fix(imports): fixing av import in test_depth.py

* Enhance documentation with Lance format details

Added information about Lance format and `lerobot-lancedb` package for multimodal AI datasets.

Signed-off-by: Quentin Lhoest <42851186+lhoestq@users.noreply.github.com>

.github/PULL_REQUEST_TEMPLATE.md

@@ -2,11 +2,6 @@
 
 Short, imperative summary (e.g., "fix(robots): handle None in sensor parser"). See [CONTRIBUTING.md](../CONTRIBUTING.md) for PR conventions.
 
-## Type / Scope
-
-- **Type**: (Bug | Feature | Docs | Performance | Test | CI | Chore)
-- **Scope**: (optional — name of module or package affected)
-
 ## Summary / Motivation
 
 - One-paragraph description of what changes and why.
@@ -19,28 +14,14 @@ Short, imperative summary (e.g., "fix(robots): handle None in sensor parser"). S
 
 ## What changed
 
-- Short, concrete bullets of the modifications (files/behaviour).
+- Short, concrete bullets explaining the functional changes (how the behavior or output differs now).
 - Short note if this introduces breaking changes and migration steps.
 
 ## How was this tested (or how to run locally)
 
-- Tests added: list new tests or test files.
+- Tests added: list new tests or test files. `pytest -q tests/ -k <keyword>`
 - Manual checks / dataset runs performed.
-- Instructions for the reviewer
-
-Example:
-
-- Ran the relevant tests:
-
-  ```bash
-  pytest -q tests/ -k <keyword>
-  ```
-
-- Reproduce with a quick example or CLI (if applicable):
-
-  ```bash
-  lerobot-train --some.option=true
-  ```
+- Instructions for the reviewer for reproducing with a quick example or CLI (if applicable)
 
 ## Checklist (required before merge)
 
@@ -48,6 +29,7 @@ Example:
 - [ ] All tests pass locally (`pytest`)
 - [ ] Documentation updated
 - [ ] CI is green
+- [ ] Community Review: I have reviewed another contributor's open PR and linked it here: # (insert PR number/link)
 
 ## Reviewer notes
 

.github/workflows/benchmark_tests.yml

@@ -83,10 +83,13 @@ 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
@@ -115,7 +118,7 @@ jobs:
             bash -c "
               hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
               lerobot-eval \
-                --policy.path=pepijn223/smolvla_libero \
+                --policy.path=lerobot/smolvla_libero \
                 --env.type=libero \
                 --env.task=libero_spatial \
                 --eval.batch_size=1 \
@@ -144,7 +147,7 @@ jobs:
             --artifacts-dir /tmp/libero-artifacts \
             --env libero \
             --task libero_spatial \
-            --policy pepijn223/smolvla_libero
+            --policy lerobot/smolvla_libero
 
       - name: Upload Libero rollout video
         if: always()
@@ -238,10 +241,13 @@ 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]
@@ -264,7 +270,7 @@ jobs:
             bash -c "
               hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
               lerobot-eval \
-                --policy.path=pepijn223/smolvla_metaworld \
+                --policy.path=lerobot/smolvla_metaworld \
                 --env.type=metaworld \
                 --env.task=metaworld-push-v3 \
                 --eval.batch_size=1 \
@@ -293,7 +299,7 @@ jobs:
             --artifacts-dir /tmp/metaworld-artifacts \
             --env metaworld \
             --task metaworld-push-v3 \
-            --policy pepijn223/smolvla_metaworld
+            --policy lerobot/smolvla_metaworld
 
       - name: Upload MetaWorld rollout video
         if: always()
@@ -310,3 +316,636 @@ 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

.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@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
     with:
       package_name: lerobot
     secrets:

.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@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
+    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # 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@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
+    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
     with:
       commit_sha: ${{ github.event.pull_request.head.sha }}
       pr_number: ${{ github.event.number }}

.github/workflows/latest_deps_tests.yml

@@ -217,6 +217,24 @@ jobs:
       - name: Run end-to-end tests
         run: make test-end-to-end
 
+  slack-notification:
+    name: Slack Notification
+    needs: [cpu-tests, gpu-tests, upgrade-lock]
+    if: always() && needs.upgrade-lock.outputs.changed == 'true'
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    env:
+      CI_SLACK_CHANNEL: ${{ secrets.CI_SLACK_CHANNEL }}
+    steps:
+      - name: Post to a Slack channel
+        uses: huggingface/hf-workflows/.github/actions/post-slack@a88e7fa2eaee28de5a4d6142381b1fb792349b67  # main
+        with:
+          slack_channel: ${{ env.CI_SLACK_CHANNEL }}
+          title: "Results of the latest dependency tests (CPU + GPU)"
+          status: ${{ (needs.cpu-tests.result == 'success' && needs.gpu-tests.result == 'success') && 'success' || 'failure' }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
   # This job creates or updates a PR with the upgraded lockfile
   open-pr:
     name: Open PR

.github/workflows/release.yml

@@ -152,13 +152,14 @@ 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 $VERSION from PyPI..."
-            uv pip install "lerobot[all]==$VERSION"
+            uv pip install --torch-backend cpu "lerobot[all]==$VERSION"
           fi
       - name: Check lerobot version
         run: uv run python -c "import lerobot; print(lerobot.__version__)"

.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 14 days with no activity.
+    This issue was closed because it has been stalled for 30 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 21 days with no activity.
+    This PR was closed because it has been stalled for 30 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 (6 months). It will be closed if no further activity occurs.
+    recent activity (1 year). 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: 180
-          days-before-issue-close: 14
+          days-before-issue-stale: 365
+          days-before-issue-close: 30
           days-before-pr-stale: 365
-          days-before-pr-close: 21
+          days-before-pr-close: 30
           delete-branch: true
           close-issue-message: ${{ env.CLOSE_ISSUE_MESSAGE }}
           close-pr-message: ${{ env.CLOSE_PR_MESSAGE }}

AGENTS.md

@@ -1,5 +1,7 @@
 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.

AGENT_GUIDE.md

@@ -0,0 +1,412 @@
+# 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).

CONTRIBUTING.md

@@ -78,6 +78,9 @@ Use the templates for required fields and examples.
 - **Issues:** Follow the [ticket template](https://github.com/huggingface/lerobot/blob/main/.github/ISSUE_TEMPLATE/bug-report.yml).
 - **Pull requests:** Rebase on `upstream/main`, use a descriptive branch (don't work on `main`), run `pre-commit` and tests locally, and follow the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md).
 
-One member of the LeRobot team will then review your contribution.
+> [!IMPORTANT]
+> Community Review Policy: To help scale our efforts and foster a collaborative environment, we ask contributors to review at least one other person's open PR before their own receives attention. This shared responsibility multiplies our review capacity and helps everyone's code get merged faster!
+
+Once you have submitted your PR and completed a peer review, a member of the LeRobot team will review your contribution.
 
 Thank you for contributing to LeRobot!

MANIFEST.in

@@ -1,3 +1,4 @@
 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

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 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).
 
 ## Inference & Evaluation
 

benchmarks/video/README.md

@@ -1,288 +0,0 @@
-# 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%**   |

benchmarks/video/run_video_benchmark.py

@@ -1,488 +0,0 @@
-#!/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))

docker/Dockerfile.benchmark.libero_plus

@@ -0,0 +1,84 @@
+# 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"]

docker/Dockerfile.benchmark.robocasa

@@ -0,0 +1,71 @@
+# 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"]

docker/Dockerfile.benchmark.robocerebra

@@ -0,0 +1,43 @@
+# 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"]

docker/Dockerfile.benchmark.robomme

@@ -0,0 +1,56 @@
+# 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"]

docker/Dockerfile.benchmark.robotwin

@@ -0,0 +1,138 @@
+# 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"]

docker/Dockerfile.benchmark.vlabench

@@ -0,0 +1,99 @@
+# 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"]

docker/Dockerfile.internal

@@ -18,9 +18,8 @@
 # docker build -f docker/Dockerfile.internal -t lerobot-internal .
 
 # Configure the base image for CI with GPU access
-# TODO(Steven): Bump these versions
-ARG CUDA_VERSION=12.4.1
-ARG OS_VERSION=22.04
+ARG CUDA_VERSION=12.8.1
+ARG OS_VERSION=24.04
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
 
 # Define Python version argument
@@ -36,16 +35,13 @@ ENV DEBIAN_FRONTEND=noninteractive \
 
 # Install Python, system dependencies, and uv (as root)
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    software-properties-common build-essential git curl \
-    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    build-essential git curl \
+    libglib2.0-0 libgl1 libegl1 ffmpeg \
     libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
     cmake pkg-config ninja-build \
-    && 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${PYTHON_VERSION}-dev \
+    python${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 \

docs/source/_toctree.yml

@@ -3,12 +3,14 @@
     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: Bring Your Own Policies
+    title: Adding a Policy
   - local: integrate_hardware
     title: Bring Your Own Hardware
   - local: hilserl
@@ -24,6 +26,12 @@
   - 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
@@ -33,6 +41,8 @@
     title: Using the Dataset Tools
   - local: dataset_subtask
     title: Using Subtasks in the Dataset
+  - local: video_encoding_parameters
+    title: Video encoding parameters
   - local: streaming_video_encoding
     title: Streaming Video Encoding
   title: "Datasets"
@@ -47,6 +57,8 @@
     title: π₀-FAST (Pi0Fast)
   - local: pi05
     title: π₀.₅ (Pi05)
+  - local: eo1
+    title: EO-1
   - local: groot
     title: NVIDIA GR00T N1.5
   - local: xvla
@@ -61,6 +73,8 @@
     title: SARM
   title: "Reward Models"
 - sections:
+  - local: inference
+    title: Policy Deployment (lerobot-rollout)
   - local: async
     title: Use Async Inference
   - local: rtc
@@ -77,10 +91,22 @@
     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
@@ -126,10 +152,6 @@
   - local: cameras
     title: Cameras
   title: "Sensors"
-- sections:
-  - local: torch_accelerators
-    title: PyTorch accelerators
-  title: "Supported Hardware"
 - sections:
   - local: notebooks
     title: Notebooks

docs/source/act.mdx

@@ -90,6 +90,6 @@ lerobot-record \
   --dataset.single_task="Your task description" \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
+  # --dataset.camera_encoder.vcodec=auto \
   --policy.path=${HF_USER}/act_policy
 ```

docs/source/bring_your_own_policies.mdx

@@ -1,60 +1,37 @@
-# Bring Your Own Policies
+# Adding a Policy
 
-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.
+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:
 
-## Step 1: Create a Policy Package
+- **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.
 
-Your custom policy should be organized as an installable Python package following LeRobot's plugin conventions.
+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.
 
-### Package Structure
+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)).
 
-Create a package with the prefix `lerobot_policy_` (IMPORTANT!) followed by your policy name:
+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.
 
-```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
-```
+---
 
-### Package Configuration
+## Anatomy of a policy
 
-Set up your `pyproject.toml`:
+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).
 
-```toml
-[project]
-name = "lerobot_policy_my_custom_policy"
-version = "0.1.0"
-dependencies = [
-    # your policy-specific dependencies
-]
-requires-python = ">= 3.12"
+### Configuration class
 
-[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.
+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.
 
 ```python
-# configuration_my_custom_policy.py
+# configuration_my_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_custom_policy")
+@PreTrainedConfig.register_subclass("my_policy")
 @dataclass
-class MyCustomPolicyConfig(PreTrainedConfig):
-    """Configuration class for MyCustomPolicy.
+class MyPolicyConfig(PreTrainedConfig):
+    """Configuration class for MyPolicy.
 
     Args:
         n_obs_steps: Number of observation steps to use as input
@@ -77,16 +54,20 @@ class MyCustomPolicyConfig(PreTrainedConfig):
             raise ValueError("n_action_steps cannot exceed horizon")
 
     def validate_features(self) -> None:
-        """Validate input/output feature compatibility."""
+        """Validate input/output feature compatibility.
+
+        Call this explicitly from your policy's __init__ — the base class does not.
+        """
         if not self.image_features:
-            raise ValueError("MyCustomPolicy requires at least one image feature.")
+            raise ValueError("MyPolicy requires at least one image feature.")
         if self.action_feature is None:
-            raise ValueError("MyCustomPolicy requires 'action' in output_features.")
+            raise ValueError("MyPolicy 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
@@ -101,8 +82,7 @@ class MyCustomPolicyConfig(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
@@ -110,32 +90,34 @@ class MyCustomPolicyConfig(PreTrainedConfig):
         return None
 ```
 
-## Step 3: Implement the Policy Class
+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.
 
-Create your policy implementation by inheriting from [`PreTrainedPolicy`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pretrained.py):
+### 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__`:
 
 ```python
-# modeling_my_custom_policy.py
+# modeling_my_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_custom_policy import MyCustomPolicyConfig
+from .configuration_my_policy import MyPolicyConfig
 
-class MyCustomPolicy(PreTrainedPolicy):
-    config_class = MyCustomPolicyConfig  # must match the string in @register_subclass
-    name = "my_custom_policy"
+class MyPolicy(PreTrainedPolicy):
+    config_class = MyPolicyConfig  # must match the string in @register_subclass
+    name = "my_policy"
 
-    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: dict[str, Any] = None):
+    def __init__(self, config: MyPolicyConfig, 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 episode state."""
+        """Reset per-episode state. Called by lerobot-eval at the start of each episode."""
         ...
 
     def get_optim_params(self) -> dict:
@@ -147,35 +129,51 @@ class MyCustomPolicy(PreTrainedPolicy):
         ...
 
     def select_action(self, batch: dict[str, torch.Tensor], **kwargs) -> torch.Tensor:
-        """Return a single action for the current timestep (called at inference)."""
+        """Return a single action for the current timestep (called every step at inference)."""
         ...
 
-    def forward(self, batch: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
+    def forward(self, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, dict | None]:
         """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": ...}
+        return loss, {"some_loss_component": some_loss_component.item()}
 ```
 
-## Step 4: Add Data Processors
+The methods called by the train/eval loops:
 
-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).
+| 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).
 
 ```python
-# processor_my_custom_policy.py
+# processor_my_policy.py
 from typing import Any
 import torch
 
 from lerobot.processor import PolicyAction, PolicyProcessorPipeline
 
 
-def make_my_custom_policy_pre_post_processors(
+def make_my_policy_pre_post_processors(
     config,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
 ) -> tuple[
@@ -187,11 +185,48 @@ def make_my_custom_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
+---
 
-Expose your classes in the package's `__init__.py`:
+## 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`:
 
 ```python
 # __init__.py
@@ -204,44 +239,148 @@ except ImportError:
         "lerobot is not installed. Please install lerobot to use this policy package."
     )
 
-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
+from .configuration_my_policy import MyPolicyConfig
+from .modeling_my_policy import MyPolicy
+from .processor_my_policy import make_my_policy_pre_post_processors
 
 __all__ = [
-    "MyCustomPolicyConfig",
-    "MyCustomPolicy",
-    "make_my_custom_policy_pre_post_processors",
+    "MyPolicyConfig",
+    "MyPolicy",
+    "make_my_policy_pre_post_processors",
 ]
 ```
 
-## Step 6: Installation and Usage
-
-### Install Your Policy Package
+### Install and use
 
 ```bash
-cd lerobot_policy_my_custom_policy
+cd lerobot_policy_my_policy
 pip install -e .
 
 # Or install from PyPI if published
-pip install lerobot_policy_my_custom_policy
+pip install lerobot_policy_my_policy
 ```
 
-### Use Your Policy
-
 Once installed, your policy automatically integrates with LeRobot's training and evaluation tools:
 
 ```bash
 lerobot-train \
-    --policy.type my_custom_policy \
+    --policy.type my_policy \
     --env.type pusht \
     --steps 200000
 ```
 
-## Examples and Community Contributions
+---
+
+## 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
 
 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)
 
-Share your policy implementations with the community! 🤗
+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. 🤗

docs/source/cheat-sheet.mdx

@@ -0,0 +1,139 @@
+# 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
+```

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.vcodec=auto \
+    # --dataset.camera_encoder.vcodec=auto \
     --display_data=true
 ```
 

docs/source/eo1.mdx

@@ -0,0 +1,168 @@
+# 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.

docs/source/groot.mdx

@@ -123,7 +123,7 @@ lerobot-record \
   --dataset.single_task="Grab and handover the red cube to the other arm" \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
+  # --dataset.camera_encoder.vcodec=auto \
   --policy.path=<user>/groot-bimanual \ # your trained model
   --dataset.episode_time_s=30 \
   --dataset.reset_time_s=10

docs/source/hardware_guide.mdx

@@ -0,0 +1,98 @@
+# 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).

docs/source/hil_data_collection.mdx

@@ -50,30 +50,30 @@ This process can be repeated iteratively: deploy, collect, fine-tune, repeat. Ea
 
 ### Teleoperator Requirements
 
-The `examples/hil` HIL scripts require **teleoperators with active motors** that can:
+The `lerobot-rollout --strategy.type=dagger` mode requires **teleoperators with active motors** that can:
 
 - Enable/disable torque programmatically
 - Move to target positions (to mirror the robot state when pausing)
 
-**Compatible teleoperators in the current `examples/hil` scripts:**
+**Compatible teleoperators:**
 
 - `openarm_mini` - OpenArm Mini
 - `so_leader` - SO100 / SO101 leader arm
 
 > [!IMPORTANT]
-> The provided `examples/hil` commands default to `bi_openarm_follower` + `openarm_mini`.
+> The provided commands default to `bi_openarm_follower` + `openarm_mini`.
 > `so_follower` + `so_leader` configs are also registered and can be used via CLI flags.
 
 ---
 
 ## Script
 
-A single script handles both synchronous and RTC-based inference. Toggle RTC with `--rtc.enabled=true`:
+Use `lerobot-rollout` with `--strategy.type=dagger` for HIL data collection. Select the inference backend with `--inference.type=sync|rtc`:
 
-| Mode                     | Flag                 | Models                |
-| ------------------------ | -------------------- | --------------------- |
-| Standard (default)       | _(no flag needed)_   | ACT, Diffusion Policy |
-| Real-Time Chunking (RTC) | `--rtc.enabled=true` | Pi0, Pi0.5, SmolVLA   |
+| Mode                     | Flag                   | Models                |
+| ------------------------ | ---------------------- | --------------------- |
+| Standard (default)       | _(no flag needed)_     | ACT, Diffusion Policy |
+| Real-Time Chunking (RTC) | `--inference.type=rtc` | Pi0, Pi0.5, SmolVLA   |
 
 ---
 
@@ -97,7 +97,7 @@ python src/lerobot/scripts/lerobot_train.py \
 **Standard inference (ACT, Diffusion Policy):**
 
 ```bash
-python examples/hil/hil_data_collection.py \
+lerobot-rollout --strategy.type=dagger \
     --robot.type=bi_openarm_follower \
     --robot.left_arm_config.port=can1 \
     --robot.left_arm_config.side=left \
@@ -108,11 +108,10 @@ python examples/hil/hil_data_collection.py \
     --teleop.port_left=/dev/ttyACM0 \
     --teleop.port_right=/dev/ttyACM1 \
     --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
-    --dataset.repo_id=your-username/hil-dataset \
+    --dataset.repo_id=your-username/rollout_hil_dataset \
     --dataset.single_task="Fold the T-shirt properly" \
     --dataset.fps=30 \
-    --dataset.episode_time_s=1000 \
-    --dataset.num_episodes=50 \
+    --strategy.num_episodes=50 \
     --interpolation_multiplier=2
 ```
 
@@ -121,11 +120,11 @@ python examples/hil/hil_data_collection.py \
 For models with high inference latency, enable RTC for smooth execution:
 
 ```bash
-python examples/hil/hil_data_collection.py \
-    --rtc.enabled=true \
-    --rtc.execution_horizon=20 \
-    --rtc.max_guidance_weight=5.0 \
-    --rtc.prefix_attention_schedule=LINEAR \
+lerobot-rollout --strategy.type=dagger \
+    --inference.type=rtc \
+    --inference.rtc.execution_horizon=20 \
+    --inference.rtc.max_guidance_weight=5.0 \
+    --inference.rtc.prefix_attention_schedule=LINEAR \
     --robot.type=bi_openarm_follower \
     --robot.left_arm_config.port=can1 \
     --robot.left_arm_config.side=left \
@@ -136,11 +135,10 @@ python examples/hil/hil_data_collection.py \
     --teleop.port_left=/dev/ttyACM0 \
     --teleop.port_right=/dev/ttyACM1 \
     --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
-    --dataset.repo_id=your-username/hil-rtc-dataset \
+    --dataset.repo_id=your-username/rollout_hil_rtc_dataset \
     --dataset.single_task="Fold the T-shirt properly" \
     --dataset.fps=30 \
-    --dataset.episode_time_s=1000 \
-    --dataset.num_episodes=50 \
+    --strategy.num_episodes=50 \
     --interpolation_multiplier=3
 ```
 
@@ -235,7 +233,7 @@ This HIL data collection approach builds on ideas from interactive imitation lea
 
 - **HG-DAgger** (Kelly et al., 2019) made this practical for robotics: a human expert monitors the robot and only intervenes when needed, rather than labeling every state. The gating between autonomous and human control is exactly the pause → takeover → return-to-policy loop used in the scripts here.
 
-- **RaC** (Hu et al., 2025) scales this loop to long-horizon tasks by explicitly decomposing interventions into **recovery** (teleoperating back to a good state) and **correction** (demonstrating the right behavior from there). This decomposition is the protocol followed by the HIL scripts in `examples/hil`.
+- **RaC** (Hu et al., 2025) scales this loop to long-horizon tasks by explicitly decomposing interventions into **recovery** (teleoperating back to a good state) and **correction** (demonstrating the right behavior from there). This decomposition is the protocol followed by the DAgger strategy in `lerobot-rollout`.
 
 - **π0.6/RECAP** (Physical Intelligence, 2025) applies the same iterative collect-and-finetune loop at scale with VLA models, showing that even large pretrained policies benefit substantially from targeted human corrections on their own failure modes. π0.6 is trained using RECAP.
 

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` 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` (defined in `lerobot/envs/configs.py`) and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
 
 <!-- prettier-ignore-start -->
 ```python
@@ -95,6 +95,7 @@ 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:
@@ -326,14 +327,22 @@ 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 the configuration of your teleoperation device (TeleoperatorConfig) under the `end_effector_bounds` field
+3. Use these values in your environment configuration under `env.processor.inverse_kinematics.end_effector_bounds` (see `InverseKinematicsConfig` in `lerobot/envs/configs.py`)
 
 **Example Configuration**
 
 ```json
-"end_effector_bounds": {
-    "max": [0.24, 0.20, 0.10],
-    "min": [0.16, -0.08, 0.03]
+{
+  "env": {
+    "processor": {
+      "inverse_kinematics": {
+        "end_effector_bounds": {
+          "max": [0.24, 0.2, 0.1],
+          "min": [0.16, -0.08, 0.03]
+        }
+      }
+    }
+  }
 }
 ```
 
@@ -404,30 +413,24 @@ 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.
 
-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.
+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:
 
 <!-- prettier-ignore-start -->
 ```python
-class SO100FollowerEndEffectorConfig(SO100FollowerConfig):
-    """Configuration for the SO100FollowerEndEffector robot."""
+class InverseKinematicsConfig:
+    """Configuration for inverse kinematics processing."""
 
-    # 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
-        }
-    )
+    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
 
-    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,
-        }
-    )
+class HILSerlProcessorConfig:
+    ...
+    # maximum gripper position that the gripper will be open at
+    max_gripper_pos: float | None = 100.0
 ```
 <!-- prettier-ignore-end -->
 
@@ -606,11 +609,11 @@ 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 `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 `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.
 
