refactor(inference): subtask = head of plan, drop AR generation

Replaced the LM-head subtask generator with a deterministic
plan-walker. The plan is already a numbered list of still-todo
subtasks (produced by ``plan_generation``); the current subtask is
just the head of that list, popped one position per chunk-dispatch
boundary.

Why
---
The previous design called ``select_message`` every ~1 Hz to
auto-regressively generate the current subtask string. On a small
training set with a chat-pretrained backbone, the LM head collapses
into n-gram loops under greedy decoding —

  "the robot arm extends and retracts and retracts and retracts
   from the beige surface and retracts from the surface"

— which then poisons the action expert's language conditioning,
because the same garbage string is what gets rendered into its
prefix. Repetition penalty / no_repeat_ngram_size help but are
bandaids.

Walking the plan deterministically eliminates the failure mode at
the root: the action expert always sees a real, recipe-aligned
subtask phrase. Inference is also interpretable now — every robot
motion traces back to a specific plan line — and we save the per-
chunk ``select_message`` round trip (~2 s on MPS).

The LM head is still trained (``high_level_subtask`` text-CE) and
still used at runtime for one-shot plan generation, interjection
replanning, and VQA. We just don't sit on the AR subtask loop.

Advance trigger: ``actions_dispatched`` increment — each new chunk
the action expert produced and dispatched advances ``plan_index``
by one. Clamps to the last plan line so the robot keeps executing
the final subtask if it runs off the end instead of going silent.
Falls back to the task string when no plan is set yet (e.g. before
``plan_generation`` has fired this episode).

Added ``_parse_plan_lines`` helper for stripping the ``"N. "``
prefix off plan items.

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

.github/PULL_REQUEST_TEMPLATE.md

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

.github/workflows/benchmark_tests.yml

@@ -83,10 +83,13 @@ jobs:
           cache-binary: false
 
       - name: Login to Docker Hub
+        if: ${{ env.DOCKERHUB_USERNAME != '' }}
         uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
         with:
           username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
           password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+        env:
+          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
 
       # Build the benchmark-specific image. The Dockerfile separates dep-install
       # from source-copy, so code-only changes skip the slow uv-sync layer
@@ -115,7 +118,7 @@ jobs:
             bash -c "
               hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
               lerobot-eval \
-                --policy.path=pepijn223/smolvla_libero \
+                --policy.path=lerobot/smolvla_libero \
                 --env.type=libero \
                 --env.task=libero_spatial \
                 --eval.batch_size=1 \
@@ -144,7 +147,7 @@ jobs:
             --artifacts-dir /tmp/libero-artifacts \
             --env libero \
             --task libero_spatial \
-            --policy pepijn223/smolvla_libero
+            --policy lerobot/smolvla_libero
 
       - name: Upload Libero rollout video
         if: always()
@@ -238,10 +241,13 @@ jobs:
           cache-binary: false
 
       - name: Login to Docker Hub
+        if: ${{ env.DOCKERHUB_USERNAME != '' }}
         uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
         with:
           username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
           password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+        env:
+          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
 
       - name: Build MetaWorld benchmark image
         uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
@@ -264,7 +270,7 @@ jobs:
             bash -c "
               hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
               lerobot-eval \
-                --policy.path=pepijn223/smolvla_metaworld \
+                --policy.path=lerobot/smolvla_metaworld \
                 --env.type=metaworld \
                 --env.task=metaworld-push-v3 \
                 --eval.batch_size=1 \
@@ -293,7 +299,7 @@ jobs:
             --artifacts-dir /tmp/metaworld-artifacts \
             --env metaworld \
             --task metaworld-push-v3 \
-            --policy pepijn223/smolvla_metaworld
+            --policy lerobot/smolvla_metaworld
 
