feat(language): tool catalog in meta/info.json + LeRobotDatasetMetadata.tools

Stores OpenAI-style function schemas at ``meta/info.json["tools"]`` so
datasets can declare which tools are available (today: just ``say``;
tomorrow: per-dataset extensions). The ``DEFAULT_TOOLS`` constant
fills in for unannotated datasets so chat-template consumers don't
have to special-case anything.

Three pieces:

- ``language.py``: ``SAY_TOOL_SCHEMA`` and ``DEFAULT_TOOLS``
  constants. Single source of truth — PR 2's writer and PR 3's
  runtime tool registry will both import from here instead of
  duplicating the dict.
- ``dataset_metadata.py``: ``LeRobotDatasetMetadata.tools`` property
  reads ``info.json["tools"]`` and falls back to ``DEFAULT_TOOLS``.
  Returns deep-copied dicts so callers can mutate the result safely.
- ``docs/source/tools.mdx``: spec page covering the catalog, per-row
  invocations, and the three-step "how to add a new tool" workflow
  (declare schema, implement, register). Linked from the docs
  toctree under the Datasets section.

This lays the groundwork for PR 2's pipeline writing the catalog out
during annotation, and PR 3's ``src/lerobot/tools/`` package shipping
runnable implementations (one file per tool — first up:
``say.py`` wrapping Kyutai's pocket-tts).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.github/workflows/benchmark_tests.yml

@@ -118,7 +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 \
@@ -147,7 +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()
@@ -270,7 +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 \
@@ -299,7 +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()
@@ -317,6 +317,115 @@ jobs:
           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\" \
+                --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
@@ -416,3 +525,421 @@ jobs:
           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] \
+                --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="${LIBERO_PLUS_TASK_IDS}" \
+            lerobot-benchmark-libero-plus:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              lerobot-eval \
+                --policy.path=\"\$LIBERO_PLUS_POLICY\" \
+                --env.type=libero_plus \
+                --env.task=\"\$LIBERO_PLUS_SUITE\" \
+                --env.task_ids=\"\$LIBERO_PLUS_TASK_IDS\" \
+                --eval.batch_size=1 \
+                --eval.n_episodes=1 \
+                --eval.use_async_envs=false \
+                --policy.device=cuda \
+                '--env.camera_name_mapping={\"agentview_image\": \"camera1\", \"robot0_eye_in_hand_image\": \"camera2\"}' \
+                --policy.empty_cameras=1 \
+                --output_dir=/tmp/eval-artifacts
+              python scripts/ci/extract_task_descriptions.py \
+                --env libero_plus --task \"\$LIBERO_PLUS_SUITE\" \
+                --output /tmp/eval-artifacts/task_descriptions.json
+            "
+
+      - name: Copy LIBERO-plus artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/libero-plus-artifacts
+          docker cp libero-plus-eval:/tmp/eval-artifacts/. /tmp/libero-plus-artifacts/ 2>/dev/null || true
+          docker rm -f libero-plus-eval || true
+
+      - name: Parse LIBERO-plus eval metrics
+        if: always()
+        run: |
+          python3 scripts/ci/parse_eval_metrics.py \
+            --artifacts-dir /tmp/libero-plus-artifacts \
+            --env libero_plus \
+            --task "${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 \
+                --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

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.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-12-4 cuda-cudart-dev-12-4 \
+         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"]

docs/source/_toctree.yml

@@ -31,8 +31,10 @@
     title: Porting Large Datasets
   - local: using_dataset_tools
     title: Using the Dataset Tools
-  - local: dataset_subtask
-    title: Using Subtasks in the Dataset
+  - local: language_and_recipes
+    title: Language Columns and Recipes
+  - local: tools
+    title: Tools
   - local: streaming_video_encoding
     title: Streaming Video Encoding
   title: "Datasets"
@@ -77,12 +79,22 @@
     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

docs/source/dataset_subtask.mdx

@@ -1,277 +0,0 @@
-# Using Subtasks in LeRobot Datasets
-
-Subtask support in robotics datasets has proven effective in improving robot reasoning and understanding. Subtasks are particularly useful for:
-
-- **Hierarchical policies**: Building policies that include subtask predictions to visualize robot reasoning in real time
-- **Reward modeling**: Helping reward models understand task progression (e.g., SARM-style stage-aware reward models)
-- **Task decomposition**: Breaking down complex manipulation tasks into atomic, interpretable steps
-
-LeRobotDataset now supports subtasks as part of its dataset structure, alongside tasks.
-
-## What are Subtasks?
-
-While a **task** describes the overall goal (e.g., "Pick up the apple and place it in the basket"), **subtasks** break down the execution into finer-grained steps:
-
-1. "Approach the apple"
-2. "Grasp the apple"
-3. "Lift the apple"
-4. "Move to basket"
-5. "Release the apple"
-
-Each frame in the dataset can be annotated with its corresponding subtask, enabling models to learn and predict these intermediate stages.
-
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/subtask-asset.png"
-  alt="An overview of subtask annotation showing how frames are labeled with intermediate subtask stages"
-  width="80%"
-/>
-
-<p>
-  <em>Figure: Overview of subtask annotation.</em>
-</p>
-
-**Reference:** _Subtask-learning based for robot self-assembly in flexible collaborative assembly in manufacturing_, Original Article, Published: 19 April 2022.
-
-## Dataset Structure
-
-Subtask information is stored in the dataset metadata:
-
-```
-my-dataset/
-├── data/
-│   └── ...
-├── meta/
-│   ├── info.json
-│   ├── stats.json
-│   ├── tasks.parquet
-│   ├── subtasks.parquet      # Subtask index → subtask string mapping
-│   └── episodes/
-│       └── ...
-└── videos/
-    └── ...
-```
-
-### Subtasks Parquet File
-
-The `meta/subtasks.parquet` file maps subtask indices to their natural language descriptions:
-
-| subtask_index | subtask (index column) |
-| ------------- | ---------------------- |
-| 0             | "Approach the apple"   |
-| 1             | "Grasp the apple"      |
-| 2             | "Lift the apple"       |
-| ...           | ...                    |
-
-### Frame-Level Annotations
-
-Each frame in the dataset can include a `subtask_index` field that references the subtasks parquet file:
-
-```python
-# Example frame data in the parquet file
-{
-    "index": 42,
-    "timestamp": 1.4,
-    "episode_index": 0,
-    "task_index": 0,
-    "subtask_index": 2,  # References "Lift the apple"
-    "observation.state": [...],
-    "action": [...],
-}
-```
-
-## Annotating Datasets with Subtasks
-
-We provide a HuggingFace Space for easily annotating any LeRobotDataset with subtasks:
-
-**[https://huggingface.co/spaces/lerobot/annotate](https://huggingface.co/spaces/lerobot/annotate)**
-
-After completing your annotation:
-
-1. Click "Push to Hub" to upload your annotated dataset
-2. You can also run the annotation space locally by following the instructions at [github.com/huggingface/lerobot-annotate](https://github.com/huggingface/lerobot-annotate)
-
-## Loading Datasets with Subtasks
-
-When you load a dataset with subtask annotations, the subtask information is automatically available:
-
-```python
-from lerobot.datasets import LeRobotDataset
-
-# Load a dataset with subtask annotations
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-# Access a sample
-sample = dataset[100]
-
-# The sample includes both task and subtask information
-print(sample["task"])        # "Collect the fruit"
-print(sample["subtask"])     # "Grasp the apple"
-print(sample["task_index"])  # tensor(0)
-print(sample["subtask_index"])  # tensor(2)
-```
-
-### Checking for Subtask Support
-
-You can check if a dataset has subtask annotations:
-
-```python
-# Check if subtasks are available
-has_subtasks = (
-    "subtask_index" in dataset.features
-    and dataset.meta.subtasks is not None
-)
-
-if has_subtasks:
-    print(f"Dataset has {len(dataset.meta.subtasks)} unique subtasks")
-    print("Subtasks:", list(dataset.meta.subtasks.index))
-```
-
-## Using Subtasks for Training
-
-### With the Tokenizer Processor
-
-The `TokenizerProcessor` automatically handles subtask tokenization for Vision-Language Action (VLA) models:
-
-```python
-from lerobot.processor import TokenizerProcessorStep
-
-# Create a tokenizer processor step
-tokenizer_processor = TokenizerProcessorStep(
-    tokenizer_name_or_path="google/paligemma-3b-pt-224",
-    padding="max_length",
-    max_length=64,
-)
-
-# The processor will automatically tokenize subtasks if present in the batch
-# and add them to the observation under:
-# - "observation.subtask.tokens"
-# - "observation.subtask.attention_mask"
-```
-
-When subtasks are available in the batch, the tokenizer processor adds:
-
-- `observation.subtask.tokens`: Tokenized subtask text
-- `observation.subtask.attention_mask`: Attention mask for the subtask tokens
-
-### DataLoader with Subtasks
-
-```python
-import torch
-from lerobot.datasets import LeRobotDataset
-
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    batch_size=16,
-    shuffle=True,
-)
-
-for batch in dataloader:
-    # Access subtask information in the batch
-    subtasks = batch["subtask"]  # List of subtask strings
-    subtask_indices = batch["subtask_index"]  # Tensor of subtask indices
-
-    # Use for training hierarchical policies or reward models
-    print(f"Batch subtasks: {set(subtasks)}")
-```
-
-## Example Datasets with Subtask Annotations
-
-Try loading a dataset with subtask annotations:
-
-```python
-from lerobot.datasets import LeRobotDataset
-
-# Example dataset with subtask annotations
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-# Explore the subtasks
-print("Available subtasks:")
-for subtask_name in dataset.meta.subtasks.index:
-    print(f"  - {subtask_name}")
-
-# Get subtask distribution
-subtask_counts = {}
-for i in range(len(dataset)):
-    sample = dataset[i]
-    subtask = sample["subtask"]
-    subtask_counts[subtask] = subtask_counts.get(subtask, 0) + 1
-
-print("\nSubtask distribution:")
-for subtask, count in sorted(subtask_counts.items(), key=lambda x: -x[1]):
-    print(f"  {subtask}: {count} frames")
-```
-
-## Use Cases
-
-### 1. Hierarchical Policy Training
-
-Train policies that predict both actions and current subtask:
-
-```python
-class HierarchicalPolicy(nn.Module):
-    def __init__(self, num_subtasks):
-        super().__init__()
-        self.action_head = nn.Linear(hidden_dim, action_dim)
-        self.subtask_head = nn.Linear(hidden_dim, num_subtasks)
-
-    def forward(self, observations):
-        features = self.encoder(observations)
-        actions = self.action_head(features)
-        subtask_logits = self.subtask_head(features)
-        return actions, subtask_logits
-```
-
-### 2. Stage-Aware Reward Modeling (SARM)
-
-Build reward models that understand task progression:
-
-```python
-# SARM predicts:
-# - Stage: Which subtask is being executed (discrete)
-# - Progress: How far along the subtask (continuous 0-1)
-
-class SARMRewardModel(nn.Module):
-    def forward(self, observations):
-        features = self.encoder(observations)
-        stage_logits = self.stage_classifier(features)
-        progress = self.progress_regressor(features)
-        return stage_logits, progress
-```
-
-### 3. Progress Visualization
-
-Monitor robot execution by tracking subtask progression:
-
-```python
-def visualize_execution(model, observations):
-    for t, obs in enumerate(observations):
-        action, subtask_logits = model(obs)
-        predicted_subtask = subtask_names[subtask_logits.argmax()]
-        print(f"t={t}: Executing '{predicted_subtask}'")
-```
-
-## API Reference
-
-### LeRobotDataset Properties
-
-| Property                    | Type                   | Description                                |
-| --------------------------- | ---------------------- | ------------------------------------------ |
-| `meta.subtasks`             | `pd.DataFrame \| None` | DataFrame mapping subtask names to indices |
-| `features["subtask_index"]` | `dict`                 | Feature spec for subtask_index if present  |
-
-### Sample Keys
-
-When subtasks are available, each sample includes:
-
-| Key             | Type           | Description                          |
-| --------------- | -------------- | ------------------------------------ |
-| `subtask_index` | `torch.Tensor` | Integer index of the current subtask |
-| `subtask`       | `str`          | Natural language subtask description |
-
-## Related Resources
-
-- [SARM Paper](https://arxiv.org/pdf/2509.25358) - Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
-- [LeRobot Annotate Space](https://huggingface.co/spaces/lerobot/annotate) - Interactive annotation tool
-- [LeRobotDataset v3.0](./lerobot-dataset-v3) - Dataset format documentation

docs/source/language_and_recipes.mdx

@@ -0,0 +1,109 @@
+# Language columns and recipes
+
+LeRobot stores reusable language annotations directly next to frame data in `data/chunk-*/file-*.parquet`.
+The two optional columns are:
+
+- `language_persistent`: a list of rows broadcast across every frame in an episode for state that remains active, such as `subtask`, `plan`, and `memory`.
+- `language_events`: a list of rows only on the exact frame where an event was emitted, such as `interjection`, `vqa`, and speech tool calls.
+
+Both columns share the same row shape (event rows omit `timestamp` because the
+frame the row sits on already provides it):
+
+```text
+role: string
+content: string | null
+style: string | null
+timestamp: float64        # persistent rows only
+camera: string | null     # observation.images.* feature key, view-dependent rows only
+tool_calls: list[Json] | null
+```
+
+The `camera` field tags rows whose `content` is grounded in a specific camera
+view. Rows of view-dependent styles (`vqa`, and the reserved `motion` /
+`trace`) MUST set `camera` to the matching `observation.images.*` feature key.
+Rows of every other style MUST leave `camera` as `null`. Pipeline writers and
+the validator enforce this via `validate_camera_field(style, camera)`.
+
+`meta/tasks.parquet` remains the canonical source for the task. The special `${task}` recipe binding always reads that task string and does not depend on language annotations.
+
+## Architecture
+
+The language stack has three layers:
+
+1. `lerobot.datasets.language` defines the schema, style registry, and `column_for_style`.
+2. `lerobot.datasets.language_render` resolves rows and renders messages.
+3. `RenderMessagesStep` turns dataset samples into `messages`, `message_streams`, and `target_message_indices`.
+
+`LeRobotDataset` stays recipe-agnostic. It passes `language_persistent` and `language_events` through when present, and unannotated datasets keep their existing behavior.
+
+## Temporal semantics
+
+Persistent styles are active after emission until replaced:
+
+- `active_at(t, style=subtask)`
+- `nth_prev(style=memory, offset=1)`
+- `nth_next(style=subtask, offset=1)`
+
+Event styles only exist on their exact timestamp:
+
+- `emitted_at(t, style=interjection)`
+- `emitted_at(t, style=vqa, role=user, camera=observation.images.top)`
+- `emitted_at(t, role=assistant, tool_name=say)`
+
+Exact event matching has no tolerance window, so writers must stamp event rows with frame timestamps from the parquet data.
+
+## View-dependent resolution
+
+For view-dependent styles (`vqa`, `motion`, `trace`), the resolver gains a
+`camera=` filter parallel to `role=` and `tool_name=`. Datasets with multiple
+cameras typically emit one (`vqa`, `user`) + (`vqa`, `assistant`) pair per
+camera at the same timestamp; without `camera=`, those resolvers see two
+matches and raise an ambiguity error. Recipes consume each camera through its
+own binding plus a matching image block, e.g.
+
+```yaml
+ask_vqa_top:
+  bindings:
+    vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
+    vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
+  messages:
+    - role: user
+      stream: high_level
+      if_present: vqa_query
+      content:
+        - { type: image, feature: observation.images.top }
+        - { type: text, text: "${vqa_query}" }
+    - { role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa }
+```
+
+Add one such sub-recipe per camera the dataset records.
+
+## Recipe anatomy
+
+Recipes are YAML files backed by `TrainingRecipe` and `MessageTurn`.
+
+```yaml
+messages:
+  - { role: user, content: "${task}", stream: high_level }
+  - { role: assistant, content: "${subtask}", stream: low_level, target: true }
+```
+
+Rendered samples use HF-style chat messages plus LeRobot sidecars:
+
+```python
+sample["messages"]
+sample["message_streams"]
+sample["target_message_indices"]
+```
+
+The renderer does not apply a tokenizer chat template. Policy processors decide how to serialize the messages for their backbone.
+
+## Blends
+
+Blend recipes select one weighted sub-recipe deterministically from the sample index.
+The canonical `recipes/pi05_hirobot.yaml` combines memory updates, interjection responses, high-level subtask prediction, low-level execution, and VQA.
+
+## Graceful absence
+
+If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.
+If an event-scoped branch is selected on a frame without the required event row, rendering returns `None`, allowing a loader to retry another sample.

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/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/tools.mdx

@@ -0,0 +1,198 @@
+# Tools
+
+LeRobot v3.1 supports **tool calls** in policies — assistant messages can
+emit structured invocations like `say(text="OK, starting now")` that the
+runtime dispatches to a real implementation (TTS, controller, logger, …).
+
+This page covers:
+
+1. Where the tool catalog lives (PR 1).
+2. How the annotation pipeline produces tool-call atoms (PR 2).
+3. How to add your own tool (PR 3).
+
+## Where tools are declared
+
+Two layers.
+
+**The catalog** — a list of OpenAI-style function schemas — lives at
+`meta/info.json["tools"]` on each dataset. Example:
+
+```json
+{
+  "features": { "...": "..." },
+  "tools": [
+    {
+      "type": "function",
+      "function": {
+        "name": "say",
+        "description": "Speak a short utterance to the user via the TTS executor.",
+        "parameters": {
+          "type": "object",
+          "properties": {
+            "text": { "type": "string", "description": "The verbatim text to speak." }
+          },
+          "required": ["text"]
+        }
+      }
+    }
+  ]
+}
+```
+
+Read it via the dataset metadata accessor:
+
+```python
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+
+meta = LeRobotDatasetMetadata(repo_id="pepijn/super_poulain_final_annotations")
+tools = meta.tools     # list[dict] — OpenAI tool schemas
+```
+
+If the dataset's `info.json` doesn't declare any tools, `meta.tools`
+returns `DEFAULT_TOOLS` from `lerobot.datasets.language` — currently a
+single-entry list with the canonical `say` schema. So unannotated
+datasets and chat-template consumers keep working without any
+configuration:
+
+```python
+prompt_str = tokenizer.apply_chat_template(
+    sample["messages"],
+    tools=meta.tools,                 # works either way
+    add_generation_prompt=False,
+    tokenize=False,
+)
+```
+
+**The implementations** — runnable Python — live under
+`src/lerobot/tools/`, one file per tool. The `say` implementation
+arrives in PR 3 and wraps Kyutai's pocket-tts model.
+
+## Per-row tool *invocations*
+
+The catalog above describes *what can be called*. The actual *call* — the
+function name plus the argument values — is stored per-row, on the
+assistant atoms in `language_events`:
+
+```python
+{
+  "role": "assistant",
+  "content": null,
+  "style": null,
+  "timestamp": 12.4,
+  "camera": null,
+  "tool_calls": [
+    { "type": "function",
+      "function": { "name": "say", "arguments": { "text": "On it." } } }
+  ]
+}
+```
+
+Recipes splice these into rendered messages via `tool_calls_from`:
+
+```yaml
+user_interjection_response:
+  bindings:
+    speech: "emitted_at(t, role=assistant, tool_name=say)"
+  messages:
+    - { role: user,      content: "${task}",         stream: high_level }
+    - { role: assistant, content: "${current_plan}", stream: high_level,
+        target: true, tool_calls_from: speech }
+```
+
+The model's training target is one assistant turn that carries both the
+plan text *and* the `say` tool call. At inference, the runtime parses
+the generated text back into structured `tool_calls` and dispatches to
+the matching implementation.
+
+## How to add your own tool
+
+Three steps. Concrete example: a `record_observation` tool the policy
+can call to capture an extra observation outside the regular control
+loop.
+
+### Step 1 — declare the schema
+
+Add an entry under `meta/info.json["tools"]`. Either edit the file
+directly on disk *before* running the annotation pipeline (it'll be
+preserved) or hand it to `lerobot-annotate` via a config flag (PR 2 —
+exact CLI lands with the pipeline change).
+
+```json
+{
+  "tools": [
+    { "type": "function", "function": { "name": "say", "...": "..." } },
+    {
+      "type": "function",
+      "function": {
+        "name": "record_observation",
+        "description": "Capture a high-resolution still image for the user.",
+        "parameters": {
+          "type": "object",
+          "properties": {
+            "label": { "type": "string", "description": "Short label for the saved image." }
+          },
+          "required": ["label"]
+        }
+      }
+    }
+  ]
+}
+```
+
+The schema follows OpenAI's function-calling convention exactly, so the
+chat template can render it natively.
+
+### Step 2 — implement the call
+
+Create `src/lerobot/tools/record_observation.py`:
+
+```python
+from .base import Tool
+from typing import Any
+
+RECORD_OBSERVATION_SCHEMA: dict[str, Any] = { "...": "..." }   # mirrors the JSON above
+
+
+class RecordObservationTool:
+    name = "record_observation"
+    schema = RECORD_OBSERVATION_SCHEMA
+
+    def __init__(self, schema: dict | None = None, output_dir: str = "."):
+        self.output_dir = output_dir
+
+    def call(self, arguments: dict) -> str:
+        label = arguments["label"]
+        # ... save the latest camera frame to <output_dir>/<label>.png ...
+        return f"saved {label}.png"
+```
+
+One file per tool keeps dependencies isolated — `record_observation`
+might pull `pillow`, while `say` (PR 3) pulls `pocket-tts`. Users
+installing only the tools they need avoid heavy transitive deps.
+
+### Step 3 — register it
+
+Add to `src/lerobot/tools/registry.py` (PR 3):
+
+```python
+from .record_observation import RecordObservationTool
+
+TOOL_REGISTRY["record_observation"] = RecordObservationTool
+```
+
+That's it. At runtime `get_tools(meta)` looks up each schema in
+`meta.tools`, instantiates the matching registered class, and returns
+a name → instance dict the dispatcher can route into.
+
+## Where this fits in the three-PR stack
+
+| Layer | PR | What lands |
+|---|---|---|
+| Catalog storage in `meta/info.json` + `meta.tools` accessor | PR 1 | This page; `SAY_TOOL_SCHEMA`, `DEFAULT_TOOLS` constants in `lerobot.datasets.language`; `LeRobotDatasetMetadata.tools` property |
+| Annotation pipeline writes `tools` to meta after a run; honors anything users pre-populated | PR 2 | `lerobot-annotate` ensures `meta/info.json["tools"]` includes the canonical `say` and merges any user-declared tools |
+| Runnable implementations under `src/lerobot/tools/`; runtime dispatcher; `say.py` wired to Kyutai's pocket-tts | PR 3 | One file per tool; `Tool` protocol; `TOOL_REGISTRY`; optional `[tools]` extra in `pyproject.toml` |
+
+If you want to use a tool *without* writing an implementation (e.g. for
+training-time chat-template formatting only), step 1 alone is enough —
+the model still learns to *generate* the call. Steps 2 and 3 are only
+needed to actually *execute* it at inference.

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.

pyproject.toml

@@ -95,7 +95,7 @@ dependencies = [
 
 # ── Feature-scoped extras ──────────────────────────────────
 dataset = [
-    "datasets>=4.0.0,<5.0.0",
+    "datasets>=4.7.0,<5.0.0",
     "pandas>=2.0.0,<3.0.0", # NOTE: Transitive dependency of datasets
     "pyarrow>=21.0.0,<30.0.0", # NOTE: Transitive dependency of datasets
     "lerobot[av-dep]",
@@ -212,6 +212,15 @@ 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

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]:
@@ -57,6 +74,24 @@ 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.
 