 **Collecting a Dataset for the reward classifier**
 
-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.
+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.
 
 To collect a dataset, you need to modify some parameters in the environment configuration based on HILSerlRobotEnvConfig.
 
@@ -658,7 +661,7 @@ Example configuration section for data collection:
   },
   "dataset": {
     "repo_id": "hf_username/dataset_name",
-    "dataset_root": "data/your_dataset",
+    "root": "data/your_dataset",
     "task": "reward_classifier_task",
     "num_episodes_to_record": 20,
     "replay_episode": null,
@@ -671,7 +674,7 @@ Example configuration section for data collection:
 
 **Reward Classifier Configuration**
 
-The reward classifier is configured using `configuration_classifier.py`. Here are the key parameters:
+The reward classifier is configured using `lerobot/rewards/classifier/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"`
@@ -689,7 +692,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
     "repo_id": "hf_username/dataset_name",
     "root": null
   },
-  "policy": {
+  "reward_model": {
     "type": "reward_classifier",
     "model_name": "helper2424/resnet10",
     "model_type": "cnn",
@@ -699,7 +702,6 @@ 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",
@@ -818,13 +820,14 @@ 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/configs/train.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/rl/train_rl.py`.
 
-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.
+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.
 
 **Starting the Learner**
 
@@ -926,7 +929,7 @@ 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`** (`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.
+- **`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.
 - **`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.
 

docs/source/hope_jr.mdx

@@ -232,7 +232,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.camera_encoder.vcodec=auto \
     --display_data=true
 ```
 
@@ -278,6 +278,6 @@ lerobot-record \
   --dataset.num_episodes=10 \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
+  # --dataset.camera_encoder.vcodec=auto \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```

docs/source/il_robots.mdx

@@ -32,6 +32,12 @@ Once you’ve gathered enough trajectories, you’ll train a neural network to i
 
 If you run into any issues at any point, jump into our [Discord community](https://discord.com/invite/s3KuuzsPFb) for support.
 
+<Tip>
+
+Want to quickly get the right commands for your setup? The [quickstart notebook](https://github.com/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb) [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb) lets you configure your robot once and generates all the commands below ready to paste.
+
+</Tip>
+
 ## Set up and Calibrate
 
 If you haven't yet set up and calibrated your robot and teleop device, please do so by following the robot-specific tutorial.
@@ -187,7 +193,7 @@ lerobot-record \
     --dataset.num_episodes=5 \
     --dataset.single_task="Grab the black cube" \
     --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
+    # --dataset.camera_encoder.vcodec=auto \
     --dataset.encoder_threads=2
 ```
 </hfoption>
@@ -503,121 +509,42 @@ hf upload ${HF_USER}/act_so101_test${CKPT} \
 
 ## Run inference and evaluate your policy
 
-You can use the `record` script from [`lerobot-record`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_record.py) with a policy checkpoint as input, to run inference and evaluate your policy. For instance, run this command or API example to run inference and record 10 evaluation episodes:
+Use `lerobot-rollout` to deploy a trained policy on your robot. You can choose different strategies depending on your needs:
 
 <hfoptions id="eval">
-<hfoption id="Command">
+<hfoption id="Base mode (no recording)">
 ```bash
-lerobot-record  \
+lerobot-rollout \
+  --strategy.type=base \
+  --policy.path=${HF_USER}/my_policy \
   --robot.type=so100_follower \
   --robot.port=/dev/ttyACM1 \
   --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
-  --robot.id=my_awesome_follower_arm \
-  --display_data=false \
-  --dataset.repo_id=${HF_USER}/eval_so100 \
-  --dataset.single_task="Put lego brick into the transparent box" \
-  --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_awesome_leader_arm \
-  --policy.path=${HF_USER}/my_policy
+  --task="Put lego brick into the transparent box" \
+  --duration=60
 ```
 </hfoption>
-<hfoption id="API example">
-
-<!-- prettier-ignore-start -->
-```python
-from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.datasets import LeRobotDataset
-from lerobot.utils.feature_utils import hw_to_dataset_features
-from lerobot.policies.act import ACTPolicy
-from lerobot.policies import make_pre_post_processors
-from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
-from lerobot.scripts.lerobot_record import record_loop
-from lerobot.common.control_utils import init_keyboard_listener
-from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
-
-
-NUM_EPISODES = 5
-FPS = 30
-EPISODE_TIME_SEC = 60
-TASK_DESCRIPTION = "My task description"
-HF_MODEL_ID = "<hf_username>/<model_repo_id>"
-HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
-
-# Create the robot configuration
-camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
-robot_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem58760434471", id="my_awesome_follower_arm", cameras=camera_config
-)
-
-# Initialize the robot
-robot = SO100Follower(robot_config)
-
-# Initialize the policy
-policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
-
-# 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_DATASET_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
-robot.connect()
-
-preprocessor, postprocessor = make_pre_post_processors(
-    policy_cfg=policy,
-    pretrained_path=HF_MODEL_ID,
-    dataset_stats=dataset.meta.stats,
-)
-
-for episode_idx in range(NUM_EPISODES):
-    log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
-
-    # Run the policy inference loop
-    record_loop(
-        robot=robot,
-        events=events,
-        fps=FPS,
-        policy=policy,
-        preprocessor=preprocessor,
-        postprocessor=postprocessor,
-        dataset=dataset,
-        control_time_s=EPISODE_TIME_SEC,
-        single_task=TASK_DESCRIPTION,
-        display_data=True,
-    )
-
-    dataset.save_episode()
-
-# Clean up
-robot.disconnect()
-dataset.push_to_hub()
+<hfoption id="Sentry mode (with recording)">
+```bash
+lerobot-rollout \
+  --strategy.type=sentry \
+  --strategy.upload_every_n_episodes=5 \
+  --policy.path=${HF_USER}/my_policy \
+  --robot.type=so100_follower \
+  --robot.port=/dev/ttyACM1 \
+  --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
+  --dataset.repo_id=${HF_USER}/eval_so100 \
+  --dataset.single_task="Put lego brick into the transparent box" \
+  --duration=600
 ```
-<!-- prettier-ignore-end -->
-
 </hfoption>
 </hfoptions>
 
-As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+The `--strategy.type` flag selects the execution mode:
 
-1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with (e.g. `outputs/train/eval_act_so101_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so101_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so101_test`).
+- `base`: Autonomous rollout with no data recording (useful for quick evaluation)
+- `sentry`: Continuous recording with auto-upload (useful for large-scale evaluation)
+- `highlight`: Ring buffer recording with keystroke save (useful for capturing interesting events)
+- `dagger`: Human-in-the-loop data collection (see [HIL Data Collection](./hil_data_collection))
+
+All strategies support `--inference.type=rtc` for smooth execution with slow VLA models (Pi0, Pi0.5, SmolVLA).

docs/source/inference.mdx

@@ -0,0 +1,261 @@
+# Policy Deployment (lerobot-rollout)
+
+`lerobot-rollout` is the single CLI for deploying trained policies on real robots. It supports multiple execution strategies and inference backends, from quick evaluation to continuous recording and human-in-the-loop data collection.
+
+## Quick Start
+
+No extra dependencies are needed beyond your robot and policy extras.
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --policy.path=lerobot/act_koch_real \
+    --robot.type=koch_follower \
+    --robot.port=/dev/ttyACM0 \
+    --task="pick up cube" \
+    --duration=30
+```
+
+This runs the policy for 30 seconds with no recording.
+
+---
+
+## Strategies
+
+Select a strategy with `--strategy.type=<name>`. Each strategy defines a different control loop with its own recording and interaction semantics.
+
+### Base (`--strategy.type=base`)
+
+Autonomous policy execution with no data recording. Use this for quick evaluation, demos, or when you only need to observe the robot.
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --task="Put lego brick into the box" \
+    --duration=60
+```
+
+| Flag             | Description                                            |
+| ---------------- | ------------------------------------------------------ |
+| `--duration`     | Run time in seconds (0 = infinite)                     |
+| `--task`         | Task description passed to the policy                  |
+| `--display_data` | Stream observations/actions to Rerun for visualization |
+
+### Sentry (`--strategy.type=sentry`)
+
+Continuous autonomous recording with periodic upload to the Hugging Face Hub. Episode boundaries are auto-computed from camera resolution and FPS so each saved episode produces a complete video file, keeping uploads efficient.
+
+Policy state (hidden state, RTC queue) persists across episode boundaries: the robot does not reset between episodes.
+
+```bash
+lerobot-rollout \
+    --strategy.type=sentry \
+    --strategy.upload_every_n_episodes=5 \
+    --policy.path=${HF_USER}/my_policy \
+    --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=${HF_USER}/rollout_eval_data \
+    --dataset.single_task="Put lego brick into the box" \
+    --duration=3600
+```
+
+| Flag                                   | Description                                                 |
+| -------------------------------------- | ----------------------------------------------------------- |
+| `--strategy.upload_every_n_episodes`   | Push to Hub every N episodes (default: 5)                   |
+| `--strategy.target_video_file_size_mb` | Target video file size for episode rotation (default: auto) |
+| `--dataset.repo_id`                    | **Required.** Hub repository for the recorded dataset       |
+| `--dataset.push_to_hub`                | Whether to push to Hub on teardown (default: true)          |
+
+### Highlight (`--strategy.type=highlight`)
+
+Autonomous rollout with on-demand recording via a memory-bounded ring buffer. The robot runs continuously while the buffer captures the last N seconds of telemetry. Press the save key to flush the buffer and start live recording; press it again to save the episode.
+
+```bash
+lerobot-rollout \
+    --strategy.type=highlight \
+    --strategy.ring_buffer_seconds=30 \
+    --strategy.save_key=s \
+    --strategy.push_key=h \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=koch_follower \
+    --robot.port=/dev/ttyACM0 \
+    --dataset.repo_id=${HF_USER}/rollout_highlight_data \
+    --dataset.single_task="Pick up the red cube"
+```
+
+**Keyboard controls:**
+
+| Key                | Action                                                   |
+| ------------------ | -------------------------------------------------------- |
+| `s` (configurable) | Start recording (flushes buffer) / stop and save episode |
+| `h` (configurable) | Push dataset to Hub                                      |
+| `ESC`              | Stop the session                                         |
+
+| Flag                                   | Description                                    |
+| -------------------------------------- | ---------------------------------------------- |
+| `--strategy.ring_buffer_seconds`       | Duration of buffered telemetry (default: 30)   |
+| `--strategy.ring_buffer_max_memory_mb` | Memory cap for the ring buffer (default: 2048) |
+| `--strategy.save_key`                  | Key to toggle recording (default: `s`)         |
+| `--strategy.push_key`                  | Key to push to Hub (default: `h`)              |
+
+### DAgger (`--strategy.type=dagger`)
+
+Human-in-the-loop data collection. Alternates between autonomous policy execution and human intervention via a teleoperator. Intervention frames are tagged with `intervention=True`. Requires a teleoperator (`--teleop.type`).
+
+See the [Human-In-the-Loop Data Collection](./hil_data_collection) guide for a detailed walkthrough.
+
+**Corrections-only mode** (default): Only human correction windows are recorded. Each correction becomes one episode.
+
+```bash
+lerobot-rollout \
+    --strategy.type=dagger \
+    --strategy.num_episodes=20 \
+    --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --robot.type=bi_openarm_follower \
+    --teleop.type=openarm_mini \
+    --dataset.repo_id=${HF_USER}/rollout_hil_data \
+    --dataset.single_task="Fold the T-shirt"
+```
+
+**Continuous recording mode** (`--strategy.record_autonomous=true`): Both autonomous and correction frames are recorded with time-based episode rotation (same as Sentry).
+
+```bash
+lerobot-rollout \
+    --strategy.type=dagger \
+    --strategy.record_autonomous=true \
+    --strategy.num_episodes=50 \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/ttyACM1 \
+    --dataset.repo_id=${HF_USER}/rollout_dagger_data \
+    --dataset.single_task="Grasp the block"
+```
+
+**Keyboard controls** (default input device):
+
+| Key     | Action                                      |
+| ------- | ------------------------------------------- |
+| `Space` | Pause / resume policy execution             |
+| `Tab`   | Start / stop human correction               |
+| `Enter` | Push dataset to Hub (corrections-only mode) |
+| `ESC`   | Stop the session                            |
+
+Foot pedal input is also supported via `--strategy.input_device=pedal`. Configure pedal codes with `--strategy.pedal.*` flags.
+
+| Flag                                 | Description                                             |
+| ------------------------------------ | ------------------------------------------------------- |
+| `--strategy.num_episodes`            | Number of correction episodes to record (default: 10)   |
+| `--strategy.record_autonomous`       | Record autonomous frames too (default: false)           |
+| `--strategy.upload_every_n_episodes` | Push to Hub every N episodes (default: 5)               |
+| `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard) |
+| `--teleop.type`                      | **Required.** Teleoperator type                         |
+
+---
+
+## Inference Backends
+
+Select a backend with `--inference.type=<name>`. All strategies work with both backends.
+
+### Sync (default)
+
+One policy call per control tick. The main loop blocks until the action is computed.
+
+Works with all policies. No extra flags needed.
+
+### Real-Time Chunking (`--inference.type=rtc`)
+
+A background thread produces action chunks asynchronously. The main control loop polls for the next ready action while the policy computes the next chunk in parallel.
+
+Use RTC with large, slow VLA models (Pi0, Pi0.5, SmolVLA) for smooth, continuous motion despite high inference latency.
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --inference.type=rtc \
+    --inference.rtc.execution_horizon=10 \
+    --inference.rtc.max_guidance_weight=10.0 \
+    --policy.path=${HF_USER}/pi0_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --task="Pick up the cube" \
+    --duration=60 \
+    --device=cuda
+```
+
+| Flag                                        | Description                                                    |
+| ------------------------------------------- | -------------------------------------------------------------- |
+| `--inference.rtc.execution_horizon`         | Steps to blend with previous chunk (default: varies by policy) |
+| `--inference.rtc.max_guidance_weight`       | Consistency enforcement strength (default: varies by policy)   |
+| `--inference.rtc.prefix_attention_schedule` | Blend schedule: `LINEAR`, `EXP`, `ONES`, `ZEROS`               |
+| `--inference.queue_threshold`               | Max queue size before backpressure (default: 30)               |
+
+See the [Real-Time Chunking](./rtc) guide for details on tuning RTC parameters.
+
+---
+
+## Common Flags
+
+| Flag                              | Description                                                       | Default |
+| --------------------------------- | ----------------------------------------------------------------- | ------- |
+| `--policy.path`                   | **Required.** HF Hub model ID or local checkpoint path            | --      |
+| `--robot.type`                    | **Required.** Robot type (e.g. `so100_follower`, `koch_follower`) | --      |
+| `--robot.port`                    | Serial port for the robot                                         | --      |
+| `--robot.cameras`                 | Camera configuration (JSON dict)                                  | --      |
+| `--fps`                           | Control loop frequency                                            | 30      |
+| `--duration`                      | Run time in seconds (0 = infinite)                                | 0       |
+| `--device`                        | Torch device (`cpu`, `cuda`, `mps`)                               | auto    |
+| `--task`                          | Task description (used when no dataset is provided)               | --      |
+| `--display_data`                  | Stream telemetry to Rerun visualization                           | false   |
+| `--display_ip` / `--display_port` | Remote Rerun server address                                       | --      |
+| `--interpolation_multiplier`      | Action interpolation factor                                       | 1       |
+| `--use_torch_compile`             | Enable `torch.compile` for inference                              | false   |
+| `--resume`                        | Resume a previous recording session                               | false   |
+| `--play_sounds`                   | Vocal synthesis for events                                        | true    |
+
+---
+
+## Programmatic Usage
+
+For custom deployments (e.g. with kinematics processors), use the rollout module API directly:
+
+```python
+from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
+from lerobot.rollout.inference import SyncInferenceConfig
+from lerobot.rollout.strategies import BaseStrategy
+from lerobot.utils.process import ProcessSignalHandler
+
+cfg = RolloutConfig(
+    robot=my_robot_config,
+    policy=my_policy_config,
+    strategy=BaseStrategyConfig(),
+    inference=SyncInferenceConfig(),
+    fps=30,
+    duration=60,
+    task="my task",
+)
+
+signal_handler = ProcessSignalHandler(use_threads=True)
+ctx = build_rollout_context(
+    cfg,
+    signal_handler.shutdown_event,
+    robot_action_processor=my_custom_action_processor,       # optional
+    robot_observation_processor=my_custom_obs_processor,     # optional
+)
+
+strategy = BaseStrategy(cfg.strategy)
+try:
+    strategy.setup(ctx)
+    strategy.run(ctx)
+finally:
+    strategy.teardown(ctx)
+```
+
+See `examples/so100_to_so100_EE/rollout.py` and `examples/phone_to_so100/rollout.py` for full examples with kinematics processors.

docs/source/installation.mdx

@@ -207,6 +207,56 @@ 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`.

docs/source/lerobot-dataset-v3.mdx

@@ -10,6 +10,7 @@ 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`
 
@@ -43,7 +44,7 @@ lerobot-record \
   --dataset.num_episodes=5 \
   --dataset.single_task="Grab the black cube" \
   --dataset.streaming_encoding=true \
-  # --dataset.vcodec=auto \
+  # --dataset.camera_encoder.vcodec=auto \
   --dataset.encoder_threads=2
 ```
 
@@ -315,3 +316,39 @@ 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/).

docs/source/libero_plus.mdx

@@ -0,0 +1,188 @@
+# 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`.

docs/source/peft_training.mdx

@@ -28,13 +28,15 @@ lerobot-train \
  --steps=100000 \
  --batch_size=32 \
  --peft.method_type=LORA \
- --peft.r=64
+ --peft.r=64 \
+ --peft.lora_alpha=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. The higher the 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
 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

docs/source/reachy2.mdx

@@ -161,7 +161,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.camera_encoder.vcodec=auto \
     --display_data=true
 ```
 
@@ -203,7 +203,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.camera_encoder.vcodec=auto \
     --display_data=true
 ```
 

docs/source/rename_map.mdx

@@ -61,17 +61,6 @@ 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.
@@ -105,10 +94,10 @@ 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", ...}'`                             |
-| 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          |
+| 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          |

docs/source/robocasa.mdx

@@ -0,0 +1,188 @@
+# 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.

docs/source/robocerebra.mdx

@@ -0,0 +1,99 @@
+# 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.

docs/source/robomme.mdx

@@ -0,0 +1,130 @@
+# 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.

docs/source/robotwin.mdx

@@ -0,0 +1,223 @@
+# 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/).

docs/source/rtc.mdx

@@ -34,7 +34,7 @@ pip install -e ".[smolvla]"
 
 ### Using RTC with Pi0
 
-You can find a complete reference implementation in [eval_with_real_robot.py](examples/rtc/eval_with_real_robot.py).
+You can use `lerobot-rollout --strategy.type=base --inference.type=rtc` for RTC deployment on real robots.
 The snippet below provides a simplified pseudo-example of how RTC operates with Pi0 in your pipeline:
 
 ```python
@@ -137,8 +137,12 @@ The script generates a visualization of the denoising process, comparing standar
 ## Testing RTC with a Real Robot
 
 ```bash
-python examples/rtc/eval_with_real_robot.py \
+lerobot-rollout \
+    --strategy.type=base \
     --policy.path=${HF_USERNAME}/policy_repo_id \
+    --inference.type=rtc \
+    --inference.rtc.execution_horizon=10 \
+    --inference.rtc.max_guidance_weight=10.0 \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58FA0834591 \
     --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
@@ -178,7 +182,7 @@ visualizer = RTCDebugVisualizer()
 # ... create plots
 ```
 
-See `examples/rtc/eval_dataset.py` for a complete example of visualization.
+See `examples/rtc/eval_dataset.py` for a complete example of offline RTC visualization.
 
 ## References
 

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/policies/sarm/processor_sarm.py`), which:
+SARM is trained through its processor (`src/lerobot/rewards/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 src/lerobot/policies/sarm/compute_rabc_weights.py \
+python -m lerobot.rewards.sarm.compute_rabc_weights \
   --dataset-repo-id your-username/your-dataset \
   --reward-model-path your-username/sarm-model \
   --visualize-only \
@@ -360,7 +360,7 @@ python src/lerobot/policies/sarm/compute_rabc_weights.py \
 <hfoption id="dense_only">
 
 ```bash
-python src/lerobot/policies/sarm/compute_rabc_weights.py \
+python -m lerobot.rewards.sarm.compute_rabc_weights \
   --dataset-repo-id your-username/your-dataset \
   --reward-model-path your-username/sarm-model \
   --visualize-only \
@@ -373,7 +373,7 @@ python src/lerobot/policies/sarm/compute_rabc_weights.py \
 <hfoption id="dual">
 
 ```bash
-python src/lerobot/policies/sarm/compute_rabc_weights.py \
+python -m lerobot.rewards.sarm.compute_rabc_weights \
   --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 src/lerobot/policies/sarm/compute_rabc_weights.py \
+python -m lerobot.rewards.sarm.compute_rabc_weights \
   --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`). 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`) if not explicitly provided. Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
 
 ```bash
 lerobot-train \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
-  --use_rabc=true \
-  --rabc_head_mode=sparse \
-  --rabc_kappa=0.01 \
+  --sample_weighting.type=rabc \
+  --sample_weighting.head_mode=sparse \
+  --sample_weighting.kappa=0.01 \
   --output_dir=outputs/train/policy_rabc \
   --batch_size=32 \
   --steps=40000
@@ -488,12 +488,13 @@ The training script automatically:
 
 **RA-BC Arguments:**
 
-| 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`                             |
+| 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`                  |
 
 ### Tuning RA-BC Kappa
 
@@ -511,30 +512,30 @@ The `kappa` parameter is the threshold that determines which samples get full we
 
 Monitor these WandB metrics during training:
 
-| Metric             | Healthy Range | Problem Indicator         |
-| ------------------ | ------------- | ------------------------- |
-| `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  |
+| 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  |
 
-**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.
+**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.
 
