest camera frame to <output_dir>/<label>.png ...
-        return f"saved {label}.png"
-```
-
-One file per tool keeps dependencies isolated — `record_observation`
-might pull `pillow`, while `say` pulls `pocket-tts`. Users installing
-only the tools they need avoid heavy transitive deps.
-
-### Step 3 — register it
-
-Add to `src/lerobot/tools/registry.py`:
-
-```python
-from .record_observation import RecordObservationTool
-
-TOOL_REGISTRY["record_observation"] = RecordObservationTool
-```
-
-That's it. At runtime `get_tools(meta)` looks up each schema in
-`meta.tools`, instantiates the matching registered class, and returns
-a name → instance dict the dispatcher can route into.
-
-If you want to use a tool _without_ writing an implementation (e.g. for
-training-time chat-template formatting only), step 1 alone is enough —
-the model still learns to _generate_ the call. Steps 2 and 3 are only
-needed to actually _execute_ it at inference.
--- a/docs/source/unitree_g1.mdx
+++ b/docs/source/unitree_g1.mdx
@@ -274,8 +274,7 @@ python src/lerobot/scripts/lerobot_train.py \
 Once trained, we recommend deploying policies using inference-time RTC:

 ```bash
-lerobot-rollout \
-  --strategy.type=base \
+python examples/rtc/eval_with_real_robot.py \
  --policy.path=your-username/your-repo-id \
  --policy.device=cuda \
  --robot.type=unitree_g1 \
--- a/docs/source/using_dataset_tools.mdx
+++ b/docs/source/using_dataset_tools.mdx
@@ -117,10 +117,10 @@ lerobot-edit-dataset \
    --repo_id lerobot/pusht_image \
    --operation.type convert_image_to_video \
    --operation.output_dir outputs/pusht_video \
-    --operation.camera_encoder.vcodec libsvtav1 \
-    --operation.camera_encoder.pix_fmt yuv420p \
-    --operation.camera_encoder.g 2 \
-    --operation.camera_encoder.crf 30
+    --operation.vcodec libsvtav1 \
+    --operation.pix_fmt yuv420p \
+    --operation.g 2 \
+    --operation.crf 30

 # Convert only specific episodes
 lerobot-edit-dataset \
@@ -147,7 +147,11 @@ lerobot-edit-dataset \
 **Parameters:**

 - `output_dir`: Custom output directory (optional - by default uses `new_repo_id` or `{repo_id}_video`)
- `camera_encoder`: Video encoder settings — all sub-fields accessible via `--operation.camera_encoder.<field>. See [Video Encoding Parameters](./video_encoding_parameters) for more details.
+- `vcodec`: Video codec to use - options: `h264`, `hevc`, `libsvtav1` (default: `libsvtav1`)
+- `pix_fmt`: Pixel format - options: `yuv420p`, `yuv444p` (default: `yuv420p`)
+- `g`: Group of pictures (GOP) size - lower values give better quality but larger files (default: 2)
+- `crf`: Constant rate factor - lower values give better quality but larger files, 0 is lossless (default: 30)
+- `fast_decode`: Fast decode tuning option (default: 0)
 - `episode_indices`: List of specific episodes to convert (default: all episodes)
 - `num_workers`: Number of parallel workers for processing (default: 4)

--- a/docs/source/video_encoding_parameters.mdx
+++ b/docs/source/video_encoding_parameters.mdx
@@ -1,117 +0,0 @@
-# Video encoding parameters
-
-When video storage is enabled, LeRobot stores each camera stream as an **MP4** file instead of saving one image file per timestep. Video encoding compresses across time, which usually cuts dataset size and I/O compared to a pile of PNG, while keeping MP4 — a format every player and loader understands.
-
-Encoding frames into an MP4 is a full FFmpeg pipeline: choice of encoder, pixel format, GOP/keyframes, quality vs. speed, and optional extra encoder flags. Most of these knobs are user-tunable through `camera_encoder`, a nested `VideoEncoderConfig` (`lerobot.configs.video.VideoEncoderConfig`) passed through PyAV.
-
-You can set these parameters from the CLI with `--dataset.camera_encoder.<field>` (e.g. with `lerobot-record` or `lerobot-rollout`). The same block applies to every camera video stream in that run.
-
-<Tip>
-  Video storage must be on for `camera_encoder` to have any effect —
-  `use_videos=True` in Python APIs, or `--dataset.video=true` on the CLI (the
-  recording default). With video off, inputs stay as images and `camera_encoder`
-  is ignored.
-</Tip>
-
-For details on **when** frames are written vs. encoded (streaming vs. post-episode), queues, and other top-level `--dataset.*` switches, see [Streaming Video Encoding](./streaming_video_encoding). For an encoding-parameter comparison and experiments, see the [video-benchmark Space](https://huggingface.co/spaces/lerobot/video-benchmark).
-
---
-
-## Example
-
-```bash
-lerobot-record \
-    --robot.type=so100_follower \
-    --robot.port=/dev/tty.usbmodem58760431541 \
-    --robot.cameras="{laptop: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-    --robot.id=black \
-    --teleop.type=so100_leader \
-    --teleop.port=/dev/tty.usbmodem58760431551 \
-    --teleop.id=blue \
-    --dataset.repo_id=<my_username>/<my_dataset_name> \
-    --dataset.num_episodes=2 \
-    --dataset.single_task="Grab the cube" \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    --dataset.camera_encoder.vcodec=h264 \
-    --dataset.camera_encoder.preset=fast \
-    --dataset.camera_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \
-    --display_data=true
-```
-
---
-
-## Tuning parameters
-
-<Tip warning={true}>
-The defaults are tuned to balance **compression ratio**, **visual quality**, and **decoding/seek speed** for typical robotics datasets. Changing them can affect both recording (CPU load, frame drops) and training (decoding throughput, image quality).
-
-Only override these parameters if you have a specific reason to, and measure the impact on your pipeline before relying on the new settings.
-
-</Tip>
-
-All flags below are prefixed with `--dataset.camera_encoder.` on the CLI.
-
-| Parameter       | Type             | Default       | Description                                                                                                                                                                            |
-| --------------- | ---------------- | ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `vcodec`        | `str`            | `"libsvtav1"` | Video codec name. `"auto"` picks the first available hardware encoder from a fixed preference list, falling back to `libsvtav1`.                                                       |
-| `pix_fmt`       | `str`            | `"yuv420p"`   | Output pixel format. Must be supported by the chosen codec in your FFmpeg build.                                                                                                       |
-| `g`             | `int`            | `2`           | GOP size — a keyframe every `g` frames. Emitted as FFmpeg option `g`.                                                                                                                  |
-| `crf`           | `int` or `float` | `30`          | Abstract quality value, mapped per codec (see the [mapping](#mapping-videoencoderconfig--ffmpeg-options) below). Lower → higher quality / larger output where the mapping is monotone. |
-| `preset`        | `int` or `str`   | `12` \*       | Encoder speed preset; meaning depends on the codec. <br/>\* When unset and `vcodec=libsvtav1`, LeRobot defaults to `12`.                                                               |
-| `fast_decode`   | `int`            | `0`           | `libsvtav1`: `0–2`, passed via `svtav1-params`. <br/>`h264` / `hevc` (software): if `>0`, sets `tune=fastdecode`. <br/>Other codecs: usually unused.                                   |
-| `video_backend` | `str`            | `"pyav"`      | Only `"pyav"` is currently implemented for video encoding.                                                                                                                             |
-| `extra_options` | `dict`           | `{}`          | Extra FFmpeg or codec specific options merged after the structured fields above. Cannot override keys already set by those fields.                                                     |
-
---
-
-## Persistence in dataset metadata
-
-After the first episode of a video stream is encoded, the encoder configuration is **persisted into the dataset metadata** (`meta/info.json`) under each video feature, alongside the values probed from the file itself. For a video feature `observation.images.<camera>`, the layout in `info.json` is:
-
-```json
-{
-  "features": {
-    "observation.images.laptop": {
-      "dtype": "video",
-      "shape": [480, 640, 3],
-      "info": {
-        "video.height": 480,
-        "video.width": 640,
-        "video.codec": "h264",
-        "video.pix_fmt": "yuv420p",
-        "video.fps": 30,
-        "video.channels": 3,
-        "video.is_depth_map": false,
-        "video.g": 2,
-        "video.crf": 30,
-        "video.preset": "fast",
-        "video.fast_decode": 0,
-        "video.video_backend": "pyav",
-        "video.extra_options": { "tune": "film", "profile:v": "high", "bf": 2 }
-      }
-    }
-  }
-}
-```
-
-Two sources contribute to the `info` block:
-
- **Stream-derived** (read back from the encoded MP4 with PyAV): `video.height`, `video.width`, `video.codec`, `video.pix_fmt`, `video.fps`, `video.channels`, `video.is_depth_map`, plus `audio.*` if an audio stream is present.
- **Encoder-derived** (taken from `VideoEncoderConfig`): `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.video_backend`, `video.extra_options`.
-
-<Tip>
-  This block is populated **once**, from the **first** episode. It assumes every
-  episode in the dataset was encoded with the same `camera_encoder`. Changing
-  encoder settings partway through a recording is not supported — the
-  `info.json` will only reflect the parameters used for the first episode.
-</Tip>
-
---
-
-## Merging datasets
-
-When aggregating datasets with `merge_datasets`, video files are concatenated as-is (no re-encoding), and encoder fields in `info.json` are merged per-key:
-
- **Stream-derived fields must match** across sources: `video.codec`, `video.pix_fmt`, `video.height`, `video.width`, `video.fps`. Otherwise FFmpeg's concat demuxer fails.
- **Encoder-tuning fields are merged loosely**: `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.extra_options`. If every source agrees, the value is kept; if not, it's set to `null` (or `{}` for `video.extra_options`) and a warning is logged.
--- a/docs/source/vlabench.mdx
+++ b/docs/source/vlabench.mdx
@@ -1,176 +0,0 @@
-# VLABench
-
-[VLABench](https://github.com/OpenMOSS/VLABench) is a large-scale benchmark for **language-conditioned robotic manipulation with long-horizon reasoning**. The upstream suite covers 100 task categories across 2,000+ objects and evaluates six dimensions of robot intelligence: mesh & texture understanding, spatial reasoning, world-knowledge transfer, semantic instruction comprehension, physical-law understanding, and long-horizon planning. Built on MuJoCo / dm_control with a Franka Panda 7-DOF arm. LeRobot exposes **43 of these tasks** through `--env.task` (21 primitives + 22 composites, see [Available tasks](#available-tasks) below).
-
- Paper: [VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning](https://arxiv.org/abs/2412.18194)
- GitHub: [OpenMOSS/VLABench](https://github.com/OpenMOSS/VLABench)
- Project website: [vlabench.github.io](https://vlabench.github.io)
- Pretrained policy: [`lerobot/smolvla_vlabench`](https://huggingface.co/lerobot/smolvla_vlabench)
-
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/vlabench.png"
-  alt="VLABench benchmark overview"
-  width="85%"
-/>
-
-## Available tasks
-
-VLABench ships two task suites covering **43 task categories** in LeRobot's `--env.task` surface:
-
-| Suite     | CLI name    | Tasks | Description                                                      |
-| --------- | ----------- | ----- | ---------------------------------------------------------------- |
-| Primitive | `primitive` | 21    | Single / few-skill combinations (select, insert, physics QA)     |
-| Composite | `composite` | 22    | Multi-step reasoning and long-horizon planning (cook, rearrange) |
-
-**Primitive tasks:** `select_fruit`, `select_toy`, `select_chemistry_tube`, `add_condiment`, `select_book`, `select_painting`, `select_drink`, `insert_flower`, `select_billiards`, `select_ingredient`, `select_mahjong`, `select_poker`, and physical-reasoning tasks (`density_qa`, `friction_qa`, `magnetism_qa`, `reflection_qa`, `simple_cuestick_usage`, `simple_seesaw_usage`, `sound_speed_qa`, `thermal_expansion_qa`, `weight_qa`).
-
-**Composite tasks:** `cluster_billiards`, `cluster_book`, `cluster_drink`, `cluster_toy`, `cook_dishes`, `cool_drink`, `find_unseen_object`, `get_coffee`, `hammer_nail`, `heat_food`, `make_juice`, `play_mahjong`, `play_math_game`, `play_poker`, `play_snooker`, `rearrange_book`, `rearrange_chemistry_tube`, `set_dining_table`, `set_study_table`, `store_food`, `take_chemistry_experiment`, `use_seesaw_complex`.
-
-`--env.task` accepts three forms:
-
- a single task name (`select_fruit`)
- a comma-separated list (`select_fruit,heat_food`)
- a suite shortcut (`primitive`, `composite`, or `primitive,composite`)
-
-## Installation
-
-VLABench is **not on PyPI** — its only distribution is the [OpenMOSS/VLABench](https://github.com/OpenMOSS/VLABench) GitHub repo — so LeRobot does not expose a `vlabench` extra. Install it manually as an editable clone, alongside the MuJoCo / dm_control pins VLABench needs, then fetch the mesh assets:
-
-```bash
-# After following the standard LeRobot installation instructions.
-
-git clone https://github.com/OpenMOSS/VLABench.git ~/VLABench
-git clone https://github.com/motion-planning/rrt-algorithms.git ~/rrt-algorithms
-pip install -e ~/VLABench -e ~/rrt-algorithms
-pip install "mujoco==3.2.2" "dm-control==1.0.22" \
-            open3d colorlog scikit-learn openai gdown
-
-python ~/VLABench/scripts/download_assets.py
-```
-
-<Tip>
-VLABench requires Linux (`sys_platform == 'linux'`) and Python 3.10+. Set the MuJoCo rendering backend before running:
-
-```bash
-export MUJOCO_GL=egl  # for headless servers (HPC, cloud)
-```
-
-</Tip>
-
-## Evaluation
-
-All eval snippets below mirror the command CI runs (see `.github/workflows/benchmark_tests.yml`). The `--rename_map` argument maps VLABench's `image` / `second_image` / `wrist_image` camera keys onto the three-camera (`camera1` / `camera2` / `camera3`) input layout the released `smolvla_vlabench` policy was trained on.
-
-### Single-task evaluation (recommended for quick iteration)
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_vlabench \
-  --env.type=vlabench \
-  --env.task=select_fruit \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
-```
-
-### Multi-task evaluation
-
-Pass a comma-separated list of tasks:
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_vlabench \
-  --env.type=vlabench \
-  --env.task=select_fruit,select_toy,add_condiment,heat_food \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
-```
-
-### Suite-wide evaluation
-
-Run an entire suite (all 21 primitives or all 22 composites):
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_vlabench \
-  --env.type=vlabench \
-  --env.task=primitive \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  --env.max_parallel_tasks=1 \
-  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
-```
-
-Or both suites:
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_vlabench \
-  --env.type=vlabench \
-  --env.task=primitive,composite \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  --env.max_parallel_tasks=1 \
-  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
-```
-
-### Recommended evaluation episodes
-
-**10 episodes per task** for reproducible benchmarking (210 total for the full primitive suite, 220 for composite). Matches the protocol in the VLABench paper.
-
-## Policy inputs and outputs
-
-**Observations:**
-
- `observation.state` — 7-dim end-effector state (position xyz + Euler xyz + gripper)
- `observation.images.image` — front camera, 480×480 HWC uint8
- `observation.images.second_image` — second camera, 480×480 HWC uint8
- `observation.images.wrist_image` — wrist camera, 480×480 HWC uint8
-
-**Actions:**
-
- Continuous control in `Box(-1, 1, shape=(7,))` — 3D position + 3D Euler orientation + 1D gripper.
-
-## Training
-
-### Datasets
-
-Pre-collected VLABench datasets in LeRobot format on the Hub:
-
- [`VLABench/vlabench_primitive_ft_lerobot_video`](https://huggingface.co/datasets/VLABench/vlabench_primitive_ft_lerobot_video) — 5,000 episodes, 128 tasks, 480×480 images.
- [`VLABench/vlabench_composite_ft_lerobot_video`](https://huggingface.co/datasets/VLABench/vlabench_composite_ft_lerobot_video) — 5,977 episodes, 167 tasks, 224×224 images.
-
-### Example training command
-
-Fine-tune a SmolVLA base on the primitive suite:
-
-```bash
-lerobot-train \
-  --policy.type=smolvla \
-  --policy.repo_id=${HF_USER}/smolvla_vlabench_primitive \
-  --policy.load_vlm_weights=true \
-  --policy.push_to_hub=true \
-  --dataset.repo_id=VLABench/vlabench_primitive_ft_lerobot_video \
-  --env.type=vlabench \
-  --env.task=select_fruit \
-  --output_dir=./outputs/smolvla_vlabench_primitive \
-  --steps=100000 \
-  --batch_size=4 \
-  --eval_freq=5000 \
-  --eval.batch_size=1 \
-  --eval.n_episodes=1 \
-  --save_freq=10000
-```
-
-## Reproducing published results
-
-The released checkpoint [`lerobot/smolvla_vlabench`](https://huggingface.co/lerobot/smolvla_vlabench) was trained on the primitive-suite dataset above and is evaluated with the [Single-task](#single-task-evaluation-recommended-for-quick-iteration) / [Suite-wide](#suite-wide-evaluation) commands. CI runs a 10-primitive-task smoke eval (one episode each) on every PR touching the benchmark.
--- a/docs/source/xvla.mdx
+++ b/docs/source/xvla.mdx
@@ -220,7 +220,7 @@ REAL_DIM = 12
 # Postprocessing: Trim 20D predictions to 12D for deployment
 ```

-See the [action_hub.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/action_hub.py) implementation for details.
+See the [action_hub.py](/home/jade_choghari/robot/lerobot/src/lerobot/policies/xvla/action_hub.py) implementation for details.

 #### Auto Action Mode (Recommended)

@@ -519,9 +519,9 @@ If you use X-VLA in your research, please cite:

 - [X-VLA Paper](https://arxiv.org/pdf/2510.10274)
 - [LeRobot Documentation](https://github.com/huggingface/lerobot)
- [Action Registry Implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/action_hub.py)
- [Processor Implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/processor_xvla.py)
- [Model Configuration](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/configuration_xvla.py)
+- [Action Registry Implementation](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/action_hub.py)
+- [Processor Implementation](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/processor_xvla.py)
+- [Model Configuration](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/configuration_xvla.py)

 ## Contributing

--- a/examples/dataset/create_progress_videos.py
+++ b/examples/dataset/create_progress_videos.py
@@ -15,12 +15,10 @@
 # limitations under the License.

 """
-Create MP4 (or GIF) videos with per-frame progress overlay for specified episodes.
+Create MP4 (or GIF) videos with sarm_progress overlay for specified episodes.

 Downloads datasets from HuggingFace, seeks directly into the episode segment
 of the source video, draws a progress line on each frame, and writes the result.
-The progress data is read from a parquet file that lives alongside the dataset
-(configurable via ``--progress-file``).

 Usage:
    python examples/dataset/create_progress_videos.py \
@@ -58,26 +56,22 @@ SCORE_FONT_SCALE = 0.8
 TASK_FONT_SCALE = 0.55


-def download_episode_metadata(
-    repo_id: str, episode: int, progress_file: str = "sarm_progress.parquet"
-) -> Path:
-    """Download only the metadata and per-frame progress file for a dataset.
+def download_episode_metadata(repo_id: str, episode: int) -> Path:
+    """Download only the metadata and sarm_progress files for a dataset.

    Args:
        repo_id: HuggingFace dataset repository ID.
        episode: Episode index (used for logging only; all meta is fetched).
-        progress_file: Filename of the per-frame progress parquet inside the
-            dataset repo.

    Returns:
        Local cache path for the downloaded snapshot.
    """
-    logging.info("[1/4] Downloading metadata + %s for %s (episode %d) ...", progress_file, repo_id, episode)
+    logging.info("[1/4] Downloading metadata for %s (episode %d) ...", repo_id, episode)
    local_path = Path(
        snapshot_download(
            repo_id=repo_id,
            repo_type="dataset",
-            allow_patterns=["meta/**", progress_file],
+            allow_patterns=["meta/**", "sarm_progress.parquet"],
            ignore_patterns=["*.mp4"],
        )
    )
@@ -221,28 +215,25 @@ def download_video_file(repo_id: str, local_path: Path, video_rel: str) -> Path:
    return video_path


-def load_progress_data(
-    local_path: Path, episode: int, progress_file: str = "sarm_progress.parquet"
-) -> np.ndarray | None:
-    """Load per-frame progress values for an episode.
+def load_progress_data(local_path: Path, episode: int) -> np.ndarray | None:
+    """Load sarm_progress values for an episode.

    Args:
        local_path: Dataset cache root.
        episode: Episode index.
-        progress_file: Filename of the per-frame progress parquet.

    Returns:
        Sorted (N, 2) array of (frame_index, progress), or None if unavailable.
    """
-    parquet_path = local_path / progress_file
+    parquet_path = local_path / "sarm_progress.parquet"
    if not parquet_path.exists():
-        logging.warning("%s not found", progress_file)
+        logging.warning("sarm_progress.parquet not found")
        return None
    df = pd.read_parquet(parquet_path)
-    logging.info("   %s columns: %s", progress_file, list(df.columns))
+    logging.info("   sarm_progress.parquet columns: %s", list(df.columns))
    episode_df = df[df["episode_index"] == episode].copy()
    if episode_df.empty:
-        logging.warning("No progress rows for episode %d in %s", episode, progress_file)
+        logging.warning("No sarm_progress rows for episode %d", episode)
        return None
    episode_df = episode_df.sort_values("frame_index")

@@ -585,7 +576,6 @@ def process_dataset(
    camera_key: str | None,
    output_dir: Path,
    create_gif: bool = False,
-    progress_file: str = "sarm_progress.parquet",
 ) -> Path | None:
    """Full pipeline: download, extract metadata, composite progress, write output.

@@ -595,8 +585,6 @@ def process_dataset(
        camera_key: Camera key to use, or None for auto-selection.
        output_dir: Directory to write output files.
        create_gif: If True, also generate a GIF from the MP4.
-        progress_file: Filename of the per-frame progress parquet inside the
-            dataset repo.

    Returns:
        Path to the final output file, or None on failure.
@@ -604,7 +592,7 @@ def process_dataset(
    safe_name = repo_id.replace("/", "_")
    logging.info("Processing: %s  |  episode %d", repo_id, episode)

-    local_path = download_episode_metadata(repo_id, episode, progress_file)
+    local_path = download_episode_metadata(repo_id, episode)
    logging.info("   Local cache: %s", local_path)

    episode_meta = load_episode_meta(local_path, episode, camera_key)
@@ -612,9 +600,9 @@ def process_dataset(

    video_path = download_video_file(repo_id, local_path, episode_meta["video_rel"])

-    progress_data = load_progress_data(local_path, episode, progress_file)
+    progress_data = load_progress_data(local_path, episode)
    if progress_data is None:
-        logging.error("Could not load progress data from %s. Skipping overlay.", progress_file)
+        logging.error("Could not load sarm_progress data. Skipping overlay.")
        return None

    logging.info("   Progress frames: %d", len(progress_data))
@@ -639,7 +627,7 @@ def process_dataset(

 def main() -> None:
    parser = argparse.ArgumentParser(
-        description="Create MP4/GIF videos with per-frame progress overlay for dataset episodes."
+        description="Create MP4/GIF videos with sarm_progress overlay for dataset episodes."
    )
    parser.add_argument(
        "--repo-id",
@@ -670,15 +658,6 @@ def main() -> None:
        action="store_true",
        help="Also generate a GIF from the MP4 output.",
    )
-    parser.add_argument(
-        "--progress-file",
-        type=str,
-        default="sarm_progress.parquet",
-        help=(
-            "Filename of the per-frame progress parquet inside the dataset repo "
-            "(default: 'sarm_progress.parquet')."
-        ),
-    )
    args = parser.parse_args()

    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
@@ -691,7 +670,6 @@ def main() -> None:
        camera_key=args.camera_key,
        output_dir=args.output_dir,
        create_gif=args.gif,
-        progress_file=args.progress_file,
    )

    if result:
--- a/examples/dataset/slurm_compute_rabc.py
+++ b/examples/dataset/slurm_compute_rabc.py
@@ -69,7 +69,7 @@ class ComputeProgressShards(PipelineStep):
        import torch
        from tqdm import tqdm