       - name: Upload MetaWorld rollout video
         if: always()
@@ -310,3 +316,630 @@ jobs:
           name: metaworld-metrics
           path: /tmp/metaworld-artifacts/metrics.json
           if-no-files-found: warn
+
+  # ── ROBOTWIN 2.0 ──────────────────────────────────────────────────────────
+  # Isolated image: full RoboTwin 2.0 stack — SAPIEN, mplib, CuRobo,
+  # pytorch3d, + simulation assets (~4 GB).
+  # Build takes ~20 min on first run; subsequent runs hit the layer cache.
+  # Requires an NVIDIA GPU runner with CUDA 12.1 drivers.
+  robotwin-integration-test:
+    name: RoboTwin 2.0 — build image + 1-episode eval
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
+      ROBOTWIN_POLICY: lerobot/smolvla_robotwin
+      ROBOTWIN_TASKS: beat_block_hammer,click_bell,handover_block,stack_blocks_two,click_alarmclock,open_microwave,adjust_bottle,lift_pot,stamp_seal,turn_switch
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+          lfs: true
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          cache-binary: false
+
+      - name: Login to Docker Hub
+        if: ${{ env.DOCKERHUB_USERNAME != '' }}
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+        env:
+          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+
+      # Build the full-install image: SAPIEN, mplib, CuRobo, pytorch3d +
+      # simulation assets (~4 GB). Layer cache lives in the runner's local
+      # Docker daemon — reused across re-runs on the same machine.
+      - name: Build RoboTwin 2.0 benchmark image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: docker/Dockerfile.benchmark.robotwin
+          push: false
+          load: true
+          tags: lerobot-benchmark-robotwin:ci
+          cache-from: type=local,src=/tmp/.buildx-cache-robotwin
+          cache-to: type=local,dest=/tmp/.buildx-cache-robotwin,mode=max
+
+      - name: Run RoboTwin 2.0 smoke eval (10 tasks, 1 episode each)
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          # Named container (no --rm) so we can docker cp artifacts out.
+          docker run --name robotwin-eval --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e ROBOTWIN_POLICY="${ROBOTWIN_POLICY}" \
+            -e ROBOTWIN_TASKS="${ROBOTWIN_TASKS}" \
+            lerobot-benchmark-robotwin:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              cd /opt/robotwin && lerobot-eval \
+                --policy.path=\"\$ROBOTWIN_POLICY\" \
+                --env.type=robotwin \
+                --env.task=\"\$ROBOTWIN_TASKS\" \
+                --eval.batch_size=1 \
+                --eval.n_episodes=1 \
+                --eval.use_async_envs=false \
+                --policy.device=cuda \
+                '--rename_map={\"observation.images.head_camera\": \"observation.images.camera1\", \"observation.images.left_camera\": \"observation.images.camera2\", \"observation.images.right_camera\": \"observation.images.camera3\"}' \
+                --output_dir=/tmp/eval-artifacts
+              python /lerobot/scripts/ci/extract_task_descriptions.py \
+                --env robotwin \
+                --task \"\$ROBOTWIN_TASKS\" \
+                --output /tmp/eval-artifacts/task_descriptions.json
+            "
+
+      - name: Copy RoboTwin artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/robotwin-artifacts
+          docker cp robotwin-eval:/tmp/eval-artifacts/. /tmp/robotwin-artifacts/ 2>/dev/null || true
+          docker rm -f robotwin-eval || true
+
+      - name: Parse RoboTwin eval metrics
+        if: always()
+        run: |
+          python3 scripts/ci/parse_eval_metrics.py \
+            --artifacts-dir /tmp/robotwin-artifacts \
+            --env robotwin \
+            --task "${ROBOTWIN_TASKS}" \
+            --policy "${ROBOTWIN_POLICY}"
+
+      - name: Upload RoboTwin rollout video
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: robotwin-rollout-video
+          path: /tmp/robotwin-artifacts/videos/
+          if-no-files-found: warn
+
+      - name: Upload RoboTwin eval metrics
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: robotwin-metrics
+          path: /tmp/robotwin-artifacts/metrics.json
+          if-no-files-found: warn
+
+  # ── ROBOCASA365 ──────────────────────────────────────────────────────────
+  # Isolated image: robocasa + robosuite installed manually as editable
+  # clones (no `lerobot[robocasa]` extra — robocasa's setup.py pins
+  # `lerobot==0.3.3`, which would shadow this repo's lerobot).
+  robocasa-integration-test:
+    name: RoboCasa365 — build image + 1-episode eval
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+          lfs: true
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          cache-binary: false
+
+      - name: Login to Docker Hub
+        if: ${{ env.DOCKERHUB_USERNAME != '' }}
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+        env:
+          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+
+      - name: Build RoboCasa365 benchmark image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: docker/Dockerfile.benchmark.robocasa
+          push: false
+          load: true
+          tags: lerobot-benchmark-robocasa:ci
+
+      - name: Run RoboCasa365 smoke eval (10 atomic tasks, 1 episode each)
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          docker run --name robocasa-eval --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            -e MUJOCO_GL=egl \
+            lerobot-benchmark-robocasa:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              lerobot-eval \
+                --policy.path=lerobot/smolvla_robocasa \
+                --env.type=robocasa \
+                --env.task=CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove,CloseToasterOvenDoor,SlideDishwasherRack,TurnOnSinkFaucet,NavigateKitchen,TurnOnElectricKettle \
+                --eval.batch_size=1 \
+                --eval.n_episodes=1 \
+                --eval.use_async_envs=false \
+                --policy.device=cuda \
+                '--rename_map={\"observation.images.robot0_agentview_left\": \"observation.images.camera1\", \"observation.images.robot0_eye_in_hand\": \"observation.images.camera2\", \"observation.images.robot0_agentview_right\": \"observation.images.camera3\"}' \
+                --output_dir=/tmp/eval-artifacts
+              python scripts/ci/extract_task_descriptions.py \
+                --env robocasa \
+                --task CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove,CloseToasterOvenDoor,SlideDishwasherRack,TurnOnSinkFaucet,NavigateKitchen,TurnOnElectricKettle \
+                --output /tmp/eval-artifacts/task_descriptions.json
+            "
+
+      - name: Copy RoboCasa365 artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/robocasa-artifacts
+          docker cp robocasa-eval:/tmp/eval-artifacts/. /tmp/robocasa-artifacts/ 2>/dev/null || true
+          docker rm -f robocasa-eval || true
+
+      - name: Parse RoboCasa365 eval metrics
+        if: always()
+        run: |
+          python3 scripts/ci/parse_eval_metrics.py \
+            --artifacts-dir /tmp/robocasa-artifacts \
+            --env robocasa \
+            --task atomic_smoke_10 \
+            --policy lerobot/smolvla_robocasa
+
+      - name: Upload RoboCasa365 rollout video
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: robocasa-rollout-video
+          path: /tmp/robocasa-artifacts/videos/
+          if-no-files-found: warn
+
+      - name: Upload RoboCasa365 eval metrics
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: robocasa-metrics
+          path: /tmp/robocasa-artifacts/metrics.json
+          if-no-files-found: warn
+
+  # ── ROBOCEREBRA ───────────────────────────────────────────────────────────
+  # Reuses the LIBERO simulator (libero_10 suite) with RoboCerebra camera
+  # defaults (image/wrist_image). The image is layered on
+  # huggingface/lerobot-gpu, which already ships [libero] as part of [all].
+  robocerebra-integration-test:
+    name: RoboCerebra — build image + 1-episode eval
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+          lfs: true
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          cache-binary: false
+
+      - name: Login to Docker Hub
+        if: ${{ env.DOCKERHUB_USERNAME != '' }}
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+        env:
+          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+
+      - name: Build RoboCerebra benchmark image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: docker/Dockerfile.benchmark.robocerebra
+          push: false
+          load: true
+          tags: lerobot-benchmark-robocerebra:ci
+          cache-from: type=local,src=/tmp/.buildx-cache-robocerebra
+          cache-to: type=local,dest=/tmp/.buildx-cache-robocerebra,mode=max
+
+      - name: Run RoboCerebra smoke eval (1 episode)
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          docker run --name robocerebra-eval --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            -e LIBERO_DATA_FOLDER=/tmp/libero_data \
+            lerobot-benchmark-robocerebra:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              lerobot-eval \
+                --policy.path=lerobot/smolvla_robocerebra \
+                --env.type=libero \
+                --env.task=libero_10 \
+                --env.fps=20 \
+                --env.obs_type=pixels_agent_pos \
+                --env.observation_height=256 \
+                --env.observation_width=256 \
+                '--env.camera_name_mapping={\"agentview_image\": \"image\", \"robot0_eye_in_hand_image\": \"wrist_image\"}' \
+                --eval.batch_size=1 \
+                --eval.n_episodes=1 \
+                --eval.use_async_envs=false \
+                --policy.device=cuda \
+                '--rename_map={\"observation.images.image\": \"observation.images.camera1\", \"observation.images.wrist_image\": \"observation.images.camera2\"}' \
+                --policy.empty_cameras=1 \
+                --output_dir=/tmp/eval-artifacts
+              python scripts/ci/extract_task_descriptions.py \
+                --env libero --task libero_10 \
+                --output /tmp/eval-artifacts/task_descriptions.json
+            "
+
+      - name: Copy RoboCerebra artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/robocerebra-artifacts
+          docker cp robocerebra-eval:/tmp/eval-artifacts/. /tmp/robocerebra-artifacts/ 2>/dev/null || true
+          docker rm -f robocerebra-eval || true
+
+      - name: Parse RoboCerebra eval metrics
+        if: always()
+        run: |
+          python3 scripts/ci/parse_eval_metrics.py \
+            --artifacts-dir /tmp/robocerebra-artifacts \
+            --env robocerebra \
+            --task libero_10 \
+            --policy lerobot/smolvla_robocerebra
+
+      - name: Upload RoboCerebra rollout video
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: robocerebra-rollout-video
+          path: /tmp/robocerebra-artifacts/videos/
+          if-no-files-found: warn
+
+      - name: Upload RoboCerebra eval metrics
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: robocerebra-metrics
+          path: /tmp/robocerebra-artifacts/metrics.json
+          if-no-files-found: warn
+
+  # ── ROBOMME ───────────────────────────────────────────────────────────────
+  # Isolated image: mani-skill/SAPIEN/Vulkan chain with gymnasium and numpy
+  # overrides (robomme can't be a pyproject extra due to numpy<2 pin).
+  robomme-integration-test:
+    name: RoboMME — build image + 1-episode eval
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
+      ROBOMME_POLICY: lerobot/smolvla_robomme
+      ROBOMME_TASKS: PickXtimes,BinFill,StopCube,MoveCube,InsertPeg,SwingXtimes,VideoUnmask,ButtonUnmask,PickHighlight,PatternLock
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        with:
+          persist-credentials: false
+          lfs: true
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          cache-binary: false
+
+      - name: Login to Docker Hub
+        if: ${{ env.DOCKERHUB_USERNAME != '' }}
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+        env:
+          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+
+      - name: Build RoboMME benchmark image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: docker/Dockerfile.benchmark.robomme
+          push: false
+          load: true
+          tags: lerobot-benchmark-robomme:ci
+
+      - name: Run RoboMME smoke eval (10 tasks, 1 episode each)
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          docker run --name robomme-eval --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            -e ROBOMME_POLICY="${ROBOMME_POLICY}" \
+            -e ROBOMME_TASKS="${ROBOMME_TASKS}" \
+            lerobot-benchmark-robomme:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              lerobot-eval \
+                --policy.path=\"\$ROBOMME_POLICY\" \
+                --env.type=robomme \
+                --env.task=\"\$ROBOMME_TASKS\" \
+                --env.dataset_split=test \
+                --env.task_ids=[0] \
+                --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

.github/workflows/documentation-upload-pr.yml

@@ -33,7 +33,7 @@ jobs:
       github.event.workflow_run.event == 'pull_request' &&
       github.event.workflow_run.conclusion == 'success' &&
       github.repository == 'huggingface/lerobot'
-    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@9ad2de8582b56c017cb530c1165116d40433f1c6  # main
     with:
       package_name: lerobot
     secrets:

.github/workflows/latest_deps_tests.yml

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

CONTRIBUTING.md

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

Makefile

@@ -178,3 +178,9 @@ test-smolvla-ete-eval:
 		--env.episode_length=5 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1
+
+# E2E annotation pipeline smoke test against a tiny in-memory fixture
+# dataset. Opt-in (not part of `make test-end-to-end`) and uses a stub VLM
+# backend, so it does not require a real model checkpoint or GPU.
+annotation-e2e:
+	uv run python -m tests.annotations.run_e2e_smoke

docker/Dockerfile.benchmark.libero_plus

@@ -0,0 +1,84 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Benchmark image for LIBERO-plus integration tests.
+# Extends the nightly GPU image (which has lerobot[all]) with the LIBERO-plus
+# fork source + its 6.4 GB perturbation assets.
+#
+# Build:  docker build -f docker/Dockerfile.benchmark.libero_plus -t lerobot-benchmark-libero-plus .
+# Run:    docker run --gpus all --rm lerobot-benchmark-libero-plus lerobot-eval ...
+
+FROM huggingface/lerobot-gpu:latest
+ENV MUJOCO_GL=egl
+
+# unzip for the 6.4 GB assets.zip; the rest are LIBERO-plus build-time extras
+# (wand / ImageMagick / fontconfig) not in the nightly base.
+USER root
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+         unzip libexpat1 libfontconfig1-dev libmagickwand-dev \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+USER user_lerobot
+
+# robosuite==1.4.1 is mandatory (the fork uses `single_arm_env` removed in
+# v1.5+). The rest are LIBERO-plus runtime deps pulled from its setup.py.
+# We install these explicitly instead of via the [libero_plus] extra because
+# the extra's `libero @ git+...` dep installs as a namespace package and then
+# clone and PYTHONPATH-override it below.
+RUN uv pip install --no-cache \
+        "robosuite==1.4.1" \
+        "bddl==1.0.1" \
+        "easydict==1.13" \
+        "mujoco==3.7.0" \
+        "matplotlib==3.10.8" \
+        "Wand==0.6.13" \
+        "scikit-image==0.25.2" \
+        "gym==0.26.2"
+
+# Clone LIBERO-plus and make it importable as `libero`. The nightly base has
+# hf-libero (10 tasks) preinstalled via lerobot[libero]; uninstall it so
+# Python resolves `import libero` to the 2402-task LIBERO-plus module instead.
+# Pinned to the current upstream main SHA so benchmark builds stay reproducible.
+ARG LIBERO_PLUS_SHA=4976dc3
+ENV LIBERO_PLUS_ROOT=/home/user_lerobot/libero-plus/libero/libero
+RUN git clone https://github.com/sylvestf/LIBERO-plus.git /home/user_lerobot/libero-plus \
+    && git -C /home/user_lerobot/libero-plus checkout ${LIBERO_PLUS_SHA} \
+    && cd /home/user_lerobot/libero-plus && uv pip install --no-cache --no-deps -e "." \
+    && (uv pip uninstall hf-libero 2>/dev/null || true)
+ENV PYTHONPATH="/home/user_lerobot/libero-plus:${PYTHONPATH}"
+
+# Perturbation textures/scenes: bddl_base_domain.py resolves XMLs via
+# DIR_PATH/../assets (package-relative, ignoring ~/.libero/config.yaml). All
+# 2402 tasks reference files that ship only in Sylvest/LIBERO-plus's
+# assets.zip (6.4 GB) under a deep author-internal prefix — extract and
+# flatten it under ${LIBERO_PLUS_ROOT}/assets.
+RUN python -c "\
+from huggingface_hub import hf_hub_download; \
+hf_hub_download(repo_id='Sylvest/LIBERO-plus', repo_type='dataset', \
+                filename='assets.zip', local_dir='/tmp/libero-plus-dl')" \
+    && unzip -q /tmp/libero-plus-dl/assets.zip -d /tmp/libero-plus-dl/extract \
+    && ASSETS_DIR=$(find /tmp/libero-plus-dl/extract -type d -name assets | head -1) \
+    && mv "${ASSETS_DIR}" ${LIBERO_PLUS_ROOT}/assets \
+    && rm -rf /tmp/libero-plus-dl
+
+# Point ~/.libero/config.yaml at the clone so LIBERO-plus's imports are
+# non-interactive (it calls input() when the config is missing).
+RUN mkdir -p /home/user_lerobot/.libero \
+    && printf "assets: ${LIBERO_PLUS_ROOT}/assets\nbddl_files: ${LIBERO_PLUS_ROOT}/bddl_files\ndatasets: ${LIBERO_PLUS_ROOT}/../datasets\ninit_states: ${LIBERO_PLUS_ROOT}/init_files\n" \
+       > /home/user_lerobot/.libero/config.yaml
+
+# Overlay the PR's source code on top of the nightly image.
+COPY --chown=user_lerobot:user_lerobot . .
+
+CMD ["/bin/bash"]