 **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 `rabc_delta_mean` and `rabc_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 `sample_weighting/delta_mean` and `sample_weighting/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)
---rabc_kappa=0.03
+--sample_weighting.kappa=0.03
 
 # Option 2: Set kappa = delta_mean + delta_std (high selectivity)
---rabc_kappa=0.05
+--sample_weighting.kappa=0.05
 
 # Option 3: Set kappa = delta_mean + 2*delta_std (very selective)
---rabc_kappa=0.07
+--sample_weighting.kappa=0.07
 ```
 
 **When RA-BC may not help:**
@@ -550,8 +551,8 @@ accelerate launch \
   src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
-  --use_rabc=true \
-  --rabc_kappa=0.01 \
+  --sample_weighting.type=rabc \
+  --sample_weighting.kappa=0.01 \
   --output_dir=outputs/train/policy_rabc \
   --batch_size=32 \
   --steps=40000
@@ -576,7 +577,7 @@ accelerate launch \
 ### RA-BC
 
 1. **Train SARM first**: RA-BC quality depends entirely on SARM quality
-2. **Monitor `rabc_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))
+2. **Monitor `sample_weight_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))
 
 ---
 

docs/source/smolvla.mdx

@@ -108,7 +108,7 @@ lerobot-record \
   --dataset.num_episodes=10 \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
+  # --dataset.camera_encoder.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

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.vcodec`                | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
+| `vcodec`                | `--dataset.camera_encoder.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`         | `60`          | Max buffered frames per camera (~2s at 30fps). Consumes RAM       |
+| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `30`          | Max buffered frames per camera (~1s 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 60 frames). When the encoder can't keep up:
+Each camera has a bounded queue (`encoder_queue_maxsize`, default 30 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.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`              |
+| 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`              |
 
 > [!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.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.                                   |
+| 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.                                                  |
 
 ## 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.vcodec=h264 --dataset.streaming_encoding=false ...
+lerobot-record --dataset.camera_encoder.vcodec=h264 --dataset.streaming_encoding=false ...
 ```
 
 ## 7. Closing note

docs/source/unitree_g1.mdx

@@ -274,7 +274,8 @@ python src/lerobot/scripts/lerobot_train.py \
 Once trained, we recommend deploying policies using inference-time RTC:
 
 ```bash
-python examples/rtc/eval_with_real_robot.py \
+lerobot-rollout \
+  --strategy.type=base \
   --policy.path=your-username/your-repo-id \
   --policy.device=cuda \
   --robot.type=unitree_g1 \
@@ -284,7 +285,7 @@ python examples/rtc/eval_with_real_robot.py \
   --task="task_description" \
   --duration=1000 \
   --fps=30 \
-  --rtc.enabled=true
+  --inference.type=rtc
 ```
 
 ---

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.vcodec libsvtav1 \
-    --operation.pix_fmt yuv420p \
-    --operation.g 2 \
-    --operation.crf 30
+    --operation.camera_encoder.vcodec libsvtav1 \
+    --operation.camera_encoder.pix_fmt yuv420p \
+    --operation.camera_encoder.g 2 \
+    --operation.camera_encoder.crf 30
 
 # Convert only specific episodes
 lerobot-edit-dataset \
@@ -147,11 +147,7 @@ lerobot-edit-dataset \
 **Parameters:**
 
 - `output_dir`: Custom output directory (optional - by default uses `new_repo_id` or `{repo_id}_video`)
-- `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)
+- `camera_encoder`: Video encoder settings — all sub-fields accessible via `--operation.camera_encoder.<field>. See [Video Encoding Parameters](./video_encoding_parameters) for more details.
 - `episode_indices`: List of specific episodes to convert (default: all episodes)
 - `num_workers`: Number of parallel workers for processing (default: 4)
 

docs/source/video_encoding_parameters.mdx

@@ -0,0 +1,117 @@
+# 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,
+        "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`, `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.

docs/source/vlabench.mdx

@@ -0,0 +1,176 @@
+# 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.