-        from lerobot.rewards.sarm.compute_rabc_weights import (
+        from lerobot.policies.sarm.compute_rabc_weights import (
            generate_all_frame_indices,
            interpolate_progress,
            load_sarm_resources,
--- a/examples/notebooks/quickstart.ipynb
+++ b/examples/notebooks/quickstart.ipynb
@@ -80,7 +80,7 @@
    "}\n",
    "\n",
    "# Dataset\n",
-    "HF_USER = \"your_hf_username\"  # `hf auth whoami` to find your username\n",
+    "HF_USER = \"your_hf_username\"  # `huggingface-cli whoami` to find your username\n",
    "DATASET_NAME = \"my_so101_dataset\"\n",
    "TASK_DESCRIPTION = \"pick and place the block\"\n",
    "NUM_EPISODES = 10\n",
@@ -291,34 +291,7 @@
    "\n",
    "Uses `POLICY_PATH` from the Configuration cell (defaults to the Hub repo ID). You can also put there the `LAST_CHECKPOINT_PATH`.\n",
    "\n",
-    "See the [inference docs](https://huggingface.co/docs/lerobot/il_robots#run-inference-and-evaluate-your-policy) for details.\n",
-    "\n",
-    "Recently ```lerobot-rollout``` was introduced, you can [read more about it here](https://huggingface.co/docs/lerobot/main/en/il_robots?eval=Base+mode+%28no+recording%29#run-inference-and-evaluate-your-policy)."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "print_cmd(\n",
-    "    \"lerobot-rollout\",\n",
-    "    \"--strategy.type=base\",\n",
-    "    f\"--policy.path={POLICY_PATH}\",\n",
-    "    f\"--robot.type={ROBOT_TYPE}\",\n",
-    "    f\"--robot.port={ROBOT_PORT}\",\n",
-    "    CAMERAS_FLAG,\n",
-    "    f'--task=\"{TASK_DESCRIPTION}\"',\n",
-    "    \"--duration=60\",\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "if you are using the V0.5.1 release you should use ```lerobot-record``` instead of rollout"
+    "See the [inference docs](https://huggingface.co/docs/lerobot/il_robots#run-inference-and-evaluate-your-policy) for details."
   ]
  },
  {
--- a/examples/omx/README.md
+++ b/examples/omx/README.md
@@ -1,136 +0,0 @@
-# OMX Follower — Cube Pick And Place Example
-
-This is an example of what is possible to do with LeRobot on a physical setup.
-It is a WIP and being used internally at LeRobot and specific to our setup, but we hope it can be a useful reference for how to use LeRobot APIs and CLIs.
-
-It includes an end-to-end example for the **OMX Follower** robot arm: pick and place a cube dataset, train a policy, and deploy it autonomously.
-
-## Hardware
-
-| Component | Value                                |
-| --------- | ------------------------------------ |
-| Robot     | OMX Follower                         |
-| Cameras   | 2× OpenCV cameras (wrist + top-down) |
-
-## Scripts
-
-| Script                 | Purpose                                                         |
-| ---------------------- | --------------------------------------------------------------- |
-| `reset_environment.py` | Standalone utility: sweep workspace, grab cube, place cube      |
-| `record_grab.py`       | Automated data collection: reset → place → record grab episodes |
-
-## Setup
-
-Make sure you have LeRobot installed in your env. (See [the installation guide](https://huggingface.co/docs/lerobot/installation))
-
-Next, we will declare some environment variables for convenience. Adjust the camera indices and robot port to match your system configuration.
-
-```bash
-export ROBOT_PORT=/dev/ttyACM0
-export TELEOP_PORT=/dev/ttyACM1
-export HF_USERNAME=<your_hf_username>
-export ROBOT_CAMERAS="{ wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30, fourcc: MJPG}, top: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30, fourcc: MJPG} }"
-```
-
-## Step 1 — Collect Data
-
-```bash
-lerobot-record \
-    --robot.type=omx_follower \
-    --robot.port=$ROBOT_PORT \
-    --robot.id=omx_follower \
-    --robot.cameras="$ROBOT_CAMERAS" \
-    --teleop.type=omx_leader \
-    --teleop.port=$TELEOP_PORT \
-    --teleop.id=omx_leader \
-    --dataset.repo_id=$HF_USERNAME/omx_pickandplace \
-    --dataset.root=data/omx_pickandplace \
-    --dataset.num_episodes=50 \
-    --dataset.single_task="Pick the cube and place it in the blue square" \
-    --dataset.streaming_encoding=true \
-    --dataset.push_to_hub=true
-```
-
-### Bonus Auto-Collect script
-
-/!\ This is specific to our setup and the task of picking and placing a cube. It is not a general-purpose data collection script. As you may notice, it doesn't require a teleop.
-
-```bash
-python -m examples.omx.record_grab \
-    --robot.type=omx_follower \
-    --robot.port=$ROBOT_PORT \
-    --robot.id=omx_follower \
-    --robot.cameras="$ROBOT_CAMERAS" \
-    --dataset.repo_id=$HF_USERNAME/omx_pickandplace \
-    --dataset.root=data/omx_pickandplace \
-    --dataset.num_episodes=50 \
-    --dataset.single_task="Pick the cube and place it in the blue square" \
-    --dataset.streaming_encoding=true \
-    --dataset.push_to_hub=true
-```
-
-Each episode:
-
-1. The arm grabs the cube from the center of the workspace and places it at a random position.
-2. The arm returns to HOME.
-3. A targeted grab is recorded: HOME → approach raised → lower onto cube → grasp → lift → carry → drop → HOME.
-
-A dataset is already available here [`maximellerbach/omx_pickandplace`](https://huggingface.co/datasets/maximellerbach/omx_pickandplace), so you can skip directly to training if you want.
-
-## Step 2 — Train
-
-To train a simple `ACT` policy on the collected dataset, you can use the `lerobot-train` CLI:
-
-```bash
-lerobot-train \
-    --dataset.repo_id=$HF_USERNAME/omx_pickandplace \
-    --policy.type=act \
-    --output_dir=outputs/train/omx_pickandplace_act \
-    --policy.device=cuda \
-    --policy.repo_id=$HF_USERNAME/omx_pickandplace_act \
-    --steps=20000 \
-    --wandb.enable=true
-```
-
-A pretrained `ACT` policy is already available here [`maximellerbach/omx_pickandplace_act`](https://huggingface.co/maximellerbach/omx_pickandplace_act).
-
-## Step 3 — Rollout
-
-Use the `lerobot-rollout` CLI with base strategy:
-
-```bash
-lerobot-rollout \
-    --strategy.type=base \
-    --robot.type=omx_follower \
-    --robot.port=$ROBOT_PORT \
-    --robot.id=omx_follower \
-    --robot.cameras="$ROBOT_CAMERAS" \
-    --policy.path=$HF_USERNAME/omx_pickandplace_act \
-```
-
-For continuous recording with automatic upload (sentry mode):
-
-```bash
-lerobot-rollout \
-    --strategy.type=sentry \
-    --strategy.upload_every_n_episodes=10 \
-    --robot.type=omx_follower \
-    --robot.port=$ROBOT_PORT \
-    --robot.id=omx_follower \
-    --robot.cameras="$ROBOT_CAMERAS" \
-    --policy.path=$HF_USERNAME/omx_pickandplace_act \
-    --dataset.repo_id=$HF_USERNAME/rollout_omx_pickandplace_act \
-```
-
-## Environment Reset Utility
-
-Those are specific to this particular physical setup. Those are scripts that execute hardcoded sequences of actions on the robot to reset the environment, which is useful for data collection and evaluation. They are not general-purpose scripts.
-
-`reset_environment.py` can be run standalone to prepare the workspace:
-
-```bash
-# Grab cube + place it at a random position on the left side
-python -m examples.omx.reset_environment --port $ROBOT_PORT --mode grab_and_place
-```
-
-It also exposes `grab_cube(robot)` and `place_cube(robot)` for use in custom scripts.
--- a/examples/omx/record_grab.py
+++ b/examples/omx/record_grab.py
@@ -1,422 +0,0 @@
-#!/usr/bin/env python3
-"""
-Auto-record grab episodes for the OMX robot arm.
-
-Each episode cycle:
-  1. grab_and_place  — grab cube from workspace center and place at a random (pan, reach) position
-  2. HOME            — return arm to home with gripper open
-  3. record_grab     — execute a targeted grab to the stored position while recording
-                       observations + actions to a LeRobotDataset
-
-Usage (run from repo root):
-    python -m examples.omx.record_grab \\
-        --robot.type=omx_follower \\
-        --robot.port=/dev/ttyACM0 \\
-        --robot.id=omx_follower \\
-        --robot.cameras="{ wrist: {type: opencv, index_or_path: 6, width: 640, height: 480, fps: 30, fourcc: MJPG}, top: {type: opencv, index_or_path: 4, width: 640, height: 480, fps: 30, fourcc: MJPG} }" \\
-        --dataset.repo_id=<hf_username>/<dataset_name> \\
-        --dataset.root=data/omx_grab \\
-        --dataset.num_episodes=50 \\
-        --dataset.single_task="Grab the cube" \\
-        --dataset.streaming_encoding=true
-"""
-
-import logging
-from dataclasses import dataclass
-from pprint import pformat
-
-import numpy as np
-
-from lerobot.cameras import CameraConfig  # noqa: F401
-from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
-from lerobot.configs import parser
-from lerobot.configs.dataset import DatasetRecordConfig
-from lerobot.datasets import (
-    LeRobotDataset,
-    VideoEncodingManager,
-    aggregate_pipeline_dataset_features,
-    create_initial_features,
-)
-from lerobot.processor import make_default_processors
-from lerobot.robots import RobotConfig, make_robot_from_config
-from lerobot.robots.omx_follower import OmxFollower
-from lerobot.utils.constants import ACTION, OBS_STR
-from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
-from lerobot.utils.robot_utils import precise_sleep
-
-from .reset_environment import (
-    APPROACH_SPEED,
-    GRIPPER_CLOSE_POS,
-    HOME_POSE,
-    PUSH_END_ELBOW_FLEX,
-    PUSH_END_SHOULDER_LIFT,
-    PUSH_START_ELBOW_FLEX,
-    PUSH_START_SHOULDER_LIFT,
-    array_to_pose,
-    grab_cube,
-    horizontal_wrist_flex,
-    move_to_pose,
-    place_cube,
-    pose_to_array,
-)
-
-# ── Grab-episode motion parameters ────────────────────────────────────────────
-
-# Shoulder-lift offset for the raised approach phase (subtracted from the target sl, arm is higher).
-GRAB_RAISE_SL_OFFSET = 20.0
-GRAB_LOWER_SPEED = 20.0
-RECORD_SPEED = 30.0
-
-# Pose the arm travels to after closing the gripper (cube held).
-GRAB_CARRY_POSE = {
-    "shoulder_pan.pos": -23.0,
-    "shoulder_lift.pos": 5.0,
-    "elbow_flex.pos": 18.0,
-    "wrist_flex.pos": -14.0,
-    "wrist_roll.pos": 0.0,
-    "gripper.pos": GRIPPER_CLOSE_POS,
-}
-
-# Per-joint jitter limits (degrees) applied to transit waypoints for human-like variation.
-# Cube-approach and carry poses are never jittered to preserve precision.
-_JITTER_LIMITS: dict[str, float] = {
-    "shoulder_pan.pos": 5.0,
-    "shoulder_lift.pos": 4.0,
-    "elbow_flex.pos": 4.0,
-    "wrist_flex.pos": 3.0,
-    "wrist_roll.pos": 2.0,
-    "gripper.pos": 0.0,
-}
-
-
-def _jitter_pose(pose: dict, rng: np.random.Generator) -> dict:
-    """Return a copy of pose with independent per-joint random perturbations."""
-    return {
-        k: v + rng.uniform(-_JITTER_LIMITS.get(k, 0.0), _JITTER_LIMITS.get(k, 0.0)) for k, v in pose.items()
-    }
-
-
-def _random_stuck_pose(rng: np.random.Generator) -> dict:
-    """Return a physically plausible stuck pose (failed grasp), gripper closed.
-
-    ef bounds are piecewise-linear in sl so the arm stays in a reachable,
-    table-safe envelope across the full sl range:
-      sl=-50 → ef ∈ [  0,  50]   (arm raised, can be bent forward)
-      sl=  0 → ef ∈ [-25,  25]   (mid reach)
-      sl= 30 → ef ∈ [-20,   0]   (arm extended, little room to flex)
-    wrist_flex is randomly offset from the horizontal value.
-    """
-    pan = float(rng.uniform(-5.0, 35.0))
-    sl = float(rng.uniform(-50.0, 30.0))
-
-    if sl <= 0.0:
-        alpha = (sl + 50.0) / 50.0  # 0 at sl=-50, 1 at sl=0
-        ef_lo = alpha * -25.0  # 0 → -25
-        ef_hi = 50.0 + alpha * -25.0  # 50 → 25
-    else:
-        alpha = sl / 30.0  # 0 at sl=0, 1 at sl=30
-        ef_lo = -25.0 + alpha * 5.0  # -25 → -20
-        ef_hi = 25.0 + alpha * -25.0  # 25 → 0
-
-    ef = float(rng.uniform(ef_lo, ef_hi))
-    wf = horizontal_wrist_flex(sl, ef) + float(rng.uniform(-15.0, 15.0))
-    return {
-        "shoulder_pan.pos": pan,
-        "shoulder_lift.pos": sl,
-        "elbow_flex.pos": ef,
-        "wrist_flex.pos": wf,
-        "wrist_roll.pos": float(rng.uniform(-15.0, 15.0)),
-        "gripper.pos": GRIPPER_CLOSE_POS,
-    }
-
-
-logger = logging.getLogger(__name__)
-
-
-@dataclass
-class OmxRecordGrabConfig:
-    robot: RobotConfig
-    dataset: DatasetRecordConfig
-    # Resume recording on an existing dataset.
-    resume: bool = False
-    # Fraction of episodes that start from a random stuck pose (gripper closed) to
-    # generate recovery data.  0.0 = disabled, 1.0 = all episodes are recovery starts.
-    recovery_prob: float = 0.5
-
-
-def record_episode_spline(
-    robot: OmxFollower,
-    waypoints: list[dict],
-    speeds: list[float],
-    dataset: LeRobotDataset,
-    task: str,
-) -> None:
-    """Execute a Catmull-Rom-style spline through waypoints, recording each frame.
-
-    Segment durations are parameterized from the maximum absolute joint delta
-    between consecutive waypoints divided by the requested segment speed,
-    producing non-uniform timing in joint space. Interior tangents are derived
-    from the adjacent per-segment velocities, with clamped (zero-velocity)
-    endpoints so the arm starts and stops smoothly. Each segment is cubic
-    Hermite, giving C1 continuity at every waypoint.
-    """
-    pts = [pose_to_array(w) for w in waypoints]
-    n = len(pts)
-
-    # Steps and duration per segment
-    n_steps_list = []
-    timestamps = []
-    for i in range(n - 1):
-        max_dist = float(np.max(np.abs(pts[i + 1] - pts[i])))
-        ns = max(1, int(max_dist / speeds[i] * dataset.fps)) if max_dist >= 0.5 else 0
-        n_steps_list.append(ns)
-        timestamps.append(ns / dataset.fps)
-
-    # Velocity tangents (deg/sec) — clamped at endpoints, Catmull-Rom for interior
-    vels = [np.zeros_like(pts[0])]
-    for i in range(1, n - 1):
-        v_prev = (pts[i] - pts[i - 1]) / timestamps[i - 1] if timestamps[i - 1] > 0 else np.zeros_like(pts[0])
-        v_next = (pts[i + 1] - pts[i]) / timestamps[i] if timestamps[i] > 0 else np.zeros_like(pts[0])
-        vels.append(0.5 * (v_prev + v_next))
-    vels.append(np.zeros_like(pts[0]))
-
-    dt = 1.0 / dataset.fps
-    for seg in range(n - 1):
-        ns = n_steps_list[seg]
-        if ns == 0:
-            continue
-        p0, p1 = pts[seg], pts[seg + 1]
-        # Scale velocity (deg/sec) to t-space tangent (deg/t-unit, where t: 0→1 over ns steps)
-        m0 = vels[seg] * timestamps[seg]
-        m1 = vels[seg + 1] * timestamps[seg]
-
-        for step in range(1, ns + 1):
-            t = step / ns
-            h00 = 2 * t**3 - 3 * t**2 + 1
-            h10 = t**3 - 2 * t**2 + t
-            h01 = -2 * t**3 + 3 * t**2
-            h11 = t**3 - t**2
-            commanded = h00 * p0 + h10 * m0 + h01 * p1 + h11 * m1
-
-            action = array_to_pose(commanded)
-            robot.send_action(action)
-            obs = robot.get_observation()
-            obs_frame = build_dataset_frame(dataset.features, obs, prefix=OBS_STR)
-            action_frame = build_dataset_frame(dataset.features, action, prefix=ACTION)
-            dataset.add_frame({**obs_frame, **action_frame, "task": task})
-            precise_sleep(dt)
-
-
-def record_grab_episode(
-    robot: OmxFollower,
-    dataset: LeRobotDataset,
-    pan: float,
-    t: float,
-    task: str,
-    recovery_start: bool = False,
-) -> None:
-    """Execute a targeted grab to the stored (pan, t) position, recording every frame.
-
-    Normal sequence (initial HOME move is NOT recorded):
-      HOME → raised approach above cube → lower → close gripper
-           → raise [jittered] → retract [jittered] → GRAB_CARRY_POSE → drop → HOME
-
-    Recovery sequence (recovery_start=True): arm is moved to a random stuck pose
-    (gripper closed) without recording, then recording begins from there:
-      stuck_pose → raised approach above cube → [normal grab sequence from there]
-
-    All segments are joined by a Catmull-Rom spline (C1-continuous velocities).
-    """
-    sl = PUSH_START_SHOULDER_LIFT + t * (PUSH_END_SHOULDER_LIFT - PUSH_START_SHOULDER_LIFT)
-    ef = PUSH_START_ELBOW_FLEX + t * (PUSH_END_ELBOW_FLEX - PUSH_START_ELBOW_FLEX)
-    sl_raised = sl - GRAB_RAISE_SL_OFFSET
-    wf_horizontal = horizontal_wrist_flex(sl, ef)
-
-    rng = np.random.default_rng()
-
-    if recovery_start:
-        stuck_pose = _random_stuck_pose(rng)
-        logger.info(f"Recovery start: {stuck_pose}")
-        move_to_pose(robot, stuck_pose, APPROACH_SPEED)
-        first_waypoints = [stuck_pose]
-        first_speeds = []
-    else:
-        jittery_start = _jitter_pose(HOME_POSE, rng)
-        move_to_pose(robot, jittery_start, APPROACH_SPEED)
-        first_waypoints = [jittery_start]
-        first_speeds = []
-
-    waypoints = first_waypoints + [
-        {  # raised approach: arm above cube
-            "shoulder_pan.pos": pan,
-            "shoulder_lift.pos": sl_raised,
-            "elbow_flex.pos": ef,
-            "wrist_flex.pos": horizontal_wrist_flex(sl_raised, ef),
-            "wrist_roll.pos": 0.0,
-            "gripper.pos": 60.0,
-        },
-        {  # lower onto cube — no jitter: precision needed
-            "shoulder_pan.pos": pan,
-            "shoulder_lift.pos": sl,
-            "elbow_flex.pos": ef,
-            "wrist_flex.pos": wf_horizontal,
-            "wrist_roll.pos": 0.0,
-            "gripper.pos": 60.0,
-        },
-        {  # close gripper — no jitter: precision needed
-            "shoulder_pan.pos": pan,
-            "shoulder_lift.pos": sl,
-            "elbow_flex.pos": ef,
-            "wrist_flex.pos": wf_horizontal,
-            "wrist_roll.pos": 0.0,
-            "gripper.pos": GRIPPER_CLOSE_POS,
-        },
-        _jitter_pose(
-            {  # raise with cube
-                "shoulder_pan.pos": pan,
-                "shoulder_lift.pos": sl_raised,
-                "elbow_flex.pos": ef,
-                "wrist_flex.pos": horizontal_wrist_flex(sl_raised, ef),
-                "wrist_roll.pos": 0.0,
-                "gripper.pos": GRIPPER_CLOSE_POS,
-            },
-            rng,
-        ),
-        _jitter_pose(
-            {  # retract: fold arm toward HOME before sweeping to carry zone
-                "shoulder_pan.pos": pan * 0.25,
-                "shoulder_lift.pos": HOME_POSE["shoulder_lift.pos"] + 5.0,
-                "elbow_flex.pos": HOME_POSE["elbow_flex.pos"] - 5.0,
-                "wrist_flex.pos": 0.0,
-                "wrist_roll.pos": 0.0,
-                "gripper.pos": GRIPPER_CLOSE_POS,
-            },
-            rng,
-        ),
-        GRAB_CARRY_POSE,  # no jitter: target drop zone
-        {**GRAB_CARRY_POSE, "gripper.pos": 60.0},  # drop cube
-        HOME_POSE,
-    ]
-    speeds = first_speeds + [
-        RECORD_SPEED,  # (HOME →) raised approach
-        GRAB_LOWER_SPEED,  # raised approach → lower
-        GRAB_LOWER_SPEED,  # lower → close gripper
-        RECORD_SPEED,  # close gripper → raise
-        RECORD_SPEED,  # raise → retract
-        RECORD_SPEED,  # retract → carry pose
-        RECORD_SPEED,  # carry pose → drop
-        RECORD_SPEED,  # drop → HOME
-    ]
-
-    record_episode_spline(robot, waypoints, speeds, dataset, task)
-
-    # Dwell at HOME for ~0.5 s before next episode
-    home_action = build_dataset_frame(dataset.features, HOME_POSE, prefix=ACTION)
-    dt = 1.0 / dataset.fps
-    for _ in range(int(dataset.fps * 0.5)):
-        robot.send_action(HOME_POSE)
-        obs = robot.get_observation()
-        obs_frame = build_dataset_frame(dataset.features, obs, prefix=OBS_STR)
-        dataset.add_frame({**obs_frame, **home_action, "task": task})
-        precise_sleep(dt)
-
-
-@parser.wrap()
-def record_grab(cfg: OmxRecordGrabConfig) -> LeRobotDataset:
-    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
-    logger.info(pformat(cfg))
-
-    robot = make_robot_from_config(cfg.robot)
-    use_videos = cfg.dataset.video
-
-    teleop_action_processor, _, robot_obs_processor = make_default_processors()
-
-    dataset_features = combine_feature_dicts(
-        aggregate_pipeline_dataset_features(
-            pipeline=teleop_action_processor,
-            initial_features=create_initial_features(action=robot.action_features),
-            use_videos=use_videos,
-        ),
-        aggregate_pipeline_dataset_features(
-            pipeline=robot_obs_processor,
-            initial_features=create_initial_features(observation=robot.observation_features),
-            use_videos=use_videos,
-        ),
-    )
-
-    num_cameras = len(robot.cameras) if hasattr(robot, "cameras") else 0
-    dataset = None
-
-    try:
-        if cfg.resume:
-            dataset = LeRobotDataset.resume(
-                cfg.dataset.repo_id,
-                root=cfg.dataset.root,
-                streaming_encoding=cfg.dataset.streaming_encoding,
-                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
-                vcodec=cfg.dataset.vcodec,
-                encoder_threads=cfg.dataset.encoder_threads,
-                image_writer_processes=cfg.dataset.num_image_writer_processes if num_cameras > 0 else 0,
-                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * num_cameras
-                if num_cameras > 0
-                else 0,
-            )
-        else:
-            cfg.dataset.stamp_repo_id()
-            dataset = LeRobotDataset.create(
-                cfg.dataset.repo_id,
-                cfg.dataset.fps,
-                root=cfg.dataset.root,
-                robot_type=robot.name,
-                features=dataset_features,
-                use_videos=use_videos,
-                streaming_encoding=cfg.dataset.streaming_encoding,
-                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
-                vcodec=cfg.dataset.vcodec,
-                encoder_threads=cfg.dataset.encoder_threads,
-                image_writer_processes=cfg.dataset.num_image_writer_processes if num_cameras > 0 else 0,
-                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * num_cameras
-                if num_cameras > 0
-                else 0,
-            )
-
-        robot.connect(calibrate=True)
-
-        rng = np.random.default_rng()
-        with VideoEncodingManager(dataset):
-            for episode_idx in range(cfg.dataset.num_episodes):
-                logger.info(f"=== Episode {episode_idx + 1}/{cfg.dataset.num_episodes} ===")
-
-                logger.info("Step 1: grabbing and placing cube...")
-                grab_cube(robot)
-                pan, t = place_cube(robot)
-                logger.info(f"Cube placed at pan={pan:.1f}, reach={t:.2f}")
-
-                recovery_start = cfg.recovery_prob > 0 and float(rng.random()) < cfg.recovery_prob
-                logger.info(f"Step 2: recording {'recovery ' if recovery_start else ''}grab episode...")
-                record_grab_episode(
-                    robot,
-                    dataset,
-                    pan,
-                    t,
-                    cfg.dataset.single_task,
-                    recovery_start=recovery_start,
-                )
-
-                dataset.save_episode()
-                logger.info(f"Episode {episode_idx + 1} saved.")
-
-    finally:
-        if dataset:
-            dataset.finalize()
-        if robot.is_connected:
-            robot.disconnect()
-
-    if cfg.dataset.push_to_hub and dataset and dataset.num_episodes > 0:
-        dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
-
-    return dataset
-
-
-if __name__ == "__main__":
-    record_grab()
--- a/examples/omx/reset_environment.py
+++ b/examples/omx/reset_environment.py
@@ -1,267 +0,0 @@
-#!/usr/bin/env python3
-"""
-Auto-reset and cube-grab utility for the OMX robot arm.
-
-Provides:
-  - grab_cube(robot): sweep workspace, center cube, close gripper
-  - place_cube(robot): carry cube to a random position, release
-
-Standalone usage (run from repo root):
-    python -m examples.omx.reset_environment --port /dev/ttyACM1 --mode grab
-    python -m examples.omx.reset_environment --port /dev/ttyACM1 --mode grab_and_place
-
-Joint range: -100 to 100 for arm joints; gripper: 50 = closed, 80 = open.
-
-To read current joint values for calibration, add after robot.connect():
-    obs = robot.get_observation()
-    print({k: round(obs[k], 1) for k in JOINT_NAMES})
-    robot.disconnect(); raise SystemExit
-
-Parallel-to-ground IK: wrist_flex = WRIST_HORIZONTAL_OFFSET - shoulder_lift - elbow_flex.
-Linear interpolation preserves this constraint between any two poses that satisfy it.
-"""
-
-import argparse
-import logging
-
-import numpy as np
-
-from lerobot.robots.omx_follower import OmxFollower, OmxFollowerConfig
-from lerobot.robots.robot import Robot
-from lerobot.utils.robot_utils import precise_sleep
-
-logger = logging.getLogger(__name__)
-
-# ── Poses ─────────────────────────────────────────────────────────────────────
-
-HOME_POSE = {
-    "shoulder_pan.pos": 0.0,
-    "shoulder_lift.pos": -50.0,
-    "elbow_flex.pos": 50.0,
-    "wrist_flex.pos": 0.0,
-    "wrist_roll.pos": 0.0,
-    "gripper.pos": 60.0,
-}
-
-SWEEP_WAYPOINTS = [
-    {
-        "shoulder_pan.pos": -60.0,
-        "shoulder_lift.pos": 50.0,
-        "elbow_flex.pos": -60.0,
-        "wrist_flex.pos": -20.0,
-        "wrist_roll.pos": 0.0,
-        "gripper.pos": 60.0,
-    },
-    {
-        "shoulder_pan.pos": -30.0,
-        "shoulder_lift.pos": 50.0,
-        "elbow_flex.pos": -60.0,
-        "wrist_flex.pos": -5.0,
-        "wrist_roll.pos": 0.0,
-        "gripper.pos": 60.0,
-    },
-    {
-        "shoulder_pan.pos": 20.0,
-        "shoulder_lift.pos": 50.0,
-        "elbow_flex.pos": -55.0,
-        "wrist_flex.pos": -5.0,
-        "wrist_roll.pos": 0.0,
-        "gripper.pos": 60.0,
-    },
-]
-
-# ── Motion parameters ─────────────────────────────────────────────────────────
-
-CONTROL_HZ = 30
-APPROACH_SPEED = 50.0
-SWEEP_SPEED = 40.0
-
-# ── Grab-sequence parameters ──────────────────────────────────────────────────
-
-GRAB_PAN = 0.0
-SWEEP_LEFT_PAN = -60.0
-SWEEP_RIGHT_PAN = 60.0
-SWEEP_END_OFFSET = 5.0  # stop before center so the cube isn't pushed past GRAB_PAN
-SWEEP_END_PAN_RANGE = (15.0, 20.0)
-
-SWEEP_LOW_SHOULDER_LIFT = 50.0
-SWEEP_LOW_ELBOW_FLEX_START = -60.0
-SWEEP_LOW_ELBOW_FLEX_END = -55.0
-
-SWEEP_HIGH_WRIST_FLEX = -20.0  # wrist tilted up during high approach to clear obstacles
-
-PUSH_START_SHOULDER_LIFT = 0.0
-PUSH_START_ELBOW_FLEX = 45.0
-PUSH_END_SHOULDER_LIFT = 50.0
-PUSH_END_ELBOW_FLEX = -50.0
-# Subtracted from shoulder_lift during the push sweep to clear the platform surface.
-# Does not affect the grab-target interpolation in record_grab.py.
-PUSH_RAISE_OFFSET = 5.0
-
-WRIST_HORIZONTAL_OFFSET = 0.0  # tune if gripper tilts during push: + tilts nose up, - down
-GRIPPER_CLOSE_POS = 50.0
-
-PLACE_LEFT_PAN_RANGE = (5.0, 30.0)  # random pan range for cube placement on the left side
-PLACE_REACH_RANGE = (0.1, 0.7)  # 0 = arm retracted (PUSH_START), 1 = fully extended (PUSH_END)
-
-JOINT_NAMES = [
-    "shoulder_pan.pos",
-    "shoulder_lift.pos",
-    "elbow_flex.pos",
-    "wrist_flex.pos",
-    "wrist_roll.pos",
-    "gripper.pos",
-]
-
-# ── Helpers ───────────────────────────────────────────────────────────────────
-
-
-def pose_to_array(pose: dict) -> np.ndarray:
-    return np.array([pose[k] for k in JOINT_NAMES])
-
-
-def array_to_pose(arr: np.ndarray) -> dict:
-    return {k: float(arr[i]) for i, k in enumerate(JOINT_NAMES)}
-
-
-def horizontal_wrist_flex(shoulder_lift: float, elbow_flex: float) -> float:
-    return WRIST_HORIZONTAL_OFFSET - shoulder_lift - elbow_flex
-
-
-def _low_sweep_pose(pan: float, elbow_flex: float, wrist_flex: float | None = None) -> dict:
-    sl = SWEEP_LOW_SHOULDER_LIFT
-    return {
-        "shoulder_pan.pos": pan,
-        "shoulder_lift.pos": sl,
-        "elbow_flex.pos": elbow_flex,
-        "wrist_flex.pos": horizontal_wrist_flex(sl, elbow_flex) if wrist_flex is None else wrist_flex,
-        "wrist_roll.pos": 0.0,
-        "gripper.pos": 60.0,
-    }
-
-
-def _high_sweep_pose(pan: float) -> dict:
-    return {**HOME_POSE, "shoulder_pan.pos": pan, "wrist_flex.pos": SWEEP_HIGH_WRIST_FLEX}
-
-
-def _push_pose(shoulder_lift: float, elbow_flex: float, pan: float = GRAB_PAN, gripper: float = 70.0) -> dict:
-    return {
-        "shoulder_pan.pos": pan,
-        "shoulder_lift.pos": shoulder_lift,
-        "elbow_flex.pos": elbow_flex,
-        "wrist_flex.pos": horizontal_wrist_flex(shoulder_lift, elbow_flex),
-        "wrist_roll.pos": 0.0,
-        "gripper.pos": gripper,
-    }
-
-
-def move_to_pose(robot: Robot, target: dict, speed: float) -> None:
-    """Interpolate from current position to target at the given speed (units/s)."""
-    obs = robot.get_observation()
-    current = np.array([obs[k] for k in JOINT_NAMES])
-    goal = pose_to_array(target)
-
-    max_distance = float(np.max(np.abs(goal - current)))
-    if max_distance < 0.5:
-        return
-
-    n_steps = max(1, int(max_distance / speed * CONTROL_HZ))
-    dt = 1.0 / CONTROL_HZ
-    for step in range(1, n_steps + 1):
-        t = step / n_steps
-        robot.send_action(array_to_pose(current + t * (goal - current)))
-        precise_sleep(dt)
-
-
-# ── Sequences ─────────────────────────────────────────────────────────────────
-
-
-def grab_cube(robot: Robot) -> None:
-    """Left sweep → right sweep → extend arm parallel to ground → close gripper."""
-    move_to_pose(robot, HOME_POSE, APPROACH_SPEED)
-
-    for pan, end_pan in [
-        (SWEEP_LEFT_PAN, GRAB_PAN - SWEEP_END_OFFSET),
-        (SWEEP_RIGHT_PAN, GRAB_PAN + SWEEP_END_OFFSET),
-    ]:
-        logger.info(f"Sweeping {'left' if pan < 0 else 'right'} → center...")
-        move_to_pose(robot, _high_sweep_pose(pan), APPROACH_SPEED)
-        move_to_pose(
-            robot, _low_sweep_pose(pan, SWEEP_LOW_ELBOW_FLEX_START, wrist_flex=-20.0), APPROACH_SPEED
-        )
-        move_to_pose(robot, _low_sweep_pose(end_pan, SWEEP_LOW_ELBOW_FLEX_END, wrist_flex=0.0), SWEEP_SPEED)
-        move_to_pose(robot, HOME_POSE, APPROACH_SPEED)
-
-    logger.info("Extending to push cube into gripper...")
-    move_to_pose(
-        robot,
-        _push_pose(PUSH_START_SHOULDER_LIFT - PUSH_RAISE_OFFSET, PUSH_START_ELBOW_FLEX),
-        APPROACH_SPEED,
-    )
-    move_to_pose(
-        robot,
-        _push_pose(PUSH_END_SHOULDER_LIFT - PUSH_RAISE_OFFSET, PUSH_END_ELBOW_FLEX),
-        SWEEP_SPEED,
-    )
-
-    logger.info("Closing gripper...")
-    move_to_pose(
-        robot,
-        _push_pose(PUSH_END_SHOULDER_LIFT, PUSH_END_ELBOW_FLEX, gripper=GRIPPER_CLOSE_POS),
-        APPROACH_SPEED,
-    )
-
-    logger.info("Grab complete.")
-
-
-def place_cube(robot: Robot) -> tuple[float, float]:
-    """Carry the cube (gripper closed) to a random position on the left side, then release.
-
-    Returns:
-        (pan, t): pan angle and reach scalar [0, 1] of the placement position.
-    """
-    pan = float(np.random.uniform(*PLACE_LEFT_PAN_RANGE))
-    t = float(np.random.uniform(*PLACE_REACH_RANGE))
-    sl = PUSH_START_SHOULDER_LIFT + t * (PUSH_END_SHOULDER_LIFT - PUSH_START_SHOULDER_LIFT)
-    ef = PUSH_START_ELBOW_FLEX + t * (PUSH_END_ELBOW_FLEX - PUSH_START_ELBOW_FLEX)
-    logger.info(f"Placing cube at pan={pan:.1f}, reach={t:.2f}...")
-
-    move_to_pose(robot, {**HOME_POSE, "gripper.pos": GRIPPER_CLOSE_POS}, APPROACH_SPEED)
-    move_to_pose(
-        robot, {**HOME_POSE, "shoulder_pan.pos": pan, "gripper.pos": GRIPPER_CLOSE_POS}, APPROACH_SPEED
-    )
-    move_to_pose(robot, _push_pose(sl, ef, pan=pan, gripper=GRIPPER_CLOSE_POS), APPROACH_SPEED)
-    move_to_pose(robot, _push_pose(sl, ef, pan=pan, gripper=80.0), APPROACH_SPEED)
-    move_to_pose(robot, HOME_POSE, APPROACH_SPEED)
-    logger.info("Place complete.")
-    return pan, t
-
-
-# ── Entry point ───────────────────────────────────────────────────────────────
-
-
-def main():
-    parser = argparse.ArgumentParser(description="OMX arm reset / grab script")
-    parser.add_argument("--port", default="/dev/ttyACM1")
-    parser.add_argument("--robot_id", default="omx_follower")
-    parser.add_argument("--mode", choices=["grab", "grab_and_place"], default="grab_and_place")
-    args = parser.parse_args()
-
-    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
-
-    robot = OmxFollower(OmxFollowerConfig(port=args.port, id=args.robot_id))
-    robot.connect(calibrate=True)
-
-    try:
-        if args.mode == "grab":
-            grab_cube(robot)
-        elif args.mode == "grab_and_place":
-            grab_cube(robot)
-            place_cube(robot)
-
-    finally:
-        robot.disconnect()
-
-
-if __name__ == "__main__":
-    main()
--- a/examples/tutorial/rl/hilserl_example.py
+++ b/examples/tutorial/rl/hilserl_example.py
@@ -4,13 +4,13 @@ from pathlib import Path
 from queue import Empty, Full

 import torch
+import torch.optim as optim

 from lerobot.datasets import LeRobotDataset
 from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
-from lerobot.policies import GaussianActorConfig
-from lerobot.policies.gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy
-from lerobot.rewards.classifier.modeling_classifier import Classifier
-from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
+from lerobot.policies import SACConfig
+from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
 from lerobot.robots.so_follower import SO100FollowerConfig
@@ -28,7 +28,7 @@ def run_learner(
    transitions_queue: mp.Queue,
    parameters_queue: mp.Queue,
    shutdown_event: mp.Event,
-    policy_learner: GaussianActorPolicy,
+    policy_learner: SACPolicy,
    online_buffer: ReplayBuffer,
    offline_buffer: ReplayBuffer,
    lr: float = 3e-4,
@@ -40,9 +40,8 @@ def run_learner(
    policy_learner.train()
    policy_learner.to(device)

-    algo_config = SACAlgorithmConfig.from_policy_config(policy_learner.config)
-    algorithm = SACAlgorithm(policy=policy_learner, config=algo_config)
-    algorithm.make_optimizers_and_scheduler()
+    # Create Adam optimizer from scratch - simple and clean
+    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)

    print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
    print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
@@ -84,26 +83,24 @@ def run_learner(
                else:
                    batch[key] = online_batch[key]

-            def batch_iter(b=batch):
-                while True:
-                    yield b
+            loss, _ = policy_learner.forward(batch)

-            stats = algorithm.update(batch_iter())
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
            training_step += 1

            if training_step % LOG_EVERY == 0:
-                log_dict = stats.to_log_dict()
                print(
-                    f"[LEARNER] Training step {training_step}, "
-                    f"critic_loss: {log_dict.get('critic', 'N/A'):.4f}, "
+                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
                    f"Buffers: Online={len(online_buffer)}, Offline={len(offline_buffer)}"
                )

            # Send updated parameters to actor every 10 training steps
            if training_step % SEND_EVERY == 0:
                try:
-                    weights = algorithm.get_weights()
-                    parameters_queue.put_nowait(weights)
+                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
+                    parameters_queue.put_nowait(state_dict)
                    print("[LEARNER] Sent updated parameters to actor")
                except Full:
                    # Missing write due to queue not being consumed (should happen rarely)
@@ -116,7 +113,7 @@ def run_actor(
    transitions_queue: mp.Queue,
    parameters_queue: mp.Queue,
    shutdown_event: mp.Event,
-    policy_actor: GaussianActorPolicy,
+    policy_actor: SACPolicy,
    reward_classifier: Classifier,
    env_cfg: HILSerlRobotEnvConfig,
    device: torch.device = "mps",
@@ -147,15 +144,15 @@ def run_actor(

            while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
                try:
-                    new_weights = parameters_queue.get_nowait()
-                    policy_actor.load_state_dict(new_weights)
+                    new_params = parameters_queue.get_nowait()
+                    policy_actor.load_state_dict(new_params)
                    print("[ACTOR] Updated policy parameters from learner")
                except Empty:  # No new updated parameters available from learner, waiting
                    pass

-                # Get action from policy (returns full action: continuous + discrete)
+                # Get action from policy
                policy_obs = make_policy_obs(obs, device=device)
-                action_tensor = policy_actor.select_action(policy_obs)
+                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
                action = action_tensor.squeeze(0).cpu().numpy()

                # Step environment
@@ -264,14 +261,14 @@ def main():
    action_features = hw_to_dataset_features(env.robot.action_features, "action")

    # Create SAC policy for action selection
-    policy_cfg = GaussianActorConfig(
+    policy_cfg = SACConfig(
        device=device,
        input_features=obs_features,
        output_features=action_features,
    )

-    policy_actor = GaussianActorPolicy(policy_cfg)
-    policy_learner = GaussianActorPolicy(policy_cfg)
+    policy_actor = SACPolicy(policy_cfg)
+    policy_learner = SACPolicy(policy_cfg)

    demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
    offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)
--- a/examples/tutorial/rl/reward_classifier_example.py
+++ b/examples/tutorial/rl/reward_classifier_example.py
@@ -1,7 +1,7 @@
 import torch

 from lerobot.datasets import LeRobotDataset
-from lerobot.rewards import RewardClassifierConfig, make_reward_model, make_reward_pre_post_processors
+from lerobot.policies import RewardClassifierConfig, make_policy, make_pre_post_processors


 def main():
@@ -22,10 +22,10 @@ def main():
        model_name="microsoft/resnet-18",
    )

-    # Make reward model, preprocessor, and optimizer
-    reward_model = make_reward_model(config, dataset_stats=dataset.meta.stats)
-    optimizer = config.get_optimizer_preset().build(reward_model.parameters())
-    preprocessor, _ = make_reward_pre_post_processors(config, dataset_stats=dataset.meta.stats)
+    # Make policy, preprocessor, and optimizer
+    policy = make_policy(config, ds_meta=dataset.meta)
+    optimizer = config.get_optimizer_preset().build(policy.parameters())
+    preprocessor, _ = make_pre_post_processors(policy_cfg=config, dataset_stats=dataset.meta.stats)

    classifier_id = "<user>/reward_classifier_hil_serl_example"

@@ -42,7 +42,7 @@ def main():
            batch = preprocessor(batch)

            # Forward pass
-            loss, output_dict = reward_model.forward(batch)
+            loss, output_dict = policy.forward(batch)

            # Backward pass and optimization
            optimizer.zero_grad()
@@ -58,8 +58,8 @@ def main():

    print("Training finished!")

-    # You can now save the trained reward model.
-    reward_model.push_to_hub(classifier_id)
+    # You can now save the trained policy.
+    policy.push_to_hub(classifier_id)


 if __name__ == "__main__":
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -59,8 +59,8 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici

 dependencies = [
    # Core ML
-    "torch>=2.7,<2.12.0",
-    "torchvision>=0.22.0,<0.27.0",
+    "torch>=2.7,<2.11.0",
+    "torchvision>=0.22.0,<0.26.0",
    "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
    "opencv-python-headless>=4.9.0,<4.14.0",
    "Pillow>=10.0.0,<13.0.0",
@@ -95,22 +95,11 @@ dependencies = [

 # ── Feature-scoped extras ──────────────────────────────────
 dataset = [
-    "datasets>=4.7.0,<5.0.0",
+    "datasets>=4.0.0,<5.0.0",
    "pandas>=2.0.0,<3.0.0", # NOTE: Transitive dependency of datasets
    "pyarrow>=21.0.0,<30.0.0", # NOTE: Transitive dependency of datasets
    "lerobot[av-dep]",
-
-    # NOTE: torchcodec wheel availability matrix (PyPI):
-    #   - linux x86_64/amd64 + macOS arm64 : wheels since 0.3.0 (the historic supported set).
-    #   - win32 x86_64                     : wheels since 0.7.0  (needs torch>=2.8).
-    #   - linux aarch64/arm64              : wheels since 0.11.0 (needs torch>=2.11).
-    #   - macOS x86_64 (Intel) and linux armv7l: no wheels in any released version -> fall through to the PyAV decoder.
-    # Each platform gets its own line so the resolver picks the minimum version that has a wheel for it.
-
-    # Other torch/torchcodec pairings (informational): 0.8.1 = ffmpeg>=8 support, 0.10 = system-wide ffmpeg support, 0.12 needs torch==2.12.
-    "torchcodec>=0.3.0,<0.12.0; (sys_platform == 'linux' and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64')",
-    "torchcodec>=0.7.0,<0.12.0; sys_platform == 'win32'",
-    "torchcodec>=0.11.0,<0.12.0; sys_platform == 'linux' and (platform_machine == 'aarch64' or platform_machine == 'arm64')",
+    "torchcodec>=0.3.0,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # NOTE: Windows support starts at version 0.7 (needs torch==2.8), ffmpeg>=8 support starts at version 0.8.1 (needs torch==2.9), system-wide ffmpeg support starts at version 0.10 (needs torch==2.10).
    "jsonlines>=4.0.0,<5.0.0",
 ]
 training = [
@@ -138,10 +127,8 @@ dataset_viz = ["lerobot[dataset]", "lerobot[viz]"]
 # Common
 av-dep = ["av>=15.0.0,<16.0.0"]
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-# NOTE: 0.9.16 links against liburdfdom_sensor.so.4, which is unavailable on Ubuntu 24.04
-# (noble ships urdfdom 3.x). Cap below 0.9.16 until system urdfdom 4.x is broadly available.
-placo-dep = ["placo>=0.9.6,<0.9.16"]
-transformers-dep = ["transformers>=5.4.0,<5.6.0"]
+placo-dep = ["placo>=0.9.6,<0.9.17"]
+transformers-dep = ["transformers==5.3.0"] # TODO(Steven): https://github.com/huggingface/lerobot/pull/3249
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
 can-dep = ["python-can>=4.2.0,<5.0.0"]
 peft-dep = ["peft>=0.18.0,<1.0.0"]
@@ -153,8 +140,6 @@ pyserial-dep = ["pyserial>=3.5,<4.0"]
 deepdiff-dep = ["deepdiff>=7.0.1,<9.0.0"]
 pynput-dep = ["pynput>=1.7.8,<1.9.0"]
 pyzmq-dep = ["pyzmq>=26.2.1,<28.0.0"]
-motorbridge-dep = ["motorbridge>=0.3.2,<0.4.0"]
-motorbridge-smart-servo-dep = ["motorbridge-smart-servo>=0.0.4,<0.1.0"]

 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
@@ -178,9 +163,6 @@ unitree_g1 = [
    "lerobot[pygame-dep]",
 ]
 reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
-# Seeed Studio reBot B601-DM follower (motorbridge / CAN) + StarArm102 / reBot Arm 102
-# leader (motorbridge-smart-servo / FashionStar UART servos).
-rebot = ["lerobot[motorbridge-dep]", "lerobot[motorbridge-smart-servo-dep]"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
    "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
@@ -212,8 +194,7 @@ groot = [
 ]
 sarm = ["lerobot[transformers-dep]", "pydantic>=2.0.0,<3.0.0", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
 xvla = ["lerobot[transformers-dep]"]
-eo1 = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]"]
-hilserl = ["lerobot[transformers-dep]", "lerobot[dataset]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]

 # Features
 async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
@@ -231,20 +212,6 @@ aloha = ["lerobot[dataset]", "gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
 pusht = ["lerobot[dataset]", "gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
 libero = ["lerobot[dataset]", "lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
 metaworld = ["lerobot[dataset]", "metaworld==3.0.0", "lerobot[scipy-dep]"]
-# NOTE: vlabench is NOT exposed as a `lerobot` extra. Its only distribution
-# is the OpenMOSS/VLABench GitHub repo (package name `VLABench`, no PyPI
-# release), so any `vlabench>=X` pip spec is unresolvable. Install it
-# manually alongside MuJoCo / dm-control — see docs/source/vlabench.mdx
-# for the recipe.
-# NOTE: robomme is NOT a pyproject extra — mani-skill hard-pins numpy<2
-# which conflicts with lerobot's numpy>=2 base pin, so the two trees can't
-# resolve into a single env. Install it only in the RoboMME Docker image
-# via `uv pip install --override` (see docker/Dockerfile.benchmark.robomme).
-# NOTE: robocasa is NOT exposed as a `lerobot` extra. Its setup.py pins
-# `lerobot==0.3.3` in install_requires, which cyclically shadows our own
-# workspace `lerobot` and makes the graph unsolvable under any resolver
-# (uv, pip). Install it manually alongside robosuite — see
-# docs/source/robocasa.mdx for the recipe.

 # All
 all = [
@@ -267,7 +234,6 @@ all = [
    "lerobot[lekiwi]",
    "lerobot[openarms]",
    "lerobot[reachy2]",
-    "lerobot[rebot]",
    "lerobot[kinematics]",
    "lerobot[intelrealsense]",
    "lerobot[diffusion]",
@@ -312,20 +278,6 @@ lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
 lerobot-rollout="lerobot.scripts.lerobot_rollout:main"

 # ---------------- Tool Configurations ----------------
-
-# cu128 wheels keep broad hardware reach; the driver floor is 570.86.
-# To use a different CUDA variant, reinstall torch with an explicit index, e.g.:
-#   uv pip install --force-reinstall torch torchvision \
-#       --index-url https://download.pytorch.org/whl/cu130
-[[tool.uv.index]]
-name = "pytorch-cu128"
-url = "https://download.pytorch.org/whl/cu128"
-explicit = true
-
-[tool.uv.sources]
-torch = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
-torchvision = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
-
 [tool.setuptools.package-data]
 lerobot = ["envs/*.json"]

--- a/scripts/ci/extract_task_descriptions.py
+++ b/scripts/ci/extract_task_descriptions.py
@@ -31,23 +31,9 @@ from __future__ import annotations

 import argparse
 import json
-import re
 import sys
 from pathlib import Path

-# LIBERO-plus derives task.language by space-joining the perturbation-variant
-# filename (grab_language_from_filename in libero/libero/benchmark/__init__.py),
-# so non-_language_ variants inherit a trailing metadata blob like
-# "view 0 0 100 0 0 initstate 0 noise 45" or "add 16". Strip those tokens so
-# the description matches the base instruction used in the training dataset.
-_LIBERO_PERTURBATION_TAIL_RE = re.compile(
-    r"(?:\s(?:view|initstate|noise|add|tb|table|light|level)(?:\s\d+)+)+$"
-)
-
-
-def _strip_libero_perturbation_tail(instruction: str) -> str:
-    return _LIBERO_PERTURBATION_TAIL_RE.sub("", instruction).strip()
-

 def _libero_descriptions(task_suite: str) -> dict[str, str]:
    from libero.libero import benchmark  # type: ignore[import-untyped]
@@ -61,10 +47,7 @@ def _libero_descriptions(task_suite: str) -> dict[str, str]:
        )
        return {}
    suite = suite_dict[task_suite]()
-    return {
-        f"{task_suite}_{i}": _strip_libero_perturbation_tail(suite.get_task(i).language)
-        for i in range(suite.n_tasks)
-    }
+    return {f"{task_suite}_{i}": suite.get_task(i).language for i in range(suite.n_tasks)}


 def _metaworld_descriptions(task_name: str) -> dict[str, str]:
@@ -74,120 +57,19 @@ def _metaworld_descriptions(task_name: str) -> dict[str, str]:
    return {f"{task_name}_0": label}


-def _robotwin_descriptions(task_names: str) -> dict[str, str]:
-    """Return descriptions for each requested RoboTwin task. Reads
-    `description/task_instruction/<task>.json` from the RoboTwin clone
-    (cwd is /opt/robotwin in CI). Falls back to the task name if missing."""
-    out: dict[str, str] = {}
-    root = Path("description/task_instruction")
-    for name in (t.strip() for t in task_names.split(",") if t.strip()):
-        desc_file = root / f"{name}.json"
-        desc = name.replace("_", " ")
-        if desc_file.is_file():
-            data = json.loads(desc_file.read_text())
-            full = data.get("full_description") or desc
-            # Strip the schema placeholders ({A}, {a}) — keep the sentence readable.
-            desc = full.replace("<", "").replace(">", "")
-        out[f"{name}_0"] = desc
-    return out
-
-
-def _robocasa_descriptions(task_spec: str) -> dict[str, str]:
-    """For each task in the comma-separated list, emit a cleaned-name label.
-
-    RoboCasa episodes carry their language instruction in the env's
-    `ep_meta['lang']`, populated per reset. Pulling it requires spinning
-    up the full kitchen env per task (~seconds each); we use the task
-    name as the key here and let the eval's episode info carry the
-    actual instruction.
-    """
-    out: dict[str, str] = {}
-    for task in (t.strip() for t in task_spec.split(",") if t.strip()):
-        # Split CamelCase into words: "CloseFridge" → "close fridge".
-        label = "".join(f" {c.lower()}" if c.isupper() else c for c in task).strip()
-        out[f"{task}_0"] = label or task
-    return out
-
-
-_ROBOMME_DESCRIPTIONS = {
-    "BinFill": "Fill the target bin with the correct number of cubes",
-    "PickXtimes": "Pick the indicated cube the specified number of times",
-    "SwingXtimes": "Swing the object the specified number of times",
-    "StopCube": "Grasp and stop the moving cube",
-    "VideoUnmask": "Pick the cube shown in the reference video",
-    "VideoUnmaskSwap": "Pick the cube matching the reference video after a swap",
-    "ButtonUnmask": "Press the button indicated by the reference",
-    "ButtonUnmaskSwap": "Press the correct button after objects are swapped",
-    "PickHighlight": "Pick the highlighted cube",
-    "VideoRepick": "Repick the cube shown in the reference video",
-    "VideoPlaceButton": "Place the cube on the button shown in the video",
-    "VideoPlaceOrder": "Place cubes in the order shown in the video",
-    "MoveCube": "Move the cube to the target location",
-    "InsertPeg": "Insert the peg into the target hole",
-    "PatternLock": "Unlock the pattern by pressing buttons in sequence",
-    "RouteStick": "Route the stick through the required waypoints",
-}
-
-
-def _robomme_descriptions(task_names: str, task_ids: list[int] | None = None) -> dict[str, str]:
-    """Return descriptions for each requested RoboMME task. Keys match the
-    video filename pattern `<task>_<task_id>` used by the eval script."""
-    if task_ids is None:
-        task_ids = [0]
-    out: dict[str, str] = {}
-    for name in (t.strip() for t in task_names.split(",") if t.strip()):
-        desc = _ROBOMME_DESCRIPTIONS.get(name, name)
-        for tid in task_ids:
-            out[f"{name}_{tid}"] = desc
-    return out
-
-
-def _vlabench_descriptions(task_spec: str) -> dict[str, str]:
-    """For each task in the comma-separated list, emit a cleaned-name label.
-
-    VLABench tasks carry language instructions on their dm_control task
-    object, but pulling them requires loading the full env per task
-    (~seconds each). The CI smoke-eval already captures the instruction
-    inside its episode info; this mapping is just enough to key
-    `metrics.json` by `<task>_0`.
-    """
-    out: dict[str, str] = {}
-    for task in (t.strip() for t in task_spec.split(",") if t.strip()):
-        out[f"{task}_0"] = task.replace("_", " ").strip()
-    return out
-
-
 def main() -> int:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--env", required=True, help="Environment family (libero, metaworld, ...)")
    parser.add_argument("--task", required=True, help="Task/suite name (e.g. libero_spatial)")
-    parser.add_argument(
-        "--task-ids",
-        type=str,
-        default=None,
-        help="Comma-separated task IDs (e.g. '0,1,2'). Default: [0]",
-    )
    parser.add_argument("--output", required=True, help="Path to write task_descriptions.json")
    args = parser.parse_args()

-    task_ids: list[int] | None = None
-    if args.task_ids:
-        task_ids = [int(x.strip()) for x in args.task_ids.split(",")]
-
    descriptions: dict[str, str] = {}
    try:
-        if args.env == ("libero", "libero_plus"):
+        if args.env == "libero":
            descriptions = _libero_descriptions(args.task)
        elif args.env == "metaworld":
            descriptions = _metaworld_descriptions(args.task)
-        elif args.env == "robotwin":
-            descriptions = _robotwin_descriptions(args.task)
-        elif args.env == "robocasa":
-            descriptions = _robocasa_descriptions(args.task)
-        elif args.env == "robomme":
-            descriptions = _robomme_descriptions(args.task, task_ids=task_ids)
-        elif args.env == "vlabench":
-            descriptions = _vlabench_descriptions(args.task)
        else:
            print(
                f"[extract_task_descriptions] No description extractor for env '{args.env}'.",
--- a/src/lerobot/cameras/opencv/camera_opencv.py
+++ b/src/lerobot/cameras/opencv/camera_opencv.py
@@ -199,13 +199,12 @@ class OpenCVCamera(Camera):
            DeviceNotConnectedError: If the camera is not connected.
        """

+        # Set FOURCC first (if specified) as it can affect available FPS/resolution options
+        if self.config.fourcc is not None:
+            self._validate_fourcc()
        if self.videocapture is None:
            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")

-        set_fourcc_after_size_and_fps = platform.system() == "Windows"
-        if self.config.fourcc is not None and not set_fourcc_after_size_and_fps:
-            self._validate_fourcc()
-
        default_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
        default_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))

@@ -223,11 +222,6 @@ class OpenCVCamera(Camera):
        else:
            self._validate_fps()

-        if self.config.fourcc is not None and set_fourcc_after_size_and_fps:
-            # On Windows with DSHOW, changing the resolution can silently override the FOURCC setting.
-            # Set FOURCC last to make sure the requested pixel format is actually enforced.
-            self._validate_fourcc()
-
    def _validate_fps(self) -> None:
        """Validates and sets the camera's frames per second (FPS)."""