docker/Dockerfile.benchmark.robocasa

@@ -0,0 +1,71 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Benchmark image for RoboCasa365 integration tests.
+# Extends the nightly GPU image (which already has all extras installed)
+# with the PR's source code and RoboCasa-specific asset setup.
+#
+# Build:  docker build -f docker/Dockerfile.benchmark.robocasa -t lerobot-benchmark-robocasa .
+# Run:    docker run --gpus all --rm lerobot-benchmark-robocasa lerobot-eval ...
+
+FROM huggingface/lerobot-gpu:latest
+
+# Install robocasa + robosuite as editable clones. pip-installing from git
+# omits data files like robocasa/models/assets/box_links/box_links_assets.json
+# (not declared in package_data), which download_kitchen_assets needs at import.
+#
+# `--no-deps` on robocasa is deliberate: its setup.py pins `lerobot==0.3.3`
+# in install_requires, which would shadow the editable lerobot baked into
+# this image. We install robocasa's actual runtime deps explicitly instead.
+# Pinned SHAs for reproducible benchmark runs. Bump when you need an
+# upstream fix; don't rely on `main`/`master` drift.
+ARG ROBOCASA_SHA=56e355ccc64389dfc1b8a61a33b9127b975ba681
+ARG ROBOSUITE_SHA=aaa8b9b214ce8e77e82926d677b4d61d55e577ab
+RUN git clone https://github.com/robocasa/robocasa.git ~/robocasa && \
+    git -C ~/robocasa checkout ${ROBOCASA_SHA} && \
+    git clone https://github.com/ARISE-Initiative/robosuite.git ~/robosuite && \
+    git -C ~/robosuite checkout ${ROBOSUITE_SHA} && \
+    uv pip install --no-cache -e ~/robocasa --no-deps && \
+    uv pip install --no-cache -e ~/robosuite && \
+    uv pip install --no-cache \
+      "numpy==2.2.5" "numba==0.61.2" "scipy==1.15.3" "mujoco==3.3.1" \
+      "pygame==2.6.1" "Pillow==12.2.0" "opencv-python==4.13.0.92" \
+      "pyyaml==6.0.3" "pynput==1.8.1" "tqdm==4.67.3" "termcolor==3.3.0" \
+      "imageio==2.37.3" "h5py==3.16.0" "lxml==6.0.4" "hidapi==0.14.0.post4" \
+      "tianshou==0.4.10" "gymnasium==1.2.3"
+
+# Set up robocasa macros and download kitchen assets. We need:
+#   - tex              : base environment textures
+#   - tex_generative   : AI-generated textures; kitchen fixture XMLs embed
+#                        refs to generative_textures/wall/tex*.png
+#                        unconditionally, so MjModel.from_xml_string fails
+#                        at reset time without them (even if the env is
+#                        constructed with generative_textures=None).
+#   - fixtures_lw      : lightwheel kitchen fixtures (fridge, counters...)
+#   - objs_lw          : lightwheel object meshes (stools, misc props)
+# We skip the objaverse/aigen object packs (~30GB combined) by pairing
+# this with --env.obj_registries=["lightwheel"] on the lerobot side.
+# The download script prompts interactively, so pipe 'y' to auto-accept.
+RUN python -m robocasa.scripts.setup_macros && \
+    yes y | python -m robocasa.scripts.download_kitchen_assets \
+      --type tex tex_generative fixtures_lw objs_lw
+
+# Overlay the PR's source code on top of the nightly image.
+COPY --chown=user_lerobot:user_lerobot . .
+
+# Re-install lerobot editably so the new source (with RoboCasaEnv registration)
+# replaces the stale package baked into the nightly image.
+RUN uv pip install --no-cache --no-deps -e .
+
+CMD ["/bin/bash"]

docker/Dockerfile.benchmark.robocerebra

@@ -0,0 +1,43 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Benchmark image for RoboCerebra integration tests.
+# RoboCerebra reuses LIBERO's simulator (libero_10 suite) with a different
+# rename_map, so this image is identical to the LIBERO benchmark image —
+# extends the nightly GPU base with LIBERO assets + the PR's source code.
+#
+# Build:  docker build -f docker/Dockerfile.benchmark.robocerebra -t lerobot-benchmark-robocerebra .
+# Run:    docker run --gpus all --rm lerobot-benchmark-robocerebra lerobot-eval ...
+
+FROM huggingface/lerobot-gpu:latest
+
+# Pre-download lerobot/libero-assets from HF Hub so nothing is fetched at
+# runtime (which times out on CI). Point the libero config at the cached path.
+# libero/libero/__init__.py calls input() when ~/.libero/config.yaml is missing,
+# so we write the config before any libero import can happen.
+RUN LIBERO_DIR=$(python -c \
+      "import importlib.util, os; s=importlib.util.find_spec('libero'); \
+       print(os.path.join(os.path.dirname(s.origin), 'libero'))") && \
+    mkdir -p /home/user_lerobot/.libero && \
+    python -c "\
+from huggingface_hub import snapshot_download; \
+snapshot_download(repo_id='lerobot/libero-assets', repo_type='dataset', \
+                  local_dir='/home/user_lerobot/.libero/assets')" && \
+    printf "assets: /home/user_lerobot/.libero/assets\nbddl_files: ${LIBERO_DIR}/bddl_files\ndatasets: ${LIBERO_DIR}/../datasets\ninit_states: ${LIBERO_DIR}/init_files\n" \
+    > /home/user_lerobot/.libero/config.yaml
+
+# Overlay the PR's source code on top of the nightly image.
+COPY --chown=user_lerobot:user_lerobot . .
+
+CMD ["/bin/bash"]

docker/Dockerfile.benchmark.robomme

@@ -0,0 +1,56 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Benchmark image for RoboMME integration tests.
+# Extends the nightly GPU image (which has lerobot[all]) with Vulkan system
+# libs for ManiSkill/SAPIEN and the robomme extra. robomme isn't in [all]
+# because mani-skill hard-pins gymnasium==0.29.1 and numpy<2.0.0 which
+# conflict with lerobot's defaults; both are safe at runtime:
+#   - gymnasium 0.29.x has the same 5-tuple step() API as 1.x (since 0.26)
+#   - numpy 1.26.4 is API-compatible with lerobot's actual usage.
+#
+# Build:  docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-benchmark-robomme .
+# Run:    docker run --gpus all --rm lerobot-benchmark-robomme lerobot-eval ...
+
+FROM huggingface/lerobot-gpu:latest
+
+# NVIDIA Container Toolkit: expose Vulkan driver capability for headless rendering.
+ENV NVIDIA_DRIVER_CAPABILITIES=all \
+    VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json
+
+# ManiSkill/SAPIEN's renderer needs Vulkan, which isn't in the base image.
+USER root
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+         libvulkan1 libvulkan-dev mesa-vulkan-drivers \
+    && mkdir -p /usr/share/vulkan/icd.d \
+    && echo '{"file_format_version":"1.0.0","ICD":{"library_path":"libGLX_nvidia.so.0","api_version":"1.3.0"}}' \
+       > /usr/share/vulkan/icd.d/nvidia_icd.json \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+USER user_lerobot
+
+# Install smolvla + av-dep via the PR's pyproject, then layer robomme on top
+# with gymnasium/numpy overrides. robomme isn't a pyproject extra because its
+# mani-skill pin conflicts with lerobot's base numpy>=2 (see pyproject.toml).
+COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
+RUN printf 'gymnasium==0.29.1\nnumpy==1.26.4\n' > /tmp/robomme_override.txt \
+    && uv pip install --no-cache --override /tmp/robomme_override.txt \
+         -e ".[smolvla,av-dep]" \
+         "robomme @ git+https://github.com/RoboMME/robomme_benchmark.git@main" \
+    && python -c "import robomme; print('robomme import OK')"
+
+# Overlay the PR's source code on top of the nightly image.
+COPY --chown=user_lerobot:user_lerobot . .
+
+CMD ["/bin/bash"]