@@ -74,21 +109,85 @@ def _robocasa_descriptions(task_spec: str) -> dict[str, str]:
     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/configs/__init__.py

@@ -23,6 +23,7 @@ Import them directly: ``from lerobot.configs.train import TrainPipelineConfig``
 
 from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
+from .recipe import MessageTurn, TrainingRecipe, load_recipe
 from .types import (
     FeatureType,
     NormalizationMode,
@@ -41,7 +42,10 @@ __all__ = [
     # Config classes
     "DatasetConfig",
     "EvalConfig",
+    "MessageTurn",
     "PeftConfig",
     "PreTrainedConfig",
+    "TrainingRecipe",
     "WandBConfig",
+    "load_recipe",
 ]

src/lerobot/configs/recipe.py

@@ -0,0 +1,193 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Literal, get_args
+
+MessageRole = Literal["user", "assistant", "system", "tool"]
+MessageStream = Literal["high_level", "low_level"]
+
+DEFAULT_BINDINGS = {
+    "subtask": "active_at(t, style=subtask)",
+    "memory": "active_at(t, style=memory)",
+    "plan": "active_at(t, style=plan)",
+    "speech": "emitted_at(t, role=assistant, tool_name=say)",
+    "interjection": "emitted_at(t, style=interjection)",
+    "vqa": "emitted_at(t, style=vqa, role=assistant)",
+    "vqa_query": "emitted_at(t, style=vqa, role=user)",
+}
+
+_PLACEHOLDER_RE = re.compile(r"\$\{([A-Za-z_][A-Za-z0-9_]*)\}")
+_VALID_ROLES = frozenset(get_args(MessageRole))
+_VALID_STREAMS = frozenset(get_args(MessageStream))
+
+
+@dataclass
+class MessageTurn:
+    """A single chat-style turn in a recipe template.
+
+    ``content`` may be a plain string, a list of HF-style multimodal blocks, or
+    ``None`` when ``tool_calls_from`` supplies tool-call payloads instead.
+    ``stream`` tags the turn for downstream filtering, ``target`` flags it as a
+    training target, and ``if_present`` skips the turn when the named binding
+    resolves to ``None``.
+    """
+
+    role: MessageRole
+    content: str | list[dict[str, Any]] | None = None
+    stream: MessageStream | None = None
+    target: bool = False
+    if_present: str | None = None
+    tool_calls_from: str | None = None
+
+    def __post_init__(self) -> None:
+        """Validate role, stream, and content after dataclass construction."""
+        if self.role not in _VALID_ROLES:
+            raise ValueError(f"Unsupported message role: {self.role!r}")
+        if self.stream is not None and self.stream not in _VALID_STREAMS:
+            raise ValueError(f"Unsupported message stream: {self.stream!r}")
+        if self.content is None and self.tool_calls_from is None:
+            raise ValueError("MessageTurn.content is required unless tool_calls_from is set.")
+        if self.content is not None and not isinstance(self.content, (str, list)):
+            raise TypeError("MessageTurn.content must be a string, a list of HF-style blocks, or None.")
+        if isinstance(self.content, list):
+            for block in self.content:
+                if not isinstance(block, dict) or "type" not in block:
+                    raise ValueError(
+                        "Multimodal content blocks must be HF-style dictionaries with a type key."
+                    )
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> MessageTurn:
+        """Construct a :class:`MessageTurn` from a plain dictionary."""
+        return cls(**data)
+
+
+@dataclass
+class TrainingRecipe:
+    """A recipe describing how to render training samples from language rows.
+
+    A recipe is either a *message recipe* (``messages`` plus optional
+    ``bindings``) or a *blend recipe* (``blend`` mapping names to weighted
+    sub-recipes). ``weight`` is only meaningful inside a blend.
+    """
+
+    messages: list[MessageTurn] | None = None
+    bindings: dict[str, str] | None = None
+    blend: dict[str, TrainingRecipe] | None = None
+    weight: float | None = None
+
+    def __post_init__(self) -> None:
+        """Validate that exactly one of ``messages`` or ``blend`` is set."""
+        if self.messages is not None and self.blend is not None:
+            raise ValueError("TrainingRecipe must set only one of messages or blend.")
+        if self.messages is None and self.blend is None:
+            raise ValueError("TrainingRecipe must set one of messages or blend.")
+
+        if self.messages is not None:
+            self._validate_message_recipe()
+        if self.blend is not None:
+            self._validate_blend_recipe()
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> TrainingRecipe:
+        """Construct a :class:`TrainingRecipe` from a nested dictionary."""
+        data = dict(data)
+        if data.get("messages") is not None:
+            data["messages"] = [
+                turn if isinstance(turn, MessageTurn) else MessageTurn.from_dict(turn)
+                for turn in data["messages"]
+            ]
+        if data.get("blend") is not None:
+            data["blend"] = {
+                name: recipe if isinstance(recipe, TrainingRecipe) else cls.from_dict(recipe)
+                for name, recipe in data["blend"].items()
+            }
+        return cls(**data)
+
+    @classmethod
+    def from_yaml(cls, path: str | Path) -> TrainingRecipe:
+        """Load a :class:`TrainingRecipe` from a YAML file at ``path``."""
+        import yaml  # type: ignore[import-untyped]
+
+        with open(path) as f:
+            data = yaml.safe_load(f)
+        if not isinstance(data, dict):
+            raise ValueError(f"Recipe YAML must contain a mapping at the top level: {path}")
+        return cls.from_dict(data)
+
+    def _validate_message_recipe(self) -> None:
+        """Ensure every templated binding is known and at least one turn is a target."""
+        assert self.messages is not None
+        known_bindings = set(DEFAULT_BINDINGS) | set(self.bindings or {}) | {"task"}
+
+        for turn in self.messages:
+            missing = self._referenced_bindings(turn) - known_bindings
+            if missing:
+                raise ValueError(f"MessageTurn references unknown binding(s): {sorted(missing)}")
+
+        if not any(turn.target for turn in self.messages):
+            raise ValueError("Message recipes must contain at least one target turn.")
+
+    def _validate_blend_recipe(self) -> None:
+        """Ensure each blend component is a non-empty, weighted message recipe."""
+        assert self.blend is not None
+        if not self.blend:
+            raise ValueError("Blend recipes must contain at least one component.")
+
+        for name, recipe in self.blend.items():
+            if recipe.blend is not None:
+                raise ValueError(f"Blend component {name!r} cannot itself define a blend.")
+            if recipe.messages is None:
+                raise ValueError(f"Blend component {name!r} must define messages.")
+            if recipe.weight is None:
+                raise ValueError(f"Blend component {name!r} must define weight.")
+            if recipe.weight <= 0:
+                raise ValueError(f"Blend component {name!r} must have a positive weight.")
+
+    def _referenced_bindings(self, turn: MessageTurn) -> set[str]:
+        """Return the binding names that ``turn`` references via placeholders or attributes."""
+        names: set[str] = set()
+        if turn.if_present is not None:
+            names.add(turn.if_present)
+        if turn.tool_calls_from is not None:
+            names.add(turn.tool_calls_from)
+        names.update(_placeholders_in_content(turn.content))
+        return names
+
+
+def _placeholders_in_content(content: str | list[dict[str, Any]] | None) -> set[str]:
+    """Return the set of ``${name}`` placeholders found anywhere in ``content``."""
+    if content is None:
+        return set()
+    if isinstance(content, str):
+        return set(_PLACEHOLDER_RE.findall(content))
+
+    names: set[str] = set()
+    for block in content:
+        for value in block.values():
+            if isinstance(value, str):
+                names.update(_PLACEHOLDER_RE.findall(value))
+    return names
+
+
+def load_recipe(path: str | Path) -> TrainingRecipe:
+    """Load a :class:`TrainingRecipe` from a YAML file at ``path``."""
+    return TrainingRecipe.from_yaml(path)

src/lerobot/configs/recipes/pi05_hirobot.yaml

@@ -0,0 +1,74 @@
+blend:
+
+  memory_update:
+    weight: 0.10
+    bindings:
+      prior_memory: "nth_prev(style=memory, offset=1)"
+      current_memory: "emitted_at(t, style=memory)"
+      completed_subtask: "nth_prev(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "Previous memory: ${prior_memory}", stream: high_level, if_present: prior_memory}
+      - {role: user, content: "Completed subtask: ${completed_subtask}", stream: high_level, if_present: completed_subtask}
+      - {role: assistant, content: "${current_memory}", stream: high_level, target: true, if_present: current_memory}
+
+  user_interjection_response:
+    weight: 0.16
+    bindings:
+      prior_plan: "nth_prev(style=plan, offset=1)"
+      current_plan: "emitted_at(t, style=plan)"
+      interjection: "emitted_at(t, style=interjection)"
+      speech: "emitted_at(t, role=assistant, tool_name=say)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "Previous plan:\n${prior_plan}", stream: high_level, if_present: prior_plan}
+      - {role: user, content: "${interjection}", stream: high_level, if_present: interjection}
+      - {role: assistant, content: "${current_plan}", stream: high_level, target: true, if_present: current_plan, tool_calls_from: speech}
+
+  high_level_subtask:
+    weight: 0.15
+    bindings:
+      next_subtask: "nth_next(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${task}\nPlan: ${plan}\nMemory: ${memory}", stream: high_level}
+      - {role: user, content: "Current subtask: ${subtask}", stream: high_level, if_present: subtask}
+      - {role: assistant, content: "${next_subtask}", stream: high_level, target: true}
+
+  low_level_execution:
+    weight: 0.35
+    messages:
+      - {role: user, content: "${task}\nPlan: ${plan}\nMemory: ${memory}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: low_level, target: true}
+
+  # VQA is view-dependent: bbox / keypoint / count answers only make sense for
+  # the camera they were grounded against. Each camera gets its own sub-recipe
+  # so the resolver can disambiguate via `camera=...` and the user-turn carries
+  # the matching image block. Adjust the camera keys (and add more sub-recipes)
+  # to match the cameras present on your dataset.
+  ask_vqa_top:
+    weight: 0.10
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.top}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
+
+  ask_vqa_wrist:
+    weight: 0.10
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.wrist)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.wrist)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.wrist}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}

src/lerobot/datasets/__init__.py

@@ -37,6 +37,14 @@ from .dataset_tools import (
 from .factory import make_dataset, resolve_delta_timestamps
 from .image_writer import safe_stop_image_writer
 from .io_utils import load_episodes, write_stats
+from .language import (
+    EVENT_ONLY_STYLES,
+    LANGUAGE_EVENTS,
+    LANGUAGE_PERSISTENT,
+    PERSISTENT_STYLES,
+    STYLE_REGISTRY,
+    column_for_style,
+)
 from .lerobot_dataset import LeRobotDataset
 from .multi_dataset import MultiLeRobotDataset
 from .pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
@@ -53,10 +61,15 @@ __all__ = [
     "CODEBASE_VERSION",
     "DEFAULT_EPISODES_PATH",
     "DEFAULT_QUANTILES",
+    "EVENT_ONLY_STYLES",
     "EpisodeAwareSampler",
+    "LANGUAGE_EVENTS",
+    "LANGUAGE_PERSISTENT",
     "LeRobotDataset",
     "LeRobotDatasetMetadata",
     "MultiLeRobotDataset",
+    "PERSISTENT_STYLES",
+    "STYLE_REGISTRY",
     "StreamingLeRobotDataset",
     "VideoEncodingManager",
     "add_features",
@@ -66,6 +79,7 @@ __all__ = [
     "convert_image_to_video_dataset",
     "create_initial_features",
     "create_lerobot_dataset_card",
+    "column_for_style",
     "delete_episodes",
     "get_feature_stats",
     "load_episodes",

src/lerobot/datasets/compute_stats.py

@@ -512,7 +512,7 @@ def compute_episode_stats(
 
     ep_stats = {}
     for key, data in episode_data.items():
-        if features[key]["dtype"] == "string":
+        if features[key]["dtype"] in {"string", "language"}:
             continue
 
         if features[key]["dtype"] in ["image", "video"]:

src/lerobot/datasets/dataset_metadata.py

@@ -34,7 +34,6 @@ from .io_utils import (
     load_episodes,
     load_info,
     load_stats,
-    load_subtasks,
     load_tasks,
     write_info,
     write_json,
@@ -177,7 +176,6 @@ class LeRobotDatasetMetadata:
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
         self.tasks = load_tasks(self.root)
-        self.subtasks = load_subtasks(self.root)
         self.episodes = load_episodes(self.root)
         self.stats = load_stats(self.root)
 
@@ -320,6 +318,28 @@ class LeRobotDatasetMetadata:
         """Keys to access visual modalities (regardless of their storage method)."""
         return [key for key, ft in self.features.items() if ft["dtype"] in ["video", "image"]]
 
+    @property
+    def tools(self) -> list[dict]:
+        """OpenAI-style tool schemas declared by this dataset.
+
+        Read from ``meta/info.json["tools"]``. Returns a copy, so callers
+        can mutate the result safely. Falls back to
+        :data:`lerobot.datasets.language.DEFAULT_TOOLS` (the canonical
+        ``say`` schema) when the dataset doesn't declare any — that way
+        unannotated datasets and chat-template consumers
+        (``apply_chat_template(messages, tools=meta.tools)``) keep
+        working out of the box.
+
+        Implementations live under :mod:`lerobot.tools` (one file per
+        tool); see ``docs/source/tools.mdx`` for the authoring guide.
+        """
+        from .language import DEFAULT_TOOLS  # noqa: PLC0415  (avoid circular import)
+
+        declared = self.info.get("tools")
+        if isinstance(declared, list) and declared:
+            return [dict(t) for t in declared]
+        return [dict(t) for t in DEFAULT_TOOLS]
+
     @property
     def names(self) -> dict[str, list | dict]:
         """Names of the various dimensions of vector modalities."""
@@ -635,7 +655,6 @@ class LeRobotDatasetMetadata:
         _validate_feature_names(features)
 
         obj.tasks = None
-        obj.subtasks = None
         obj.episodes = None
         obj.stats = None
         obj.info = create_empty_dataset_info(

src/lerobot/datasets/dataset_reader.py

@@ -295,9 +295,4 @@ class DatasetReader:
         task_idx = item["task_index"].item()
         item["task"] = self._meta.tasks.iloc[task_idx].name
 
-        # add subtask information if available
-        if "subtask_index" in self._meta.features and self._meta.subtasks is not None:
-            subtask_idx = item["subtask_index"].item()
-            item["subtask"] = self._meta.subtasks.iloc[subtask_idx].name
-
         return item

src/lerobot/datasets/feature_utils.py

@@ -22,6 +22,12 @@ from PIL import Image as PILImage
 from lerobot.utils.constants import DEFAULT_FEATURES
 from lerobot.utils.utils import is_valid_numpy_dtype_string
 
+from .language import (
+    LANGUAGE_PERSISTENT,
+    is_language_column,
+    language_events_column_feature,
+    language_persistent_column_feature,
+)
 from .utils import (
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
@@ -45,7 +51,13 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
     """
     hf_features = {}
     for key, ft in features.items():
-        if ft["dtype"] == "video":
+        if is_language_column(key):
+            hf_features[key] = (
+                language_persistent_column_feature()
+                if key == LANGUAGE_PERSISTENT
+                else language_events_column_feature()
+            )
+        elif ft["dtype"] == "video":
             continue
         elif ft["dtype"] == "image":
             hf_features[key] = datasets.Image()
@@ -242,6 +254,8 @@ def validate_feature_dtype_and_shape(
         return validate_feature_image_or_video(name, expected_shape, value)
     elif expected_dtype == "string":
         return validate_feature_string(name, value)
+    elif expected_dtype == "language":
+        return ""
     else:
         raise NotImplementedError(f"The feature dtype '{expected_dtype}' is not implemented yet.")
 

src/lerobot/datasets/io_utils.py

@@ -34,7 +34,6 @@ from lerobot.utils.utils import SuppressProgressBars, flatten_dict, unflatten_di
 from .utils import (
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_EPISODES_PATH,
-    DEFAULT_SUBTASKS_PATH,
     DEFAULT_TASKS_PATH,
     EPISODES_DIR,
     INFO_PATH,
@@ -189,14 +188,6 @@ def load_tasks(local_dir: Path) -> pandas.DataFrame:
     return tasks
 
 
-def load_subtasks(local_dir: Path) -> pandas.DataFrame | None:
-    """Load subtasks from subtasks.parquet if it exists."""
-    subtasks_path = local_dir / DEFAULT_SUBTASKS_PATH
-    if subtasks_path.exists():
-        return pd.read_parquet(subtasks_path)
-    return None
-
-
 def write_episodes(episodes: Dataset, local_dir: Path) -> None:
     """Write episode metadata to a parquet file in the LeRobot v3.0 format.
     This function writes episode-level metadata to a single parquet file.
@@ -268,11 +259,13 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to
         dict: The batch with items converted to torch tensors.
     """
     for key in items_dict:
+        if key in {"language_persistent", "language_events"}:
+            continue
         first_item = items_dict[key][0]
         if isinstance(first_item, PILImage.Image):
             to_tensor = transforms.ToTensor()
             items_dict[key] = [to_tensor(img) for img in items_dict[key]]
-        elif first_item is None:
+        elif first_item is None or isinstance(first_item, dict):
             pass
         else:
             items_dict[key] = [x if isinstance(x, str) else torch.tensor(x) for x in items_dict[key]]
@@ -308,7 +301,11 @@ def item_to_torch(item: dict) -> dict:
         dict: Dictionary with all tensor-like items converted to torch.Tensor.
     """
     for key, val in item.items():
-        if isinstance(val, (np.ndarray | list)) and key not in ["task"]:
+        if isinstance(val, (np.ndarray | list)) and key not in [
+            "task",
+            "language_persistent",
+            "language_events",
+        ]:
             # Convert numpy arrays and lists to torch tensors
             item[key] = torch.tensor(val)
     return item

src/lerobot/datasets/language.py

@@ -0,0 +1,236 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from typing import Literal
+
+import datasets
+import pyarrow as pa
+
+LANGUAGE_PERSISTENT = "language_persistent"
+LANGUAGE_EVENTS = "language_events"
+LANGUAGE_COLUMNS = (LANGUAGE_PERSISTENT, LANGUAGE_EVENTS)
+PERSISTENT_ROW_FIELDS = ("role", "content", "style", "timestamp", "camera", "tool_calls")
+EVENT_ROW_FIELDS = ("role", "content", "style", "camera", "tool_calls")
+
+CORE_STYLES = {
+    "subtask",
+    "plan",
+    "memory",
+    "motion",
+    "interjection",
+    "vqa",
+    "trace",
+    "task_aug",
+}
+EXTENDED_STYLES = set()
+STYLE_REGISTRY = CORE_STYLES | EXTENDED_STYLES
+
+PERSISTENT_STYLES = {"subtask", "plan", "memory", "motion", "task_aug"}
+EVENT_ONLY_STYLES = {"interjection", "vqa", "trace"}
+
+# Styles whose ``content`` is grounded in a specific camera view. Rows of these
+# styles MUST carry a non-null ``camera`` referencing an ``observation.images.*``
+# feature key. Rows of every other style MUST have ``camera=None``. ``motion``
+# is intentionally NOT in this set: motion primitives are described in
+# robot-frame (joint / Cartesian) terms, not pixel space, so they are
+# camera-agnostic. ``trace`` is the pixel-trajectory event style and IS
+# view-dependent. The ``camera`` field nevertheless lives on
+# ``PERSISTENT_ROW_FIELDS`` too so the schema, validator, and resolver
+# behave symmetrically across the two columns; persistent rows simply
+# always have ``camera=None`` in practice today.
+VIEW_DEPENDENT_STYLES = {"vqa", "trace"}
+
+LanguageColumn = Literal["language_persistent", "language_events"]
+
+
+def _json_arrow_type() -> pa.DataType:
+    """Return the Arrow JSON type, falling back to ``string`` on older pyarrow."""
+    return pa.json_() if hasattr(pa, "json_") else pa.string()
+
+
+def _json_feature() -> object:
+    """Return the HF ``datasets`` JSON feature, falling back to a string value."""
+    return datasets.Json() if hasattr(datasets, "Json") else datasets.Value("string")
+
+
+def language_persistent_row_arrow_type() -> pa.StructType:
+    """Return the Arrow struct type for a single persistent language row.
+
+    Persistent rows carry their own ``timestamp`` because they represent a state
+    that became active at a specific moment and remains active until superseded.
+    """
+    return pa.struct(
+        [
+            pa.field("role", pa.string(), nullable=False),
+            pa.field("content", pa.string(), nullable=True),
+            pa.field("style", pa.string(), nullable=True),
+            pa.field("timestamp", pa.float64(), nullable=False),
+            pa.field("camera", pa.string(), nullable=True),
+            pa.field("tool_calls", pa.list_(_json_arrow_type()), nullable=True),
+        ]
+    )
+
+
+def language_event_row_arrow_type() -> pa.StructType:
+    """Return the Arrow struct type for a single event language row.
+
+    Event rows have no ``timestamp`` field: each event is stored on the dataset
+    row whose frame timestamp is the event's firing time.
+    """
+    return pa.struct(
+        [
+            pa.field("role", pa.string(), nullable=False),
+            pa.field("content", pa.string(), nullable=True),
+            pa.field("style", pa.string(), nullable=True),
+            pa.field("camera", pa.string(), nullable=True),
+            pa.field("tool_calls", pa.list_(_json_arrow_type()), nullable=True),
+        ]
+    )
+
+
+def language_persistent_arrow_type() -> pa.ListType:
+    """Return the Arrow list type for the ``language_persistent`` column."""
+    return pa.list_(language_persistent_row_arrow_type())
+
+
+def language_events_arrow_type() -> pa.ListType:
+    """Return the Arrow list type for the ``language_events`` column."""
+    return pa.list_(language_event_row_arrow_type())
+
+
+def language_persistent_row_feature() -> dict[str, object]:
+    """Return the HF ``datasets`` feature mapping for a persistent language row."""
+    return {
+        "role": datasets.Value("string"),
+        "content": datasets.Value("string"),
+        "style": datasets.Value("string"),
+        "timestamp": datasets.Value("float64"),
+        "camera": datasets.Value("string"),
+        "tool_calls": datasets.List(_json_feature()),
+    }
+
+
+def language_event_row_feature() -> dict[str, object]:
+    """Return the HF ``datasets`` feature mapping for an event language row."""
+    return {
+        "role": datasets.Value("string"),
+        "content": datasets.Value("string"),
+        "style": datasets.Value("string"),
+        "camera": datasets.Value("string"),
+        "tool_calls": datasets.List(_json_feature()),
+    }
+
+
+def language_persistent_column_feature() -> datasets.List:
+    """Return the HF ``datasets`` feature for the ``language_persistent`` column."""
+    return datasets.List(language_persistent_row_feature())
+
+
+def language_events_column_feature() -> datasets.List:
+    """Return the HF ``datasets`` feature for the ``language_events`` column."""
+    return datasets.List(language_event_row_feature())
+
+
+def language_feature_info() -> dict[str, dict]:
+    """Return the ``info["features"]`` entries for both language columns."""
+    return {
+        LANGUAGE_PERSISTENT: {"dtype": "language", "shape": (1,), "names": None},
+        LANGUAGE_EVENTS: {"dtype": "language", "shape": (1,), "names": None},
+    }
+
+
+def is_language_column(key: str) -> bool:
+    """Return ``True`` if ``key`` is one of the dataset's language column names."""
+    return key in LANGUAGE_COLUMNS
+
+
+def is_view_dependent_style(style: str | None) -> bool:
+    """Return ``True`` if rows of ``style`` must be tagged with a ``camera`` key."""
+    return style in VIEW_DEPENDENT_STYLES
+
+
+def validate_camera_field(style: str | None, camera: str | None) -> None:
+    """Enforce the ``camera`` invariant: required iff ``style`` is view-dependent.
+
+    Raises ``ValueError`` if a view-dependent style is missing ``camera`` or if
+    a non-view-dependent style carries one. Pipeline writers and the validator
+    should call this on every emitted row.
+    """
+    if is_view_dependent_style(style):
+        if not camera:
+            raise ValueError(
+                f"Rows of view-dependent style {style!r} require a non-empty 'camera' "
+                f"field referencing an 'observation.images.*' feature key."
+            )
+    elif camera is not None:
+        raise ValueError(
+            f"Rows of style {style!r} must have camera=None; got camera={camera!r}."
+        )
+
+
+# --- Tool registry --------------------------------------------------------
+# Tools declared on a dataset live in ``meta/info.json["tools"]`` as a list
+# of OpenAI-style function schemas. The runtime / training stack reads them
+# through :class:`LeRobotDatasetMetadata.tools` (with these constants as
+# fallback when the dataset doesn't declare any). Implementations live
+# under :mod:`lerobot.tools` (one file per tool); see
+# ``docs/source/tools.mdx`` for the authoring guide.
+
+SAY_TOOL_SCHEMA: dict = {
+    "type": "function",
+    "function": {
+        "name": "say",
+        "description": "Speak a short utterance to the user via the TTS executor.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "text": {
+                    "type": "string",
+                    "description": "The verbatim text to speak.",
+                }
+            },
+            "required": ["text"],
+        },
+    },
+}
+"""Canonical schema for the ``say`` tool emitted by the steerable
+annotation pipeline (PR 2 Module 2). Single source of truth — PR 2's
+writer, PR 3's runtime tool registry, and the dataset visualizer all
+import this constant rather than duplicating the dict."""
+
+DEFAULT_TOOLS: list[dict] = [SAY_TOOL_SCHEMA]
+"""Fallback tools list. Returned by ``LeRobotDatasetMetadata.tools``
+when ``meta/info.json["tools"]`` is unset, so unannotated datasets and
+chat-template consumers (``apply_chat_template(messages, tools=...)``)
+keep working out of the box."""
+
+
+def column_for_style(style: str | None) -> LanguageColumn:
+    """Map a language style to the column where rows of that style are stored.
+
+    Styles in :data:`PERSISTENT_STYLES` route to :data:`LANGUAGE_PERSISTENT`.
+    Styles in :data:`EVENT_ONLY_STYLES` and the implicit ``None`` style route
+    to :data:`LANGUAGE_EVENTS`.
+    """
+    if style is None:
+        return LANGUAGE_EVENTS
+    if style in PERSISTENT_STYLES:
+        return LANGUAGE_PERSISTENT
+    if style in EVENT_ONLY_STYLES:
+        return LANGUAGE_EVENTS
+    raise ValueError(f"Unknown language style: {style!r}")