--- a/src/lerobot/cameras/realsense/camera_realsense.py
+++ b/src/lerobot/cameras/realsense/camera_realsense.py
@@ -17,7 +17,6 @@ Provides the RealSenseCamera class for capturing frames from Intel RealSense cam
 """

 import logging
-import sys
 import time
 from threading import Event, Lock, Thread
 from typing import TYPE_CHECKING, Any
@@ -42,7 +41,6 @@ from ..utils import get_cv2_rotation
 from .configuration_realsense import RealSenseCameraConfig

 logger = logging.getLogger(__name__)
-pkg_name = "pyrealsense2-macosx" if sys.platform == "darwin" else "pyrealsense2"


 class RealSenseCamera(Camera):
@@ -116,7 +114,7 @@ class RealSenseCamera(Camera):
        Args:
            config: The configuration settings for the camera.
        """
-        require_package(pkg_name, extra="intelrealsense", import_name="pyrealsense2")
+        require_package("pyrealsense2", extra="intelrealsense")
        super().__init__(config)

        self.config = config
--- a/src/lerobot/common/train_utils.py
+++ b/src/lerobot/common/train_utils.py
@@ -99,7 +99,6 @@ def save_checkpoint(
        optimizer (Optimizer | None, optional): The optimizer to save the state from. Defaults to None.
        scheduler (LRScheduler | None, optional): The scheduler to save the state from. Defaults to None.
        preprocessor: The preprocessor/pipeline to save. Defaults to None.
-        postprocessor: The postprocessor/pipeline to save. Defaults to None.
    """
    pretrained_dir = checkpoint_dir / PRETRAINED_MODEL_DIR
    policy.save_pretrained(pretrained_dir)
--- a/src/lerobot/common/wandb_utils.py
+++ b/src/lerobot/common/wandb_utils.py
@@ -41,12 +41,8 @@ def cfg_to_group(
            return tag
        return tag[:max_tag_length]

-    if cfg.is_reward_model_training:
-        trainable_tag = f"reward_model:{cfg.reward_model.type}"
-    else:
-        trainable_tag = f"policy:{cfg.policy.type}"
    lst = [
-        trainable_tag,
+        f"policy:{cfg.policy.type}",
        f"seed:{cfg.seed}",
    ]
    if cfg.dataset is not None:
--- a/src/lerobot/configs/init.py
+++ b/src/lerobot/configs/init.py
@@ -24,7 +24,6 @@ Import them directly: ``from lerobot.configs.train import TrainPipelineConfig``
 from .dataset import DatasetRecordConfig
 from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
-from .recipe import MessageTurn, TrainingRecipe, load_recipe
 from .types import (
    FeatureType,
    NormalizationMode,
@@ -32,12 +31,6 @@ from .types import (
    PolicyFeature,
    RTCAttentionSchedule,
 )
-from .video import (
-    VALID_VIDEO_CODECS,
-    VIDEO_ENCODER_INFO_KEYS,
-    VideoEncoderConfig,
-    camera_encoder_defaults,
-)

 __all__ = [
    # Types
@@ -50,16 +43,7 @@ __all__ = [
    "DatasetRecordConfig",
    "DatasetConfig",
    "EvalConfig",
-    "MessageTurn",
    "PeftConfig",
    "PreTrainedConfig",
-    "TrainingRecipe",
    "WandBConfig",
-    "load_recipe",
-    "VideoEncoderConfig",
-    # Defaults
-    "camera_encoder_defaults",
-    # Constants
-    "VALID_VIDEO_CODECS",
-    "VIDEO_ENCODER_INFO_KEYS",
 ]
--- a/src/lerobot/configs/dataset.py
+++ b/src/lerobot/configs/dataset.py
@@ -18,8 +18,6 @@ from dataclasses import dataclass, field
 from datetime import datetime
 from pathlib import Path

-from .video import VideoEncoderConfig, camera_encoder_defaults
-

@dataclass
 class DatasetRecordConfig:
@@ -57,9 +55,10 @@ class DatasetRecordConfig:
    # Number of episodes to record before batch encoding videos
    # Set to 1 for immediate encoding (default behavior), or higher for batched encoding
    video_encoding_batch_size: int = 1
-    # Video encoder settings for camera MP4s (codec, quality, GOP, etc.). Tuned via CLI nested keys,
-    # e.g. ``--dataset.camera_encoder.vcodec=h264`` (see ``VideoEncoderConfig``).
-    camera_encoder: VideoEncoderConfig = field(default_factory=camera_encoder_defaults)
+    # Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1', 'auto',
+    # or hardware-specific: 'h264_videotoolbox', 'h264_nvenc', 'h264_vaapi', 'h264_qsv'.
+    # Use 'auto' to auto-detect the best available hardware encoder.
+    vcodec: str = "libsvtav1"
    # Enable streaming video encoding: encode frames in real-time during capture instead
    # of writing PNG images first. Makes save_episode() near-instant. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding
    streaming_encoding: bool = False
@@ -69,13 +68,10 @@ class DatasetRecordConfig:
    # Number of threads per encoder instance. None = auto (codec default).
    # Lower values reduce CPU usage, maps to 'lp' (via svtav1-params) for libsvtav1 and 'threads' for h264/hevc..
    encoder_threads: int | None = None
+    # Rename map for the observation to override the image and state keys
+    rename_map: dict[str, str] = field(default_factory=dict)

-    def stamp_repo_id(self) -> None:
-        """Append a date-time tag to ``repo_id`` so each recording session gets a unique name.
-
-        Must be called explicitly at dataset *creation* time — not on resume,
-        where the existing ``repo_id`` (already stamped) must be preserved.
-        """
+    def __post_init__(self) -> None:
        if self.repo_id:
            timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
            self.repo_id = f"{self.repo_id}_{timestamp}"
--- a/src/lerobot/configs/default.py
+++ b/src/lerobot/configs/default.py
@@ -17,7 +17,7 @@
 from dataclasses import dataclass, field

 from lerobot.transforms import ImageTransformsConfig
-from lerobot.utils.import_utils import get_safe_default_video_backend
+from lerobot.utils.import_utils import get_safe_default_codec


@dataclass
@@ -34,7 +34,7 @@ class DatasetConfig:
    image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)
    revision: str | None = None
    use_imagenet_stats: bool = True
-    video_backend: str = field(default_factory=get_safe_default_video_backend)
+    video_backend: str = field(default_factory=get_safe_default_codec)
    # When True, video frames are returned as uint8 tensors (0-255) instead of float32 (0.0-1.0).
    # This reduces memory and speeds up DataLoader IPC. The training pipeline handles the conversion.
    return_uint8: bool = False
@@ -117,9 +117,3 @@ class PeftConfig:
    # the rank used for the adapter. In general a higher rank means more trainable parameters and closer to full
    # fine-tuning.
    r: int = 16
-
-    # Alpha parameter for LoRA scaling (scaling = lora_alpha / r).
-    # In general, a higher alpha means stronger adaptation signal.
-    # If None, the PEFT library defaults to alpha=8, which may dampen high-rank adapters.
-    # Common values are r (alpha == rank) or 2*r.
-    lora_alpha: int | None = None
--- a/src/lerobot/configs/eval.py
+++ b/src/lerobot/configs/eval.py
@@ -18,8 +18,8 @@ from logging import getLogger
 from pathlib import Path

 from lerobot import envs, policies  # noqa: F401
+from lerobot.configs import parser

-from . import parser
 from .default import EvalConfig
 from .policies import PreTrainedConfig

@@ -46,11 +46,8 @@ class EvalPipelineConfig:
        # HACK: We parse again the cli args here to get the pretrained path if there was one.
        policy_path = parser.get_path_arg("policy")
        if policy_path:
-            yaml_overrides = parser.get_yaml_overrides("policy")
-            cli_overrides = parser.get_cli_overrides("policy") or []
-            self.policy = PreTrainedConfig.from_pretrained(
-                policy_path, cli_overrides=yaml_overrides + cli_overrides
-            )
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
            self.policy.pretrained_path = Path(policy_path)

        else:
--- a/src/lerobot/configs/parser.py
+++ b/src/lerobot/configs/parser.py
@@ -13,10 +13,8 @@
 # limitations under the License.
 import importlib
 import inspect
-import json
 import pkgutil
 import sys
-import tempfile
 from argparse import ArgumentError
 from collections.abc import Callable, Iterable, Sequence
 from functools import wraps
@@ -26,7 +24,6 @@ from types import ModuleType
 from typing import Any, TypeVar, cast

 import draccus
-import yaml  # type: ignore[import-untyped]

 from lerobot.utils.utils import has_method

@@ -35,29 +32,6 @@ F = TypeVar("F", bound=Callable[..., object])
 PATH_KEY = "path"
 PLUGIN_DISCOVERY_SUFFIX = "discover_packages_path"

-# Storage for path args extracted from YAML/JSON config files, so that
-# get_path_arg() can find them even when they weren't passed via CLI.
-_config_path_args: dict[str, str] = {}
-
-# Storage for non-path YAML overrides so validate() can pass them to from_pretrained.
-_config_yaml_overrides: dict[str, list[str]] = {}
-
-
-def _flatten_to_cli_args(d: dict, prefix: str = "") -> list[str]:
-    """Recursively flatten a nested dict to CLI-style args (e.g. {"lr": 1e-4} -> ["--lr=0.0001"])."""
-    args = []
-    for key, value in d.items():
-        if key in (PATH_KEY, draccus.CHOICE_TYPE_KEY):
-            continue
-        full_key = f"{prefix}.{key}" if prefix else key
-        if isinstance(value, bool):
-            value = str(value).lower()
-        if isinstance(value, dict):
-            args.extend(_flatten_to_cli_args(value, full_key))
-        elif value is not None and not isinstance(value, list):
-            args.append(f"--{full_key}={value}")
-    return args
-

 def get_cli_overrides(field_name: str, args: Sequence[str] | None = None) -> list[str] | None:
    """Parses arguments from cli at a given nested attribute level.
@@ -171,14 +145,7 @@ def load_plugin(plugin_path: str) -> None:


 def get_path_arg(field_name: str, args: Sequence[str] | None = None) -> str | None:
-    result = parse_arg(f"{field_name}.{PATH_KEY}", args)
-    if result is None:
-        result = _config_path_args.get(field_name)
-    return result
-
-
-def get_yaml_overrides(field_name: str) -> list[str]:
-    return _config_yaml_overrides.get(field_name, [])
+    return parse_arg(f"{field_name}.{PATH_KEY}", args)


 def get_type_arg(field_name: str, args: Sequence[str] | None = None) -> str | None:
@@ -225,52 +192,6 @@ def filter_path_args(fields_to_filter: str | list[str], args: Sequence[str] | No
    return filtered_args


-def extract_path_fields_from_config(config_path: str, path_fields: list[str]) -> str:
-    """Extract `path` fields from a YAML/JSON config before draccus processes it.
-
-    When a user specifies e.g. ``policy.path: lerobot/smolvla_base`` in a YAML config,
-    draccus will fail because ``path`` is not a valid field on policy config classes.
-    This function extracts those path values, stores them in ``_config_path_args`` for
-    later retrieval by ``get_path_arg()``, and returns a cleaned temp config file path.
-    """
-    config_file = Path(config_path)
-    suffix = config_file.suffix.lower()
-
-    if suffix in (".yaml", ".yml"):
-        with open(config_file) as f:
-            config_data = yaml.safe_load(f)
-    elif suffix == ".json":
-        with open(config_file) as f:
-            config_data = json.load(f)
-    else:
-        return config_path
-
-    if not isinstance(config_data, dict):
-        return config_path
-
-    modified = False
-    for field in path_fields:
-        if field in config_data and isinstance(config_data[field], dict) and PATH_KEY in config_data[field]:
-            _config_path_args[field] = str(config_data[field].pop(PATH_KEY))
-            remaining = config_data[field]
-            if remaining:
-                _config_yaml_overrides[field] = _flatten_to_cli_args(remaining)
-            else:
-                del config_data[field]
-            modified = True
-
-    if not modified:
-        return config_path
-
-    # Write cleaned config to a temp file
-    with tempfile.NamedTemporaryFile(mode="w", suffix=suffix, delete=False) as tmp:
-        if suffix in (".yaml", ".yml"):
-            yaml.dump(config_data, tmp, default_flow_style=False)
-        else:
-            json.dump(config_data, tmp, indent=2)
-    return tmp.name
-
-
 def wrap(config_path: Path | None = None) -> Callable[[F], F]:
    """
    HACK: Similar to draccus.wrap but does three additional things:
@@ -304,9 +225,6 @@ def wrap(config_path: Path | None = None) -> Callable[[F], F]:
                if has_method(argtype, "__get_path_fields__"):
                    path_fields = argtype.__get_path_fields__()
                    cli_args = filter_path_args(path_fields, cli_args)
-                    # Also extract path fields from the YAML/JSON config file
-                    if config_path_cli:
-                        config_path_cli = extract_path_fields_from_config(config_path_cli, path_fields)
                if has_method(argtype, "from_pretrained") and config_path_cli:
                    cli_args = filter_arg("config_path", cli_args)
                    cfg = argtype.from_pretrained(config_path_cli, cli_args=cli_args)
--- a/src/lerobot/configs/recipe.py
+++ b/src/lerobot/configs/recipe.py
@@ -1,206 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import annotations
-
-import re
-from dataclasses import dataclass
-from pathlib import Path
-from typing import Any, Literal, get_args
-
-MessageRole = Literal["user", "assistant", "system", "tool"]
-MessageStream = Literal["high_level", "low_level"]
-
-DEFAULT_BINDINGS = {
-    "subtask": "active_at(t, style=subtask)",
-    "memory": "active_at(t, style=memory)",
-    "plan": "active_at(t, style=plan)",
-    "speech": "emitted_at(t, role=assistant, tool_name=say)",
-    "interjection": "emitted_at(t, style=interjection)",
-    "vqa": "emitted_at(t, style=vqa, role=assistant)",
-    "vqa_query": "emitted_at(t, style=vqa, role=user)",
-}
-
-PLACEHOLDER_RE = re.compile(r"\$\{([A-Za-z_][A-Za-z0-9_]*)\}")
-"""``${name}`` placeholder pattern used by both recipe binding-reference
-discovery (here) and rendered-message substitution (in ``language_render``)."""
-
-_VALID_ROLES = frozenset(get_args(MessageRole))
-_VALID_STREAMS = frozenset(get_args(MessageStream))
-
-
-@dataclass
-class MessageTurn:
-    """A single chat-style turn in a recipe template.
-
-    ``content`` may be a plain string, a list of HF-style multimodal blocks, or
-    ``None`` when ``tool_calls_from`` supplies tool-call payloads instead.
-    ``stream`` tags the turn for downstream filtering, ``target`` flags it as a
-    training target, and ``if_present`` skips the turn when the named binding
-    resolves to ``None``.
-    """
-
-    role: MessageRole
-    content: str | list[dict[str, Any]] | None = None
-    stream: MessageStream | None = None
-    target: bool = False
-    if_present: str | None = None
-    tool_calls_from: str | None = None
-
-    def __post_init__(self) -> None:
-        """Validate role, stream, and content after dataclass construction."""
-        if self.role not in _VALID_ROLES:
-            raise ValueError(f"Unsupported message role: {self.role!r}")
-        # ``stream`` is typed Optional only so the dataclass can keep its
-        # field ordering, but recipes must always tag every turn with a
-        # stream — the renderer's ``_validate_rendered`` would reject
-        # ``None`` later on. Fail at construction so the bad recipe is
-        # caught at YAML load time rather than at the first sample.
-        if self.stream is None:
-            raise ValueError(
-                f"MessageTurn(role={self.role!r}) is missing a stream — "
-                f"every turn must declare one of {sorted(_VALID_STREAMS)}."
-            )
-        if self.stream not in _VALID_STREAMS:
-            raise ValueError(f"Unsupported message stream: {self.stream!r}")
-        if self.content is None and self.tool_calls_from is None:
-            raise ValueError("MessageTurn.content is required unless tool_calls_from is set.")
-        if self.content is not None and not isinstance(self.content, (str, list)):
-            raise TypeError("MessageTurn.content must be a string, a list of HF-style blocks, or None.")
-        if isinstance(self.content, list):
-            for block in self.content:
-                if not isinstance(block, dict) or "type" not in block:
-                    raise ValueError(
-                        "Multimodal content blocks must be HF-style dictionaries with a type key."
-                    )
-
-    @classmethod
-    def from_dict(cls, data: dict[str, Any]) -> MessageTurn:
-        """Construct a :class:`MessageTurn` from a plain dictionary."""
-        return cls(**data)
-
-
-@dataclass
-class TrainingRecipe:
-    """A recipe describing how to render training samples from language rows.
-
-    A recipe is either a *message recipe* (``messages`` plus optional
-    ``bindings``) or a *blend recipe* (``blend`` mapping names to weighted
-    sub-recipes). ``weight`` is only meaningful inside a blend.
-    """
-
-    messages: list[MessageTurn] | None = None
-    bindings: dict[str, str] | None = None
-    blend: dict[str, TrainingRecipe] | None = None
-    weight: float | None = None
-
-    def __post_init__(self) -> None:
-        """Validate that exactly one of ``messages`` or ``blend`` is set."""
-        if self.messages is not None and self.blend is not None:
-            raise ValueError("TrainingRecipe must set only one of messages or blend.")
-        if self.messages is None and self.blend is None:
-            raise ValueError("TrainingRecipe must set one of messages or blend.")
-
-        if self.messages is not None:
-            self._validate_message_recipe()
-        if self.blend is not None:
-            self._validate_blend_recipe()
-
-    @classmethod
-    def from_dict(cls, data: dict[str, Any]) -> TrainingRecipe:
-        """Construct a :class:`TrainingRecipe` from a nested dictionary."""
-        data = dict(data)
-        if data.get("messages") is not None:
-            data["messages"] = [
-                turn if isinstance(turn, MessageTurn) else MessageTurn.from_dict(turn)
-                for turn in data["messages"]
-            ]
-        if data.get("blend") is not None:
-            data["blend"] = {
-                name: recipe if isinstance(recipe, TrainingRecipe) else cls.from_dict(recipe)
-                for name, recipe in data["blend"].items()
-            }
-        return cls(**data)
-
-    @classmethod
-    def from_yaml(cls, path: str | Path) -> TrainingRecipe:
-        """Load a :class:`TrainingRecipe` from a YAML file at ``path``."""
-        import yaml  # type: ignore[import-untyped]
-
-        with open(path) as f:
-            data = yaml.safe_load(f)
-        if not isinstance(data, dict):
-            raise ValueError(f"Recipe YAML must contain a mapping at the top level: {path}")
-        return cls.from_dict(data)
-
-    def _validate_message_recipe(self) -> None:
-        """Ensure every templated binding is known and at least one turn is a target."""
-        assert self.messages is not None
-        known_bindings = set(DEFAULT_BINDINGS) | set(self.bindings or {}) | {"task"}
-
-        for turn in self.messages:
-            missing = self._referenced_bindings(turn) - known_bindings
-            if missing:
-                raise ValueError(f"MessageTurn references unknown binding(s): {sorted(missing)}")
-
-        if not any(turn.target for turn in self.messages):
-            raise ValueError("Message recipes must contain at least one target turn.")
-
-    def _validate_blend_recipe(self) -> None:
-        """Ensure each blend component is a non-empty, weighted message recipe."""
-        assert self.blend is not None
-        if not self.blend:
-            raise ValueError("Blend recipes must contain at least one component.")
-
-        for name, recipe in self.blend.items():
-            if recipe.blend is not None:
-                raise ValueError(f"Blend component {name!r} cannot itself define a blend.")
-            if recipe.messages is None:
-                raise ValueError(f"Blend component {name!r} must define messages.")
-            if recipe.weight is None:
-                raise ValueError(f"Blend component {name!r} must define weight.")
-            if recipe.weight <= 0:
-                raise ValueError(f"Blend component {name!r} must have a positive weight.")
-
-    def _referenced_bindings(self, turn: MessageTurn) -> set[str]:
-        """Return the binding names that ``turn`` references via placeholders or attributes."""
-        names: set[str] = set()
-        if turn.if_present is not None:
-            names.add(turn.if_present)
-        if turn.tool_calls_from is not None:
-            names.add(turn.tool_calls_from)
-        names.update(_placeholders_in_content(turn.content))
-        return names
-
-
-def _placeholders_in_content(content: str | list[dict[str, Any]] | None) -> set[str]:
-    """Return the set of ``${name}`` placeholders found anywhere in ``content``."""
-    if content is None:
-        return set()
-    if isinstance(content, str):
-        return set(PLACEHOLDER_RE.findall(content))
-
-    names: set[str] = set()
-    for block in content:
-        for value in block.values():
-            if isinstance(value, str):
-                names.update(PLACEHOLDER_RE.findall(value))
-    return names
-
-
-def load_recipe(path: str | Path) -> TrainingRecipe:
-    """Load a :class:`TrainingRecipe` from a YAML file at ``path``."""
-    return TrainingRecipe.from_yaml(path)
--- a/src/lerobot/configs/rewards.py
+++ b/src/lerobot/configs/rewards.py
@@ -1,164 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import abc
-import builtins
-import json
-import logging
-import os
-import tempfile
-from dataclasses import dataclass, field
-from pathlib import Path
-from typing import Any, TypeVar
-
-import draccus
-from huggingface_hub import hf_hub_download
-from huggingface_hub.constants import CONFIG_NAME
-from huggingface_hub.errors import HfHubHTTPError
-
-from lerobot.optim.optimizers import OptimizerConfig
-from lerobot.optim.schedulers import LRSchedulerConfig
-from lerobot.utils.device_utils import auto_select_torch_device, is_torch_device_available
-from lerobot.utils.hub import HubMixin
-
-from .types import PolicyFeature
-
-T = TypeVar("T", bound="RewardModelConfig")
-logger = logging.getLogger(__name__)
-
-
-@dataclass
-class RewardModelConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
-    """Base configuration for reward models.
-
-    Args:
-    input_features: A dictionary defining the PolicyFeature of the input data for the reward. The key represents
-        the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-    output_features: A dictionary defining the PolicyFeature of the output data for the reward. The key represents
-        the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-    """
-
-    # Reuses PolicyFeature
-    input_features: dict[str, PolicyFeature] = field(default_factory=dict)
-    output_features: dict[str, PolicyFeature] = field(default_factory=dict)
-
-    device: str | None = None
-
-    pretrained_path: str | None = None
-
-    push_to_hub: bool = False
-    repo_id: str | None = None
-
-    # Hub metadata
-    license: str | None = None
-    tags: list[str] | None = None
-    private: bool | None = None
-
-    def __post_init__(self) -> None:
-        if not self.device or not is_torch_device_available(self.device):
-            auto_device = auto_select_torch_device()
-            logger.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
-            self.device = auto_device.type
-
-    @property
-    def type(self) -> str:
-        choice_name = self.get_choice_name(self.__class__)
-        if not isinstance(choice_name, str):
-            raise TypeError(f"Expected string from get_choice_name, got {type(choice_name)}")
-        return choice_name
-
-    @property
-    def observation_delta_indices(self) -> list | None:  # type: ignore[type-arg]
-        return None
-
-    @property
-    def action_delta_indices(self) -> list | None:  # type: ignore[type-arg]
-        return None
-
-    @property
-    def reward_delta_indices(self) -> list | None:  # type: ignore[type-arg]
-        return None
-
-    def get_optimizer_preset(self) -> OptimizerConfig | None:
-        """Default optimizer for this reward model, or ``None`` for zero-shot models."""
-        return None
-
-    def get_scheduler_preset(self) -> LRSchedulerConfig | None:
-        return None
-
-    def validate_features(self) -> None:
-        pass
-
-    def _save_pretrained(self, save_directory: Path) -> None:
-        with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
-            draccus.dump(self, f, indent=4)
-
-    @classmethod
-    def from_pretrained(
-        cls: builtins.type[T],
-        pretrained_name_or_path: str | Path,
-        *,
-        force_download: bool = False,
-        resume_download: bool | None = None,
-        proxies: dict[Any, Any] | None = None,
-        token: str | bool | None = None,
-        cache_dir: str | Path | None = None,
-        local_files_only: bool = False,
-        revision: str | None = None,
-        **reward_kwargs: Any,
-    ) -> T:
-        model_id = str(pretrained_name_or_path)
-        config_file: str | None = None
-        if Path(model_id).is_dir():
-            if CONFIG_NAME in os.listdir(model_id):
-                config_file = os.path.join(model_id, CONFIG_NAME)
-            else:
-                logger.error(f"{CONFIG_NAME} not found in {Path(model_id).resolve()}")
-        else:
-            try:
-                config_file = hf_hub_download(
-                    repo_id=model_id,
-                    filename=CONFIG_NAME,
-                    revision=revision,
-                    cache_dir=cache_dir,
-                    force_download=force_download,
-                    proxies=proxies,
-                    resume_download=resume_download,
-                    token=token,
-                    local_files_only=local_files_only,
-                )
-            except HfHubHTTPError as e:
-                raise FileNotFoundError(
-                    f"{CONFIG_NAME} not found on the HuggingFace Hub in {model_id}"
-                ) from e
-
-        if config_file is None:
-            raise FileNotFoundError(f"{CONFIG_NAME} not found in {model_id}")
-
-        # HACK: Parse the original config to get the config subclass, so that we can
-        # apply cli overrides.
-        with draccus.config_type("json"):
-            orig_config = draccus.parse(cls, config_file, args=[])
-
-        with open(config_file) as f:
-            config = json.load(f)
-
-        config.pop("type", None)
-        with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
-            json.dump(config, f)
-            config_file = f.name
-
-        cli_overrides = reward_kwargs.pop("cli_overrides", [])
-        with draccus.config_type("json"):
-            return draccus.parse(orig_config.__class__, config_file, args=cli_overrides)
--- a/src/lerobot/configs/train.py
+++ b/src/lerobot/configs/train.py
@@ -13,9 +13,7 @@
 # limitations under the License.
 import builtins
 import datetime as dt
-import json
 import os
-import tempfile
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any
@@ -25,60 +23,21 @@ from huggingface_hub import hf_hub_download
 from huggingface_hub.errors import HfHubHTTPError

 from lerobot import envs
+from lerobot.configs import parser
 from lerobot.optim import LRSchedulerConfig, OptimizerConfig
 from lerobot.utils.hub import HubMixin
-from lerobot.utils.sample_weighting import SampleWeightingConfig

-from . import parser
 from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
-from .rewards import RewardModelConfig

 TRAIN_CONFIG_NAME = "train_config.json"


-def _migrate_legacy_rabc_fields(config: dict[str, Any]) -> dict[str, Any] | None:
-    """Return migrated payload for legacy RA-BC fields, or None when no migration is needed."""
-    legacy_fields = (
-        "use_rabc",
-        "rabc_progress_path",
-        "rabc_kappa",
-        "rabc_epsilon",
-        "rabc_head_mode",
-    )
-    if not any(key in config for key in legacy_fields):
-        return None
-
-    migrated_config = dict(config)
-    use_rabc = bool(migrated_config.pop("use_rabc", False))
-    rabc_progress_path = migrated_config.pop("rabc_progress_path", None)
-    rabc_kappa = migrated_config.pop("rabc_kappa", None)
-    rabc_epsilon = migrated_config.pop("rabc_epsilon", None)
-    rabc_head_mode = migrated_config.pop("rabc_head_mode", None)
-
-    # New configs may already define sample_weighting explicitly. In that case,
-    # legacy fields are ignored after being stripped from the payload.
-    if migrated_config.get("sample_weighting") is None and use_rabc:
-        sample_weighting: dict[str, Any] = {"type": "rabc"}
-        if rabc_progress_path is not None:
-            sample_weighting["progress_path"] = rabc_progress_path
-        if rabc_kappa is not None:
-            sample_weighting["kappa"] = rabc_kappa
-        if rabc_epsilon is not None:
-            sample_weighting["epsilon"] = rabc_epsilon
-        if rabc_head_mode is not None:
-            sample_weighting["head_mode"] = rabc_head_mode
-        migrated_config["sample_weighting"] = sample_weighting
-
-    return migrated_config
-
-
@dataclass
 class TrainPipelineConfig(HubMixin):
    dataset: DatasetConfig
    env: envs.EnvConfig | None = None
    policy: PreTrainedConfig | None = None
-    reward_model: RewardModelConfig | None = None
    # Set `dir` to where you would like to save all of the run outputs. If you run another training session
    # with the same value for `dir` its contents will be overwritten unless you set `resume` to true.
    output_dir: Path | None = None
@@ -113,44 +72,27 @@ class TrainPipelineConfig(HubMixin):
    wandb: WandBConfig = field(default_factory=WandBConfig)
    peft: PeftConfig | None = None

-    # Sample weighting configuration (e.g., for RA-BC training)
-    sample_weighting: SampleWeightingConfig | None = None
+    # RA-BC (Reward-Aligned Behavior Cloning) parameters
+    use_rabc: bool = False  # Enable reward-weighted training
+    rabc_progress_path: str | None = None  # Path to precomputed SARM progress parquet file
+    rabc_kappa: float = 0.01  # Hard threshold for high-quality samples
+    rabc_epsilon: float = 1e-6  # Small constant for numerical stability
+    rabc_head_mode: str | None = "sparse"  # For dual-head models: "sparse" or "dense"

    # Rename map for the observation to override the image and state keys
    rename_map: dict[str, str] = field(default_factory=dict)
    checkpoint_path: Path | None = field(init=False, default=None)

-    @property
-    def is_reward_model_training(self) -> bool:
-        """True when the config targets a reward model rather than a policy."""
-        return self.reward_model is not None
-
-    @property
-    def trainable_config(self) -> PreTrainedConfig | RewardModelConfig:
-        """Return whichever config (policy or reward_model) is active."""
-        if self.is_reward_model_training:
-            return self.reward_model  # type: ignore[return-value]
-        return self.policy  # type: ignore[return-value]
-
    def validate(self) -> None:
        # HACK: We parse again the cli args here to get the pretrained paths if there was some.
        policy_path = parser.get_path_arg("policy")
-        reward_model_path = parser.get_path_arg("reward_model")
-
-        if reward_model_path:
-            cli_overrides = parser.get_cli_overrides("reward_model")
-            self.reward_model = RewardModelConfig.from_pretrained(
-                reward_model_path, cli_overrides=cli_overrides
-            )
-            self.reward_model.pretrained_path = str(Path(reward_model_path))
-        elif policy_path:
-            yaml_overrides = parser.get_yaml_overrides("policy")
-            cli_overrides = parser.get_cli_overrides("policy") or []
-            self.policy = PreTrainedConfig.from_pretrained(
-                policy_path, cli_overrides=yaml_overrides + cli_overrides
-            )
+        if policy_path:
+            # Only load the policy config
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
            self.policy.pretrained_path = Path(policy_path)
        elif self.resume:
+            # The entire train config is already loaded, we just need to get the checkpoint dir
            config_path = parser.parse_arg("config_path")
            if not config_path:
                raise ValueError(
@@ -166,22 +108,18 @@ class TrainPipelineConfig(HubMixin):
            policy_dir = Path(config_path).parent
            if self.policy is not None:
                self.policy.pretrained_path = policy_dir
-            if self.reward_model is not None:
-                self.reward_model.pretrained_path = str(policy_dir)
            self.checkpoint_path = policy_dir.parent

-        if self.policy is None and self.reward_model is None:
+        if self.policy is None:
            raise ValueError(
-                "Neither policy nor reward_model is configured. "
-                "Please specify one with `--policy.path` or `--reward_model.path`."
+                "Policy is not configured. Please specify a pretrained policy with `--policy.path`."
            )

-        active_cfg = self.trainable_config
        if not self.job_name:
            if self.env is None:
-                self.job_name = f"{active_cfg.type}"
+                self.job_name = f"{self.policy.type}"
            else:
-                self.job_name = f"{self.env.type}_{active_cfg.type}"
+                self.job_name = f"{self.env.type}_{self.policy.type}"

        if not self.resume and isinstance(self.output_dir, Path) and self.output_dir.is_dir():
            raise FileExistsError(
@@ -199,16 +137,26 @@ class TrainPipelineConfig(HubMixin):
        if not self.use_policy_training_preset and (self.optimizer is None or self.scheduler is None):
            raise ValueError("Optimizer and Scheduler must be set when the policy presets are not used.")
        elif self.use_policy_training_preset and not self.resume:
-            self.optimizer = active_cfg.get_optimizer_preset()
-            self.scheduler = active_cfg.get_scheduler_preset()
+            self.optimizer = self.policy.get_optimizer_preset()
+            self.scheduler = self.policy.get_scheduler_preset()

-        if hasattr(active_cfg, "push_to_hub") and active_cfg.push_to_hub and not active_cfg.repo_id:
-            raise ValueError("'repo_id' argument missing. Please specify it to push the model to the hub.")
+        if self.policy.push_to_hub and not self.policy.repo_id:
+            raise ValueError(
+                "'policy.repo_id' argument missing. Please specify it to push the model to the hub."
+            )
+
+        if self.use_rabc and not self.rabc_progress_path:
+            # Auto-detect from dataset path
+            repo_id = self.dataset.repo_id
+            if self.dataset.root:
+                self.rabc_progress_path = str(Path(self.dataset.root) / "sarm_progress.parquet")
+            else:
+                self.rabc_progress_path = f"hf://datasets/{repo_id}/sarm_progress.parquet"

    @classmethod
    def __get_path_fields__(cls) -> list[str]:
-        """Keys for draccus pretrained-path loading."""
-        return ["policy", "reward_model"]
+        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
+        return ["policy"]

    def to_dict(self) -> dict[str, Any]:
        return draccus.encode(self)  # type: ignore[no-any-return]  # because of the third-party library draccus uses Any as the return type
@@ -259,16 +207,12 @@ class TrainPipelineConfig(HubMixin):
                ) from e

        cli_args = kwargs.pop("cli_args", [])
-        # Legacy RA-BC migration only applies to framework-saved checkpoints (always JSON).
-        # Hand-written YAML/TOML configs are expected to use the current sample_weighting schema.
-        if config_file is not None and config_file.endswith(".json"):
-            with open(config_file) as f:
-                config = json.load(f)
-            migrated_config = _migrate_legacy_rabc_fields(config)
-            if migrated_config is not None:
-                with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
-                    json.dump(migrated_config, f)
-                    config_file = f.name
-
        with draccus.config_type("json"):
            return draccus.parse(cls, config_file, args=cli_args)
+
+
+@dataclass(kw_only=True)
+class TrainRLServerPipelineConfig(TrainPipelineConfig):
+    # NOTE: In RL, we don't need an offline dataset
+    # TODO: Make `TrainPipelineConfig.dataset` optional
+    dataset: DatasetConfig | None = None  # type: ignore[assignment] # because the parent class has made it's type non-optional
--- a/src/lerobot/configs/video.py
+++ b/src/lerobot/configs/video.py
@@ -1,235 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# Note: We subclass str so that serialization is straightforward
-# https://stackoverflow.com/questions/24481852/serialising-an-enum-member-to-json
-
-"""Video encoder configurations."""
-
-from __future__ import annotations
-
-import logging
-from dataclasses import dataclass, field
-from typing import Any
-
-from lerobot.utils.import_utils import require_package
-
-logger = logging.getLogger(__name__)
-
-# List of hardware encoders to probe for auto-selection. Availability depends on the platform and the chosen video backend.
-# Determines the order of preference for auto-selection when vcodec="auto" is used.
-HW_VIDEO_CODECS = [
-    "h264_videotoolbox",  # macOS
-    "hevc_videotoolbox",  # macOS
-    "h264_nvenc",  # NVIDIA GPU
-    "hevc_nvenc",  # NVIDIA GPU
-    "h264_vaapi",  # Linux Intel/AMD
-    "h264_qsv",  # Intel Quick Sync
-]
-VALID_VIDEO_CODECS: frozenset[str] = frozenset({"h264", "hevc", "libsvtav1", "auto", *HW_VIDEO_CODECS})
-# Aliases for legacy video codec names.
-VIDEO_CODECS_ALIASES: dict[str, str] = {"av1": "libsvtav1"}
-
-
-LIBSVTAV1_DEFAULT_PRESET: int = 12
-
-# Keys persisted under ``features[*]["info"]`` as ``video.<name>`` (from :class:`VideoEncoderConfig`).
-# ``vcodec``` and ``pix_fmt`` are derived from the video stream directly.
-VIDEO_ENCODER_INFO_FIELD_NAMES: frozenset[str] = frozenset(
-    {"g", "crf", "preset", "fast_decode", "extra_options", "video_backend"}
-)
-VIDEO_ENCODER_INFO_KEYS: frozenset[str] = frozenset(
-    f"video.{name}" for name in VIDEO_ENCODER_INFO_FIELD_NAMES
-)
-
-
-@dataclass
-class VideoEncoderConfig:
-    """Video encoder configuration.
-
-    Attributes:
-        vcodec: Video encoder name. ``"auto"`` is resolved during
-            construction (HW encoder if available, else ``libsvtav1``).
-        pix_fmt: Pixel format (e.g. ``"yuv420p"``).
-        g: GOP size (keyframe interval).
-        crf: Quality level — mapped to the native quality parameter of the
-            codec (``crf`` for software, ``qp`` for NVENC/VAAPI,
-            ``q:v`` for VideoToolbox, ``global_quality`` for QSV).
-        preset: Speed/quality preset. Accepted type is per-codec.
-        fast_decode: Fast-decode tuning. For ``libsvtav1`` this is a level (0-2)
-            embedded in ``svtav1-params``. For ``h264`` and ``hevc`` non-zero values
-            set ``tune=fastdecode``. Ignored for other codecs.
-        video_backend: Python to be used for encoding. Only ``"pyav"``
-            is currently supported.
-        extra_options: Free-form dictionary of additional video encoder options
-            (e.g. ``{"tune": "film", "profile:v": "high", "bf": 2}``).
-    """
-
-    vcodec: str = "libsvtav1"  # TODO(CarolinePascal): rename to codec ?
-    pix_fmt: str = "yuv420p"
-    g: int | None = 2
-    crf: int | float | None = 30
-    preset: int | str | None = None
-    fast_decode: int = 0
-    # TODO(CarolinePascal): add torchcodec support + find a way to unify the
-    # two backends (encoding and decoding).
-    video_backend: str = "pyav"
-    extra_options: dict[str, Any] = field(default_factory=dict)
-
-    def __post_init__(self) -> None:
-        self.resolve_vcodec()
-        # Empty-constructor ergonomics: ``VideoEncoderConfig()`` must "just work".
-        if self.preset is None and self.vcodec == "libsvtav1":
-            self.preset = LIBSVTAV1_DEFAULT_PRESET
-        self.validate()
-
-    @classmethod
-    def from_video_info(cls, video_info: dict | None) -> VideoEncoderConfig:
-        """Reconstruct a :class:`VideoEncoderConfig` from a video feature's ``info`` block.
-        Missing or ``None`` values fall back to the class defaults.
-        """
-        video_info = video_info or {}
-        kwargs: dict[str, Any] = {}
-
-        for src_key, dst_field in (("video.codec", "vcodec"), ("video.pix_fmt", "pix_fmt")):
-            value = video_info.get(src_key)
-            if value is not None:
-                kwargs[dst_field] = value
-
-        for field_name in VIDEO_ENCODER_INFO_FIELD_NAMES:
-            value = video_info.get(f"video.{field_name}")
-            if value is None:
-                continue
-            # Persisted as ``{}`` after merges with disagreeing sources — treat as default.
-            if field_name == "extra_options" and not value:
-                continue
-            kwargs[field_name] = value
-
-        return cls(**kwargs)
-
-    def detect_available_encoders(self, encoders: list[str] | str) -> list[str]:
-        """Return the subset of available encoders based on the specified video backend.
-
-        Args:
-            encoders: List of encoder names to detect. If a string, it is converted to a list.
-        Returns:
-            List of available encoder names. If the video backend is not "pyav", returns an empty list.
-        """
-        if self.video_backend == "pyav":
-            require_package("av", extra="dataset")
-            from lerobot.datasets import detect_available_encoders_pyav
-
-            return detect_available_encoders_pyav(encoders)
-        return []
-
-    def validate(self) -> None:
-        """Validate the video encoder configuration."""
-        if self.video_backend == "pyav":
-            require_package("av", extra="dataset")
-            from lerobot.datasets import check_video_encoder_parameters_pyav
-
-            check_video_encoder_parameters_pyav(self.vcodec, self.pix_fmt, self.get_codec_options())
-
-    def resolve_vcodec(self) -> None:
-        """Check ``vcodec`` and, when it is ``"auto"``, pick a concrete encoder.
-
-        For ``"auto"``, the first hardware encoder in the preference list that is available is chosen; if none are available, ``libsvtav1`` is used. If the
-        resolved codec (explicit or after auto-selection) is not available, raises ``ValueError``.
-
-        Stream-derived canonical codec names listed in :data:`VIDEO_CODECS_ALIASES` are
-        rewritten to their corresponding encoder name (e.g. ``"av1"`` → ``"libsvtav1"``).
-        """
-        self.vcodec = VIDEO_CODECS_ALIASES.get(self.vcodec, self.vcodec)
-        if self.vcodec not in VALID_VIDEO_CODECS:
-            raise ValueError(f"Invalid vcodec '{self.vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
-        if self.vcodec == "auto":
-            available = self.detect_available_encoders(HW_VIDEO_CODECS)
-            for encoder in HW_VIDEO_CODECS:
-                if encoder in available:
-                    logger.info(f"Auto-selected video codec: {encoder}")
-                    self.vcodec = encoder
-                    return
-            logger.warning("No hardware encoder available, falling back to software encoder 'libsvtav1'")
-            self.vcodec = "libsvtav1"
-
-        if self.detect_available_encoders(self.vcodec):
-            logger.info(f"Using video codec: {self.vcodec}")
-            return
-        raise ValueError(f"Unsupported video codec: {self.vcodec} with video backend {self.video_backend}")
-
-    def get_codec_options(
-        self, encoder_threads: int | None = None, as_strings: bool = False
-    ) -> dict[str, Any]:
-        """Translate the tuning fields to codec-specific options.
-
-        ``VideoEncoderConfig.extra_options`` are merged last but never override a structured field.
-
-        Args:
-            encoder_threads: Number of encoder threads set globally for all VideoEncoderConfigs.
-                For libsvtav1, this is mapped to ``lp`` via ``svtav1-params``.
-                For h264/hevc, this is mapped to ``threads``.
-                Hardware encoders ignore this parameter.
-            as_strings: If ``True``, casts values to strings.
-        """
-        opts: dict[str, Any] = {}
-
-        def set_if(key: str, value: Any) -> None:
-            if value is not None:
-                opts[key] = value if not as_strings else str(value)
-
-        # GOP size is not a codec-specific option, so it is always set.
-        set_if("g", self.g)
-
-        if self.vcodec == "libsvtav1":
-            set_if("crf", self.crf)
-            set_if("preset", self.preset)
-            svtav1_parts: list[str] = []
-            if self.fast_decode is not None:
-                svtav1_parts.append(f"fast-decode={max(0, min(2, self.fast_decode))}")
-            if encoder_threads is not None:
-                svtav1_parts.append(f"lp={encoder_threads}")
-            if svtav1_parts:
-                opts["svtav1-params"] = ":".join(svtav1_parts)
-        elif self.vcodec in ("h264", "hevc"):
-            set_if("crf", self.crf)
-            set_if("preset", self.preset)
-            if self.fast_decode:
-                opts["tune"] = "fastdecode"
-            set_if("threads", encoder_threads)
-        elif self.vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
-            if self.crf is not None:
-                opts["q:v"] = max(1, min(100, 100 - self.crf * 2))
-        elif self.vcodec in ("h264_nvenc", "hevc_nvenc"):
-            opts["rc"] = 0
-            set_if("qp", self.crf)
-            set_if("preset", self.preset)
-        elif self.vcodec == "h264_vaapi":
-            set_if("qp", self.crf)
-        elif self.vcodec == "h264_qsv":
-            set_if("global_quality", self.crf)
-            set_if("preset", self.preset)
-        else:
-            set_if("crf", self.crf)
-            set_if("preset", self.preset)
-
-        # Extra options are merged last but never override structured fields (values are kept as given).
-        for k, v in self.extra_options.items():
-            if k not in opts:
-                set_if(k, v)
-
-        return opts
-
-
-def camera_encoder_defaults() -> VideoEncoderConfig:
-    """Return a :class:`VideoEncoderConfig` with RGB-camera defaults."""
-    return VideoEncoderConfig()
--- a/src/lerobot/datasets/init.py
+++ b/src/lerobot/datasets/init.py
@@ -31,25 +31,15 @@ from .dataset_tools import (
    modify_features,
    modify_tasks,
    recompute_stats,
-    reencode_dataset,
    remove_feature,
    split_dataset,
 )
 from .factory import make_dataset, resolve_delta_timestamps
 from .image_writer import safe_stop_image_writer
 from .io_utils import load_episodes, write_stats
-from .language import (
-    EVENT_ONLY_STYLES,
-    LANGUAGE_EVENTS,
-    LANGUAGE_PERSISTENT,
-    PERSISTENT_STYLES,
-    STYLE_REGISTRY,
-    column_for_style,
-)
 from .lerobot_dataset import LeRobotDataset
 from .multi_dataset import MultiLeRobotDataset
 from .pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
-from .pyav_utils import check_video_encoder_parameters_pyav, detect_available_encoders_pyav
 from .sampler import EpisodeAwareSampler
 from .streaming_dataset import StreamingLeRobotDataset
 from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card
@@ -63,19 +53,12 @@ __all__ = [
    "CODEBASE_VERSION",
    "DEFAULT_EPISODES_PATH",
    "DEFAULT_QUANTILES",
-    "EVENT_ONLY_STYLES",
    "EpisodeAwareSampler",
-    "LANGUAGE_EVENTS",
-    "LANGUAGE_PERSISTENT",
    "LeRobotDataset",
    "LeRobotDatasetMetadata",
    "MultiLeRobotDataset",
-    "PERSISTENT_STYLES",
-    "STYLE_REGISTRY",
    "StreamingLeRobotDataset",
    "VideoEncodingManager",
-    "check_video_encoder_parameters_pyav",
-    "detect_available_encoders_pyav",
    "add_features",
    "aggregate_datasets",
    "aggregate_pipeline_dataset_features",
@@ -83,7 +66,6 @@ __all__ = [
    "convert_image_to_video_dataset",
    "create_initial_features",
    "create_lerobot_dataset_card",
-    "column_for_style",
    "delete_episodes",
    "get_feature_stats",
    "load_episodes",
@@ -92,7 +74,6 @@ __all__ = [
    "modify_features",
    "modify_tasks",
    "recompute_stats",
-    "reencode_dataset",
    "remove_feature",
    "resolve_delta_timestamps",
    "safe_stop_image_writer",
--- a/src/lerobot/datasets/aggregate.py
+++ b/src/lerobot/datasets/aggregate.py
@@ -15,7 +15,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-import copy
 import logging
 import shutil
 from pathlib import Path
@@ -24,11 +23,9 @@ import datasets
 import pandas as pd
 import tqdm