docs/source/xvla.mdx

@@ -220,7 +220,7 @@ REAL_DIM = 12
 # Postprocessing: Trim 20D predictions to 12D for deployment
 ```
 
-See the [action_hub.py](/home/jade_choghari/robot/lerobot/src/lerobot/policies/xvla/action_hub.py) implementation for details.
+See the [action_hub.py](https://github.com/huggingface/lerobot/blob/main/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/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)
+- [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)
 
 ## Contributing
 

examples/dataset/slurm_compute_rabc.py

@@ -69,7 +69,7 @@ class ComputeProgressShards(PipelineStep):
         import torch
         from tqdm import tqdm
 
-        from lerobot.policies.sarm.compute_rabc_weights import (
+        from lerobot.rewards.sarm.compute_rabc_weights import (
             generate_all_frame_indices,
             interpolate_progress,
             load_sarm_resources,

examples/hil/hil_utils.py

@@ -1,226 +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.
-
-"""Shared utilities for Human-in-the-Loop data collection scripts."""
-
-import logging
-import time
-from dataclasses import dataclass, field
-from pathlib import Path
-
-from lerobot.common.control_utils import is_headless
-from lerobot.processor import (
-    IdentityProcessorStep,
-    RobotAction,
-    RobotObservation,
-    RobotProcessorPipeline,
-    observation_to_transition,
-    robot_action_observation_to_transition,
-    transition_to_observation,
-    transition_to_robot_action,
-)
-from lerobot.robots import Robot
-from lerobot.teleoperators import Teleoperator
-from lerobot.utils.robot_utils import precise_sleep
-
-logger = logging.getLogger(__name__)
-
-
-@dataclass
-class HILDatasetConfig:
-    repo_id: str
-    single_task: str
-    root: str | Path | None = None
-    fps: int = 30
-    episode_time_s: float = 120
-    num_episodes: int = 50
-    video: bool = True
-    push_to_hub: bool = True
-    private: bool = False
-    tags: list[str] | None = None
-    num_image_writer_processes: int = 0
-    num_image_writer_threads_per_camera: int = 4
-    video_encoding_batch_size: int = 1
-    vcodec: str = "auto"
-    streaming_encoding: bool = True
-    encoder_queue_maxsize: int = 30
-    encoder_threads: int | None = None
-    rename_map: dict[str, str] = field(default_factory=dict)
-
-
-def teleop_has_motor_control(teleop: Teleoperator) -> bool:
-    """Check if teleoperator has motor control capabilities."""
-    return all(hasattr(teleop, attr) for attr in ("enable_torque", "disable_torque", "write_goal_positions"))
-
-
-def teleop_disable_torque(teleop: Teleoperator) -> None:
-    """Disable teleop torque if supported."""
-    if hasattr(teleop, "disable_torque"):
-        teleop.disable_torque()
-
-
-def teleop_enable_torque(teleop: Teleoperator) -> None:
-    """Enable teleop torque if supported."""
-    if hasattr(teleop, "enable_torque"):
-        teleop.enable_torque()
-
-
-def teleop_smooth_move_to(teleop: Teleoperator, target_pos: dict, duration_s: float = 2.0, fps: int = 50):
-    """Smoothly move teleop to target position if motor control is available."""
-    if not teleop_has_motor_control(teleop):
-        logger.warning("Teleop does not support motor control - cannot mirror robot position")
-        return
-
-    teleop_enable_torque(teleop)
-    current = teleop.get_action()
-    steps = max(int(duration_s * fps), 1)
-
-    for step in range(steps + 1):
-        t = step / steps
-        interp = {}
-        for k in current:
-            if k in target_pos:
-                interp[k] = current[k] * (1 - t) + target_pos[k] * t
-            else:
-                interp[k] = current[k]
-        teleop.write_goal_positions(interp)
-        time.sleep(1 / fps)
-
-
-def init_keyboard_listener():
-    """Initialize keyboard listener with HIL controls."""
-    events = {
-        "exit_early": False,
-        "rerecord_episode": False,
-        "stop_recording": False,
-        "policy_paused": False,
-        "correction_active": False,
-        "resume_policy": False,
-        "in_reset": False,
-        "start_next_episode": False,
-    }
-
-    if is_headless():
-        logger.warning("Headless environment - keyboard controls unavailable")
-        return None, events
-
-    from pynput import keyboard
-
-    def on_press(key):
-        try:
-            if events["in_reset"]:
-                if key in [keyboard.Key.space, keyboard.Key.right]:
-                    logger.info("[HIL] Starting next episode...")
-                    events["start_next_episode"] = True
-                elif hasattr(key, "char") and key.char == "c":
-                    events["start_next_episode"] = True
-                elif key == keyboard.Key.esc:
-                    logger.info("[HIL] ESC - Stop recording, pushing to hub...")
-                    events["stop_recording"] = True
-                    events["start_next_episode"] = True
-            else:
-                if key == keyboard.Key.space:
-                    if not events["policy_paused"] and not events["correction_active"]:
-                        logger.info("[HIL] PAUSED - Press 'c' to take control or 'p' to resume policy")
-                        events["policy_paused"] = True
-                elif hasattr(key, "char") and key.char == "c":
-                    if events["policy_paused"] and not events["correction_active"]:
-                        logger.info("[HIL] Taking control...")
-                        events["start_next_episode"] = True
-                elif hasattr(key, "char") and key.char == "p":
-                    if events["policy_paused"] or events["correction_active"]:
-                        logger.info("[HIL] Resuming policy...")
-                        events["resume_policy"] = True
-                elif key == keyboard.Key.right:
-                    logger.info("[HIL] End episode")
-                    events["exit_early"] = True
-                elif key == keyboard.Key.left:
-                    logger.info("[HIL] Re-record episode")
-                    events["rerecord_episode"] = True
-                    events["exit_early"] = True
-                elif key == keyboard.Key.esc:
-                    logger.info("[HIL] ESC - Stop recording...")
-                    events["stop_recording"] = True
-                    events["exit_early"] = True
-        except Exception as e:
-            logger.info(f"Key error: {e}")
-
-    listener = keyboard.Listener(on_press=on_press)
-    listener.start()
-    return listener, events
-
-
-def make_identity_processors():
-    """Create identity processors for recording."""
-    teleop_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-        steps=[IdentityProcessorStep()],
-        to_transition=robot_action_observation_to_transition,
-        to_output=transition_to_robot_action,
-    )
-    obs_proc = RobotProcessorPipeline[RobotObservation, RobotObservation](
-        steps=[IdentityProcessorStep()],
-        to_transition=observation_to_transition,
-        to_output=transition_to_observation,
-    )
-    return teleop_proc, obs_proc
-
-
-def reset_loop(robot: Robot, teleop: Teleoperator, events: dict, fps: int):
-    """Reset period where human repositions environment."""
-    logger.info("[HIL] RESET")
-
-    events["in_reset"] = True
-    events["start_next_episode"] = False
-
-    obs = robot.get_observation()
-    robot_pos = {k: v for k, v in obs.items() if k.endswith(".pos") and k in robot.observation_features}
-    teleop_smooth_move_to(teleop, robot_pos, duration_s=2.0, fps=50)
-
-    logger.info("Press any key to enable teleoperation")
-    while not events["start_next_episode"] and not events["stop_recording"]:
-        precise_sleep(0.05)
-
-    if events["stop_recording"]:
-        return
-
-    events["start_next_episode"] = False
-    teleop_disable_torque(teleop)
-    logger.info("Teleop enabled - press any key to start episode")
-
-    while not events["start_next_episode"] and not events["stop_recording"]:
-        loop_start = time.perf_counter()
-        action = teleop.get_action()
-        robot.send_action(action)
-        precise_sleep(1 / fps - (time.perf_counter() - loop_start))
-
-    events["in_reset"] = False
-    events["start_next_episode"] = False
-    events["exit_early"] = False
-    events["policy_paused"] = False
-    events["correction_active"] = False
-    events["resume_policy"] = False
-
-
-def print_controls(rtc: bool = False):
-    """Print control instructions."""
-    mode = "Human-in-the-Loop Data Collection" + (" (RTC)" if rtc else "")
-    logger.info(
-        "%s\n  Controls:\n"
-        "    SPACE  - Pause policy\n"
-        "    c      - Take control\n"
-        "    p      - Resume policy after pause/correction\n"
-        "    →      - End episode\n"
-        "    ESC    - Stop and push to hub",
-        mode,
-    )

examples/lekiwi/evaluate.py

@@ -14,17 +14,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.common.control_utils import init_keyboard_listener
+import logging
+import time
+
+from lerobot.common.control_utils import init_keyboard_listener, predict_action
 from lerobot.datasets import LeRobotDataset
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
+from lerobot.policies.utils import make_robot_action
 from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.constants import ACTION, OBS_STR
-from lerobot.utils.feature_utils import hw_to_dataset_features
+from lerobot.utils.feature_utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 NUM_EPISODES = 2
 FPS = 30
@@ -35,6 +39,9 @@ HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
 
 
 def main():
+    # NOTE: For production policy deployment, use `lerobot-rollout` CLI instead.
+    # This script provides a self-contained example for educational purposes.
+
     # Create the robot configuration & robot
     robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
 
@@ -83,43 +90,67 @@ def main():
             raise ValueError("Robot is not connected!")
 
         print("Starting evaluate loop...")
+        control_interval = 1 / FPS
         recorded_episodes = 0
         while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
             log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
 
-            # Main record loop
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                policy=policy,
-                preprocessor=preprocessor,  # Pass the pre and post policy processors
-                postprocessor=postprocessor,
-                dataset=dataset,
-                control_time_s=EPISODE_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=teleop_action_processor,
-                robot_action_processor=robot_action_processor,
-                robot_observation_processor=robot_observation_processor,
-            )
+            # Inline evaluation loop: predict actions and send to robot
+            timestamp = 0
+            start_episode_t = time.perf_counter()
+            while timestamp < EPISODE_TIME_SEC:
+                start_loop_t = time.perf_counter()
+
+                if events["exit_early"]:
+                    events["exit_early"] = False
+                    break
+
+                # Get robot observation
+                obs = robot.get_observation()
+                obs_processed = robot_observation_processor(obs)
+                observation_frame = build_dataset_frame(dataset.features, obs_processed, prefix=OBS_STR)
+
+                # Predict action using the policy
+                action_tensor = predict_action(
+                    observation=observation_frame,
+                    policy=policy,
+                    device=policy.config.device,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    use_amp=policy.config.device.type == "cuda",
+                    task=TASK_DESCRIPTION,
+                    robot_type=robot.name,
+                )
+
+                # Convert policy output to robot action dict
+                action_values = make_robot_action(action_tensor, dataset.features)
+
+                # Process and send action to robot
+                robot_action_to_send = robot_action_processor((action_values, obs))
+                robot.send_action(robot_action_to_send)
+
+                # Write to dataset
+                action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
+                frame = {**observation_frame, **action_frame, "task": TASK_DESCRIPTION}
+                dataset.add_frame(frame)
+
+                log_rerun_data(observation=obs_processed, action=action_values)
+
+                dt_s = time.perf_counter() - start_loop_t
+                sleep_time_s = control_interval - dt_s
+                if sleep_time_s < 0:
+                    logging.warning(
+                        f"Evaluate loop is running slower ({1 / dt_s:.1f} Hz) than the target FPS ({FPS} Hz)."
+                    )
+                precise_sleep(max(sleep_time_s, 0.0))
+                timestamp = time.perf_counter() - start_episode_t
 
             # Reset the environment if not stopping or re-recording
             if not events["stop_recording"] and (
                 (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
             ):
                 log_say("Reset the environment")
-                record_loop(
-                    robot=robot,
-                    events=events,
-                    fps=FPS,
-                    control_time_s=EPISODE_TIME_SEC,
-                    single_task=TASK_DESCRIPTION,
-                    display_data=True,
-                    teleop_action_processor=teleop_action_processor,
-                    robot_action_processor=robot_action_processor,
-                    robot_observation_processor=robot_observation_processor,
-                )
+                log_say("Waiting for environment reset, press right arrow key when ready...")
 
             if events["rerecord_episode"]:
                 log_say("Re-record episode")

examples/lekiwi/record.py

@@ -45,9 +45,6 @@ def main():
     leader_arm = SO100Leader(leader_arm_config)
     keyboard = KeyboardTeleop(keyboard_config)
 
-    # TODO(Steven): Update this example to use pipelines
-    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
-
     # Configure the dataset features
     action_features = hw_to_dataset_features(robot.action_features, ACTION)
     obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
@@ -77,6 +74,10 @@ def main():
         if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
             raise ValueError("Robot or teleop is not connected!")
 
+        teleop_action_processor, robot_action_processor, robot_observation_processor = (
+            make_default_processors()
+        )
+
         print("Starting record loop...")
         recorded_episodes = 0
         while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
@@ -87,14 +88,14 @@ def main():
                 robot=robot,
                 events=events,
                 fps=FPS,
+                teleop_action_processor=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
                 dataset=dataset,
                 teleop=[leader_arm, keyboard],
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
-                teleop_action_processor=teleop_action_processor,
-                robot_action_processor=robot_action_processor,
-                robot_observation_processor=robot_observation_processor,
             )
 
             # Reset the environment if not stopping or re-recording
@@ -106,13 +107,13 @@ def main():
                     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=[leader_arm, keyboard],
                     control_time_s=RESET_TIME_SEC,
                     single_task=TASK_DESCRIPTION,
                     display_data=True,
-                    teleop_action_processor=teleop_action_processor,
-                    robot_action_processor=robot_action_processor,
-                    robot_observation_processor=robot_observation_processor,
                 )
 
             if events["rerecord_episode"]:

examples/lekiwi/rollout.py

@@ -0,0 +1,77 @@
+# !/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.
+
+"""Run a trained policy on LeKiwi without recording (base rollout).
+
+Uses the rollout engine's :class:`BaseStrategy` (autonomous execution,
+no dataset) with :class:`SyncInferenceConfig` (inline policy call per
+control tick).  For a CLI entry point with the same capabilities plus
+recording, upload, and human-in-the-loop variants, see ``lerobot-rollout``.
+"""
+
+from lerobot.configs import PreTrainedConfig
+from lerobot.robots.lekiwi import LeKiwiClientConfig
+from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
+from lerobot.rollout.inference import SyncInferenceConfig
+from lerobot.rollout.strategies import BaseStrategy
+from lerobot.utils.process import ProcessSignalHandler
+from lerobot.utils.utils import init_logging
+
+FPS = 30
+DURATION_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+
+
+def main():
+    init_logging()
+
+    # Robot: LeKiwi client — make sure lekiwi_host is already running on the robot.
+    robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
+
+    # Policy: load the pretrained config.  ``pretrained_path`` is read downstream
+    # by ``build_rollout_context`` to reload the full model.
+    policy_config = PreTrainedConfig.from_pretrained(HF_MODEL_ID)
+    policy_config.pretrained_path = HF_MODEL_ID
+
+    # Assemble the rollout config: base strategy (no recording) + sync inference.
+    cfg = RolloutConfig(
+        robot=robot_config,
+        policy=policy_config,
+        strategy=BaseStrategyConfig(),
+        inference=SyncInferenceConfig(),
+        fps=FPS,
+        duration=DURATION_SEC,
+        task=TASK_DESCRIPTION,
+    )
+
+    # Graceful Ctrl-C: the strategy loop exits when shutdown_event is set.
+    signal_handler = ProcessSignalHandler(use_threads=True)
+
+    # Build the context (connects robot, loads policy, wires the inference strategy).
+    # No custom processors here — LeKiwi runs on raw joint features.
+    ctx = build_rollout_context(cfg, signal_handler.shutdown_event)
+
+    strategy = BaseStrategy(cfg.strategy)
+    try:
+        strategy.setup(ctx)
+        strategy.run(ctx)
+    finally:
+        strategy.teardown(ctx)
+
+
+if __name__ == "__main__":
+    main()

examples/notebooks/quickstart.ipynb

@@ -0,0 +1,369 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# 🤗 LeRobot Quickstart\n",
+    "\n",
+    "Calibration → teleoperation → data collection → training → evaluation.\n",
+    "\n",
+    "Install the required dependencies: `pip install -e .[notebook,dataset,training,viz,hardware]`.\n",
+    "\n",
+    "**How to use:**\n",
+    "1. Edit the **Configuration** cell with your settings.\n",
+    "2. Run all cells (`Run All`).\n",
+    "3. Each section prints a ready-to-paste terminal command - copy it and run it.\n",
+    "\n",
+    "Each setup is different, please refer to the [LeRobot documentation](https://huggingface.co/docs/lerobot/il_robots) for more details on each step and available options. <br>\n",
+    "Feel free to make this notebook your own and adapt it to your needs!"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## Utils"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def _cameras_arg(cameras: dict) -> str:\n",
+    "    if not cameras:\n",
+    "        return \"\"\n",
+    "    entries = [f\"{n}: {{{', '.join(f'{k}: {v}' for k, v in cfg.items())}}}\" for n, cfg in cameras.items()]\n",
+    "    return \"{ \" + \", \".join(entries) + \" }\"\n",
+    "\n",
+    "\n",
+    "def print_cmd(*parts: str) -> None:\n",
+    "    \"\"\"Print a shell command with line continuations, skipping empty parts.\"\"\"\n",
+    "    non_empty = [p for p in parts if p]\n",
+    "    print(\" \\\\\\n    \".join(non_empty))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## Configuration\n",
+    "\n",
+    "Edit this cell, then **Run All** to generate all commands below."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Robot (follower) - run `lerobot-find-port` to discover the port\n",
+    "ROBOT_TYPE = \"so101_follower\"\n",
+    "ROBOT_PORT = \"/dev/ttyACM0\"\n",
+    "ROBOT_ID = \"my_follower_arm\"\n",
+    "\n",
+    "# Teleop (leader) - run `lerobot-find-port` to discover the port\n",
+    "TELEOP_TYPE = \"so101_leader\"\n",
+    "TELEOP_PORT = \"/dev/ttyACM1\"\n",
+    "TELEOP_ID = \"my_leader_arm\"\n",
+    "\n",
+    "# Cameras - set to {} to disable\n",
+    "# Run `lerobot-find-cameras opencv` to list available cameras and their indices\n",
+    "CAMERAS = {\n",
+    "    \"top\": {\"type\": \"opencv\", \"index_or_path\": 2, \"width\": 640, \"height\": 480, \"fps\": 30},\n",
+    "    \"wrist\": {\"type\": \"opencv\", \"index_or_path\": 4, \"width\": 640, \"height\": 480, \"fps\": 30},\n",
+    "}\n",
+    "\n",
+    "# Dataset\n",
+    "HF_USER = \"your_hf_username\"  # `hf auth whoami` to find your username\n",
+    "DATASET_NAME = \"my_so101_dataset\"\n",
+    "TASK_DESCRIPTION = \"pick and place the block\"\n",
+    "NUM_EPISODES = 10\n",
+    "\n",
+    "# Training\n",
+    "POLICY_TYPE = \"act\"  # act, diffusion, smolvla, ...\n",
+    "POLICY_DEVICE = \"cuda\"  # cuda / cpu / mps\n",
+    "TRAIN_STEPS = 10_000\n",
+    "SAVE_FREQ = 2_000\n",
+    "OUTPUT_DIR = f\"outputs/train/{DATASET_NAME}\"\n",
+    "\n",
+    "# Inference - Hub repo ID or local checkpoint path\n",
+    "# e.g. set to f\"{OUTPUT_DIR}/checkpoints/last\" to use a local checkpoint\n",
+    "POLICY_PATH = f\"{HF_USER}/{DATASET_NAME}_{POLICY_TYPE}\"\n",
+    "LAST_CHECKPOINT_PATH = f\"{OUTPUT_DIR}/checkpoints/last\"\n",
+    "\n",
+    "# Derived\n",
+    "DATASET_REPO_ID = f\"{HF_USER}/{DATASET_NAME}\"\n",
+    "DATASET_ROOT = f\"data/{DATASET_NAME}\"\n",
+    "POLICY_REPO_ID = f\"{HF_USER}/{DATASET_NAME}_{POLICY_TYPE}\"\n",
+    "EVAL_REPO_ID = f\"{HF_USER}/eval_{DATASET_NAME}\"\n",
+    "CAMERAS_ARG = _cameras_arg(CAMERAS)\n",
+    "CAMERAS_FLAG = f'--robot.cameras=\"{CAMERAS_ARG}\"' if CAMERAS_ARG else \"\"\n",
+    "\n",
+    "print(f\"Robot  : {ROBOT_TYPE} @ {ROBOT_PORT}\")\n",
+    "print(f\"Teleop : {TELEOP_TYPE} @ {TELEOP_PORT}\")\n",
+    "print(f\"Cameras: {list(CAMERAS) or 'none'}\")\n",
+    "print(f\"Dataset: {DATASET_REPO_ID} ({NUM_EPISODES} episodes) saved to {DATASET_ROOT}\")\n",
+    "print(f\"Policy : {POLICY_TYPE} -> {POLICY_REPO_ID}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## 1. Calibration\n",
+    "\n",
+    "Run once per arm before first use."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Follower\n",
+    "print_cmd(\n",
+    "    \"lerobot-calibrate\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    f\"--robot.id={ROBOT_ID}\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Leader\n",
+    "print_cmd(\n",
+    "    \"lerobot-calibrate\",\n",
+    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
+    "    f\"--teleop.port={TELEOP_PORT}\",\n",
+    "    f\"--teleop.id={TELEOP_ID}\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## 2. Teleoperation\n",
+    "\n",
+    "See the [teleoperation docs](https://huggingface.co/docs/lerobot/il_robots#teleoperate) and the [cameras guide](https://huggingface.co/docs/lerobot/cameras) for more options."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print_cmd(\n",
+    "    \"lerobot-teleoperate\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    f\"--robot.id={ROBOT_ID}\",\n",
+    "    CAMERAS_FLAG,\n",
+    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
+    "    f\"--teleop.port={TELEOP_PORT}\",\n",
+    "    f\"--teleop.id={TELEOP_ID}\",\n",
+    "    \"--display_data=true\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## 3. Record Dataset\n",
+    "\n",
+    "See the [recording docs](https://huggingface.co/docs/lerobot/il_robots#record-a-dataset) for tips on gathering good data."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print_cmd(\n",
+    "    \"lerobot-record\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    f\"--robot.id={ROBOT_ID}\",\n",
+    "    CAMERAS_FLAG,\n",
+    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
+    "    f\"--teleop.port={TELEOP_PORT}\",\n",
+    "    f\"--teleop.id={TELEOP_ID}\",\n",
+    "    f\"--dataset.repo_id={DATASET_REPO_ID}\",\n",
+    "    f\"--dataset.num_episodes={NUM_EPISODES}\",\n",
+    "    f'--dataset.single_task=\"{TASK_DESCRIPTION}\"',\n",
+    "    \"--dataset.streaming_encoding=true\",\n",
+    "    \"--display_data=true\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Resume a previously interrupted recording session\n",
+    "print_cmd(\n",
+    "    \"lerobot-record\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    f\"--robot.id={ROBOT_ID}\",\n",
+    "    CAMERAS_FLAG,\n",
+    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
+    "    f\"--teleop.port={TELEOP_PORT}\",\n",
+    "    f\"--teleop.id={TELEOP_ID}\",\n",
+    "    f\"--dataset.repo_id={DATASET_REPO_ID}\",\n",
+    "    f\"--dataset.root={DATASET_ROOT}\",\n",
+    "    f\"--dataset.num_episodes={NUM_EPISODES}\",\n",
+    "    f'--dataset.single_task=\"{TASK_DESCRIPTION}\"',\n",
+    "    \"--dataset.streaming_encoding=true\",\n",
+    "    \"--display_data=true\",\n",
+    "    \"--resume=true\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## 4. Train Policy\n",
+    "\n",
+    "See the [training docs](https://huggingface.co/docs/lerobot/il_robots#train-a-policy) for configuration options and tips."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print_cmd(\n",
+    "    \"lerobot-train\",\n",
+    "    f\"--dataset.repo_id={DATASET_REPO_ID}\",\n",
+    "    f\"--policy.type={POLICY_TYPE}\",\n",
+    "    f\"--policy.device={POLICY_DEVICE}\",\n",
+    "    f\"--policy.repo_id={POLICY_REPO_ID}\",\n",
+    "    f\"--output_dir={OUTPUT_DIR}\",\n",
+    "    f\"--steps={TRAIN_STEPS}\",\n",
+    "    f\"--save_freq={SAVE_FREQ}\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Resume a previously interrupted training session\n",
+    "print_cmd(\n",
+    "    \"lerobot-train\",\n",
+    "    f\"--config_path={LAST_CHECKPOINT_PATH}/pretrained_model/train_config.json\",\n",
+    "    \"--resume=true\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## 5. Inference\n",
+    "\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"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print_cmd(\n",
+    "    \"lerobot-record\",\n",
+    "    f\"--policy.path={POLICY_PATH}\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    f\"--robot.id={ROBOT_ID}\",\n",
+    "    CAMERAS_FLAG,\n",
+    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
+    "    f\"--teleop.port={TELEOP_PORT}\",\n",
+    "    f\"--teleop.id={TELEOP_ID}\",\n",
+    "    f\"--dataset.repo_id={EVAL_REPO_ID}\",\n",
+    "    f\"--dataset.num_episodes={NUM_EPISODES}\",\n",
+    "    f'--dataset.single_task=\"{TASK_DESCRIPTION}\"',\n",
+    "    \"--dataset.streaming_encoding=true\",\n",
+    ")"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "lerobot (3.12.3)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}

examples/omx/README.md

@@ -0,0 +1,136 @@
+# 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.

examples/omx/record_grab.py

@@ -0,0 +1,422 @@
+#!/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()

examples/omx/reset_environment.py

@@ -0,0 +1,267 @@
+#!/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()

examples/phone_to_so100/evaluate.py

@@ -14,13 +14,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import logging
+import time
+
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.common.control_utils import init_keyboard_listener, predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
+from lerobot.policies.utils import make_robot_action
 from lerobot.processor import (
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
@@ -34,11 +38,12 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.feature_utils import combine_feature_dicts
+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 lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 NUM_EPISODES = 5
 FPS = 30
@@ -49,6 +54,9 @@ HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
 
 
 def main():
+    # NOTE: For production policy deployment, use `lerobot-rollout` CLI instead.
+    # This script provides a self-contained example for educational purposes.
+
     # Create the robot configuration & robot
     camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
     robot_config = SO100FollowerConfig(
@@ -143,43 +151,67 @@ def main():
             raise ValueError("Robot is not connected!")
 
         print("Starting evaluate loop...")
+        control_interval = 1 / FPS
         episode_idx = 0
         for episode_idx in range(NUM_EPISODES):
             log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
-            # Main record loop
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                policy=policy,
-                preprocessor=preprocessor,  # Pass the pre and post policy processors
-                postprocessor=postprocessor,
-                dataset=dataset,
-                control_time_s=EPISODE_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=make_default_teleop_action_processor(),
-                robot_action_processor=robot_ee_to_joints_processor,
-                robot_observation_processor=robot_joints_to_ee_pose_processor,
-            )
+            # Inline evaluation loop: predict actions and send to robot
+            timestamp = 0
+            start_episode_t = time.perf_counter()
+            while timestamp < EPISODE_TIME_SEC:
+                start_loop_t = time.perf_counter()
+
+                if events["exit_early"]:
+                    events["exit_early"] = False
+                    break
+
+                # Get robot observation
+                obs = robot.get_observation()
+                obs_processed = robot_joints_to_ee_pose_processor(obs)
+                observation_frame = build_dataset_frame(dataset.features, obs_processed, prefix=OBS_STR)
+
+                # Predict action using the policy
+                action_tensor = predict_action(
+                    observation=observation_frame,
+                    policy=policy,