docker/Dockerfile.benchmark.robotwin

@@ -0,0 +1,138 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Benchmark image for RoboTwin 2.0 integration tests.
+# Extends the nightly GPU image with the RoboTwin simulator stack:
+#   sapien/mplib/pytorch3d + NVlabs CuRobo + embodiments.zip + objects.zip
+# (~3.96 GB of assets; background_texture.zip ~11 GB skipped for smoke eval).
+#
+# Build: docker build -f docker/Dockerfile.benchmark.robotwin -t lerobot-benchmark-robotwin .
+# Run:   docker run --gpus all --rm lerobot-benchmark-robotwin \
+#            lerobot-eval --env.type=robotwin --env.task=beat_block_hammer ...
+
+FROM huggingface/lerobot-gpu:latest
+
+ENV NVIDIA_DRIVER_CAPABILITIES=all \
+    VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json \
+    ROBOTWIN_ROOT=/opt/robotwin
+
+# The nightly base is CUDA -base (no compiler, no Vulkan loader). CuRobo's
+# `pip install -e .` runs nvcc, and SAPIEN renders via Vulkan — add both.
+USER root
+# Pinned upstream SHA for reproducible benchmark runs. Bump when we need
+# an upstream fix; don't rely on `main` drift.
+ARG ROBOTWIN_SHA=0aeea2d669c0f8516f4d5785f0aa33ba812c14b4
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+         cuda-nvcc-12-4 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 +31,12 @@
     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: annotation_pipeline
+    title: Annotation Pipeline
   - local: streaming_video_encoding
     title: Streaming Video Encoding
   title: "Datasets"
@@ -77,10 +81,22 @@
     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/annotation_pipeline.mdx

@@ -0,0 +1,161 @@
+# Annotation Pipeline
+
+`lerobot-annotate` populates the two language columns introduced by the
+[Language Columns and Recipes](./language_and_recipes) page —
+`language_persistent` and `language_events` — directly into
+`data/chunk-*/file-*.parquet`. There is no flavor namespace and no sidecar
+file tree: multiple revisions of a dataset mean multiple dataset copies.
+
+## What the pipeline produces
+
+Three modules write into a per-episode staging tree, then a single writer
+rewrites the data shards in place:
+
+| Style / atom                                | Column                | Module   |
+| ------------------------------------------- | --------------------- | -------- |
+| `subtask` (Pi0.7-style "how, not what")     | `language_persistent` | Module 1 |
+| `plan` (initial + refresh on interjection)  | `language_persistent` | Module 1 |
+| `memory` (MEM-style compression)            | `language_persistent` | Module 1 |
+| `interjection`                              | `language_events`     | Module 2 |
+| speech tool-call atom (`style=null`, `say`) | `language_events`     | Module 2 |
+| `vqa` (user / assistant pair)               | `language_events`     | Module 3 |
+
+The writer drops the legacy `subtask_index` column. It does **not** add a
+`tools` column to the parquet — the tool catalog lives at
+`meta/info.json["tools"]` instead (see [Tools](./tools)). After every
+annotation run the pipeline ensures the canonical `say` schema is
+present in that list, preserving any tools the user pre-declared. Chat-
+template consumers read the catalog through
+`LeRobotDatasetMetadata.tools` and pass it to
+`apply_chat_template(messages, tools=meta.tools, ...)`.
+
+If you want to declare additional tools for a dataset before annotation
+runs, edit `meta/info.json["tools"]` directly — the pipeline preserves
+anything already there. Implementations of those tools live under
+`src/lerobot/tools/`; one file per tool, registered via
+`TOOL_REGISTRY`. See the [Tools](./tools) doc for the authoring guide.
+
+## How to run it locally or on SLURM
+
+Install the extra and invoke the console script:
+
+```bash
+uv sync --extra annotations
+uv run lerobot-annotate \
+  --repo_id=imstevenpmwork/super_poulain_draft \
+  --vlm.backend=vllm \
+  --vlm.model_id=Qwen/Qwen3.6-27B-FP8 \
+  --vlm.tensor_parallel_size=2
+```
+
+The pipeline attaches actual camera footage to every Module 1/2/3 prompt
+by default, decoded from the dataset's first `observation.images.*`
+stream. Override with `--vlm.camera_key=observation.images.<name>` to
+pin a specific viewpoint. Datasets with no video tracks fall back to
+text-only prompts automatically.
+
+**Module 1 sees the whole episode as one video block.** Subtask
+decomposition gets a `{"type":"video", "video":[<frames>]}` block
+covering the entire demonstration; Qwen-VL pools temporally on its own
+and decides where to cut. There is no keyframe stride or count knob —
+`--module_1.max_video_frames` (default 32) only caps the frames packed
+into the video block as a model-capacity bound. Module 2 attaches a
+single still frame at the interjection timestamp; Module 3 attaches the
+exact emission frame to each VQA pair.
+
+The executor picks `LocalPipelineExecutor` for small datasets and
+`SlurmPipelineExecutor` for large ones based on
+`--executor.auto_threshold` (default 32 episodes). Force local with
+`--executor.force_local=true`. SLURM jobs honour `--executor.slurm_partition`,
+`--executor.slurm_gpus`, and `--executor.slurm_time`.
+
+## Style-to-recipe consumer mapping
+
+The pipeline produces exactly the styles consumed by
+`src/lerobot/configs/recipes/pi05_hirobot.yaml`:
+
+- `low_level_execution`, `high_level_subtask`, `memory_update` consume
+  `subtask`/`plan`/`memory` from `language_persistent`.
+- `user_interjection_response` consumes `interjection` events plus the
+  paired speech atom (merged into one assistant target turn via
+  `tool_calls_from`) and the same-timestamp `plan` refresh.
+- `ask_vqa` consumes the `(vqa, user)` and `(vqa, assistant)` pairs from
+  `language_events`.
+
+## Why the design is scoped to the canonical recipe
+
+Two things drive the scope:
+
+1. **Persistent state vs exact-event split.** Persistent rows (`subtask`,
+   `plan`, `memory`) broadcast per episode and answer "what state is in
+   force at this frame?". Event rows (`interjection`, `vqa`, speech) only
+   appear on the exact frame whose timestamp matches the emission. The
+   pipeline writes timestamps taken straight from the source parquet — no
+   floating-point recomputation.
+2. **One Qwen-VL pass.** All three modules share a single VLM client
+   (vLLM if available, transformers fallback) so the cost is one model
+   load per dataset, not three.
+
+## Module independence and staged reruns
+
+Each module writes its raw output to
+`<root>/.annotate_staging/episode_{N:06d}/<module>.jsonl`. That makes
+prompt iteration cheap — re-running one module overwrites only its own
+JSONL file before the writer composes the final parquet. Modules can be
+disabled via `--module_1.enabled=false` (and similarly for 2 and 3) to
+test them in isolation.
+
+## Validation/report checks before final write
+
+Before the writer runs, `StagingValidator` checks:
+
+- exact frame-timestamp alignment for every event row;
+- no orphan speech / interjection pairs;
+- `plan` is refreshed at every interjection timestamp;
+- `memory` rows fall on subtask boundaries (warning, not error);
+- VQA assistant `content` parses as JSON in one of the
+  bbox / keypoint / count / attribute / spatial shapes;
+- every row routes to the column dictated by `column_for_style(style)`.
+
+Errors abort the writer (`--skip_validation=true` overrides for debugging).
+
+## Paper inspirations per module
+
+- **Module 1 — subtasks.** Hi Robot ([Shi 2025](https://arxiv.org/abs/2502.19417))
+  atom granularity ("pick up one piece of lettuce", "place bowl to box");
+  Pi0.7 ([Physical Intelligence 2025](https://pi.website/pi07)) "how, not
+  what" detail.
+- **Module 1 — memory.** MEM ([Torne 2026](https://arxiv.org/abs/2603.03596))
+  compression directive: keep only minimal relevant information; functional
+  outcomes preserved, specific attributes dropped.
+- **Module 2 — interjections.** Hi Robot scenario taxonomy: negative task,
+  situated correction, specific constraint, preference. Speech is a
+  tool-call-only atom (`tool_calls=[{type:function, function:{name:"say",
+arguments:{text:...}}}]`).
+- **Module 3 — VQA.** ECoT ([Zawalski 2024](https://arxiv.org/abs/2407.08693))
+  grounded features (bounding boxes in pixel `[x_min, y_min, x_max, y_max]`,
+  keypoints) and Steerable Policies' multi-abstraction grounding.
+
+Future maintainers should adjust the prompt templates in
+`src/lerobot/annotations/steerable_pipeline/prompts/` against these
+references rather than rewriting from scratch.
+
+## Compute and list-size estimates
+
+Per episode, the pipeline issues O(`max_steps`) Module 1 calls,
+O(`max_interjections_per_episode`) Module 2 calls, and
+O(`vqa_emission_hz × episode_seconds`) Module 3 calls. With defaults
+(8 subtasks, 1 interjection, 1 Hz × 3 pairs) and 30-second episodes, that
+is ~50 VLM calls per episode. `language_persistent` per episode is ~10s of
+KB at most (parquet dictionary-encodes one entry per episode);
+`language_events` is empty on most frames and is bounded by the number of
+emissions, not `num_frames × num_emissions`.
+
+## Reproducibility via seed and prompt hashes
+
+`--seed` (default 1729) feeds the per-episode RNGs that select interjection
+timestamps and VQA question types. Combined with the deterministic prompt
+templates checked into `prompts/`, two runs at the same seed against the
+same dataset and the same model checkpoint produce byte-identical staging
+artifacts. Prompt edits are recorded by file hash; future tooling can pin
+expected `(seed, prompt_hash)` pairs into the dataset card.