-from lerobot.configs import VIDEO_ENCODER_INFO_KEYS
-
 from .compute_stats import aggregate_stats
 from .dataset_metadata import LeRobotDatasetMetadata
-from .feature_utils import features_equal_for_merge, get_hf_features_from_features
+from .feature_utils import get_hf_features_from_features
 from .io_utils import (
    get_file_size_in_mb,
    get_parquet_file_size_in_mb,
@@ -49,54 +46,11 @@ from .utils import (
 from .video_utils import concatenate_video_files, get_video_duration_in_s


-def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> dict[str, dict]:
-    """Create a merged video feature info dictionary for aggregation. The video encoder info is merged field-by-field: each key is kept only when every source agrees; otherwise that key is set to ``null`` (or ``{}`` for ``video.extra_options``) and a warning is logged.
-
-    Args:
-        all_metadata: List of LeRobotDatasetMetadata objects to merge.
-
-    Returns:
-        dict: A dictionary of merged video feature info.
-    """
-    merged_info = copy.deepcopy(all_metadata[0].features)
-    video_keys = [k for k in merged_info if merged_info[k].get("dtype") == "video"]
-
-    for vk in video_keys:
-        video_infos = [m.features.get(vk, {}).get("info") or {} for m in all_metadata]
-        base_video_info = video_infos[0]
-
-        merged_encoder_info: dict = {}
-        fallback_keys: list[str] = []
-        for info_key in VIDEO_ENCODER_INFO_KEYS:
-            values = [info.get(info_key, None) for info in video_infos]
-            first_value = values[0]
-            all_match = all(v == first_value for v in values[1:])
-
-            if all_match:
-                merged_encoder_info[info_key] = first_value
-            else:
-                fallback_keys.append(info_key)
-                merged_encoder_info[info_key] = {} if info_key == "video.extra_options" else None
-
-        if fallback_keys:
-            logging.warning(
-                f"Merging heterogeneous or incomplete video encoder metadata for feature {vk}. "
-                f"Setting these keys to null: {fallback_keys}.",
-            )
-
-        merged_info[vk]["info"] = {**base_video_info, **merged_encoder_info}
-        # TODO(CarolinePascal): make this variable once we have support for other video backends.
-        merged_info[vk]["info"]["video.video_backend"] = "pyav"
-
-    return merged_info
-
-
 def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
    """Validates that all dataset metadata have consistent properties.

    Ensures all datasets have the same fps, robot_type, and features to guarantee
    compatibility when aggregating them into a single dataset.
-    Video encoder info is not considered for validation but is merged during aggregation in ``merge_video_feature_info_for_aggregate``.

    Args:
        all_metadata: List of LeRobotDatasetMetadata objects to validate.
@@ -120,7 +74,7 @@ def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
            raise ValueError(
                f"Same robot_type is expected, but got robot_type={meta.robot_type} instead of {robot_type}."
            )
-        if not features_equal_for_merge(features, meta.features):
+        if features != meta.features:
            raise ValueError(
                f"Same features is expected, but got features={meta.features} instead of {features}."
            )
@@ -143,8 +97,8 @@ def update_data_df(df, src_meta, dst_meta):
        pd.DataFrame: Updated DataFrame with adjusted indices.
    """

-    df["episode_index"] = df["episode_index"] + dst_meta.info.total_episodes
-    df["index"] = df["index"] + dst_meta.info.total_frames
+    df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]
+    df["index"] = df["index"] + dst_meta.info["total_frames"]

    src_task_names = src_meta.tasks.index.take(df["task_index"].to_numpy())
    df["task_index"] = dst_meta.tasks.loc[src_task_names, "task_index"].to_numpy()
@@ -271,9 +225,9 @@ def update_meta_data(
        # Clean up temporary columns
        df = df.drop(columns=["_orig_chunk", "_orig_file"])

-    df["dataset_from_index"] = df["dataset_from_index"] + dst_meta.info.total_frames
-    df["dataset_to_index"] = df["dataset_to_index"] + dst_meta.info.total_frames
-    df["episode_index"] = df["episode_index"] + dst_meta.info.total_episodes
+    df["dataset_from_index"] = df["dataset_from_index"] + dst_meta.info["total_frames"]
+    df["dataset_to_index"] = df["dataset_to_index"] + dst_meta.info["total_frames"]
+    df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]

    return df

@@ -283,8 +237,8 @@ def aggregate_datasets(
    aggr_repo_id: str,
    roots: list[Path] | None = None,
    aggr_root: Path | None = None,
-    data_files_size_in_mb: int | None = None,
-    video_files_size_in_mb: int | None = None,
+    data_files_size_in_mb: float | None = None,
+    video_files_size_in_mb: float | None = None,
    chunk_size: int | None = None,
 ):
    """Aggregates multiple LeRobot datasets into a single unified dataset.
@@ -320,8 +274,7 @@ def aggregate_datasets(
            LeRobotDatasetMetadata(repo_id, root=root) for repo_id, root in zip(repo_ids, roots, strict=False)
        ]
    )
-    fps, robot_type, _ = validate_all_metadata(all_metadata)
-    features = merge_video_feature_info_for_aggregate(all_metadata)
+    fps, robot_type, features = validate_all_metadata(all_metadata)
    video_keys = [key for key in features if features[key]["dtype"] == "video"]

    dst_meta = LeRobotDatasetMetadata.create(
@@ -360,8 +313,8 @@ def aggregate_datasets(
        # to avoid interference between different source datasets
        data_idx.pop("src_to_dst", None)

-        dst_meta.info.total_episodes += src_meta.total_episodes
-        dst_meta.info.total_frames += src_meta.total_frames
+        dst_meta.info["total_episodes"] += src_meta.total_episodes
+        dst_meta.info["total_frames"] += src_meta.total_frames

    finalize_aggregation(dst_meta, all_metadata)
    logging.info("Aggregation complete.")
@@ -379,6 +332,7 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
        videos_idx: Dictionary tracking video chunk and file indices.
        video_files_size_in_mb: Maximum size for video files in MB (defaults to DEFAULT_VIDEO_FILE_SIZE_IN_MB)
        chunk_size: Maximum number of files per chunk (defaults to DEFAULT_CHUNK_SIZE)
+
    Returns:
        dict: Updated videos_idx with current chunk and file indices.
    """
@@ -460,11 +414,9 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
                current_dst_duration = dst_file_durations.get(dst_key, 0)
                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_dst_duration
                videos_idx[key]["src_to_dst"][(src_chunk_idx, src_file_idx)] = dst_key
-                # TODO(CarolinePascal): Move the check before the loop to avoid failing in the middle + add possibility to re-encode the video if the check fails
                concatenate_video_files(
                    [dst_path, src_path],
                    dst_path,
-                    compatibility_check=True,
                )
                # Update duration of this destination file
                dst_file_durations[dst_key] = current_dst_duration + src_duration
@@ -688,10 +640,14 @@ def finalize_aggregation(aggr_meta, all_metadata):
    write_tasks(aggr_meta.tasks, aggr_meta.root)

    logging.info("write info")
-    aggr_meta.info.total_tasks = len(aggr_meta.tasks)
-    aggr_meta.info.total_episodes = sum(m.total_episodes for m in all_metadata)
-    aggr_meta.info.total_frames = sum(m.total_frames for m in all_metadata)
-    aggr_meta.info.splits = {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"}
+    aggr_meta.info.update(
+        {
+            "total_tasks": len(aggr_meta.tasks),
+            "total_episodes": sum(m.total_episodes for m in all_metadata),
+            "total_frames": sum(m.total_frames for m in all_metadata),
+            "splits": {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"},
+        }
+    )
    write_info(aggr_meta.info, aggr_meta.root)

    logging.info("write stats")
--- a/src/lerobot/datasets/compute_stats.py
+++ b/src/lerobot/datasets/compute_stats.py
@@ -512,7 +512,7 @@ def compute_episode_stats(

    ep_stats = {}
    for key, data in episode_data.items():
-        if features[key]["dtype"] in {"string", "language"}:
+        if features[key]["dtype"] == "string":
            continue

        if features[key]["dtype"] in ["image", "video"]:
--- a/src/lerobot/datasets/dataset_metadata.py
+++ b/src/lerobot/datasets/dataset_metadata.py
@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import contextlib
-from collections.abc import Callable
 from pathlib import Path

 import numpy as np
@@ -24,7 +23,6 @@ import pyarrow as pa
 import pyarrow.parquet as pq
 from huggingface_hub import snapshot_download

-from lerobot.configs import VideoEncoderConfig
 from lerobot.utils.constants import DEFAULT_FEATURES, HF_LEROBOT_HOME, HF_LEROBOT_HUB_CACHE
 from lerobot.utils.feature_utils import _validate_feature_names
 from lerobot.utils.utils import flatten_dict
@@ -36,14 +34,16 @@ from .io_utils import (
    load_episodes,
    load_info,
    load_stats,
+    load_subtasks,
    load_tasks,
    write_info,
+    write_json,
    write_stats,
    write_tasks,
 )
-from .language import DEFAULT_TOOLS, LANGUAGE_COLUMNS
 from .utils import (
    DEFAULT_EPISODES_PATH,
+    INFO_PATH,
    check_version_compatibility,
    get_safe_version,
    has_legacy_hub_download_metadata,
@@ -177,6 +177,7 @@ class LeRobotDatasetMetadata:
        self.info = load_info(self.root)
        check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
        self.tasks = load_tasks(self.root)
+        self.subtasks = load_subtasks(self.root)
        self.episodes = load_episodes(self.root)
        self.stats = load_stats(self.root)

@@ -190,29 +191,6 @@ class LeRobotDatasetMetadata:
        if self.episodes is None:
            self._load_metadata()

-    def filter_episodes(
-        self,
-        predicate: Callable[[dict], bool],
-        candidates: list[int] | None = None,
-    ) -> list[int]:
-        """Filter episodes whose metadata satisfies a given predicate.
-
-        Args:
-            predicate: Predicate over per-episode metadata rows used to select episodes.
-            candidates: Optional list of episode indices to restrict evaluation to.
-
-        Returns:
-            List of sorted episode indices that satisfy the predicate.
-        """
-        self.ensure_readable()
-        if candidates is not None:
-            candidate_set = set(candidates)
-            combined = lambda ep: ep["episode_index"] in candidate_set and predicate(ep)  # noqa: E731
-        else:
-            combined = predicate
-        filtered = self.episodes.filter(combined, keep_in_memory=True, load_from_cache_file=False)
-        return sorted(int(idx) for idx in filtered["episode_index"])
-
    def _pull_from_repo(
        self,
        allow_patterns: list[str] | str | None = None,
@@ -250,7 +228,7 @@ class LeRobotDatasetMetadata:
    @property
    def _version(self) -> packaging.version.Version:
        """Codebase version used to create this dataset."""
-        return packaging.version.parse(self.info.codebase_version)
+        return packaging.version.parse(self.info["codebase_version"])

    def get_data_file_path(self, ep_index: int) -> Path:
        """Return the relative parquet file path for the given episode index.
@@ -305,27 +283,27 @@ class LeRobotDatasetMetadata:
    @property
    def data_path(self) -> str:
        """Formattable string for the parquet files."""
-        return self.info.data_path
+        return self.info["data_path"]

    @property
    def video_path(self) -> str | None:
        """Formattable string for the video files."""
-        return self.info.video_path
+        return self.info["video_path"]

    @property
    def robot_type(self) -> str | None:
        """Robot type used in recording this dataset."""
-        return self.info.robot_type
+        return self.info["robot_type"]

    @property
    def fps(self) -> int:
        """Frames per second used during data collection."""
-        return self.info.fps
+        return self.info["fps"]

    @property
    def features(self) -> dict[str, dict]:
        """All features contained in the dataset."""
-        return self.info.features
+        return self.info["features"]

    @property
    def image_keys(self) -> list[str]:
@@ -342,49 +320,6 @@ class LeRobotDatasetMetadata:
        """Keys to access visual modalities (regardless of their storage method)."""
        return [key for key, ft in self.features.items() if ft["dtype"] in ["video", "image"]]

-    @property
-    def has_language_columns(self) -> bool:
-        """Return ``True`` if the dataset declares any language column.
-
-        Used to gate language-aware code paths (collate, render step) so
-        unannotated datasets keep PyTorch's default collate behavior.
-        """
-        return any(col in self.features for col in LANGUAGE_COLUMNS)
-
-    @property
-    def tools(self) -> list[dict]:
-        """OpenAI-style tool schemas declared by this dataset.
-
-        Read from ``meta/info.json["tools"]``. Returns a copy, so callers
-        can mutate the result safely. Falls back to
-        :data:`lerobot.datasets.language.DEFAULT_TOOLS` (the canonical
-        ``say`` schema) when the dataset doesn't declare any — that way
-        unannotated datasets and chat-template consumers
-        (``apply_chat_template(messages, tools=meta.tools)``) keep
-        working out of the box.
-
-        Implementations live under :mod:`lerobot.tools` (one file per
-        tool); see ``docs/source/tools.mdx`` for the authoring guide.
-        """
-        declared = self.info.tools
-        if declared:
-            return [dict(t) for t in declared]
-        return [dict(t) for t in DEFAULT_TOOLS]
-
-    @tools.setter
-    def tools(self, value: list[dict] | None) -> None:
-        """Persist a tool catalog to ``meta/info.json`` and reload metadata.
-
-        Writes ``value`` into the on-disk ``info.json`` (or clears the
-        ``tools`` key when ``value`` is ``None`` or empty), then reloads
-        ``self.info`` so the in-memory metadata matches what's on disk.
-        Saves callers from hand-editing ``info.json`` and re-instantiating
-        the metadata object.
-        """
-        self.info.tools = [dict(t) for t in value] if value else None
-        write_info(self.info, self.root)
-        self.info = load_info(self.root)
-
    @property
    def names(self) -> dict[str, list | dict]:
        """Names of the various dimensions of vector modalities."""
@@ -398,32 +333,32 @@ class LeRobotDatasetMetadata:
    @property
    def total_episodes(self) -> int:
        """Total number of episodes available."""
-        return self.info.total_episodes
+        return self.info["total_episodes"]

    @property
    def total_frames(self) -> int:
        """Total number of frames saved in this dataset."""
-        return self.info.total_frames
+        return self.info["total_frames"]

    @property
    def total_tasks(self) -> int:
        """Total number of different tasks performed in this dataset."""
-        return self.info.total_tasks
+        return self.info["total_tasks"]

    @property
    def chunks_size(self) -> int:
        """Max number of files per chunk."""
-        return self.info.chunks_size
+        return self.info["chunks_size"]

    @property
    def data_files_size_in_mb(self) -> int:
        """Max size of data file in mega bytes."""
-        return self.info.data_files_size_in_mb
+        return self.info["data_files_size_in_mb"]

    @property
    def video_files_size_in_mb(self) -> int:
        """Max size of video file in mega bytes."""
-        return self.info.video_files_size_in_mb
+        return self.info["video_files_size_in_mb"]

    def get_task_index(self, task: str) -> int | None:
        """
@@ -567,33 +502,20 @@ class LeRobotDatasetMetadata:
        self._save_episode_metadata(episode_dict)

        # Update info
-        self.info.total_episodes += 1
-        self.info.total_frames += episode_length
-        self.info.total_tasks = len(self.tasks)
-        self.info.splits = {"train": f"0:{self.info.total_episodes}"}
+        self.info["total_episodes"] += 1
+        self.info["total_frames"] += episode_length
+        self.info["total_tasks"] = len(self.tasks)
+        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}

        write_info(self.info, self.root)

        self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats is not None else episode_stats
        write_stats(self.stats, self.root)

-    def update_video_info(
-        self,
-        video_key: str | None = None,
-        camera_encoder: VideoEncoderConfig | None = None,
-    ) -> None:
-        """Populate per-feature video info in ``info.json``.
-
+    def update_video_info(self, video_key: str | None = None) -> None:
+        """
        Warning: this function writes info from first episode videos, implicitly assuming that all videos have
        been encoded the same way. Also, this means it assumes the first episode exists.
-
-        Args:
-            video_key: If provided, only update this video key. Otherwise update
-                all video keys in the dataset.
-            camera_encoder: Encoder configuration used to produce the
-                videos. When provided, its fields are recorded as
-                ``video.<field>`` entries alongside the stream-derived
-                ``video.*`` entries (see :func:`get_video_info`).
        """
        if video_key is not None and video_key not in self.video_keys:
            raise ValueError(f"Video key {video_key} not found in dataset")
@@ -602,7 +524,7 @@ class LeRobotDatasetMetadata:
        for key in video_keys:
            if not self.features[key].get("info", None):
                video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
-                self.info.features[key]["info"] = get_video_info(video_path, camera_encoder=camera_encoder)
+                self.info["features"][key]["info"] = get_video_info(video_path)

    def update_chunk_settings(
        self,
@@ -624,17 +546,17 @@ class LeRobotDatasetMetadata:
        if chunks_size is not None:
            if chunks_size <= 0:
                raise ValueError(f"chunks_size must be positive, got {chunks_size}")
-            self.info.chunks_size = chunks_size
+            self.info["chunks_size"] = chunks_size

        if data_files_size_in_mb is not None:
            if data_files_size_in_mb <= 0:
                raise ValueError(f"data_files_size_in_mb must be positive, got {data_files_size_in_mb}")
-            self.info.data_files_size_in_mb = data_files_size_in_mb
+            self.info["data_files_size_in_mb"] = data_files_size_in_mb

        if video_files_size_in_mb is not None:
            if video_files_size_in_mb <= 0:
                raise ValueError(f"video_files_size_in_mb must be positive, got {video_files_size_in_mb}")
-            self.info.video_files_size_in_mb = video_files_size_in_mb
+            self.info["video_files_size_in_mb"] = video_files_size_in_mb

        # Update the info file on disk
        write_info(self.info, self.root)
@@ -713,6 +635,7 @@ class LeRobotDatasetMetadata:
        _validate_feature_names(features)

        obj.tasks = None
+        obj.subtasks = None
        obj.episodes = None
        obj.stats = None
        obj.info = create_empty_dataset_info(
@@ -730,7 +653,7 @@ class LeRobotDatasetMetadata:
                f"Features contain video keys {obj.video_keys}, but 'use_videos' is set to False. "
                "Either remove video features from the features dict, or set 'use_videos=True'."
            )
-        write_info(obj.info, obj.root)
+        write_json(obj.info, obj.root / INFO_PATH)
        obj.revision = None
        obj._pq_writer = None
        obj.latest_episode = None
--- a/src/lerobot/datasets/dataset_reader.py
+++ b/src/lerobot/datasets/dataset_reader.py
@@ -295,4 +295,9 @@ class DatasetReader:
        task_idx = item["task_index"].item()
        item["task"] = self._meta.tasks.iloc[task_idx].name

+        # add subtask information if available
+        if "subtask_index" in self._meta.features and self._meta.subtasks is not None:
+            subtask_idx = item["subtask_index"].item()
+            item["subtask"] = self._meta.subtasks.iloc[subtask_idx].name
+
        return item
--- a/src/lerobot/datasets/dataset_tools.py
+++ b/src/lerobot/datasets/dataset_tools.py
@@ -26,7 +26,7 @@ This module provides utilities for:
 import logging
 import shutil
 from collections.abc import Callable
-from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor, as_completed
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path

 import datasets
@@ -36,7 +36,6 @@ import pyarrow.parquet as pq
 import torch
 from tqdm import tqdm

-from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults
 from lerobot.utils.constants import ACTION, HF_LEROBOT_HOME, OBS_IMAGE, OBS_STATE
 from lerobot.utils.utils import flatten_dict

@@ -61,14 +60,9 @@ from .utils import (
    DEFAULT_DATA_FILE_SIZE_IN_MB,
    DEFAULT_DATA_PATH,
    DEFAULT_EPISODES_PATH,
-    VIDEO_DIR,
    update_chunk_file_indices,
 )
-from .video_utils import (
-    encode_video_frames,
-    get_video_info,
-    reencode_video,
-)
+from .video_utils import encode_video_frames, get_video_info


 def _load_episode_with_stats(src_dataset: LeRobotDataset, episode_idx: int) -> dict:
@@ -101,11 +95,6 @@ def delete_episodes(
 ) -> LeRobotDataset:
    """Delete episodes from a LeRobotDataset and create a new dataset.

-    Video segments that need re-encoding (because the source file mixes kept and
-    deleted episodes) are re-encoded with the source dataset's existing encoder
-    settings — read back from ``meta/info.json`` — so the output dataset stays
-    consistent with its own metadata.
-
    Args:
        dataset: The source LeRobotDataset.
        episode_indices: List of episode indices to delete.
@@ -168,11 +157,6 @@ def split_dataset(
 ) -> dict[str, LeRobotDataset]:
    """Split a LeRobotDataset into multiple smaller datasets.

-    Video segments that need re-encoding (because the source file mixes episodes
-    that fall into different splits) are re-encoded with the source dataset's
-    existing encoder settings — read back from ``meta/info.json`` — so each
-    output split stays consistent with its own metadata.
-
    Args:
        dataset: The source LeRobotDataset to split.
        splits: Either a dict mapping split names to episode indices, or a dict mapping
@@ -594,7 +578,8 @@ def _keep_episodes_from_video_with_av(
    output_path: Path,
    episodes_to_keep: list[tuple[int, int]],
    fps: float,
-    camera_encoder: VideoEncoderConfig,
+    vcodec: str = "libsvtav1",
+    pix_fmt: str = "yuv420p",
 ) -> None:
    """Keep only specified episodes from a video file using PyAV.

@@ -608,7 +593,8 @@ def _keep_episodes_from_video_with_av(
            Ranges are half-open intervals: [start_frame, end_frame), where start_frame
            is inclusive and end_frame is exclusive.
        fps: Frame rate of the video.
-        camera_encoder: Video encoder settings used to re-encode the kept frames.
+        vcodec: Video codec to use for encoding.
+        pix_fmt: Pixel format for output video.
    """
    from fractions import Fraction

@@ -633,13 +619,12 @@ def _keep_episodes_from_video_with_av(

    # Convert fps to Fraction for PyAV compatibility.
    fps_fraction = Fraction(fps).limit_denominator(1000)
-    codec_options = camera_encoder.get_codec_options(as_strings=True)
-    v_out = out.add_stream(camera_encoder.vcodec, rate=fps_fraction, options=codec_options)
+    v_out = out.add_stream(vcodec, rate=fps_fraction)

    # PyAV type stubs don't distinguish video streams from audio/subtitle streams.
    v_out.width = v_in.codec_context.width
    v_out.height = v_in.codec_context.height
-    v_out.pix_fmt = camera_encoder.pix_fmt
+    v_out.pix_fmt = pix_fmt

    # Set time_base to match the frame rate for proper timestamp handling.
    v_out.time_base = Fraction(1, int(fps))
@@ -702,14 +687,14 @@ def _copy_and_reindex_videos(
    src_dataset: LeRobotDataset,
    dst_meta: LeRobotDatasetMetadata,
    episode_mapping: dict[int, int],
+    vcodec: str = "libsvtav1",
+    pix_fmt: str = "yuv420p",
 ) -> dict[int, dict]:
    """Copy and filter video files, only re-encoding files with deleted episodes.

    For video files that only contain kept episodes, we copy them directly.
    For files with mixed kept/deleted episodes, we use PyAV filters to efficiently
-    re-encode only the desired segments. The encoder used for re-encoding is
-    derived per video key from the source dataset's ``meta/info.json`` so the
-    destination metadata keeps describing the videos accurately.
+    re-encode only the desired segments.

    Args:
        src_dataset: Source dataset to copy from
@@ -726,9 +711,6 @@ def _copy_and_reindex_videos(

    for video_key in src_dataset.meta.video_keys:
        logging.info(f"Processing videos for {video_key}")
-        camera_encoder = VideoEncoderConfig.from_video_info(
-            src_dataset.meta.info.features.get(video_key, {}).get("info")
-        )

        if dst_meta.video_path is None:
            raise ValueError("Destination metadata has no video_path defined")
@@ -810,7 +792,8 @@ def _copy_and_reindex_videos(
                    dst_video_path,
                    episodes_to_keep_ranges,
                    src_dataset.meta.fps,
-                    camera_encoder,
+                    vcodec,
+                    pix_fmt,
                )

                cumulative_ts = 0.0
@@ -914,10 +897,14 @@ def _copy_and_reindex_episodes_metadata(

    dst_meta.finalize()

-    dst_meta.info.total_episodes = len(episode_mapping)
-    dst_meta.info.total_frames = total_frames
-    dst_meta.info.total_tasks = len(dst_meta.tasks) if dst_meta.tasks is not None else 0
-    dst_meta.info.splits = {"train": f"0:{len(episode_mapping)}"}
+    dst_meta.info.update(
+        {
+            "total_episodes": len(episode_mapping),
+            "total_frames": total_frames,
+            "total_tasks": len(dst_meta.tasks) if dst_meta.tasks is not None else 0,
+            "splits": {"train": f"0:{len(episode_mapping)}"},
+        }
+    )
    write_info(dst_meta.info, dst_meta.root)

    if not all_stats:
@@ -1082,20 +1069,21 @@ def _copy_episodes_metadata_and_stats(
    if episodes_dir.exists():
        shutil.copytree(episodes_dir, dst_episodes_dir, dirs_exist_ok=True)

-    dst_meta.info.total_episodes = src_dataset.meta.total_episodes
-    dst_meta.info.total_frames = src_dataset.meta.total_frames
-    dst_meta.info.total_tasks = src_dataset.meta.total_tasks
-    # Preserve original splits if available, otherwise create default
-    dst_meta.info.splits = (
-        src_dataset.meta.info.splits
-        if src_dataset.meta.info.splits
-        else {"train": f"0:{src_dataset.meta.total_episodes}"}
+    dst_meta.info.update(
+        {
+            "total_episodes": src_dataset.meta.total_episodes,
+            "total_frames": src_dataset.meta.total_frames,
+            "total_tasks": src_dataset.meta.total_tasks,
+            "splits": src_dataset.meta.info.get("splits", {"train": f"0:{src_dataset.meta.total_episodes}"}),
+        }
    )

    if dst_meta.video_keys and src_dataset.meta.video_keys:
        for key in dst_meta.video_keys:
            if key in src_dataset.meta.features:
-                dst_meta.info.features[key]["info"] = src_dataset.meta.info.features[key].get("info", {})
+                dst_meta.info["features"][key]["info"] = src_dataset.meta.info["features"][key].get(
+                    "info", {}
+                )

    write_info(dst_meta.info, dst_meta.root)

@@ -1281,7 +1269,11 @@ def _estimate_frame_size_via_calibration(
    episode_indices: list[int],
    temp_dir: Path,
    fps: int,
-    camera_encoder: VideoEncoderConfig,
+    vcodec: str,
+    pix_fmt: str,
+    g: int,
+    crf: int,
+    fast_decode: int,
    num_calibration_frames: int = 30,
 ) -> float:
    """Estimate MB per frame by encoding a small calibration sample.
@@ -1295,7 +1287,11 @@ def _estimate_frame_size_via_calibration(
        episode_indices: List of episode indices being processed.
        temp_dir: Temporary directory for calibration files.
        fps: Frames per second for video encoding.
-        camera_encoder: Video encoder settings used for calibration encoding.
+        vcodec: Video codec (libsvtav1, h264, hevc).
+        pix_fmt: Pixel format (yuv420p, etc.).
+        g: GOP size (group of pictures).
+        crf: Constant Rate Factor (quality).
+        fast_decode: Fast decode tuning parameter.
        num_calibration_frames: Number of frames to use for calibration (default: 30).

    Returns:
@@ -1331,7 +1327,11 @@ def _estimate_frame_size_via_calibration(
            imgs_dir=calibration_dir,
            video_path=calibration_video_path,
            fps=fps,
-            camera_encoder=camera_encoder,
+            vcodec=vcodec,
+            pix_fmt=pix_fmt,
+            g=g,
+            crf=crf,
+            fast_decode=fast_decode,
            overwrite=True,
        )

@@ -1525,7 +1525,7 @@ def modify_tasks(
    write_tasks(new_task_df, root)

    # Update info.json
-    dataset.meta.info.total_tasks = len(unique_tasks)
+    dataset.meta.info["total_tasks"] = len(unique_tasks)
    write_info(dataset.meta.info, root)

    # Reload metadata to reflect changes
@@ -1649,7 +1649,11 @@ def convert_image_to_video_dataset(
    dataset: LeRobotDataset,
    output_dir: Path | None = None,
    repo_id: str | None = None,
-    camera_encoder: VideoEncoderConfig | None = None,
+    vcodec: str = "libsvtav1",
+    pix_fmt: str = "yuv420p",
+    g: int = 2,
+    crf: int = 30,
+    fast_decode: int = 0,
    episode_indices: list[int] | None = None,
    num_workers: int = 4,
    max_episodes_per_batch: int | None = None,
@@ -1664,8 +1668,11 @@ def convert_image_to_video_dataset(
        dataset: The source LeRobot dataset with images
        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
-        camera_encoder: Video encoder settings
-            (``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`).
+        vcodec: Video codec (default: libsvtav1)
+        pix_fmt: Pixel format (default: yuv420p)
+        g: Group of pictures size (default: 2)
+        crf: Constant rate factor (default: 30)
+        fast_decode: Fast decode tuning (default: 0)
        episode_indices: List of episode indices to convert (None = all episodes)
        num_workers: Number of threads for parallel processing (default: 4)
        max_episodes_per_batch: Maximum episodes per video batch to avoid memory issues (None = no limit)
@@ -1674,9 +1681,6 @@ def convert_image_to_video_dataset(
    Returns:
        New LeRobotDataset with images encoded as videos
    """
-    if camera_encoder is None:
-        camera_encoder = camera_encoder_defaults()
-
    # Check that it's an image dataset
    if len(dataset.meta.video_keys) > 0:
        raise ValueError(
@@ -1700,10 +1704,7 @@ def convert_image_to_video_dataset(
    logging.info(
        f"Converting {len(episode_indices)} episodes with {len(img_keys)} cameras from {dataset.repo_id}"
    )
-    logging.info(
-        f"Video codec: {camera_encoder.vcodec}, pixel format: {camera_encoder.pix_fmt}, "
-        f"GOP: {camera_encoder.g}, CRF: {camera_encoder.crf}"
-    )
+    logging.info(f"Video codec: {vcodec}, pixel format: {pix_fmt}, GOP: {g}, CRF: {crf}")

    # Create new features dict, converting image features to video features
    new_features = {}
@@ -1773,7 +1774,11 @@ def convert_image_to_video_dataset(
                episode_indices=episode_indices,
                temp_dir=temp_dir,
                fps=fps,
-                camera_encoder=camera_encoder,
+                vcodec=vcodec,
+                pix_fmt=pix_fmt,
+                g=g,
+                crf=crf,
+                fast_decode=fast_decode,
            )

            logging.info(f"Processing camera: {img_key}")
@@ -1815,7 +1820,11 @@ def convert_image_to_video_dataset(
                    imgs_dir=imgs_dir,
                    video_path=video_path,
                    fps=fps,
-                    camera_encoder=camera_encoder,
+                    vcodec=vcodec,
+                    pix_fmt=pix_fmt,
+                    g=g,
+                    crf=crf,
+                    fast_decode=fast_decode,
                    overwrite=True,
                )

@@ -1849,10 +1858,10 @@ def convert_image_to_video_dataset(
        episodes_df.to_parquet(episodes_path, index=False)

        # Update metadata info
-        new_meta.info.total_episodes = len(episode_indices)
-        new_meta.info.total_frames = sum(ep["length"] for ep in all_episode_metadata.values())
-        new_meta.info.total_tasks = dataset.meta.total_tasks
-        new_meta.info.splits = {"train": f"0:{len(episode_indices)}"}
+        new_meta.info["total_episodes"] = len(episode_indices)
+        new_meta.info["total_frames"] = sum(ep["length"] for ep in all_episode_metadata.values())
+        new_meta.info["total_tasks"] = dataset.meta.total_tasks
+        new_meta.info["splits"] = {"train": f"0:{len(episode_indices)}"}

        # Update video info for all image keys (now videos)
        # We need to manually set video info since update_video_info() checks video_keys first
@@ -1861,9 +1870,7 @@ def convert_image_to_video_dataset(
                video_path = new_meta.root / new_meta.video_path.format(
                    video_key=img_key, chunk_index=0, file_index=0
                )
-                new_meta.info.features[img_key]["info"] = get_video_info(
-                    video_path, camera_encoder=camera_encoder
-                )
+                new_meta.info["features"][img_key]["info"] = get_video_info(video_path)

        write_info(new_meta.info, new_meta.root)

@@ -1886,83 +1893,3 @@ def convert_image_to_video_dataset(

    # Return new dataset
    return LeRobotDataset(repo_id=repo_id, root=output_dir)
-
-
-def _reencode_video_worker(args: tuple) -> Path:
-    """Picklable worker for :func:`reencode_dataset`'s process pool."""
-    video_path, camera_encoder, encoder_threads = args
-    reencode_video(
-        input_video_path=video_path,
-        output_video_path=video_path,
-        camera_encoder=camera_encoder,
-        encoder_threads=encoder_threads,
-        overwrite=True,
-    )
-    return video_path
-
-
-def reencode_dataset(
-    dataset: LeRobotDataset,
-    camera_encoder: VideoEncoderConfig,
-    encoder_threads: int | None = None,
-    num_workers: int | None = None,
-) -> LeRobotDataset:
-    """Re-encode every video in a dataset with a new set of encoding parameters.
-
-    Videos are re-encoded in-place and the video information in ``info.json`` is refreshed.
-
-    Args:
-        dataset: An existing :class:`LeRobotDataset` whose videos will be
-            re-encoded.
-        camera_encoder: Target encoder configuration applied to every video
-            file.
-        encoder_threads: Per-encoder thread count forwarded to
-            :func:`reencode_video`. ``None`` lets the codec decide.
-        num_workers: Number of parallel processes. ``None`` or ``0`` means
-            sequential (no multiprocessing); ``1+`` spawns a
-            :class:`~concurrent.futures.ProcessPoolExecutor`.
-
-    Returns:
-        The same :class:`LeRobotDataset` instance with its metadata updated
-        on disk.
-    """
-    meta = dataset.meta
-    video_paths_list = []
-
-    # Only re-encode if the videos are not already encoded with the given video encoding parameters
-    for video_key in meta.video_keys:
-        current_info = meta.info.features[video_key].get("info", {})
-        current_encoder = VideoEncoderConfig.from_video_info(current_info)
-        if current_encoder != camera_encoder:
-            video_paths_list.extend((meta.root / VIDEO_DIR / video_key).rglob("*.mp4"))
-        else:
-            logging.info(f"{video_key} videos are already encoded with {camera_encoder}. Nothing to do.")
-
-    if len(video_paths_list) == 0:
-        logging.warning("Dataset has no videos to re-encode.")
-        return dataset
-    logging.info(f"Re-encoding {len(video_paths_list)} video file(s) with {camera_encoder}")
-
-    worker_args = [(vp, camera_encoder, encoder_threads) for vp in video_paths_list]
-    if num_workers and num_workers > 1:
-        with ProcessPoolExecutor(max_workers=num_workers) as pool:
-            futures = [pool.submit(_reencode_video_worker, args) for args in worker_args]
-            for future in tqdm(
-                as_completed(futures),
-                total=len(futures),
-                desc="Re-encoding videos",
-            ):
-                future.result()
-    else:
-        for args in tqdm(worker_args, desc="Re-encoding videos"):
-            _reencode_video_worker(args)
-
-    # Refresh video info in metadata for every video key.
-    for vid_key in meta.video_keys:
-        video_path = meta.root / meta.get_video_file_path(0, vid_key)
-        meta.info.features[vid_key]["info"] = get_video_info(video_path, camera_encoder=camera_encoder)
-
-    write_info(meta.info, meta.root)
-    logging.info("Dataset metadata updated.")
-
-    return dataset
--- a/src/lerobot/datasets/dataset_writer.py
+++ b/src/lerobot/datasets/dataset_writer.py
@@ -31,8 +31,6 @@ import PIL.Image
 import pyarrow.parquet as pq
 import torch

-from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults
-
 from .compute_stats import compute_episode_stats
 from .dataset_metadata import LeRobotDatasetMetadata
 from .feature_utils import (
@@ -67,19 +65,14 @@ def _encode_video_worker(
    episode_index: int,
    root: Path,
    fps: int,
-    camera_encoder: VideoEncoderConfig | None = None,
+    vcodec: str = "libsvtav1",
    encoder_threads: int | None = None,
 ) -> Path:
    temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
    fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
    img_dir = (root / fpath).parent
    encode_video_frames(
-        img_dir,
-        temp_path,
-        fps,
-        camera_encoder=camera_encoder,
-        encoder_threads=encoder_threads,
-        overwrite=True,
+        img_dir, temp_path, fps, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
    )
    shutil.rmtree(img_dir)
    return temp_path
@@ -96,22 +89,20 @@ class DatasetWriter:
        self,
        meta: LeRobotDatasetMetadata,
        root: Path,
-        camera_encoder: VideoEncoderConfig | None,
+        vcodec: str,
        encoder_threads: int | None,
        batch_encoding_size: int,
        streaming_encoder: StreamingVideoEncoder | None = None,
        initial_frames: int = 0,
    ):
-        """Initialize the writer with metadata, codec, and encoder config.
+        """Initialize the writer with metadata, codec, and encoding config.