src/lerobot/datasets/language_render.py

@@ -0,0 +1,593 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import copy
+import hashlib
+import re
+from collections.abc import Sequence
+from typing import Any
+
+from lerobot.configs.recipe import DEFAULT_BINDINGS, TrainingRecipe
+
+from .language import (
+    EVENT_ONLY_STYLES,
+    LANGUAGE_PERSISTENT,
+    PERSISTENT_STYLES,
+    column_for_style,
+)
+
+LanguageRow = dict[str, Any]
+RenderedMessages = dict[str, list[Any]]
+
+_RESOLVER_RE = re.compile(r"^(?P<name>[A-Za-z_][A-Za-z0-9_]*)\((?P<args>.*)\)$")
+_PLACEHOLDER_RE = re.compile(r"\$\{([A-Za-z_][A-Za-z0-9_]*)\}")
+
+
+def active_at(
+    t: float,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow] | None = None,
+    style: str | None = None,
+    role: str | None = None,
+    tool_name: str | None = None,
+    camera: str | None = None,
+) -> LanguageRow | None:
+    """Return the persistent row of ``style`` that is active at time ``t``.
+
+    A persistent row is "active" at ``t`` when its own ``timestamp`` is the
+    most recent one ``<= t`` for the given ``style``/``role``/``tool_name``/
+    ``camera`` selector. ``events`` is accepted for resolver-signature
+    uniformity but is not consulted: only persistent styles are valid here.
+    """
+    _validate_persistent_resolver("active_at", style)
+    matches = _matching_rows(
+        persistent, style=style, role=role, tool_name=tool_name, camera=camera
+    )
+    matches = [row for row in matches if _timestamp(row) <= t]
+    return _select_latest(
+        matches, style=style, role=role, tool_name=tool_name, camera=camera
+    )
+
+
+def emitted_at(
+    t: float,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow],
+    style: str | None = None,
+    role: str | None = None,
+    tool_name: str | None = None,
+    camera: str | None = None,
+) -> LanguageRow | None:
+    """Return the row of ``style`` emitted at exactly time ``t``.
+
+    For persistent styles, this matches persistent rows whose own ``timestamp``
+    equals ``t``. For event styles, the ``events`` list is assumed to come from
+    the dataset row at frame ``t`` (event rows carry no timestamp of their own),
+    so all matching event rows are considered emitted at ``t``. ``camera``
+    filters by the row's ``camera`` field — required to disambiguate when
+    multiple view-dependent rows share ``(t, role)`` across cameras.
+    """
+    column = column_for_style(style)
+    if column == LANGUAGE_PERSISTENT:
+        matches = [
+            row
+            for row in _matching_rows(
+                persistent, style=style, role=role, tool_name=tool_name, camera=camera
+            )
+            if _timestamp(row) == t
+        ]
+        return _select_one(
+            matches,
+            style=style,
+            role=role,
+            tool_name=tool_name,
+            camera=camera,
+            sort_key=_persistent_sort_key,
+        )
+    matches = _matching_rows(
+        events, style=style, role=role, tool_name=tool_name, camera=camera
+    )
+    return _select_one(
+        matches,
+        style=style,
+        role=role,
+        tool_name=tool_name,
+        camera=camera,
+        sort_key=_event_sort_key,
+    )
+
+
+def nth_prev(
+    t: float,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow] | None = None,
+    style: str | None = None,
+    offset: int = 1,
+    role: str | None = None,
+    tool_name: str | None = None,
+    camera: str | None = None,
+) -> LanguageRow | None:
+    """Return the persistent row that was active ``offset`` steps before ``t``.
+
+    Walks back through chronologically sorted persistent rows of ``style``
+    (filtered by optional ``role``/``tool_name``/``camera``) and returns the
+    one ``offset`` positions before the row active at ``t``. Only valid for
+    persistent styles.
+    """
+    return _nth_relative(
+        t,
+        persistent=persistent,
+        style=style,
+        offset=-offset,
+        role=role,
+        tool_name=tool_name,
+        camera=camera,
+        resolver_name="nth_prev",
+    )
+
+
+def nth_next(
+    t: float,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow] | None = None,
+    style: str | None = None,
+    offset: int = 1,
+    role: str | None = None,
+    tool_name: str | None = None,
+    camera: str | None = None,
+) -> LanguageRow | None:
+    """Return the persistent row that becomes active ``offset`` steps after ``t``.
+
+    Walks forward through chronologically sorted persistent rows of ``style``
+    (filtered by optional ``role``/``tool_name``/``camera``) and returns the
+    one ``offset`` positions after the row active at ``t``. Only valid for
+    persistent styles.
+    """
+    return _nth_relative(
+        t,
+        persistent=persistent,
+        style=style,
+        offset=offset,
+        role=role,
+        tool_name=tool_name,
+        camera=camera,
+        resolver_name="nth_next",
+    )
+
+
+def render_sample(
+    *,
+    recipe: TrainingRecipe,
+    persistent: Sequence[LanguageRow] | None,
+    events: Sequence[LanguageRow] | None,
+    t: float,
+    sample_idx: int,
+    task: str | None = None,
+    dataset_ctx: Any | None = None,
+) -> RenderedMessages | None:
+    """Render the chat-style messages for a single dataset sample.
+
+    Resolves the recipe's bindings against ``persistent`` and ``events`` rows
+    at frame timestamp ``t``, then expands the recipe's message templates.
+    Returns ``None`` if the resolved sample contains no target message.
+    """
+    persistent_rows = _normalize_rows(persistent or [])
+    event_rows = _normalize_rows(events or [])
+    selected_recipe = _select_recipe(recipe, sample_idx)
+    bindings = _resolve_bindings(
+        selected_recipe,
+        persistent=persistent_rows,
+        events=event_rows,
+        t=t,
+        sample_idx=sample_idx,
+        task=task,
+        dataset_ctx=dataset_ctx,
+    )
+    return _render_message_recipe(selected_recipe, bindings)
+
+
+def _select_recipe(recipe: TrainingRecipe, sample_idx: int) -> TrainingRecipe:
+    """Pick a deterministic blend component for ``sample_idx`` (or return ``recipe``)."""
+    if recipe.blend is None:
+        return recipe
+
+    total_weight = sum(component.weight or 0.0 for component in recipe.blend.values())
+    if total_weight <= 0:
+        raise ValueError("Blend weights must sum to a positive value.")
+
+    digest = hashlib.blake2b(str(sample_idx).encode(), digest_size=8).digest()
+    draw = int.from_bytes(digest, "big") / 2**64 * total_weight
+    cumulative = 0.0
+    last_component: TrainingRecipe | None = None
+    for component in recipe.blend.values():
+        last_component = component
+        cumulative += component.weight or 0.0
+        if draw < cumulative:
+            return component
+    assert last_component is not None
+    return last_component
+
+
+def _resolve_bindings(
+    recipe: TrainingRecipe,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow],
+    t: float,
+    sample_idx: int,
+    task: str | None,
+    dataset_ctx: Any | None,
+) -> dict[str, LanguageRow | str | None]:
+    """Resolve every binding in ``recipe`` (plus ``task``) at time ``t``."""
+    bindings: dict[str, LanguageRow | str | None] = {
+        "task": _resolve_task(
+            task, dataset_ctx, persistent=persistent, sample_idx=sample_idx
+        ),
+    }
+    specs = {**DEFAULT_BINDINGS, **(recipe.bindings or {})}
+    for name, spec in specs.items():
+        bindings[name] = _resolve_spec(spec, persistent=persistent, events=events, t=t)
+    return bindings
+
+
+def _resolve_task(
+    task: str | None,
+    dataset_ctx: Any | None,
+    *,
+    persistent: Sequence[LanguageRow] = (),
+    sample_idx: int = 0,
+) -> str | None:
+    """Return the task string for ``sample_idx``.
+
+    Resolution order:
+
+    1. Explicit ``task`` override (caller-supplied) wins.
+    2. If ``persistent`` contains rows of style ``task_aug`` (role=user),
+       deterministically pick one by ``sample_idx`` so each frame of an
+       episode rotates through the available rephrasings across an epoch.
+       This realizes Xiao 2022 / CAST-style task-prompt diversity without
+       changing ``meta/tasks.parquet`` and without forcing recipes to opt
+       in: ``${task}`` automatically picks a rephrasing when one exists,
+       and falls back to the canonical task otherwise. Recipes that want
+       the literal canonical task can override the binding.
+    3. Otherwise read the canonical task from ``dataset_ctx`` (which is
+       backed by ``meta/tasks.parquet``).
+    """
+    if task is not None:
+        return task
+
+    aug_rows = [
+        r
+        for r in persistent
+        if r.get("style") == "task_aug" and r.get("role") == "user"
+    ]
+    if aug_rows:
+        # Deterministic, blake2b-based pick keyed on sample_idx so the
+        # rotation is reproducible across runs (Python's built-in ``hash``
+        # is process-randomized).
+        digest = hashlib.blake2b(
+            f"task_aug:{sample_idx}".encode(), digest_size=8
+        ).digest()
+        idx = int.from_bytes(digest, "big") % len(aug_rows)
+        chosen = aug_rows[idx].get("content")
+        if chosen:
+            return str(chosen)
+
+    if dataset_ctx is None:
+        return None
+    if isinstance(dataset_ctx, dict):
+        return dataset_ctx.get("task")
+    return getattr(dataset_ctx, "task", None)
+
+
+def _resolve_spec(
+    spec: str,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow],
+    t: float,
+) -> LanguageRow | None:
+    """Parse a single binding's resolver expression and dispatch to its function."""
+    match = _RESOLVER_RE.match(spec.strip())
+    if match is None:
+        raise ValueError(f"Invalid resolver expression: {spec!r}")
+    name = match.group("name")
+    kwargs = _parse_resolver_args(match.group("args"))
+    kwargs.pop("t_arg", None)
+
+    resolvers = {
+        "active_at": active_at,
+        "emitted_at": emitted_at,
+        "nth_prev": nth_prev,
+        "nth_next": nth_next,
+    }
+    if name not in resolvers:
+        raise ValueError(f"Unknown language resolver: {name!r}")
+    return resolvers[name](t, persistent=persistent, events=events, **kwargs)
+
+
+def _parse_resolver_args(args: str) -> dict[str, Any]:
+    """Parse a comma-separated resolver argument list into a kwargs dict."""
+    kwargs: dict[str, Any] = {}
+    if not args.strip():
+        return kwargs
+
+    parts = [part.strip() for part in args.split(",") if part.strip()]
+    for part in parts:
+        if part == "t":
+            kwargs["t_arg"] = True
+            continue
+        if "=" not in part:
+            raise ValueError(f"Invalid resolver argument: {part!r}")
+        key, value = (item.strip() for item in part.split("=", 1))
+        if key == "offset":
+            kwargs[key] = int(value)
+        else:
+            kwargs[key] = value.strip("\"'")
+    return kwargs
+
+
+def _render_message_recipe(
+    recipe: TrainingRecipe,
+    bindings: dict[str, LanguageRow | str | None],
+) -> RenderedMessages | None:
+    """Expand ``recipe.messages`` into rendered chat messages using ``bindings``."""
+    assert recipe.messages is not None
+    messages: list[dict[str, Any]] = []
+    streams: list[str | None] = []
+    target_indices: list[int] = []
+
+    for turn in recipe.messages:
+        if turn.if_present is not None and bindings.get(turn.if_present) is None:
+            continue
+
+        message = {"role": turn.role}
+        if turn.content is not None:
+            message["content"] = _render_content(turn.content, bindings)
+
+        if turn.tool_calls_from is not None:
+            row = bindings.get(turn.tool_calls_from)
+            tool_calls = row.get("tool_calls") if isinstance(row, dict) else None
+            if tool_calls:
+                message["tool_calls"] = copy.deepcopy(tool_calls)
+
+        message_idx = len(messages)
+        messages.append(message)
+        streams.append(turn.stream)
+        if turn.target:
+            target_indices.append(message_idx)
+
+    if not target_indices:
+        return None
+
+    rendered = {
+        "messages": messages,
+        "message_streams": streams,
+        "target_message_indices": target_indices,
+    }
+    _validate_rendered(rendered)
+    return rendered
+
+
+def _render_content(
+    content: str | list[dict[str, Any]],
+    bindings: dict[str, LanguageRow | str | None],
+) -> str | list[dict[str, Any]]:
+    """Substitute bindings into a string or each string field of multimodal blocks."""
+    if isinstance(content, str):
+        return _substitute(content, bindings)
+
+    rendered_blocks = []
+    for block in content:
+        rendered_block = copy.deepcopy(block)
+        for key, value in rendered_block.items():
+            if isinstance(value, str):
+                rendered_block[key] = _substitute(value, bindings)
+        rendered_blocks.append(rendered_block)
+    return rendered_blocks
+
+
+def _substitute(template: str, bindings: dict[str, LanguageRow | str | None]) -> str:
+    """Replace ``${name}`` placeholders in ``template`` with their bound values."""
+
+    def replace(match: re.Match[str]) -> str:
+        """Resolve a single ``${name}`` match to its bound string value."""
+        name = match.group(1)
+        if name not in bindings:
+            raise ValueError(f"Unknown template binding: {name!r}")
+        value = bindings[name]
+        if value is None:
+            return ""
+        if isinstance(value, dict):
+            content = value.get("content")
+            return "" if content is None else str(content)
+        return str(value)
+
+    return _PLACEHOLDER_RE.sub(replace, template)
+
+
+def _validate_rendered(rendered: RenderedMessages) -> None:
+    """Sanity-check the rendered output for stream/target alignment."""
+    messages = rendered["messages"]
+    streams = rendered["message_streams"]
+    target_indices = rendered["target_message_indices"]
+
+    if len(streams) != len(messages):
+        raise ValueError("message_streams must be aligned with messages.")
+    if not target_indices:
+        raise ValueError("Rendered samples must contain at least one target message.")
+    for idx in target_indices:
+        if idx < 0 or idx >= len(messages):
+            raise ValueError(f"Target message index {idx} is out of bounds.")
+    for idx, stream in enumerate(streams):
+        if stream is None:
+            raise ValueError(f"Rendered message {idx} has no stream.")
+
+
+def _nth_relative(
+    t: float,
+    *,
+    persistent: Sequence[LanguageRow],
+    style: str | None,
+    offset: int,
+    role: str | None,
+    tool_name: str | None,
+    camera: str | None,
+    resolver_name: str,
+) -> LanguageRow | None:
+    """Shared body for ``nth_prev`` / ``nth_next`` with signed ``offset``."""
+    _validate_persistent_resolver(resolver_name, style)
+    if abs(offset) < 1:
+        raise ValueError(f"{resolver_name} offset must be non-zero.")
+
+    rows = sorted(
+        _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera),
+        key=_persistent_sort_key,
+    )
+    if not rows:
+        return None
+
+    anchor_idx = None
+    for idx, row in enumerate(rows):
+        if _timestamp(row) <= t:
+            anchor_idx = idx
+        else:
+            break
+
+    target_idx = (offset - 1 if offset > 0 else None) if anchor_idx is None else anchor_idx + offset
+
+    if target_idx is None or target_idx < 0 or target_idx >= len(rows):
+        return None
+    return rows[target_idx]
+
+
+def _validate_persistent_resolver(resolver_name: str, style: str | None) -> None:
+    """Reject calls with missing or event-only ``style`` for persistent resolvers."""
+    if style is None:
+        raise ValueError(f"{resolver_name} requires a persistent style.")
+    if style in EVENT_ONLY_STYLES:
+        raise ValueError(f"{resolver_name} cannot be used with event-only style {style!r}.")
+    if style not in PERSISTENT_STYLES:
+        column_for_style(style)
+
+
+def _matching_rows(
+    rows: Sequence[LanguageRow],
+    *,
+    style: str | None,
+    role: str | None,
+    tool_name: str | None,
+    camera: str | None,
+) -> list[LanguageRow]:
+    """Return ``rows`` filtered by optional ``style``/``role``/``tool_name``/``camera`` selectors."""
+    return [
+        row
+        for row in rows
+        if (style is None or row.get("style") == style)
+        and (role is None or row.get("role") == role)
+        and (tool_name is None or _row_has_tool_name(row, tool_name))
+        and (camera is None or row.get("camera") == camera)
+    ]
+
+
+def _select_latest(
+    rows: Sequence[LanguageRow],
+    *,
+    style: str | None,
+    role: str | None,
+    tool_name: str | None,
+    camera: str | None,
+) -> LanguageRow | None:
+    """Return the row tied for the latest ``timestamp`` (disambiguated by selectors)."""
+    if not rows:
+        return None
+    rows = sorted(rows, key=_persistent_sort_key)
+    latest_ts = _timestamp(rows[-1])
+    return _select_one(
+        [row for row in rows if _timestamp(row) == latest_ts],
+        style=style,
+        role=role,
+        tool_name=tool_name,
+        camera=camera,
+        sort_key=_persistent_sort_key,
+    )
+
+
+def _select_one(
+    rows: Sequence[LanguageRow],
+    *,
+    style: str | None,
+    role: str | None,
+    tool_name: str | None,
+    camera: str | None,
+    sort_key: Any,
+) -> LanguageRow | None:
+    """Return the single matching row, or raise if the selectors are ambiguous."""
+    if not rows:
+        return None
+    if len(rows) > 1 and role is None and tool_name is None and camera is None:
+        raise ValueError(
+            f"Ambiguous resolver for style={style!r}; add role=..., tool_name=..., "
+            f"or camera=... to disambiguate."
+        )
+    return sorted(rows, key=sort_key)[0]
+
+
+def _persistent_sort_key(row: LanguageRow) -> tuple[float, str, str]:
+    """Sort key for persistent rows: ``(timestamp, style, role)``."""
+    return (_timestamp(row), row.get("style") or "", row.get("role") or "")
+
+
+def _event_sort_key(row: LanguageRow) -> tuple[str, str]:
+    """Sort key for event rows: ``(style, role)`` (timestamp is implicit in the frame)."""
+    return (row.get("style") or "", row.get("role") or "")
+
+
+def _timestamp(row: LanguageRow) -> float:
+    """Extract a row's ``timestamp`` as a Python float (unwrapping numpy scalars)."""
+    value = row["timestamp"]
+    return float(value.item() if hasattr(value, "item") else value)
+
+
+def _row_has_tool_name(row: LanguageRow, tool_name: str) -> bool:
+    """Return ``True`` if any of the row's tool calls invokes ``tool_name``."""
+    for tool_call in row.get("tool_calls") or []:
+        if isinstance(tool_call, str):
+            continue
+        function = tool_call.get("function") if isinstance(tool_call, dict) else None
+        if isinstance(function, dict) and function.get("name") == tool_name:
+            return True
+    return False
+
+
+def _normalize_rows(rows: Sequence[Any]) -> list[LanguageRow]:
+    """Convert pyarrow scalars / mappings into a fresh list of plain dict rows."""
+    normalized = []
+    for row in rows:
+        if row is None:
+            continue
+        if hasattr(row, "as_py"):
+            row = row.as_py()
+        if not isinstance(row, dict):
+            raise TypeError(f"Language rows must be dictionaries, got {type(row).__name__}.")
+        normalized.append(dict(row))
+    return normalized

src/lerobot/datasets/utils.py

@@ -83,7 +83,6 @@ VIDEO_DIR = "videos"
 
 CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
 DEFAULT_TASKS_PATH = "meta/tasks.parquet"
-DEFAULT_SUBTASKS_PATH = "meta/subtasks.parquet"
 DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"

src/lerobot/envs/configs.py

@@ -331,6 +331,7 @@ class LiberoEnv(EnvConfig):
     camera_name_mapping: dict[str, str] | None = None
     observation_height: int = 360
     observation_width: int = 360
+    is_libero_plus: bool = False
     features: dict[str, PolicyFeature] = field(
         default_factory=lambda: {
             ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
@@ -432,6 +433,7 @@ class LiberoEnv(EnvConfig):
             control_mode=self.control_mode,
             episode_length=self.episode_length,
             camera_name_mapping=self.camera_name_mapping,
+            is_libero_plus=self.is_libero_plus,
         )
 
     def get_env_processors(self):
@@ -571,6 +573,71 @@ class RoboCasaEnv(EnvConfig):
         )
 