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/il_robots.mdx

@@ -32,6 +32,12 @@ Once you’ve gathered enough trajectories, you’ll train a neural network to i
 
 If you run into any issues at any point, jump into our [Discord community](https://discord.com/invite/s3KuuzsPFb) for support.
 
+<Tip>
+
+Want to quickly get the right commands for your setup? The [quickstart notebook](https://github.com/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb) [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb) lets you configure your robot once and generates all the commands below ready to paste.
+
+</Tip>
+
 ## Set up and Calibrate
 
 If you haven't yet set up and calibrated your robot and teleop device, please do so by following the robot-specific tutorial.

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/robocasa.mdx

@@ -0,0 +1,188 @@
+# RoboCasa365
+
+[RoboCasa365](https://robocasa.ai) is a large-scale simulation framework for training and benchmarking **generalist robots** in everyday kitchen tasks. It ships 365 diverse manipulation tasks across 2,500 kitchen environments, 3,200+ object assets and 600+ hours of human demonstration data, on a PandaOmron 12-DOF mobile manipulator (Franka arm on a holonomic base).
+
+- Paper: [RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots](https://arxiv.org/abs/2406.02523)
+- GitHub: [robocasa/robocasa](https://github.com/robocasa/robocasa)
+- Project website: [robocasa.ai](https://robocasa.ai)
+- Pretrained policy: [`lerobot/smolvla_robocasa`](https://huggingface.co/lerobot/smolvla_robocasa)
+- Single-task dataset (CloseFridge): [`pepijn223/robocasa_CloseFridge`](https://huggingface.co/datasets/pepijn223/robocasa_CloseFridge)
+
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/robocasa-banner.webp"
+  alt="RoboCasa365 benchmark overview"
+  width="85%"
+/>
+
+## Available tasks
+
+RoboCasa365 organizes its 365 tasks into two families and three upstream benchmark groups that LeRobot exposes as first-class `--env.task` shortcuts:
+
+| Family    | Tasks | Description                                                                     |
+| --------- | ----- | ------------------------------------------------------------------------------- |
+| Atomic    | ~65   | Single-skill tasks: pick-and-place, door/drawer manipulation, appliance control |
+| Composite | ~300  | Multi-step tasks across 60+ categories: cooking, cleaning, organizing, etc.     |
+
+**Atomic task examples:** `CloseFridge`, `OpenDrawer`, `OpenCabinet`, `TurnOnMicrowave`, `TurnOffStove`, `NavigateKitchen`, `PickPlaceCounterToStove`.
+
+**Composite task categories:** baking, boiling, brewing, chopping, clearing table, defrosting food, loading dishwasher, making tea, microwaving food, washing dishes, and more.
+
+`--env.task` accepts three forms:
+
+- a single task name (`CloseFridge`)
+- a comma-separated list (`CloseFridge,OpenBlenderLid,PickPlaceCoffee`)
+- a benchmark-group shortcut — `atomic_seen`, `composite_seen`, `composite_unseen`, `pretrain50`, `pretrain100`, `pretrain200`, `pretrain300` — which auto-expands to the upstream task list and auto-sets the dataset `split` (`target` or `pretrain`).
+
+## Installation
+
+RoboCasa and its dependency `robosuite` are not published on PyPI, and RoboCasa's own `setup.py` hardcodes `lerobot==0.3.3`, which conflicts with this repo's `lerobot`. LeRobot therefore does **not** expose a `robocasa` extra — install the two packages manually as editable clones (using `--no-deps` on `robocasa` to skip its shadowed `lerobot` pin):
+
+```bash
+# After following the standard LeRobot installation instructions.
+
+git clone https://github.com/robocasa/robocasa.git ~/robocasa
+git clone https://github.com/ARISE-Initiative/robosuite.git ~/robosuite
+pip install -e ~/robocasa --no-deps
+pip install -e ~/robosuite
+
+# Robocasa's runtime deps (the ones its setup.py would have pulled, minus
+# the bad lerobot pin).
+pip install numpy numba scipy mujoco pygame Pillow opencv-python \
+            pyyaml pynput tqdm termcolor imageio h5py lxml hidapi \
+            tianshou gymnasium
+
+python -m robocasa.scripts.setup_macros
+# Lightweight assets (lightwheel object meshes + textures). Enough for
+# the default env out of the box.
+python -m robocasa.scripts.download_kitchen_assets \
+  --type tex tex_generative fixtures_lw objs_lw
+# Optional: full objaverse/aigen registries (~30GB) for richer object
+# variety. Enable at eval time via --env.obj_registries (see below).
+# python -m robocasa.scripts.download_kitchen_assets --type objs_objaverse
+```
+
+<Tip>
+RoboCasa requires MuJoCo. Set the rendering backend before training or evaluation:
+
+```bash
+export MUJOCO_GL=egl  # for headless servers (HPC, cloud)
+```
+
+</Tip>
+
+### Object registries
+
+By default the env samples objects only from the `lightwheel` registry (what `--type objs_lw` ships), which avoids a `Probabilities contain NaN` crash when the objaverse / aigen packs aren't on disk. If you've downloaded the full asset set, enable the full registry at runtime:
+
+```bash
+--env.obj_registries='[objaverse,lightwheel]'
+```
+
+## Evaluation
+
+All eval snippets below mirror the CI command (see `.github/workflows/benchmark_tests.yml`). The `--rename_map` argument maps RoboCasa's native camera keys (`robot0_agentview_left` / `robot0_eye_in_hand` / `robot0_agentview_right`) onto the three-camera (`camera1` / `camera2` / `camera3`) input layout the released `smolvla_robocasa` policy was trained on.
+
+### Single-task evaluation (recommended for quick iteration)
+
+```bash
+lerobot-eval \
+  --policy.path=lerobot/smolvla_robocasa \
+  --env.type=robocasa \
+  --env.task=CloseFridge \
+  --eval.batch_size=1 \
+  --eval.n_episodes=20 \
+  --eval.use_async_envs=false \
+  --policy.device=cuda \
+  '--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
+```
+
+### Multi-task evaluation
+
+Pass a comma-separated list of tasks:
+
+```bash
+lerobot-eval \
+  --policy.path=lerobot/smolvla_robocasa \
+  --env.type=robocasa \
+  --env.task=CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove \
+  --eval.batch_size=1 \
+  --eval.n_episodes=20 \
+  --eval.use_async_envs=false \
+  --policy.device=cuda \
+  '--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
+```
+
+### Benchmark-group evaluation
+
+Run an entire upstream group (e.g. all 18 `atomic_seen` tasks with `split=target`):
+
+```bash
+lerobot-eval \
+  --policy.path=lerobot/smolvla_robocasa \
+  --env.type=robocasa \
+  --env.task=atomic_seen \
+  --eval.batch_size=1 \
+  --eval.n_episodes=20 \
+  --eval.use_async_envs=false \
+  --policy.device=cuda \
+  '--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
+```
+
+### Recommended evaluation episodes
+
+**20 episodes per task** for reproducible benchmarking. Matches the protocol used in published results.
+
+## Policy inputs and outputs
+
+**Observations** (raw RoboCasa camera names are preserved verbatim):
+
+- `observation.state` — 16-dim proprioceptive state (base position, base quaternion, relative end-effector position, relative end-effector quaternion, gripper qpos)
+- `observation.images.robot0_agentview_left` — left agent view, 256×256 HWC uint8
+- `observation.images.robot0_eye_in_hand` — wrist camera view, 256×256 HWC uint8
+- `observation.images.robot0_agentview_right` — right agent view, 256×256 HWC uint8
+
+**Actions:**
+
+- Continuous control in `Box(-1, 1, shape=(12,))` — base motion (4D) + control mode (1D) + end-effector position (3D) + end-effector rotation (3D) + gripper (1D).
+
+## Training
+
+### Single-task example
+
+A ready-to-use single-task dataset is on the Hub:
+[`pepijn223/robocasa_CloseFridge`](https://huggingface.co/datasets/pepijn223/robocasa_CloseFridge).
+
+Fine-tune a SmolVLA base on `CloseFridge`:
+
+```bash
+lerobot-train \
+  --policy.type=smolvla \
+  --policy.repo_id=${HF_USER}/smolvla_robocasa_CloseFridge \
+  --policy.load_vlm_weights=true \
+  --policy.push_to_hub=true \
+  --dataset.repo_id=pepijn223/robocasa_CloseFridge \
+  --env.type=robocasa \
+  --env.task=CloseFridge \
+  --output_dir=./outputs/smolvla_robocasa_CloseFridge \
+  --steps=100000 \
+  --batch_size=4 \
+  --eval_freq=5000 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=5 \
+  --save_freq=10000
+```
+
+Evaluate the resulting checkpoint:
+
+```bash
+lerobot-eval \
+  --policy.path=${HF_USER}/smolvla_robocasa_CloseFridge \
+  --env.type=robocasa \
+  --env.task=CloseFridge \
+  --eval.batch_size=1 \
+  --eval.n_episodes=20
+```
+
+## Reproducing published results
+
+The released checkpoint [`lerobot/smolvla_robocasa`](https://huggingface.co/lerobot/smolvla_robocasa) is evaluated with the commands in the [Evaluation](#evaluation) section. CI runs a 10-atomic-task smoke eval (one episode each) on every PR touching the benchmark, picking fixture-centric tasks that don't require the objaverse asset pack.