        Args:
            meta: Dataset metadata instance (used for feature schema, chunk
                settings, and episode persistence).
            root: Local dataset root directory.
-            camera_encoder: Video encoder settings applied to all cameras.
-                ``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`.
-            encoder_threads: Number of encoder threads (global). ``None``
-                lets the codec decide.
+            vcodec: Video codec for encoding (e.g. ``'libsvtav1'``, ``'h264'``).
+            encoder_threads: Threads per encoder instance. ``None`` for auto.
            batch_encoding_size: Number of episodes to accumulate before
                batch-encoding videos.
            streaming_encoder: Optional pre-built :class:`StreamingVideoEncoder`
@@ -120,7 +111,7 @@ class DatasetWriter:
        """
        self._meta = meta
        self._root = root
-        self._camera_encoder = camera_encoder or camera_encoder_defaults()
+        self._vcodec = vcodec
        self._encoder_threads = encoder_threads
        self._batch_encoding_size = batch_encoding_size
        self._streaming_encoder = streaming_encoder
@@ -250,14 +241,7 @@ class DatasetWriter:
        for key, ft in self._meta.features.items():
            if key in ["index", "episode_index", "task_index"] or ft["dtype"] in ["image", "video"]:
                continue
-            stacked_values = np.stack(episode_buffer[key])
-
-            # `shape=(1,)` numeric features are serialized as `datasets.Value`, which expects scalars.
-            # Normalizing to `(N,)` keeps save semantics stable across dependency versions.
-            if tuple(ft["shape"]) == (1,) and ft["dtype"] != "string":
-                stacked_values = stacked_values.reshape(episode_length)
-
-            episode_buffer[key] = stacked_values
+            episode_buffer[key] = np.stack(episode_buffer[key])

        # Wait for image writer to end, so that episode stats over images can be computed
        self._wait_image_writer()
@@ -300,7 +284,7 @@ class DatasetWriter:
                            episode_index,
                            self._root,
                            self._meta.fps,
-                            self._camera_encoder,
+                            self._vcodec,
                            self._encoder_threads,
                        ): video_key
                        for video_key in self._meta.video_keys
@@ -511,7 +495,7 @@ class DatasetWriter:

        # Update video info (only needed when first episode is encoded)
        if episode_index == 0:
-            self._meta.update_video_info(video_key, camera_encoder=self._camera_encoder)
+            self._meta.update_video_info(video_key)
            write_info(self._meta.info, self._meta.root)

        metadata = {
@@ -580,12 +564,7 @@ class DatasetWriter:
    def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path:
        """Use ffmpeg to convert frames stored as png into mp4 videos."""
        return _encode_video_worker(
-            video_key,
-            episode_index,
-            self._root,
-            self._meta.fps,
-            self._camera_encoder,
-            self._encoder_threads,
+            video_key, episode_index, self._root, self._meta.fps, self._vcodec, self._encoder_threads
        )

    def close_writer(self) -> None:
--- a/src/lerobot/datasets/factory.py
+++ b/src/lerobot/datasets/factory.py
@@ -19,7 +19,6 @@ from pprint import pformat
 import torch

 from lerobot.configs import PreTrainedConfig
-from lerobot.configs.rewards import RewardModelConfig
 from lerobot.configs.train import TrainPipelineConfig
 from lerobot.transforms import ImageTransforms
 from lerobot.utils.constants import ACTION, IMAGENET_STATS, OBS_PREFIX, REWARD
@@ -31,14 +30,12 @@ from .streaming_dataset import StreamingLeRobotDataset


 def resolve_delta_timestamps(
-    cfg: PreTrainedConfig | RewardModelConfig, ds_meta: LeRobotDatasetMetadata
+    cfg: PreTrainedConfig, ds_meta: LeRobotDatasetMetadata
 ) -> dict[str, list] | None:
-    """Resolves delta_timestamps by reading from the 'delta_indices' properties of the config.
+    """Resolves delta_timestamps by reading from the 'delta_indices' properties of the PreTrainedConfig.

    Args:
-        cfg (PreTrainedConfig | RewardModelConfig): The config to read delta_indices from. Both
-            ``PreTrainedConfig`` and concrete ``RewardModelConfig`` subclasses expose the
-            ``{observation,action,reward}_delta_indices`` properties used below.
+        cfg (PreTrainedConfig): The PreTrainedConfig to read delta_indices from.
        ds_meta (LeRobotDatasetMetadata): The dataset from which features and fps are used to build
            delta_timestamps against.

@@ -85,7 +82,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
        ds_meta = LeRobotDatasetMetadata(
            cfg.dataset.repo_id, root=cfg.dataset.root, revision=cfg.dataset.revision
        )
-        delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, ds_meta)
+        delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
        if not cfg.dataset.streaming:
            dataset = LeRobotDataset(
                cfg.dataset.repo_id,
--- a/src/lerobot/datasets/feature_utils.py
+++ b/src/lerobot/datasets/feature_utils.py
@@ -13,30 +13,21 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import logging
 from pprint import pformat

 import datasets
 import numpy as np
 from PIL import Image as PILImage

-from lerobot.configs import VIDEO_ENCODER_INFO_KEYS
 from lerobot.utils.constants import DEFAULT_FEATURES
 from lerobot.utils.utils import is_valid_numpy_dtype_string

-from .language import (
-    LANGUAGE_PERSISTENT,
-    is_language_column,
-    language_events_column_feature,
-    language_persistent_column_feature,
-)
 from .utils import (
    DEFAULT_CHUNK_SIZE,
    DEFAULT_DATA_FILE_SIZE_IN_MB,
    DEFAULT_DATA_PATH,
    DEFAULT_VIDEO_FILE_SIZE_IN_MB,
    DEFAULT_VIDEO_PATH,
-    DatasetInfo,
 )


@@ -54,13 +45,7 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
    """
    hf_features = {}
    for key, ft in features.items():
-        if is_language_column(key):
-            hf_features[key] = (
-                language_persistent_column_feature()
-                if key == LANGUAGE_PERSISTENT
-                else language_events_column_feature()
-            )
-        elif ft["dtype"] == "video":
+        if ft["dtype"] == "video":
            continue
        elif ft["dtype"] == "image":
            hf_features[key] = datasets.Image()
@@ -93,8 +78,8 @@ def create_empty_dataset_info(
    chunks_size: int | None = None,
    data_files_size_in_mb: int | None = None,
    video_files_size_in_mb: int | None = None,
-) -> DatasetInfo:
-    """Create a template ``DatasetInfo`` object for a new dataset's ``meta/info.json``.
+) -> dict:
+    """Create a template dictionary for a new dataset's `info.json`.

    Args:
        codebase_version (str): The version of the LeRobot codebase.
@@ -102,59 +87,25 @@ def create_empty_dataset_info(
        features (dict): The LeRobot features dictionary for the dataset.
        use_videos (bool): Whether the dataset will store videos.
        robot_type (str | None): The type of robot used, if any.
-        chunks_size (int | None): Max files per chunk directory. Defaults to ``DEFAULT_CHUNK_SIZE``.
-        data_files_size_in_mb (int | None): Max parquet file size in MB. Defaults to ``DEFAULT_DATA_FILE_SIZE_IN_MB``.
-        video_files_size_in_mb (int | None): Max video file size in MB. Defaults to ``DEFAULT_VIDEO_FILE_SIZE_IN_MB``.

    Returns:
-        DatasetInfo: A typed dataset information object with initial metadata.
+        dict: A dictionary with the initial dataset metadata.
    """
-    return DatasetInfo(
-        codebase_version=codebase_version,
-        fps=fps,
-        features=features,
-        robot_type=robot_type,
-        chunks_size=chunks_size or DEFAULT_CHUNK_SIZE,
-        data_files_size_in_mb=data_files_size_in_mb or DEFAULT_DATA_FILE_SIZE_IN_MB,
-        video_files_size_in_mb=video_files_size_in_mb or DEFAULT_VIDEO_FILE_SIZE_IN_MB,
-        data_path=DEFAULT_DATA_PATH,
-        video_path=DEFAULT_VIDEO_PATH if use_videos else None,
-    )
-
-
-def features_equal_for_merge(features_a: dict[str, dict], features_b: dict[str, dict]) -> bool:
-    """Return whether two LeRobotDatasetMetadata ``features`` dicts are compatible for aggregation.
-
-    For video features, keys under ``info`` related to video encoding parameters are ignored during
-    comparison as they do not prevent aggregation.
-    """
-
-    def _without_encoder_info_keys(feature: dict) -> dict:
-        filtered = dict(feature)
-        filtered_info = filtered.get("info")
-        if isinstance(filtered_info, dict):
-            filtered["info"] = {
-                info_key: info_value
-                for info_key, info_value in filtered_info.items()
-                if info_key not in VIDEO_ENCODER_INFO_KEYS
-            }
-        return filtered
-
-    if set(features_a) != set(features_b):
-        return False
-    for key in features_a:
-        fa_key = features_a[key]
-        fb_key = features_b[key]
-        if fa_key.get("dtype") != fb_key.get("dtype"):
-            return False
-        if fa_key.get("dtype") != "video":
-            if fa_key != fb_key:
-                return False
-            continue
-
-        if _without_encoder_info_keys(fa_key) != _without_encoder_info_keys(fb_key):
-            return False
-    return True
+    return {
+        "codebase_version": codebase_version,
+        "robot_type": robot_type,
+        "total_episodes": 0,
+        "total_frames": 0,
+        "total_tasks": 0,
+        "chunks_size": chunks_size or DEFAULT_CHUNK_SIZE,
+        "data_files_size_in_mb": data_files_size_in_mb or DEFAULT_DATA_FILE_SIZE_IN_MB,
+        "video_files_size_in_mb": video_files_size_in_mb or DEFAULT_VIDEO_FILE_SIZE_IN_MB,
+        "fps": fps,
+        "splits": {},
+        "data_path": DEFAULT_DATA_PATH,
+        "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
+        "features": features,
+    }


 def check_delta_timestamps(
@@ -291,8 +242,6 @@ def validate_feature_dtype_and_shape(
        return validate_feature_image_or_video(name, expected_shape, value)
    elif expected_dtype == "string":
        return validate_feature_string(name, value)
-    elif expected_dtype == "language":
-        return validate_feature_language(name, value)
    else:
        raise NotImplementedError(f"The feature dtype '{expected_dtype}' is not implemented yet.")

@@ -372,30 +321,6 @@ def validate_feature_string(name: str, value: str) -> str:
    return ""


-def validate_feature_language(name: str, value) -> str:
-    """Validate a feature that is expected to hold language annotations.
-
-    Language columns (``language_persistent`` / ``language_events``) are
-    populated after recording by the annotation pipeline, not at record time.
-    Any value supplied here is dropped before the frame is written, so a
-    non-empty value almost certainly signals a mistake. We warn rather than
-    fail to keep recording resilient.
-
-    Args:
-        name (str): The name of the feature.
-        value: The value to validate.
-
-    Returns:
-        str: Always an empty string — language values are non-fatal.
-    """
-    if value is not None:
-        logging.warning(
-            f"The feature '{name}' is a 'language' column populated by the annotation pipeline, "
-            f"not at record time. The provided value will be dropped."
-        )
-    return ""
-
-
 def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict) -> None:
    """Validate the episode buffer before it's written to disk.

--- a/src/lerobot/datasets/io_utils.py
+++ b/src/lerobot/datasets/io_utils.py
@@ -31,15 +31,14 @@ from torchvision import transforms
 from lerobot.utils.io_utils import load_json, write_json
 from lerobot.utils.utils import SuppressProgressBars, flatten_dict, unflatten_dict

-from .language import LANGUAGE_COLUMNS
 from .utils import (
    DEFAULT_DATA_FILE_SIZE_IN_MB,
    DEFAULT_EPISODES_PATH,
+    DEFAULT_SUBTASKS_PATH,
    DEFAULT_TASKS_PATH,
    EPISODES_DIR,
    INFO_PATH,
    STATS_PATH,
-    DatasetInfo,
    serialize_dict,
 )

@@ -116,21 +115,25 @@ def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
    return dataset


-def write_info(info: DatasetInfo, local_dir: Path) -> None:
-    write_json(info.to_dict(), local_dir / INFO_PATH)
+def write_info(info: dict, local_dir: Path) -> None:
+    write_json(info, local_dir / INFO_PATH)


-def load_info(local_dir: Path) -> DatasetInfo:
+def load_info(local_dir: Path) -> dict:
    """Load dataset info metadata from its standard file path.

+    Also converts shape lists to tuples for consistency.
+
    Args:
        local_dir (Path): The root directory of the dataset.

    Returns:
-        DatasetInfo: The typed dataset information object.
+        dict: The dataset information dictionary.
    """
-    raw = load_json(local_dir / INFO_PATH)
-    return DatasetInfo.from_dict(raw)
+    info = load_json(local_dir / INFO_PATH)
+    for ft in info["features"].values():
+        ft["shape"] = tuple(ft["shape"])
+    return info


 def write_stats(stats: dict, local_dir: Path) -> None:
@@ -186,6 +189,14 @@ def load_tasks(local_dir: Path) -> pandas.DataFrame:
    return tasks


+def load_subtasks(local_dir: Path) -> pandas.DataFrame | None:
+    """Load subtasks from subtasks.parquet if it exists."""
+    subtasks_path = local_dir / DEFAULT_SUBTASKS_PATH
+    if subtasks_path.exists():
+        return pd.read_parquet(subtasks_path)
+    return None
+
+
 def write_episodes(episodes: Dataset, local_dir: Path) -> None:
    """Write episode metadata to a parquet file in the LeRobot v3.0 format.
    This function writes episode-level metadata to a single parquet file.
@@ -257,13 +268,11 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to
        dict: The batch with items converted to torch tensors.
    """
    for key in items_dict:
-        if key in LANGUAGE_COLUMNS:
-            continue
        first_item = items_dict[key][0]
        if isinstance(first_item, PILImage.Image):
            to_tensor = transforms.ToTensor()
            items_dict[key] = [to_tensor(img) for img in items_dict[key]]
-        elif first_item is None or isinstance(first_item, dict):
+        elif first_item is None:
            pass
        else:
            items_dict[key] = [x if isinstance(x, str) else torch.tensor(x) for x in items_dict[key]]
@@ -298,9 +307,8 @@ def item_to_torch(item: dict) -> dict:
    Returns:
        dict: Dictionary with all tensor-like items converted to torch.Tensor.
    """
-    skip_keys = {"task", *LANGUAGE_COLUMNS}
    for key, val in item.items():
-        if isinstance(val, (np.ndarray | list)) and key not in skip_keys:
+        if isinstance(val, (np.ndarray | list)) and key not in ["task"]:
            # Convert numpy arrays and lists to torch tensors
            item[key] = torch.tensor(val)
    return item
--- a/src/lerobot/datasets/language.py
+++ b/src/lerobot/datasets/language.py
@@ -1,242 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import annotations
-
-from typing import Literal
-
-import datasets
-import pyarrow as pa
-
-LANGUAGE_PERSISTENT = "language_persistent"
-LANGUAGE_EVENTS = "language_events"
-LANGUAGE_COLUMNS = (LANGUAGE_PERSISTENT, LANGUAGE_EVENTS)
-PERSISTENT_ROW_FIELDS = ("role", "content", "style", "timestamp", "camera", "tool_calls")
-EVENT_ROW_FIELDS = ("role", "content", "style", "camera", "tool_calls")
-
-CORE_STYLES = {
-    "subtask",
-    "plan",
-    "memory",
-    "motion",
-    "interjection",
-    "vqa",
-    "trace",
-    "task_aug",
-}
-# Project-local styles can be registered at import time by appending to
-# ``EXTENDED_STYLES`` before ``column_for_style`` is called. Anything added
-# here is treated as a known style alongside ``CORE_STYLES`` for resolver
-# validation. Empty by default — populate from a downstream module that
-# also extends ``PERSISTENT_STYLES`` or ``EVENT_ONLY_STYLES`` to declare
-# the new style's column.
-EXTENDED_STYLES: set[str] = set()
-STYLE_REGISTRY = CORE_STYLES | EXTENDED_STYLES
-
-PERSISTENT_STYLES = {"subtask", "plan", "memory", "motion", "task_aug"}
-EVENT_ONLY_STYLES = {"interjection", "vqa", "trace"}
-
-# Styles whose ``content`` is grounded in a specific camera view. Rows of these
-# styles MUST carry a non-null ``camera`` referencing an ``observation.images.*``
-# feature key. Rows of every other style MUST have ``camera=None``. ``motion``
-# is intentionally NOT in this set: motion primitives are described in
-# robot-frame (joint / Cartesian) terms, not pixel space, so they are
-# camera-agnostic. ``trace`` is the pixel-trajectory event style and IS
-# view-dependent. The ``camera`` field nevertheless lives on
-# ``PERSISTENT_ROW_FIELDS`` too so the schema, validator, and resolver
-# behave symmetrically across the two columns; persistent rows simply
-# always have ``camera=None`` in practice today.
-VIEW_DEPENDENT_STYLES = {"vqa", "trace"}
-
-LanguageColumn = Literal["language_persistent", "language_events"]
-
-
-def _json_arrow_type() -> pa.DataType:
-    """Return the Arrow JSON type, falling back to ``string`` on older pyarrow."""
-    return pa.json_() if hasattr(pa, "json_") else pa.string()
-
-
-def _json_feature() -> object:
-    """Return the HF ``datasets`` JSON feature, falling back to a string value."""
-    return datasets.Json() if hasattr(datasets, "Json") else datasets.Value("string")
-
-
-def language_persistent_row_arrow_type() -> pa.StructType:
-    """Return the Arrow struct type for a single persistent language row.
-
-    Persistent rows carry their own ``timestamp`` because they represent a state
-    that became active at a specific moment and remains active until superseded.
-    ``timestamp`` is ``float32`` to match the timestamp dtype LeRobotDataset
-    uses for frame data.
-    """
-    return pa.struct(
-        [
-            pa.field("role", pa.string(), nullable=False),
-            pa.field("content", pa.string(), nullable=True),
-            pa.field("style", pa.string(), nullable=True),
-            pa.field("timestamp", pa.float32(), nullable=False),
-            pa.field("camera", pa.string(), nullable=True),
-            pa.field("tool_calls", pa.list_(_json_arrow_type()), nullable=True),
-        ]
-    )
-
-
-def language_event_row_arrow_type() -> pa.StructType:
-    """Return the Arrow struct type for a single event language row.
-
-    Event rows have no ``timestamp`` field: each event is stored on the dataset
-    row whose frame timestamp is the event's firing time.
-    """
-    return pa.struct(
-        [
-            pa.field("role", pa.string(), nullable=False),
-            pa.field("content", pa.string(), nullable=True),
-            pa.field("style", pa.string(), nullable=True),
-            pa.field("camera", pa.string(), nullable=True),
-            pa.field("tool_calls", pa.list_(_json_arrow_type()), nullable=True),
-        ]
-    )
-
-
-def language_persistent_arrow_type() -> pa.ListType:
-    """Return the Arrow list type for the ``language_persistent`` column."""
-    return pa.list_(language_persistent_row_arrow_type())
-
-
-def language_events_arrow_type() -> pa.ListType:
-    """Return the Arrow list type for the ``language_events`` column."""
-    return pa.list_(language_event_row_arrow_type())
-
-
-def language_persistent_row_feature() -> dict[str, object]:
-    """Return the HF ``datasets`` feature mapping for a persistent language row."""
-    return {
-        "role": datasets.Value("string"),
-        "content": datasets.Value("string"),
-        "style": datasets.Value("string"),
-        "timestamp": datasets.Value("float32"),
-        "camera": datasets.Value("string"),
-        "tool_calls": datasets.List(_json_feature()),
-    }
-
-
-def language_event_row_feature() -> dict[str, object]:
-    """Return the HF ``datasets`` feature mapping for an event language row."""
-    return {
-        "role": datasets.Value("string"),
-        "content": datasets.Value("string"),
-        "style": datasets.Value("string"),
-        "camera": datasets.Value("string"),
-        "tool_calls": datasets.List(_json_feature()),
-    }
-
-
-def language_persistent_column_feature() -> datasets.List:
-    """Return the HF ``datasets`` feature for the ``language_persistent`` column."""
-    return datasets.List(language_persistent_row_feature())
-
-
-def language_events_column_feature() -> datasets.List:
-    """Return the HF ``datasets`` feature for the ``language_events`` column."""
-    return datasets.List(language_event_row_feature())
-
-
-def language_feature_info() -> dict[str, dict]:
-    """Return the ``info["features"]`` entries for both language columns."""
-    return {
-        LANGUAGE_PERSISTENT: {"dtype": "language", "shape": (1,), "names": None},
-        LANGUAGE_EVENTS: {"dtype": "language", "shape": (1,), "names": None},
-    }
-
-
-def is_language_column(key: str) -> bool:
-    """Return ``True`` if ``key`` is one of the dataset's language column names."""
-    return key in LANGUAGE_COLUMNS
-
-
-def is_view_dependent_style(style: str | None) -> bool:
-    """Return ``True`` if rows of ``style`` must be tagged with a ``camera`` key."""
-    return style in VIEW_DEPENDENT_STYLES
-
-
-def validate_camera_field(style: str | None, camera: str | None) -> None:
-    """Enforce the ``camera`` invariant: required iff ``style`` is view-dependent.
-
-    Raises ``ValueError`` if a view-dependent style is missing ``camera`` or if
-    a non-view-dependent style carries one. Pipeline writers and the validator
-    should call this on every emitted row.
-    """
-    if is_view_dependent_style(style):
-        if not camera:
-            raise ValueError(
-                f"Rows of view-dependent style {style!r} require a non-empty 'camera' "
-                f"field referencing an 'observation.images.*' feature key."
-            )
-    elif camera is not None:
-        raise ValueError(f"Rows of style {style!r} must have camera=None; got camera={camera!r}.")
-
-
-# --- Tool registry --------------------------------------------------------
-# Tools declared on a dataset live in ``meta/info.json["tools"]`` as a list
-# of OpenAI-style function schemas. The runtime / training stack reads them
-# through :class:`LeRobotDatasetMetadata.tools` (with these constants as
-# fallback when the dataset doesn't declare any). Implementations live
-# under :mod:`lerobot.tools` (one file per tool); see
-# ``docs/source/tools.mdx`` for the authoring guide.
-
-SAY_TOOL_SCHEMA: dict = {
-    "type": "function",
-    "function": {
-        "name": "say",
-        "description": "Speak a short utterance to the user via the TTS executor.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "text": {
-                    "type": "string",
-                    "description": "The verbatim text to speak.",
-                }
-            },
-            "required": ["text"],
-        },
-    },
-}
-"""Canonical schema for the ``say`` tool emitted by the steerable
-annotation pipeline (PR 2 Module 2). Single source of truth — PR 2's
-writer, PR 3's runtime tool registry, and the dataset visualizer all
-import this constant rather than duplicating the dict."""
-
-DEFAULT_TOOLS: list[dict] = [SAY_TOOL_SCHEMA]
-"""Fallback tools list. Returned by ``LeRobotDatasetMetadata.tools``
-when ``meta/info.json["tools"]`` is unset, so unannotated datasets and
-chat-template consumers (``apply_chat_template(messages, tools=...)``)
-keep working out of the box."""
-
-
-def column_for_style(style: str | None) -> LanguageColumn:
-    """Map a language style to the column where rows of that style are stored.
-
-    Styles in :data:`PERSISTENT_STYLES` route to :data:`LANGUAGE_PERSISTENT`.
-    Styles in :data:`EVENT_ONLY_STYLES` and the implicit ``None`` style route
-    to :data:`LANGUAGE_EVENTS`.
-    """
-    if style is None:
-        return LANGUAGE_EVENTS
-    if style in PERSISTENT_STYLES:
-        return LANGUAGE_PERSISTENT
-    if style in EVENT_ONLY_STYLES:
-        return LANGUAGE_EVENTS
-    raise ValueError(f"Unknown language style: {style!r}")
--- a/src/lerobot/datasets/language_render.py
+++ b/src/lerobot/datasets/language_render.py
@@ -1,545 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import annotations
-
-import copy
-import hashlib
-import re
-from collections.abc import Sequence
-from typing import Any
-
-from lerobot.configs.recipe import DEFAULT_BINDINGS, PLACEHOLDER_RE, TrainingRecipe
-from lerobot.utils.utils import unwrap_scalar
-
-from .language import LANGUAGE_PERSISTENT, column_for_style
-
-LanguageRow = dict[str, Any]
-RenderedMessages = dict[str, list[Any]]
-
-_RESOLVER_RE = re.compile(r"^(?P<name>[A-Za-z_][A-Za-z0-9_]*)\((?P<args>.*)\)$")
-
-
-def active_at(
-    t: float,
-    *,
-    persistent: Sequence[LanguageRow],
-    style: str | None = None,
-    role: str | None = None,
-    tool_name: str | None = None,
-    camera: str | None = None,
-) -> LanguageRow | None:
-    """Return the persistent row of ``style`` that is active at time ``t``.
-
-    A persistent row is "active" at ``t`` when its own ``timestamp`` is the
-    most recent one ``<= t`` for the given ``style``/``role``/``tool_name``/
-    ``camera`` selector. Only valid for persistent styles.
-    """
-    _validate_persistent_resolver("active_at", style)
-    matches = [
-        row
-        for row in _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera)
-        if _timestamp(row) <= t
-    ]
-    if not matches:
-        return None
-    latest_ts = max(_timestamp(row) for row in matches)
-    return _select_one(
-        [row for row in matches if _timestamp(row) == latest_ts],
-        style=style,
-        role=role,
-        tool_name=tool_name,
-        camera=camera,
-    )
-
-
-EMITTED_AT_TOLERANCE_S = 0.1
-"""Half-window for matching persistent rows to a frame timestamp in
-``emitted_at``. Persistent timestamps come from parquet (float32) and ``t``
-is also a float32 from parquet, so in the ideal hot path an exact match
-would suffice — but any caller that derives ``t`` arithmetically (e.g.
-``frame_idx / fps``) breaks bit-equality. A 0.1 s tolerance covers
-common arithmetic drift without admitting frames that are visibly far
-apart at typical control rates (30–100 Hz). This does mean two persistent
-rows of the same selector emitted within 0.1 s of each other cannot be
-told apart by ``emitted_at`` — acceptable because persistent annotations
-(subtask / plan / memory transitions) change on a human-action timescale,
-not at the camera frame rate."""
-
-
-def emitted_at(
-    t: float,
-    *,
-    persistent: Sequence[LanguageRow],
-    events: Sequence[LanguageRow],
-    style: str | None = None,
-    role: str | None = None,
-    tool_name: str | None = None,
-    camera: str | None = None,
-) -> LanguageRow | None:
-    """Return the row of ``style`` emitted at exactly time ``t``.
-
-    For persistent styles, this matches persistent rows whose own ``timestamp``
-    is within ``EMITTED_AT_TOLERANCE_S`` of ``t`` (see that constant for why
-    we use a tolerance instead of bit-equality). For event styles, the
-    ``events`` list is assumed to come from the dataset row at frame ``t``
-    (event rows carry no timestamp of their own), so all matching event rows
-    are considered emitted at ``t``. ``camera`` filters by the row's
-    ``camera`` field — required to disambiguate when multiple view-dependent
-    rows share ``(t, role)`` across cameras.
-    """
-    if column_for_style(style) == LANGUAGE_PERSISTENT:
-        matches = [
-            row
-            for row in _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera)
-            if abs(_timestamp(row) - t) <= EMITTED_AT_TOLERANCE_S
-        ]
-    else:
-        matches = _matching_rows(events, style=style, role=role, tool_name=tool_name, camera=camera)
-    return _select_one(matches, style=style, role=role, tool_name=tool_name, camera=camera)
-
-
-def nth_prev(
-    t: float,
-    *,
-    persistent: Sequence[LanguageRow],
-    style: str | None = None,
-    offset: int = 1,
-    role: str | None = None,
-    tool_name: str | None = None,
-    camera: str | None = None,
-) -> LanguageRow | None:
-    """Return the persistent row that was active ``offset`` steps before ``t``.
-
-    Walks back through chronologically sorted persistent rows of ``style``
-    (filtered by optional ``role``/``tool_name``/``camera``) and returns the
-    one ``offset`` positions before the row active at ``t``. Only valid for
-    persistent styles.
-    """
-    return _nth_relative("nth_prev", t, persistent, style, -offset, role, tool_name, camera)
-
-
-def nth_next(
-    t: float,
-    *,
-    persistent: Sequence[LanguageRow],
-    style: str | None = None,
-    offset: int = 1,
-    role: str | None = None,
-    tool_name: str | None = None,
-    camera: str | None = None,
-) -> LanguageRow | None:
-    """Return the persistent row that becomes active ``offset`` steps after ``t``.
-
-    Walks forward through chronologically sorted persistent rows of ``style``
-    (filtered by optional ``role``/``tool_name``/``camera``) and returns the
-    one ``offset`` positions after the row active at ``t``. Only valid for
-    persistent styles.
-    """
-    return _nth_relative("nth_next", t, persistent, style, offset, role, tool_name, camera)
-
-
-def render_sample(
-    *,
-    recipe: TrainingRecipe,
-    persistent: Sequence[LanguageRow] | None,
-    events: Sequence[LanguageRow] | None,
-    t: float,
-    sample_idx: int,
-    task: str | None = None,
-    dataset_ctx: Any | None = None,
-) -> RenderedMessages | None:
-    """Render the chat-style messages for a single dataset sample.
-
-    Resolves the recipe's bindings against ``persistent`` and ``events`` rows
-    at frame timestamp ``t``, then expands the recipe's message templates.
-    Returns ``None`` if the resolved sample contains no target message.
-    """
-    persistent_rows = _normalize_rows(persistent or [])
-    event_rows = _normalize_rows(events or [])
-    selected_recipe = _select_recipe(recipe, sample_idx)
-    bindings = _resolve_bindings(
-        selected_recipe,
-        persistent=persistent_rows,
-        events=event_rows,
-        t=t,
-        sample_idx=sample_idx,
-        task=task,
-        dataset_ctx=dataset_ctx,
-    )
-    return _render_message_recipe(selected_recipe, bindings)
-
-
-def _select_recipe(recipe: TrainingRecipe, sample_idx: int) -> TrainingRecipe:
-    """Pick a deterministic blend component for ``sample_idx`` (or return ``recipe``)."""
-    if recipe.blend is None:
-        return recipe
-
-    total_weight = sum(component.weight or 0.0 for component in recipe.blend.values())
-    if total_weight <= 0:
-        raise ValueError("Blend weights must sum to a positive value.")
-
-    digest = hashlib.blake2b(str(sample_idx).encode(), digest_size=8).digest()
-    draw = int.from_bytes(digest, "big") / 2**64 * total_weight
-    cumulative = 0.0
-    last_component: TrainingRecipe | None = None
-    for component in recipe.blend.values():
-        last_component = component
-        cumulative += component.weight or 0.0
-        if draw < cumulative:
-            return component
-    assert last_component is not None
-    return last_component
-
-
-def _resolve_bindings(
-    recipe: TrainingRecipe,
-    *,
-    persistent: Sequence[LanguageRow],
-    events: Sequence[LanguageRow],
-    t: float,
-    sample_idx: int,
-    task: str | None,
-    dataset_ctx: Any | None,
-) -> dict[str, LanguageRow | str | None]:
-    """Resolve every binding in ``recipe`` (plus ``task``) at time ``t``."""
-    bindings: dict[str, LanguageRow | str | None] = {
-        "task": _resolve_task(task, dataset_ctx, persistent=persistent, sample_idx=sample_idx),
-    }
-    specs = {**DEFAULT_BINDINGS, **(recipe.bindings or {})}
-    for name, spec in specs.items():
-        bindings[name] = _resolve_spec(spec, persistent=persistent, events=events, t=t)
-    return bindings
-
-
-def _resolve_task(
-    task: str | None,
-    dataset_ctx: Any | None,
-    *,
-    persistent: Sequence[LanguageRow] = (),
-    sample_idx: int = 0,
-) -> str | None:
-    """Return the task string for ``sample_idx``.
-
-    Resolution order:
-
-    1. Explicit ``task`` override (caller-supplied) wins.
-    2. If ``persistent`` contains rows of style ``task_aug`` (role=user),
-       deterministically pick one by ``sample_idx`` so each frame of an
-       episode rotates through the available rephrasings across an epoch.
-       This realizes Xiao 2022 / CAST-style task-prompt diversity without
-       changing ``meta/tasks.parquet`` and without forcing recipes to opt
-       in: ``${task}`` automatically picks a rephrasing when one exists,
-       and falls back to the canonical task otherwise. Recipes that want
-       the literal canonical task can override the binding.
-    3. Otherwise read the canonical task from ``dataset_ctx`` (which is
-       backed by ``meta/tasks.parquet``).
-    """
-    if task is not None:
-        return task
-
-    aug_rows = [r for r in persistent if r.get("style") == "task_aug" and r.get("role") == "user"]
-    if aug_rows:
-        # Deterministic, blake2b-based pick keyed on sample_idx so the
-        # rotation is reproducible across runs (Python's built-in ``hash``
-        # is process-randomized).
-        digest = hashlib.blake2b(f"task_aug:{sample_idx}".encode(), digest_size=8).digest()
-        idx = int.from_bytes(digest, "big") % len(aug_rows)
-        chosen = aug_rows[idx].get("content")
-        if chosen:
-            return str(chosen)
-
-    if dataset_ctx is None:
-        return None
-    if isinstance(dataset_ctx, dict):
-        return dataset_ctx.get("task")
-    return getattr(dataset_ctx, "task", None)
-
-
-def _resolve_spec(
-    spec: str,
-    *,
-    persistent: Sequence[LanguageRow],
-    events: Sequence[LanguageRow],
-    t: float,
-) -> LanguageRow | None:
-    """Parse a single binding's resolver expression and dispatch to its function."""
-    match = _RESOLVER_RE.match(spec.strip())
-    if match is None:
-        raise ValueError(f"Invalid resolver expression: {spec!r}")
-    name = match.group("name")
-    kwargs = _parse_resolver_args(match.group("args"))
-    kwargs.pop("t_arg", None)
-
-    if name == "emitted_at":
-        return emitted_at(t, persistent=persistent, events=events, **kwargs)
-    if name == "active_at":
-        return active_at(t, persistent=persistent, **kwargs)
-    if name == "nth_prev":
-        return nth_prev(t, persistent=persistent, **kwargs)
-    if name == "nth_next":
-        return nth_next(t, persistent=persistent, **kwargs)
-    raise ValueError(f"Unknown language resolver: {name!r}")
-
-
-def _parse_resolver_args(args: str) -> dict[str, Any]:
-    """Parse a comma-separated resolver argument list into a kwargs dict."""
-    kwargs: dict[str, Any] = {}
-    if not args.strip():
-        return kwargs
-
-    parts = [part.strip() for part in args.split(",") if part.strip()]
-    for part in parts:
-        if part == "t":
-            kwargs["t_arg"] = True
-            continue
-        if "=" not in part:
-            raise ValueError(f"Invalid resolver argument: {part!r}")
-        key, value = (item.strip() for item in part.split("=", 1))
-        if key == "offset":
-            kwargs[key] = int(value)
-        else:
-            kwargs[key] = value.strip("\"'")
-    return kwargs
-
-
-def _render_message_recipe(
-    recipe: TrainingRecipe,
-    bindings: dict[str, LanguageRow | str | None],
-) -> RenderedMessages | None:
-    """Expand ``recipe.messages`` into rendered chat messages using ``bindings``."""
-    assert recipe.messages is not None
-    messages: list[dict[str, Any]] = []
-    streams: list[str | None] = []
-    target_indices: list[int] = []
-
-    for turn in recipe.messages:
-        if turn.if_present is not None and bindings.get(turn.if_present) is None:
-            continue
-
-        message = {"role": turn.role}
-        if turn.content is not None:
-            message["content"] = _render_content(turn.content, bindings)
-
-        if turn.tool_calls_from is not None:
-            row = bindings.get(turn.tool_calls_from)
-            tool_calls = row.get("tool_calls") if isinstance(row, dict) else None
-            if tool_calls:
-                message["tool_calls"] = copy.deepcopy(tool_calls)
-
-        message_idx = len(messages)
-        messages.append(message)
-        streams.append(turn.stream)
-        if turn.target:
-            target_indices.append(message_idx)
-
-    if not target_indices:
-        return None
-
-    rendered = {
-        "messages": messages,
-        "message_streams": streams,
-        "target_message_indices": target_indices,
-    }
-    _validate_rendered(rendered)
-    return rendered
-
-
-def _render_content(
-    content: str | list[dict[str, Any]],
-    bindings: dict[str, LanguageRow | str | None],
-) -> str | list[dict[str, Any]]:
-    """Substitute bindings into a string or each string field of multimodal blocks."""
-    if isinstance(content, str):
-        return _substitute(content, bindings)
-
-    rendered_blocks = []
-    for block in content:
-        rendered_block = copy.deepcopy(block)
-        for key, value in rendered_block.items():
-            if isinstance(value, str):
-                rendered_block[key] = _substitute(value, bindings)
-        rendered_blocks.append(rendered_block)
-    return rendered_blocks
-
-
-def _substitute(template: str, bindings: dict[str, LanguageRow | str | None]) -> str:
-    """Replace ``${name}`` placeholders in ``template`` with their bound values."""
-
-    def replace(match: re.Match[str]) -> str:
-        """Resolve a single ``${name}`` match to its bound string value."""
-        name = match.group(1)
-        if name not in bindings:
-            raise ValueError(f"Unknown template binding: {name!r}")
-        value = bindings[name]
-        if value is None:
-            return ""
-        if isinstance(value, dict):
-            content = value.get("content")
-            return "" if content is None else str(content)
-        return str(value)
-
-    return PLACEHOLDER_RE.sub(replace, template)
-
-
-def _validate_rendered(rendered: RenderedMessages) -> None:
-    """Sanity-check the rendered output for stream/target alignment."""
-    messages = rendered["messages"]
-    streams = rendered["message_streams"]
-    target_indices = rendered["target_message_indices"]
-
-    if len(streams) != len(messages):
-        raise ValueError("message_streams must be aligned with messages.")
-    if not target_indices:
-        raise ValueError("Rendered samples must contain at least one target message.")
-    for idx in target_indices:
-        if idx < 0 or idx >= len(messages):
-            raise ValueError(f"Target message index {idx} is out of bounds.")
-    # ``stream`` is enforced non-None at MessageTurn construction time
-    # (see ``MessageTurn.__post_init__``), so a missing stream here would
-    # mean the dataclass invariant was bypassed; no need to re-check.
-
-
-def _nth_relative(
-    name: str,
-    t: float,
-    persistent: Sequence[LanguageRow],
-    style: str | None,
-    offset: int,
-    role: str | None,
-    tool_name: str | None,
-    camera: str | None,
-) -> LanguageRow | None:
-    """Shared body for ``nth_prev`` / ``nth_next`` with signed ``offset``."""
-    _validate_persistent_resolver(name, style)
-    if abs(offset) < 1:
-        raise ValueError(f"{name} offset must be non-zero.")
-
-    rows = sorted(
-        _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera),
-        key=_row_sort_key,
-    )
-    if not rows:
-        return None
-
-    anchor_idx = None
-    for idx, row in enumerate(rows):
-        if _timestamp(row) <= t:
-            anchor_idx = idx
-        else:
-            break
-
-    target_idx = (offset - 1 if offset > 0 else None) if anchor_idx is None else anchor_idx + offset
-
-    if target_idx is None or target_idx < 0 or target_idx >= len(rows):
-        return None
-    return rows[target_idx]
-
-
-def _validate_persistent_resolver(name: str, style: str | None) -> None:
-    """Reject calls with missing or event-only ``style`` for persistent resolvers."""
-    if style is None:
-        raise ValueError(f"{name} requires a persistent style.")
-    if column_for_style(style) != LANGUAGE_PERSISTENT:
-        raise ValueError(f"{name} cannot be used with event-only style {style!r}.")
-
-
-def _matching_rows(
-    rows: Sequence[LanguageRow],
-    *,
-    style: str | None,
-    role: str | None,
-    tool_name: str | None,
-    camera: str | None,
-) -> list[LanguageRow]:
-    """Return ``rows`` filtered by optional ``style``/``role``/``tool_name``/``camera`` selectors."""
-    return [
-        row
-        for row in rows
-        if (style is None or row.get("style") == style)
-        and (role is None or row.get("role") == role)
-        and (tool_name is None or _row_has_tool_name(row, tool_name))
-        and (camera is None or row.get("camera") == camera)
-    ]
-
-
-def _select_one(
-    rows: Sequence[LanguageRow],
-    *,
-    style: str | None,
-    role: str | None,
-    tool_name: str | None,
-    camera: str | None,
-) -> LanguageRow | None:
-    """Return the single matching row, or raise if the resolver is ambiguous.
-
-    Multiple matches always raise — even when the caller already passed
-    some selectors — because remaining ambiguity means the data has
-    several rows that look identical to the resolver and the caller
-    needs to pin down a specific one (e.g. add ``camera=...`` for VQA
-    rows shared across cameras).
-    """
-    if not rows:
-        return None
-    if len(rows) > 1:
-        raise ValueError(
-            f"Ambiguous resolver for style={style!r} role={role!r} "
-            f"tool_name={tool_name!r} camera={camera!r}: {len(rows)} matching rows. "
-            f"Add a selector that distinguishes them."
-        )
-    return rows[0]
-
-
-def _row_sort_key(row: LanguageRow) -> tuple[float, str, str]:
-    """Stable sort key for both persistent and event rows.
-
-    Event rows lack ``timestamp`` (it is implicit in the frame), so default
-    to ``0.0`` — within a single frame all event rows share the same sort
-    bucket and are tiebroken by ``(style, role)``.
-    """
-    timestamp = row.get("timestamp")
-    ts = float(unwrap_scalar(timestamp)) if timestamp is not None else 0.0
-    return (ts, row.get("style") or "", row.get("role") or "")
-
-
-def _timestamp(row: LanguageRow) -> float:
-    """Extract a row's ``timestamp`` as a Python float (unwrapping numpy scalars)."""
-    return float(unwrap_scalar(row["timestamp"]))
-
-
-def _row_has_tool_name(row: LanguageRow, tool_name: str) -> bool:
-    """Return ``True`` if any of the row's tool calls invokes ``tool_name``."""
-    for tool_call in row.get("tool_calls") or []:
-        if isinstance(tool_call, str):
-            continue
-        function = tool_call.get("function") if isinstance(tool_call, dict) else None
-        if isinstance(function, dict) and function.get("name") == tool_name:
-            return True
-    return False
-
-
-def _normalize_rows(rows: Sequence[Any]) -> list[LanguageRow]:
-    """Convert pyarrow scalars / mappings into a fresh list of plain dict rows."""
-    normalized = []
-    for row in rows:
-        if row is None:
-            continue
-        if hasattr(row, "as_py"):
-            row = row.as_py()
-        if not isinstance(row, dict):
-            raise TypeError(f"Language rows must be dictionaries, got {type(row).__name__}.")
-        normalized.append(dict(row))
-    return normalized
--- a/src/lerobot/datasets/lerobot_dataset.py
+++ b/src/lerobot/datasets/lerobot_dataset.py
@@ -24,7 +24,6 @@ import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError

-from lerobot.configs import VideoEncoderConfig
 from lerobot.utils.constants import HF_LEROBOT_HUB_CACHE

 from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
@@ -37,7 +36,8 @@ from .utils import (
 )
 from .video_utils import (
    StreamingVideoEncoder,
-    get_safe_default_video_backend,
+    get_safe_default_codec,
+    resolve_vcodec,
 )

 logger = logging.getLogger(__name__)
@@ -49,7 +49,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        repo_id: str,
        root: str | Path | None = None,
        episodes: list[int] | None = None,
-        episode_filter: Callable[[dict], bool] | None = None,
        image_transforms: Callable | None = None,
        delta_timestamps: dict[str, list[float]] | None = None,
        tolerance_s: float = 1e-4,
@@ -59,10 +58,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
        video_backend: str | None = None,
        return_uint8: bool = False,
        batch_encoding_size: int = 1,
-        camera_encoder: VideoEncoderConfig | None = None,
-        encoder_threads: int | None = None,
+        vcodec: str = "libsvtav1",
        streaming_encoding: bool = False,
        encoder_queue_maxsize: int = 30,
+        encoder_threads: int | None = None,
    ):
        """
        2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -154,11 +153,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
                ``$HF_LEROBOT_HOME/hub``.
            episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                their episode_index in this list. Defaults to None.
-            episode_filter (Callable[[dict], bool] | None, optional): Predicate over per-episode
-                metadata rows used to select episodes. Evaluated against ``meta/`` without ``stats`` keys
-                (e.g.``task_index``, ``episode_index``, ``length``, ``from_timestamp``, ``to_timestamp``).
-                Intersected with ``episodes`` when both are set. Example: ``lambda ep: ep["length"] >= 100``.
-                Defaults to None.
            image_transforms (Callable | None, optional):
                Transform applied to visual modalities inside `__getitem__` after image decoding / tensor
                conversion. This works for both image-backed and video-backed observations and can later be
@@ -183,15 +177,16 @@ class LeRobotDataset(torch.utils.data.Dataset):
                You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision.
            batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
-            camera_encoder (VideoEncoderConfig | None, optional): Video encoder settings for cameras
-                (codec, quality, etc.). When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults`
-                is used by the writer.
-            encoder_threads (int | None, optional): Number of encoder threads (global). ``None`` lets the
-                codec decide.
+            vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
+                'libsvtav1', 'auto', or hardware-specific codecs like 'h264_videotoolbox', 'h264_nvenc'.
+                Defaults to 'libsvtav1'. Use 'auto' to auto-detect the best available hardware encoder.
            streaming_encoding (bool, optional): If True, encode video frames in real-time during capture
                instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
            encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
                streaming encoding. Defaults to 30 (~1s at 30fps).
+            encoder_threads (int | None, optional): Number of threads per encoder instance. None lets the
+                codec auto-detect (default). Lower values reduce CPU usage per encoder. Maps to 'lp' (via svtav1-params) for
+                libsvtav1 and 'threads' for h264/hevc.