 
+@EnvConfig.register_subclass("vlabench")
+@dataclass
+class VLABenchEnv(EnvConfig):
+    task: str = "select_fruit"
+    fps: int = 10
+    episode_length: int = 500
+    obs_type: str = "pixels_agent_pos"
+    render_mode: str = "rgb_array"
+    render_resolution: tuple[int, int] = (480, 480)
+    robot: str = "franka"
+    action_mode: str = "eef"
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
+        }
+    )
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            ACTION: ACTION,
+            "agent_pos": OBS_STATE,
+            "pixels/image": f"{OBS_IMAGES}.image",
+            "pixels/second_image": f"{OBS_IMAGES}.second_image",
+            "pixels/wrist_image": f"{OBS_IMAGES}.wrist_image",
+        }
+    )
+
+    def __post_init__(self):
+        h, w = self.render_resolution
+        if self.obs_type == "pixels":
+            self.features["pixels/image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+            self.features["pixels/second_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+            self.features["pixels/wrist_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+        elif self.obs_type == "pixels_agent_pos":
+            self.features["pixels/image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+            self.features["pixels/second_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+            self.features["pixels/wrist_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(7,))
+        else:
+            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+            "render_resolution": self.render_resolution,
+            "robot": self.robot,
+            "max_episode_steps": self.episode_length,
+            "action_mode": self.action_mode,
+        }
+
+    def create_envs(self, n_envs: int, use_async_envs: bool = False):
+        from .vlabench import create_vlabench_envs
+
+        if self.task is None:
+            raise ValueError("VLABenchEnv requires a task to be specified")
+        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
+        return create_vlabench_envs(
+            task=self.task,
+            n_envs=n_envs,
+            gym_kwargs=self.gym_kwargs,
+            env_cls=env_cls,
+        )
+
+
 @EnvConfig.register_subclass("isaaclab_arena")
 @dataclass
 class IsaaclabArenaEnv(HubEnvConfig):
@@ -649,3 +716,171 @@ class IsaaclabArenaEnv(HubEnvConfig):
             ),
             PolicyProcessorPipeline(steps=[]),
         )
+
+
+@EnvConfig.register_subclass("libero_plus")
+@dataclass
+class LiberoPlusEnv(LiberoEnv):
+    """Config for LIBERO-plus robustness benchmark evaluation.
+
+    LIBERO-plus extends LIBERO with 7 perturbation dimensions (camera viewpoints,
+    object layouts, robot initial states, language instructions, lighting, background
+    textures, sensor noise) producing ~10k task variants.
+
+    The gym interface is identical to LIBERO so this class reuses ``LiberoEnv``
+    entirely — only the registered name and default task suite differ.
+
+    Install: see docker/Dockerfile.benchmark.libero_plus — LIBERO-plus ships
+    as a namespace package from a git fork and must be cloned + PYTHONPATH'd
+    rather than installed as a pyproject extra.
+
+    See Also:
+        https://github.com/sylvestf/LIBERO-plus
+    """
+
+    task: str = "libero_spatial"
+    is_libero_plus: bool = True
+
+
+@EnvConfig.register_subclass("robotwin")
+@dataclass
+class RoboTwinEnvConfig(EnvConfig):
+    """Configuration for RoboTwin 2.0 benchmark environments.
+
+    RoboTwin 2.0 is a dual-arm manipulation benchmark with 50 tasks built on the
+    SAPIEN simulator. The robot is an Aloha-AgileX bimanual platform with 14 DOF
+    (7 per arm). All three cameras are enabled by default.
+
+    See: https://robotwin-platform.github.io
+    Dataset: https://huggingface.co/datasets/lerobot/robotwin_unified
+    """
+
+    task: str = "beat_block_hammer"  # single task or comma-separated list
+    fps: int = 25
+    episode_length: int = 300
+    obs_type: str = "pixels_agent_pos"
+    render_mode: str = "rgb_array"
+    # Available cameras from RoboTwin's aloha-agilex embodiment: head_camera
+    # (torso-mounted) + left_camera / right_camera (wrists).
+    camera_names: str = "head_camera,left_camera,right_camera"
+    # Match the D435 dims in task_config/demo_clean.yml (_camera_config.yml).
+    # Gym's vector-env concatenate pre-allocates buffers of this shape, so it
+    # must equal what SAPIEN actually renders.
+    observation_height: int = 240
+    observation_width: int = 320
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
+        }
+    )
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            ACTION: ACTION,
+            "pixels/head_camera": f"{OBS_IMAGES}.head_camera",
+            "pixels/left_camera": f"{OBS_IMAGES}.left_camera",
+            "pixels/right_camera": f"{OBS_IMAGES}.right_camera",
+            "agent_pos": OBS_STATE,
+        }
+    )
+
+    def __post_init__(self):
+        cam_list = [c.strip() for c in self.camera_names.split(",") if c.strip()]
+        for cam in cam_list:
+            self.features[f"pixels/{cam}"] = PolicyFeature(
+                type=FeatureType.VISUAL,
+                shape=(self.observation_height, self.observation_width, 3),
+            )
+            # Keep features_map entry if already set (default_factory); add if missing.
+            key = f"pixels/{cam}"
+            if key not in self.features_map:
+                self.features_map[key] = f"{OBS_IMAGES}.{cam}"
+
+        if self.obs_type == "pixels_agent_pos":
+            self.features["agent_pos"] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(14,),  # 14 DOF: 7 per arm
+            )
+        elif self.obs_type != "pixels":
+            raise ValueError(
+                f"Unsupported obs_type '{self.obs_type}'. "
+                "RoboTwinEnvConfig supports 'pixels' and 'pixels_agent_pos'."
+            )
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {}
+
+    def create_envs(self, n_envs: int, use_async_envs: bool = True):
+        from lerobot.envs.robotwin import create_robotwin_envs
+
+        if not self.task:
+            raise ValueError("RoboTwinEnvConfig requires `task` to be specified.")
+
+        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
+        cam_list = [c.strip() for c in self.camera_names.split(",") if c.strip()]
+        return create_robotwin_envs(
+            task=self.task,
+            n_envs=n_envs,
+            env_cls=env_cls,
+            camera_names=cam_list,
+            observation_height=self.observation_height,
+            observation_width=self.observation_width,
+            episode_length=self.episode_length,
+        )
+
+
+@EnvConfig.register_subclass("robomme")
+@dataclass
+class RoboMMEEnv(EnvConfig):
+    """RoboMME memory-augmented manipulation benchmark (ManiSkill/SAPIEN).
+
+    16 tasks across 4 suites: Counting, Permanence, Reference, Imitation.
+    Dataset: lerobot/robomme (LeRobot v3.0, 1,600 episodes).
+    Benchmark: https://github.com/RoboMME/robomme_benchmark
+
+    Requires the `robomme` git package installed separately (Linux only);
+    see docker/Dockerfile.benchmark.robomme for the canonical install.
+    """
+
+    task: str = "PickXtimes"
+    fps: int = 10
+    episode_length: int = 300
+    action_space: str = "joint_angle"  # or "ee_pose" (7-D)
+    dataset_split: str = "test"  # "train" | "val" | "test"
+    task_ids: list[int] | None = None
+    features: dict[str, PolicyFeature] = field(default_factory=dict)
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            ACTION: ACTION,
+            "pixels/image": f"{OBS_IMAGES}.image",
+            "pixels/wrist_image": f"{OBS_IMAGES}.wrist_image",
+            "agent_pos": OBS_STATE,
+        }
+    )
+
+    def __post_init__(self):
+        action_dim = 8 if self.action_space == "joint_angle" else 7
+        self.features = {
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_dim,)),
+            "pixels/image": PolicyFeature(type=FeatureType.VISUAL, shape=(256, 256, 3)),
+            "pixels/wrist_image": PolicyFeature(type=FeatureType.VISUAL, shape=(256, 256, 3)),
+            "agent_pos": PolicyFeature(type=FeatureType.STATE, shape=(8,)),
+        }
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {}
+
+    def create_envs(self, n_envs: int, use_async_envs: bool = True):
+        from lerobot.envs.robomme import create_robomme_envs
+
+        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
+        return create_robomme_envs(
+            task=self.task,
+            n_envs=n_envs,
+            action_space_type=self.action_space,
+            dataset=self.dataset_split,
+            episode_length=self.episode_length,
+            task_ids=self.task_ids,
+            env_cls=env_cls,
+        )

src/lerobot/envs/libero.py

@@ -16,6 +16,7 @@
 from __future__ import annotations
 
 import os
+import re
 from collections import defaultdict
 from collections.abc import Callable, Iterable, Mapping, Sequence
 from functools import partial
@@ -56,14 +57,34 @@ def _select_task_ids(total_tasks: int, task_ids: Iterable[int] | None) -> list[i
     return ids
 
 
-def get_task_init_states(task_suite: Any, i: int) -> np.ndarray:
-    init_states_path = (
-        Path(get_libero_path("init_states"))
-        / task_suite.tasks[i].problem_folder
-        / task_suite.tasks[i].init_states_file
-    )
-    init_states = torch.load(init_states_path, weights_only=False)  # nosec B614
-    return init_states
+# LIBERO-plus perturbation variants encode the perturbation in the filename
+# but on disk only the base `.pruned_init` exists — strip the suffix to match
+# LIBERO-plus's own suite.get_task_init_states() (we reimplement it here so we
+# can pass weights_only=False for PyTorch 2.6+ numpy pickles).
+_LIBERO_PERTURBATION_SUFFIX_RE = re.compile(r"_(?:language|view|light)_[^.]*|_(?:table|tb)_\d+")
+
+
+def get_task_init_states(task_suite: Any, i: int, is_libero_plus: bool = False) -> np.ndarray:
+    task = task_suite.tasks[i]
+    filename = Path(task.init_states_file)
+    root = Path(get_libero_path("init_states"))
+
+    if not is_libero_plus:
+        init_states_path = root / task.problem_folder / filename.name
+        return torch.load(init_states_path, weights_only=False)  # nosec B614
+
+    # LIBERO-plus: `_add_` / `_level` variants store extra-object layouts under
+    # libero_newobj/ as a flat array that must be reshaped to (1, -1).
+    if "_add_" in filename.name or "_level" in filename.name:
+        init_states_path = root / "libero_newobj" / task.problem_folder / filename.name
+        init_states = torch.load(init_states_path, weights_only=False)  # nosec B614
+        return init_states.reshape(1, -1)
+
+    # LIBERO-plus perturbation variants encode the perturbation in the filename
+    # but on disk only the base `.pruned_init` exists — strip the suffix to match.
+    stripped = _LIBERO_PERTURBATION_SUFFIX_RE.sub("", filename.stem) + filename.suffix
+    init_states_path = root / task.problem_folder / stripped
+    return torch.load(init_states_path, weights_only=False)  # nosec B614
 
 
 def get_libero_dummy_action():
@@ -105,9 +126,11 @@ class LiberoEnv(gym.Env):
         camera_name_mapping: dict[str, str] | None = None,
         num_steps_wait: int = 10,
         control_mode: str = "relative",
+        is_libero_plus: bool = False,
     ):
         super().__init__()
         self.task_id = task_id
+        self.is_libero_plus = is_libero_plus
         self.obs_type = obs_type
         self.render_mode = render_mode
         self.observation_width = observation_width
@@ -134,7 +157,11 @@ class LiberoEnv(gym.Env):
         self.episode_index = episode_index
         self.episode_length = episode_length
         # Load once and keep
-        self._init_states = get_task_init_states(task_suite, self.task_id) if self.init_states else None
+        self._init_states = (
+            get_task_init_states(task_suite, self.task_id, is_libero_plus=self.is_libero_plus)
+            if self.init_states
+            else None
+        )
         self._reset_stride = n_envs  # when performing a reset, append `_reset_stride` to `init_state_id`.
 
         self.init_state_id = self.episode_index  # tie each sub-env to a fixed init state
@@ -367,6 +394,7 @@ def _make_env_fns(
     gym_kwargs: Mapping[str, Any],
     control_mode: str,
     camera_name_mapping: dict[str, str] | None = None,
+    is_libero_plus: bool = False,
 ) -> list[Callable[[], LiberoEnv]]:
     """Build n_envs factory callables for a single (suite, task_id)."""
 
@@ -383,6 +411,7 @@ def _make_env_fns(
             n_envs=n_envs,
             control_mode=control_mode,
             camera_name_mapping=camera_name_mapping,
+            is_libero_plus=is_libero_plus,
             **local_kwargs,
         )
 
@@ -405,6 +434,7 @@ def create_libero_envs(
     control_mode: str = "relative",
     episode_length: int | None = None,
     camera_name_mapping: dict[str, str] | None = None,
+    is_libero_plus: bool = False,
 ) -> dict[str, dict[int, Any]]:
     """
     Create vectorized LIBERO environments with a consistent return shape.
@@ -463,6 +493,7 @@ def create_libero_envs(
                 gym_kwargs=gym_kwargs,
                 control_mode=control_mode,
                 camera_name_mapping=camera_name_mapping,
+                is_libero_plus=is_libero_plus,
             )
             if is_async:
                 lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)

src/lerobot/envs/robomme.py

@@ -0,0 +1,245 @@
+"""RoboMME environment wrapper for LeRobot evaluation.
+
+Wraps the RoboMME ``BenchmarkEnvBuilder`` into a Gymnasium-compatible
+``VectorEnv`` suitable for ``lerobot_eval``.
+
+RoboMME tasks:
+  Counting:    BinFill, PickXtimes, SwingXtimes, StopCube
+  Permanence:  VideoUnmask, VideoUnmaskSwap, ButtonUnmask, ButtonUnmaskSwap
+  Reference:   PickHighlight, VideoRepick, VideoPlaceButton, VideoPlaceOrder
+  Imitation:   MoveCube, InsertPeg, PatternLock, RouteStick
+
+Dataset: lerobot/robomme (LeRobot v3.0, 1,600 episodes)
+Install: see docker/Dockerfile.benchmark.robomme  (Linux only — mani-skill vs numpy pin conflict)
+Benchmark: https://github.com/RoboMME/robomme_benchmark
+"""
+
+from __future__ import annotations
+
+from collections.abc import Callable, Sequence
+from functools import partial
+from typing import Any
+
+import gymnasium as gym
+import numpy as np
+from gymnasium import spaces
+
+from .utils import _LazyAsyncVectorEnv
+
+ROBOMME_TASKS = [
+    "BinFill",
+    "PickXtimes",
+    "SwingXtimes",
+    "StopCube",
+    "VideoUnmask",
+    "VideoUnmaskSwap",
+    "ButtonUnmask",
+    "ButtonUnmaskSwap",
+    "PickHighlight",
+    "VideoRepick",
+    "VideoPlaceButton",
+    "VideoPlaceOrder",
+    "MoveCube",
+    "InsertPeg",
+    "PatternLock",
+    "RouteStick",
+]
+
+
+class RoboMMEGymEnv(gym.Env):
+    """Thin Gymnasium wrapper around a single RoboMME episode env."""
+
+    metadata = {"render_modes": ["rgb_array"], "render_fps": 10}
+
+    def __init__(
+        self,
+        task: str = "PickXtimes",
+        action_space_type: str = "joint_angle",
+        dataset: str = "test",
+        episode_idx: int = 0,
+        max_steps: int = 300,
+    ):
+        super().__init__()
+        from robomme.env_record_wrapper import BenchmarkEnvBuilder
+
+        self._task = task
+        self._action_space_type = action_space_type
+        self._dataset = dataset
+        self._episode_idx = episode_idx
+        self._max_steps = max_steps
+        self._max_episode_steps = max_steps
+
+        self._builder = BenchmarkEnvBuilder(
+            env_id=task,
+            dataset=dataset,
+            action_space=action_space_type,
+            gui_render=False,
+            max_steps=max_steps,
+        )
+        self._env = None
+        self._last_raw_obs: dict | None = None
+
+        action_dim = 8 if action_space_type == "joint_angle" else 7
+        self.action_space = spaces.Box(low=-1.0, high=1.0, shape=(action_dim,), dtype=np.float32)
+        # `pixels` must be a nested Dict so `preprocess_observation()` in
+        # envs/utils.py picks it up and maps each camera to
+        # `observation.images.<cam>`. A flat layout (`pixels/image`,
+        # `pixels/wrist_image`) silently drops every image from the batch.
+        self.observation_space = spaces.Dict(
+            {
+                "pixels": spaces.Dict(
+                    {
+                        "image": spaces.Box(0, 255, shape=(256, 256, 3), dtype=np.uint8),
+                        "wrist_image": spaces.Box(0, 255, shape=(256, 256, 3), dtype=np.uint8),
+                    }
+                ),
+                "agent_pos": spaces.Box(-np.inf, np.inf, shape=(8,), dtype=np.float32),
+            }
+        )
+
+    def reset(self, *, seed=None, options=None):
+        super().reset(seed=seed)
+        self._env = self._builder.make_env_for_episode(
+            episode_idx=self._episode_idx,
+            max_steps=self._max_steps,
+        )
+        obs, info = self._env.reset()
+        self._last_raw_obs = obs
+        return self._convert_obs(obs), self._convert_info(info)
+
+    def step(self, action):
+        obs, reward, terminated, truncated, info = self._env.step(action)
+        self._last_raw_obs = obs
+
+        terminated_bool = bool(terminated.item()) if hasattr(terminated, "item") else bool(terminated)
+        truncated_bool = bool(truncated.item()) if hasattr(truncated, "item") else bool(truncated)
+
+        status = info.get("status", "ongoing")
+        is_success = status == "success"
+        conv_info = self._convert_info(info)
+        conv_info["is_success"] = is_success
+
+        return self._convert_obs(obs), float(reward), terminated_bool, truncated_bool, conv_info
+
+    def render(self) -> np.ndarray | None:
+        """Return the front camera image from the last observation for video recording."""
+        if self._last_raw_obs is None:
+            return np.zeros((256, 256, 3), dtype=np.uint8)
+        front = self._last_raw_obs.get("front_rgb_list")
+        if front is None:
+            return np.zeros((256, 256, 3), dtype=np.uint8)
+        frame = front[-1] if isinstance(front, list) else front
+        return np.asarray(frame, dtype=np.uint8)
+
+    def _convert_obs(self, obs: dict) -> dict:
+        front_rgb = (
+            obs["front_rgb_list"][-1] if isinstance(obs["front_rgb_list"], list) else obs["front_rgb_list"]
+        )
+        wrist_rgb = (
+            obs["wrist_rgb_list"][-1] if isinstance(obs["wrist_rgb_list"], list) else obs["wrist_rgb_list"]
+        )
+        joint_state = (
+            obs["joint_state_list"][-1]
+            if isinstance(obs["joint_state_list"], list)
+            else obs["joint_state_list"]
+        )
+        gripper_state = (
+            obs["gripper_state_list"][-1]
+            if isinstance(obs["gripper_state_list"], list)
+            else obs["gripper_state_list"]
+        )
+
+        front_rgb = np.asarray(front_rgb, dtype=np.uint8)
+        wrist_rgb = np.asarray(wrist_rgb, dtype=np.uint8)
+        joint = np.asarray(joint_state, dtype=np.float32).flatten()[:7]
+        gripper = np.asarray(gripper_state, dtype=np.float32).flatten()[:1]
+        state = np.concatenate([joint, gripper])
+
+        return {
+            "pixels": {"image": front_rgb, "wrist_image": wrist_rgb},
+            "agent_pos": state,
+        }
+
+    def _convert_info(self, info: dict) -> dict:
+        return {
+            "status": info.get("status", "ongoing"),
+            "task_goal": info.get("task_goal", ""),
+        }
+
+
+def _make_env_fns(
+    *,
+    task: str,
+    n_envs: int,
+    action_space_type: str,
+    dataset: str,
+    episode_length: int,
+    task_id: int,
+) -> list[Callable[[], RoboMMEGymEnv]]:
+    """Build n_envs factory callables for one RoboMME task id."""
+
+    def _make_one(episode_index: int) -> RoboMMEGymEnv:
+        return RoboMMEGymEnv(
+            task=task,
+            action_space_type=action_space_type,
+            dataset=dataset,
+            episode_idx=episode_index,
+            max_steps=episode_length,
+        )
+
+    return [partial(_make_one, task_id + i) for i in range(n_envs)]
+
+
+def create_robomme_envs(
+    task: str,
+    n_envs: int = 1,
+    action_space_type: str = "joint_angle",
+    dataset: str = "test",
+    episode_length: int = 300,
+    task_ids: list[int] | None = None,
+    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
+) -> dict[str, dict[int, gym.vector.VectorEnv]]:
+    """Create vectorized RoboMME environments for evaluation.
+
+    `task` may be a single RoboMME task name (e.g. "PickXtimes") or a
+    comma-separated list (e.g. "PickXtimes,BinFill,StopCube"). Each task
+    becomes its own suite in the returned mapping.
+
+    Returns {suite_name: {task_id: VectorEnv}} matching lerobot's expected format.
+    """
+    if env_cls is None or not callable(env_cls):
+        raise ValueError("env_cls must be a callable that wraps a list of env factory callables.")
+    if not isinstance(n_envs, int) or n_envs <= 0:
+        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
+
+    if task_ids is None:
+        task_ids = [0]
+
+    task_names = [t.strip() for t in task.split(",") if t.strip()]
+    is_async = env_cls is gym.vector.AsyncVectorEnv
+    cached_obs_space: spaces.Space | None = None
+    cached_act_space: spaces.Space | None = None
+    cached_metadata: dict[str, Any] | None = None
+    out: dict[str, dict[int, gym.vector.VectorEnv]] = {}
+    for task_name in task_names:
+        envs_by_task: dict[int, gym.vector.VectorEnv] = {}
+        for task_id in task_ids:
+            fns = _make_env_fns(
+                task=task_name,
+                n_envs=n_envs,
+                action_space_type=action_space_type,
+                dataset=dataset,
+                episode_length=episode_length,
+                task_id=task_id,
+            )
+            if is_async:
+                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
+                if cached_obs_space is None:
+                    cached_obs_space = lazy.observation_space
+                    cached_act_space = lazy.action_space
+                    cached_metadata = lazy.metadata
+                envs_by_task[task_id] = lazy
+            else:
+                envs_by_task[task_id] = env_cls(fns)
+        out[task_name] = envs_by_task
+    return out