+                    device=policy.config.device,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    use_amp=policy.config.device.type == "cuda",
+                    task=TASK_DESCRIPTION,
+                    robot_type=robot.name,
+                )
+
+                # Convert policy output to robot action dict
+                action_values = make_robot_action(action_tensor, dataset.features)
+
+                # Process and send action to robot (EE -> joints via IK)
+                robot_action_to_send = robot_ee_to_joints_processor((action_values, obs))
+                robot.send_action(robot_action_to_send)
+
+                # Write to dataset
+                action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
+                frame = {**observation_frame, **action_frame, "task": TASK_DESCRIPTION}
+                dataset.add_frame(frame)
+
+                log_rerun_data(observation=obs_processed, action=action_values)
+
+                dt_s = time.perf_counter() - start_loop_t
+                sleep_time_s = control_interval - dt_s
+                if sleep_time_s < 0:
+                    logging.warning(
+                        f"Evaluate loop is running slower ({1 / dt_s:.1f} Hz) than the target FPS ({FPS} Hz)."
+                    )
+                precise_sleep(max(sleep_time_s, 0.0))
+                timestamp = time.perf_counter() - start_episode_t
 
             # 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,
-                    control_time_s=EPISODE_TIME_SEC,
-                    single_task=TASK_DESCRIPTION,
-                    display_data=True,
-                    teleop_action_processor=make_default_teleop_action_processor(),
-                    robot_action_processor=robot_ee_to_joints_processor,
-                    robot_observation_processor=robot_joints_to_ee_pose_processor,
-                )
+                log_say("Waiting for environment reset, press right arrow key when ready...")
 
             if events["rerecord_episode"]:
                 log_say("Re-record episode")
@@ -190,7 +222,6 @@ def main():
 
             # Save episode
             dataset.save_episode()
-            episode_idx += 1
     finally:
         # Clean up
         log_say("Stop recording")

examples/phone_to_so100/record.py

@@ -65,14 +65,15 @@ def main():
     robot = SO100Follower(robot_config)
     phone = Phone(teleop_config)
 
-    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo:
+    #   https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
     kinematics_solver = RobotKinematics(
         urdf_path="./SO101/so101_new_calib.urdf",
         target_frame_name="gripper_frame_link",
         joint_names=list(robot.bus.motors.keys()),
     )
 
-    # Build pipeline to convert phone action to EE action
+    # Build pipeline to convert phone action to EE action (with gripper velocity mapped to joint).
     phone_to_robot_ee_pose_processor = RobotProcessorPipeline[
         tuple[RobotAction, RobotObservation], RobotAction
     ](
@@ -94,7 +95,7 @@ def main():
         to_output=transition_to_robot_action,
     )
 
-    # Build pipeline to convert EE action to joints action
+    # Build pipeline to convert EE action to joints action (IK).
     robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
         steps=[
             InverseKinematicsEEToJoints(
@@ -107,7 +108,7 @@ def main():
         to_output=transition_to_robot_action,
     )
 
-    # Build pipeline to convert joint observation to EE observation
+    # Build pipeline to convert joint observation to EE observation (FK).
     robot_joints_to_ee_pose = RobotProcessorPipeline[RobotObservation, RobotObservation](
         steps=[
             ForwardKinematicsJointsToEE(
@@ -118,13 +119,12 @@ def main():
         to_output=transition_to_observation,
     )
 
-    # Create the dataset
+    # Create the dataset, deriving features from the pipelines so the on-disk schema
+    # matches exactly what the pipelines produce at runtime.
     dataset = LeRobotDataset.create(
         repo_id=HF_REPO_ID,
         fps=FPS,
         features=combine_feature_dicts(
-            # Run the feature contract of the pipelines
-            # This tells you how the features would look like after the pipeline steps
             aggregate_pipeline_dataset_features(
                 pipeline=phone_to_robot_ee_pose_processor,
                 initial_features=create_initial_features(action=phone.action_features),
@@ -163,14 +163,14 @@ def main():
                 robot=robot,
                 events=events,
                 fps=FPS,
+                teleop_action_processor=phone_to_robot_ee_pose_processor,
+                robot_action_processor=robot_ee_to_joints_processor,
+                robot_observation_processor=robot_joints_to_ee_pose,
                 teleop=phone,
                 dataset=dataset,
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
-                teleop_action_processor=phone_to_robot_ee_pose_processor,
-                robot_action_processor=robot_ee_to_joints_processor,
-                robot_observation_processor=robot_joints_to_ee_pose,
             )
 
             # Reset the environment if not stopping or re-recording
@@ -182,13 +182,13 @@ def main():
                     robot=robot,
                     events=events,
                     fps=FPS,
+                    teleop_action_processor=phone_to_robot_ee_pose_processor,
+                    robot_action_processor=robot_ee_to_joints_processor,
+                    robot_observation_processor=robot_joints_to_ee_pose,
                     teleop=phone,
                     control_time_s=RESET_TIME_SEC,
                     single_task=TASK_DESCRIPTION,
                     display_data=True,
-                    teleop_action_processor=phone_to_robot_ee_pose_processor,
-                    robot_action_processor=robot_ee_to_joints_processor,
-                    robot_observation_processor=robot_joints_to_ee_pose,
                 )
 
             if events["rerecord_episode"]:

examples/phone_to_so100/rollout.py

@@ -0,0 +1,126 @@
+# !/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.
+
+"""Run a trained EE-space policy on SO100 (phone-trained) without recording.
+
+Mirrors ``examples/so100_to_so100_EE/rollout.py`` — the model was trained
+with phone teleoperation in EE space, so at deployment we only need the
+joint↔EE conversion on the robot side; the phone is not used.
+
+Uses :class:`BaseStrategy` (no recording) + :class:`SyncInferenceConfig`
+(inline policy call).  For recording during rollout, switch to Sentry,
+Highlight, or DAgger via ``lerobot-rollout --strategy.type=...``.
+"""
+
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.configs import PreTrainedConfig
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import (
+    RobotProcessorPipeline,
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
+from lerobot.rollout.inference import SyncInferenceConfig
+from lerobot.rollout.strategies import BaseStrategy
+from lerobot.types import RobotAction, RobotObservation
+from lerobot.utils.process import ProcessSignalHandler
+from lerobot.utils.utils import init_logging
+
+FPS = 30
+DURATION_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+
+
+def main():
+    init_logging()
+
+    camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem58760434471",
+        id="my_awesome_follower_arm",
+        cameras=camera_config,
+        use_degrees=True,
+    )
+
+    # Peek at motor names once to build the kinematic solver.
+    temp_robot = SO100Follower(robot_config)
+    motor_names = list(temp_robot.bus.motors.keys())
+
+    kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=motor_names,
+    )
+
+    robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=motor_names)],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+
+    robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=motor_names,
+                initial_guess_current_joints=True,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    policy_config = PreTrainedConfig.from_pretrained(HF_MODEL_ID)
+    policy_config.pretrained_path = HF_MODEL_ID
+
+    cfg = RolloutConfig(
+        robot=robot_config,
+        policy=policy_config,
+        strategy=BaseStrategyConfig(),
+        inference=SyncInferenceConfig(),
+        fps=FPS,
+        duration=DURATION_SEC,
+        task=TASK_DESCRIPTION,
+    )
+
+    signal_handler = ProcessSignalHandler(use_threads=True)
+
+    ctx = build_rollout_context(
+        cfg,
+        signal_handler.shutdown_event,
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose_processor,
+    )
+
+    strategy = BaseStrategy(cfg.strategy)
+    try:
+        strategy.setup(ctx)
+        strategy.run(ctx)
+    finally:
+        strategy.teardown(ctx)
+
+
+if __name__ == "__main__":
+    main()

examples/rtc/eval_with_real_robot.py

@@ -1,673 +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.
-
-"""
-Demo script showing how to use Real-Time Chunking (RTC) with action chunking policies on real robots.
-
-This script demonstrates:
-1. Creating a robot and policy (SmolVLA, Pi0, etc.) with RTC
-2. Consuming actions from the policy while the robot executes
-3. Periodically requesting new action chunks in the background using threads
-4. Managing action buffers and timing for real-time operation
-
-For simulation environments, see eval_with_simulation.py
-
-Usage:
-    # Run RTC with Real robot with RTC
-    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --policy.device=mps \
-        --rtc.enabled=true \
-        --rtc.execution_horizon=20 \
-        --robot.type=so100_follower \
-        --robot.port=/dev/tty.usbmodem58FA0834591 \
-        --robot.id=so100_follower \
-        --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-        --task="Move green small object into the purple platform" \
-        --duration=120
-
-    # Run RTC with Real robot without RTC
-    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --policy.device=mps \
-        --rtc.enabled=false \
-        --robot.type=so100_follower \
-        --robot.port=/dev/tty.usbmodem58FA0834591 \
-        --robot.id=so100_follower \
-        --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-        --task="Move green small object into the purple platform" \
-        --duration=120
-
-    # Run RTC with Real robot with pi0.5 policy
-    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/pi05_check_rtc \
-        --policy.device=mps \
-        --rtc.enabled=true \
-        --rtc.execution_horizon=20 \
-        --robot.type=so100_follower \
-        --robot.port=/dev/tty.usbmodem58FA0834591 \
-        --robot.id=so100_follower \
-        --robot.cameras="{ gripper: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}}" \
-        --task="Move green small object into the purple platform" \
-        --duration=120
-
-    # Run RTC with bi_openarm_follower (dual-arm OpenArms) and pi0.5 policy
-    python examples/rtc/eval_with_real_robot.py \
-        --policy.path=lerobot-data-collection/folding_final \
-        --robot.type=bi_openarm_follower \
-        --robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}}' \
-        --robot.left_arm_config.port=can0 \
-        --robot.left_arm_config.side=left \
-        --robot.left_arm_config.can_interface=socketcan \
-        --robot.left_arm_config.disable_torque_on_disconnect=true \
-        --robot.left_arm_config.max_relative_target=8.0 \
-        --robot.right_arm_config.port=can1 \
-        --robot.right_arm_config.side=right \
-        --robot.right_arm_config.can_interface=socketcan \
-        --robot.right_arm_config.disable_torque_on_disconnect=true \
-        --robot.right_arm_config.max_relative_target=8.0 \
-        --task="Fold the T-shirt properly" \
-        --fps=30 \
-        --duration=2000 \
-        --interpolation_multiplier=3 \
-        --rtc.enabled=true \
-        --rtc.execution_horizon=20 \
-        --rtc.max_guidance_weight=5.0 \
-        --rtc.prefix_attention_schedule=LINEAR \
-        --device=cuda
-"""
-
-import logging
-import math
-import sys
-import time
-import traceback
-from dataclasses import dataclass, field
-from threading import Event, Lock, Thread
-
-import torch
-from torch import Tensor
-
-from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
-from lerobot.cameras.realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.cameras.zmq import ZMQCameraConfig  # noqa: F401
-from lerobot.configs import PreTrainedConfig, RTCAttentionSchedule, parser
-from lerobot.policies import get_policy_class, make_pre_post_processors
-from lerobot.policies.rtc import ActionInterpolator, ActionQueue, LatencyTracker, RTCConfig
-from lerobot.processor import (
-    NormalizerProcessorStep,
-    RelativeActionsProcessorStep,
-    TransitionKey,
-    create_transition,
-    make_default_robot_action_processor,
-    make_default_robot_observation_processor,
-    to_relative_actions,
-)
-from lerobot.rl.process import ProcessSignalHandler
-from lerobot.robots import (  # noqa: F401
-    Robot,
-    RobotConfig,
-    bi_openarm_follower,
-    bi_so_follower,
-    koch_follower,
-    so_follower,
-    unitree_g1,
-)
-from lerobot.robots.utils import make_robot_from_config
-from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
-from lerobot.utils.feature_utils import build_dataset_frame, hw_to_dataset_features
-from lerobot.utils.hub import HubMixin
-from lerobot.utils.utils import init_logging
-
-logging.basicConfig(level=logging.INFO)
-logger = logging.getLogger(__name__)
-
-
-class RobotWrapper:
-    def __init__(self, robot: Robot):
-        self.robot = robot
-        self.lock = Lock()
-
-    def get_observation(self) -> dict[str, Tensor]:
-        with self.lock:
-            return self.robot.get_observation()
-
-    def send_action(self, action: Tensor):
-        with self.lock:
-            self.robot.send_action(action)
-
-    def observation_features(self) -> list[str]:
-        with self.lock:
-            return self.robot.observation_features
-
-    def action_features(self) -> list[str]:
-        with self.lock:
-            return self.robot.action_features
-
-
-@dataclass
-class RTCDemoConfig(HubMixin):
-    """Configuration for RTC demo with action chunking policies and real robots."""
-
-    # Policy configuration
-    policy: PreTrainedConfig | None = None
-
-    # Robot configuration
-    robot: RobotConfig | None = None
-
-    # RTC configuration
-    rtc: RTCConfig = field(
-        default_factory=lambda: RTCConfig(
-            execution_horizon=10,
-            max_guidance_weight=1.0,
-            prefix_attention_schedule=RTCAttentionSchedule.EXP,
-        )
-    )
-
-    # Demo parameters
-    duration: float = 30.0  # Duration to run the demo (seconds)
-    fps: float = 10.0  # Action execution frequency (Hz)
-    interpolation_multiplier: int = 1  # Control rate multiplier (1=off, 2=2x, 3=3x)
-
-    # Compute device
-    device: str | None = None  # Device to run on (cuda, cpu, auto)
-
-    # Get new actions horizon. The amount of executed steps after which will be requested new actions.
-    # It should be higher than inference delay + execution horizon.
-    action_queue_size_to_get_new_actions: int = 30
-
-    # Task to execute
-    task: str = field(default="", metadata={"help": "Task to execute"})
-
-    # Torch compile configuration
-    use_torch_compile: bool = field(
-        default=False,
-        metadata={"help": "Use torch.compile for faster inference (PyTorch 2.0+)"},
-    )
-
-    torch_compile_backend: str = field(
-        default="inductor",
-        metadata={"help": "Backend for torch.compile (inductor, aot_eager, cudagraphs)"},
-    )
-
-    torch_compile_mode: str = field(
-        default="default",
-        metadata={"help": "Compilation mode (default, reduce-overhead, max-autotune)"},
-    )
-
-    torch_compile_disable_cudagraphs: bool = field(
-        default=True,
-        metadata={
-            "help": "Disable CUDA graphs in torch.compile. Required due to in-place tensor "
-            "operations in denoising loop (x_t += dt * v_t) which cause tensor aliasing issues."
-        },
-    )
-
-    def __post_init__(self):
-        # 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:
-            cli_overrides = parser.get_cli_overrides("policy")
-            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
-            self.policy.pretrained_path = policy_path
-        else:
-            raise ValueError("Policy path is required")
-
-        # Validate that robot configuration is provided
-        if self.robot is None:
-            raise ValueError("Robot configuration must be provided")
-
-    @classmethod
-    def __get_path_fields__(cls) -> list[str]:
-        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
-        return ["policy"]
-
-
-def is_image_key(k: str) -> bool:
-    return k.startswith(OBS_IMAGES)
-
-
-def _reanchor_relative_rtc_prefix(
-    prev_actions_absolute: Tensor,
-    current_state: Tensor,
-    relative_step: RelativeActionsProcessorStep,
-    normalizer_step: NormalizerProcessorStep | None,
-    policy_device: torch.device | str,
-) -> Tensor:
-    """Convert absolute leftovers into model-space for relative-action RTC policies.
-
-    When a policy uses relative actions, the RTC prefix (leftover actions from
-    the previous chunk) is stored in absolute space. Before feeding it back to
-    the policy we need to re-express it relative to the *current* robot state
-    and then re-normalize.
-    """
-    state = current_state.detach().cpu()
-    if state.dim() == 1:
-        state = state.unsqueeze(0)
-
-    action_cpu = prev_actions_absolute.detach().cpu()
-    mask = relative_step._build_mask(action_cpu.shape[-1])
-    relative_actions = to_relative_actions(action_cpu, state, mask)
-
-    transition = create_transition(action=relative_actions)
-    if normalizer_step is not None:
-        transition = normalizer_step(transition)
-
-    return transition[TransitionKey.ACTION].to(policy_device)
-
-
-def get_actions(
-    policy,
-    robot: RobotWrapper,
-    robot_observation_processor,
-    action_queue: ActionQueue,
-    shutdown_event: Event,
-    cfg: RTCDemoConfig,
-):
-    """Thread function to request action chunks from the policy.
-
-    Args:
-        policy: The policy instance (SmolVLA, Pi0, etc.)
-        robot: The robot instance for getting observations
-        robot_observation_processor: Processor for raw robot observations
-        action_queue: Queue to put new action chunks
-        shutdown_event: Event to signal shutdown
-        cfg: Demo configuration
-    """
-    try:
-        logger.info("[GET_ACTIONS] Starting get actions thread")
-
-        latency_tracker = LatencyTracker()  # Track latency of action chunks
-        fps = cfg.fps
-        time_per_chunk = 1.0 / fps
-
-        # Only keep .pos joints + camera streams if the policy was trained on positions,
-        # not the full pos/vel/torque state the robot exposes.
-        observation_features_hw = {
-            key: value
-            for key, value in robot.observation_features().items()
-            if key.endswith(".pos") or isinstance(value, tuple)
-        }
-
-        dataset_features = hw_to_dataset_features(observation_features_hw, "observation")
-        policy_device = policy.config.device
-
-        # Load preprocessor and postprocessor from pretrained files
-        # The stats are embedded in the processor .safetensors files
-        logger.info(f"[GET_ACTIONS] Loading preprocessor/postprocessor from {cfg.policy.pretrained_path}")
-
-        preprocessor, postprocessor = make_pre_post_processors(
-            policy_cfg=cfg.policy,
-            pretrained_path=cfg.policy.pretrained_path,
-            dataset_stats=None,  # Will load from pretrained processor files
-            preprocessor_overrides={
-                "device_processor": {"device": cfg.policy.device},
-            },
-        )
-
-        logger.info("[GET_ACTIONS] Preprocessor/postprocessor loaded successfully with embedded stats")
-
-        relative_step = next(
-            (s for s in preprocessor.steps if isinstance(s, RelativeActionsProcessorStep) and s.enabled),
-            None,
-        )
-        normalizer_step = next(
-            (s for s in preprocessor.steps if isinstance(s, NormalizerProcessorStep)),
-            None,
-        )
-        if relative_step is not None:
-            if relative_step.action_names is None:
-                cfg_names = getattr(cfg.policy, "action_feature_names", None)
-                if cfg_names:
-                    relative_step.action_names = list(cfg_names)
-                else:
-                    relative_step.action_names = [
-                        k for k in robot.robot.action_features if k.endswith(".pos")
-                    ]
-            logger.info("[GET_ACTIONS] Relative actions enabled: will re-anchor RTC prefix")
-
-        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
-
-        if not cfg.rtc.enabled:
-            get_actions_threshold = 0
-
-        while not shutdown_event.is_set():
-            if action_queue.qsize() <= get_actions_threshold:
-                current_time = time.perf_counter()
-                action_index_before_inference = action_queue.get_action_index()
-                prev_actions = action_queue.get_left_over()
-
-                inference_latency = latency_tracker.max()
-                inference_delay = math.ceil(inference_latency / time_per_chunk)
-
-                obs = robot.get_observation()
-
-                # Apply robot observation processor
-                obs_processed = robot_observation_processor(obs)
-
-                obs_with_policy_features = build_dataset_frame(
-                    dataset_features, obs_processed, prefix="observation"
-                )
-
-                for name in obs_with_policy_features:
-                    obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
-                    if "image" in name:
-                        obs_with_policy_features[name] = (
-                            obs_with_policy_features[name].type(torch.float32) / 255
-                        )
-                        obs_with_policy_features[name] = (
-                            obs_with_policy_features[name].permute(2, 0, 1).contiguous()
-                        )
-                    obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0)
-                    obs_with_policy_features[name] = obs_with_policy_features[name].to(policy_device)
-
-                obs_with_policy_features["task"] = [cfg.task]  # Task should be a list, not a string!
-                obs_with_policy_features["robot_type"] = (
-                    robot.robot.name if hasattr(robot.robot, "name") else ""
-                )
-
-                preproceseded_obs = preprocessor(obs_with_policy_features)
-
-                # Re-anchor leftover actions for relative-action policies.
-                # We need the *postprocessed* (absolute) leftover, not the original
-                # (normalized/relative) one that get_left_over() returns.
-                if (
-                    prev_actions is not None
-                    and relative_step is not None
-                    and OBS_STATE in obs_with_policy_features
-                ):
-                    with action_queue.lock:
-                        if action_queue.queue is not None:
-                            prev_actions_abs = action_queue.queue[action_queue.last_index :].clone()
-                        else:
-                            prev_actions_abs = None
-                    if prev_actions_abs is not None and prev_actions_abs.numel() > 0:
-                        prev_actions = _reanchor_relative_rtc_prefix(
-                            prev_actions_absolute=prev_actions_abs,
-                            current_state=obs_with_policy_features[OBS_STATE],
-                            relative_step=relative_step,
-                            normalizer_step=normalizer_step,
-                            policy_device=policy_device,
-                        )
-
-                # Generate actions WITH RTC
-                actions = policy.predict_action_chunk(
-                    preproceseded_obs,
-                    inference_delay=inference_delay,
-                    prev_chunk_left_over=prev_actions,
-                )
-
-                # Store original actions (before postprocessing) for RTC
-                original_actions = actions.squeeze(0).clone()
-
-                postprocessed_actions = postprocessor(actions)
-
-                postprocessed_actions = postprocessed_actions.squeeze(0)
-
-                new_latency = time.perf_counter() - current_time
-                new_delay = math.ceil(new_latency / time_per_chunk)
-                latency_tracker.add(new_latency)
-
-                if cfg.action_queue_size_to_get_new_actions < cfg.rtc.execution_horizon + new_delay:
-                    logger.warning(
-                        "[GET_ACTIONS] cfg.action_queue_size_to_get_new_actions Too small, It should be higher than inference delay + execution horizon."
-                    )
-
-                action_queue.merge(
-                    original_actions, postprocessed_actions, new_delay, action_index_before_inference
-                )
-            else:
-                # Small sleep to prevent busy waiting
-                time.sleep(0.1)
-
-        logger.info("[GET_ACTIONS] get actions thread shutting down")
-    except Exception as e:
-        logger.error(f"[GET_ACTIONS] Fatal exception in get_actions thread: {e}")
-        logger.error(traceback.format_exc())
-        sys.exit(1)
-
-
-def actor_control(
-    robot: RobotWrapper,
-    robot_action_processor,
-    action_queue: ActionQueue,
-    shutdown_event: Event,
-    cfg: RTCDemoConfig,
-):
-    """Thread function to execute actions on the robot.
-
-    Args:
-        robot: The robot instance
-        action_queue: Queue to get actions from
-        shutdown_event: Event to signal shutdown
-        cfg: Demo configuration
-    """
-    try:
-        logger.info("[ACTOR] Starting actor thread")
-
-        action_keys = [k for k in robot.action_features() if k.endswith(".pos")]
-
-        action_count = 0
-        interpolator = ActionInterpolator(multiplier=cfg.interpolation_multiplier)
-        action_interval = interpolator.get_control_interval(cfg.fps)
-
-        while not shutdown_event.is_set():
-            start_time = time.perf_counter()
-
-            if interpolator.needs_new_action():
-                new_action = action_queue.get()
-                if new_action is not None:
-                    interpolator.add(new_action.cpu())
-
-            action = interpolator.get()
-            if action is not None:
-                action = action.cpu()
-                action_dict = {key: action[i].item() for i, key in enumerate(action_keys)}
-                action_processed = robot_action_processor((action_dict, None))
-                robot.send_action(action_processed)
-                action_count += 1
-
-            dt_s = time.perf_counter() - start_time
-            time.sleep(max(0, (action_interval - dt_s) - 0.001))
-
-        logger.info(f"[ACTOR] Actor thread shutting down. Total actions executed: {action_count}")
-    except Exception as e:
-        logger.error(f"[ACTOR] Fatal exception in actor_control thread: {e}")
-        logger.error(traceback.format_exc())
-        sys.exit(1)
-
-
-def _apply_torch_compile(policy, cfg: RTCDemoConfig):
-    """Apply torch.compile to the policy's predict_action_chunk method.
-
-    Args:
-        policy: Policy instance to compile
-        cfg: Configuration containing torch compile settings
-
-    Returns:
-        Policy with compiled predict_action_chunk method
-    """
-
-    # PI models handle their own compilation
-    if policy.type == "pi05" or policy.type == "pi0":
-        return policy
-
-    try:
-        # Check if torch.compile is available (PyTorch 2.0+)
-        if not hasattr(torch, "compile"):
-            logger.warning(
-                f"torch.compile is not available. Requires PyTorch 2.0+. "
-                f"Current version: {torch.__version__}. Skipping compilation."
-            )
-            return policy
-
-        logger.info("Applying torch.compile to predict_action_chunk...")
-        logger.info(f"  Backend: {cfg.torch_compile_backend}")
-        logger.info(f"  Mode: {cfg.torch_compile_mode}")
-        logger.info(f"  Disable CUDA graphs: {cfg.torch_compile_disable_cudagraphs}")
-
-        # Compile the predict_action_chunk method
-        # - CUDA graphs disabled to prevent tensor aliasing from in-place ops (x_t += dt * v_t)
-        compile_kwargs = {
-            "backend": cfg.torch_compile_backend,
-            "mode": cfg.torch_compile_mode,
-        }
-
-        # Disable CUDA graphs if requested (prevents tensor aliasing issues)
-        if cfg.torch_compile_disable_cudagraphs:
-            compile_kwargs["options"] = {"triton.cudagraphs": False}
-
-        original_method = policy.predict_action_chunk
-        compiled_method = torch.compile(original_method, **compile_kwargs)
-        policy.predict_action_chunk = compiled_method
-        logger.info("✓ Successfully compiled predict_action_chunk")
-
-    except Exception as e:
-        logger.error(f"Failed to apply torch.compile: {e}")
-        logger.warning("Continuing without torch.compile")
-
-    return policy
-
-
-@parser.wrap()
-def demo_cli(cfg: RTCDemoConfig):
-    """Main entry point for RTC demo with draccus configuration."""
-
-    # Initialize logging
-    init_logging()
-
-    logger.info(f"Using device: {cfg.device}")
-
-    # Setup signal handler for graceful shutdown
-    signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
-    shutdown_event = signal_handler.shutdown_event
-
-    policy = None
-    robot = None
-    get_actions_thread = None
-    actor_thread = None
-
-    policy_class = get_policy_class(cfg.policy.type)
-
-    # Load config and set compile_model for pi0/pi05 models
-    config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
-
-    if cfg.policy.type == "pi05" or cfg.policy.type == "pi0":
-        config.compile_model = cfg.use_torch_compile
-