docs/source/robocerebra.mdx

@@ -0,0 +1,99 @@
+# RoboCerebra
+
+[RoboCerebra](https://robocerebra-project.github.io/) is a long-horizon manipulation benchmark that evaluates **high-level reasoning, planning, and memory** in VLAs. Episodes chain multiple sub-goals with language-grounded intermediate instructions, built on top of LIBERO's simulator stack (MuJoCo + robosuite, Franka Panda 7-DOF).
+
+- Paper: [RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation](https://arxiv.org/abs/2506.06677)
+- Project website: [robocerebra-project.github.io](https://robocerebra-project.github.io/)
+- Dataset: [`lerobot/robocerebra_unified`](https://huggingface.co/datasets/lerobot/robocerebra_unified) — LeRobot v3.0, 6,660 episodes / 571,116 frames at 20 fps, 1,728 language-grounded sub-tasks.
+- Pretrained policy: [`lerobot/smolvla_robocerebra`](https://huggingface.co/lerobot/smolvla_robocerebra)
+
+## Available tasks
+
+RoboCerebra reuses LIBERO's simulator, so evaluation runs against the LIBERO `libero_10` long-horizon suite:
+
+| Suite     | CLI name    | Tasks | Description                                                   |
+| --------- | ----------- | ----- | ------------------------------------------------------------- |
+| LIBERO-10 | `libero_10` | 10    | Long-horizon kitchen/living room tasks chaining 3–6 sub-goals |
+
+Each RoboCerebra episode in the dataset is segmented into multiple sub-tasks with natural-language instructions, which the unified dataset exposes as independent supervision signals.
+
+## Installation
+
+RoboCerebra piggybacks on LIBERO, so the `libero` extra is all you need:
+
+```bash
+pip install -e ".[libero]"
+```
+
+<Tip>
+RoboCerebra requires Linux (MuJoCo / robosuite). Set the rendering backend before training or evaluation:
+
+```bash
+export MUJOCO_GL=egl  # for headless servers (HPC, cloud)
+```
+
+</Tip>
+
+## Evaluation
+
+RoboCerebra eval runs against LIBERO's `libero_10` suite with RoboCerebra's camera naming (`image` + `wrist_image`) and an extra empty-camera slot so a three-view-trained policy receives the expected input layout:
+
+```bash
+lerobot-eval \
+  --policy.path=lerobot/smolvla_robocerebra \
+  --env.type=libero \
+  --env.task=libero_10 \
+  --env.fps=20 \
+  --env.obs_type=pixels_agent_pos \
+  --env.observation_height=256 \
+  --env.observation_width=256 \
+  '--env.camera_name_mapping={"agentview_image": "image", "robot0_eye_in_hand_image": "wrist_image"}' \
+  --eval.batch_size=1 \
+  --eval.n_episodes=10 \
+  --eval.use_async_envs=false \
+  --policy.device=cuda \
+  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.wrist_image": "observation.images.camera2"}' \
+  --policy.empty_cameras=1
+```
+
+### Recommended evaluation episodes
+
+**10 episodes per task** across the `libero_10` suite (100 total) for reproducible benchmarking. Matches the protocol used in the RoboCerebra paper.
+
+## Policy inputs and outputs
+
+**Observations:**
+
+- `observation.state` — 8-dim proprioceptive state (7 joint positions + gripper)
+- `observation.images.image` — third-person view, 256×256 HWC uint8
+- `observation.images.wrist_image` — wrist-mounted camera view, 256×256 HWC uint8
+
+**Actions:**
+
+- Continuous control in `Box(-1, 1, shape=(7,))` — end-effector delta (6D) + gripper (1D)
+
+## Training
+
+The unified dataset at [`lerobot/robocerebra_unified`](https://huggingface.co/datasets/lerobot/robocerebra_unified) exposes two RGB streams and language-grounded sub-task annotations:
+
+| Feature                          | Shape         | Description          |
+| -------------------------------- | ------------- | -------------------- |
+| `observation.images.image`       | (256, 256, 3) | Third-person view    |
+| `observation.images.wrist_image` | (256, 256, 3) | Wrist-mounted camera |
+| `observation.state`              | (8,)          | Joint pos + gripper  |
+| `action`                         | (7,)          | EEF delta + gripper  |
+
+Fine-tune a SmolVLA base on it:
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/smolvla_base \
+  --dataset.repo_id=lerobot/robocerebra_unified \
+  --env.type=libero \
+  --env.task=libero_10 \
+  --output_dir=outputs/smolvla_robocerebra
+```
+
+## Reproducing published results
+
+The released checkpoint [`lerobot/smolvla_robocerebra`](https://huggingface.co/lerobot/smolvla_robocerebra) was trained on `lerobot/robocerebra_unified` and evaluated with the command in the [Evaluation](#evaluation) section. CI runs the same command with `--eval.n_episodes=1` as a smoke test on every PR touching the benchmark.

docs/source/robomme.mdx

@@ -0,0 +1,130 @@
+# RoboMME
+
+[RoboMME](https://robomme.github.io) is a memory-augmented manipulation benchmark built on ManiSkill (SAPIEN). It evaluates a robot's ability to retain and use information across an episode — counting, object permanence, reference, and imitation.
+
+- **16 tasks** across 4 memory-skill suites
+- **1,600 training demos** (100 per task, 50 val, 50 test)
+- **Dataset**: [`lerobot/robomme`](https://huggingface.co/datasets/lerobot/robomme) — LeRobot v3.0, 768K frames at 10 fps
+- **Simulator**: ManiSkill / SAPIEN, Panda arm, Linux only
+
+![RoboMME benchmark tasks overview](https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04639/gradient.png)
+
+## Tasks
+
+| Suite                             | Tasks                                                         |
+| --------------------------------- | ------------------------------------------------------------- |
+| **Counting** (temporal memory)    | BinFill, PickXtimes, SwingXtimes, StopCube                    |
+| **Permanence** (spatial memory)   | VideoUnmask, VideoUnmaskSwap, ButtonUnmask, ButtonUnmaskSwap  |
+| **Reference** (object memory)     | PickHighlight, VideoRepick, VideoPlaceButton, VideoPlaceOrder |
+| **Imitation** (procedural memory) | MoveCube, InsertPeg, PatternLock, RouteStick                  |
+
+## Installation
+
+> RoboMME requires **Linux** (ManiSkill/SAPIEN uses Vulkan rendering). Docker is recommended to isolate dependency conflicts.
+
+### Native (Linux)
+
+```bash
+pip install --override <(printf 'gymnasium==0.29.1\nnumpy==1.26.4\n') \
+  -e '.[smolvla,av-dep]' \
+  'robomme @ git+https://github.com/RoboMME/robomme_benchmark.git@main'
+```
+
+> **Dependency note**: `mani-skill` (pulled by `robomme`) pins `gymnasium==0.29.1` and `numpy<2.0.0`, which conflict with lerobot's base `numpy>=2.0.0`. That's why `robomme` is not a pyproject extra — use the override install above, or the Docker approach below to avoid conflicts entirely.
+
+### Docker (recommended)
+
+```bash
+# Build base image first (from repo root)
+docker build -f docker/Dockerfile.eval-base -t lerobot-eval-base .
+
+# Build RoboMME eval image (applies gymnasium + numpy pin overrides)
+docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-robomme .
+```
+
+The `docker/Dockerfile.benchmark.robomme` image overrides `gymnasium==0.29.1` and `numpy==1.26.4` after lerobot's install. Both versions are runtime-safe for lerobot's actual API usage.
+
+## Running Evaluation
+
+### Default (single task, single episode)
+
+```bash
+lerobot-eval \
+    --policy.path=<your_policy_repo> \
+    --env.type=robomme \
+    --env.task=PickXtimes \
+    --env.dataset_split=test \
+    --env.task_ids=[0] \
+    --eval.batch_size=1 \
+    --eval.n_episodes=1
+```
+
+### Multi-task evaluation
+
+Evaluate multiple tasks in one run by comma-separating task names. Use `task_ids` to control which episodes are evaluated per task. Recommended: 50 episodes per task for the test split.
+
+```bash
+lerobot-eval \
+    --policy.path=<your_policy_repo> \
+    --env.type=robomme \
+    --env.task=PickXtimes,BinFill,StopCube,MoveCube,InsertPeg \
+    --env.dataset_split=test \
+    --env.task_ids=[0,1,2,3,4,5,6,7,8,9] \
+    --eval.batch_size=1 \
+    --eval.n_episodes=50
+```
+
+### Key CLI options for `env.type=robomme`
+
+| Option               | Default       | Description                                        |
+| -------------------- | ------------- | -------------------------------------------------- |
+| `env.task`           | `PickXtimes`  | Any of the 16 task names above (comma-separated)   |
+| `env.dataset_split`  | `test`        | `train`, `val`, or `test`                          |
+| `env.action_space`   | `joint_angle` | `joint_angle` (8-D) or `ee_pose` (7-D)             |
+| `env.episode_length` | `300`         | Max steps per episode                              |
+| `env.task_ids`       | `null`        | List of episode indices to evaluate (null = `[0]`) |
+
+## Dataset
+
+The dataset [`lerobot/robomme`](https://huggingface.co/datasets/lerobot/robomme) is in **LeRobot v3.0 format** and can be loaded directly:
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+dataset = LeRobotDataset("lerobot/robomme")
+```
+
+### Dataset features
+
+| Feature            | Shape         | Description                     |
+| ------------------ | ------------- | ------------------------------- |
+| `image`            | (256, 256, 3) | Front camera RGB                |
+| `wrist_image`      | (256, 256, 3) | Wrist camera RGB                |
+| `actions`          | (8,)          | Joint angles + gripper          |
+| `state`            | (8,)          | Joint positions + gripper state |
+| `simple_subgoal`   | str           | High-level language annotation  |
+| `grounded_subgoal` | str           | Grounded language annotation    |
+| `episode_index`    | int           | Episode ID                      |
+| `frame_index`      | int           | Frame within episode            |
+
+### Feature key alignment (training)
+
+The env wrapper exposes `pixels/image` and `pixels/wrist_image` as observation keys. The `features_map` in `RoboMMEEnv` maps these to `observation.images.image` and `observation.images.wrist_image` for the policy. State is exposed as `agent_pos` and maps to `observation.state`.
+
+The dataset's `image` and `wrist_image` columns already align with the policy input keys, so no renaming is needed when fine-tuning.
+
+## Action Spaces
+
+| Type          | Dim | Description                                               |
+| ------------- | --- | --------------------------------------------------------- |
+| `joint_angle` | 8   | 7 joint angles + 1 gripper (−1 closed, +1 open, absolute) |
+| `ee_pose`     | 7   | xyz + roll/pitch/yaw + gripper                            |
+
+Set via `--env.action_space=joint_angle` (default) or `--env.action_space=ee_pose`.
+
+## Platform Notes
+
+- **Linux only**: ManiSkill requires SAPIEN/Vulkan. macOS and Windows are not supported.
+- **GPU recommended**: Rendering is CPU-capable but slow; CUDA + Vulkan gives full speed.
+- **gymnasium / numpy conflict**: See installation note above. Docker image handles this automatically.
+- **ManiSkill fork**: `robomme` depends on a specific ManiSkill fork (`YinpeiDai/ManiSkill`), pulled in automatically via the `robomme` package.