src/lerobot/envs/robotwin.py

@@ -0,0 +1,488 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import annotations
+
+import importlib
+import logging
+from collections import defaultdict
+from collections.abc import Callable, Sequence
+from functools import partial
+from typing import Any
+
+import gymnasium as gym
+import numpy as np
+import torch
+from gymnasium import spaces
+
+from lerobot.types import RobotObservation
+
+from .utils import _LazyAsyncVectorEnv
+
+logger = logging.getLogger(__name__)
+
+# Camera names as used by RoboTwin 2.0. The wrapper appends "_rgb" when looking
+# up keys in get_obs() output (e.g. "head_camera" → "head_camera_rgb").
+ROBOTWIN_CAMERA_NAMES: tuple[str, ...] = (
+    "head_camera",
+    "left_camera",
+    "right_camera",
+)
+
+ACTION_DIM = 14  # 7 DOF × 2 arms
+ACTION_LOW = -1.0
+ACTION_HIGH = 1.0
+DEFAULT_EPISODE_LENGTH = 300
+# D435 dims from task_config/_camera_config.yml (what demo_clean.yml selects).
+DEFAULT_CAMERA_H = 240
+DEFAULT_CAMERA_W = 320
+
+# Task list from RoboTwin 2.0's `envs/` directory — mirrors upstream exactly
+# (50 tasks as of main; earlier revisions had 60 with a different split).
+# Keep this in sync with:
+#   gh api /repos/RoboTwin-Platform/RoboTwin/contents/envs --paginate \
+#     | jq -r '.[].name' | grep -E '\.py$' | grep -v '^_' | sed 's/\.py$//'
+ROBOTWIN_TASKS: tuple[str, ...] = (
+    "adjust_bottle",
+    "beat_block_hammer",
+    "blocks_ranking_rgb",
+    "blocks_ranking_size",
+    "click_alarmclock",
+    "click_bell",
+    "dump_bin_bigbin",
+    "grab_roller",
+    "handover_block",
+    "handover_mic",
+    "hanging_mug",
+    "lift_pot",
+    "move_can_pot",
+    "move_pillbottle_pad",
+    "move_playingcard_away",
+    "move_stapler_pad",
+    "open_laptop",
+    "open_microwave",
+    "pick_diverse_bottles",
+    "pick_dual_bottles",
+    "place_a2b_left",
+    "place_a2b_right",
+    "place_bread_basket",
+    "place_bread_skillet",
+    "place_burger_fries",
+    "place_can_basket",
+    "place_cans_plasticbox",
+    "place_container_plate",
+    "place_dual_shoes",
+    "place_empty_cup",
+    "place_fan",
+    "place_mouse_pad",
+    "place_object_basket",
+    "place_object_scale",
+    "place_object_stand",
+    "place_phone_stand",
+    "place_shoe",
+    "press_stapler",
+    "put_bottles_dustbin",
+    "put_object_cabinet",
+    "rotate_qrcode",
+    "scan_object",
+    "shake_bottle",
+    "shake_bottle_horizontally",
+    "stack_blocks_three",
+    "stack_blocks_two",
+    "stack_bowls_three",
+    "stack_bowls_two",
+    "stamp_seal",
+    "turn_switch",
+)
+
+
+_ROBOTWIN_SETUP_CACHE: dict[str, dict[str, Any]] = {}
+
+
+def _load_robotwin_setup_kwargs(task_name: str) -> dict[str, Any]:
+    """Build the kwargs dict RoboTwin's setup_demo expects.
+
+    Mirrors the config loading done by RoboTwin's ``script/eval_policy.py``:
+    reads ``task_config/demo_clean.yml``, resolves the embodiment file from
+    ``_embodiment_config.yml``, loads the robot's own ``config.yml``, and
+    reads camera dimensions from ``_camera_config.yml``.
+
+    Uses ``aloha-agilex`` single-robot dual-arm by default (the only embodiment
+    used by beat_block_hammer and most smoke-test tasks).
+    """
+    if task_name in _ROBOTWIN_SETUP_CACHE:
+        return dict(_ROBOTWIN_SETUP_CACHE[task_name])
+
+    import os
+
+    import yaml  # type: ignore[import-untyped]
+    from envs import CONFIGS_PATH  # type: ignore[import-not-found]
+
+    task_config = "demo_clean"
+    with open(os.path.join(CONFIGS_PATH, f"{task_config}.yml"), encoding="utf-8") as f:
+        args = yaml.safe_load(f)
+
+    # Resolve embodiment — demo_clean.yml uses [aloha-agilex] (dual-arm single robot)
+    with open(os.path.join(CONFIGS_PATH, "_embodiment_config.yml"), encoding="utf-8") as f:
+        embodiment_types = yaml.safe_load(f)
+    embodiment = args.get("embodiment", ["aloha-agilex"])
+    if len(embodiment) == 1:
+        robot_file = embodiment_types[embodiment[0]]["file_path"]
+        args["left_robot_file"] = robot_file
+        args["right_robot_file"] = robot_file
+        args["dual_arm_embodied"] = True
+    elif len(embodiment) == 3:
+        args["left_robot_file"] = embodiment_types[embodiment[0]]["file_path"]
+        args["right_robot_file"] = embodiment_types[embodiment[1]]["file_path"]
+        args["embodiment_dis"] = embodiment[2]
+        args["dual_arm_embodied"] = False
+    else:
+        raise ValueError(f"embodiment must have 1 or 3 items, got {len(embodiment)}")
+
+    with open(os.path.join(args["left_robot_file"], "config.yml"), encoding="utf-8") as f:
+        args["left_embodiment_config"] = yaml.safe_load(f)
+    with open(os.path.join(args["right_robot_file"], "config.yml"), encoding="utf-8") as f:
+        args["right_embodiment_config"] = yaml.safe_load(f)
+
+    # Camera dimensions
+    with open(os.path.join(CONFIGS_PATH, "_camera_config.yml"), encoding="utf-8") as f:
+        camera_config = yaml.safe_load(f)
+    head_cam = args["camera"]["head_camera_type"]
+    args["head_camera_h"] = camera_config[head_cam]["h"]
+    args["head_camera_w"] = camera_config[head_cam]["w"]
+
+    # Headless overrides
+    args["render_freq"] = 0
+    args["task_name"] = task_name
+    args["task_config"] = task_config
+
+    _ROBOTWIN_SETUP_CACHE[task_name] = args
+    return dict(args)
+
+
+def _load_robotwin_task(task_name: str) -> type:
+    """Dynamically import and return a RoboTwin 2.0 task class.
+
+    RoboTwin tasks live in ``envs/<task_name>.py`` relative to the repository
+    root and are expected to be on ``sys.path`` after installation.
+    """
+    try:
+        module = importlib.import_module(f"envs.{task_name}")
+    except ModuleNotFoundError as e:
+        raise ModuleNotFoundError(
+            f"Could not import RoboTwin task '{task_name}'. "
+            "Ensure RoboTwin 2.0 is installed and its 'envs/' directory is on PYTHONPATH. "
+            "See the RoboTwin installation guide: https://robotwin-platform.github.io/doc/usage/robotwin-install.html"
+        ) from e
+    task_cls = getattr(module, task_name, None)
+    if task_cls is None:
+        raise AttributeError(f"Task class '{task_name}' not found in envs/{task_name}.py")
+    return task_cls
+
+
+class RoboTwinEnv(gym.Env):
+    """Gymnasium wrapper around a single RoboTwin 2.0 task.
+
+    RoboTwin uses a custom SAPIEN-based API (``setup_demo`` / ``get_obs`` /
+    ``take_action`` / ``check_success``) rather than the standard gym interface.
+    This class bridges that API to Gymnasium so that ``lerobot-eval`` can drive
+    RoboTwin exactly like LIBERO or Meta-World.
+
+    The underlying SAPIEN environment is created lazily on the first ``reset()``
+    call *inside the worker process*.  This is required for
+    ``gym.vector.AsyncVectorEnv`` compatibility: SAPIEN allocates EGL/GPU
+    contexts that must not be forked from the parent process.
+
+    Observations
+    ------------
+    The ``pixels`` dict uses the raw RoboTwin camera names as keys (e.g.
+    ``"head_camera"``, ``"left_camera"``). ``preprocess_observation`` in
+    ``envs/utils.py`` then converts these to ``observation.images.<cam>``.
+
+    Actions
+    -------
+    14-dim float32 array in ``[-1, 1]`` (joint-space, 7 DOF per arm).
+
+    Autograd
+    --------
+    ``setup_demo`` and ``take_action`` drive CuRobo's Newton trajectory
+    optimizer, which calls ``cost.backward()`` internally. lerobot_eval wraps
+    the rollout in ``torch.no_grad()``, so both call sites re-enable grad.
+    """
+
+    metadata = {"render_modes": ["rgb_array"], "render_fps": 25}
+
+    def __init__(
+        self,
+        task_name: str,
+        episode_index: int = 0,
+        n_envs: int = 1,
+        camera_names: Sequence[str] = ROBOTWIN_CAMERA_NAMES,
+        observation_height: int | None = None,
+        observation_width: int | None = None,
+        episode_length: int = DEFAULT_EPISODE_LENGTH,
+        render_mode: str = "rgb_array",
+    ):
+        super().__init__()
+        self.task_name = task_name
+        self.task = task_name  # used by add_envs_task() in utils.py
+        self.task_description = task_name.replace("_", " ")
+        self.episode_index = episode_index
+        self._reset_stride = n_envs
+        self.camera_names = list(camera_names)
+        # Default to D435 dims (the camera type baked into task_config/demo_clean.yml).
+        # The YAML-driven lookup is deferred to reset() so construction doesn't
+        # import RoboTwin's `envs` module — fast-tests run without RoboTwin installed.
+        self.observation_height = observation_height or DEFAULT_CAMERA_H
+        self.observation_width = observation_width or DEFAULT_CAMERA_W
+        self.episode_length = episode_length
+        self._max_episode_steps = episode_length  # lerobot_eval.rollout reads this
+        self.render_mode = render_mode
+
+        self._env: Any | None = None  # deferred — created on first reset() inside worker
+        self._step_count: int = 0
+        self._black_frame = np.zeros((self.observation_height, self.observation_width, 3), dtype=np.uint8)
+
+        image_spaces = {
+            cam: spaces.Box(
+                low=0,
+                high=255,
+                shape=(self.observation_height, self.observation_width, 3),
+                dtype=np.uint8,
+            )
+            for cam in self.camera_names
+        }
+        self.observation_space = spaces.Dict(
+            {
+                "pixels": spaces.Dict(image_spaces),
+                "agent_pos": spaces.Box(low=-np.inf, high=np.inf, shape=(ACTION_DIM,), dtype=np.float32),
+            }
+        )
+        self.action_space = spaces.Box(
+            low=ACTION_LOW, high=ACTION_HIGH, shape=(ACTION_DIM,), dtype=np.float32
+        )
+
+    def _ensure_env(self) -> None:
+        """Create the SAPIEN environment on first use.
+
+        Called inside the worker subprocess after fork(), so each worker gets
+        its own EGL/GPU context rather than inheriting a stale one from the
+        parent process (which causes crashes with AsyncVectorEnv).
+        """
+        if self._env is not None:
+            return
+        task_cls = _load_robotwin_task(self.task_name)
+        self._env = task_cls()
+
+    def _get_obs(self) -> RobotObservation:
+        assert self._env is not None, "_get_obs called before _ensure_env()"
+        raw = self._env.get_obs()
+        cameras_raw = raw.get("observation", {})
+
+        images: dict[str, np.ndarray] = {}
+        for cam in self.camera_names:
+            cam_data = cameras_raw.get(cam)
+            img = cam_data.get("rgb") if cam_data else None
+            if img is None:
+                images[cam] = self._black_frame
+                continue
+            img = np.asarray(img, dtype=np.uint8)
+            if img.ndim == 2:
+                img = np.stack([img, img, img], axis=-1)
+            elif img.shape[-1] != 3:
+                img = img[..., :3]
+            images[cam] = img
+
+        ja = raw.get("joint_action") or {}
+        vec = ja.get("vector")
+        if vec is not None:
+            arr = np.asarray(vec, dtype=np.float32).ravel()
+            joint_state = (
+                arr[:ACTION_DIM] if arr.size >= ACTION_DIM else np.zeros(ACTION_DIM, dtype=np.float32)
+            )
+        else:
+            joint_state = np.zeros(ACTION_DIM, dtype=np.float32)
+
+        return {"pixels": images, "agent_pos": joint_state}
+
+    def reset(self, seed: int | None = None, **kwargs) -> tuple[RobotObservation, dict]:
+        self._ensure_env()
+        super().reset(seed=seed)
+        assert self._env is not None  # set by _ensure_env() above
+
+        actual_seed = self.episode_index if seed is None else seed
+        setup_kwargs = _load_robotwin_setup_kwargs(self.task_name)
+        setup_kwargs.update(seed=actual_seed, is_test=True)
+        with torch.enable_grad():
+            self._env.setup_demo(**setup_kwargs)
+        self.episode_index += self._reset_stride
+        self._step_count = 0
+
+        obs = self._get_obs()
+        return obs, {"is_success": False, "task": self.task_name}
+
+    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
+        assert self._env is not None, "step() called before reset()"
+        if action.ndim != 1 or action.shape[0] != ACTION_DIM:
+            raise ValueError(f"Expected 1-D action of shape ({ACTION_DIM},), got {action.shape}")
+
+        with torch.enable_grad():
+            if hasattr(self._env, "take_action"):
+                self._env.take_action(action)
+            else:
+                self._env.step(action)
+
+        self._step_count += 1
+
+        is_success = bool(getattr(self._env, "eval_success", False))
+        if not is_success and hasattr(self._env, "check_success"):
+            is_success = bool(self._env.check_success())
+
+        obs = self._get_obs()
+        reward = float(is_success)
+        terminated = is_success
+        truncated = self._step_count >= self.episode_length
+
+        info: dict[str, Any] = {
+            "task": self.task_name,
+            "is_success": is_success,
+            "step": self._step_count,
+        }
+        if terminated or truncated:
+            info["final_info"] = {
+                "task": self.task_name,
+                "is_success": is_success,
+            }
+            self.reset()
+
+        return obs, reward, terminated, truncated, info
+
+    def render(self) -> np.ndarray:
+        self._ensure_env()
+        obs = self._get_obs()
+        # Prefer head camera for rendering; fall back to first available.
+        if "head_camera" in obs["pixels"]:
+            return obs["pixels"]["head_camera"]
+        return next(iter(obs["pixels"].values()))
+
+    def close(self) -> None:
+        if self._env is not None:
+            if hasattr(self._env, "close_env"):
+                import contextlib
+
+                with contextlib.suppress(TypeError):
+                    self._env.close_env()
+            self._env = None
+
+
+# ---- Multi-task factory --------------------------------------------------------
+
+
+def _make_env_fns(
+    *,
+    task_name: str,
+    n_envs: int,
+    camera_names: list[str],
+    observation_height: int,
+    observation_width: int,
+    episode_length: int,
+) -> list[Callable[[], RoboTwinEnv]]:
+    """Return n_envs factory callables for a single task."""
+
+    def _make_one(episode_index: int) -> RoboTwinEnv:
+        return RoboTwinEnv(
+            task_name=task_name,
+            episode_index=episode_index,
+            n_envs=n_envs,
+            camera_names=camera_names,
+            observation_height=observation_height,
+            observation_width=observation_width,
+            episode_length=episode_length,
+        )
+
+    return [partial(_make_one, i) for i in range(n_envs)]
+
+
+def create_robotwin_envs(
+    task: str,
+    n_envs: int,
+    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
+    camera_names: Sequence[str] = ROBOTWIN_CAMERA_NAMES,
+    observation_height: int = DEFAULT_CAMERA_H,
+    observation_width: int = DEFAULT_CAMERA_W,
+    episode_length: int = DEFAULT_EPISODE_LENGTH,
+) -> dict[str, dict[int, Any]]:
+    """Create vectorized RoboTwin 2.0 environments.
+
+    Returns:
+        ``dict[task_name][0] -> VectorEnv`` — one entry per task, each wrapping
+        ``n_envs`` parallel rollouts.
+
+    Args:
+        task: Comma-separated list of task names (e.g. ``"beat_block_hammer"``
+            or ``"beat_block_hammer,click_bell"``).
+        n_envs: Number of parallel rollouts per task.
+        env_cls: Vector env constructor (e.g. ``gym.vector.AsyncVectorEnv``).
+        camera_names: Cameras to include in observations.
+        observation_height: Pixel height for all cameras.
+        observation_width: Pixel width for all cameras.
+        episode_length: Max steps before truncation.
+    """
+    if env_cls is None or not callable(env_cls):
+        raise ValueError("env_cls must be callable (e.g. gym.vector.AsyncVectorEnv).")
+    if not isinstance(n_envs, int) or n_envs <= 0:
+        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
+
+    task_names = [t.strip() for t in str(task).split(",") if t.strip()]
+    if not task_names:
+        raise ValueError("`task` must contain at least one RoboTwin task name.")
+
+    unknown = [t for t in task_names if t not in ROBOTWIN_TASKS]
+    if unknown:
+        raise ValueError(f"Unknown RoboTwin tasks: {unknown}. Available tasks: {sorted(ROBOTWIN_TASKS)}")
+
+    logger.info(
+        "Creating RoboTwin envs | tasks=%s | n_envs(per task)=%d",
+        task_names,
+        n_envs,
+    )
+
+    is_async = env_cls is gym.vector.AsyncVectorEnv
+    cached_obs_space: spaces.Space | None = None
+    cached_act_space: spaces.Space | None = None
+    cached_metadata: dict[str, Any] | None = None
+
+    out: dict[str, dict[int, Any]] = defaultdict(dict)
+    for task_name in task_names:
+        fns = _make_env_fns(
+            task_name=task_name,
+            n_envs=n_envs,
+            camera_names=list(camera_names),
+            observation_height=observation_height,
+            observation_width=observation_width,
+            episode_length=episode_length,
+        )
+        if is_async:
+            lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
+            if cached_obs_space is None:
+                cached_obs_space = lazy.observation_space
+                cached_act_space = lazy.action_space
+                cached_metadata = lazy.metadata
+            out[task_name][0] = lazy
+        else:
+            out[task_name][0] = env_cls(fns)
+        logger.info("Built vec env | task=%s | n_envs=%d", task_name, n_envs)
+
+    return {k: dict(v) for k, v in out.items()}