        Note:
            Write-mode parameters (``streaming_encoding``, ``batch_encoding_size``) passed to
@@ -204,11 +199,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
        self.reader = None
        self.set_image_transforms(image_transforms)
        self.delta_timestamps = delta_timestamps
+        self.episodes = episodes
        self.tolerance_s = tolerance_s
        self.revision = revision if revision else CODEBASE_VERSION
-        self._video_backend = video_backend if video_backend else get_safe_default_video_backend()
+        self._video_backend = video_backend if video_backend else get_safe_default_codec()
        self._return_uint8 = return_uint8
        self._batch_encoding_size = batch_encoding_size
+        self._vcodec = resolve_vcodec(vcodec)
        self._encoder_threads = encoder_threads

        if self._requested_root is not None:
@@ -221,23 +218,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        self.root = self.meta.root
        self.revision = self.meta.revision

-        if episodes is not None and any(
-            episode >= self.meta.total_episodes or episode < 0 for episode in episodes
-        ):
-            logger.warning(
-                f"Some episodes in the provided episodes list are out of range for this dataset ({self.meta.total_episodes})."
-            )
-
-        if episode_filter is not None:
-            resolved = self.meta.filter_episodes(episode_filter, candidates=episodes)
-            if not resolved:
-                raise ValueError(
-                    "The episode filter did not match any episode. Make sure the filter and episodes list are valid and compatible."
-                )
-            logger.info(f"The episode filter matched {len(resolved)} episode(s).")
-            episodes = resolved
-        self.episodes = episodes
-
        # Create reader (hf_dataset loaded below)
        self.reader = DatasetReader(
            meta=self.meta,
@@ -271,15 +251,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
            streaming_enc = None
            if streaming_encoding and len(self.meta.video_keys) > 0:
                streaming_enc = self._build_streaming_encoder(
-                    self.meta.fps,
-                    camera_encoder,
-                    encoder_queue_maxsize,
-                    encoder_threads,
+                    self.meta.fps, self._vcodec, encoder_queue_maxsize, encoder_threads
                )
            self.writer = DatasetWriter(
                meta=self.meta,
                root=self.root,
-                camera_encoder=camera_encoder,
+                vcodec=self._vcodec,
                encoder_threads=encoder_threads,
                batch_encoding_size=batch_encoding_size,
                streaming_encoder=streaming_enc,
@@ -321,13 +298,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
    @staticmethod
    def _build_streaming_encoder(
        fps: int,
-        camera_encoder: VideoEncoderConfig | None,
+        vcodec: str,
        encoder_queue_maxsize: int,
        encoder_threads: int | None,
    ) -> StreamingVideoEncoder:
        return StreamingVideoEncoder(
            fps=fps,
-            camera_encoder=camera_encoder,
+            vcodec=vcodec,
+            pix_fmt="yuv420p",
+            g=2,
+            crf=30,
+            preset=None,
            queue_maxsize=encoder_queue_maxsize,
            encoder_threads=encoder_threads,
        )
@@ -644,13 +625,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
        image_writer_threads: int = 0,
        video_backend: str | None = None,
        batch_encoding_size: int = 1,
-        camera_encoder: VideoEncoderConfig | None = None,
+        vcodec: str = "libsvtav1",
        metadata_buffer_size: int = 10,
        streaming_encoding: bool = False,
        encoder_queue_maxsize: int = 30,
        encoder_threads: int | None = None,
-        video_files_size_in_mb: int | None = None,
-        data_files_size_in_mb: int | None = None,
    ) -> "LeRobotDataset":
        """Create a new LeRobotDataset from scratch for recording data.

@@ -675,20 +654,20 @@ class LeRobotDataset(torch.utils.data.Dataset):
            video_backend: Video decoding backend (used when reading back).
            batch_encoding_size: Number of episodes to accumulate before
                batch-encoding videos. ``1`` means encode immediately.
-            camera_encoder: Video encoder settings for cameras (codec, quality, etc.).
-                When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` is used.
-            encoder_threads: Number of encoder threads (global). ``None``
-                lets the codec decide.
+            vcodec: Video codec for encoding. Options include ``'libsvtav1'``,
+                ``'h264'``, ``'hevc'``, ``'auto'``.
            metadata_buffer_size: Number of episode metadata records to buffer
                before flushing to parquet.
            streaming_encoding: If ``True``, encode video frames in real-time
                during capture instead of writing images first.
            encoder_queue_maxsize: Max buffered frames per camera when using
                streaming encoding.
+            encoder_threads: Threads per encoder instance. ``None`` for auto.

        Returns:
            A new :class:`LeRobotDataset` in write mode.
        """
+        vcodec = resolve_vcodec(vcodec)
        obj = cls.__new__(cls)
        obj.meta = LeRobotDatasetMetadata.create(
            repo_id=repo_id,
@@ -698,8 +677,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
            root=root,
            use_videos=use_videos,
            metadata_buffer_size=metadata_buffer_size,
-            video_files_size_in_mb=video_files_size_in_mb,
-            data_files_size_in_mb=data_files_size_in_mb,
        )
        obj.repo_id = obj.meta.repo_id
        obj._requested_root = obj.meta.root
@@ -709,23 +686,23 @@ class LeRobotDataset(torch.utils.data.Dataset):
        obj.image_transforms = None
        obj.delta_timestamps = None
        obj.episodes = None
-        obj._video_backend = video_backend if video_backend is not None else get_safe_default_video_backend()
+        obj._video_backend = video_backend if video_backend is not None else get_safe_default_codec()
        obj._return_uint8 = False
        obj._batch_encoding_size = batch_encoding_size
+        obj._vcodec = vcodec
        obj._encoder_threads = encoder_threads

        # Reader is lazily created on first access (write-only mode)
        obj.reader = None

+        # Create writer
        streaming_enc = None
        if streaming_encoding and len(obj.meta.video_keys) > 0:
-            streaming_enc = cls._build_streaming_encoder(
-                fps, camera_encoder, encoder_queue_maxsize, encoder_threads
-            )
+            streaming_enc = cls._build_streaming_encoder(fps, vcodec, encoder_queue_maxsize, encoder_threads)
        obj.writer = DatasetWriter(
            meta=obj.meta,
            root=obj.root,
-            camera_encoder=camera_encoder,
+            vcodec=vcodec,
            encoder_threads=encoder_threads,
            batch_encoding_size=batch_encoding_size,
            streaming_encoder=streaming_enc,
@@ -748,12 +725,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
        force_cache_sync: bool = False,
        video_backend: str | None = None,
        batch_encoding_size: int = 1,
-        camera_encoder: VideoEncoderConfig | None = None,
-        encoder_threads: int | None = None,
+        vcodec: str = "libsvtav1",
        image_writer_processes: int = 0,
        image_writer_threads: int = 0,
        streaming_encoding: bool = False,
        encoder_queue_maxsize: int = 30,
+        encoder_threads: int | None = None,
    ) -> "LeRobotDataset":
        """Resume recording on an existing dataset.

@@ -776,15 +753,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
            video_backend: Video decoding backend for reading back data.
            batch_encoding_size: Number of episodes to accumulate before
                batch-encoding videos.
-            camera_encoder: Video encoder settings for cameras (codec, quality, etc.).
-                When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` is used.
-            encoder_threads: Number of encoder threads (global). ``None``
-                lets the codec decide.
+            vcodec: Video codec for encoding.
            image_writer_processes: Subprocesses for async image writing.
            image_writer_threads: Threads for async image writing.
            streaming_encoding: If ``True``, encode video in real-time during
                capture.
            encoder_queue_maxsize: Max buffered frames per camera for streaming.
+            encoder_threads: Threads per encoder instance. ``None`` for auto.

        Returns:
            A :class:`LeRobotDataset` in write mode, ready to append episodes.
@@ -795,6 +770,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                "Writing into the revision-safe Hub snapshot cache (used when root=None) would corrupt "
                "the shared cache. Please provide a local directory path."
            )
+        vcodec = resolve_vcodec(vcodec)
        obj = cls.__new__(cls)
        obj.repo_id = repo_id
        obj._requested_root = Path(root)
@@ -803,9 +779,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
        obj.image_transforms = None
        obj.delta_timestamps = None
        obj.episodes = None
-        obj._video_backend = video_backend if video_backend else get_safe_default_video_backend()
+        obj._video_backend = video_backend if video_backend else get_safe_default_codec()
        obj._return_uint8 = False
        obj._batch_encoding_size = batch_encoding_size
+        obj._vcodec = vcodec
+        obj._encoder_threads = encoder_threads

        if obj._requested_root is not None:
            obj._requested_root.mkdir(exist_ok=True, parents=True)
@@ -814,22 +792,21 @@ class LeRobotDataset(torch.utils.data.Dataset):
        obj.meta = LeRobotDatasetMetadata(
            obj.repo_id, obj._requested_root, obj.revision, force_cache_sync=force_cache_sync
        )
-
-        obj._encoder_threads = encoder_threads
        obj.root = obj.meta.root

        # Reader is lazily created on first access (write-only mode)
        obj.reader = None

+        # Create writer for appending
        streaming_enc = None
        if streaming_encoding and len(obj.meta.video_keys) > 0:
            streaming_enc = cls._build_streaming_encoder(
-                obj.meta.fps, camera_encoder, encoder_queue_maxsize, encoder_threads
+                obj.meta.fps, vcodec, encoder_queue_maxsize, encoder_threads
            )
        obj.writer = DatasetWriter(
            meta=obj.meta,
            root=obj.root,
-            camera_encoder=camera_encoder,
+            vcodec=vcodec,
            encoder_threads=encoder_threads,
            batch_encoding_size=batch_encoding_size,
            streaming_encoder=streaming_enc,
--- a/src/lerobot/datasets/multi_dataset.py
+++ b/src/lerobot/datasets/multi_dataset.py
@@ -123,7 +123,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):

        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
        """
-        return self._datasets[0].meta.info.fps
+        return self._datasets[0].meta.info["fps"]

    @property
    def video(self) -> bool:
@@ -133,7 +133,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):

        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
        """
-        return len(self._datasets[0].meta.video_keys) > 0
+        return self._datasets[0].meta.info.get("video", False)

    @property
    def features(self) -> datasets.Features:
--- a/src/lerobot/datasets/pyav_utils.py
+++ b/src/lerobot/datasets/pyav_utils.py
@@ -1,174 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""PyAV-based compatibility checks for :class:`VideoEncoderConfig`.
-
-Centralises all :mod:`av` introspection of the bundled FFmpeg build.
-Checks degrade to a no-op when the target codec isn't available locally.
-"""
-
-import functools
-import logging
-from typing import Any
-
-import av
-
-logger = logging.getLogger(__name__)
-
-FFMPEG_NUMERIC_OPTION_TYPES = ("INT", "INT64", "UINT64", "FLOAT", "DOUBLE")
-FFMPEG_INTEGER_OPTION_TYPES = ("INT", "INT64", "UINT64")
-
-
-@functools.cache
-def get_codec(vcodec: str) -> av.codec.Codec | None:
-    """PyAV write-mode ``Codec`` for *vcodec*, or ``None`` if unavailable."""
-    try:
-        return av.codec.Codec(vcodec, "w")
-    except Exception:
-        return None
-
-
-@functools.cache
-def _get_codec_options_by_name(vcodec: str) -> dict[str, av.option.Option]:
-    """Private-option name → PyAV ``Option`` for *vcodec* (empty if unavailable)."""
-    codec = get_codec(vcodec)
-    if codec is None:
-        return {}
-    return {opt.name: opt for opt in codec.descriptor.options}
-
-
-@functools.cache
-def _get_codec_video_formats(vcodec: str) -> tuple[str, ...]:
-    """Pixel formats accepted by *vcodec* in PyAV's preferred order (empty if unknown)."""
-    codec = get_codec(vcodec)
-    if codec is None:
-        return ()
-    return tuple(fmt.name for fmt in (codec.video_formats or []))
-
-
-def detect_available_encoders_pyav(encoders: list[str] | str) -> list[str]:
-    """Return the subset of *encoders* available as video encoders in the local FFmpeg build.
-
-    Each name is probed directly via :func:`get_codec`; input order is preserved.
-    """
-    if isinstance(encoders, str):
-        encoders = [encoders]
-
-    available: list[str] = []
-    for name in encoders:
-        codec = get_codec(name)
-        if codec is not None and codec.type == "video":
-            available.append(name)
-        else:
-            logger.debug("encoder '%s' not available as video encoder", name)
-    return available
-
-
-def _check_option_value(vcodec: str, label: str, value: Any, opt: av.option.Option) -> None:
-    """Range-check numeric *value* and choice-check string *value* against *opt*."""
-    type_name = opt.type.name
-    if type_name in FFMPEG_NUMERIC_OPTION_TYPES:
-        if isinstance(value, bool):
-            raise ValueError(
-                f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
-            )
-        elif isinstance(value, str):
-            try:
-                num_val = float(value)
-            except ValueError as e:
-                raise ValueError(
-                    f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
-                ) from e
-        elif isinstance(value, (float, int)):
-            num_val = value
-        else:
-            raise ValueError(
-                f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
-            )
-
-        # Check integer type compatibility
-        if type_name in FFMPEG_INTEGER_OPTION_TYPES and not num_val.is_integer():
-            raise ValueError(
-                f"{label}={num_val!r} must be an integer for codec {vcodec!r} "
-                f"(FFmpeg option {opt.name!r} is {type_name}); float values are not allowed."
-            )
-
-        # Check numeric range compatibility
-        lo, hi = float(opt.min), float(opt.max)
-        if lo < hi and not (lo <= num_val <= hi):
-            raise ValueError(
-                f"{label}={num_val} is out of range for codec {vcodec!r}; must be in [{lo}, {hi}]"
-            )
-
-    elif type_name == "STRING":
-        if isinstance(value, bool):
-            raise ValueError(f"{label}={value!r} is not a valid string value for codec {vcodec!r}.")
-        if isinstance(value, str):
-            str_val = value
-        elif isinstance(value, (int, float)):
-            str_val = str(value)
-        else:
-            raise ValueError(f"{label}={value!r} has unsupported type for STRING option on codec {vcodec!r}")
-
-        # Check string choice compatibility
-        choices = [c.name for c in (opt.choices or [])]
-        if choices and str_val not in choices:
-            raise ValueError(
-                f"{label}={str_val!r} is not a supported choice for codec "
-                f"{vcodec!r}; valid choices: {choices}"
-            )
-    else:
-        return
-
-
-def _check_pixel_format(vcodec: str, pix_fmt: str) -> None:
-    formats = _get_codec_video_formats(vcodec)
-    if formats and pix_fmt not in formats:
-        raise ValueError(
-            f"pix_fmt={pix_fmt!r} is not supported by codec {vcodec!r}; "
-            f"supported pixel formats: {list(formats)}"
-        )
-
-
-def _check_codec_options(vcodec: str, codec_options: dict[str, Any]) -> None:
-    """Validate merged encoder options (typed) against the codec's published AVOptions."""
-    supported_options = _get_codec_options_by_name(vcodec)
-    for key, value in codec_options.items():
-        # GOP size is not a codec-specific option, it has to be validated separately.
-        if key == "g":
-            if isinstance(value, bool) or not isinstance(value, int) or value < 1:
-                raise ValueError(f"g={value!r} must be a positive integer for codec {vcodec!r}")
-            continue
-        if key not in supported_options:
-            continue
-        _check_option_value(vcodec, key, value, supported_options[key])
-
-
-def check_video_encoder_parameters_pyav(vcodec: str, pix_fmt: str, codec_options: dict[str, Any]) -> None:
-    """Verify *config* is compatible with the bundled FFmpeg build.
-
-    Checks pixel format, abstract tuning-field compatibility, and each merged
-    encoder option from :meth:`~lerobot.configs.video.VideoEncoderConfig.get_codec_options`
-    against PyAV (including numeric ``extra_options`` present in that dict).
-    No-op when ``config.vcodec`` isn't in the local FFmpeg build.
-
-    Raises:
-        ValueError: on the first incompatibility encountered.
-    """
-    options = _get_codec_options_by_name(vcodec)
-    if not options:
-        raise ValueError(f"Codec {vcodec!r} is not available in the bundled FFmpeg build")
-    _check_pixel_format(vcodec, pix_fmt)
-    _check_codec_options(vcodec, codec_options)
--- a/src/lerobot/datasets/streaming_dataset.py
+++ b/src/lerobot/datasets/streaming_dataset.py
@@ -434,7 +434,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):

    def _make_padding_camera_frame(self, camera_key: str):
        """Variable-shape padding frame for given camera keys, given in (H, W, C)"""
-        return torch.zeros(self.meta.info.features[camera_key]["shape"]).permute(-1, 0, 1)
+        return torch.zeros(self.meta.info["features"][camera_key]["shape"]).permute(-1, 0, 1)

    def _get_video_frame_padding_mask(
        self,
--- a/src/lerobot/datasets/utils.py
+++ b/src/lerobot/datasets/utils.py
@@ -14,11 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import contextlib
-import dataclasses
 import importlib.resources
 import json
 import logging
-from dataclasses import dataclass, field
 from pathlib import Path

 import datasets
@@ -72,12 +70,9 @@ class ForwardCompatibilityError(CompatibilityError):
        super().__init__(message)


-logger = logging.getLogger(__name__)
-
-
 DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
-DEFAULT_DATA_FILE_SIZE_IN_MB = 100  # Max size per file
-DEFAULT_VIDEO_FILE_SIZE_IN_MB = 200  # Max size per file
+DEFAULT_DATA_FILE_SIZE_IN_MB = 50  # Max size per file
+DEFAULT_VIDEO_FILE_SIZE_IN_MB = 100  # Max size per file

 INFO_PATH = "meta/info.json"
 STATS_PATH = "meta/stats.json"
@@ -88,6 +83,7 @@ VIDEO_DIR = "videos"

 CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
 DEFAULT_TASKS_PATH = "meta/tasks.parquet"
+DEFAULT_SUBTASKS_PATH = "meta/subtasks.parquet"
 DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
@@ -98,130 +94,6 @@ LEGACY_EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
 LEGACY_TASKS_PATH = "meta/tasks.jsonl"


-@dataclass
-class DatasetInfo:
-    """Typed representation of the ``meta/info.json`` file for a LeRobot dataset.
-
-    Replaces the previously untyped ``dict`` returned by ``load_info()`` and
-    created by ``create_empty_dataset_info()``.  Using a dataclass provides
-    explicit field definitions, IDE auto-completion, and validation at
-    construction time.
-    """
-
-    codebase_version: str
-    fps: int
-    features: dict[str, dict]
-
-    # Episode / frame counters — start at zero for new datasets
-    total_episodes: int = 0
-    total_frames: int = 0
-    total_tasks: int = 0
-
-    # Storage settings
-    chunks_size: int = field(default=DEFAULT_CHUNK_SIZE)
-    data_files_size_in_mb: int = field(default=DEFAULT_DATA_FILE_SIZE_IN_MB)
-    video_files_size_in_mb: int = field(default=DEFAULT_VIDEO_FILE_SIZE_IN_MB)
-
-    # File path templates
-    data_path: str = field(default=DEFAULT_DATA_PATH)
-    video_path: str | None = field(default=DEFAULT_VIDEO_PATH)
-
-    # Optional metadata
-    robot_type: str | None = None
-    splits: dict[str, str] = field(default_factory=dict)
-    # OpenAI-style tool schemas declared by the dataset. ``None`` means the
-    # dataset doesn't declare any — readers fall back to ``DEFAULT_TOOLS``.
-    tools: list[dict] | None = None
-
-    def __post_init__(self) -> None:
-        # Coerce feature shapes from list to tuple — JSON deserialisation
-        # returns lists, but the rest of the codebase expects tuples.
-        for ft in self.features.values():
-            if isinstance(ft.get("shape"), list):
-                ft["shape"] = tuple(ft["shape"])
-
-        if self.fps <= 0:
-            raise ValueError(f"fps must be positive, got {self.fps}")
-        if self.chunks_size <= 0:
-            raise ValueError(f"chunks_size must be positive, got {self.chunks_size}")
-        if self.data_files_size_in_mb <= 0:
-            raise ValueError(f"data_files_size_in_mb must be positive, got {self.data_files_size_in_mb}")
-        if self.video_files_size_in_mb <= 0:
-            raise ValueError(f"video_files_size_in_mb must be positive, got {self.video_files_size_in_mb}")
-
-    def to_dict(self) -> dict:
-        """Return a JSON-serialisable dict.
-
-        Converts tuple shapes back to lists so ``json.dump`` can handle them.
-        Drops ``tools`` when unset so existing datasets keep a clean
-        ``info.json``.
-        """
-        d = dataclasses.asdict(self)
-        for ft in d["features"].values():
-            if isinstance(ft.get("shape"), tuple):
-                ft["shape"] = list(ft["shape"])
-        if d.get("tools") is None:
-            d.pop("tools", None)
-        return d
-
-    @classmethod
-    def from_dict(cls, data: dict) -> "DatasetInfo":
-        """Construct from a raw dict (e.g. loaded directly from JSON).
-
-        Unknown keys are ignored for forward compatibility with datasets that
-        carry additional fields (e.g. ``total_videos`` from v2.x). A warning is
-        logged when such fields are present.
-        """
-        known = {f.name for f in dataclasses.fields(cls)}
-        unknown = sorted(k for k in data if k not in known)
-        if unknown:
-            logger.warning(f"Unknown fields in DatasetInfo: {unknown}. These will be ignored.")
-        return cls(**{k: v for k, v in data.items() if k in known})
-
-    # ---------------------------------------------------------------------------
-    # Temporary dict-style compatibility layer
-    # Allows existing ``info["key"]`` call-sites to keep working without changes.
-    # Once all callers have been migrated to attribute access, remove these.
-    # ---------------------------------------------------------------------------
-    def __getitem__(self, key: str):
-        import warnings
-
-        warnings.warn(
-            f"Accessing DatasetInfo with dict-style syntax info['{key}'] is deprecated. "
-            f"Use attribute access info.{key} instead.",
-            DeprecationWarning,
-            stacklevel=2,
-        )
-        try:
-            return getattr(self, key)
-        except AttributeError as err:
-            raise KeyError(key) from err
-
-    def __setitem__(self, key: str, value) -> None:
-        import warnings
-
-        warnings.warn(
-            f"Setting DatasetInfo with dict-style syntax info['{key}'] = ... is deprecated. "
-            f"Use attribute assignment info.{key} = ... instead.",
-            DeprecationWarning,
-            stacklevel=2,
-        )
-        if not hasattr(self, key):
-            raise KeyError(f"DatasetInfo has no field '{key}'")
-        setattr(self, key, value)
-
-    def __contains__(self, key: str) -> bool:
-        """Check if a field exists (dict-like interface)."""
-        return hasattr(self, key)
-
-    def get(self, key: str, default=None):
-        """Get attribute value with default fallback (dict-like interface)."""
-        try:
-            return getattr(self, key)
-        except AttributeError:
-            return default
-
-
 def has_legacy_hub_download_metadata(root: Path) -> bool:
    """Return ``True`` when *root* looks like a legacy Hub ``local_dir`` mirror.

@@ -422,7 +294,7 @@ def create_branch(repo_id: str, *, branch: str, repo_type: str | None = None) ->

 def create_lerobot_dataset_card(
    tags: list | None = None,
-    dataset_info: DatasetInfo | None = None,
+    dataset_info: dict | None = None,
    **kwargs,
 ) -> DatasetCard:
    """Create a `DatasetCard` for a LeRobot dataset.
@@ -433,7 +305,7 @@ def create_lerobot_dataset_card(

    Args:
        tags (list | None): A list of tags to add to the dataset card.
-        dataset_info (DatasetInfo | None): The dataset's info object, which will
+        dataset_info (dict | None): The dataset's info dictionary, which will
            be displayed on the card.
        **kwargs: Additional keyword arguments to populate the card template.

@@ -446,7 +318,7 @@ def create_lerobot_dataset_card(
        card_tags += tags
    if dataset_info:
        dataset_structure = "[meta/info.json](meta/info.json):\n"
-        dataset_structure += f"```json\n{json.dumps(dataset_info.to_dict(), indent=4)}\n```\n"
+        dataset_structure += f"```json\n{json.dumps(dataset_info, indent=4)}\n```\n"
        kwargs = {**kwargs, "dataset_structure": dataset_structure}
    card_data = DatasetCardData(
        license=kwargs.get("license"),
--- a/src/lerobot/datasets/video_utils.py
+++ b/src/lerobot/datasets/video_utils.py
@@ -17,14 +17,12 @@ import contextlib
 import glob
 import importlib
 import logging
-import os
 import queue
 import shutil
 import tempfile
 import threading
 import warnings
-from collections import OrderedDict
-from dataclasses import asdict, dataclass, field
+from dataclasses import dataclass, field
 from fractions import Fraction
 from pathlib import Path
 from threading import Lock
@@ -35,17 +33,90 @@ import fsspec
 import numpy as np
 import pyarrow as pa
 import torch
+import torchvision
 from datasets.features.features import register_feature
 from PIL import Image

-from lerobot.configs import (
-    VideoEncoderConfig,
-    camera_encoder_defaults,
-)
-from lerobot.utils.import_utils import get_safe_default_video_backend
+from lerobot.utils.import_utils import get_safe_default_codec

 logger = logging.getLogger(__name__)

+# List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
+# Determines the order of preference for auto-selection when vcodec="auto" is used.
+HW_ENCODERS = [
+    "h264_videotoolbox",  # macOS
+    "hevc_videotoolbox",  # macOS
+    "h264_nvenc",  # NVIDIA GPU
+    "hevc_nvenc",  # NVIDIA GPU
+    "h264_vaapi",  # Linux Intel/AMD
+    "h264_qsv",  # Intel Quick Sync
+]
+
+VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1", "auto"} | set(HW_ENCODERS)
+
+
+def _get_codec_options(
+    vcodec: str,
+    g: int | None = 2,
+    crf: int | None = 30,
+    preset: int | None = None,
+) -> dict:
+    """Build codec-specific options dict for video encoding."""
+    options = {}
+
+    # GOP size (keyframe interval) - supported by VideoToolbox and software encoders
+    if g is not None and (vcodec in ("h264_videotoolbox", "hevc_videotoolbox") or vcodec not in HW_ENCODERS):
+        options["g"] = str(g)
+
+    # Quality control (codec-specific parameter names)
+    if crf is not None:
+        if vcodec in ("h264", "hevc", "libsvtav1"):
+            options["crf"] = str(crf)
+        elif vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
+            quality = max(1, min(100, int(100 - crf * 2)))
+            options["q:v"] = str(quality)
+        elif vcodec in ("h264_nvenc", "hevc_nvenc"):
+            options["rc"] = "constqp"
+            options["qp"] = str(crf)
+        elif vcodec in ("h264_vaapi",):
+            options["qp"] = str(crf)
+        elif vcodec in ("h264_qsv",):
+            options["global_quality"] = str(crf)
+
+    # Preset (only for libsvtav1)
+    if vcodec == "libsvtav1":
+        options["preset"] = str(preset) if preset is not None else "12"
+
+    return options
+
+
+def detect_available_hw_encoders() -> list[str]:
+    """Probe PyAV/FFmpeg for available hardware video encoders."""
+    available = []
+    for codec_name in HW_ENCODERS:
+        try:
+            av.codec.Codec(codec_name, "w")
+            available.append(codec_name)
+        except Exception:  # nosec B110
+            logger.debug("HW encoder '%s' not available", codec_name)  # nosec B110
+    return available
+
+
+def resolve_vcodec(vcodec: str) -> str:
+    """Validate vcodec and resolve 'auto' to best available HW encoder, fallback to libsvtav1."""
+    if vcodec not in VALID_VIDEO_CODECS:
+        raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
+    if vcodec != "auto":
+        logger.info(f"Using video codec: {vcodec}")
+        return vcodec
+    available = detect_available_hw_encoders()
+    for encoder in HW_ENCODERS:
+        if encoder in available:
+            logger.info(f"Auto-selected video codec: {encoder}")
+            return encoder
+    logger.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
+    return "libsvtav1"
+

 def decode_video_frames(
    video_path: Path | str,
@@ -61,9 +132,7 @@ def decode_video_frames(
        video_path (Path): Path to the video file.
        timestamps (list[float]): List of timestamps to extract frames.
        tolerance_s (float): Allowed deviation in seconds for frame retrieval.
-        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available
-            in the platform; otherwise, defaults to "pyav". The legacy value "video_reader" is
-            accepted for one release as an alias for "pyav" and will be removed in a future version.
+        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available in the platform; otherwise, defaults to "pyav".
        return_uint8 (bool): If True, return raw uint8 frames without float32 normalization.
            This reduces memory for DataLoader IPC; normalization can be done on GPU afterward.

@@ -73,90 +142,88 @@ def decode_video_frames(
    Currently supports torchcodec on cpu and pyav.
    """
    if backend is None:
-        backend = get_safe_default_video_backend()
+        backend = get_safe_default_codec()
    if backend == "torchcodec":
        return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s, return_uint8=return_uint8)
-    elif backend == "pyav":
-        return decode_video_frames_pyav(video_path, timestamps, tolerance_s, return_uint8=return_uint8)
-    elif backend == "video_reader":
-        logger.warning("backend='video_reader' is deprecated and now aliases to 'pyav'.")
-        return decode_video_frames_pyav(video_path, timestamps, tolerance_s, return_uint8=return_uint8)
+    elif backend in ["pyav", "video_reader"]:
+        return decode_video_frames_torchvision(
+            video_path, timestamps, tolerance_s, backend, return_uint8=return_uint8
+        )
    else:
        raise ValueError(f"Unsupported video backend: {backend}")


-def decode_video_frames_pyav(
+def decode_video_frames_torchvision(
    video_path: Path | str,
    timestamps: list[float],
    tolerance_s: float,
+    backend: str = "pyav",
    log_loaded_timestamps: bool = False,
    return_uint8: bool = False,
 ) -> torch.Tensor:
-    """Loads frames associated to the requested timestamps of a video using PyAV.
+    """Loads frames associated to the requested timestamps of a video

-    This is the fallback decoder for platforms where torchcodec has no wheel (currently macOS
-    x86_64 and linux armv7l — see the torchcodec block in pyproject.toml for the full matrix).
-    On supported platforms, prefer `decode_video_frames_torchcodec`, which is faster and supports
-    accurate seek.
+    The backend can be either "pyav" (default) or "video_reader".
+    "video_reader" requires installing torchvision from source, see:
+    https://github.com/pytorch/vision/blob/main/torchvision/csrc/io/decoder/gpu/README.rst
+    (note that you need to compile against ffmpeg<4.3)

-    PyAV doesn't support accurate seek: we seek to the nearest preceding keyframe and decode
-    forward until we have covered the requested timestamp range. The number of key frames in a
-    video can be adjusted at encoding time to trade off decoding speed against file size.
+    While both use cpu, "video_reader" is supposedly faster than "pyav" but requires additional setup.
+    For more info on video decoding, see `benchmark/video/README.md`

-    Args:
-        video_path: Path to the video file.
-        timestamps: List of timestamps (in seconds) to extract frames for.
-        tolerance_s: Allowed deviation in seconds between a queried timestamp and the closest
-            decoded frame.
-        log_loaded_timestamps: When True, log every decoded frame's timestamp at INFO level.
-        return_uint8: When True, return raw uint8 frames (C, H, W). Otherwise, return float32 in
-            [0, 1] range.
+    See torchvision doc for more info on these two backends:
+    https://pytorch.org/vision/0.18/index.html?highlight=backend#torchvision.set_video_backend

-    Returns:
-        torch.Tensor of shape (len(timestamps), C, H, W).
+    Note: Video benefits from inter-frame compression. Instead of storing every frame individually,
+    the encoder stores a reference frame (or a key frame) and subsequent frames as differences relative to
+    that key frame. As a consequence, to access a requested frame, we need to load the preceding key frame,
+    and all subsequent frames until reaching the requested frame. The number of key frames in a video
+    can be adjusted during encoding to take into account decoding time and video size in bytes.
    """
-    # TODO(rcadene): also load audio stream at the same time
    video_path = str(video_path)

+    # set backend
+    keyframes_only = False
+    torchvision.set_video_backend(backend)
+    if backend == "pyav":
+        keyframes_only = True  # pyav doesn't support accurate seek
+
+    # set a video stream reader
+    # TODO(rcadene): also load audio stream at the same time
+    reader = torchvision.io.VideoReader(video_path, "video")
+
    # set the first and last requested timestamps
    # Note: previous timestamps are usually loaded, since we need to access the previous key frame
    first_ts = min(timestamps)
    last_ts = max(timestamps)

-    loaded_frames: list[torch.Tensor] = []
-    loaded_ts: list[float] = []
+    # access closest key frame of the first requested frame
+    # Note: closest key frame timestamp is usually smaller than `first_ts` (e.g. key frame can be the first frame of the video)
+    # for details on what `seek` is doing see: https://pyav.basswood-io.com/docs/stable/api/container.html?highlight=inputcontainer#av.container.InputContainer.seek
+    reader.seek(first_ts, keyframes_only=keyframes_only)

-    # Seek + decode. `container.seek(offset)` with no `stream` argument expects the offset in
-    # av.time_base units (microseconds). `backward=True` lands us on the nearest keyframe at or
-    # before `first_ts`, so we can then decode forward until we cover `last_ts`. See:
-    # https://pyav.basswood-io.com/docs/stable/api/container.html#av.container.InputContainer.seek
-    with av.open(video_path) as container:
-        stream = container.streams.video[0]
-        container.seek(int(first_ts * av.time_base), backward=True)
+    # load all frames until last requested frame
+    loaded_frames = []
+    loaded_ts = []
+    for frame in reader:
+        current_ts = frame["pts"]
+        if log_loaded_timestamps:
+            logger.info(f"frame loaded at timestamp={current_ts:.4f}")
+        loaded_frames.append(frame["data"])
+        loaded_ts.append(current_ts)
+        if current_ts >= last_ts:
+            break

-        for frame in container.decode(stream):
-            if frame.pts is None:
-                continue
-            current_ts = float(frame.pts * stream.time_base)
-            if log_loaded_timestamps:
-                logger.info(f"frame loaded at timestamp={current_ts:.4f}")
-            # Convert to CHW uint8 to match torchcodec's output layout.
-            arr = frame.to_ndarray(format="rgb24")  # H, W, 3
-            loaded_frames.append(torch.from_numpy(arr).permute(2, 0, 1).contiguous())
-            loaded_ts.append(current_ts)
-            if current_ts >= last_ts:
-                break
+    if backend == "pyav":
+        reader.container.close()

-    if not loaded_frames:
-        raise FrameTimestampError(
-            f"No frames could be decoded from {video_path} in the timestamp range [{first_ts}, {last_ts}]."
-        )
+    reader = None

    query_ts = torch.tensor(timestamps)
-    loaded_ts_t = torch.tensor(loaded_ts)
+    loaded_ts = torch.tensor(loaded_ts)

    # compute distances between each query timestamp and timestamps of all loaded frames
-    dist = torch.cdist(query_ts[:, None], loaded_ts_t[:, None], p=1)
+    dist = torch.cdist(query_ts[:, None], loaded_ts[:, None], p=1)
    min_, argmin_ = dist.min(1)

    is_within_tol = min_ < tolerance_s
@@ -167,14 +234,14 @@ def decode_video_frames_pyav(
            " This might be due to synchronization issues with timestamps during data collection."
            " To be safe, we advise to ignore this item during training."
            f"\nqueried timestamps: {query_ts}"
-            f"\nloaded timestamps: {loaded_ts_t}"
+            f"\nloaded timestamps: {loaded_ts}"
            f"\nvideo: {video_path}"
-            f"\nbackend: pyav"
+            f"\nbackend: {backend}"
        )

    # get closest frames to the query timestamps
    closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
-    closest_ts = loaded_ts_t[argmin_]
+    closest_ts = loaded_ts[argmin_]

    if log_loaded_timestamps:
        logger.info(f"{closest_ts=}")
@@ -193,70 +260,15 @@ def decode_video_frames_pyav(
    return closest_frames


-DEFAULT_DECODER_CACHE_SIZE = 100
-"""Default LRU capacity for :class:`VideoDecoderCache`.
-
-Sized to comfortably hold a small rolling window of episodes worth of decoders
-(typical recipes: 2-4 cameras per episode × tens of episodes in flight) while
-bounding host RAM. Each cached entry retains a torchcodec ``VideoDecoder`` plus
-an open ``fsspec`` file handle — on the order of a few MB per entry. Override
-via the ``LEROBOT_VIDEO_DECODER_CACHE_SIZE`` env var or by passing ``max_size``
-to the constructor (``None`` restores the legacy unbounded behaviour).
-"""
-
-
-def _default_max_cache_size() -> int | None:
-    raw = os.environ.get("LEROBOT_VIDEO_DECODER_CACHE_SIZE")
-    if raw is None:
-        return DEFAULT_DECODER_CACHE_SIZE
-    raw = raw.strip().lower()
-    if raw in ("", "none", "unbounded", "-1"):
-        return None
-    try:
-        value = int(raw)
-    except ValueError as e:
-        raise ValueError(
-            f"LEROBOT_VIDEO_DECODER_CACHE_SIZE must be an integer, 'none', or '-1'; got {raw!r}"
-        ) from e
-    if value <= 0:
-        raise ValueError(f"LEROBOT_VIDEO_DECODER_CACHE_SIZE must be positive; got {value}")
-    return value
-
-
 class VideoDecoderCache:
-    """Thread-safe LRU cache for torchcodec ``VideoDecoder`` instances.
+    """Thread-safe cache for video decoders to avoid expensive re-initialization."""