docs/source/robotwin.mdx

@@ -0,0 +1,223 @@
+# RoboTwin 2.0
+
+RoboTwin 2.0 is a **large-scale dual-arm manipulation benchmark** built on the SAPIEN physics engine. It provides a standardized evaluation protocol for bimanual robotic policies across 50 tasks (as of upstream `main`) with strong domain randomization (clutter, lighting, background, tabletop height, and language instructions).
+
+- Paper: [RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation](https://arxiv.org/abs/2506.18088)
+- GitHub: [RoboTwin-Platform/RoboTwin](https://github.com/RoboTwin-Platform/RoboTwin)
+- Leaderboard: [robotwin-platform.github.io/leaderboard](https://robotwin-platform.github.io/leaderboard)
+- Dataset: [lerobot/robotwin_unified](https://huggingface.co/datasets/lerobot/robotwin_unified)
+
+![RoboTwin 2.0 benchmark overview](https://www.aitntnews.com/pictures/2025/7/8/9a7f79cb-5ba9-11f0-8581-fa163e47d677.png)
+
+## Overview
+
+| Property      | Value                                                    |
+| ------------- | -------------------------------------------------------- |
+| Tasks         | 50 dual-arm manipulation tasks                           |
+| Robot         | Aloha-AgileX bimanual (14 DOF, 7 per arm)                |
+| Action space  | 14-dim joint-space, continuous in `[-1, 1]`              |
+| Cameras       | `head_camera`, `left_camera`, `right_camera`             |
+| Simulator     | SAPIEN (not MuJoCo)                                      |
+| Eval protocol | 100 episodes/task, 50 demo_clean demonstrations          |
+| Eval settings | **Easy** (`demo_clean`) and **Hard** (`demo_randomized`) |
+
+## Available tasks
+
+RoboTwin 2.0 ships 50 dual-arm manipulation tasks in its upstream `envs/` directory. The canonical list is the `ROBOTWIN_TASKS` tuple in `src/lerobot/envs/robotwin.py`, mirrored verbatim from the upstream repo. Example tasks:
+
+| Task                     | CLI name                 | Category          |
+| ------------------------ | ------------------------ | ----------------- |
+| Beat block with hammer   | `beat_block_hammer`      | Tool use          |
+| Click bell / alarm clock | `click_bell`             | Precision press   |
+| Stack blocks (2 / 3)     | `stack_blocks_two/three` | Stacking          |
+| Stack bowls (2 / 3)      | `stack_bowls_two/three`  | Stacking          |
+| Handover block / mic     | `handover_block`         | Bimanual coord.   |
+| Lift pot                 | `lift_pot`               | Bimanual lift     |
+| Shake bottle             | `shake_bottle`           | Continuous motion |
+| Turn switch              | `turn_switch`            | Articulated obj   |
+| Stamp seal               | `stamp_seal`             | Precision place   |
+| Scan object              | `scan_object`            | Mobile manip.     |
+
+Pass a comma-separated list to `--env.task` to run multiple tasks in a single eval sweep.
+
+<Tip warning={true}>
+  `open_laptop` is currently broken upstream (its `check_success()` uses
+  `self.arm_tag`, which is only set inside the scripted-expert `play_once()`
+  path and therefore unavailable during normal policy eval). Avoid it until the
+  upstream bug is fixed, or patch the task to default `self.arm_tag = "left"` in
+  `load_actors()`.
+</Tip>
+
+## Dataset
+
+The RoboTwin 2.0 dataset is available in **LeRobot v3.0 format** on the Hugging Face Hub:
+
+```
+lerobot/robotwin_unified
+```
+
+It contains over 100,000 pre-collected trajectories across all 50 tasks (79.6 GB, Apache 2.0 license). No format conversion is needed — it is already in the correct LeRobot v3.0 schema with video observations and action labels.
+
+You can load it directly with the HF Datasets library:
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("lerobot/robotwin_unified", split="train")
+```
+
+## Installation
+
+RoboTwin 2.0 requires **Linux** with an NVIDIA GPU (CUDA 12.1 recommended). Installation takes approximately 20 minutes.
+
+### 1. Create a conda environment
+
+```bash
+conda create -n robotwin python=3.10 -y
+conda activate robotwin
+```
+
+### 2. Install LeRobot
+
+```bash
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e "."
+```
+
+### 3. Install RoboTwin 2.0
+
+```bash
+git clone https://github.com/RoboTwin-Platform/RoboTwin.git
+cd RoboTwin
+bash script/_install.sh
+bash script/_download_assets.sh
+```
+
+The install script handles all Python dependencies including SAPIEN, CuRobo, mplib, and pytorch3d.
+
+<Tip warning={true}>
+If the automated install fails, install manually:
+
+```bash
+pip install -r requirements.txt
+pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"
+cd envs && git clone https://github.com/NVlabs/curobo.git && cd curobo
+pip install -e . --no-build-isolation
+```
+
+Then apply the required mplib fix: in `mplib/planner.py` line 807, remove `or collide` from the conditional.
+
+</Tip>
+
+### 4. Add RoboTwin to PYTHONPATH
+
+The RoboTwin task modules must be importable by LeRobot. From within the `RoboTwin/` directory:
+
+```bash
+export PYTHONPATH="${PYTHONPATH}:$(pwd)"
+```
+
+Add this to your shell profile to make it permanent.
+
+## Evaluation
+
+### Standard evaluation (recommended)
+
+Evaluate a policy on a single task with the official protocol (100 episodes):
+
+```bash
+lerobot-eval \
+  --policy.path="your-hf-policy-id" \
+  --env.type=robotwin \
+  --env.task=beat_block_hammer \
+  --eval.batch_size=1 \
+  --eval.n_episodes=100
+```
+
+### Single-task quick check
+
+```bash
+lerobot-eval \
+  --policy.path="your-hf-policy-id" \
+  --env.type=robotwin \
+  --env.task=beat_block_hammer \
+  --eval.batch_size=1 \
+  --eval.n_episodes=5
+```
+
+### Multi-task sweep
+
+Evaluate on several tasks in one run:
+
+```bash
+lerobot-eval \
+  --policy.path="your-hf-policy-id" \
+  --env.type=robotwin \
+  --env.task=beat_block_hammer,click_bell,handover_block,stack_blocks_two \
+  --eval.batch_size=1 \
+  --eval.n_episodes=100
+```
+
+### Full benchmark (all 50 tasks)
+
+```bash
+lerobot-eval \
+  --policy.path="your-hf-policy-id" \
+  --env.type=robotwin \
+  --env.task=adjust_bottle,beat_block_hammer,blocks_ranking_rgb,blocks_ranking_size,click_alarmclock,click_bell,dump_bin_bigbin,grab_roller,handover_block,handover_mic,hanging_mug,lift_pot,move_can_pot,move_pillbottle_pad,move_playingcard_away,move_stapler_pad,open_microwave,pick_diverse_bottles,pick_dual_bottles,place_a2b_left,place_a2b_right,place_bread_basket,place_bread_skillet,place_burger_fries,place_can_basket,place_cans_plasticbox,place_container_plate,place_dual_shoes,place_empty_cup,place_fan,place_mouse_pad,place_object_basket,place_object_scale,place_object_stand,place_phone_stand,place_shoe,press_stapler,put_bottles_dustbin,put_object_cabinet,rotate_qrcode,scan_object,shake_bottle,shake_bottle_horizontally,stack_blocks_three,stack_blocks_two,stack_bowls_three,stack_bowls_two,stamp_seal,turn_switch \
+  --eval.batch_size=1 \
+  --eval.n_episodes=100
+```
+
+<Tip>
+  `open_laptop` is intentionally omitted above because of the upstream
+  `self.arm_tag` bug (see the **Available tasks** section). Re-add it once the
+  upstream fix lands.
+</Tip>
+
+## Camera configuration
+
+By default, all three cameras are included:
+
+| Camera key     | Description                    |
+| -------------- | ------------------------------ |
+| `head_camera`  | Torso-mounted overhead view    |
+| `left_camera`  | Left arm wrist-mounted camera  |
+| `right_camera` | Right arm wrist-mounted camera |
+
+To use a subset of cameras, override `--env.camera_names`:
+
+```bash
+lerobot-eval \
+  --policy.path="your-hf-policy-id" \
+  --env.type=robotwin \
+  --env.task=beat_block_hammer \
+  --env.camera_names="head_camera,left_camera" \
+  --eval.batch_size=1 \
+  --eval.n_episodes=10
+```
+
+## Environment config reference
+
+Key parameters for `RoboTwinEnvConfig`:
+
+| Parameter            | Default                                  | Description                        |
+| -------------------- | ---------------------------------------- | ---------------------------------- |
+| `task`               | `"beat_block_hammer"`                    | Comma-separated task name(s)       |
+| `fps`                | `25`                                     | Simulation FPS                     |
+| `episode_length`     | `300`                                    | Max steps per episode              |
+| `obs_type`           | `"pixels_agent_pos"`                     | `"pixels"` or `"pixels_agent_pos"` |
+| `camera_names`       | `"head_camera,left_camera,right_camera"` | Comma-separated active cameras     |
+| `observation_height` | `240`                                    | Camera pixel height                |
+| `observation_width`  | `320`                                    | Camera pixel width                 |
+
+## Leaderboard submission
+
+Results can be submitted to the [RoboTwin 2.0 leaderboard](https://robotwin-platform.github.io/leaderboard). The official protocol requires:
+
+- Training on 50 `demo_clean` demonstrations per task
+- Evaluating 100 episodes per task
+- Reporting success rate separately for **Easy** (`demo_clean`) and **Hard** (`demo_randomized`) settings
+
+For submission instructions, refer to the [RoboTwin 2.0 documentation](https://robotwin-platform.github.io/doc/).