src/lerobot/envs/vlabench.py

@@ -0,0 +1,589 @@
+#!/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.
+"""VLABench environment wrapper for LeRobot.
+
+VLABench is a large-scale benchmark for language-conditioned robotic manipulation
+with long-horizon reasoning, built on MuJoCo/dm_control.
+
+- Paper: https://arxiv.org/abs/2412.18194
+- GitHub: https://github.com/OpenMOSS/VLABench
+- Website: https://vlabench.github.io
+"""
+
+from __future__ import annotations
+
+import contextlib
+import logging
+from collections import defaultdict
+from collections.abc import Callable, Sequence
+from typing import Any
+
+import cv2
+import gymnasium as gym
+import numpy as np
+from gymnasium import spaces
+from scipy.spatial.transform import Rotation
+
+from lerobot.types import RobotObservation
+
+from .utils import _LazyAsyncVectorEnv
+
+logger = logging.getLogger(__name__)
+
+ACTION_DIM = 7  # pos(3) + euler(3) + gripper(1)
+ACTION_LOW = np.array([-1.0, -1.0, -1.0, -1.0, -1.0, -1.0, 0.0], dtype=np.float32)
+ACTION_HIGH = np.array([1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0], dtype=np.float32)
+
+# Default max episode steps per task type
+DEFAULT_MAX_EPISODE_STEPS = 500
+
+# VLABench task suites
+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",
+    # Physical series
+    "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",
+]
+
+SUITE_TASKS: dict[str, list[str]] = {
+    "primitive": PRIMITIVE_TASKS,
+    "composite": COMPOSITE_TASKS,
+}
+
+
+class VLABenchEnv(gym.Env):
+    """Gymnasium wrapper for VLABench environments.
+
+    Wraps the dm_control-based VLABench simulator behind a standard gym.Env interface.
+    Supports multiple cameras (front, second, wrist) and end-effector control.
+    """
+
+    metadata = {"render_modes": ["rgb_array"], "render_fps": 10}
+
+    def __init__(
+        self,
+        task: str = "select_fruit",
+        obs_type: str = "pixels_agent_pos",
+        render_mode: str = "rgb_array",
+        render_resolution: tuple[int, int] = (480, 480),
+        robot: str = "franka",
+        max_episode_steps: int = DEFAULT_MAX_EPISODE_STEPS,
+        action_mode: str = "eef",
+    ):
+        super().__init__()
+        self.task = task
+        self.obs_type = obs_type
+        self.render_mode = render_mode
+        self.render_resolution = render_resolution
+        self.robot = robot
+        self._max_episode_steps = max_episode_steps
+        self.action_mode = action_mode
+
+        # Deferred — created on first reset() inside worker subprocess to avoid
+        # inheriting stale GPU/EGL contexts when AsyncVectorEnv spawns workers.
+        # We never cache `env.physics`: dm_control exposes it as a weakref
+        # proxy that goes stale across resets (rebuilds the sim), so we always
+        # refetch it via `self._env.physics` at the call site.
+        self._env = None
+        self.task_description = ""  # populated on first reset
+        # Cached world-frame XYZ of the robot base link. The VLABench datasets
+        # log both `observation.state` positions and `actions` positions in
+        # robot-base frame (see VLABench/scripts/convert_to_lerobot.py which
+        # subtracts `robot_frame_pos` from ee_pos). The robot is attached at a
+        # fixed offset per task so this is safe to cache once per env build.
+        self._robot_base_xyz: np.ndarray | None = None
+
+        h, w = self.render_resolution
+
+        if self.obs_type == "state":
+            raise NotImplementedError(
+                "The 'state' observation type is not supported in VLABenchEnv. "
+                "Please use 'pixels' or 'pixels_agent_pos'."
+            )
+        elif self.obs_type == "pixels":
+            self.observation_space = spaces.Dict(
+                {
+                    "pixels": spaces.Dict(
+                        {
+                            "image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
+                            "second_image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
+                            "wrist_image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
+                        }
+                    ),
+                }
+            )
+        elif self.obs_type == "pixels_agent_pos":
+            self.observation_space = spaces.Dict(
+                {
+                    "pixels": spaces.Dict(
+                        {
+                            "image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
+                            "second_image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
+                            "wrist_image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
+                        }
+                    ),
+                    "agent_pos": spaces.Box(low=-np.inf, high=np.inf, shape=(7,), dtype=np.float64),
+                }
+            )
+        else:
+            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
+
+        self.action_space = spaces.Box(low=ACTION_LOW, high=ACTION_HIGH, dtype=np.float32)
+
+    # Max attempts to rebuild the underlying env when MuJoCo throws
+    # `PhysicsError` (e.g. mjWARN_BADQACC) during VLABench's 20-step
+    # reset warm-up. Some random task/layout samples land in unstable
+    # initial configurations; re-sampling the layout almost always
+    # gives a stable one. A handful of upstream tasks (notably
+    # `select_mahjong`) have layout samplers that diverge often enough
+    # to need >>5 retries, so we pick a generous ceiling.
+    _ENSURE_ENV_MAX_ATTEMPTS = 20
+
+    def _ensure_env(self) -> None:
+        """Create the underlying VLABench env on first use.
+
+        Called inside the worker subprocess after fork(), so each worker gets
+        its own clean rendering context rather than inheriting a stale one from
+        the parent process (which causes crashes with AsyncVectorEnv).
+
+        Retries on `PhysicsError`: VLABench's `LM4ManipDMEnv.reset()` runs 20
+        warm-up `step()` calls while toggling gravity/fluids to let the scene
+        settle; for some random layouts MuJoCo's integrator diverges and
+        raises `mjWARN_BADQACC`. Re-sampling the layout almost always yields
+        a stable one, so we retry a number of times before giving up. Between
+        attempts we reseed NumPy's global RNG from OS entropy so the upstream
+        task sampler explores fresh initial states — without this, retries
+        can replay the same diverging configuration when the sampler is
+        deterministic given the current RNG state.
+        """
+        if self._env is not None:
+            return
+
+        import VLABench.robots  # noqa: F401  # type: ignore[import-untyped]
+        import VLABench.tasks  # noqa: F401  # type: ignore[import-untyped]
+        from dm_control.rl.control import PhysicsError  # type: ignore[import-untyped]
+        from VLABench.envs import load_env  # type: ignore[import-untyped]
+
+        h, w = self.render_resolution
+        last_exc: PhysicsError | None = None
+        for attempt in range(1, self._ENSURE_ENV_MAX_ATTEMPTS + 1):
+            try:
+                env = load_env(task=self.task, robot=self.robot, render_resolution=(h, w))
+                self._env = env
+                break
+            except PhysicsError as exc:
+                last_exc = exc
+                logger.warning(
+                    "PhysicsError on attempt %d/%d while building task '%s': %s. Retrying with fresh layout…",
+                    attempt,
+                    self._ENSURE_ENV_MAX_ATTEMPTS,
+                    self.task,
+                    exc,
+                )
+                np.random.seed(None)
+        if self._env is None:
+            assert last_exc is not None
+            raise RuntimeError(
+                f"VLABench task '{self.task}' failed to produce a stable "
+                f"initial layout after {self._ENSURE_ENV_MAX_ATTEMPTS} "
+                f"attempts. This task's upstream sampler diverges too "
+                f"often for the configured robot; consider removing it "
+                f"from the eval set. Last physics error: {last_exc}"
+            ) from last_exc
+
+        # Extract task description from the dm_control task
+        task_obj = self._env.task
+        if hasattr(task_obj, "task_description"):
+            self.task_description = task_obj.task_description
+        elif hasattr(task_obj, "language_instruction"):
+            self.task_description = task_obj.language_instruction
+        else:
+            self.task_description = self.task
+
+        # Cache robot base world position so `_build_ctrl_from_action` and
+        # `_get_obs` can translate between robot-frame (dataset) and
+        # world-frame (dm_control) without hitting physics every call.
+        try:
+            self._robot_base_xyz = np.asarray(self._env.get_robot_frame_position(), dtype=np.float64).reshape(
+                3
+            )
+        except Exception:
+            # Fallback to VLABench's default Franka base position.
+            self._robot_base_xyz = np.array([0.0, -0.4, 0.78], dtype=np.float64)
+
+    def _get_obs(self) -> dict:
+        """Get current observation from the environment."""
+        assert self._env is not None
+
+        obs = self._env.get_observation()
+        h, w = self.render_resolution
+
+        def _to_hwc3(arr: np.ndarray) -> np.ndarray:
+            """Coerce any camera array to the declared (h, w, 3) uint8 shape."""
+            a = np.asarray(arr)
+            # Drop a leading singleton batch dim if present.
+            while a.ndim > 3 and a.shape[0] == 1:
+                a = a[0]
+            if a.ndim == 3 and a.shape[0] in (1, 3, 4) and a.shape[-1] not in (1, 3, 4):
+                # CHW → HWC
+                a = np.transpose(a, (1, 2, 0))
+            if a.ndim == 2:
+                a = np.stack([a] * 3, axis=-1)
+            if a.ndim != 3:
+                return np.zeros((h, w, 3), dtype=np.uint8)
+            # Force 3 channels.
+            if a.shape[-1] == 1:
+                a = np.repeat(a, 3, axis=-1)
+            elif a.shape[-1] == 4:
+                a = a[..., :3]
+            elif a.shape[-1] != 3:
+                return np.zeros((h, w, 3), dtype=np.uint8)
+            if a.shape[:2] != (h, w):
+                a = cv2.resize(a, (w, h), interpolation=cv2.INTER_AREA)
+            return a.astype(np.uint8)
+
+        # Extract camera images — VLABench returns (n_cameras, C, H, W) or individual arrays
+        raw_frames: list[np.ndarray] = []
+        if "rgb" in obs:
+            rgb = obs["rgb"]
+            if isinstance(rgb, np.ndarray):
+                if rgb.ndim == 4:
+                    raw_frames = [rgb[i] for i in range(rgb.shape[0])]
+                elif rgb.ndim == 3:
+                    raw_frames = [rgb]
+
+        image_keys = ["image", "second_image", "wrist_image"]
+        images: dict[str, np.ndarray] = {}
+        for i, key in enumerate(image_keys):
+            if i < len(raw_frames):
+                images[key] = _to_hwc3(raw_frames[i])
+            else:
+                images[key] = np.zeros((h, w, 3), dtype=np.uint8)
+
+        # Convert VLABench's raw ee_state `[pos_world(3), quat_wxyz(4), open(1)]`
+        # to the dataset's observation.state layout `[pos_robot(3), euler_xyz(3),
+        # gripper(1)]`. See VLABench/scripts/convert_to_lerobot.py — positions
+        # are stored in robot-base frame and orientations as scipy extrinsic
+        # 'xyz' euler angles.
+        raw = np.asarray(obs.get("ee_state", np.zeros(8)), dtype=np.float64).ravel()
+        pos_world = raw[:3] if raw.size >= 3 else np.zeros(3, dtype=np.float64)
+        quat_wxyz = raw[3:7] if raw.size >= 7 else np.array([1.0, 0.0, 0.0, 0.0], dtype=np.float64)
+        gripper = float(raw[7]) if raw.size >= 8 else 0.0
+
+        base = self._robot_base_xyz if self._robot_base_xyz is not None else np.zeros(3, dtype=np.float64)
+        pos_robot = pos_world - base
+        euler_xyz = Rotation.from_quat([quat_wxyz[1], quat_wxyz[2], quat_wxyz[3], quat_wxyz[0]]).as_euler(
+            "xyz", degrees=False
+        )
+
+        ee_state = np.concatenate([pos_robot, euler_xyz, [gripper]]).astype(np.float64)
+
+        if self.obs_type == "pixels":
+            return {"pixels": images}
+        elif self.obs_type == "pixels_agent_pos":
+            return {
+                "pixels": images,
+                "agent_pos": ee_state.astype(np.float64),
+            }
+        else:
+            raise ValueError(f"Unknown obs_type: {self.obs_type}")
+
+    # ---- Action adaptation (EEF → joint ctrl) --------------------------------
+    #
+    # The HF vlabench datasets log 7D actions
+    # `[x, y, z (robot frame), rx, ry, rz (scipy extrinsic xyz), gripper]`,
+    # exactly matching VLABench's own eval pipeline (evaluator.base):
+    #   pos, euler, g = policy(...)
+    #   quat = euler_to_quaternion(*euler)      # extrinsic xyz -> wxyz
+    #   _, qpos = robot.get_qpos_from_ee_pos(physics, pos=pos + base, quat=quat)
+    #   env.step(np.concatenate([qpos, [g, g]]))
+    #
+    # VLABench's dm_control task writes `data.ctrl[:] = action` directly — for
+    # Franka that's 9 entries (7 arm joints + 2 gripper fingers). We mirror the
+    # above conversion so the policy's EEF commands actually drive the robot.
+
+    _FRANKA_FINGER_OPEN = 0.04  # qpos when gripper fully open
+
+    def _build_ctrl_from_action(self, action: np.ndarray, ctrl_dim: int) -> np.ndarray:
+        """Convert a 7D EEF action into the `ctrl_dim`-sized joint command vector.
+
+        For the Franka default (ctrl_dim=9): 7 arm joint qposes (via IK) +
+        2 gripper finger qposes (open/closed based on the gripper scalar).
+        If the action is already joint-space (shape matches ctrl_dim), pass
+        through.
+        """
+        if action.shape[0] == ctrl_dim:
+            return action.astype(np.float64, copy=False)
+
+        if action.shape[0] != 7:
+            # Unknown layout — fall back to zero-pad so the sim doesn't crash.
+            padded = np.zeros(ctrl_dim, dtype=np.float64)
+            padded[: min(action.shape[0], ctrl_dim)] = action[:ctrl_dim]
+            return padded
+
+        from dm_control.utils.inverse_kinematics import qpos_from_site_pose
+
+        # Action position is in robot-base frame (see convert_to_lerobot.py);
+        # dm_control's IK expects a world-frame target.
+        base = self._robot_base_xyz if self._robot_base_xyz is not None else np.zeros(3, dtype=np.float64)
+        pos_world = np.asarray(action[:3], dtype=np.float64) + base
+        rx, ry, rz = float(action[3]), float(action[4]), float(action[5])
+        gripper = float(np.clip(action[6], 0.0, 1.0))
+
+        # Dataset euler is scipy extrinsic 'xyz' (same as VLABench's
+        # `euler_to_quaternion`). scipy emits `[x, y, z, w]`; dm_control's IK
+        # and MuJoCo use `[w, x, y, z]`, so reorder.
+        qxyzw = Rotation.from_euler("xyz", [rx, ry, rz], degrees=False).as_quat()
+        quat = np.array([qxyzw[3], qxyzw[0], qxyzw[1], qxyzw[2]], dtype=np.float64)
+
+        assert self._env is not None
+        robot = self._env.task.robot
+        site_name = robot.end_effector_site.full_identifier
+
+        # inplace=False so IK doesn't mutate physics state mid-step — we only
+        # want the solved qpos. Fetch a fresh physics handle — caching it can
+        # yield a stale weakref after a reset.
+        ik_result = qpos_from_site_pose(
+            self._env.physics,
+            site_name=site_name,
+            target_pos=pos_world,
+            target_quat=quat,
+            inplace=False,
+            max_steps=100,
+        )
+        n_dof = robot.n_dof  # 7 for Franka
+        arm_qpos = ik_result.qpos[:n_dof]
+
+        # Dataset gripper convention: 1 = open (finger qpos = 0.04),
+        # 0 = closed (finger qpos = 0.0). See VLABench/scripts/convert_to_lerobot.py
+        # where `trajectory[i][-1] > 0.03` is encoded as `1`.
+        finger_qpos = gripper * self._FRANKA_FINGER_OPEN
+
+        ctrl = np.zeros(ctrl_dim, dtype=np.float64)
+        ctrl[:n_dof] = arm_qpos
+        # Remaining entries are gripper fingers (usually 2 for Franka).
+        ctrl[n_dof:] = finger_qpos
+        return ctrl
+
+    def reset(self, seed=None, **kwargs) -> tuple[RobotObservation, dict[str, Any]]:
+        self._ensure_env()
+        assert self._env is not None
+        super().reset(seed=seed)
+
+        if seed is not None:
+            self._seed_inner_env(int(self.np_random.integers(0, 2**31 - 1)))
+
+        self._env.reset()
+
+        observation = self._get_obs()
+        info = {"is_success": False}
+        return observation, info
+
+    def _seed_inner_env(self, seed: int) -> None:
+        """Propagate `seed` to the inner dm_control env. `Environment.reset()`
+        doesn't accept a seed, so we re-seed the task and environment
+        `RandomState`s directly. Best-effort: silently skipped when the
+        expected attributes are absent on a given VLABench version.
+        """
+        for owner_attr, rng_attr in (("task", "random"), (None, "_random_state")):
+            owner = getattr(self._env, owner_attr) if owner_attr else self._env
+            rng = getattr(owner, rng_attr, None)
+            rng_seed = getattr(rng, "seed", None)
+            if callable(rng_seed):
+                rng_seed(seed)
+
+    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
+        from dm_control.rl.control import PhysicsError  # type: ignore[import-untyped]
+
+        self._ensure_env()
+        assert self._env is not None
+
+        if action.ndim != 1:
+            raise ValueError(
+                f"Expected action to be 1-D (shape (action_dim,)), "
+                f"but got shape {action.shape} with ndim={action.ndim}"
+            )
+
+        if self.action_mode not in ("eef", "joint", "delta_eef"):
+            raise ValueError(f"Unknown action_mode: {self.action_mode}")
+
+        # Always refetch physics — dm_control returns a weakref proxy that can
+        # go stale across resets.
+        physics = self._env.physics
+        ctrl_dim = int(physics.data.ctrl.shape[0])
+        ctrl = self._build_ctrl_from_action(action, ctrl_dim)
+        try:
+            timestep = self._env.step(ctrl)
+        except PhysicsError as exc:
+            # Physics integrator diverged (e.g. mjWARN_BADQACC). Treat it as
+            # a graceful failed termination rather than a hard crash — the
+            # rest of the multi-task eval should still run.
+            logger.warning(
+                "PhysicsError during step on task '%s': %s. Terminating episode.",
+                self.task,
+                exc,
+            )
+            observation = self._get_obs()
+            info = {"task": self.task, "is_success": False, "physics_error": True}
+            # Drop the stale env so the next reset() rebuilds it cleanly.
+            with contextlib.suppress(Exception):
+                self._env.close()
+            self._env = None
+            return observation, 0.0, True, False, info
+
+        # Extract reward from dm_control timestep
+        reward = float(timestep.reward) if timestep.reward is not None else 0.0
+
+        # Check success via the task's termination condition
+        is_success = False
+        if hasattr(self._env, "task") and hasattr(self._env.task, "should_terminate_episode"):
+            is_success = bool(self._env.task.should_terminate_episode(self._env.physics))
+
+        terminated = is_success
+        truncated = False
+        info = {
+            "task": self.task,
+            "is_success": is_success,
+        }
+
+        observation = self._get_obs()
+
+        if terminated:
+            self.reset()
+
+        return observation, reward, terminated, truncated, info
+
+    def render(self) -> np.ndarray:
+        self._ensure_env()
+        obs = self._get_obs()
+        return obs["pixels"]["image"]
+
+    def close(self):
+        if self._env is not None:
+            self._env.close()
+            self._env = None
+
+
+# ---- Main API ----------------------------------------------------------------
+
+
+def create_vlabench_envs(
+    task: str,
+    n_envs: int,
+    gym_kwargs: dict[str, Any] | None = None,
+    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
+) -> dict[str, dict[int, Any]]:
+    """
+    Create vectorized VLABench environments with a consistent return shape.
+
+    Returns:
+        dict[suite_name][task_id] -> vec_env (env_cls([...]) with exactly n_envs factories)
+
+    Notes:
+        - n_envs is the number of rollouts *per task*.
+        - `task` can be a suite name ("primitive", "composite"), a comma-separated list of
+          suite names, or individual task names (e.g. "select_fruit,heat_food").
+    """
+    if env_cls is None or not callable(env_cls):
+        raise ValueError("env_cls must be a callable that wraps a list of environment factory callables.")
+    if not isinstance(n_envs, int) or n_envs <= 0:
+        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
+
+    gym_kwargs = dict(gym_kwargs or {})
+    task_groups = [t.strip() for t in task.split(",") if t.strip()]
+    if not task_groups:
+        raise ValueError("`task` must contain at least one VLABench task or suite name.")
+
+    logger.info(
+        "Creating VLABench envs | task_groups=%s | n_envs(per task)=%d",
+        task_groups,
+        n_envs,
+    )
+
+    is_async = env_cls is gym.vector.AsyncVectorEnv
+    cached_obs_space = None
+    cached_act_space = None
+    cached_metadata = None
+    out: dict[str, dict[int, Any]] = defaultdict(dict)
+
+    for group in task_groups:
+        # Check if it's a suite name, otherwise treat as individual task
+        tasks = SUITE_TASKS.get(group, [group])
+
+        for tid, task_name in enumerate(tasks):
+            logger.info(
+                "Building vec env | group=%s | task_id=%d | task=%s",
+                group,
+                tid,
+                task_name,
+            )
+
+            fns = [(lambda tn=task_name: VLABenchEnv(task=tn, **gym_kwargs)) for _ in range(n_envs)]
+
+            if is_async:
+                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
+                if cached_obs_space is None:
+                    cached_obs_space = lazy.observation_space
+                    cached_act_space = lazy.action_space
+                    cached_metadata = lazy.metadata
+                out[group][tid] = lazy
+            else:
+                out[group][tid] = env_cls(fns)
+
+    return {group: dict(task_map) for group, task_map in out.items()}

src/lerobot/processor/__init__.py

@@ -93,6 +93,7 @@ from .relative_action_processor import (
     to_relative_actions,
 )
 from .rename_processor import RenameObservationsProcessorStep, rename_stats
+from .render_messages_processor import RenderMessagesStep
 from .tokenizer_processor import ActionTokenizerProcessorStep, TokenizerProcessorStep
 
 __all__ = [
@@ -128,6 +129,7 @@ __all__ = [
     "make_default_robot_observation_processor",
     "AbsoluteActionsProcessorStep",
     "RelativeActionsProcessorStep",
+    "RenderMessagesStep",
     "MapDeltaActionToRobotActionStep",
     "MapTensorToDeltaActionDictStep",
     "NewLineTaskProcessorStep",

src/lerobot/processor/batch_processor.py

@@ -174,6 +174,24 @@ class AddBatchDimensionComplementaryDataStep(ComplementaryDataProcessorStep):
             task_index_value = complementary_data["task_index"]
             if isinstance(task_index_value, Tensor) and task_index_value.dim() == 0:
                 complementary_data["task_index"] = task_index_value.unsqueeze(0)
+
+        complementary_data.pop("language_persistent", None)
+        complementary_data.pop("language_events", None)
+
+        if "messages" in complementary_data:
+            messages = complementary_data["messages"]
+            if isinstance(messages, list) and (not messages or isinstance(messages[0], dict)):
+                complementary_data["messages"] = [messages]
+
+        if "message_streams" in complementary_data:
+            streams = complementary_data["message_streams"]
+            if isinstance(streams, list) and (not streams or isinstance(streams[0], str)):
+                complementary_data["message_streams"] = [streams]
+
+        if "target_message_indices" in complementary_data:
+            indices = complementary_data["target_message_indices"]
+            if isinstance(indices, list) and (not indices or isinstance(indices[0], int)):
+                complementary_data["target_message_indices"] = [indices]
         return complementary_data
 
     def transform_features(

src/lerobot/processor/converters.py

@@ -167,12 +167,35 @@ def _extract_complementary_data(batch: dict[str, Any]) -> dict[str, Any]:
     """
     pad_keys = {k: v for k, v in batch.items() if "_is_pad" in k}
     task_key = {"task": batch["task"]} if "task" in batch else {}
-    subtask_key = {"subtask": batch["subtask"]} if "subtask" in batch else {}
     index_key = {"index": batch["index"]} if "index" in batch else {}
     task_index_key = {"task_index": batch["task_index"]} if "task_index" in batch else {}
     episode_index_key = {"episode_index": batch["episode_index"]} if "episode_index" in batch else {}
+    timestamp_key = {"timestamp": batch["timestamp"]} if "timestamp" in batch else {}
+    language_persistent_key = (
+        {"language_persistent": batch["language_persistent"]} if "language_persistent" in batch else {}
+    )
+    language_events_key = {"language_events": batch["language_events"]} if "language_events" in batch else {}
+    messages_key = {"messages": batch["messages"]} if "messages" in batch else {}
+    message_streams_key = {"message_streams": batch["message_streams"]} if "message_streams" in batch else {}
+    target_message_indices_key = (
+        {"target_message_indices": batch["target_message_indices"]}
+        if "target_message_indices" in batch
+        else {}
+    )
 
-    return {**pad_keys, **task_key, **subtask_key, **index_key, **task_index_key, **episode_index_key}
+    return {
+        **pad_keys,
+        **task_key,
+        **index_key,
+        **task_index_key,
+        **episode_index_key,
+        **timestamp_key,
+        **language_persistent_key,
+        **language_events_key,
+        **messages_key,
+        **message_streams_key,
+        **target_message_indices_key,
+    }
 
 
 def create_transition(

src/lerobot/processor/render_messages_processor.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any
+
+from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.configs.recipe import TrainingRecipe
+from lerobot.datasets.language import LANGUAGE_EVENTS, LANGUAGE_PERSISTENT
+from lerobot.datasets.language_render import render_sample
+from lerobot.types import EnvTransition, TransitionKey
+
+from .pipeline import ProcessorStep, ProcessorStepRegistry
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="render_messages_processor")
+class RenderMessagesStep(ProcessorStep):
+    """Processor step that turns raw language columns into rendered chat messages.
+
+    Reads ``language_persistent`` and ``language_events`` from the transition's
+    complementary data, renders them through ``recipe`` at the sample timestamp,
+    and replaces the raw columns with the resulting ``messages`` /
+    ``message_streams`` / ``target_message_indices`` keys.
+    """
+
+    recipe: TrainingRecipe
+    dataset_ctx: Any | None = None
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition | None:
+        """Render messages for a single transition; return ``None`` to drop it."""
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+        persistent = complementary_data.get(LANGUAGE_PERSISTENT) or []
+        events = complementary_data.get(LANGUAGE_EVENTS) or []
+
+        if not persistent and not events:
+            return transition
+
+        timestamp = complementary_data.get("timestamp")
+        if timestamp is None:
+            raise KeyError("RenderMessagesStep requires sample timestamp in complementary data.")
+
+        sample_idx = complementary_data.get("index", 0)
+        rendered = render_sample(
+            recipe=self.recipe,
+            persistent=persistent,
+            events=events,
+            t=_scalar(timestamp),
+            sample_idx=int(_scalar(sample_idx)),
+            task=complementary_data.get("task"),
+            dataset_ctx=self.dataset_ctx,
+        )
+        if rendered is None:
+            return None
+
+        new_transition = transition.copy()
+        new_complementary_data = dict(complementary_data)
+        new_complementary_data.pop(LANGUAGE_PERSISTENT, None)
+        new_complementary_data.pop(LANGUAGE_EVENTS, None)
+        new_complementary_data.update(rendered)
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
+        return new_transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """Pass features through unchanged; rendering only touches complementary data."""
+        return features
+
+
+def _scalar(value: Any) -> float | int:
+    """Unwrap a tensor/array/single-element list into a Python scalar."""
+    if hasattr(value, "item"):
+        return value.item()
+    if isinstance(value, list) and len(value) == 1:
+        return _scalar(value[0])
+    return value

src/lerobot/scripts/lerobot_train.py

@@ -47,6 +47,7 @@ from lerobot.datasets import EpisodeAwareSampler, make_dataset
 from lerobot.envs import close_envs, make_env, make_env_pre_post_processors
 from lerobot.optim.factory import make_optimizer_and_scheduler
 from lerobot.policies import PreTrainedPolicy, make_policy, make_pre_post_processors
+from lerobot.utils.collate import lerobot_collate_fn
 from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.logging_utils import AverageMeter, MetricsTracker
 from lerobot.utils.random_utils import set_seed
@@ -386,6 +387,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         sampler=sampler,
         pin_memory=device.type == "cuda",
         drop_last=False,
+        collate_fn=lerobot_collate_fn,
         prefetch_factor=cfg.prefetch_factor if cfg.num_workers > 0 else None,
         persistent_workers=cfg.persistent_workers and cfg.num_workers > 0,
     )

src/lerobot/utils/collate.py

@@ -0,0 +1,54 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from typing import Any
+
+from torch.utils.data._utils.collate import default_collate
+
+from lerobot.datasets.language import LANGUAGE_COLUMNS
+
+_PYTHON_LIST_KEYS = {"messages", "message_streams", "target_message_indices"}
+
+
+def lerobot_collate_fn(batch: list[dict[str, Any] | None]) -> dict[str, Any] | None:
+    """Collate function that preserves Python-list and language fields as lists.
+
+    Drops ``None`` samples (e.g. recipes that yielded no target message), keeps
+    rendered-message and language fields as plain Python lists, and delegates
+    every other key to PyTorch's ``default_collate``.
+    """
+    batch = [sample for sample in batch if sample is not None]
+    if not batch:
+        return None
+
+    preserved = {
+        key: [sample[key] for sample in batch if key in sample]
+        for key in _PYTHON_LIST_KEYS
+        if any(key in sample for sample in batch)
+    }
+    tensorizable = [
+        {
+            key: value
+            for key, value in sample.items()
+            if key not in _PYTHON_LIST_KEYS and key not in LANGUAGE_COLUMNS
+        }
+        for sample in batch
+    ]
+    collated = default_collate(tensorizable)
+    collated.update(preserved)
+    return collated

tests/configs/test_recipe.py

@@ -0,0 +1,32 @@
+#!/usr/bin/env python
+
+from pathlib import Path
+
+import pytest
+
+from lerobot.configs.recipe import MessageTurn, TrainingRecipe
+
+
+def test_message_recipe_validates_unknown_binding():
+    with pytest.raises(ValueError, match="unknown binding"):
+        TrainingRecipe(
+            messages=[
+                MessageTurn(role="user", content="${missing}", stream="high_level"),
+                MessageTurn(role="assistant", content="ok", stream="high_level", target=True),
+            ]
+        )
+
+
+def test_canonical_recipe_loads():
+    recipe = TrainingRecipe.from_yaml(Path("src/lerobot/configs/recipes/pi05_hirobot.yaml"))
+
+    assert recipe.blend is not None
+    assert set(recipe.blend) == {
+        "memory_update",
+        "user_interjection_response",
+        "high_level_subtask",
+        "low_level_execution",
+        "ask_vqa_top",
+        "ask_vqa_wrist",
+    }
+    assert sum(component.weight for component in recipe.blend.values()) == pytest.approx(0.96)