-    if config.use_peft:
-        from peft import PeftConfig, PeftModel
-
-        peft_pretrained_path = cfg.policy.pretrained_path
-        peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
-
-        policy = policy_class.from_pretrained(
-            pretrained_name_or_path=peft_config.base_model_name_or_path, config=config
-        )
-        policy = PeftModel.from_pretrained(policy, peft_pretrained_path, config=peft_config)
-    else:
-        policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
-
-    # Turn on RTC
-    policy.config.rtc_config = cfg.rtc
-
-    # Init RTC processort, as by default if RTC disabled in the config
-    # The processor won't be created
-    policy.init_rtc_processor()
-
-    assert policy.name in ["smolvla", "pi05", "pi0"], "Only smolvla, pi05, and pi0 are supported for RTC"
-
-    policy = policy.to(cfg.device)
-    policy.eval()
-
-    # Apply torch.compile to predict_action_chunk method if enabled
-    if cfg.use_torch_compile:
-        policy = _apply_torch_compile(policy, cfg)
-
-    # Create robot
-    logger.info(f"Initializing robot: {cfg.robot.type}")
-    robot = make_robot_from_config(cfg.robot)
-    robot.connect()
-    robot_wrapper = RobotWrapper(robot)
-
-    # Create robot observation processor
-    robot_observation_processor = make_default_robot_observation_processor()
-    robot_action_processor = make_default_robot_action_processor()
-
-    # Create action queue for communication between threads
-    action_queue = ActionQueue(cfg.rtc)
-
-    # Start chunk requester thread
-    get_actions_thread = Thread(
-        target=get_actions,
-        args=(policy, robot_wrapper, robot_observation_processor, action_queue, shutdown_event, cfg),
-        daemon=True,
-        name="GetActions",
-    )
-    get_actions_thread.start()
-    logger.info("Started get actions thread")
-
-    # Start action executor thread
-    actor_thread = Thread(
-        target=actor_control,
-        args=(robot_wrapper, robot_action_processor, action_queue, shutdown_event, cfg),
-        daemon=True,
-        name="Actor",
-    )
-    actor_thread.start()
-    logger.info("Started actor thread")
-
-    logger.info("Started stop by duration thread")
-
-    # Main thread monitors for duration or shutdown
-    logger.info(f"Running demo for {cfg.duration} seconds...")
-    start_time = time.time()
-
-    while not shutdown_event.is_set() and (time.time() - start_time) < cfg.duration:
-        time.sleep(10)
-
-        # Log queue status periodically
-        if int(time.time() - start_time) % 5 == 0:
-            logger.info(f"[MAIN] Action queue size: {action_queue.qsize()}")
-
-        if time.time() - start_time > cfg.duration:
-            break
-
-    logger.info("Demo duration reached or shutdown requested")
-
-    # Signal shutdown
-    shutdown_event.set()
-
-    # Wait for threads to finish
-    if get_actions_thread and get_actions_thread.is_alive():
-        logger.info("Waiting for chunk requester thread to finish...")
-        get_actions_thread.join()
-
-    if actor_thread and actor_thread.is_alive():
-        logger.info("Waiting for action executor thread to finish...")
-        actor_thread.join()
-
-    # Cleanup robot
-    if robot:
-        robot.disconnect()
-        logger.info("Robot disconnected")
-
-    logger.info("Cleanup completed")
-
-
-if __name__ == "__main__":
-    demo_cli()
-    logging.info("RTC demo finished")

examples/so100_to_so100_EE/evaluate.py

@@ -14,13 +14,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import logging
+import time
+
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.common.control_utils import init_keyboard_listener, predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
+from lerobot.policies.utils import make_robot_action
 from lerobot.processor import (
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
@@ -34,11 +38,12 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.feature_utils import combine_feature_dicts
+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 lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 NUM_EPISODES = 5
 FPS = 30
@@ -49,6 +54,9 @@ HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
 
 
 def main():
+    # NOTE: For production policy deployment, use `lerobot-rollout` CLI instead.
+    # This script provides a self-contained example for educational purposes.
+
     # Create the robot configuration & robot
     camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
     robot_config = SO100FollowerConfig(
@@ -143,43 +151,67 @@ def main():
             raise ValueError("Robot is not connected!")
 
         print("Starting evaluate loop...")
+        control_interval = 1 / FPS
         episode_idx = 0
         for episode_idx in range(NUM_EPISODES):
             log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
-            # Main record loop
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                policy=policy,
-                preprocessor=preprocessor,  # Pass the pre and post policy processors
-                postprocessor=postprocessor,
-                dataset=dataset,
-                control_time_s=EPISODE_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=make_default_teleop_action_processor(),
-                robot_action_processor=robot_ee_to_joints_processor,
-                robot_observation_processor=robot_joints_to_ee_pose_processor,
-            )
+            # Inline evaluation loop: predict actions and send to robot
+            timestamp = 0
+            start_episode_t = time.perf_counter()
+            while timestamp < EPISODE_TIME_SEC:
+                start_loop_t = time.perf_counter()
+
+                if events["exit_early"]:
+                    events["exit_early"] = False
+                    break
+
+                # Get robot observation
+                obs = robot.get_observation()
+                obs_processed = robot_joints_to_ee_pose_processor(obs)
+                observation_frame = build_dataset_frame(dataset.features, obs_processed, prefix=OBS_STR)
+
+                # Predict action using the policy
+                action_tensor = predict_action(
+                    observation=observation_frame,
+                    policy=policy,
+                    device=policy.config.device,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    use_amp=policy.config.device.type == "cuda",
+                    task=TASK_DESCRIPTION,
+                    robot_type=robot.name,
+                )
+
+                # Convert policy output to robot action dict
+                action_values = make_robot_action(action_tensor, dataset.features)
+
+                # Process and send action to robot (EE -> joints via IK)
+                robot_action_to_send = robot_ee_to_joints_processor((action_values, obs))
+                robot.send_action(robot_action_to_send)
+
+                # Write to dataset
+                action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
+                frame = {**observation_frame, **action_frame, "task": TASK_DESCRIPTION}
+                dataset.add_frame(frame)
+
+                log_rerun_data(observation=obs_processed, action=action_values)
+
+                dt_s = time.perf_counter() - start_loop_t
+                sleep_time_s = control_interval - dt_s
+                if sleep_time_s < 0:
+                    logging.warning(
+                        f"Evaluate loop is running slower ({1 / dt_s:.1f} Hz) than the target FPS ({FPS} Hz)."
+                    )
+                precise_sleep(max(sleep_time_s, 0.0))
+                timestamp = time.perf_counter() - start_episode_t
 
             # 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,
-                    control_time_s=EPISODE_TIME_SEC,
-                    single_task=TASK_DESCRIPTION,
-                    display_data=True,
-                    teleop_action_processor=make_default_teleop_action_processor(),
-                    robot_action_processor=robot_ee_to_joints_processor,
-                    robot_observation_processor=robot_joints_to_ee_pose_processor,
-                )
+                log_say("Waiting for environment reset, press right arrow key when ready...")
 
             if events["rerecord_episode"]:
                 log_say("Re-record episode")
@@ -190,7 +222,6 @@ def main():
 
             # Save episode
             dataset.save_episode()
-            episode_idx += 1
     finally:
         # Clean up
         log_say("Stop recording")

examples/so100_to_so100_EE/record.py

@@ -62,21 +62,20 @@ def main():
     follower = SO100Follower(follower_config)
     leader = SO100Leader(leader_config)
 
-    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo:
+    #   https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
     follower_kinematics_solver = RobotKinematics(
         urdf_path="./SO101/so101_new_calib.urdf",
         target_frame_name="gripper_frame_link",
         joint_names=list(follower.bus.motors.keys()),
     )
-
-    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
     leader_kinematics_solver = RobotKinematics(
         urdf_path="./SO101/so101_new_calib.urdf",
         target_frame_name="gripper_frame_link",
         joint_names=list(leader.bus.motors.keys()),
     )
 
-    # Build pipeline to convert follower joints to EE observation
+    # Build pipeline to convert follower joints to EE observation.
     follower_joints_to_ee = RobotProcessorPipeline[RobotObservation, RobotObservation](
         steps=[
             ForwardKinematicsJointsToEE(
@@ -87,7 +86,7 @@ def main():
         to_output=transition_to_observation,
     )
 
-    # Build pipeline to convert leader joints to EE action
+    # Build pipeline to convert leader joints to EE action.
     leader_joints_to_ee = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
         steps=[
             ForwardKinematicsJointsToEE(
@@ -98,9 +97,9 @@ def main():
         to_output=transition_to_robot_action,
     )
 
-    # Build pipeline to convert EE action to follower joints
+    # Build pipeline to convert EE action to follower joints (with safety bounds).
     ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-        [
+        steps=[
             EEBoundsAndSafety(
                 end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
                 max_ee_step_m=0.10,
@@ -115,13 +114,12 @@ def main():
         to_output=transition_to_robot_action,
     )
 
-    # Create the dataset
+    # Create the dataset, deriving features from the pipelines so the on-disk schema
+    # matches exactly what the pipelines produce at runtime.
     dataset = LeRobotDataset.create(
         repo_id=HF_REPO_ID,
         fps=FPS,
         features=combine_feature_dicts(
-            # Run the feature contract of the pipelines
-            # This tells you how the features would look like after the pipeline steps
             aggregate_pipeline_dataset_features(
                 pipeline=leader_joints_to_ee,
                 initial_features=create_initial_features(action=leader.action_features),
@@ -144,7 +142,7 @@ def main():
 
     # Initialize the keyboard listener and rerun visualization
     listener, events = init_keyboard_listener()
-    init_rerun(session_name="recording_phone")
+    init_rerun(session_name="recording_so100_ee")
 
     try:
         if not leader.is_connected or not follower.is_connected:
@@ -160,14 +158,14 @@ def main():
                 robot=follower,
                 events=events,
                 fps=FPS,
+                teleop_action_processor=leader_joints_to_ee,
+                robot_action_processor=ee_to_follower_joints,
+                robot_observation_processor=follower_joints_to_ee,
                 teleop=leader,
                 dataset=dataset,
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
-                teleop_action_processor=leader_joints_to_ee,
-                robot_action_processor=ee_to_follower_joints,
-                robot_observation_processor=follower_joints_to_ee,
             )
 
             # Reset the environment if not stopping or re-recording
@@ -179,13 +177,13 @@ def main():
                     robot=follower,
                     events=events,
                     fps=FPS,
+                    teleop_action_processor=leader_joints_to_ee,
+                    robot_action_processor=ee_to_follower_joints,
+                    robot_observation_processor=follower_joints_to_ee,
                     teleop=leader,
                     control_time_s=RESET_TIME_SEC,
                     single_task=TASK_DESCRIPTION,
                     display_data=True,
-                    teleop_action_processor=leader_joints_to_ee,
-                    robot_action_processor=ee_to_follower_joints,
-                    robot_observation_processor=follower_joints_to_ee,
                 )
 
             if events["rerecord_episode"]:

examples/so100_to_so100_EE/rollout.py

@@ -0,0 +1,134 @@
+# !/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.
+
+"""Run a trained EE-space policy on SO100 without recording (base rollout).
+
+Uses the rollout engine's :class:`BaseStrategy` (autonomous execution,
+no dataset) with :class:`SyncInferenceConfig` (inline policy call per
+control tick).  The custom observation/action processors convert between
+joint space (robot hardware) and end-effector space (policy I/O) via
+forward/inverse kinematics.
+"""
+
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.configs import PreTrainedConfig
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import (
+    RobotProcessorPipeline,
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
+from lerobot.rollout.inference import SyncInferenceConfig
+from lerobot.rollout.strategies import BaseStrategy
+from lerobot.types import RobotAction, RobotObservation
+from lerobot.utils.process import ProcessSignalHandler
+from lerobot.utils.utils import init_logging
+
+FPS = 30
+DURATION_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+
+
+def main():
+    init_logging()
+
+    # Robot configuration — the rollout engine will connect it inside build_rollout_context.
+    camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411",
+        id="my_awesome_follower_arm",
+        cameras=camera_config,
+        use_degrees=True,
+    )
+
+    # Kinematic solver: we need the motor-name list, so peek at the robot once.
+    # (The rollout engine owns the connected instance; we only use this for introspection.)
+    temp_robot = SO100Follower(robot_config)
+    motor_names = list(temp_robot.bus.motors.keys())
+
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo:
+    #   https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=motor_names,
+    )
+
+    # Joint-space observation → EE-space observation (consumed by the policy).
+    robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=motor_names)],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+
+    # EE-space action (produced by the policy) → joint-space action (sent to robot).
+    robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=motor_names,
+                initial_guess_current_joints=True,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    # Policy config (full model is loaded inside build_rollout_context).
+    policy_config = PreTrainedConfig.from_pretrained(HF_MODEL_ID)
+    policy_config.pretrained_path = HF_MODEL_ID
+
+    cfg = RolloutConfig(
+        robot=robot_config,
+        policy=policy_config,
+        strategy=BaseStrategyConfig(),
+        inference=SyncInferenceConfig(),
+        fps=FPS,
+        duration=DURATION_SEC,
+        task=TASK_DESCRIPTION,
+    )
+
+    signal_handler = ProcessSignalHandler(use_threads=True)
+
+    # Pass the EE kinematic processors via kwargs; the defaults (identity) would
+    # otherwise skip the joint↔EE conversion and the policy would receive the
+    # wrong observation/action space.
+    ctx = build_rollout_context(
+        cfg,
+        signal_handler.shutdown_event,
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose_processor,
+    )
+
+    strategy = BaseStrategy(cfg.strategy)
+    try:
+        strategy.setup(ctx)
+        strategy.run(ctx)
+    finally:
+        strategy.teardown(ctx)
+
+
+if __name__ == "__main__":
+    main()

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 SACConfig
-from lerobot.policies.sac.modeling_sac import SACPolicy
-from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
+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.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: SACPolicy,
+    policy_learner: GaussianActorPolicy,
     online_buffer: ReplayBuffer,
     offline_buffer: ReplayBuffer,
     lr: float = 3e-4,
@@ -40,8 +40,9 @@ def run_learner(
     policy_learner.train()
     policy_learner.to(device)
 
-    # Create Adam optimizer from scratch - simple and clean
-    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)
+    algo_config = SACAlgorithmConfig.from_policy_config(policy_learner.config)
+    algorithm = SACAlgorithm(policy=policy_learner, config=algo_config)
+    algorithm.make_optimizers_and_scheduler()
 
     print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
     print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
@@ -83,24 +84,26 @@ def run_learner(
                 else:
                     batch[key] = online_batch[key]
 
-            loss, _ = policy_learner.forward(batch)
+            def batch_iter(b=batch):
+                while True:
+                    yield b
 
-            optimizer.zero_grad()
-            loss.backward()
-            optimizer.step()
+            stats = algorithm.update(batch_iter())
             training_step += 1
 
             if training_step % LOG_EVERY == 0:
+                log_dict = stats.to_log_dict()
                 print(
-                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
+                    f"[LEARNER] Training step {training_step}, "
+                    f"critic_loss: {log_dict.get('critic', 'N/A'):.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:
-                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
-                    parameters_queue.put_nowait(state_dict)
+                    weights = algorithm.get_weights()
+                    parameters_queue.put_nowait(weights)
                     print("[LEARNER] Sent updated parameters to actor")
                 except Full:
                     # Missing write due to queue not being consumed (should happen rarely)
@@ -113,7 +116,7 @@ def run_actor(
     transitions_queue: mp.Queue,
     parameters_queue: mp.Queue,
     shutdown_event: mp.Event,
-    policy_actor: SACPolicy,
+    policy_actor: GaussianActorPolicy,
     reward_classifier: Classifier,
     env_cfg: HILSerlRobotEnvConfig,
     device: torch.device = "mps",
@@ -144,15 +147,15 @@ def run_actor(
 
             while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
                 try:
-                    new_params = parameters_queue.get_nowait()
-                    policy_actor.load_state_dict(new_params)
+                    new_weights = parameters_queue.get_nowait()
+                    policy_actor.load_state_dict(new_weights)
                     print("[ACTOR] Updated policy parameters from learner")
                 except Empty:  # No new updated parameters available from learner, waiting
                     pass
 
-                # Get action from policy
+                # Get action from policy (returns full action: continuous + discrete)
                 policy_obs = make_policy_obs(obs, device=device)
-                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
+                action_tensor = policy_actor.select_action(policy_obs)
                 action = action_tensor.squeeze(0).cpu().numpy()
 
                 # Step environment
@@ -261,14 +264,14 @@ def main():
     action_features = hw_to_dataset_features(env.robot.action_features, "action")
 
     # Create SAC policy for action selection
-    policy_cfg = SACConfig(
+    policy_cfg = GaussianActorConfig(
         device=device,
         input_features=obs_features,
         output_features=action_features,
     )
 
-    policy_actor = SACPolicy(policy_cfg)
-    policy_learner = SACPolicy(policy_cfg)
+    policy_actor = GaussianActorPolicy(policy_cfg)
+    policy_learner = GaussianActorPolicy(policy_cfg)
 
     demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
     offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)

examples/tutorial/rl/reward_classifier_example.py

@@ -1,7 +1,7 @@
 import torch
 
 from lerobot.datasets import LeRobotDataset
-from lerobot.policies import RewardClassifierConfig, make_policy, make_pre_post_processors
+from lerobot.rewards import RewardClassifierConfig, make_reward_model, make_reward_pre_post_processors
 
 
 def main():
@@ -22,10 +22,10 @@ def main():
         model_name="microsoft/resnet-18",
     )
 
-    # 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)
+    # 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)
 
     classifier_id = "<user>/reward_classifier_hil_serl_example"
 
@@ -42,7 +42,7 @@ def main():
             batch = preprocessor(batch)
 
             # Forward pass
-            loss, output_dict = policy.forward(batch)
+            loss, output_dict = reward_model.forward(batch)
 
             # Backward pass and optimization
             optimizer.zero_grad()
@@ -58,8 +58,8 @@ def main():
 
     print("Training finished!")
 
-    # You can now save the trained policy.
-    policy.push_to_hub(classifier_id)
+    # You can now save the trained reward model.
+    reward_model.push_to_hub(classifier_id)
 
 
 if __name__ == "__main__":

pyproject.toml

@@ -59,8 +59,8 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 
 dependencies = [
     # Core ML
-    "torch>=2.7,<2.11.0",
-    "torchvision>=0.22.0,<0.26.0",
+    "torch>=2.7,<2.12.0",
+    "torchvision>=0.22.0,<0.27.0",
     "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",
@@ -99,7 +99,18 @@ dataset = [
     "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]",
-    "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).
+
+    # 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')",
     "jsonlines>=4.0.0,<5.0.0",
 ]
 training = [
@@ -108,9 +119,9 @@ training = [
     "wandb>=0.24.0,<0.25.0",
 ]
 hardware = [
-    "pynput>=1.7.8,<1.9.0",
-    "pyserial>=3.5,<4.0",
-    "deepdiff>=7.0.1,<9.0.0",
+    "lerobot[pynput-dep]",
+    "lerobot[pyserial-dep]",
+    "lerobot[deepdiff-dep]",
 ]
 viz = [
     "rerun-sdk>=0.24.0,<0.27.0",
@@ -128,7 +139,7 @@ dataset_viz = ["lerobot[dataset]", "lerobot[viz]"]
 av-dep = ["av>=15.0.0,<16.0.0"]
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.9.17"]
-transformers-dep = ["transformers==5.3.0"] # TODO(Steven): https://github.com/huggingface/lerobot/pull/3249
+transformers-dep = ["transformers>=5.4.0,<5.6.0"]
 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"]
@@ -136,10 +147,14 @@ scipy-dep = ["scipy>=1.14.0,<2.0.0"]
 diffusers-dep = ["diffusers>=0.27.2,<0.36.0"]
 qwen-vl-utils-dep = ["qwen-vl-utils>=0.0.11,<0.1.0"]
 matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0", "contourpy>=1.3.0,<2.0.0"] # NOTE: Explicitly listing contourpy helps the resolver converge faster.
+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"]
 
 # Motors
-feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
-dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
+feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
+dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
 damiao = ["lerobot[can-dep]"]
 robstride = ["lerobot[can-dep]"]
 
@@ -147,10 +162,11 @@ robstride = ["lerobot[can-dep]"]
 openarms = ["lerobot[damiao]"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
-lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
+lekiwi = ["lerobot[feetech]", "lerobot[pyzmq-dep]"]
 unitree_g1 = [
     # "unitree-sdk2==1.0.1",
-    "pyzmq>=26.2.1,<28.0.0",
+    "lerobot[pyzmq-dep]",
+    "lerobot[pyserial-dep]",
     "onnxruntime>=1.16.0,<2.0.0",
     "onnx>=1.16.0,<2.0.0",
     "meshcat>=0.3.0,<0.4.0",
@@ -189,14 +205,16 @@ 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]"]
-hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-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]"]
 
 # Features
 async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
 peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
 
 # Development
-dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1", "ruff>=0.14.1"]
+dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1", "ruff>=0.14.1", "lerobot[notebook]"]
+notebook = ["jupyter>=1.0.0,<2.0.0", "ipykernel>=6.0.0,<7.0.0"]
 test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
@@ -206,6 +224,20 @@ 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 = [
@@ -269,8 +301,23 @@ lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
 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"]
 

scripts/ci/extract_task_descriptions.py

@@ -31,9 +31,23 @@ 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]
@@ -47,7 +61,10 @@ def _libero_descriptions(task_suite: str) -> dict[str, str]:
         )
         return {}
     suite = suite_dict[task_suite]()
-    return {f"{task_suite}_{i}": suite.get_task(i).language for i in range(suite.n_tasks)}
+    return {
+        f"{task_suite}_{i}": _strip_libero_perturbation_tail(suite.get_task(i).language)
+        for i in range(suite.n_tasks)
+    }
 
 
 def _metaworld_descriptions(task_name: str) -> dict[str, str]:
@@ -57,19 +74,120 @@ 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":
+        if args.env == ("libero", "libero_plus"):
             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}'.",

src/lerobot/cameras/opencv/camera_opencv.py

@@ -430,7 +430,7 @@ class OpenCVCamera(Camera):
         Internal loop run by the background thread for asynchronous reading.
 
         On each iteration:
-        1. Reads a color frame
+        1. Reads a color frame (blocking call)
         2. Stores result in latest_frame and updates timestamp (thread-safe)
         3. Sets new_frame_event to notify listeners
 
@@ -439,8 +439,9 @@ class OpenCVCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
+        stop_event = self.stop_event
         failure_count = 0
-        while not self.stop_event.is_set():
+        while not stop_event.is_set():
             try:
                 raw_frame = self._read_from_hardware()
                 processed_frame = self._postprocess_image(raw_frame)
@@ -478,6 +479,8 @@ class OpenCVCamera(Camera):
 
         if self.thread is not None and self.thread.is_alive():
             self.thread.join(timeout=2.0)
+            if self.thread.is_alive():
+                logger.warning(f"{self} read thread did not terminate within timeout.")
 
         self.thread = None
         self.stop_event = None

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -33,7 +33,7 @@ import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
 
 from lerobot.utils.decorators import check_if_not_connected
-from lerobot.utils.import_utils import _reachy2_sdk_available
+from lerobot.utils.import_utils import _reachy2_sdk_available, require_package
 
 if TYPE_CHECKING or _reachy2_sdk_available:
     from reachy2_sdk.media.camera import CameraView
@@ -76,6 +76,7 @@ class Reachy2Camera(Camera):
         Args:
             config: The configuration settings for the camera.
         """
+        require_package("reachy2_sdk", extra="reachy2")
         super().__init__(config)
 
         self.config = config

src/lerobot/cameras/realsense/camera_realsense.py

@@ -17,18 +17,21 @@ 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 Any
+from typing import TYPE_CHECKING, Any
 
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
 from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
-try:
-    import pyrealsense2 as rs  # type: ignore  # TODO: add type stubs for pyrealsense2
-except Exception as e:
-    logging.info(f"Could not import realsense: {e}")
+from lerobot.utils.import_utils import _pyrealsense2_available, require_package
+
+if TYPE_CHECKING or _pyrealsense2_available:
+    import pyrealsense2 as rs
+else:
+    rs = None
 
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.errors import DeviceNotConnectedError
@@ -39,6 +42,7 @@ 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):
@@ -112,7 +116,7 @@ class RealSenseCamera(Camera):
         Args:
             config: The configuration settings for the camera.
         """
-
+        require_package(pkg_name, extra="intelrealsense", import_name="pyrealsense2")
         super().__init__(config)
 
         self.config = config
@@ -328,8 +332,8 @@ class RealSenseCamera(Camera):
         from the camera hardware via the RealSense pipeline.
 
         Returns:
-            np.ndarray: The depth map as a NumPy array (height, width)
-                  of type `np.uint16` (raw depth values in millimeters) and rotation.
+            np.ndarray: The depth map as a NumPy array (height, width, 1)
+                  of type `np.uint16` (raw depth values in millimeters).
 
         Raises:
             DeviceNotConnectedError: If the camera is not connected.
@@ -461,8 +465,8 @@ class RealSenseCamera(Camera):
         Internal loop run by the background thread for asynchronous reading.
 
         On each iteration:
-        1. Reads a color frame with 500ms timeout
-        2. Stores result in latest_frame and updates timestamp (thread-safe)
+        1. Reads a color/depth frame (blocking call with 10s timeout)
+        2. Stores result in latest_color_frame/latest_depth_frame and updates timestamp (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
@@ -470,8 +474,9 @@ class RealSenseCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
+        stop_event = self.stop_event
         failure_count = 0
-        while not self.stop_event.is_set():
+        while not stop_event.is_set():
             try:
                 frame = self._read_from_hardware()
                 color_frame_raw = frame.get_color_frame()
@@ -482,6 +487,8 @@ class RealSenseCamera(Camera):
                     depth_frame_raw = frame.get_depth_frame()
                     depth_frame = np.asanyarray(depth_frame_raw.get_data())