-    Cached entries hold a ``VideoDecoder`` plus the open ``fsspec`` file handle
-    backing it. When the cache is full and a new path is requested, the
-    least-recently-used entry is evicted and its file handle is closed. This
-    bounds host-RAM growth when iterating over datasets with many distinct
-    video files (otherwise each ``DataLoader`` worker pins every decoder it has
-    ever opened until the process exits).
-
-    Args:
-        max_size: Maximum number of decoders to retain. ``None`` disables
-            eviction and restores legacy unbounded behaviour. Defaults to the
-            value of ``LEROBOT_VIDEO_DECODER_CACHE_SIZE`` if set, otherwise
-            :data:`DEFAULT_DECODER_CACHE_SIZE`.
-    """
-
-    _SENTINEL: ClassVar[object] = object()
-
-    def __init__(self, max_size: int | None | object = _SENTINEL):
-        if max_size is VideoDecoderCache._SENTINEL:
-            max_size = _default_max_cache_size()
-        if max_size is not None and max_size <= 0:
-            raise ValueError(f"max_size must be positive or None; got {max_size}")
-        self.max_size: int | None = max_size  # type: ignore[assignment]
-        self._cache: OrderedDict[str, tuple[Any, Any]] = OrderedDict()
+    def __init__(self):
+        self._cache: dict[str, tuple[Any, Any]] = {}
        self._lock = Lock()

-    def __contains__(self, video_path: object) -> bool:
-        with self._lock:
-            return str(video_path) in self._cache
-
    def get_decoder(self, video_path: str):
-        """Get a cached decoder or create a new one, evicting LRU if at capacity."""
+        """Get a cached decoder or create a new one."""
        if importlib.util.find_spec("torchcodec"):
            from torchcodec.decoders import VideoDecoder
        else:
@@ -268,36 +280,18 @@ class VideoDecoderCache:
        video_path = str(video_path)

        with self._lock:
-            entry = self._cache.get(video_path)
-            if entry is not None:
-                self._cache.move_to_end(video_path)
-                return entry[0]
-
-            file_handle = fsspec.open(video_path).__enter__()
-            try:
+            if video_path not in self._cache:
+                file_handle = fsspec.open(video_path).__enter__()
                decoder = VideoDecoder(file_handle, seek_mode="approximate")
-            except Exception:
-                file_handle.close()
-                raise
-            self._cache[video_path] = (decoder, file_handle)
+                self._cache[video_path] = (decoder, file_handle)

-            # Evict LRU entries until we are back under the cap. We close
-            # evicted file handles immediately; the associated ``VideoDecoder``
-            # is released to the GC when its last reference goes away.
-            if self.max_size is not None:
-                while len(self._cache) > self.max_size:
-                    _evicted_path, (_evicted_decoder, evicted_handle) = self._cache.popitem(last=False)
-                    with contextlib.suppress(Exception):
-                        evicted_handle.close()
-
-            return decoder
+            return self._cache[video_path][0]

    def clear(self):
-        """Clear the cache and close all file handles."""
+        """Clear the cache and close file handles."""
        with self._lock:
            for _, file_handle in self._cache.values():
-                with contextlib.suppress(Exception):
-                    file_handle.close()
+                file_handle.close()
            self._cache.clear()

    def size(self) -> int:
@@ -406,17 +400,18 @@ def encode_video_frames(
    imgs_dir: Path | str,
    video_path: Path | str,
    fps: int,
-    camera_encoder: VideoEncoderConfig | None = None,
-    encoder_threads: int | None = None,
-    *,
+    vcodec: str = "libsvtav1",
+    pix_fmt: str = "yuv420p",
+    g: int | None = 2,
+    crf: int | None = 30,
+    fast_decode: int = 0,
    log_level: int | None = av.logging.WARNING,
    overwrite: bool = False,
+    preset: int | None = None,
+    encoder_threads: int | None = None,
 ) -> None:
    """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    if camera_encoder is None:
-        camera_encoder = camera_encoder_defaults()
-    vcodec = camera_encoder.vcodec
-    pix_fmt = camera_encoder.pix_fmt
+    vcodec = resolve_vcodec(vcodec)

    video_path = Path(video_path)
    imgs_dir = Path(imgs_dir)
@@ -427,18 +422,42 @@ def encode_video_frames(

    video_path.parent.mkdir(parents=True, exist_ok=True)

+    # Encoders/pixel formats incompatibility check
+    if (vcodec == "libsvtav1" or vcodec == "hevc") and pix_fmt == "yuv444p":
+        logger.warning(
+            f"Incompatible pixel format 'yuv444p' for codec {vcodec}, auto-selecting format 'yuv420p'"
+        )
+        pix_fmt = "yuv420p"
+
    # Get input frames
    template = "frame-" + ("[0-9]" * 6) + ".png"
    input_list = sorted(
        glob.glob(str(imgs_dir / template)), key=lambda x: int(x.split("-")[-1].split(".")[0])
    )

+    # Define video output frame size (assuming all input frames are the same size)
    if len(input_list) == 0:
        raise FileNotFoundError(f"No images found in {imgs_dir}.")
    with Image.open(input_list[0]) as dummy_image:
        width, height = dummy_image.size

-    video_options = camera_encoder.get_codec_options(encoder_threads, as_strings=True)
+    # Define video codec options
+    video_options = _get_codec_options(vcodec, g, crf, preset)
+
+    if fast_decode:
+        key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
+        value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
+        video_options[key] = value
+
+    if encoder_threads is not None:
+        if vcodec == "libsvtav1":
+            lp_param = f"lp={encoder_threads}"
+            if "svtav1-params" in video_options:
+                video_options["svtav1-params"] += f":{lp_param}"
+            else:
+                video_options["svtav1-params"] = lp_param
+        else:
+            video_options["threads"] = str(encoder_threads)

    # Set logging level
    if log_level is not None:
@@ -474,97 +493,8 @@ def encode_video_frames(
        raise OSError(f"Video encoding did not work. File not found: {video_path}.")


-def reencode_video(
-    input_video_path: Path | str,
-    output_video_path: Path | str,
-    camera_encoder: VideoEncoderConfig | None = None,
-    encoder_threads: int | None = None,
-    log_level: int | None = av.logging.WARNING,
-    overwrite: bool = False,
-) -> None:
-    """Re-encode a video file using the given encoder configuration.
-
-    Args:
-        input_video_path: Existing video file to read.
-        output_video_path: Path for the re-encoded file.
-        camera_encoder: Encoder configuration. Defaults to :func:`camera_encoder_defaults`.
-        encoder_threads: Optional thread count forwarded to :meth:`VideoEncoderConfig.get_codec_options`.
-        log_level: libav log level while encoding, or ``None`` to leave logging unchanged. Defaults to WARNING.
-        overwrite: When ``False`` and ``output_video_path`` already exists, skip and log a warning.
-    """
-
-    camera_encoder = camera_encoder or camera_encoder_defaults()
-
-    output_video_path = Path(output_video_path)
-
-    if output_video_path.exists() and not overwrite:
-        logger.warning(f"Video file already exists: {output_video_path}. Skipping re-encode.")
-        return
-
-    output_video_path.parent.mkdir(parents=True, exist_ok=True)
-
-    video_options = camera_encoder.get_codec_options(encoder_threads, as_strings=True)
-    vcodec = camera_encoder.vcodec
-    pix_fmt = camera_encoder.pix_fmt
-
-    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_named_file:
-        tmp_output_video_path = tmp_named_file.name
-
-    if log_level is not None:
-        logging.getLogger("libav").setLevel(log_level)
-
-    try:
-        with av.open(input_video_path, mode="r") as src:
-            try:
-                in_stream = src.streams.video[0]
-            except IndexError as e:
-                raise ValueError(f"No video stream in {input_video_path}") from e
-
-            fps = (
-                in_stream.base_rate
-            )  # We allow fractional fps though LeRobotDataset only supports integer fps
-            width = int(in_stream.width)
-            height = int(in_stream.height)
-
-            with av.open(
-                tmp_output_video_path,
-                mode="w",
-                options={
-                    "movflags": "faststart"
-                },  # faststart is to move the metadata to the beginning of the file to speed up loading
-            ) as dst:
-                out_stream = dst.add_stream(vcodec, fps, options=video_options)
-                out_stream.pix_fmt = pix_fmt
-                out_stream.width = width
-                out_stream.height = height
-
-                for frame in src.decode(in_stream):
-                    frame = frame.reformat(width=width, height=height, format=pix_fmt)
-                    packet = out_stream.encode(frame)
-                    if packet:
-                        dst.mux(packet)
-
-                packet = out_stream.encode()
-                if packet:
-                    dst.mux(packet)
-
-        shutil.move(tmp_output_video_path, output_video_path)
-    except Exception:
-        Path(tmp_output_video_path).unlink(missing_ok=True)
-        raise
-    finally:
-        if log_level is not None:
-            av.logging.restore_default_callback()
-
-    if not output_video_path.exists():
-        raise OSError(f"Video re-encoding did not work. File not found: {output_video_path}.")
-
-
 def concatenate_video_files(
-    input_video_paths: list[Path | str],
-    output_video_path: Path,
-    overwrite: bool = True,
-    compatibility_check: bool = False,
+    input_video_paths: list[Path | str], output_video_path: Path, overwrite: bool = True
 ):
    """
    Concatenate multiple video files into a single video file using pyav.
@@ -577,7 +507,6 @@ def concatenate_video_files(
        input_video_paths: Ordered list of input video file paths to concatenate.
        output_video_path: Path to the output video file.
        overwrite: Whether to overwrite the output video file if it already exists. Default is True.
-        compatibility_check: Whether to check if the input videos are compatible. Default is False.

    Note:
        - Creates a temporary directory for intermediate files that is cleaned up after use.
@@ -596,22 +525,6 @@ def concatenate_video_files(
    if len(input_video_paths) == 0:
        raise FileNotFoundError("No input video paths provided.")

-    # This check may be skipped at recording time as videos are encoded with the same encoder config.
-    if compatibility_check:
-        reference_video_info = get_video_info(input_video_paths[0])
-        for input_path in input_video_paths[1:]:
-            video_info = get_video_info(input_path)
-            if (
-                video_info["video.height"] != reference_video_info["video.height"]
-                or video_info["video.width"] != reference_video_info["video.width"]
-                or video_info["video.fps"] != reference_video_info["video.fps"]
-                or video_info["video.codec"] != reference_video_info["video.codec"]
-                or video_info["video.pix_fmt"] != reference_video_info["video.pix_fmt"]
-            ):
-                raise ValueError(
-                    f"Input video {input_path} is not compatible with the reference video {input_video_paths[0]}."
-                )
-
    # Create a temporary .ffconcat file to list the input video paths
    with tempfile.NamedTemporaryFile(mode="w", suffix=".ffconcat", delete=False) as tmp_concatenate_file:
        tmp_concatenate_file.write("ffconcat version 1.0\n")
@@ -678,20 +591,26 @@ class _CameraEncoderThread(threading.Thread):
        fps: int,
        vcodec: str,
        pix_fmt: str,
-        codec_options: dict[str, str],
+        g: int | None,
+        crf: int | None,
+        preset: int | None,
        frame_queue: queue.Queue,
        result_queue: queue.Queue,
        stop_event: threading.Event,
+        encoder_threads: int | None = None,
    ):
        super().__init__(daemon=True)
        self.video_path = video_path
        self.fps = fps
        self.vcodec = vcodec
        self.pix_fmt = pix_fmt
-        self.codec_options = codec_options
+        self.g = g
+        self.crf = crf
+        self.preset = preset
        self.frame_queue = frame_queue
        self.result_queue = result_queue
        self.stop_event = stop_event
+        self.encoder_threads = encoder_threads

    def run(self) -> None:
        from .compute_stats import RunningQuantileStats, auto_downsample_height_width
@@ -727,9 +646,19 @@ class _CameraEncoderThread(threading.Thread):
                # Open container on first frame (to get width/height)
                if container is None:
                    height, width = frame_data.shape[:2]
+                    video_options = _get_codec_options(self.vcodec, self.g, self.crf, self.preset)
+                    if self.encoder_threads is not None:
+                        if self.vcodec == "libsvtav1":
+                            lp_param = f"lp={self.encoder_threads}"
+                            if "svtav1-params" in video_options:
+                                video_options["svtav1-params"] += f":{lp_param}"
+                            else:
+                                video_options["svtav1-params"] = lp_param
+                        else:
+                            video_options["threads"] = str(self.encoder_threads)
                    Path(self.video_path).parent.mkdir(parents=True, exist_ok=True)
                    container = av.open(str(self.video_path), "w")
-                    output_stream = container.add_stream(self.vcodec, self.fps, options=self.codec_options)
+                    output_stream = container.add_stream(self.vcodec, self.fps, options=video_options)
                    output_stream.pix_fmt = self.pix_fmt
                    output_stream.width = width
                    output_stream.height = height
@@ -795,24 +724,22 @@ class StreamingVideoEncoder:
    def __init__(
        self,
        fps: int,
-        camera_encoder: VideoEncoderConfig | None = None,
+        vcodec: str = "libsvtav1",
+        pix_fmt: str = "yuv420p",
+        g: int | None = 2,
+        crf: int | None = 30,
+        preset: int | None = None,
        queue_maxsize: int = 30,
        encoder_threads: int | None = None,
    ):
-        """
-        Args:
-            fps: Frames per second for the output videos.
-            camera_encoder: Video encoder settings applied to all cameras.
-                When ``None``, :func:`camera_encoder_defaults` is used.
-            encoder_threads: Number of encoder threads (global setting).
-                ``None`` lets the codec decide.
-            queue_maxsize: Max frames to buffer per camera before
-                back-pressure drops frames.
-        """
        self.fps = fps
-        self._camera_encoder = camera_encoder or camera_encoder_defaults()
-        self._encoder_threads = encoder_threads
+        self.vcodec = resolve_vcodec(vcodec)
+        self.pix_fmt = pix_fmt
+        self.g = g
+        self.crf = crf
+        self.preset = preset
        self.queue_maxsize = queue_maxsize
+        self.encoder_threads = encoder_threads

        self._frame_queues: dict[str, queue.Queue] = {}
        self._result_queues: dict[str, queue.Queue] = {}
@@ -843,17 +770,18 @@ class StreamingVideoEncoder:
            temp_video_dir = Path(tempfile.mkdtemp(dir=temp_dir))
            video_path = temp_video_dir / f"{video_key.replace('/', '_')}_streaming.mp4"

-            vcodec = self._camera_encoder.vcodec
-            codec_options = self._camera_encoder.get_codec_options(self._encoder_threads, as_strings=True)
            encoder_thread = _CameraEncoderThread(
                video_path=video_path,
                fps=self.fps,
-                vcodec=vcodec,
-                pix_fmt=self._camera_encoder.pix_fmt,
-                codec_options=codec_options,
+                vcodec=self.vcodec,
+                pix_fmt=self.pix_fmt,
+                g=self.g,
+                crf=self.crf,
+                preset=self.preset,
                frame_queue=frame_queue,
                result_queue=result_queue,
                stop_event=stop_event,
+                encoder_threads=self.encoder_threads,
            )
            encoder_thread.start()

@@ -1058,18 +986,8 @@ def get_audio_info(video_path: Path | str) -> dict:
    return audio_info


-def get_video_info(
-    video_path: Path | str,
-    camera_encoder: VideoEncoderConfig | None = None,
-) -> dict:
-    """Build the ``video.*`` / ``audio.*`` info dict persisted in ``info.json``.
-
-    Args:
-        video_path: Path to the encoded video file to probe.
-        camera_encoder: If provided, record the exact encoder settings used to encode this
-            video. Stream-derived values take precedence — encoder fields are only written for keys
-            not already populated from the video file itself.
-    """
+def get_video_info(video_path: Path | str) -> dict:
+    # Set logging level
    logging.getLogger("libav").setLevel(av.logging.WARNING)

    # Getting video stream information
@@ -1100,14 +1018,6 @@ def get_video_info(
    # Adding audio stream information
    video_info.update(**get_audio_info(video_path))

-    # Add additional encoder configuration if provided
-    if camera_encoder is not None:
-        for field_name, field_value in asdict(camera_encoder).items():
-            # vcodec is already populated from the video stream
-            if field_name == "vcodec":
-                continue
-            video_info.setdefault(f"video.{field_name}", field_value)
-
    return video_info


--- a/src/lerobot/envs/configs.py
+++ b/src/lerobot/envs/configs.py
@@ -331,7 +331,6 @@ class LiberoEnv(EnvConfig):
    camera_name_mapping: dict[str, str] | None = None
    observation_height: int = 360
    observation_width: int = 360
-    is_libero_plus: bool = False
    features: dict[str, PolicyFeature] = field(
        default_factory=lambda: {
            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
@@ -433,7 +432,6 @@ class LiberoEnv(EnvConfig):
            control_mode=self.control_mode,
            episode_length=self.episode_length,
            camera_name_mapping=self.camera_name_mapping,
-            is_libero_plus=self.is_libero_plus,
        )

    def get_env_processors(self):
@@ -498,146 +496,6 @@ class MetaworldEnv(EnvConfig):
        )


-@EnvConfig.register_subclass("robocasa")
-@dataclass
-class RoboCasaEnv(EnvConfig):
-    task: str = "CloseFridge"
-    fps: int = 20
-    episode_length: int = 1000
-    obs_type: str = "pixels_agent_pos"
-    render_mode: str = "rgb_array"
-    camera_name: str = "robot0_agentview_left,robot0_eye_in_hand,robot0_agentview_right"
-    observation_height: int = 256
-    observation_width: int = 256
-    visualization_height: int = 512
-    visualization_width: int = 512
-    split: str | None = None
-    # Object-mesh registries to sample from. Upstream default is
-    # ("objaverse", "lightwheel"), but objaverse is ~30GB and the CI image
-    # only ships the lightwheel pack. Override to include objaverse once
-    # you've run `python -m robocasa.scripts.download_kitchen_assets
-    # --type objaverse` locally.
-    obj_registries: list[str] = field(default_factory=lambda: ["lightwheel"])
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(12,))}
-    )
-    features_map: dict[str, str] = field(default_factory=lambda: {ACTION: ACTION, "agent_pos": OBS_STATE})
-
-    def __post_init__(self):
-        if self.obs_type not in ("pixels", "pixels_agent_pos"):
-            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
-
-        # Preserve raw RoboCasa camera names end-to-end (e.g.
-        # `observation.images.robot0_agentview_left`). This matches the
-        # naming convention used by the RoboCasa datasets on the Hub, so
-        # trained policies don't need a `--rename_map` at eval time.
-        cams = [c.strip() for c in self.camera_name.split(",") if c.strip()]
-        for cam in cams:
-            self.features[f"pixels/{cam}"] = PolicyFeature(
-                type=FeatureType.VISUAL,
-                shape=(self.observation_height, self.observation_width, 3),
-            )
-            self.features_map[f"pixels/{cam}"] = f"{OBS_IMAGES}.{cam}"
-
-        if self.obs_type == "pixels_agent_pos":
-            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(16,))
-
-    @property
-    def gym_kwargs(self) -> dict:
-        kwargs: dict[str, Any] = {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-            "observation_height": self.observation_height,
-            "observation_width": self.observation_width,
-            "visualization_height": self.visualization_height,
-            "visualization_width": self.visualization_width,
-        }
-        if self.split is not None:
-            kwargs["split"] = self.split
-        return kwargs
-
-    def create_envs(self, n_envs: int, use_async_envs: bool = False):
-        from .robocasa import create_robocasa_envs
-
-        if self.task is None:
-            raise ValueError("RoboCasaEnv requires a task to be specified")
-        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
-        return create_robocasa_envs(
-            task=self.task,
-            n_envs=n_envs,
-            camera_name=self.camera_name,
-            gym_kwargs=self.gym_kwargs,
-            env_cls=env_cls,
-            episode_length=self.episode_length,
-            obj_registries=tuple(self.obj_registries),
-        )
-
-
-@EnvConfig.register_subclass("vlabench")
-@dataclass
-class VLABenchEnv(EnvConfig):
-    task: str = "select_fruit"
-    fps: int = 10
-    episode_length: int = 500
-    obs_type: str = "pixels_agent_pos"
-    render_mode: str = "rgb_array"
-    render_resolution: tuple[int, int] = (480, 480)
-    robot: str = "franka"
-    action_mode: str = "eef"
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            ACTION: ACTION,
-            "agent_pos": OBS_STATE,
-            "pixels/image": f"{OBS_IMAGES}.image",
-            "pixels/second_image": f"{OBS_IMAGES}.second_image",
-            "pixels/wrist_image": f"{OBS_IMAGES}.wrist_image",
-        }
-    )
-
-    def __post_init__(self):
-        h, w = self.render_resolution
-        if self.obs_type == "pixels":
-            self.features["pixels/image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-            self.features["pixels/second_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-            self.features["pixels/wrist_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-        elif self.obs_type == "pixels_agent_pos":
-            self.features["pixels/image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-            self.features["pixels/second_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-            self.features["pixels/wrist_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(7,))
-        else:
-            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-            "render_resolution": self.render_resolution,
-            "robot": self.robot,
-            "max_episode_steps": self.episode_length,
-            "action_mode": self.action_mode,
-        }
-
-    def create_envs(self, n_envs: int, use_async_envs: bool = False):
-        from .vlabench import create_vlabench_envs
-
-        if self.task is None:
-            raise ValueError("VLABenchEnv requires a task to be specified")
-        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
-        return create_vlabench_envs(
-            task=self.task,
-            n_envs=n_envs,
-            gym_kwargs=self.gym_kwargs,
-            env_cls=env_cls,
-        )
-
-
@EnvConfig.register_subclass("isaaclab_arena")
@dataclass
 class IsaaclabArenaEnv(HubEnvConfig):
@@ -716,171 +574,3 @@ class IsaaclabArenaEnv(HubEnvConfig):
            ),
            PolicyProcessorPipeline(steps=[]),
        )
-
-
-@EnvConfig.register_subclass("libero_plus")
-@dataclass
-class LiberoPlusEnv(LiberoEnv):
-    """Config for LIBERO-plus robustness benchmark evaluation.
-
-    LIBERO-plus extends LIBERO with 7 perturbation dimensions (camera viewpoints,
-    object layouts, robot initial states, language instructions, lighting, background
-    textures, sensor noise) producing ~10k task variants.
-
-    The gym interface is identical to LIBERO so this class reuses ``LiberoEnv``
-    entirely — only the registered name and default task suite differ.
-
-    Install: see docker/Dockerfile.benchmark.libero_plus — LIBERO-plus ships
-    as a namespace package from a git fork and must be cloned + PYTHONPATH'd
-    rather than installed as a pyproject extra.
-
-    See Also:
-        https://github.com/sylvestf/LIBERO-plus
-    """
-
-    task: str = "libero_spatial"
-    is_libero_plus: bool = True
-
-
-@EnvConfig.register_subclass("robotwin")
-@dataclass
-class RoboTwinEnvConfig(EnvConfig):
-    """Configuration for RoboTwin 2.0 benchmark environments.
-
-    RoboTwin 2.0 is a dual-arm manipulation benchmark with 50 tasks built on the
-    SAPIEN simulator. The robot is an Aloha-AgileX bimanual platform with 14 DOF
-    (7 per arm). All three cameras are enabled by default.
-
-    See: https://robotwin-platform.github.io
-    Dataset: https://huggingface.co/datasets/lerobot/robotwin_unified
-    """
-
-    task: str = "beat_block_hammer"  # single task or comma-separated list
-    fps: int = 25
-    episode_length: int = 300
-    obs_type: str = "pixels_agent_pos"
-    render_mode: str = "rgb_array"
-    # Available cameras from RoboTwin's aloha-agilex embodiment: head_camera
-    # (torso-mounted) + left_camera / right_camera (wrists).
-    camera_names: str = "head_camera,left_camera,right_camera"
-    # Match the D435 dims in task_config/demo_clean.yml (_camera_config.yml).
-    # Gym's vector-env concatenate pre-allocates buffers of this shape, so it
-    # must equal what SAPIEN actually renders.
-    observation_height: int = 240
-    observation_width: int = 320
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            ACTION: ACTION,
-            "pixels/head_camera": f"{OBS_IMAGES}.head_camera",
-            "pixels/left_camera": f"{OBS_IMAGES}.left_camera",
-            "pixels/right_camera": f"{OBS_IMAGES}.right_camera",
-            "agent_pos": OBS_STATE,
-        }
-    )
-
-    def __post_init__(self):
-        cam_list = [c.strip() for c in self.camera_names.split(",") if c.strip()]
-        for cam in cam_list:
-            self.features[f"pixels/{cam}"] = PolicyFeature(
-                type=FeatureType.VISUAL,
-                shape=(self.observation_height, self.observation_width, 3),
-            )
-            # Keep features_map entry if already set (default_factory); add if missing.
-            key = f"pixels/{cam}"
-            if key not in self.features_map:
-                self.features_map[key] = f"{OBS_IMAGES}.{cam}"
-
-        if self.obs_type == "pixels_agent_pos":
-            self.features["agent_pos"] = PolicyFeature(
-                type=FeatureType.STATE,
-                shape=(14,),  # 14 DOF: 7 per arm
-            )
-        elif self.obs_type != "pixels":
-            raise ValueError(
-                f"Unsupported obs_type '{self.obs_type}'. "
-                "RoboTwinEnvConfig supports 'pixels' and 'pixels_agent_pos'."
-            )
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {}
-
-    def create_envs(self, n_envs: int, use_async_envs: bool = True):
-        from lerobot.envs.robotwin import create_robotwin_envs
-
-        if not self.task:
-            raise ValueError("RoboTwinEnvConfig requires `task` to be specified.")
-
-        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
-        cam_list = [c.strip() for c in self.camera_names.split(",") if c.strip()]
-        return create_robotwin_envs(
-            task=self.task,
-            n_envs=n_envs,
-            env_cls=env_cls,
-            camera_names=cam_list,
-            observation_height=self.observation_height,
-            observation_width=self.observation_width,
-            episode_length=self.episode_length,
-        )
-
-
-@EnvConfig.register_subclass("robomme")
-@dataclass
-class RoboMMEEnv(EnvConfig):
-    """RoboMME memory-augmented manipulation benchmark (ManiSkill/SAPIEN).
-
-    16 tasks across 4 suites: Counting, Permanence, Reference, Imitation.
-    Dataset: lerobot/robomme (LeRobot v3.0, 1,600 episodes).
-    Benchmark: https://github.com/RoboMME/robomme_benchmark
-
-    Requires the `robomme` git package installed separately (Linux only);
-    see docker/Dockerfile.benchmark.robomme for the canonical install.
-    """
-
-    task: str = "PickXtimes"
-    fps: int = 10
-    episode_length: int = 300
-    action_space: str = "joint_angle"  # or "ee_pose" (7-D)
-    dataset_split: str = "test"  # "train" | "val" | "test"
-    task_ids: list[int] | None = None
-    features: dict[str, PolicyFeature] = field(default_factory=dict)
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            ACTION: ACTION,
-            "pixels/image": f"{OBS_IMAGES}.image",
-            "pixels/wrist_image": f"{OBS_IMAGES}.wrist_image",
-            "agent_pos": OBS_STATE,
-        }
-    )
-
-    def __post_init__(self):
-        action_dim = 8 if self.action_space == "joint_angle" else 7
-        self.features = {
-            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_dim,)),
-            "pixels/image": PolicyFeature(type=FeatureType.VISUAL, shape=(256, 256, 3)),
-            "pixels/wrist_image": PolicyFeature(type=FeatureType.VISUAL, shape=(256, 256, 3)),
-            "agent_pos": PolicyFeature(type=FeatureType.STATE, shape=(8,)),
-        }
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {}
-
-    def create_envs(self, n_envs: int, use_async_envs: bool = True):
-        from lerobot.envs.robomme import create_robomme_envs
-
-        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
-        return create_robomme_envs(
-            task=self.task,
-            n_envs=n_envs,
-            action_space_type=self.action_space,
-            dataset=self.dataset_split,
-            episode_length=self.episode_length,
-            task_ids=self.task_ids,
-            env_cls=env_cls,
-        )
--- a/src/lerobot/envs/libero.py
+++ b/src/lerobot/envs/libero.py
@@ -16,7 +16,6 @@
 from __future__ import annotations

 import os
-import re
 from collections import defaultdict
 from collections.abc import Callable, Iterable, Mapping, Sequence
 from functools import partial
@@ -32,7 +31,20 @@ from libero.libero.envs import OffScreenRenderEnv

 from lerobot.types import RobotObservation

-from .utils import _LazyAsyncVectorEnv, parse_camera_names
+from .utils import _LazyAsyncVectorEnv
+
+
+def _parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
+    """Normalize camera_name into a non-empty list of strings."""
+    if isinstance(camera_name, str):
+        cams = [c.strip() for c in camera_name.split(",") if c.strip()]
+    elif isinstance(camera_name, (list | tuple)):
+        cams = [str(c).strip() for c in camera_name if str(c).strip()]
+    else:
+        raise TypeError(f"camera_name must be str or sequence[str], got {type(camera_name).__name__}")
+    if not cams:
+        raise ValueError("camera_name resolved to an empty list.")
+    return cams


 def _get_suite(name: str) -> benchmark.Benchmark:
@@ -57,34 +69,14 @@ def _select_task_ids(total_tasks: int, task_ids: Iterable[int] | None) -> list[i
    return ids


-# LIBERO-plus perturbation variants encode the perturbation in the filename
-# but on disk only the base `.pruned_init` exists — strip the suffix to match
-# LIBERO-plus's own suite.get_task_init_states() (we reimplement it here so we
-# can pass weights_only=False for PyTorch 2.6+ numpy pickles).
-_LIBERO_PERTURBATION_SUFFIX_RE = re.compile(r"_(?:language|view|light)_[^.]*|_(?:table|tb)_\d+")
-
-
-def get_task_init_states(task_suite: Any, i: int, is_libero_plus: bool = False) -> np.ndarray:
-    task = task_suite.tasks[i]
-    filename = Path(task.init_states_file)
-    root = Path(get_libero_path("init_states"))
-
-    if not is_libero_plus:
-        init_states_path = root / task.problem_folder / filename.name
-        return torch.load(init_states_path, weights_only=False)  # nosec B614
-
-    # LIBERO-plus: `_add_` / `_level` variants store extra-object layouts under
-    # libero_newobj/ as a flat array that must be reshaped to (1, -1).
-    if "_add_" in filename.name or "_level" in filename.name:
-        init_states_path = root / "libero_newobj" / task.problem_folder / filename.name
-        init_states = torch.load(init_states_path, weights_only=False)  # nosec B614
-        return init_states.reshape(1, -1)
-
-    # LIBERO-plus perturbation variants encode the perturbation in the filename
-    # but on disk only the base `.pruned_init` exists — strip the suffix to match.
-    stripped = _LIBERO_PERTURBATION_SUFFIX_RE.sub("", filename.stem) + filename.suffix
-    init_states_path = root / task.problem_folder / stripped
-    return torch.load(init_states_path, weights_only=False)  # nosec B614
+def get_task_init_states(task_suite: Any, i: int) -> np.ndarray:
+    init_states_path = (
+        Path(get_libero_path("init_states"))
+        / task_suite.tasks[i].problem_folder
+        / task_suite.tasks[i].init_states_file
+    )
+    init_states = torch.load(init_states_path, weights_only=False)  # nosec B614
+    return init_states


 def get_libero_dummy_action():
@@ -126,11 +118,9 @@ class LiberoEnv(gym.Env):
        camera_name_mapping: dict[str, str] | None = None,
        num_steps_wait: int = 10,
        control_mode: str = "relative",
-        is_libero_plus: bool = False,
    ):
        super().__init__()
        self.task_id = task_id
-        self.is_libero_plus = is_libero_plus
        self.obs_type = obs_type
        self.render_mode = render_mode
        self.observation_width = observation_width
@@ -138,7 +128,7 @@ class LiberoEnv(gym.Env):
        self.visualization_width = visualization_width
        self.visualization_height = visualization_height
        self.init_states = init_states
-        self.camera_name = parse_camera_names(
+        self.camera_name = _parse_camera_names(
            camera_name
        )  # agentview_image (main) or robot0_eye_in_hand_image (wrist)

@@ -157,11 +147,7 @@ class LiberoEnv(gym.Env):
        self.episode_index = episode_index
        self.episode_length = episode_length
        # Load once and keep
-        self._init_states = (
-            get_task_init_states(task_suite, self.task_id, is_libero_plus=self.is_libero_plus)
-            if self.init_states
-            else None
-        )
+        self._init_states = get_task_init_states(task_suite, self.task_id) if self.init_states else None
        self._reset_stride = n_envs  # when performing a reset, append `_reset_stride` to `init_state_id`.

        self.init_state_id = self.episode_index  # tie each sub-env to a fixed init state
@@ -394,7 +380,6 @@ def _make_env_fns(
    gym_kwargs: Mapping[str, Any],
    control_mode: str,
    camera_name_mapping: dict[str, str] | None = None,
-    is_libero_plus: bool = False,
 ) -> list[Callable[[], LiberoEnv]]:
    """Build n_envs factory callables for a single (suite, task_id)."""

@@ -411,7 +396,6 @@ def _make_env_fns(
            n_envs=n_envs,
            control_mode=control_mode,
            camera_name_mapping=camera_name_mapping,
-            is_libero_plus=is_libero_plus,
            **local_kwargs,
        )

@@ -434,7 +418,6 @@ def create_libero_envs(
    control_mode: str = "relative",
    episode_length: int | None = None,
    camera_name_mapping: dict[str, str] | None = None,
-    is_libero_plus: bool = False,
 ) -> dict[str, dict[int, Any]]:
    """
    Create vectorized LIBERO environments with a consistent return shape.
@@ -454,7 +437,7 @@ def create_libero_envs(
    gym_kwargs = dict(gym_kwargs or {})
    task_ids_filter = gym_kwargs.pop("task_ids", None)  # optional: limit to specific tasks

-    camera_names = parse_camera_names(camera_name)
+    camera_names = _parse_camera_names(camera_name)
    suite_names = [s.strip() for s in str(task).split(",") if s.strip()]
    if not suite_names:
        raise ValueError("`task` must contain at least one LIBERO suite name.")
@@ -479,7 +462,6 @@ def create_libero_envs(
        # Probe once and reuse to avoid creating a temp env per task.
        cached_obs_space: spaces.Space | None = None
        cached_act_space: spaces.Space | None = None
-        cached_metadata: dict[str, Any] | None = None

        for tid in selected:
            fns = _make_env_fns(
@@ -493,14 +475,12 @@ def create_libero_envs(
                gym_kwargs=gym_kwargs,
                control_mode=control_mode,
                camera_name_mapping=camera_name_mapping,
-                is_libero_plus=is_libero_plus,
            )
            if is_async:
-                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
+                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space)
                if cached_obs_space is None:
                    cached_obs_space = lazy.observation_space
                    cached_act_space = lazy.action_space
-                    cached_metadata = lazy.metadata
                out[suite_name][tid] = lazy
            else:
                out[suite_name][tid] = env_cls(fns)
--- a/src/lerobot/envs/metaworld.py
+++ b/src/lerobot/envs/metaworld.py
@@ -311,7 +311,6 @@ def create_metaworld_envs(
    is_async = env_cls is gym.vector.AsyncVectorEnv
    cached_obs_space = None
    cached_act_space = None
-    cached_metadata = None
    out: dict[str, dict[int, Any]] = defaultdict(dict)

    for group in task_groups:
@@ -325,11 +324,10 @@ def create_metaworld_envs(
            fns = [(lambda tn=task_name: MetaworldEnv(task=tn, **gym_kwargs)) for _ in range(n_envs)]