tests/datasets/test_language.py

@@ -0,0 +1,152 @@
+#!/usr/bin/env python
+
+import numpy as np
+import pandas as pd
+import pyarrow as pa
+import pytest
+
+from lerobot.datasets import LeRobotDataset
+from lerobot.datasets.io_utils import write_info
+from lerobot.datasets.language import (
+    EVENT_ONLY_STYLES,
+    LANGUAGE_EVENTS,
+    LANGUAGE_PERSISTENT,
+    PERSISTENT_STYLES,
+    STYLE_REGISTRY,
+    VIEW_DEPENDENT_STYLES,
+    column_for_style,
+    is_view_dependent_style,
+    language_events_arrow_type,
+    language_feature_info,
+    language_persistent_arrow_type,
+    validate_camera_field,
+)
+from lerobot.datasets.utils import DEFAULT_DATA_PATH
+
+
+def test_language_arrow_schema_has_expected_fields():
+    persistent_row_type = language_persistent_arrow_type().value_type
+    event_row_type = language_events_arrow_type().value_type
+
+    assert isinstance(persistent_row_type, pa.StructType)
+    assert persistent_row_type.names == [
+        "role",
+        "content",
+        "style",
+        "timestamp",
+        "camera",
+        "tool_calls",
+    ]
+
+    assert isinstance(event_row_type, pa.StructType)
+    assert event_row_type.names == ["role", "content", "style", "camera", "tool_calls"]
+
+
+def test_style_registry_routes_columns():
+    assert {"subtask", "plan", "memory", "motion", "task_aug"} == PERSISTENT_STYLES
+    assert {"interjection", "vqa", "trace"} == EVENT_ONLY_STYLES
+    assert PERSISTENT_STYLES | EVENT_ONLY_STYLES <= STYLE_REGISTRY
+
+    assert column_for_style("subtask") == LANGUAGE_PERSISTENT
+    assert column_for_style("plan") == LANGUAGE_PERSISTENT
+    assert column_for_style("memory") == LANGUAGE_PERSISTENT
+    assert column_for_style("motion") == LANGUAGE_PERSISTENT
+    assert column_for_style("task_aug") == LANGUAGE_PERSISTENT
+    assert column_for_style("interjection") == LANGUAGE_EVENTS
+    assert column_for_style("vqa") == LANGUAGE_EVENTS
+    assert column_for_style("trace") == LANGUAGE_EVENTS
+    assert column_for_style(None) == LANGUAGE_EVENTS
+
+
+def test_view_dependent_styles():
+    # motion lives in PERSISTENT_STYLES and is described in robot-frame
+    # (joint / Cartesian) terms, so it is NOT view-dependent. Only vqa
+    # (event) and trace (event, pixel-trajectory) carry a camera tag.
+    assert {"vqa", "trace"} == VIEW_DEPENDENT_STYLES
+    assert is_view_dependent_style("vqa")
+    assert is_view_dependent_style("trace")
+    assert not is_view_dependent_style("motion")
+    assert not is_view_dependent_style("subtask")
+    assert not is_view_dependent_style("plan")
+    assert not is_view_dependent_style("interjection")
+    assert not is_view_dependent_style(None)
+
+
+def test_validate_camera_field_requires_camera_for_view_dependent_styles():
+    validate_camera_field("vqa", "observation.images.top")
+    validate_camera_field("trace", "observation.images.front")
+    with pytest.raises(ValueError, match="view-dependent"):
+        validate_camera_field("vqa", None)
+    with pytest.raises(ValueError, match="view-dependent"):
+        validate_camera_field("trace", "")
+
+
+def test_validate_camera_field_rejects_camera_on_non_view_dependent_styles():
+    validate_camera_field("subtask", None)
+    validate_camera_field("plan", None)
+    validate_camera_field("memory", None)
+    validate_camera_field("motion", None)
+    validate_camera_field("interjection", None)
+    validate_camera_field(None, None)
+    with pytest.raises(ValueError, match="must have camera=None"):
+        validate_camera_field("subtask", "observation.images.top")
+    with pytest.raises(ValueError, match="must have camera=None"):
+        validate_camera_field("motion", "observation.images.top")
+    with pytest.raises(ValueError, match="must have camera=None"):
+        validate_camera_field("interjection", "observation.images.top")
+    with pytest.raises(ValueError, match="must have camera=None"):
+        validate_camera_field(None, "observation.images.top")
+
+
+def test_unknown_style_rejected():
+    with pytest.raises(ValueError, match="Unknown language style"):
+        column_for_style("surprise")
+
+
+def test_lerobot_dataset_passes_language_columns_through(tmp_path, empty_lerobot_dataset_factory):
+    root = tmp_path / "language_dataset"
+    dataset = empty_lerobot_dataset_factory(
+        root=root,
+        features={"state": {"dtype": "float32", "shape": (2,), "names": None}},
+        use_videos=False,
+    )
+    dataset.add_frame({"state": np.array([0.0, 1.0], dtype=np.float32), "task": "tidy"})
+    dataset.add_frame({"state": np.array([1.0, 2.0], dtype=np.float32), "task": "tidy"})
+    dataset.save_episode()
+    dataset.finalize()
+
+    persistent = [
+        {
+            "role": "assistant",
+            "content": "reach for the cup",
+            "style": "subtask",
+            "timestamp": 0.0,
+            "camera": None,
+            "tool_calls": None,
+        }
+    ]
+    event = {
+        "role": "user",
+        "content": "what is visible?",
+        "style": "vqa",
+        "camera": "observation.images.top",
+        "tool_calls": None,
+    }
+    data_path = root / DEFAULT_DATA_PATH.format(chunk_index=0, file_index=0)
+    df = pd.read_parquet(data_path)
+    df[LANGUAGE_PERSISTENT] = [persistent, persistent]
+    df[LANGUAGE_EVENTS] = [[event], []]
+    df.to_parquet(data_path)
+
+    info = dataset.meta.info
+    info["features"].update(language_feature_info())
+    write_info(info, root)
+
+    reloaded = LeRobotDataset(repo_id=dataset.repo_id, root=root)
+
+    first = reloaded[0]
+    second = reloaded[1]
+    assert first[LANGUAGE_PERSISTENT] == persistent
+    assert first[LANGUAGE_EVENTS] == [event]
+    assert second[LANGUAGE_PERSISTENT] == persistent
+    assert second[LANGUAGE_EVENTS] == []

tests/datasets/test_language_render.py

@@ -0,0 +1,388 @@
+#!/usr/bin/env python
+
+from pathlib import Path
+
+import pytest
+
+from lerobot.configs.recipe import MessageTurn, TrainingRecipe
+from lerobot.datasets.language_render import active_at, emitted_at, nth_next, nth_prev, render_sample
+
+
+def persistent_row(role, content, style, timestamp, tool_calls=None, camera=None):
+    return {
+        "role": role,
+        "content": content,
+        "style": style,
+        "timestamp": timestamp,
+        "camera": camera,
+        "tool_calls": tool_calls,
+    }
+
+
+def event_row(role, content, style, tool_calls=None, camera=None):
+    return {
+        "role": role,
+        "content": content,
+        "style": style,
+        "camera": camera,
+        "tool_calls": tool_calls,
+    }
+
+
+PERSISTENT = [
+    persistent_row("assistant", "plan 0", "plan", 0.0),
+    persistent_row("assistant", "memory 0", "memory", 0.0),
+    persistent_row("assistant", "subtask 0", "subtask", 0.0),
+    persistent_row("assistant", "memory 1", "memory", 1.0),
+    persistent_row("assistant", "subtask 1", "subtask", 1.0),
+]
+EVENTS_AT_1 = [
+    event_row("user", "what is visible?", "vqa", camera="observation.images.top"),
+    event_row("assistant", '{"count": 2}', "vqa", camera="observation.images.top"),
+]
+EVENTS_AT_2 = [
+    event_row("user", "skip wiping", "interjection"),
+    event_row(
+        "assistant",
+        None,
+        None,
+        [{"type": "function", "function": {"name": "say", "arguments": {"text": "Skipping wiping."}}}],
+    ),
+]
+# Same emission tick, two cameras: triggers per-camera disambiguation in
+# resolvers, mirroring how Module 3 of the annotation pipeline writes one
+# (vqa, user) + (vqa, assistant) pair per camera.
+EVENTS_AT_3_TWO_CAMERAS = [
+    event_row("user", "how many cups (top)?", "vqa", camera="observation.images.top"),
+    event_row("assistant", '{"count": 3}', "vqa", camera="observation.images.top"),
+    event_row("user", "how many cups (wrist)?", "vqa", camera="observation.images.wrist"),
+    event_row("assistant", '{"count": 1}', "vqa", camera="observation.images.wrist"),
+]
+
+
+def test_resolver_temporal_semantics():
+    assert active_at(0.5, persistent=PERSISTENT, style="subtask")["content"] == "subtask 0"
+    assert active_at(1.0, persistent=PERSISTENT, style="subtask")["content"] == "subtask 1"
+    assert emitted_at(0.5, persistent=PERSISTENT, events=[], style="vqa", role="assistant") is None
+    assert (
+        emitted_at(1.0, persistent=PERSISTENT, events=EVENTS_AT_1, style="vqa", role="assistant")["content"]
+        == '{"count": 2}'
+    )
+
+
+def test_persistent_relative_resolvers_reject_event_styles():
+    with pytest.raises(ValueError, match="event-only"):
+        active_at(1.0, persistent=PERSISTENT, style="vqa")
+    with pytest.raises(ValueError, match="event-only"):
+        nth_prev(1.0, persistent=PERSISTENT, style="interjection")
+
+
+def test_nth_prev_and_next():
+    assert nth_prev(1.0, persistent=PERSISTENT, style="subtask", offset=1)["content"] == "subtask 0"
+    assert nth_next(0.0, persistent=PERSISTENT, style="subtask", offset=1)["content"] == "subtask 1"
+
+
+def test_substitution_if_present_multimodal_and_tool_calls():
+    recipe = TrainingRecipe(
+        messages=[
+            MessageTurn(
+                role="user",
+                content=[
+                    {"type": "image", "feature": "observation.images.top"},
+                    {"type": "text", "text": "${task}: ${interjection}"},
+                ],
+                stream="high_level",
+                if_present="interjection",
+            ),
+            MessageTurn(
+                role="assistant",
+                content="${plan}",
+                stream="high_level",
+                target=True,
+                tool_calls_from="speech",
+            ),
+        ],
+        bindings={"plan": "active_at(t, style=plan)"},
+    )
+
+    rendered = render_sample(
+        recipe=recipe,
+        persistent=PERSISTENT,
+        events=EVENTS_AT_2,
+        t=2.0,
+        sample_idx=0,
+        task="clean kitchen",
+    )
+
+    assert rendered["messages"][0]["content"][1]["text"] == "clean kitchen: skip wiping"
+    assert rendered["messages"][1]["content"] == "plan 0"
+    assert rendered["messages"][1]["tool_calls"][0]["function"]["name"] == "say"
+    assert rendered["message_streams"] == ["high_level", "high_level"]
+    assert rendered["target_message_indices"] == [1]
+
+
+def test_exact_event_miss_returns_none_when_target_skips():
+    recipe = TrainingRecipe(
+        messages=[
+            MessageTurn(role="user", content="${vqa_query}", stream="high_level", if_present="vqa_query"),
+            MessageTurn(
+                role="assistant",
+                content="${vqa}",
+                stream="high_level",
+                target=True,
+                if_present="vqa",
+            ),
+        ]
+    )
+
+    assert (
+        render_sample(recipe=recipe, persistent=PERSISTENT, events=EVENTS_AT_2, t=0.0, sample_idx=0) is None
+    )
+
+
+def test_deterministic_blend_sampling():
+    recipe = TrainingRecipe(
+        blend={
+            "a": TrainingRecipe(
+                weight=1.0,
+                messages=[
+                    MessageTurn(role="user", content="${task}", stream="high_level"),
+                    MessageTurn(role="assistant", content="a", stream="high_level", target=True),
+                ],
+            ),
+            "b": TrainingRecipe(
+                weight=1.0,
+                messages=[
+                    MessageTurn(role="user", content="${task}", stream="high_level"),
+                    MessageTurn(role="assistant", content="b", stream="high_level", target=True),
+                ],
+            ),
+        }
+    )
+
+    first = render_sample(
+        recipe=recipe, persistent=PERSISTENT, events=EVENTS_AT_2, t=0.0, sample_idx=123, task="x"
+    )
+    second = render_sample(
+        recipe=recipe, persistent=PERSISTENT, events=EVENTS_AT_2, t=0.0, sample_idx=123, task="x"
+    )
+    assert first == second
+
+
+def test_emitted_at_filters_vqa_by_camera():
+    top = emitted_at(
+        3.0,
+        persistent=PERSISTENT,
+        events=EVENTS_AT_3_TWO_CAMERAS,
+        style="vqa",
+        role="assistant",
+        camera="observation.images.top",
+    )
+    wrist = emitted_at(
+        3.0,
+        persistent=PERSISTENT,
+        events=EVENTS_AT_3_TWO_CAMERAS,
+        style="vqa",
+        role="assistant",
+        camera="observation.images.wrist",
+    )
+    assert top["content"] == '{"count": 3}'
+    assert wrist["content"] == '{"count": 1}'
+
+
+def test_emitted_at_raises_on_ambiguous_per_camera_vqa():
+    with pytest.raises(ValueError, match="Ambiguous resolver"):
+        emitted_at(
+            3.0,
+            persistent=PERSISTENT,
+            events=EVENTS_AT_3_TWO_CAMERAS,
+            style="vqa",
+            role="assistant",
+        )
+
+
+def test_per_camera_blend_renders_both_views():
+    recipe = TrainingRecipe(
+        blend={
+            "top": TrainingRecipe(
+                weight=1.0,
+                bindings={
+                    "vqa_query": (
+                        "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
+                    ),
+                    "vqa": (
+                        "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
+                    ),
+                },
+                messages=[
+                    MessageTurn(
+                        role="user",
+                        content=[
+                            {"type": "image", "feature": "observation.images.top"},
+                            {"type": "text", "text": "${vqa_query}"},
+                        ],
+                        stream="high_level",
+                        if_present="vqa_query",
+                    ),
+                    MessageTurn(
+                        role="assistant",
+                        content="${vqa}",
+                        stream="high_level",
+                        target=True,
+                        if_present="vqa",
+                    ),
+                ],
+            ),
+            "wrist": TrainingRecipe(
+                weight=1.0,
+                bindings={
+                    "vqa_query": (
+                        "emitted_at(t, style=vqa, role=user, camera=observation.images.wrist)"
+                    ),
+                    "vqa": (
+                        "emitted_at(t, style=vqa, role=assistant, camera=observation.images.wrist)"
+                    ),
+                },
+                messages=[
+                    MessageTurn(
+                        role="user",
+                        content=[
+                            {"type": "image", "feature": "observation.images.wrist"},
+                            {"type": "text", "text": "${vqa_query}"},
+                        ],
+                        stream="high_level",
+                        if_present="vqa_query",
+                    ),
+                    MessageTurn(
+                        role="assistant",
+                        content="${vqa}",
+                        stream="high_level",
+                        target=True,
+                        if_present="vqa",
+                    ),
+                ],
+            ),
+        }
+    )
+
+    rendered_top = render_sample(
+        recipe=recipe.blend["top"],
+        persistent=PERSISTENT,
+        events=EVENTS_AT_3_TWO_CAMERAS,
+        t=3.0,
+        sample_idx=0,
+    )
+    rendered_wrist = render_sample(
+        recipe=recipe.blend["wrist"],
+        persistent=PERSISTENT,
+        events=EVENTS_AT_3_TWO_CAMERAS,
+        t=3.0,
+        sample_idx=0,
+    )
+
+    assert rendered_top["messages"][0]["content"][0]["feature"] == "observation.images.top"
+    assert rendered_top["messages"][0]["content"][1]["text"] == "how many cups (top)?"
+    assert rendered_top["messages"][1]["content"] == '{"count": 3}'
+
+    assert rendered_wrist["messages"][0]["content"][0]["feature"] == "observation.images.wrist"
+    assert rendered_wrist["messages"][0]["content"][1]["text"] == "how many cups (wrist)?"
+    assert rendered_wrist["messages"][1]["content"] == '{"count": 1}'
+
+
+def test_resolve_task_picks_rephrasing_deterministically_per_sample():
+    rephrasings = [
+        persistent_row("user", "tidy the kitchen", "task_aug", 0.0),
+        persistent_row("user", "please clean up the kitchen", "task_aug", 0.0),
+        persistent_row("user", "kitchen needs tidying", "task_aug", 0.0),
+        persistent_row("user", "make the kitchen clean", "task_aug", 0.0),
+    ]
+    recipe = TrainingRecipe(
+        messages=[
+            MessageTurn(role="user", content="${task}", stream="high_level"),
+            MessageTurn(role="assistant", content="ok", stream="high_level", target=True),
+        ]
+    )
+
+    # No explicit task override → resolver consults persistent rows.
+    seen: set[str] = set()
+    for sample_idx in range(64):
+        rendered = render_sample(
+            recipe=recipe,
+            persistent=rephrasings,
+            events=[],
+            t=0.0,
+            sample_idx=sample_idx,
+            dataset_ctx={"task": "canonical kitchen task"},
+        )
+        seen.add(rendered["messages"][0]["content"])
+    # Every rephrasing should be reachable across enough samples.
+    assert seen == {r["content"] for r in rephrasings}
+    # Same sample_idx → same pick (determinism).
+    a = render_sample(
+        recipe=recipe, persistent=rephrasings, events=[], t=0.0, sample_idx=42,
+        dataset_ctx={"task": "canonical"},
+    )
+    b = render_sample(
+        recipe=recipe, persistent=rephrasings, events=[], t=0.0, sample_idx=42,
+        dataset_ctx={"task": "canonical"},
+    )
+    assert a["messages"][0]["content"] == b["messages"][0]["content"]
+
+
+def test_resolve_task_falls_back_to_canonical_without_rephrasings():
+    recipe = TrainingRecipe(
+        messages=[
+            MessageTurn(role="user", content="${task}", stream="high_level"),
+            MessageTurn(role="assistant", content="ok", stream="high_level", target=True),
+        ]
+    )
+    rendered = render_sample(
+        recipe=recipe,
+        persistent=PERSISTENT,  # no task_aug rows
+        events=[],
+        t=0.0,
+        sample_idx=0,
+        dataset_ctx={"task": "clean the kitchen"},
+    )
+    assert rendered["messages"][0]["content"] == "clean the kitchen"
+
+
+def test_resolve_task_explicit_override_beats_rephrasings():
+    rephrasings = [
+        persistent_row("user", "rephrased one", "task_aug", 0.0),
+        persistent_row("user", "rephrased two", "task_aug", 0.0),
+    ]
+    recipe = TrainingRecipe(
+        messages=[
+            MessageTurn(role="user", content="${task}", stream="high_level"),
+            MessageTurn(role="assistant", content="ok", stream="high_level", target=True),
+        ]
+    )
+    rendered = render_sample(
+        recipe=recipe,
+        persistent=rephrasings,
+        events=[],
+        t=0.0,
+        sample_idx=0,
+        task="explicit override wins",
+        dataset_ctx={"task": "canonical"},
+    )
+    assert rendered["messages"][0]["content"] == "explicit override wins"
+
+
+def test_canonical_recipe_can_render_low_level_branch():
+    recipe = TrainingRecipe.from_yaml(Path("src/lerobot/configs/recipes/pi05_hirobot.yaml"))
+    low_level = TrainingRecipe(blend={"low": recipe.blend["low_level_execution"]})
+
+    rendered = render_sample(
+        recipe=low_level,
+        persistent=PERSISTENT,
+        events=[],
+        t=0.5,
+        sample_idx=0,
+        task="clean kitchen",
+    )
+
+    assert rendered["messages"][-1] == {"role": "assistant", "content": "subtask 0"}
+    assert rendered["message_streams"][-1] == "low_level"
+    assert rendered["target_message_indices"] == [1]

tests/datasets/test_subtask_dataset.py

@@ -1,193 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-Tests for subtask functionality in LeRobotDataset.
-
-These tests verify that:
-- Subtask information is correctly loaded from datasets that have subtask data
-- The __getitem__ method correctly adds subtask strings to returned items
-- Subtask handling gracefully handles missing data
-"""
-
-import pytest
-
-pytest.importorskip("pandas", reason="pandas is required (install lerobot[dataset])")
-
-import pandas as pd  # noqa: E402
-import torch
-
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-
-class TestSubtaskDataset:
-    """Tests for subtask handling in LeRobotDataset."""
-
-    @pytest.fixture
-    def subtask_dataset(self):
-        """Load the test subtask dataset from the hub."""
-        # Use lerobot/pusht-subtask dataset with episode 1
-        return LeRobotDataset(
-            repo_id="lerobot/pusht-subtask",
-            episodes=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11],
-        )
-
-    def test_subtask_dataset_loads(self, subtask_dataset):
-        """Test that the subtask dataset loads successfully."""
-        assert subtask_dataset is not None
-        assert len(subtask_dataset) > 0
-
-    def test_subtask_metadata_loaded(self, subtask_dataset):
-        """Test that subtask metadata is loaded when present in dataset."""
-        # The dataset should have subtasks metadata loaded
-        assert subtask_dataset.meta.subtasks is not None
-        assert isinstance(subtask_dataset.meta.subtasks, pd.DataFrame)
-
-    def test_subtask_index_in_features(self, subtask_dataset):
-        """Test that subtask_index is a feature when dataset has subtasks."""
-        assert "subtask_index" in subtask_dataset.features
-
-    def test_getitem_returns_subtask_string(self, subtask_dataset):
-        """Test that __getitem__ correctly adds subtask string to returned item."""
-        item = subtask_dataset[0]
-
-        # Subtask should be present in the returned item
-        assert "subtask" in item
-        assert isinstance(item["subtask"], str)
-        assert len(item["subtask"]) > 0  # Should not be empty
-
-    def test_getitem_has_subtask_index(self, subtask_dataset):
-        """Test that __getitem__ includes subtask_index."""
-        item = subtask_dataset[0]
-
-        assert "subtask_index" in item
-        assert isinstance(item["subtask_index"], torch.Tensor)
-
-    def test_subtask_index_maps_to_valid_subtask(self, subtask_dataset):
-        """Test that subtask_index correctly maps to a subtask in metadata."""
-        item = subtask_dataset[0]
-
-        subtask_idx = item["subtask_index"].item()
-        subtask_from_metadata = subtask_dataset.meta.subtasks.iloc[subtask_idx].name
-
-        assert item["subtask"] == subtask_from_metadata
-
-    def test_all_items_have_subtask(self, subtask_dataset):
-        """Test that all items in the dataset have subtask information."""
-        for i in range(min(len(subtask_dataset), 5)):  # Check first 5 items
-            item = subtask_dataset[i]
-            assert "subtask" in item
-            assert isinstance(item["subtask"], str)
-
-    def test_task_and_subtask_coexist(self, subtask_dataset):
-        """Test that both task and subtask are present in returned items."""
-        item = subtask_dataset[0]
-
-        # Both task and subtask should be present
-        assert "task" in item
-        assert "subtask" in item
-        assert isinstance(item["task"], str)
-        assert isinstance(item["subtask"], str)
-
-
-class TestSubtaskDatasetMissing:
-    """Tests for graceful handling when subtask data is missing."""
-
-    @pytest.fixture
-    def dataset_without_subtasks(self, tmp_path, empty_lerobot_dataset_factory):
-        """Create a dataset without subtask information."""
-        features = {"state": {"dtype": "float32", "shape": (2,), "names": None}}
-        dataset = empty_lerobot_dataset_factory(root=tmp_path / "no_subtask", features=features)
-
-        # Add some frames and save
-        for _ in range(5):
-            dataset.add_frame({"state": torch.randn(2), "task": "Test task"})
-        dataset.save_episode()
-        dataset.finalize()
-
-        # Reload the dataset
-        return LeRobotDataset(dataset.repo_id, root=dataset.root)
-
-    def test_no_subtask_in_features(self, dataset_without_subtasks):
-        """Test that subtask_index is not in features when not provided."""
-        assert "subtask_index" not in dataset_without_subtasks.features
-
-    def test_getitem_without_subtask(self, dataset_without_subtasks):
-        """Test that __getitem__ works when subtask is not present."""
-        item = dataset_without_subtasks[0]
-
-        # Item should still be retrievable
-        assert item is not None
-        assert "state" in item
-        assert "task" in item
-
-        # Subtask should NOT be present
-        assert "subtask" not in item
-
-    def test_subtasks_metadata_is_none(self, dataset_without_subtasks):
-        """Test that subtasks metadata is None when not present."""
-        assert dataset_without_subtasks.meta.subtasks is None
-
-
-class TestSubtaskEdgeCases:
-    """Edge case tests for subtask handling."""
-
-    def test_subtask_with_multiple_episodes(self):
-        """Test subtask handling with multiple episodes if available."""
-        try:
-            dataset = LeRobotDataset(
-                repo_id="lerobot/pusht-subtask",
-                episodes=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11],
-            )
-        except Exception:
-            pytest.skip("Could not load test-subtask dataset")
-
-        # Check first and last items have valid subtasks
-        first_item = dataset[0]
-        last_item = dataset[len(dataset) - 1]
-
-        assert "subtask" in first_item
-        assert "subtask" in last_item
-        assert isinstance(first_item["subtask"], str)
-        assert isinstance(last_item["subtask"], str)
-
-    def test_subtask_index_consistency(self):
-        """Test that same subtask_index returns same subtask string."""
-        try:
-            dataset = LeRobotDataset(
-                repo_id="lerobot/pusht-subtask",
-                episodes=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11],
-            )
-        except Exception:
-            pytest.skip("Could not load test-subtask dataset")
-
-        if len(dataset) < 2:
-            pytest.skip("Dataset too small for this test")
-
-        # Collect subtask_index to subtask mappings
-        subtask_map = {}
-        for i in range(min(len(dataset), 10)):
-            item = dataset[i]
-            idx = item["subtask_index"].item()
-            subtask = item["subtask"]
-
-            if idx in subtask_map:
-                # Same index should always return same subtask
-                assert subtask_map[idx] == subtask, (
-                    f"Inconsistent subtask for index {idx}: '{subtask_map[idx]}' vs '{subtask}'"
-                )
-            else:
-                subtask_map[idx] = subtask