                     processed_depth_frame = self._postprocess_image(depth_frame, depth_frame=True)
+                    if processed_depth_frame.ndim == 2:  # (H, W) -> (H, W, 1)
+                        processed_depth_frame = processed_depth_frame[..., np.newaxis]
 
                 capture_time = time.perf_counter()
 
@@ -518,6 +525,8 @@ class RealSenseCamera(Camera):
 
         if self.thread is not None and self.thread.is_alive():
             self.thread.join(timeout=2.0)
+            if self.thread.is_alive():  # pragma: no cover
+                logger.warning(f"{self} read thread did not terminate within timeout.")
 
         self.thread = None
         self.stop_event = None
@@ -528,7 +537,6 @@ class RealSenseCamera(Camera):
             self.latest_timestamp = None
             self.new_frame_event.clear()
 
-    # NOTE(Steven): Missing implementation for depth for now
     @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
@@ -571,7 +579,6 @@ class RealSenseCamera(Camera):
 
         return frame
 
-    # NOTE(Steven): Missing implementation for depth for now
     @check_if_not_connected
     def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent (color) frame captured immediately (Peeking).
@@ -607,6 +614,71 @@ class RealSenseCamera(Camera):
 
         return frame
 
+    @check_if_not_connected
+    def async_read_depth(self, timeout_ms: float = 200) -> NDArray[Any]:
+        """Read the latest depth frame asynchronously, in metric meters.
+
+        Mirrors :meth:`async_read` but returns the depth stream rather than the
+        color stream. Output is ``np.uint16`` of shape ``(H, W, 1)``.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If ``use_depth`` is ``False`` for this camera, or if
+                the background read thread is not running.
+            TimeoutError: If no frame becomes available within ``timeout_ms``.
+        """
+        if not self.use_depth:
+            raise RuntimeError(f"{self}: cannot read depth — camera was configured with use_depth=False.")
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
+            raise TimeoutError(f"Timed out waiting for depth frame from camera {self} after {timeout_ms} ms.")
+
+        with self.frame_lock:
+            depth_frame = self.latest_depth_frame
+            self.new_frame_event.clear()
+
+        if depth_frame is None:
+            raise RuntimeError(f"Internal error: Event set but no depth frame available for {self}.")
+
+        return depth_frame
+
+    @check_if_not_connected
+    def read_latest_depth(self, max_age_ms: int = 500) -> NDArray[Any]:
+        """Return the most recent depth frame in metric meters (peeking).
+
+        Non-blocking counterpart of :meth:`read_latest` for the depth stream.
+        Output is ``np.uint16`` of shape ``(H, W, 1)`` in millimeters.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If ``use_depth`` is ``False`` for this camera, or if
+                no depth frame has been captured yet.
+            TimeoutError: If the latest depth frame is older than ``max_age_ms``.
+        """
+        if not self.use_depth:
+            raise RuntimeError(f"{self}: cannot read depth — camera was configured with use_depth=False.")
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        with self.frame_lock:
+            depth_frame = self.latest_depth_frame
+            timestamp = self.latest_timestamp
+
+        if depth_frame is None or timestamp is None:
+            raise RuntimeError(f"{self} has not captured any depth frames yet.")
+
+        age_ms = (time.perf_counter() - timestamp) * 1e3
+        if age_ms > max_age_ms:
+            raise TimeoutError(
+                f"{self} latest depth frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
+            )
+
+        return depth_frame
+
     def disconnect(self) -> None:
         """
         Disconnects from the camera, stops the pipeline, and cleans up resources.

src/lerobot/cameras/zmq/camera_zmq.py

@@ -28,12 +28,19 @@ import json
 import logging
 import time
 from threading import Event, Lock, Thread
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 import cv2
 import numpy as np
 from numpy.typing import NDArray
 
+from lerobot.utils.import_utils import _zmq_available, require_package
+
+if TYPE_CHECKING or _zmq_available:
+    import zmq
+else:
+    zmq = None
+
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.errors import DeviceNotConnectedError
 
@@ -74,8 +81,8 @@ class ZMQCamera(Camera):
     """
 
     def __init__(self, config: ZMQCameraConfig):
+        require_package("pyzmq", extra="pyzmq-dep", import_name="zmq")
         super().__init__(config)
-        import zmq
 
         self.config = config
         self.server_address = config.server_address
@@ -117,8 +124,6 @@ class ZMQCamera(Camera):
         logger.info(f"Connecting to {self}...")
 
         try:
-            import zmq
-
             self.context = zmq.Context()
             self.socket = self.context.socket(zmq.SUB)
             self.socket.setsockopt_string(zmq.SUBSCRIBE, "")
@@ -180,11 +185,8 @@ class ZMQCamera(Camera):
 
         try:
             message = self.socket.recv_string()
-        except Exception as e:
-            # zmq is lazy-imported in connect(), so check by name to avoid a top-level import
-            if type(e).__name__ == "Again":
-                raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
-            raise
+        except zmq.Again as e:
+            raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
 
         # Decode JSON message
         data = json.loads(message)
@@ -247,8 +249,9 @@ class ZMQCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized.")
 
+        stop_event = self.stop_event
         failure_count = 0
-        while not self.stop_event.is_set():
+        while not stop_event.is_set():
             try:
                 frame = self._read_from_hardware()
                 capture_time = time.perf_counter()
@@ -290,6 +293,8 @@ class ZMQCamera(Camera):
 
         if self.thread is not None and self.thread.is_alive():
             self.thread.join(timeout=2.0)
+            if self.thread.is_alive():
+                logger.warning(f"{self} read thread did not terminate within timeout.")
 
         self.thread = None
         self.stop_event = None

src/lerobot/common/control_utils.py

@@ -28,6 +28,12 @@ import numpy as np
 import torch
 
 from lerobot.policies import PreTrainedPolicy, prepare_observation_for_inference
+from lerobot.utils.import_utils import _deepdiff_available, require_package
+
+if TYPE_CHECKING or _deepdiff_available:
+    from deepdiff import DeepDiff
+else:
+    DeepDiff = None
 
 if TYPE_CHECKING:
     from lerobot.datasets import LeRobotDataset
@@ -217,10 +223,7 @@ def sanity_check_dataset_robot_compatibility(
     Raises:
         ValueError: If any of the checked metadata fields do not match.
     """
-    from lerobot.utils.import_utils import require_package
-
-    require_package("deepdiff", extra="hardware")
-    from deepdiff import DeepDiff
+    require_package("deepdiff", extra="deepdiff-dep")
 
     from lerobot.utils.constants import DEFAULT_FEATURES
 

src/lerobot/common/train_utils.py

@@ -99,6 +99,7 @@ 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)

src/lerobot/common/wandb_utils.py

@@ -41,8 +41,12 @@ 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 = [
-        f"policy:{cfg.policy.type}",
+        trainable_tag,
         f"seed:{cfg.seed}",
     ]
     if cfg.dataset is not None:

src/lerobot/configs/__init__.py

@@ -21,6 +21,7 @@ are intentionally NOT re-exported here to avoid circular dependencies
 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 .types import (
@@ -30,6 +31,14 @@ from .types import (
     PolicyFeature,
     RTCAttentionSchedule,
 )
+from .video import (
+    VALID_VIDEO_CODECS,
+    VIDEO_ENCODER_INFO_KEYS,
+    DepthEncoderConfig,
+    VideoEncoderConfig,
+    camera_encoder_defaults,
+    depth_encoder_defaults,
+)
 
 __all__ = [
     # Types
@@ -39,9 +48,18 @@ __all__ = [
     "PolicyFeature",
     "RTCAttentionSchedule",
     # Config classes
+    "DatasetRecordConfig",
     "DatasetConfig",
     "EvalConfig",
     "PeftConfig",
     "PreTrainedConfig",
     "WandBConfig",
+    "VideoEncoderConfig",
+    "DepthEncoderConfig",
+    # Defaults
+    "camera_encoder_defaults",
+    "depth_encoder_defaults",
+    # Constants
+    "VALID_VIDEO_CODECS",
+    "VIDEO_ENCODER_INFO_KEYS",
 ]

src/lerobot/configs/dataset.py

@@ -0,0 +1,83 @@
+# 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.
+
+"""Shared dataset recording configuration used by both ``lerobot-record`` and ``lerobot-rollout``."""
+
+from dataclasses import dataclass, field
+from datetime import datetime
+from pathlib import Path
+
+from .video import DepthEncoderConfig, VideoEncoderConfig, camera_encoder_defaults, depth_encoder_defaults
+
+
+@dataclass
+class DatasetRecordConfig:
+    # Dataset identifier. By convention it should match '{hf_username}/{dataset_name}' (e.g. `lerobot/test`).
+    repo_id: str = ""
+    # A short but accurate description of the task performed during the recording (e.g. "Pick the Lego block and drop it in the box on the right.")
+    single_task: str = ""
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    root: str | Path | None = None
+    # Limit the frames per second.
+    fps: int = 30
+    # Number of seconds for data recording for each episode.
+    episode_time_s: int | float = 60
+    # Number of seconds for resetting the environment after each episode.
+    reset_time_s: int | float = 60
+    # Number of episodes to record.
+    num_episodes: int = 50
+    # Encode frames in the dataset into video
+    video: bool = True
+    # Upload dataset to Hugging Face hub.
+    push_to_hub: bool = True
+    # Upload on private repository on the Hugging Face hub.
+    private: bool = False
+    # Add tags to your dataset on the hub.
+    tags: list[str] | None = None
+    # Number of subprocesses handling the saving of frames as PNG. Set to 0 to use threads only;
+    # set to ≥1 to use subprocesses, each using threads to write images. The best number of processes
+    # and threads depends on your system. We recommend 4 threads per camera with 0 processes.
+    # If fps is unstable, adjust the thread count. If still unstable, try using 1 or more subprocesses.
+    num_image_writer_processes: int = 0
+    # Number of threads writing the frames as png images on disk, per camera.
+    # Too many threads might cause unstable teleoperation fps due to main thread being blocked.
+    # Not enough threads might cause low camera fps.
+    num_image_writer_threads_per_camera: int = 4
+    # 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 encoder settings for depth-map MP4s (codec, quality, GOP, etc.). Tuned via CLI nested keys.
+    depth_encoder: DepthEncoderConfig = field(default_factory=depth_encoder_defaults)
+    # 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
+    # Maximum number of frames to buffer per camera when using streaming encoding.
+    # ~1s buffer at 30fps. Provides backpressure if the encoder can't keep up.
+    encoder_queue_maxsize: int = 30
+    # 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
+
+    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.
+        """
+        if self.repo_id:
+            timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+            self.repo_id = f"{self.repo_id}_{timestamp}"

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_codec
+from lerobot.utils.import_utils import get_safe_default_video_backend
 
 
 @dataclass
@@ -34,7 +34,10 @@ 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_codec)
+    video_backend: str = field(default_factory=get_safe_default_video_backend)
+    # 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
     streaming: bool = False
 
     def __post_init__(self) -> None:
@@ -114,3 +117,9 @@ 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

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
 
@@ -46,8 +46,11 @@ 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:
-            cli_overrides = parser.get_cli_overrides("policy")
-            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            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
+            )
             self.policy.pretrained_path = Path(policy_path)
 
         else:

src/lerobot/configs/parser.py

@@ -13,8 +13,10 @@
 # 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
@@ -24,6 +26,7 @@ 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
 
@@ -32,6 +35,29 @@ 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.
@@ -145,7 +171,14 @@ def load_plugin(plugin_path: str) -> None:
 
 
 def get_path_arg(field_name: str, args: Sequence[str] | None = None) -> str | None:
-    return parse_arg(f"{field_name}.{PATH_KEY}", args)
+    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, [])
 
 
 def get_type_arg(field_name: str, args: Sequence[str] | None = None) -> str | None:
@@ -192,6 +225,52 @@ 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:
@@ -225,6 +304,9 @@ 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)

src/lerobot/configs/rewards.py

@@ -0,0 +1,164 @@
+# 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)

src/lerobot/configs/train.py

@@ -13,7 +13,9 @@
 # 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
@@ -23,21 +25,60 @@ 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
@@ -56,6 +97,8 @@ class TrainPipelineConfig(HubMixin):
     # Number of workers for the dataloader.
     num_workers: int = 4
     batch_size: int = 8
+    prefetch_factor: int = 4
+    persistent_workers: bool = True
     steps: int = 100_000
     eval_freq: int = 20_000
     log_freq: int = 200
@@ -70,27 +113,44 @@ class TrainPipelineConfig(HubMixin):
     wandb: WandBConfig = field(default_factory=WandBConfig)
     peft: PeftConfig | 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"
+    # Sample weighting configuration (e.g., for RA-BC training)
+    sample_weighting: SampleWeightingConfig | None = None
 
     # 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")
-        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)
+        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
+            )
             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(
@@ -106,18 +166,22 @@ 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:
+        if self.policy is None and self.reward_model is None:
             raise ValueError(
-                "Policy is not configured. Please specify a pretrained policy with `--policy.path`."
+                "Neither policy nor reward_model is configured. "
+                "Please specify one with `--policy.path` or `--reward_model.path`."
             )
 
+        active_cfg = self.trainable_config
         if not self.job_name:
             if self.env is None:
-                self.job_name = f"{self.policy.type}"
+                self.job_name = f"{active_cfg.type}"
             else:
-                self.job_name = f"{self.env.type}_{self.policy.type}"
+                self.job_name = f"{self.env.type}_{active_cfg.type}"
 
         if not self.resume and isinstance(self.output_dir, Path) and self.output_dir.is_dir():
             raise FileExistsError(
@@ -135,26 +199,16 @@ 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 = self.policy.get_optimizer_preset()
-            self.scheduler = self.policy.get_scheduler_preset()
+            self.optimizer = active_cfg.get_optimizer_preset()
+            self.scheduler = active_cfg.get_scheduler_preset()
 
-        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"
+        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.")
 
     @classmethod
     def __get_path_fields__(cls) -> list[str]:
-        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
-        return ["policy"]
+        """Keys for draccus pretrained-path loading."""
+        return ["policy", "reward_model"]
 
     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
@@ -205,12 +259,16 @@ 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

src/lerobot/configs/video.py

@@ -0,0 +1,315 @@
+# 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, fields
+from typing import Any, ClassVar
+
+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", "ffv1", "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
+)
+
+# Default depth quantization and encoding parameters.
+DEPTH_QUANT_BITS: int = 12
+DEPTH_QMAX: int = (1 << DEPTH_QUANT_BITS) - 1  # 4095
+
+DEFAULT_DEPTH_MIN: float = 0.01
+DEFAULT_DEPTH_MAX: float = 10.0
+DEFAULT_DEPTH_SHIFT: float = 3.5
+DEFAULT_DEPTH_USE_LOG: bool = True
+DEFAULT_DEPTH_PIX_FMT: str = "gray12le"
+
+# Depth-specific tuning fields persisted under ``features[*]["info"]`` as ``video.<name>``.
+DEPTH_ENCODER_INFO_FIELD_NAMES: frozenset[str] = frozenset({"depth_min", "depth_max", "shift", "use_log"})
+
+
+@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)
+
+    # Source-data channel count this encoder is expected to handle (3 for RGB,
+    # 1 for depth, etc.)
+    _DEFAULT_CHANNELS: ClassVar[int] = 3
+
+    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(), channels=self._DEFAULT_CHANNELS
+            )
+
+    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)
+        elif self.vcodec == "ffv1":
+            # Lossless intra-frame codec. ``crf``/``preset``/``fast_decode``
+            # are not meaningful.
+            set_if("threads", encoder_threads)
+        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()
+
+
+@dataclass
+class DepthEncoderConfig(VideoEncoderConfig):
+    """Encoder configuration for depth-map streams.
+
+    Inherits the full :class:`VideoEncoderConfig` surface (codec, GOP, CRF,
+    preset, ``extra_options``…) and adds the four parameters of the depth
+    quantizer.
+
+    Defaults flip ``vcodec`` to ``"hevc"`` (Main 12 profile) and ``pix_fmt``
+    to ``"gray12le"``.
+
+
+    Attributes:
+        depth_min: Minimum depth in physical units (e.g. metres) represented
+            by quantum ``0``.
+        depth_max: Maximum depth represented by quantum :data:`DEPTH_QMAX`.
+        shift: Pre-log offset for numerical stability near zero.
+        use_log: ``True`` for logarithmic quantization (default; matches
+            sensor error profile), ``False`` for linear.
+    """
+
+    vcodec: str = "hevc"
+    pix_fmt: str = "gray12le"
+
+    depth_min: float = DEFAULT_DEPTH_MIN
+    depth_max: float = DEFAULT_DEPTH_MAX
+    shift: float = DEFAULT_DEPTH_SHIFT
+    use_log: bool = DEFAULT_DEPTH_USE_LOG
+
+    _DEFAULT_CHANNELS: ClassVar[int] = 1
+
+    @classmethod
+    def from_video_info(cls, video_info: dict | None) -> DepthEncoderConfig:
+        """Reconstruct a :class:`DepthEncoderConfig` from a depth feature's ``info`` block.
+
+        Reuses :meth:`VideoEncoderConfig.from_video_info` for the base
+        codec/tuning fields and then layers the depth-specific tuning
+        (``depth_min`` / ``depth_max`` / ``shift`` / ``use_log``) on top.
+        Missing keys fall back to the class defaults.
+        """
+        base = VideoEncoderConfig.from_video_info(video_info)
+        kwargs: dict[str, Any] = {f.name: getattr(base, f.name) for f in fields(base) if f.init}
+
+        video_info = video_info or {}
+        for name in DEPTH_ENCODER_INFO_FIELD_NAMES:
+            value = video_info.get(f"video.{name}")
+            if value is not None:
+                kwargs[name] = value
+        return cls(**kwargs)
+
+
+def depth_encoder_defaults() -> DepthEncoderConfig:
+    """Return a :class:`DepthEncoderConfig` with depth-camera defaults."""
+    return DepthEncoderConfig()

src/lerobot/datasets/__init__.py

@@ -40,6 +40,7 @@ from .io_utils import load_episodes, write_stats
 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
@@ -59,6 +60,8 @@ __all__ = [
     "MultiLeRobotDataset",
     "StreamingLeRobotDataset",
     "VideoEncodingManager",
+    "check_video_encoder_parameters_pyav",
+    "detect_available_encoders_pyav",
     "add_features",
     "aggregate_datasets",
     "aggregate_pipeline_dataset_features",

src/lerobot/datasets/aggregate.py

@@ -15,6 +15,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import copy
 import logging
 import shutil
 from pathlib import Path
@@ -23,9 +24,11 @@ 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 get_hf_features_from_features
+from .feature_utils import features_equal_for_merge, get_hf_features_from_features
 from .io_utils import (
     get_file_size_in_mb,
     get_parquet_file_size_in_mb,
@@ -46,11 +49,54 @@ 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.
@@ -74,7 +120,7 @@ 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 features != meta.features:
+        if not features_equal_for_merge(features, meta.features):
             raise ValueError(
                 f"Same features is expected, but got features={meta.features} instead of {features}."
             )
@@ -97,8 +143,8 @@ 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()
@@ -225,9 +271,9 @@ 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
 
@@ -237,8 +283,8 @@ def aggregate_datasets(
     aggr_repo_id: str,
     roots: list[Path] | None = None,
     aggr_root: Path | None = None,
-    data_files_size_in_mb: float | None = None,
-    video_files_size_in_mb: float | None = None,
+    data_files_size_in_mb: int | None = None,
+    video_files_size_in_mb: int | None = None,
     chunk_size: int | None = None,
 ):
     """Aggregates multiple LeRobot datasets into a single unified dataset.
@@ -274,7 +320,8 @@ def aggregate_datasets(
             LeRobotDatasetMetadata(repo_id, root=root) for repo_id, root in zip(repo_ids, roots, strict=False)
         ]
     )
-    fps, robot_type, features = validate_all_metadata(all_metadata)
+    fps, robot_type, _ = validate_all_metadata(all_metadata)
+    features = merge_video_feature_info_for_aggregate(all_metadata)
     video_keys = [key for key in features if features[key]["dtype"] == "video"]
 
     dst_meta = LeRobotDatasetMetadata.create(
@@ -313,8 +360,8 @@ 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.")
@@ -332,7 +379,6 @@ 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.
     """
@@ -414,9 +460,11 @@ 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
@@ -640,14 +688,10 @@ def finalize_aggregation(aggr_meta, all_metadata):
     write_tasks(aggr_meta.tasks, aggr_meta.root)
 
     logging.info("write info")
-    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)}"},
-        }
-    )
+    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)}"}
     write_info(aggr_meta.info, aggr_meta.root)
 
     logging.info("write stats")

src/lerobot/datasets/compute_stats.py

@@ -550,8 +550,10 @@ def _validate_stat_value(value: np.ndarray, key: str, feature_key: str) -> None:
     if key == "count" and value.shape != (1,):
         raise ValueError(f"Shape of 'count' must be (1), but is {value.shape} instead.")
 
-    if "image" in feature_key and key != "count" and value.shape != (3, 1, 1):
-        raise ValueError(f"Shape of quantile '{key}' must be (3,1,1), but is {value.shape} instead.")
+    if "image" in feature_key and key != "count" and value.shape not in ((3, 1, 1), (1, 1, 1)):
+        raise ValueError(
+            f"Shape of quantile '{key}' must be (3,1,1) or (1,1,1) but is {value.shape} instead."
+        )
 
 
 def _assert_type_and_shape(stats_list: list[dict[str, dict]]):

src/lerobot/datasets/dataset_metadata.py

@@ -14,6 +14,7 @@
 # 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
@@ -23,6 +24,7 @@ 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
@@ -37,13 +39,11 @@ from .io_utils import (
     load_subtasks,
     load_tasks,
     write_info,
-    write_json,
     write_stats,
     write_tasks,
 )
 from .utils import (
     DEFAULT_EPISODES_PATH,
-    INFO_PATH,
     check_version_compatibility,
     get_safe_version,
     has_legacy_hub_download_metadata,
@@ -191,6 +191,29 @@ 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,
@@ -228,7 +251,7 @@ 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.
@@ -283,27 +306,27 @@ 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]:
@@ -315,6 +338,25 @@ class LeRobotDatasetMetadata:
         """Keys to access visual modalities stored as videos."""
         return [key for key, ft in self.features.items() if ft["dtype"] == "video"]
 
+    @property
+    def depth_keys(self) -> list[str]:
+        """Keys to access depth-map modalities stored as videos or images.
+
+        A depth key is a feature whose ``info`` dict carries ``"is_depth_map": True``
+        (or the legacy ``"video.is_depth_map"`` inside ``info`` or ``video_info``).
+        """
+
+        def _is_depth(ft: dict) -> bool:
+            info = ft.get("info") or {}
+            video_info = ft.get("video_info") or {}
+            return (
+                info.get("is_depth_map", False)
+                or info.get("video.is_depth_map", False)
+                or video_info.get("video.is_depth_map", False)
+            )
+
+        return [key for key, ft in self.features.items() if _is_depth(ft)]
+
     @property
     def camera_keys(self) -> list[str]:
         """Keys to access visual modalities (regardless of their storage method)."""
@@ -333,32 +375,32 @@ 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:
         """
@@ -502,29 +544,46 @@ 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) -> None:
-        """
+    def update_video_info(
+        self,
+        video_key: str | None = None,
+        video_encoder: VideoEncoderConfig | None = None,
+    ) -> None:
+        """Populate per-feature video info in ``info.json``.
+
         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")
 
         video_keys = [video_key] if video_key is not None else self.video_keys
         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)
+            existing = self.features[key].get("info") or {}
+            # Skip only if real video info has already been written. The ``is_depth_map`` entry (created at feature creation) is not blocking.
+            if set(existing.keys()) - {"is_depth_map"}:
+                continue
+            video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
+            new_info = get_video_info(video_path, video_encoder=video_encoder)
+            self.info.features[key]["info"] = {**existing, **new_info}
 
     def update_chunk_settings(
         self,
@@ -546,17 +605,17 @@ 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)
@@ -653,7 +712,7 @@ 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_json(obj.info, obj.root / INFO_PATH)
+        write_info(obj.info, obj.root)
         obj.revision = None
         obj._pq_writer = None
         obj.latest_episode = None

src/lerobot/datasets/dataset_reader.py

@@ -16,12 +16,16 @@
 """Private reader component for LeRobotDataset. Handles random-access reading (HF dataset, delta indices, video decoding)."""
 
 from collections.abc import Callable
+from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
 
 import datasets
 import torch
 
+from lerobot.configs.video import DepthEncoderConfig
+
 from .dataset_metadata import LeRobotDatasetMetadata
+from .depth_utils import dequantize_depth
 from .feature_utils import (
     check_delta_timestamps,
     get_delta_indices,
@@ -49,6 +53,7 @@ class DatasetReader:
         video_backend: str,
         delta_timestamps: dict[str, list[float]] | None,
         image_transforms: Callable | None,
+        return_uint8: bool = False,
     ):
         """Initialize the reader with metadata, filtering, and transform config.
 
@@ -73,6 +78,7 @@ class DatasetReader:
         self._tolerance_s = tolerance_s
         self._video_backend = video_backend
         self._image_transforms = image_transforms
+        self._return_uint8 = return_uint8
 
         self.hf_dataset: datasets.Dataset | None = None
         self._absolute_to_relative_idx: dict[int, int] | None = None
@@ -83,6 +89,12 @@ class DatasetReader:
             check_delta_timestamps(delta_timestamps, meta.fps, tolerance_s)
             self.delta_indices = get_delta_indices(delta_timestamps, meta.fps)
 
+        ##TODO(CarolinePascal): Should we rather use a more lightweight structure ?
+        self._depth_encoder_configs: dict[str, DepthEncoderConfig] = {
+            vid_key: DepthEncoderConfig.from_video_info(self._meta.features[vid_key].get("info"))
+            for vid_key in self._meta.depth_keys
+        }
+
     def try_load(self) -> bool:
         """Attempt to load from local cache. Returns True if data is sufficient."""
         try:
@@ -105,10 +117,8 @@ class DatasetReader:
         """Build absolute-to-relative index mapping from loaded hf_dataset."""
         self._absolute_to_relative_idx = None
         if self.episodes is not None and self.hf_dataset is not None:
-            self._absolute_to_relative_idx = {
-                abs_idx.item() if isinstance(abs_idx, torch.Tensor) else abs_idx: rel_idx
-                for rel_idx, abs_idx in enumerate(self.hf_dataset["index"])
-            }
+            indices = self.hf_dataset.data.column("index").to_numpy()
+            self._absolute_to_relative_idx = dict(zip(indices.tolist(), range(len(indices)), strict=True))
 
     @property
     def num_frames(self) -> int:
@@ -235,16 +245,41 @@ class DatasetReader:
         Segmentation Fault.
         """
         ep = self._meta.episodes[ep_idx]
-        item = {}
-        for vid_key, query_ts in query_timestamps.items():
+
+        def _decode_single(vid_key: str, query_ts: list[float]) -> tuple[str, torch.Tensor]:
             from_timestamp = ep[f"videos/{vid_key}/from_timestamp"]
             shifted_query_ts = [from_timestamp + ts for ts in query_ts]
-
             video_path = self.root / self._meta.get_video_file_path(ep_idx, vid_key)
-            frames = decode_video_frames(video_path, shifted_query_ts, self._tolerance_s, self._video_backend)
-            item[vid_key] = frames.squeeze(0)
+            frames = decode_video_frames(
+                video_path,
+                shifted_query_ts,
+                self._tolerance_s,
+                self._video_backend,
+                return_uint8=self._return_uint8,