            if is_async:
-                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
+                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space)
                if cached_obs_space is None:
                    cached_obs_space = lazy.observation_space
                    cached_act_space = lazy.action_space
-                    cached_metadata = lazy.metadata
                out[group][tid] = lazy
            else:
                out[group][tid] = env_cls(fns)
--- a/src/lerobot/envs/robocasa.py
+++ b/src/lerobot/envs/robocasa.py
@@ -1,425 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-from __future__ import annotations
-
-import logging
-from collections import defaultdict
-from collections.abc import Callable, Sequence
-from functools import partial
-from typing import Any
-
-import gymnasium as gym
-import numpy as np
-from gymnasium import spaces
-
-from lerobot.types import RobotObservation
-
-from .utils import _LazyAsyncVectorEnv, parse_camera_names
-
-logger = logging.getLogger(__name__)
-
-# Dimensions for the flat action/state vectors used by the LeRobot wrapper.
-# These correspond to the PandaOmron robot in RoboCasa365.
-OBS_STATE_DIM = 16  # base_pos(3) + base_quat(4) + ee_pos_rel(3) + ee_quat_rel(4) + gripper_qpos(2)
-ACTION_DIM = 12  # base_motion(4) + control_mode(1) + ee_pos(3) + ee_rot(3) + gripper(1)
-ACTION_LOW = -1.0
-ACTION_HIGH = 1.0
-
-# Default PandaOmron cameras. We surface these raw names directly as
-# `observation.images.<name>` so the LeRobot dataset/policy keys match
-# RoboCasa's native convention (no implicit renaming).
-DEFAULT_CAMERAS = [
-    "robot0_agentview_left",
-    "robot0_eye_in_hand",
-    "robot0_agentview_right",
-]
-
-# Object-mesh registries to sample from. RoboCasa's upstream default is
-# ("objaverse", "lightwheel"), but the objaverse pack is huge (~30GB) and
-# most users — including our CI image — only download the lightwheel pack
-# (`--type objs_lw` in `download_kitchen_assets`). When a sampled object
-# category has zero candidates in every registry, robocasa crashes with
-# `ValueError: Probabilities contain NaN` (0/0 divide in the probability
-# normalization). Restricting to registries that are actually on disk
-# avoids the NaN and matches what the asset download provides.
-DEFAULT_OBJ_REGISTRIES: tuple[str, ...] = ("lightwheel",)
-
-# Task-group shortcuts accepted as `--env.task`. When the user passes one of
-# these names, we expand it to the upstream RoboCasa task list and auto-set
-# the dataset split. Individual task names (optionally comma-separated) still
-# take precedence; this only triggers on an exact group-name match.
-_TASK_GROUP_SPLITS = {
-    "atomic_seen": "target",
-    "composite_seen": "target",
-    "composite_unseen": "target",
-    "pretrain50": "pretrain",
-    "pretrain100": "pretrain",
-    "pretrain200": "pretrain",
-    "pretrain300": "pretrain",
-}
-
-
-def _resolve_tasks(task: str) -> tuple[list[str], str | None]:
-    """Resolve a `--env.task` value to (task_names, split_override).
-
-    If `task` is a known task-group name (e.g. `atomic_seen`, `pretrain100`),
-    expand it via `robocasa.utils.dataset_registry.{TARGET,PRETRAINING}_TASKS`
-    and return the matching split. Otherwise treat `task` as a single task or
-    comma-separated list and leave the split untouched (None).
-    """
-    key = task.strip()
-    if key in _TASK_GROUP_SPLITS:
-        from robocasa.utils.dataset_registry import PRETRAINING_TASKS, TARGET_TASKS
-
-        combined = {**TARGET_TASKS, **PRETRAINING_TASKS}
-        if key not in combined:
-            raise ValueError(
-                f"Task group '{key}' is not available in this version of robocasa. "
-                f"Known groups: {sorted(combined.keys())}."
-            )
-        return list(combined[key]), _TASK_GROUP_SPLITS[key]
-
-    names = [t.strip() for t in task.split(",") if t.strip()]
-    if not names:
-        raise ValueError("`task` must contain at least one RoboCasa task name.")
-    return names, None
-
-
-def convert_action(flat_action: np.ndarray) -> dict[str, Any]:
-    """Split a flat (12,) action vector into a RoboCasa action dict.
-
-    Layout: base_motion(4) + control_mode(1) + ee_pos(3) + ee_rot(3) + gripper(1)
-    """
-    return {
-        "action.base_motion": flat_action[0:4],
-        "action.control_mode": flat_action[4:5],
-        "action.end_effector_position": flat_action[5:8],
-        "action.end_effector_rotation": flat_action[8:11],
-        "action.gripper_close": flat_action[11:12],
-    }
-
-
-class RoboCasaEnv(gym.Env):
-    """LeRobot gym.Env wrapper for RoboCasa365 kitchen environments.
-
-    Wraps RoboCasaGymEnv from the robocasa package and converts its
-    dict-based observations and actions into the flat arrays LeRobot expects.
-    Raw RoboCasa camera names are preserved verbatim under `pixels/<cam>`.
-    """
-
-    metadata = {"render_modes": ["rgb_array"], "render_fps": 20}
-
-    def __init__(
-        self,
-        task: str,
-        camera_name: str | Sequence[str] = ",".join(DEFAULT_CAMERAS),
-        obs_type: str = "pixels_agent_pos",
-        render_mode: str = "rgb_array",
-        observation_width: int = 256,
-        observation_height: int = 256,
-        visualization_width: int = 512,
-        visualization_height: int = 512,
-        split: str | None = None,
-        episode_length: int | None = None,
-        obj_registries: Sequence[str] = DEFAULT_OBJ_REGISTRIES,
-        episode_index: int = 0,
-    ):
-        super().__init__()
-        self.task = task
-        self.obs_type = obs_type
-        self.render_mode = render_mode
-        self.observation_width = observation_width
-        self.observation_height = observation_height
-        self.visualization_width = visualization_width
-        self.visualization_height = visualization_height
-        self.split = split
-        self.obj_registries = tuple(obj_registries)
-        # Per-worker index (0..n_envs-1) used to spread the user-provided
-        # seed across factories so each sub-env explores a distinct layout
-        # even when the same seed is passed to `reset()`.
-        self.episode_index = int(episode_index)
-
-        self.camera_name = parse_camera_names(camera_name)
-
-        self._max_episode_steps = episode_length if episode_length is not None else 1000
-
-        # Deferred — created on first reset() inside the worker subprocess
-        # to avoid inheriting stale GPU/EGL contexts across fork().
-        self._env: Any = None
-        self.task_description = ""
-
-        images = {
-            cam: spaces.Box(
-                low=0,
-                high=255,
-                shape=(self.observation_height, self.observation_width, 3),
-                dtype=np.uint8,
-            )
-            for cam in self.camera_name
-        }
-
-        if self.obs_type == "pixels":
-            self.observation_space = spaces.Dict({"pixels": spaces.Dict(images)})
-        elif self.obs_type == "pixels_agent_pos":
-            self.observation_space = spaces.Dict(
-                {
-                    "pixels": spaces.Dict(images),
-                    "agent_pos": spaces.Box(
-                        low=-np.inf,
-                        high=np.inf,
-                        shape=(OBS_STATE_DIM,),
-                        dtype=np.float32,
-                    ),
-                }
-            )
-        else:
-            raise ValueError(f"Unsupported obs_type '{self.obs_type}'. Use 'pixels' or 'pixels_agent_pos'.")
-
-        self.action_space = spaces.Box(
-            low=ACTION_LOW,
-            high=ACTION_HIGH,
-            shape=(ACTION_DIM,),
-            dtype=np.float32,
-        )
-
-    def _ensure_env(self) -> None:
-        """Create the underlying RoboCasaGymEnv on first use.
-
-        Called inside the worker subprocess after fork(), so each worker gets
-        its own clean rendering context rather than inheriting a stale one from
-        the parent process (which causes crashes with AsyncVectorEnv).
-        """
-        if self._env is not None:
-            return
-        from robocasa.wrappers.gym_wrapper import RoboCasaGymEnv
-
-        # RoboCasaGymEnv defaults split="test", which create_env rejects
-        # (only None/"all"/"pretrain"/"target" are valid). Always pass a
-        # valid value so we don't hit that default. Extra kwargs are
-        # forwarded to the underlying kitchen env via create_env/robosuite.make.
-        self._env = RoboCasaGymEnv(
-            env_name=self.task,
-            camera_widths=self.observation_width,
-            camera_heights=self.observation_height,
-            split=self.split if self.split is not None else "all",
-            obj_registries=self.obj_registries,
-        )
-
-        ep_meta = self._env.env.get_ep_meta()
-        self.task_description = ep_meta.get("lang", self.task)
-
-    def _format_raw_obs(self, raw_obs: dict) -> RobotObservation:
-        """Convert RoboCasaGymEnv observation dict to LeRobot format."""
-        # RoboCasaGymEnv emits camera frames under "video.<cam>".
-        images = {cam: raw_obs[f"video.{cam}"] for cam in self.camera_name if f"video.{cam}" in raw_obs}
-
-        if self.obs_type == "pixels":
-            return {"pixels": images}
-
-        # `state.*` keys come from PandaOmronKeyConverter inside the wrapper.
-        agent_pos = np.concatenate(
-            [
-                raw_obs.get("state.base_position", np.zeros(3)),
-                raw_obs.get("state.base_rotation", np.zeros(4)),
-                raw_obs.get("state.end_effector_position_relative", np.zeros(3)),
-                raw_obs.get("state.end_effector_rotation_relative", np.zeros(4)),
-                raw_obs.get("state.gripper_qpos", np.zeros(2)),
-            ],
-            axis=-1,
-        ).astype(np.float32)
-
-        return {"pixels": images, "agent_pos": agent_pos}
-
-    def render(self) -> np.ndarray:
-        self._ensure_env()
-        assert self._env is not None
-        return self._env.render()
-
-    def reset(self, seed=None, **kwargs):
-        self._ensure_env()
-        assert self._env is not None
-        super().reset(seed=seed)
-        # Spread the seed across workers so n_envs factories don't all
-        # roll the same scene. With an explicit user seed we shift it by
-        # episode_index; with no seed we fall back to episode_index so
-        # each worker is still distinct rather than inheriting the same
-        # global RNG state.
-        worker_seed = seed + self.episode_index if seed is not None else self.episode_index
-        raw_obs, info = self._env.reset(seed=worker_seed)
-
-        ep_meta = self._env.env.get_ep_meta()
-        self.task_description = ep_meta.get("lang", self.task)
-
-        observation = self._format_raw_obs(raw_obs)
-        info = {"is_success": False}
-        return observation, info
-
-    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
-        self._ensure_env()
-        assert self._env is not None
-        if action.ndim != 1:
-            raise ValueError(
-                f"Expected action to be 1-D (shape (action_dim,)), "
-                f"but got shape {action.shape} with ndim={action.ndim}"
-            )
-
-        action_dict = convert_action(action)
-        raw_obs, reward, done, truncated, info = self._env.step(action_dict)
-
-        is_success = bool(info.get("success", False))
-        terminated = done or is_success
-        info.update({"task": self.task, "done": done, "is_success": is_success})
-
-        observation = self._format_raw_obs(raw_obs)
-        if terminated:
-            info["final_info"] = {
-                "task": self.task,
-                "done": bool(done),
-                "is_success": bool(is_success),
-            }
-            self.reset()
-
-        return observation, reward, terminated, truncated, info
-
-    def close(self):
-        if self._env is not None:
-            self._env.close()
-
-
-def _make_env_fns(
-    *,
-    task: str,
-    n_envs: int,
-    camera_names: list[str],
-    obs_type: str,
-    render_mode: str,
-    observation_width: int,
-    observation_height: int,
-    visualization_width: int,
-    visualization_height: int,
-    split: str | None,
-    episode_length: int | None,
-    obj_registries: Sequence[str],
-) -> list[Callable[[], RoboCasaEnv]]:
-    """Build n_envs factory callables for a single task.
-
-    Each factory carries a distinct ``episode_index`` (``0..n_envs-1``) so
-    ``RoboCasaEnv.reset()`` can derive a per-worker seed series from the
-    user-provided seed.
-    """
-
-    def _make_env(episode_index: int) -> RoboCasaEnv:
-        return RoboCasaEnv(
-            task=task,
-            camera_name=camera_names,
-            obs_type=obs_type,
-            render_mode=render_mode,
-            observation_width=observation_width,
-            observation_height=observation_height,
-            visualization_width=visualization_width,
-            visualization_height=visualization_height,
-            split=split,
-            episode_length=episode_length,
-            obj_registries=obj_registries,
-            episode_index=episode_index,
-        )
-
-    return [partial(_make_env, i) for i in range(n_envs)]
-
-
-def create_robocasa_envs(
-    task: str,
-    n_envs: int,
-    gym_kwargs: dict[str, Any] | None = None,
-    camera_name: str | Sequence[str] = ",".join(DEFAULT_CAMERAS),
-    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
-    episode_length: int | None = None,
-    obj_registries: Sequence[str] = DEFAULT_OBJ_REGISTRIES,
-) -> dict[str, dict[int, Any]]:
-    """Create vectorized RoboCasa365 environments with a consistent return shape.
-
-    Returns:
-        dict[task_name][task_id] -> vec_env (env_cls([...]) with exactly n_envs factories)
-
-    `task` can be:
-      - a single task name (e.g. `CloseFridge`)
-      - a comma-separated list of task names (e.g. `CloseFridge,PickPlaceCoffee`)
-      - a benchmark-group shortcut (`atomic_seen`, `composite_seen`,
-        `composite_unseen`, `pretrain50`, `pretrain100`, `pretrain200`,
-        `pretrain300`), which auto-expands to the upstream task list and
-        auto-sets the dataset `split` ("target" or "pretrain").
-    """
-    if env_cls is None or not callable(env_cls):
-        raise ValueError("env_cls must be a callable that wraps a list of environment factory callables.")
-    if not isinstance(n_envs, int) or n_envs <= 0:
-        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
-
-    gym_kwargs = dict(gym_kwargs or {})
-    obs_type = gym_kwargs.pop("obs_type", "pixels_agent_pos")
-    render_mode = gym_kwargs.pop("render_mode", "rgb_array")
-    observation_width = gym_kwargs.pop("observation_width", 256)
-    observation_height = gym_kwargs.pop("observation_height", 256)
-    visualization_width = gym_kwargs.pop("visualization_width", 512)
-    visualization_height = gym_kwargs.pop("visualization_height", 512)
-    split = gym_kwargs.pop("split", None)
-
-    camera_names = parse_camera_names(camera_name)
-    task_names, group_split = _resolve_tasks(str(task))
-    if group_split is not None and split is None:
-        split = group_split
-
-    logger.info(
-        "Creating RoboCasa envs | tasks=%s | split=%s | n_envs(per task)=%d",
-        task_names,
-        split,
-        n_envs,
-    )
-
-    is_async = env_cls is gym.vector.AsyncVectorEnv
-
-    cached_obs_space: spaces.Space | None = None
-    cached_act_space: spaces.Space | None = None
-    cached_metadata: dict[str, Any] | None = None
-    out: dict[str, dict[int, Any]] = defaultdict(dict)
-
-    for task_name in task_names:
-        fns = _make_env_fns(
-            task=task_name,
-            n_envs=n_envs,
-            camera_names=camera_names,
-            obs_type=obs_type,
-            render_mode=render_mode,
-            observation_width=observation_width,
-            observation_height=observation_height,
-            visualization_width=visualization_width,
-            visualization_height=visualization_height,
-            split=split,
-            episode_length=episode_length,
-            obj_registries=obj_registries,
-        )
-
-        if is_async:
-            lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
-            if cached_obs_space is None:
-                cached_obs_space = lazy.observation_space
-                cached_act_space = lazy.action_space
-                cached_metadata = lazy.metadata
-            out[task_name][0] = lazy
-        else:
-            out[task_name][0] = env_cls(fns)
-        logger.info("Built vec env | task=%s | n_envs=%d", task_name, n_envs)
-
-    return {name: dict(task_map) for name, task_map in out.items()}
--- a/src/lerobot/envs/robomme.py
+++ b/src/lerobot/envs/robomme.py
@@ -1,245 +0,0 @@
-"""RoboMME environment wrapper for LeRobot evaluation.
-
-Wraps the RoboMME ``BenchmarkEnvBuilder`` into a Gymnasium-compatible
-``VectorEnv`` suitable for ``lerobot_eval``.
-
-RoboMME tasks:
-  Counting:    BinFill, PickXtimes, SwingXtimes, StopCube
-  Permanence:  VideoUnmask, VideoUnmaskSwap, ButtonUnmask, ButtonUnmaskSwap
-  Reference:   PickHighlight, VideoRepick, VideoPlaceButton, VideoPlaceOrder
-  Imitation:   MoveCube, InsertPeg, PatternLock, RouteStick
-
-Dataset: lerobot/robomme (LeRobot v3.0, 1,600 episodes)
-Install: see docker/Dockerfile.benchmark.robomme  (Linux only — mani-skill vs numpy pin conflict)
-Benchmark: https://github.com/RoboMME/robomme_benchmark
-"""
-
-from __future__ import annotations
-
-from collections.abc import Callable, Sequence
-from functools import partial
-from typing import Any
-
-import gymnasium as gym
-import numpy as np
-from gymnasium import spaces
-
-from .utils import _LazyAsyncVectorEnv
-
-ROBOMME_TASKS = [
-    "BinFill",
-    "PickXtimes",
-    "SwingXtimes",
-    "StopCube",
-    "VideoUnmask",
-    "VideoUnmaskSwap",
-    "ButtonUnmask",
-    "ButtonUnmaskSwap",
-    "PickHighlight",
-    "VideoRepick",
-    "VideoPlaceButton",
-    "VideoPlaceOrder",
-    "MoveCube",
-    "InsertPeg",
-    "PatternLock",
-    "RouteStick",
-]
-
-
-class RoboMMEGymEnv(gym.Env):
-    """Thin Gymnasium wrapper around a single RoboMME episode env."""
-
-    metadata = {"render_modes": ["rgb_array"], "render_fps": 10}
-
-    def __init__(
-        self,
-        task: str = "PickXtimes",
-        action_space_type: str = "joint_angle",
-        dataset: str = "test",
-        episode_idx: int = 0,
-        max_steps: int = 300,
-    ):
-        super().__init__()
-        from robomme.env_record_wrapper import BenchmarkEnvBuilder
-
-        self._task = task
-        self._action_space_type = action_space_type
-        self._dataset = dataset
-        self._episode_idx = episode_idx
-        self._max_steps = max_steps
-        self._max_episode_steps = max_steps
-
-        self._builder = BenchmarkEnvBuilder(
-            env_id=task,
-            dataset=dataset,
-            action_space=action_space_type,
-            gui_render=False,
-            max_steps=max_steps,
-        )
-        self._env = None
-        self._last_raw_obs: dict | None = None
-
-        action_dim = 8 if action_space_type == "joint_angle" else 7
-        self.action_space = spaces.Box(low=-1.0, high=1.0, shape=(action_dim,), dtype=np.float32)
-        # `pixels` must be a nested Dict so `preprocess_observation()` in
-        # envs/utils.py picks it up and maps each camera to
-        # `observation.images.<cam>`. A flat layout (`pixels/image`,
-        # `pixels/wrist_image`) silently drops every image from the batch.
-        self.observation_space = spaces.Dict(
-            {
-                "pixels": spaces.Dict(
-                    {
-                        "image": spaces.Box(0, 255, shape=(256, 256, 3), dtype=np.uint8),
-                        "wrist_image": spaces.Box(0, 255, shape=(256, 256, 3), dtype=np.uint8),
-                    }
-                ),
-                "agent_pos": spaces.Box(-np.inf, np.inf, shape=(8,), dtype=np.float32),
-            }
-        )
-
-    def reset(self, *, seed=None, options=None):
-        super().reset(seed=seed)
-        self._env = self._builder.make_env_for_episode(
-            episode_idx=self._episode_idx,
-            max_steps=self._max_steps,
-        )
-        obs, info = self._env.reset()
-        self._last_raw_obs = obs
-        return self._convert_obs(obs), self._convert_info(info)
-
-    def step(self, action):
-        obs, reward, terminated, truncated, info = self._env.step(action)
-        self._last_raw_obs = obs
-
-        terminated_bool = bool(terminated.item()) if hasattr(terminated, "item") else bool(terminated)
-        truncated_bool = bool(truncated.item()) if hasattr(truncated, "item") else bool(truncated)
-
-        status = info.get("status", "ongoing")
-        is_success = status == "success"
-        conv_info = self._convert_info(info)
-        conv_info["is_success"] = is_success
-
-        return self._convert_obs(obs), float(reward), terminated_bool, truncated_bool, conv_info
-
-    def render(self) -> np.ndarray | None:
-        """Return the front camera image from the last observation for video recording."""
-        if self._last_raw_obs is None:
-            return np.zeros((256, 256, 3), dtype=np.uint8)
-        front = self._last_raw_obs.get("front_rgb_list")
-        if front is None:
-            return np.zeros((256, 256, 3), dtype=np.uint8)
-        frame = front[-1] if isinstance(front, list) else front
-        return np.asarray(frame, dtype=np.uint8)
-
-    def _convert_obs(self, obs: dict) -> dict:
-        front_rgb = (
-            obs["front_rgb_list"][-1] if isinstance(obs["front_rgb_list"], list) else obs["front_rgb_list"]
-        )
-        wrist_rgb = (
-            obs["wrist_rgb_list"][-1] if isinstance(obs["wrist_rgb_list"], list) else obs["wrist_rgb_list"]
-        )
-        joint_state = (
-            obs["joint_state_list"][-1]
-            if isinstance(obs["joint_state_list"], list)
-            else obs["joint_state_list"]
-        )
-        gripper_state = (
-            obs["gripper_state_list"][-1]
-            if isinstance(obs["gripper_state_list"], list)
-            else obs["gripper_state_list"]
-        )
-
-        front_rgb = np.asarray(front_rgb, dtype=np.uint8)
-        wrist_rgb = np.asarray(wrist_rgb, dtype=np.uint8)
-        joint = np.asarray(joint_state, dtype=np.float32).flatten()[:7]
-        gripper = np.asarray(gripper_state, dtype=np.float32).flatten()[:1]
-        state = np.concatenate([joint, gripper])
-
-        return {
-            "pixels": {"image": front_rgb, "wrist_image": wrist_rgb},
-            "agent_pos": state,
-        }
-
-    def _convert_info(self, info: dict) -> dict:
-        return {
-            "status": info.get("status", "ongoing"),
-            "task_goal": info.get("task_goal", ""),
-        }
-
-
-def _make_env_fns(
-    *,
-    task: str,
-    n_envs: int,
-    action_space_type: str,
-    dataset: str,
-    episode_length: int,
-    task_id: int,
-) -> list[Callable[[], RoboMMEGymEnv]]:
-    """Build n_envs factory callables for one RoboMME task id."""
-
-    def _make_one(episode_index: int) -> RoboMMEGymEnv:
-        return RoboMMEGymEnv(
-            task=task,
-            action_space_type=action_space_type,
-            dataset=dataset,
-            episode_idx=episode_index,
-            max_steps=episode_length,
-        )
-
-    return [partial(_make_one, task_id + i) for i in range(n_envs)]
-
-
-def create_robomme_envs(
-    task: str,
-    n_envs: int = 1,
-    action_space_type: str = "joint_angle",
-    dataset: str = "test",
-    episode_length: int = 300,
-    task_ids: list[int] | None = None,
-    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
-) -> dict[str, dict[int, gym.vector.VectorEnv]]:
-    """Create vectorized RoboMME environments for evaluation.
-
-    `task` may be a single RoboMME task name (e.g. "PickXtimes") or a
-    comma-separated list (e.g. "PickXtimes,BinFill,StopCube"). Each task
-    becomes its own suite in the returned mapping.
-
-    Returns {suite_name: {task_id: VectorEnv}} matching lerobot's expected format.
-    """
-    if env_cls is None or not callable(env_cls):
-        raise ValueError("env_cls must be a callable that wraps a list of env factory callables.")
-    if not isinstance(n_envs, int) or n_envs <= 0:
-        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
-
-    if task_ids is None:
-        task_ids = [0]
-
-    task_names = [t.strip() for t in task.split(",") if t.strip()]
-    is_async = env_cls is gym.vector.AsyncVectorEnv
-    cached_obs_space: spaces.Space | None = None
-    cached_act_space: spaces.Space | None = None
-    cached_metadata: dict[str, Any] | None = None
-    out: dict[str, dict[int, gym.vector.VectorEnv]] = {}
-    for task_name in task_names:
-        envs_by_task: dict[int, gym.vector.VectorEnv] = {}
-        for task_id in task_ids:
-            fns = _make_env_fns(
-                task=task_name,
-                n_envs=n_envs,
-                action_space_type=action_space_type,
-                dataset=dataset,
-                episode_length=episode_length,
-                task_id=task_id,
-            )
-            if is_async:
-                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
-                if cached_obs_space is None:
-                    cached_obs_space = lazy.observation_space
-                    cached_act_space = lazy.action_space
-                    cached_metadata = lazy.metadata
-                envs_by_task[task_id] = lazy
-            else:
-                envs_by_task[task_id] = env_cls(fns)
-        out[task_name] = envs_by_task
-    return out
--- a/src/lerobot/envs/robotwin.py
+++ b/src/lerobot/envs/robotwin.py
@@ -1,488 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-from __future__ import annotations
-
-import importlib
-import logging
-from collections import defaultdict
-from collections.abc import Callable, Sequence
-from functools import partial
-from typing import Any
-
-import gymnasium as gym
-import numpy as np
-import torch
-from gymnasium import spaces
-
-from lerobot.types import RobotObservation
-
-from .utils import _LazyAsyncVectorEnv
-
-logger = logging.getLogger(__name__)
-
-# Camera names as used by RoboTwin 2.0. The wrapper appends "_rgb" when looking
-# up keys in get_obs() output (e.g. "head_camera" → "head_camera_rgb").
-ROBOTWIN_CAMERA_NAMES: tuple[str, ...] = (
-    "head_camera",
-    "left_camera",
-    "right_camera",
-)
-
-ACTION_DIM = 14  # 7 DOF × 2 arms
-ACTION_LOW = -1.0
-ACTION_HIGH = 1.0
-DEFAULT_EPISODE_LENGTH = 300
-# D435 dims from task_config/_camera_config.yml (what demo_clean.yml selects).
-DEFAULT_CAMERA_H = 240
-DEFAULT_CAMERA_W = 320
-
-# Task list from RoboTwin 2.0's `envs/` directory — mirrors upstream exactly
-# (50 tasks as of main; earlier revisions had 60 with a different split).
-# Keep this in sync with:
-#   gh api /repos/RoboTwin-Platform/RoboTwin/contents/envs --paginate \
-#     | jq -r '.[].name' | grep -E '\.py$' | grep -v '^_' | sed 's/\.py$//'
-ROBOTWIN_TASKS: tuple[str, ...] = (
-    "adjust_bottle",
-    "beat_block_hammer",
-    "blocks_ranking_rgb",
-    "blocks_ranking_size",
-    "click_alarmclock",
-    "click_bell",
-    "dump_bin_bigbin",
-    "grab_roller",
-    "handover_block",
-    "handover_mic",
-    "hanging_mug",
-    "lift_pot",
-    "move_can_pot",
-    "move_pillbottle_pad",
-    "move_playingcard_away",
-    "move_stapler_pad",
-    "open_laptop",
-    "open_microwave",
-    "pick_diverse_bottles",
-    "pick_dual_bottles",
-    "place_a2b_left",
-    "place_a2b_right",
-    "place_bread_basket",
-    "place_bread_skillet",
-    "place_burger_fries",
-    "place_can_basket",
-    "place_cans_plasticbox",
-    "place_container_plate",
-    "place_dual_shoes",
-    "place_empty_cup",
-    "place_fan",
-    "place_mouse_pad",
-    "place_object_basket",
-    "place_object_scale",
-    "place_object_stand",
-    "place_phone_stand",
-    "place_shoe",
-    "press_stapler",
-    "put_bottles_dustbin",
-    "put_object_cabinet",
-    "rotate_qrcode",
-    "scan_object",
-    "shake_bottle",
-    "shake_bottle_horizontally",
-    "stack_blocks_three",
-    "stack_blocks_two",
-    "stack_bowls_three",
-    "stack_bowls_two",
-    "stamp_seal",
-    "turn_switch",
-)
-
-
-_ROBOTWIN_SETUP_CACHE: dict[str, dict[str, Any]] = {}
-
-
-def _load_robotwin_setup_kwargs(task_name: str) -> dict[str, Any]:
-    """Build the kwargs dict RoboTwin's setup_demo expects.
-
-    Mirrors the config loading done by RoboTwin's ``script/eval_policy.py``:
-    reads ``task_config/demo_clean.yml``, resolves the embodiment file from
-    ``_embodiment_config.yml``, loads the robot's own ``config.yml``, and
-    reads camera dimensions from ``_camera_config.yml``.
-
-    Uses ``aloha-agilex`` single-robot dual-arm by default (the only embodiment
-    used by beat_block_hammer and most smoke-test tasks).
-    """
-    if task_name in _ROBOTWIN_SETUP_CACHE:
-        return dict(_ROBOTWIN_SETUP_CACHE[task_name])
-
-    import os
-
-    import yaml  # type: ignore[import-untyped]
-    from envs import CONFIGS_PATH  # type: ignore[import-not-found]
-
-    task_config = "demo_clean"
-    with open(os.path.join(CONFIGS_PATH, f"{task_config}.yml"), encoding="utf-8") as f:
-        args = yaml.safe_load(f)
-
-    # Resolve embodiment — demo_clean.yml uses [aloha-agilex] (dual-arm single robot)
-    with open(os.path.join(CONFIGS_PATH, "_embodiment_config.yml"), encoding="utf-8") as f:
-        embodiment_types = yaml.safe_load(f)
-    embodiment = args.get("embodiment", ["aloha-agilex"])
-    if len(embodiment) == 1:
-        robot_file = embodiment_types[embodiment[0]]["file_path"]
-        args["left_robot_file"] = robot_file
-        args["right_robot_file"] = robot_file
-        args["dual_arm_embodied"] = True
-    elif len(embodiment) == 3:
-        args["left_robot_file"] = embodiment_types[embodiment[0]]["file_path"]
-        args["right_robot_file"] = embodiment_types[embodiment[1]]["file_path"]
-        args["embodiment_dis"] = embodiment[2]
-        args["dual_arm_embodied"] = False
-    else:
-        raise ValueError(f"embodiment must have 1 or 3 items, got {len(embodiment)}")
-
-    with open(os.path.join(args["left_robot_file"], "config.yml"), encoding="utf-8") as f:
-        args["left_embodiment_config"] = yaml.safe_load(f)
-    with open(os.path.join(args["right_robot_file"], "config.yml"), encoding="utf-8") as f:
-        args["right_embodiment_config"] = yaml.safe_load(f)
-
-    # Camera dimensions
-    with open(os.path.join(CONFIGS_PATH, "_camera_config.yml"), encoding="utf-8") as f:
-        camera_config = yaml.safe_load(f)
-    head_cam = args["camera"]["head_camera_type"]
-    args["head_camera_h"] = camera_config[head_cam]["h"]
-    args["head_camera_w"] = camera_config[head_cam]["w"]
-
-    # Headless overrides
-    args["render_freq"] = 0
-    args["task_name"] = task_name
-    args["task_config"] = task_config
-
-    _ROBOTWIN_SETUP_CACHE[task_name] = args
-    return dict(args)
-
-
-def _load_robotwin_task(task_name: str) -> type:
-    """Dynamically import and return a RoboTwin 2.0 task class.
-
-    RoboTwin tasks live in ``envs/<task_name>.py`` relative to the repository
-    root and are expected to be on ``sys.path`` after installation.
-    """
-    try:
-        module = importlib.import_module(f"envs.{task_name}")
-    except ModuleNotFoundError as e:
-        raise ModuleNotFoundError(
-            f"Could not import RoboTwin task '{task_name}'. "
-            "Ensure RoboTwin 2.0 is installed and its 'envs/' directory is on PYTHONPATH. "
-            "See the RoboTwin installation guide: https://robotwin-platform.github.io/doc/usage/robotwin-install.html"
-        ) from e
-    task_cls = getattr(module, task_name, None)
-    if task_cls is None:
-        raise AttributeError(f"Task class '{task_name}' not found in envs/{task_name}.py")
-    return task_cls
-
-
-class RoboTwinEnv(gym.Env):
-    """Gymnasium wrapper around a single RoboTwin 2.0 task.
-
-    RoboTwin uses a custom SAPIEN-based API (``setup_demo`` / ``get_obs`` /
-    ``take_action`` / ``check_success``) rather than the standard gym interface.
-    This class bridges that API to Gymnasium so that ``lerobot-eval`` can drive
-    RoboTwin exactly like LIBERO or Meta-World.
-
-    The underlying SAPIEN environment is created lazily on the first ``reset()``
-    call *inside the worker process*.  This is required for
-    ``gym.vector.AsyncVectorEnv`` compatibility: SAPIEN allocates EGL/GPU
-    contexts that must not be forked from the parent process.
-
-    Observations
-    ------------
-    The ``pixels`` dict uses the raw RoboTwin camera names as keys (e.g.
-    ``"head_camera"``, ``"left_camera"``). ``preprocess_observation`` in
-    ``envs/utils.py`` then converts these to ``observation.images.<cam>``.
-
-    Actions
-    -------
-    14-dim float32 array in ``[-1, 1]`` (joint-space, 7 DOF per arm).
-
-    Autograd
-    --------
-    ``setup_demo`` and ``take_action`` drive CuRobo's Newton trajectory
-    optimizer, which calls ``cost.backward()`` internally. lerobot_eval wraps
-    the rollout in ``torch.no_grad()``, so both call sites re-enable grad.
-    """
-
-    metadata = {"render_modes": ["rgb_array"], "render_fps": 25}
-
-    def __init__(
-        self,
-        task_name: str,
-        episode_index: int = 0,
-        n_envs: int = 1,
-        camera_names: Sequence[str] = ROBOTWIN_CAMERA_NAMES,
-        observation_height: int | None = None,
-        observation_width: int | None = None,
-        episode_length: int = DEFAULT_EPISODE_LENGTH,
-        render_mode: str = "rgb_array",
-    ):
-        super().__init__()
-        self.task_name = task_name
-        self.task = task_name  # used by add_envs_task() in utils.py
-        self.task_description = task_name.replace("_", " ")
-        self.episode_index = episode_index
-        self._reset_stride = n_envs
-        self.camera_names = list(camera_names)
-        # Default to D435 dims (the camera type baked into task_config/demo_clean.yml).
-        # The YAML-driven lookup is deferred to reset() so construction doesn't
-        # import RoboTwin's `envs` module — fast-tests run without RoboTwin installed.
-        self.observation_height = observation_height or DEFAULT_CAMERA_H
-        self.observation_width = observation_width or DEFAULT_CAMERA_W
-        self.episode_length = episode_length
-        self._max_episode_steps = episode_length  # lerobot_eval.rollout reads this
-        self.render_mode = render_mode
-
-        self._env: Any | None = None  # deferred — created on first reset() inside worker
-        self._step_count: int = 0
-        self._black_frame = np.zeros((self.observation_height, self.observation_width, 3), dtype=np.uint8)
-
-        image_spaces = {
-            cam: spaces.Box(
-                low=0,
-                high=255,
-                shape=(self.observation_height, self.observation_width, 3),
-                dtype=np.uint8,
-            )
-            for cam in self.camera_names
-        }
-        self.observation_space = spaces.Dict(
-            {
-                "pixels": spaces.Dict(image_spaces),
-                "agent_pos": spaces.Box(low=-np.inf, high=np.inf, shape=(ACTION_DIM,), dtype=np.float32),
-            }
-        )
-        self.action_space = spaces.Box(
-            low=ACTION_LOW, high=ACTION_HIGH, shape=(ACTION_DIM,), dtype=np.float32
-        )
-
-    def _ensure_env(self) -> None:
-        """Create the SAPIEN environment on first use.
-
-        Called inside the worker subprocess after fork(), so each worker gets
-        its own EGL/GPU context rather than inheriting a stale one from the
-        parent process (which causes crashes with AsyncVectorEnv).
-        """
-        if self._env is not None:
-            return
-        task_cls = _load_robotwin_task(self.task_name)
-        self._env = task_cls()
-
-    def _get_obs(self) -> RobotObservation:
-        assert self._env is not None, "_get_obs called before _ensure_env()"
-        raw = self._env.get_obs()
-        cameras_raw = raw.get("observation", {})
-
-        images: dict[str, np.ndarray] = {}
-        for cam in self.camera_names:
-            cam_data = cameras_raw.get(cam)
-            img = cam_data.get("rgb") if cam_data else None
-            if img is None:
-                images[cam] = self._black_frame
-                continue
-            img = np.asarray(img, dtype=np.uint8)
-            if img.ndim == 2:
-                img = np.stack([img, img, img], axis=-1)
-            elif img.shape[-1] != 3:
-                img = img[..., :3]
-            images[cam] = img
-
-        ja = raw.get("joint_action") or {}
-        vec = ja.get("vector")
-        if vec is not None:
-            arr = np.asarray(vec, dtype=np.float32).ravel()
-            joint_state = (
-                arr[:ACTION_DIM] if arr.size >= ACTION_DIM else np.zeros(ACTION_DIM, dtype=np.float32)
-            )
-        else:
-            joint_state = np.zeros(ACTION_DIM, dtype=np.float32)
-
-        return {"pixels": images, "agent_pos": joint_state}
-
-    def reset(self, seed: int | None = None, **kwargs) -> tuple[RobotObservation, dict]:
-        self._ensure_env()
-        super().reset(seed=seed)
-        assert self._env is not None  # set by _ensure_env() above
-
-        actual_seed = self.episode_index if seed is None else seed
-        setup_kwargs = _load_robotwin_setup_kwargs(self.task_name)
-        setup_kwargs.update(seed=actual_seed, is_test=True)
-        with torch.enable_grad():
-            self._env.setup_demo(**setup_kwargs)
-        self.episode_index += self._reset_stride
-        self._step_count = 0
-
-        obs = self._get_obs()
-        return obs, {"is_success": False, "task": self.task_name}
-
-    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
-        assert self._env is not None, "step() called before reset()"
-        if action.ndim != 1 or action.shape[0] != ACTION_DIM:
-            raise ValueError(f"Expected 1-D action of shape ({ACTION_DIM},), got {action.shape}")
-
-        with torch.enable_grad():
-            if hasattr(self._env, "take_action"):
-                self._env.take_action(action)
-            else:
-                self._env.step(action)
-
-        self._step_count += 1
-
-        is_success = bool(getattr(self._env, "eval_success", False))
-        if not is_success and hasattr(self._env, "check_success"):
-            is_success = bool(self._env.check_success())
-
-        obs = self._get_obs()
-        reward = float(is_success)
-        terminated = is_success
-        truncated = self._step_count >= self.episode_length
-
-        info: dict[str, Any] = {
-            "task": self.task_name,
-            "is_success": is_success,
-            "step": self._step_count,
-        }
-        if terminated or truncated:
-            info["final_info"] = {
-                "task": self.task_name,
-                "is_success": is_success,
-            }
-            self.reset()
-
-        return obs, reward, terminated, truncated, info
-
-    def render(self) -> np.ndarray:
-        self._ensure_env()
-        obs = self._get_obs()
-        # Prefer head camera for rendering; fall back to first available.
-        if "head_camera" in obs["pixels"]:
-            return obs["pixels"]["head_camera"]
-        return next(iter(obs["pixels"].values()))
-
-    def close(self) -> None:
-        if self._env is not None:
-            if hasattr(self._env, "close_env"):
-                import contextlib
-
-                with contextlib.suppress(TypeError):
-                    self._env.close_env()
-            self._env = None
-
-
-# ---- Multi-task factory --------------------------------------------------------
-
-
-def _make_env_fns(
-    *,
-    task_name: str,
-    n_envs: int,
-    camera_names: list[str],
-    observation_height: int,
-    observation_width: int,
-    episode_length: int,
-) -> list[Callable[[], RoboTwinEnv]]:
-    """Return n_envs factory callables for a single task."""
-
-    def _make_one(episode_index: int) -> RoboTwinEnv:
-        return RoboTwinEnv(
-            task_name=task_name,
-            episode_index=episode_index,
-            n_envs=n_envs,
-            camera_names=camera_names,
-            observation_height=observation_height,
-            observation_width=observation_width,
-            episode_length=episode_length,
-        )
-
-    return [partial(_make_one, i) for i in range(n_envs)]
-
-
-def create_robotwin_envs(
-    task: str,
-    n_envs: int,
-    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
-    camera_names: Sequence[str] = ROBOTWIN_CAMERA_NAMES,
-    observation_height: int = DEFAULT_CAMERA_H,
-    observation_width: int = DEFAULT_CAMERA_W,
-    episode_length: int = DEFAULT_EPISODE_LENGTH,
-) -> dict[str, dict[int, Any]]:
-    """Create vectorized RoboTwin 2.0 environments.
-
-    Returns:
-        ``dict[task_name][0] -> VectorEnv`` — one entry per task, each wrapping
-        ``n_envs`` parallel rollouts.
-
-    Args:
-        task: Comma-separated list of task names (e.g. ``"beat_block_hammer"``
-            or ``"beat_block_hammer,click_bell"``).
-        n_envs: Number of parallel rollouts per task.
-        env_cls: Vector env constructor (e.g. ``gym.vector.AsyncVectorEnv``).
-        camera_names: Cameras to include in observations.
-        observation_height: Pixel height for all cameras.
-        observation_width: Pixel width for all cameras.
-        episode_length: Max steps before truncation.
-    """
-    if env_cls is None or not callable(env_cls):
-        raise ValueError("env_cls must be callable (e.g. gym.vector.AsyncVectorEnv).")
-    if not isinstance(n_envs, int) or n_envs <= 0:
-        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
-
-    task_names = [t.strip() for t in str(task).split(",") if t.strip()]
-    if not task_names:
-        raise ValueError("`task` must contain at least one RoboTwin task name.")
-
-    unknown = [t for t in task_names if t not in ROBOTWIN_TASKS]
-    if unknown:
-        raise ValueError(f"Unknown RoboTwin tasks: {unknown}. Available tasks: {sorted(ROBOTWIN_TASKS)}")
-
-    logger.info(
-        "Creating RoboTwin envs | tasks=%s | n_envs(per task)=%d",
-        task_names,
-        n_envs,
-    )
-
-    is_async = env_cls is gym.vector.AsyncVectorEnv
-    cached_obs_space: spaces.Space | None = None
-    cached_act_space: spaces.Space | None = None
-    cached_metadata: dict[str, Any] | None = None
-
-    out: dict[str, dict[int, Any]] = defaultdict(dict)
-    for task_name in task_names:
-        fns = _make_env_fns(
-            task_name=task_name,
-            n_envs=n_envs,
-            camera_names=list(camera_names),
-            observation_height=observation_height,
-            observation_width=observation_width,
-            episode_length=episode_length,
-        )
-        if is_async:
-            lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
-            if cached_obs_space is None:
-                cached_obs_space = lazy.observation_space
-                cached_act_space = lazy.action_space
-                cached_metadata = lazy.metadata
-            out[task_name][0] = lazy
-        else:
-            out[task_name][0] = env_cls(fns)
-        logger.info("Built vec env | task=%s | n_envs=%d", task_name, n_envs)
-
-    return {k: dict(v) for k, v in out.items()}
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Steven Palma	8e21268c29	test: add dataset guard + fix imports	2026-04-20 00:36:02 +02:00
Steven Palma	4130d4a4a5	update docs + docstrings + examples + add minimal test	2026-04-19 23:53:53 +02:00
Steven Palma	47bb840a55	add context guards	2026-04-19 23:21:14 +02:00
Steven Palma	9519ff5e09	Merge branch 'main' into feat/decouple_record_script Signed-off-by: Steven Palma <imstevenpmwork@ieee.org>	2026-04-19 22:48:08 +02:00
Steven Palma	32a27cae8a	filesize default change + more logs + filesize mb based episode + go back to init pos + rerun log + date end of repo_id	2026-04-19 16:50:19 +02:00
Steven Palma	8cee56e2d6	fix pre-commit	2026-04-17 16:46:58 +02:00
Steven Palma	a76874f35e	test dagger	2026-04-17 16:46:38 +02:00
Steven Palma	35bb2c7459	simplify dagger	2026-04-17 15:55:03 +02:00
Steven Palma	051f6c6803	Merge branch 'main' into feat/decouple_record_script	2026-04-17 14:25:18 +02:00
Steven Palma	04ae0312a2	HW tests fixes	2026-04-16 17:29:22 +02:00
Steven Palma	cc634de9e7	add docstrings	2026-04-16 16:40:33 +02:00
Steven Palma	3eda5712d3	some more iterations	2026-04-16 15:52:23 +02:00
Steven Palma	783ec6e232	minor improvements	2026-04-16 14:34:22 +02:00
Steven Palma	4e3175ff15	address review	2026-04-15 19:31:53 +02:00
Steven Palma	edd7fc52a8	feat: introduce inference engine strategy	2026-04-15 17:51:44 +02:00
Steven Palma	0f0f8b8961	imports and comments	2026-04-15 16:28:56 +02:00
Steven Palma	79db54dc34	Merge branch 'main' into feat/decouple_record_script	2026-04-15 11:06:45 +02:00
Steven Palma	6ae07878f7	Merge branch 'main' into feat/decouple_record_script	2026-04-14 22:54:29 +02:00
Steven Palma	10d05e03bc	Merge branch 'main' into feat/decouple_record_script	2026-04-14 21:35:26 +02:00
Steven Palma	f2c29d78cf	more improvements and fixes	2026-04-14 17:51:03 +02:00
Steven Palma	8bc47e4318	target review	2026-04-14 17:14:09 +02:00
Steven Palma	49f32b9796	some more iterations	2026-04-14 16:34:52 +02:00
Steven Palma	f55782f9f7	pre-commit run	2026-04-14 15:42:19 +02:00
Steven Palma	05a2604d6e	first iteration	2026-04-14 15:42:04 +02:00