tests/envs/test_robotwin.py

@@ -0,0 +1,282 @@
+#!/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.
+"""Unit tests for the RoboTwin 2.0 Gymnasium wrapper.
+
+These tests mock out the SAPIEN-based RoboTwin runtime (task modules +
+YAML config loader) so they run without the full RoboTwin installation
+(SAPIEN, CuRobo, mplib, asset downloads, etc.).
+"""
+
+from __future__ import annotations
+
+from contextlib import contextmanager
+from unittest.mock import MagicMock, patch
+
+import gymnasium as gym
+import numpy as np
+import pytest
+
+from lerobot.envs.robotwin import (
+    ACTION_DIM,
+    ROBOTWIN_CAMERA_NAMES,
+    ROBOTWIN_TASKS,
+    RoboTwinEnv,
+    create_robotwin_envs,
+)
+
+# ---------------------------------------------------------------------------
+# Fixtures / helpers
+# ---------------------------------------------------------------------------
+
+
+def _make_mock_task_env(
+    height: int = 240,
+    width: int = 320,
+    cameras: tuple[str, ...] = ROBOTWIN_CAMERA_NAMES,
+) -> MagicMock:
+    """Return a mock that mimics the RoboTwin task class API.
+
+    RoboTwin's real get_obs returns
+        {"observation": {cam: {"rgb": img}}, "joint_action": {"vector": np.ndarray}, ...}
+    so the mock follows the same nested shape.
+    """
+    obs_dict = {
+        "observation": {cam: {"rgb": np.zeros((height, width, 3), dtype=np.uint8)} for cam in cameras},
+        "joint_action": {"vector": np.zeros(ACTION_DIM, dtype=np.float32)},
+        "endpose": {},
+    }
+
+    mock = MagicMock()
+    mock.get_obs.return_value = obs_dict
+    mock.setup_demo.return_value = None
+    mock.take_action.return_value = None
+    mock.eval_success = False
+    mock.check_success.return_value = False
+    mock.close_env.return_value = None
+    return mock
+
+
+@contextmanager
+def _patch_runtime(mock_task_instance: MagicMock):
+    """Patch both the task-class loader and the YAML config loader so the
+    env can construct + reset without a real RoboTwin install."""
+    task_cls = MagicMock(return_value=mock_task_instance)
+    fake_setup = {
+        "head_camera_h": 240,
+        "head_camera_w": 320,
+        "left_embodiment_config": {},
+        "right_embodiment_config": {},
+        "left_robot_file": "",
+        "right_robot_file": "",
+        "dual_arm_embodied": True,
+        "render_freq": 0,
+        "task_name": "beat_block_hammer",
+        "task_config": "demo_clean",
+    }
+    with (
+        patch("lerobot.envs.robotwin._load_robotwin_task", return_value=task_cls),
+        patch("lerobot.envs.robotwin._load_robotwin_setup_kwargs", return_value=fake_setup),
+    ):
+        yield
+
+
+# ---------------------------------------------------------------------------
+# RoboTwinEnv unit tests
+# ---------------------------------------------------------------------------
+
+
+class TestRoboTwinEnv:
+    def test_observation_space_shape(self):
+        """observation_space should have the configured h×w×3 for every camera."""
+        h, w = 240, 320
+        env = RoboTwinEnv(
+            task_name="beat_block_hammer",
+            observation_height=h,
+            observation_width=w,
+            camera_names=["head_camera", "left_camera"],
+        )
+        pixels_space = env.observation_space["pixels"]
+        assert pixels_space["head_camera"].shape == (h, w, 3)
+        assert pixels_space["left_camera"].shape == (h, w, 3)
+        assert "right_camera" not in pixels_space
+
+    def test_action_space(self):
+        env = RoboTwinEnv(task_name="beat_block_hammer")
+        assert env.action_space.shape == (ACTION_DIM,)
+        assert env.action_space.dtype == np.float32
+
+    def test_reset_returns_correct_obs_keys(self):
+        mock_task = _make_mock_task_env()
+        env = RoboTwinEnv(task_name="beat_block_hammer")
+        with _patch_runtime(mock_task):
+            obs, info = env.reset()
+
+        assert "pixels" in obs
+        for cam in ROBOTWIN_CAMERA_NAMES:
+            assert cam in obs["pixels"], f"Missing camera '{cam}' in obs"
+        assert "agent_pos" in obs
+        assert obs["agent_pos"].shape == (ACTION_DIM,)
+        assert info["is_success"] is False
+
+    def test_reset_calls_setup_demo(self):
+        mock_task = _make_mock_task_env()
+        env = RoboTwinEnv(task_name="beat_block_hammer")
+        with _patch_runtime(mock_task):
+            env.reset(seed=42)
+        # setup_demo receives the full YAML-derived kwargs plus seed + is_test;
+        # we only assert the caller-provided bits.
+        assert mock_task.setup_demo.call_count == 1
+        call_kwargs = mock_task.setup_demo.call_args.kwargs
+        assert call_kwargs["seed"] == 42
+        assert call_kwargs["is_test"] is True
+
+    def test_step_returns_correct_types(self):
+        mock_task = _make_mock_task_env()
+        env = RoboTwinEnv(task_name="beat_block_hammer")
+        action = np.zeros(ACTION_DIM, dtype=np.float32)
+        with _patch_runtime(mock_task):
+            env.reset()
+            obs, reward, terminated, truncated, info = env.step(action)
+
+        assert isinstance(obs, dict)
+        assert isinstance(reward, float)
+        assert isinstance(terminated, bool)
+        assert isinstance(truncated, bool)
+        assert isinstance(info, dict)
+
+    def test_step_wrong_action_shape_raises(self):
+        mock_task = _make_mock_task_env()
+        env = RoboTwinEnv(task_name="beat_block_hammer")
+        bad_action = np.zeros(7, dtype=np.float32)  # wrong dim
+        with _patch_runtime(mock_task):
+            env.reset()
+            with pytest.raises(ValueError, match="Expected 1-D action"):
+                env.step(bad_action)
+
+    def test_success_terminates_episode(self):
+        mock_task = _make_mock_task_env()
+        mock_task.check_success.return_value = True
+        env = RoboTwinEnv(task_name="beat_block_hammer")
+        action = np.zeros(ACTION_DIM, dtype=np.float32)
+        with _patch_runtime(mock_task):
+            env.reset()
+            _, _, terminated, _, info = env.step(action)
+        assert terminated is True
+        assert info["is_success"] is True
+
+    def test_truncation_after_episode_length(self):
+        mock_task = _make_mock_task_env()
+        env = RoboTwinEnv(task_name="beat_block_hammer", episode_length=2)
+        action = np.zeros(ACTION_DIM, dtype=np.float32)
+        with _patch_runtime(mock_task):
+            env.reset()
+            env.step(action)  # step 1
+            _, _, _, truncated, _ = env.step(action)  # step 2 → truncated
+        assert truncated is True
+
+    def test_close_calls_close_env(self):
+        mock_task = _make_mock_task_env()
+        env = RoboTwinEnv(task_name="beat_block_hammer")
+        with _patch_runtime(mock_task):
+            env.reset()
+            env.close()
+        mock_task.close_env.assert_called_once()
+
+    def test_black_frame_for_missing_camera(self):
+        """If a camera key is absent from get_obs(), a black frame is returned."""
+        # Mock exposes only head_camera; we ask for both head_camera + left_camera.
+        mock_task = _make_mock_task_env(height=10, width=10, cameras=("head_camera",))
+        env = RoboTwinEnv(
+            task_name="beat_block_hammer",
+            camera_names=["head_camera", "left_camera"],
+            observation_height=10,
+            observation_width=10,
+        )
+        with _patch_runtime(mock_task):
+            obs, _ = env.reset()
+        assert obs["pixels"]["left_camera"].shape == (10, 10, 3)
+        assert obs["pixels"]["left_camera"].sum() == 0
+
+    def test_task_and_task_description_attributes(self):
+        env = RoboTwinEnv(task_name="beat_block_hammer")
+        assert env.task == "beat_block_hammer"
+        assert isinstance(env.task_description, str)
+
+    def test_deferred_init_env_is_none_before_reset(self):
+        env = RoboTwinEnv(task_name="beat_block_hammer")
+        assert env._env is None  # noqa: SLF001  (testing internal state)
+
+
+# ---------------------------------------------------------------------------
+# create_robotwin_envs tests
+# ---------------------------------------------------------------------------
+
+
+class TestCreateRoboTwinEnvs:
+    def test_returns_correct_structure(self):
+        mock_task = _make_mock_task_env()
+        with _patch_runtime(mock_task):
+            envs = create_robotwin_envs(
+                task="beat_block_hammer",
+                n_envs=1,
+                env_cls=gym.vector.SyncVectorEnv,
+            )
+        assert "beat_block_hammer" in envs
+        assert 0 in envs["beat_block_hammer"]
+        assert isinstance(envs["beat_block_hammer"][0], gym.vector.SyncVectorEnv)
+
+    def test_multi_task(self):
+        mock_task = _make_mock_task_env()
+        with _patch_runtime(mock_task):
+            envs = create_robotwin_envs(
+                task="beat_block_hammer,click_bell",
+                n_envs=1,
+                env_cls=gym.vector.SyncVectorEnv,
+            )
+        assert set(envs.keys()) == {"beat_block_hammer", "click_bell"}
+
+    def test_unknown_task_raises(self):
+        with pytest.raises(ValueError, match="Unknown RoboTwin tasks"):
+            create_robotwin_envs(
+                task="not_a_real_task",
+                n_envs=1,
+                env_cls=gym.vector.SyncVectorEnv,
+            )
+
+    def test_invalid_n_envs_raises(self):
+        with pytest.raises(ValueError, match="n_envs must be a positive int"):
+            create_robotwin_envs(
+                task="beat_block_hammer",
+                n_envs=0,
+                env_cls=gym.vector.SyncVectorEnv,
+            )
+
+
+# ---------------------------------------------------------------------------
+# ROBOTWIN_TASKS list
+# ---------------------------------------------------------------------------
+
+
+def test_task_list_not_empty():
+    assert len(ROBOTWIN_TASKS) >= 50
+
+
+def test_all_tasks_are_strings():
+    assert all(isinstance(t, str) and t for t in ROBOTWIN_TASKS)
+
+
+def test_no_duplicate_tasks():
+    assert len(ROBOTWIN_TASKS) == len(set(ROBOTWIN_TASKS))

tests/processor/test_render_messages_processor.py

@@ -0,0 +1,56 @@
+#!/usr/bin/env python
+
+import torch
+
+from lerobot.configs.recipe import MessageTurn, TrainingRecipe
+from lerobot.processor.converters import create_transition
+from lerobot.processor.render_messages_processor import RenderMessagesStep
+from lerobot.types import TransitionKey
+
+
+def test_render_messages_step_noops_without_language_columns():
+    recipe = TrainingRecipe(
+        messages=[
+            MessageTurn(role="user", content="${task}", stream="high_level"),
+            MessageTurn(role="assistant", content="${subtask}", stream="low_level", target=True),
+        ]
+    )
+    transition = create_transition(complementary_data={"task": "do it"})
+
+    assert RenderMessagesStep(recipe)(transition) == transition
+
+
+def test_render_messages_step_renders_and_drops_raw_language():
+    recipe = TrainingRecipe(
+        messages=[
+            MessageTurn(role="user", content="${task}", stream="high_level"),
+            MessageTurn(role="assistant", content="${subtask}", stream="low_level", target=True),
+        ]
+    )
+    transition = create_transition(
+        complementary_data={
+            "task": "do it",
+            "timestamp": torch.tensor(0.0),
+            "index": torch.tensor(7),
+            "language_persistent": [
+                {
+                    "role": "assistant",
+                    "content": "reach carefully",
+                    "style": "subtask",
+                    "timestamp": 0.0,
+                    "camera": None,
+                    "tool_calls": None,
+                }
+            ],
+            "language_events": [],
+        }
+    )
+
+    out = RenderMessagesStep(recipe)(transition)
+    data = out[TransitionKey.COMPLEMENTARY_DATA]
+
+    assert "language_persistent" not in data
+    assert "language_events" not in data
+    assert data["messages"][-1]["content"] == "reach carefully"
+    assert data["message_streams"] == ["high_level", "low_level"]
+    assert data["target_message_indices"] == [1]

tests/test_robomme_env.py

@@ -0,0 +1,232 @@
+# 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.
+"""Unit tests for the RoboMME env wrapper and config.
+
+RoboMME requires Linux + ManiSkill (Vulkan/SAPIEN), so tests that touch the
+env wrapper mock the ``robomme`` package. Tests that only exercise the
+dataclass config run without any mocking.
+"""
+
+from __future__ import annotations
+
+import sys
+from types import ModuleType
+from unittest.mock import MagicMock
+
+import numpy as np
+
+
+def _install_robomme_stub():
+    """Register a minimal stub for the ``robomme`` package on sys.modules."""
+    stub = ModuleType("robomme")
+    wrapper_stub = ModuleType("robomme.env_record_wrapper")
+
+    class FakeBuilder:
+        def __init__(self, **kwargs):
+            pass
+
+        def make_env_for_episode(self, episode_idx: int, max_steps: int):
+            env = MagicMock()
+            obs = {
+                "front_rgb_list": [np.zeros((256, 256, 3), dtype=np.uint8)],
+                "wrist_rgb_list": [np.zeros((256, 256, 3), dtype=np.uint8)],
+                "joint_state_list": [np.zeros(7, dtype=np.float32)],
+                "gripper_state_list": [np.zeros(2, dtype=np.float32)],
+            }
+            env.reset.return_value = (obs, {"status": "ongoing", "task_goal": "pick the cube"})
+            env.step.return_value = (obs, 0.0, False, False, {"status": "ongoing", "task_goal": ""})
+            return env
+
+    wrapper_stub.BenchmarkEnvBuilder = FakeBuilder
+    stub.env_record_wrapper = wrapper_stub
+    sys.modules["robomme"] = stub
+    sys.modules["robomme.env_record_wrapper"] = wrapper_stub
+
+
+def _uninstall_robomme_stub():
+    sys.modules.pop("robomme", None)
+    sys.modules.pop("robomme.env_record_wrapper", None)
+
+
+# ---------------------------------------------------------------------------
+# Config tests (no sim required)
+# ---------------------------------------------------------------------------
+
+
+def test_robomme_env_config_defaults():
+    from lerobot.envs.configs import RoboMMEEnv
+
+    cfg = RoboMMEEnv()
+    assert cfg.task == "PickXtimes"
+    assert cfg.fps == 10
+    assert cfg.episode_length == 300
+    assert cfg.action_space == "joint_angle"
+    assert cfg.dataset_split == "test"
+    assert cfg.task_ids is None
+
+
+def test_robomme_env_config_type():
+    from lerobot.envs.configs import RoboMMEEnv
+
+    cfg = RoboMMEEnv()
+    assert cfg.type == "robomme"
+
+
+def test_robomme_features_map():
+    from lerobot.envs.configs import RoboMMEEnv
+    from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
+
+    cfg = RoboMMEEnv()
+    assert cfg.features_map[ACTION] == ACTION
+    assert cfg.features_map["pixels/image"] == f"{OBS_IMAGES}.image"
+    assert cfg.features_map["pixels/wrist_image"] == f"{OBS_IMAGES}.wrist_image"
+    assert cfg.features_map["agent_pos"] == OBS_STATE
+
+
+def test_robomme_features_action_dim_joint_angle():
+    from lerobot.envs.configs import RoboMMEEnv
+    from lerobot.utils.constants import ACTION
+
+    cfg = RoboMMEEnv(action_space="joint_angle")
+    assert cfg.features[ACTION].shape == (8,)
+
+
+def test_robomme_features_action_dim_ee_pose():
+    """`ee_pose` uses a 7-D action; __post_init__ sets the correct shape."""
+    from lerobot.envs.configs import RoboMMEEnv
+    from lerobot.utils.constants import ACTION
+
+    cfg = RoboMMEEnv(action_space="ee_pose")
+    assert cfg.features[ACTION].shape == (7,)
+
+
+# ---------------------------------------------------------------------------
+# Obs conversion (pure Python, no sim)
+# ---------------------------------------------------------------------------
+
+
+def test_convert_obs_list_format():
+    """_convert_obs takes the last element from list-format obs fields and
+    emits a nested ``pixels`` dict (image, wrist_image) plus ``agent_pos``.
+
+    The nested layout is required so ``preprocess_observation()`` in
+    ``envs/utils.py`` maps each camera to ``observation.images.<cam>``.
+    """
+    _install_robomme_stub()
+    try:
+        from lerobot.envs.robomme import RoboMMEGymEnv
+
+        env = RoboMMEGymEnv.__new__(RoboMMEGymEnv)
+
+        front = np.full((256, 256, 3), 42, dtype=np.uint8)
+        wrist = np.full((256, 256, 3), 7, dtype=np.uint8)
+        joints = np.arange(7, dtype=np.float32)
+        gripper = np.array([0.5, 0.5], dtype=np.float32)
+
+        obs_raw = {
+            "front_rgb_list": [np.zeros_like(front), front],
+            "wrist_rgb_list": [np.zeros_like(wrist), wrist],
+            "joint_state_list": [np.zeros(7, dtype=np.float32), joints],
+            "gripper_state_list": [np.zeros(2, dtype=np.float32), gripper],
+        }
+
+        result = env._convert_obs(obs_raw)
+        np.testing.assert_array_equal(result["pixels"]["image"], front)
+        np.testing.assert_array_equal(result["pixels"]["wrist_image"], wrist)
+        assert result["agent_pos"].shape == (8,)
+        np.testing.assert_array_almost_equal(result["agent_pos"][:7], joints)
+        assert result["agent_pos"][7] == gripper[0]
+    finally:
+        _uninstall_robomme_stub()
+
+
+def test_convert_obs_array_format():
+    """_convert_obs also handles non-list (direct array) obs."""
+    _install_robomme_stub()
+    try:
+        from lerobot.envs.robomme import RoboMMEGymEnv
+
+        env = RoboMMEGymEnv.__new__(RoboMMEGymEnv)
+
+        front = np.zeros((256, 256, 3), dtype=np.uint8)
+        obs_raw = {
+            "front_rgb_list": front,
+            "wrist_rgb_list": front,
+            "joint_state_list": np.zeros(7, dtype=np.float32),
+            "gripper_state_list": np.zeros(2, dtype=np.float32),
+        }
+        result = env._convert_obs(obs_raw)
+        assert result["pixels"]["image"].shape == (256, 256, 3)
+        assert result["pixels"]["wrist_image"].shape == (256, 256, 3)
+        assert result["agent_pos"].shape == (8,)
+    finally:
+        _uninstall_robomme_stub()
+
+
+# ---------------------------------------------------------------------------
+# create_robomme_envs (mocked sim)
+# ---------------------------------------------------------------------------
+
+
+def test_create_robomme_envs_returns_correct_structure():
+    """Single task -> {task_name: {task_id: VectorEnv}} with one entry per task_id."""
+    _install_robomme_stub()
+    try:
+        from lerobot.envs.robomme import create_robomme_envs
+
+        env_cls = MagicMock(return_value=MagicMock())
+        result = create_robomme_envs(
+            task="PickXtimes",
+            n_envs=1,
+            task_ids=[0, 1],
+            env_cls=env_cls,
+        )
+
+        assert "PickXtimes" in result
+        assert 0 in result["PickXtimes"]
+        assert 1 in result["PickXtimes"]
+        assert env_cls.call_count == 2
+    finally:
+        _uninstall_robomme_stub()
+
+
+def test_create_robomme_envs_multi_task():
+    """Comma-separated task list produces one suite per task."""
+    _install_robomme_stub()
+    try:
+        from lerobot.envs.robomme import create_robomme_envs
+
+        env_cls = MagicMock(return_value=MagicMock())
+        result = create_robomme_envs(
+            task="PickXtimes,BinFill,StopCube",
+            n_envs=1,
+            env_cls=env_cls,
+        )
+
+        assert set(result.keys()) == {"PickXtimes", "BinFill", "StopCube"}
+    finally:
+        _uninstall_robomme_stub()
+
+
+def test_create_robomme_envs_raises_on_invalid_env_cls():
+    _install_robomme_stub()
+    try:
+        import pytest
+
+        from lerobot.envs.robomme import create_robomme_envs
+
+        with pytest.raises(ValueError, match="env_cls must be a callable"):
+            create_robomme_envs(task="PickXtimes", n_envs=1, env_cls=None)
+    finally:
+        _uninstall_robomme_stub()

tests/utils/test_collate.py

@@ -0,0 +1,36 @@
+#!/usr/bin/env python
+
+import torch
+
+from lerobot.utils.collate import lerobot_collate_fn
+
+
+def test_lerobot_collate_preserves_messages_and_drops_raw_language():
+    batch = [
+        {
+            "index": torch.tensor(0),
+            "messages": [{"role": "assistant", "content": "a"}],
+            "message_streams": ["low_level"],
+            "target_message_indices": [0],
+            "language_persistent": [{"content": "raw"}],
+            "language_events": [],
+        },
+        {
+            "index": torch.tensor(1),
+            "messages": [{"role": "assistant", "content": "b"}],
+            "message_streams": ["low_level"],
+            "target_message_indices": [0],
+            "language_persistent": [{"content": "raw"}],
+            "language_events": [],
+        },
+    ]
+
+    out = lerobot_collate_fn(batch)
+
+    assert out["index"].tolist() == [0, 1]
+    assert out["messages"][0][0]["content"] == "a"
+    assert out["messages"][1][0]["content"] == "b"
+    assert out["message_streams"] == [["low_level"], ["low_level"]]
+    assert out["target_message_indices"] == [[0], [0]]
+    assert "language_persistent" not in out
+    assert "language_events" not in out