+                is_depth=vid_key in self._meta.depth_keys,
+            )
+            if vid_key in self._meta.depth_keys:
+                depth_encoder = self._depth_encoder_configs[vid_key]
+                frames = dequantize_depth(
+                    frames,
+                    depth_min=depth_encoder.depth_min,
+                    depth_max=depth_encoder.depth_max,
+                    shift=depth_encoder.shift,
+                    use_log=depth_encoder.use_log,
+                    output_tensor=True,
+                )
+            return vid_key, frames.squeeze(0)
 
-        return item
+        items = list(query_timestamps.items())
+
+        # Single camera: no threading overhead
+        if len(items) <= 1:
+            return {vid_key: _decode_single(vid_key, query_ts)[1] for vid_key, query_ts in items}
+
+        # Multi-camera: decode in parallel (video decoding releases the GIL)
+        with ThreadPoolExecutor(max_workers=len(items)) as pool:
+            futures = [pool.submit(_decode_single, k, ts) for k, ts in items]
+            return dict(f.result() for f in futures)
 
     def get_item(self, idx) -> dict:
         """Core __getitem__ logic. Assumes hf_dataset is loaded.

src/lerobot/datasets/dataset_tools.py

@@ -36,6 +36,7 @@ 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
 
@@ -62,7 +63,10 @@ from .utils import (
     DEFAULT_EPISODES_PATH,
     update_chunk_file_indices,
 )
-from .video_utils import encode_video_frames, get_video_info
+from .video_utils import (
+    encode_video_frames,
+    get_video_info,
+)
 
 
 def _load_episode_with_stats(src_dataset: LeRobotDataset, episode_idx: int) -> dict:
@@ -95,6 +99,11 @@ 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.
@@ -157,6 +166,11 @@ 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
@@ -578,8 +592,7 @@ def _keep_episodes_from_video_with_av(
     output_path: Path,
     episodes_to_keep: list[tuple[int, int]],
     fps: float,
-    vcodec: str = "libsvtav1",
-    pix_fmt: str = "yuv420p",
+    camera_encoder: VideoEncoderConfig,
 ) -> None:
     """Keep only specified episodes from a video file using PyAV.
 
@@ -593,8 +606,7 @@ 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.
-        vcodec: Video codec to use for encoding.
-        pix_fmt: Pixel format for output video.
+        camera_encoder: Video encoder settings used to re-encode the kept frames.
     """
     from fractions import Fraction
 
@@ -619,12 +631,13 @@ def _keep_episodes_from_video_with_av(
 
     # Convert fps to Fraction for PyAV compatibility.
     fps_fraction = Fraction(fps).limit_denominator(1000)
-    v_out = out.add_stream(vcodec, rate=fps_fraction)
+    codec_options = camera_encoder.get_codec_options(as_strings=True)
+    v_out = out.add_stream(camera_encoder.vcodec, rate=fps_fraction, options=codec_options)
 
     # 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 = pix_fmt
+    v_out.pix_fmt = camera_encoder.pix_fmt
 
     # Set time_base to match the frame rate for proper timestamp handling.
     v_out.time_base = Fraction(1, int(fps))
@@ -687,14 +700,14 @@ 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.
+    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.
 
     Args:
         src_dataset: Source dataset to copy from
@@ -711,6 +724,9 @@ 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")
@@ -792,8 +808,7 @@ def _copy_and_reindex_videos(
                     dst_video_path,
                     episodes_to_keep_ranges,
                     src_dataset.meta.fps,
-                    vcodec,
-                    pix_fmt,
+                    camera_encoder,
                 )
 
                 cumulative_ts = 0.0
@@ -897,14 +912,10 @@ def _copy_and_reindex_episodes_metadata(
 
     dst_meta.finalize()
 
-    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)}"},
-        }
-    )
+    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)}"}
     write_info(dst_meta.info, dst_meta.root)
 
     if not all_stats:
@@ -1069,21 +1080,20 @@ def _copy_episodes_metadata_and_stats(
     if episodes_dir.exists():
         shutil.copytree(episodes_dir, dst_episodes_dir, dirs_exist_ok=True)
 
-    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}"}),
-        }
+    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}"}
     )
 
     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)
 
@@ -1269,11 +1279,7 @@ def _estimate_frame_size_via_calibration(
     episode_indices: list[int],
     temp_dir: Path,
     fps: int,
-    vcodec: str,
-    pix_fmt: str,
-    g: int,
-    crf: int,
-    fast_decode: int,
+    camera_encoder: VideoEncoderConfig,
     num_calibration_frames: int = 30,
 ) -> float:
     """Estimate MB per frame by encoding a small calibration sample.
@@ -1287,11 +1293,7 @@ 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.
-        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.
+        camera_encoder: Video encoder settings used for calibration encoding.
         num_calibration_frames: Number of frames to use for calibration (default: 30).
 
     Returns:
@@ -1327,11 +1329,7 @@ def _estimate_frame_size_via_calibration(
             imgs_dir=calibration_dir,
             video_path=calibration_video_path,
             fps=fps,
-            vcodec=vcodec,
-            pix_fmt=pix_fmt,
-            g=g,
-            crf=crf,
-            fast_decode=fast_decode,
+            video_encoder=camera_encoder,
             overwrite=True,
         )
 
@@ -1525,7 +1523,7 @@ 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
@@ -1649,11 +1647,7 @@ def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
     output_dir: Path | None = None,
     repo_id: str | None = None,
-    vcodec: str = "libsvtav1",
-    pix_fmt: str = "yuv420p",
-    g: int = 2,
-    crf: int = 30,
-    fast_decode: int = 0,
+    camera_encoder: VideoEncoderConfig | None = None,
     episode_indices: list[int] | None = None,
     num_workers: int = 4,
     max_episodes_per_batch: int | None = None,
@@ -1668,11 +1662,8 @@ 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.
-        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)
+        camera_encoder: Video encoder settings
+            (``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`).
         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)
@@ -1681,6 +1672,9 @@ 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(
@@ -1704,7 +1698,10 @@ 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: {vcodec}, pixel format: {pix_fmt}, GOP: {g}, CRF: {crf}")
+    logging.info(
+        f"Video codec: {camera_encoder.vcodec}, pixel format: {camera_encoder.pix_fmt}, "
+        f"GOP: {camera_encoder.g}, CRF: {camera_encoder.crf}"
+    )
 
     # Create new features dict, converting image features to video features
     new_features = {}
@@ -1774,11 +1771,7 @@ def convert_image_to_video_dataset(
                 episode_indices=episode_indices,
                 temp_dir=temp_dir,
                 fps=fps,
-                vcodec=vcodec,
-                pix_fmt=pix_fmt,
-                g=g,
-                crf=crf,
-                fast_decode=fast_decode,
+                camera_encoder=camera_encoder,
             )
 
             logging.info(f"Processing camera: {img_key}")
@@ -1820,11 +1813,7 @@ def convert_image_to_video_dataset(
                     imgs_dir=imgs_dir,
                     video_path=video_path,
                     fps=fps,
-                    vcodec=vcodec,
-                    pix_fmt=pix_fmt,
-                    g=g,
-                    crf=crf,
-                    fast_decode=fast_decode,
+                    video_encoder=camera_encoder,
                     overwrite=True,
                 )
 
@@ -1858,10 +1847,10 @@ 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
@@ -1870,7 +1859,9 @@ 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)
+                new_meta.info.features[img_key]["info"] = get_video_info(
+                    video_path, video_encoder=camera_encoder
+                )
 
         write_info(new_meta.info, new_meta.root)
 

src/lerobot/datasets/dataset_writer.py

@@ -31,6 +31,13 @@ import PIL.Image
 import pyarrow.parquet as pq
 import torch
 
+from lerobot.configs import (
+    DepthEncoderConfig,
+    VideoEncoderConfig,
+    camera_encoder_defaults,
+    depth_encoder_defaults,
+)
+
 from .compute_stats import compute_episode_stats
 from .dataset_metadata import LeRobotDatasetMetadata
 from .feature_utils import (
@@ -46,6 +53,7 @@ from .io_utils import (
     write_info,
 )
 from .utils import (
+    DEFAULT_DEPTH_PATH,
     DEFAULT_EPISODES_PATH,
     DEFAULT_IMAGE_PATH,
     update_chunk_file_indices,
@@ -65,14 +73,24 @@ def _encode_video_worker(
     episode_index: int,
     root: Path,
     fps: int,
-    vcodec: str = "libsvtav1",
+    video_encoder: VideoEncoderConfig | None = None,
     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)
+    path_template = (
+        DEFAULT_DEPTH_PATH
+        if video_encoder is not None and isinstance(video_encoder, DepthEncoderConfig)
+        else DEFAULT_IMAGE_PATH
+    )
+    fpath = path_template.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, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
+        img_dir,
+        temp_path,
+        fps,
+        video_encoder=video_encoder,
+        encoder_threads=encoder_threads,
+        overwrite=True,
     )
     shutil.rmtree(img_dir)
     return temp_path
@@ -89,20 +107,25 @@ class DatasetWriter:
         self,
         meta: LeRobotDatasetMetadata,
         root: Path,
-        vcodec: str,
+        camera_encoder: VideoEncoderConfig | None,
+        depth_encoder: DepthEncoderConfig | None,
         encoder_threads: int | None,
         batch_encoding_size: int,
         streaming_encoder: StreamingVideoEncoder | None = None,
         initial_frames: int = 0,
     ):
-        """Initialize the writer with metadata, codec, and encoding config.
+        """Initialize the writer with metadata, codec, and encoder config.
 
         Args:
             meta: Dataset metadata instance (used for feature schema, chunk
                 settings, and episode persistence).
             root: Local dataset root directory.
-            vcodec: Video codec for encoding (e.g. ``'libsvtav1'``, ``'h264'``).
-            encoder_threads: Threads per encoder instance. ``None`` for auto.
+            camera_encoder: Video encoder settings applied to all cameras.
+                ``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`.
+            depth_encoder: Video encoder settings applied to all **depth** cameras.
+                ``None`` uses :func:`~lerobot.configs.depth_encoder_defaults`.
+            encoder_threads: Number of encoder threads (global). ``None``
+                lets the codec decide.
             batch_encoding_size: Number of episodes to accumulate before
                 batch-encoding videos.
             streaming_encoder: Optional pre-built :class:`StreamingVideoEncoder`
@@ -111,7 +134,8 @@ class DatasetWriter:
         """
         self._meta = meta
         self._root = root
-        self._vcodec = vcodec
+        self._camera_encoder = camera_encoder or camera_encoder_defaults()
+        self._depth_encoder = depth_encoder or depth_encoder_defaults()
         self._encoder_threads = encoder_threads
         self._batch_encoding_size = batch_encoding_size
         self._streaming_encoder = streaming_encoder
@@ -136,7 +160,8 @@ class DatasetWriter:
         return ep_buffer
 
     def _get_image_file_path(self, episode_index: int, image_key: str, frame_index: int) -> Path:
-        fpath = DEFAULT_IMAGE_PATH.format(
+        path_template = DEFAULT_DEPTH_PATH if image_key in self._meta.depth_keys else DEFAULT_IMAGE_PATH
+        fpath = path_template.format(
             image_key=image_key, episode_index=episode_index, frame_index=frame_index
         )
         return self._root / fpath
@@ -186,6 +211,7 @@ class DatasetWriter:
         if frame_index == 0 and self._streaming_encoder is not None:
             self._streaming_encoder.start_episode(
                 video_keys=list(self._meta.video_keys),
+                depth_video_keys=set(self._meta.video_keys) & set(self._meta.depth_keys),
                 temp_dir=self._root,
             )
 
@@ -284,7 +310,9 @@ class DatasetWriter:
                             episode_index,
                             self._root,
                             self._meta.fps,
-                            self._vcodec,
+                            self._depth_encoder
+                            if video_key in self._meta.depth_keys
+                            else self._camera_encoder,
                             self._encoder_threads,
                         ): video_key
                         for video_key in self._meta.video_keys
@@ -495,7 +523,12 @@ class DatasetWriter:
 
         # Update video info (only needed when first episode is encoded)
         if episode_index == 0:
-            self._meta.update_video_info(video_key)
+            self._meta.update_video_info(
+                video_key,
+                video_encoder=self._depth_encoder
+                if video_key in self._meta.depth_keys
+                else self._camera_encoder,
+            )
             write_info(self._meta.info, self._meta.root)
 
         metadata = {
@@ -562,9 +595,15 @@ class DatasetWriter:
             self.image_writer.wait_until_done()
 
     def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path:
-        """Use ffmpeg to convert frames stored as png into mp4 videos."""
+        """Use ffmpeg to convert frames stored as png/tiff into mp4 videos."""
+        is_depth = video_key in self._meta.depth_keys
         return _encode_video_worker(
-            video_key, episode_index, self._root, self._meta.fps, self._vcodec, self._encoder_threads
+            video_key,
+            episode_index,
+            self._root,
+            self._meta.fps,
+            self._depth_encoder if is_depth else self._camera_encoder,
+            self._encoder_threads,
         )
 
     def close_writer(self) -> None:
@@ -597,7 +636,7 @@ class DatasetWriter:
 
     def cleanup_interrupted_episode(self, episode_index: int) -> None:
         """Remove temporary image directories for an interrupted episode."""
-        for key in self._meta.video_keys:
+        for key in self._meta.camera_keys:
             img_dir = self._get_image_file_path(
                 episode_index=episode_index, image_key=key, frame_index=0
             ).parent

src/lerobot/datasets/depth_utils.py

@@ -0,0 +1,214 @@
+#!/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.
+"""
+Depth encoding/decoding helpers for :class:`VideoEncoderConfig`.
+"""
+
+import math
+from typing import Literal
+
+import av
+import numpy as np
+import torch
+from numpy.typing import NDArray
+
+from lerobot.configs.video import (
+    DEFAULT_DEPTH_MAX,
+    DEFAULT_DEPTH_MIN,
+    DEFAULT_DEPTH_PIX_FMT,
+    DEFAULT_DEPTH_SHIFT,
+    DEFAULT_DEPTH_USE_LOG,
+    DEPTH_QMAX,
+)
+
+from .pyav_utils import write_u16_plane
+
+_MM_PER_METRE = 1000.0
+_UINT16_MAX = 65535
+
+
+def _validate_log_quant_params(depth_min: float, shift: float) -> None:
+    """Ensure ``log(depth_min + shift)`` is finite."""
+    if depth_min + shift <= 0:
+        raise ValueError(
+            f"depth_min + shift must be positive for logarithmic quantization, "
+            f"got depth_min={depth_min} + shift={shift} = {depth_min + shift}"
+        )
+
+
+def _depth_input_to_float32_and_unit(
+    depth: NDArray[np.integer] | NDArray[np.floating],
+    input_unit: Literal["auto", "m", "mm"],
+) -> tuple[NDArray[np.float32], Literal["m", "mm"]]:
+    """Convert depth to float32 in the chosen unit, and return the resolved unit."""
+    resolved_unit = (
+        ("m" if np.issubdtype(depth.dtype, np.floating) else "mm") if input_unit == "auto" else input_unit
+    )
+    return depth.astype(np.float32, order="K"), resolved_unit
+
+
+def quantize_depth(
+    depth: NDArray[np.uint16] | NDArray[np.float32] | torch.Tensor,
+    depth_min: float = DEFAULT_DEPTH_MIN,
+    depth_max: float = DEFAULT_DEPTH_MAX,
+    shift: float = DEFAULT_DEPTH_SHIFT,
+    use_log: bool = DEFAULT_DEPTH_USE_LOG,
+    pix_fmt: str = DEFAULT_DEPTH_PIX_FMT,
+    video_backend: str | None = "pyav",
+    input_unit: Literal["auto", "m", "mm"] = "auto",
+) -> NDArray[np.uint16] | av.VideoFrame:
+    """Quantize depth to 12-bit codes (``uint16``, values ``0…DEPTH_QMAX``).
+
+    Depth maps are packed into 12-bit integer frames so they fit in standard
+    high-bit-depth pixel formats (e.g. ``yuv420p12le`` / ``gray12le``)
+    and can be encoded by widely supported video codecs (HEVC Main 12, ffv1).
+    Logarithmic quantization is the default because it allocates more quanta
+    to near-range depth, which matches the (1/depth) error profile of typical
+    depth sensors. Math is ported from BEHAVIOR-1K's ``obs_utils.py``.
+
+    **Input units**:
+
+    - ``input_unit="auto"`` (default): infer from dtype (floating = m, non-floating = mm).
+    - ``input_unit="mm"``: interpret input values as millimetres.
+    - ``input_unit="m"``: interpret input values as metres.
+
+    Quantization math runs in the **resolved input unit**.
+
+    ``depth_min``, ``depth_max``, and ``shift`` are always in **metres**.
+
+    Args:
+        depth: Depth map; ``torch.Tensor`` is moved to CPU for conversion.
+        depth_min: Depth (metres) at quantum ``0``.
+        depth_max: Depth (metres) at quantum :data:`DEPTH_QMAX`.
+        shift: Depth shift (metres); used in log mode. Must satisfy ``depth_min + shift > 0``.
+        use_log: If ``True`` (default), quantize in log space.
+        video_backend: Video backend to use for encoding. Defaults to "pyav".
+        input_unit: Input unit policy (``"auto"``, ``"mm"``, ``"m"``).
+
+    Returns:
+        ``numpy.ndarray``, ``dtype=uint16``, same shape as ``depth``, values in
+        ``[0, DEPTH_QMAX]``.
+
+    Raises:
+        ValueError: If ``input_unit`` is not ``"auto"``, ``"mm"``, or ``"m"``.
+        ValueError: If ``use_log=True`` and ``depth_min + shift <= 0``.
+    """
+    if input_unit not in ("auto", "m", "mm"):
+        raise ValueError(f"input_unit must be 'auto', 'm', or 'mm', got {input_unit!r}")
+
+    if isinstance(depth, torch.Tensor):
+        depth = depth.detach().cpu().numpy()
+
+    # Squeeze single-channel dim: (H, W, 1) or (1, H, W) → (H, W)
+    if depth.ndim == 3 and (depth.shape[-1] == 1 or depth.shape[0] == 1):
+        depth = depth.squeeze()
+
+    depth_f, resolved_unit = _depth_input_to_float32_and_unit(depth, input_unit=input_unit)
+
+    # Convert depth_min, depth_max, and shift to the resolved input unit.
+    depth_min_u = np.float32(depth_min) if resolved_unit == "m" else np.float32(depth_min * _MM_PER_METRE)
+    depth_max_u = np.float32(depth_max) if resolved_unit == "m" else np.float32(depth_max * _MM_PER_METRE)
+    shift_u = np.float32(shift) if resolved_unit == "m" else np.float32(shift * _MM_PER_METRE)
+
+    # Normalization and quantization is performed in the resolved input unit.
+    if use_log:
+        _validate_log_quant_params(depth_min, shift)
+        log_min = math.log(float(depth_min_u + shift_u))
+        log_max = math.log(float(depth_max_u + shift_u))
+        norm = (np.log(depth_f + shift_u) - log_min) / (log_max - log_min)
+    else:
+        norm = (depth_f - depth_min_u) / (depth_max_u - depth_min_u)
+
+    quantized = np.rint(norm * DEPTH_QMAX).clip(0, DEPTH_QMAX).astype(np.uint16, copy=False)
+
+    if video_backend == "pyav":
+        frame = av.VideoFrame.from_ndarray(quantized, format=pix_fmt)
+        write_u16_plane(frame.planes[0], quantized)
+        return frame
+    else:
+        return quantized
+
+
+def dequantize_depth(
+    quantized: NDArray[np.uint16] | av.VideoFrame,
+    depth_min: float = DEFAULT_DEPTH_MIN,
+    depth_max: float = DEFAULT_DEPTH_MAX,
+    shift: float = DEFAULT_DEPTH_SHIFT,
+    use_log: bool = DEFAULT_DEPTH_USE_LOG,
+    pix_fmt: str = DEFAULT_DEPTH_PIX_FMT,
+    output_unit: Literal["m", "mm"] = "mm",
+    output_tensor: bool = False,
+) -> NDArray[np.uint16] | NDArray[np.float32] | torch.Tensor:
+    """Inverse of :func:`quantize_depth`.
+
+    Tuning arguments **must match** :func:`quantize_depth`.
+
+    Decoding inverts the same normalized code mapping as :func:`quantize_depth`
+    using ``depth_min`` / ``depth_max`` / ``shift`` (in metres), then returns
+    the requested output unit.
+
+    Args:
+        quantized: 12-bit codes ``[0, DEPTH_QMAX]``, ``dtype=uint16``.
+        depth_min, depth_max, shift, use_log: Same as :func:`quantize_depth` (metres).
+        output_unit: ``\"mm\"`` returns ``uint16`` millimetres (``rint``, clip
+            ``[0, 65535]``). ``\"m\"`` returns ``float32`` metres in
+            ``[depth_min, depth_max]``.
+        output_tensor: If True, return a torch.Tensor instead of a numpy array.
+
+    Returns:
+        Depth map in the requested unit and dtype.
+
+    Raises:
+        ValueError: If ``use_log=True`` and ``depth_min + shift <= 0``.
+        ValueError: If ``output_unit`` is not ``\"m\"`` or ``\"mm\"``.
+    """
+    if output_unit not in ("m", "mm"):
+        raise ValueError(f"output_unit must be 'm' or 'mm', got {output_unit!r}")
+
+    if isinstance(quantized, av.VideoFrame):
+        quantized = quantized.to_ndarray(format=pix_fmt)
+
+    norm = np.asarray(quantized, dtype=np.float32, order="K") / DEPTH_QMAX
+
+    depth_min_m = np.float32(depth_min)
+    depth_max_m = np.float32(depth_max)
+    shift_m = np.float32(shift)
+
+    # The de-normalization and de-quantization is performed in meters (convenience choice).
+    if use_log:
+        _validate_log_quant_params(depth_min, shift)
+        log_min = math.log(float(depth_min_m + shift_m))
+        log_max = math.log(float(depth_max_m + shift_m))
+        depth_m = np.exp(norm * (log_max - log_min) + log_min) - shift_m
+    else:
+        depth_m = norm * (depth_max_m - depth_min_m) + depth_min_m
+    depth_m = np.clip(depth_m, depth_min_m, depth_max_m).astype(np.float32, copy=False)
+
+    # Add single-channel dim: (H, W) → (H, W, 1)
+    if depth_m.ndim == 2:
+        depth_m = depth_m[..., np.newaxis]
+
+    # Return depth as float32 meters.
+    if output_unit == "m":
+        return torch.from_numpy(depth_m) if output_tensor else depth_m
+
+    # Return depth as uint16 millimeters.
+    mm = np.rint(depth_m * _MM_PER_METRE).clip(0, _UINT16_MAX).astype(np.uint16, copy=False)
+    if output_tensor:
+        # torch.uint16 support is very limited, we convert to float32 instead.
+        return torch.from_numpy(mm.astype(np.float32))
+    else:
+        return mm

src/lerobot/datasets/factory.py

@@ -19,6 +19,7 @@ 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
@@ -30,12 +31,14 @@ from .streaming_dataset import StreamingLeRobotDataset
 
 
 def resolve_delta_timestamps(
-    cfg: PreTrainedConfig, ds_meta: LeRobotDatasetMetadata
+    cfg: PreTrainedConfig | RewardModelConfig, ds_meta: LeRobotDatasetMetadata
 ) -> dict[str, list] | None:
-    """Resolves delta_timestamps by reading from the 'delta_indices' properties of the PreTrainedConfig.
+    """Resolves delta_timestamps by reading from the 'delta_indices' properties of the config.
 
     Args:
-        cfg (PreTrainedConfig): The PreTrainedConfig to read delta_indices from.
+        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.
         ds_meta (LeRobotDatasetMetadata): The dataset from which features and fps are used to build
             delta_timestamps against.
 
@@ -82,7 +85,7 @@ 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.policy, ds_meta)
+        delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, ds_meta)
         if not cfg.dataset.streaming:
             dataset = LeRobotDataset(
                 cfg.dataset.repo_id,
@@ -92,6 +95,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 image_transforms=image_transforms,
                 revision=cfg.dataset.revision,
                 video_backend=cfg.dataset.video_backend,
+                return_uint8=True,
                 tolerance_s=cfg.tolerance_s,
             )
         else:
@@ -104,6 +108,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 revision=cfg.dataset.revision,
                 max_num_shards=cfg.num_workers,
                 tolerance_s=cfg.tolerance_s,
+                return_uint8=True,
             )
     else:
         raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")

src/lerobot/datasets/feature_utils.py

@@ -19,6 +19,7 @@ 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
 
@@ -28,6 +29,7 @@ from .utils import (
     DEFAULT_DATA_PATH,
     DEFAULT_VIDEO_FILE_SIZE_IN_MB,
     DEFAULT_VIDEO_PATH,
+    DatasetInfo,
 )
 
 
@@ -78,8 +80,8 @@ 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,
-) -> dict:
-    """Create a template dictionary for a new dataset's `info.json`.
+) -> DatasetInfo:
+    """Create a template ``DatasetInfo`` object for a new dataset's ``meta/info.json``.
 
     Args:
         codebase_version (str): The version of the LeRobot codebase.
@@ -87,25 +89,59 @@ 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:
-        dict: A dictionary with the initial dataset metadata.
+        DatasetInfo: A typed dataset information object with initial metadata.
     """
-    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,
-    }
+    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
 
 
 def check_delta_timestamps(
@@ -285,7 +321,7 @@ def validate_feature_image_or_video(
 
     Args:
         name (str): The name of the feature.
-        expected_shape (list[str]): The expected shape (C, H, W).
+        expected_shape (list[str]): The expected shape, e.g. (C, H, W) or (H, W, C).
         value: The image data to validate.
 
     Returns:

src/lerobot/datasets/image_writer.py

@@ -30,22 +30,53 @@ def safe_stop_image_writer(func):
     def wrapper(*args, **kwargs):
         try:
             return func(*args, **kwargs)
-        except Exception as e:
+        except BaseException:
             dataset = kwargs.get("dataset")
             writer = getattr(dataset, "writer", None) if dataset else None
             if writer is not None and writer.image_writer is not None:
                 logger.warning("Waiting for image writer to terminate...")
                 writer.image_writer.stop()
-            raise e
+            raise
 
     return wrapper
 
 
 def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True) -> PIL.Image.Image:
-    # TODO(aliberts): handle 1 channel and 4 for depth images
-    if image_array.ndim != 3:
-        raise ValueError(f"The array has {image_array.ndim} dimensions, but 3 is expected for an image.")
+    """Convert a NumPy array to a PIL Image, preserving precision for grayscale.
 
+    Behaviour by shape:
+
+    - ``(H, W)`` or ``(1, H, W)`` / ``(H, W, 1)``: single-channel grayscale.
+      The native dtype is preserved using the matching PIL mode
+      (``I;16`` / ``F``). This is the path used for raw depth maps (no rescaling, clamping, or downcasting)
+    - ``(3, H, W)`` / ``(H, W, 3)``: RGB. Channels-first inputs are transposed
+      to channels-last. Float inputs in ``[0, 1]`` are scaled to ``uint8``
+      (existing behaviour, gated by ``range_check``).
+
+    Other shapes / channel counts raise ``NotImplementedError`` or
+    ``ValueError``.
+    """
+    # TODO(CarolinePascal): 4 dimensions RGB-D images
+    if image_array.ndim not in (2, 3):
+        raise ValueError(f"The array has {image_array.ndim} dimensions, but 2 or 3 is expected for an image.")
+
+    # Squeeze 3D single-channel inputs to 2D so depth maps work whether the
+    # caller emits (H, W), (1, H, W), or (H, W, 1).
+    if image_array.ndim == 3:
+        if image_array.shape[0] == 1:
+            image_array = image_array[0]
+        elif image_array.shape[-1] == 1:
+            image_array = image_array[..., 0]
+
+    if image_array.ndim == 2:
+        if image_array.dtype not in [np.uint16, np.float32]:
+            raise ValueError(
+                f"Unsupported single-channel image dtype: {image_array.dtype}. "
+                f"Supported dtypes: {sorted(str(d) for d in [np.uint16, np.float32])}."
+            )
+        return PIL.Image.fromarray(np.ascontiguousarray(image_array))
+
+    # 3D path: must be RGB (3 channels), channels-first or channels-last.
     if image_array.shape[0] == 3:
         # Transpose from pytorch convention (C, H, W) to (H, W, C)
         image_array = image_array.transpose(1, 2, 0)
@@ -71,13 +102,28 @@ def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True)
     return PIL.Image.fromarray(image_array)
 
 
+def save_kwargs_for_path(fpath: Path, compress_level: int) -> dict:
+    """Pick the right format-specific kwargs for :meth:`PIL.Image.Image.save`.
+
+    PNG uses ``compress_level`` (0-9, zlib). TIFF uses ``compression`` (raw) for lossless raw depth maps.
+    """
+    suffix = Path(fpath).suffix.lower()
+    if suffix == ".png":
+        return {"compress_level": compress_level}
+    if suffix in (".tif", ".tiff"):
+        return {"compression": "raw"}
+    return {}
+
+
 def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path, compress_level: int = 1):
     """
     Saves a NumPy array or PIL Image to a file.
 
     This function handles both NumPy arrays and PIL Image objects, converting
     the former to a PIL Image before saving. It includes error handling for
-    the save operation.
+    the save operation. The output format is inferred from the *fpath*
+    extension: ``.png`` → PNG with ``compress_level``, ``.tiff`` / ``.tif``
+    → lossless raw depth maps (TIFF).
 
     Args:
         image (np.ndarray | PIL.Image.Image): The image data to save.
@@ -101,7 +147,7 @@ def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path, compress_level
             img = image
         else:
             raise TypeError(f"Unsupported image type: {type(image)}")
-        img.save(fpath, compress_level=compress_level)
+        img.save(fpath, **save_kwargs_for_path(fpath, compress_level))
     except Exception as e:
         logger.error("Error writing image %s: %s", fpath, e)