docs/source/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.

examples/annotation/run_hf_job.py

@@ -0,0 +1,69 @@
+#!/usr/bin/env python
+"""Launch ``lerobot-annotate`` on a Hugging Face job (vllm + Qwen3.6 MoE).
+
+Spawns one ``h200x2`` job that:
+
+  1. installs this branch of ``lerobot`` plus the annotation extras,
+  2. boots two vllm servers (one per GPU) with Qwen3.6-35B-A3B-FP8,
+  3. runs Module 1/2/3 across the dataset (per-camera VQA via PR 3471),
+  4. uploads the annotated dataset to ``--push_to_hub``.
+
+Usage:
+
+    HF_TOKEN=hf_... uv run python examples/annotation/run_hf_job.py
+
+Adjust ``CMD`` below to point at your own dataset / target hub repo.
+"""
+
+import os
+
+from huggingface_hub import get_token, run_job
+
+token = os.environ.get("HF_TOKEN") or get_token()
+if not token:
+    raise RuntimeError("No HF token. Run `huggingface-cli login` or `export HF_TOKEN=hf_...`")
+
+CMD = (
+    "apt-get update -qq && apt-get install -y -qq git ffmpeg && "
+    "pip install --no-deps "
+    "'lerobot @ git+https://github.com/huggingface/lerobot.git@feat/language-annotation-pipeline' && "
+    "pip install --upgrade-strategy only-if-needed "
+    "datasets pyarrow av jsonlines draccus gymnasium torchcodec mergedeep pyyaml-include toml typing-inspect && "
+    "export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 && "
+    "export VLLM_VIDEO_BACKEND=pyav && "
+    "lerobot-annotate "
+    "--repo_id=imstevenpmwork/super_poulain_draft "
+    "--vlm.backend=openai "
+    "--vlm.model_id=Qwen/Qwen3.6-35B-A3B-FP8 "
+    "--vlm.parallel_servers=2 "
+    "--vlm.num_gpus=2 "
+    '--vlm.serve_command="vllm serve Qwen/Qwen3.6-35B-A3B-FP8 '
+    "--tensor-parallel-size 1 --max-model-len 32768 "
+    '--gpu-memory-utilization 0.8 --uvicorn-log-level warning --port {port}" '
+    "--vlm.serve_ready_timeout_s=1800 "
+    "--vlm.client_concurrency=128 "
+    "--vlm.max_new_tokens=512 "
+    "--vlm.temperature=0.7 "
+    "--executor.episode_parallelism=16 "
+    "--vlm.chat_template_kwargs='{\"enable_thinking\": false}' "
+    "--vlm.camera_key=observation.images.wrist "
+    "--module_1.frames_per_second=1.0 "
+    "--module_1.use_video_url=true "
+    "--module_1.use_video_url_fps=1.0 "
+    "--module_1.derive_task_from_video=always "
+    "--module_1.n_task_rephrasings=30 "
+    "--module_2.max_interjections_per_episode=6 "
+    "--module_3.K=3 "
+    "--module_3.vqa_emission_hz=1.0 "
+    "--push_to_hub=pepijn223/super_poulain_full_tool3"
+)
+
+job = run_job(
+    image="vllm/vllm-openai:latest",
+    command=["bash", "-c", CMD],
+    flavor="h200x2",
+    secrets={"HF_TOKEN": token},
+    timeout="2h",
+)
+print(f"Job URL: {job.url}")
+print(f"Job ID:  {job.id}")

examples/notebooks/quickstart.ipynb

@@ -0,0 +1,342 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# 🤗 LeRobot Quickstart\n",
+    "\n",
+    "Calibration → teleoperation → data collection → training → evaluation.\n",
+    "\n",
+    "Install the required dependencies: `pip install -e .[notebook,dataset,training,viz,hardware]`.\n",
+    "\n",
+    "**How to use:**\n",
+    "1. Edit the **Configuration** cell with your settings.\n",
+    "2. Run all cells (`Run All`).\n",
+    "3. Each section prints a ready-to-paste terminal command - copy it and run it.\n",
+    "\n",
+    "Each setup is different, please refer to the [LeRobot documentation](https://huggingface.co/docs/lerobot/il_robots) for more details on each step and available options. <br>\n",
+    "Feel free to make this notebook your own and adapt it to your needs!"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## Utils"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def _cameras_arg(cameras: dict) -> str:\n",
+    "    if not cameras:\n",
+    "        return \"\"\n",
+    "    entries = [f\"{n}: {{{', '.join(f'{k}: {v}' for k, v in cfg.items())}}}\" for n, cfg in cameras.items()]\n",
+    "    return \"{ \" + \", \".join(entries) + \" }\"\n",
+    "\n",
+    "\n",
+    "def print_cmd(*parts: str) -> None:\n",
+    "    \"\"\"Print a shell command with line continuations, skipping empty parts.\"\"\"\n",
+    "    non_empty = [p for p in parts if p]\n",
+    "    print(\" \\\\\\n    \".join(non_empty))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## Configuration\n",
+    "\n",
+    "Edit this cell, then **Run All** to generate all commands below."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Robot (follower) - run `lerobot-find-port` to discover the port\n",
+    "ROBOT_TYPE = \"so101_follower\"\n",
+    "ROBOT_PORT = \"/dev/ttyACM0\"\n",
+    "ROBOT_ID = \"my_follower_arm\"\n",
+    "\n",
+    "# Teleop (leader) - run `lerobot-find-port` to discover the port\n",
+    "TELEOP_TYPE = \"so101_leader\"\n",
+    "TELEOP_PORT = \"/dev/ttyACM1\"\n",
+    "TELEOP_ID = \"my_leader_arm\"\n",
+    "\n",
+    "# Cameras - set to {} to disable\n",
+    "# Run `lerobot-find-cameras opencv` to list available cameras and their indices\n",
+    "CAMERAS = {\n",
+    "    \"top\": {\"type\": \"opencv\", \"index_or_path\": 2, \"width\": 640, \"height\": 480, \"fps\": 30},\n",
+    "    \"wrist\": {\"type\": \"opencv\", \"index_or_path\": 4, \"width\": 640, \"height\": 480, \"fps\": 30},\n",
+    "}\n",
+    "\n",
+    "# Dataset\n",
+    "HF_USER = \"your_hf_username\"  # `huggingface-cli whoami` to find your username\n",
+    "DATASET_NAME = \"my_so101_dataset\"\n",
+    "TASK_DESCRIPTION = \"pick and place the block\"\n",
+    "NUM_EPISODES = 10\n",
+    "\n",
+    "# Training\n",
+    "POLICY_TYPE = \"act\"  # act, diffusion, smolvla, ...\n",
+    "POLICY_DEVICE = \"cuda\"  # cuda / cpu / mps\n",
+    "TRAIN_STEPS = 10_000\n",
+    "SAVE_FREQ = 2_000\n",
+    "OUTPUT_DIR = f\"outputs/train/{DATASET_NAME}\"\n",
+    "\n",
+    "# Inference - Hub repo ID or local checkpoint path\n",
+    "# e.g. set to f\"{OUTPUT_DIR}/checkpoints/last\" to use a local checkpoint\n",
+    "POLICY_PATH = f\"{HF_USER}/{DATASET_NAME}_{POLICY_TYPE}\"\n",
+    "LAST_CHECKPOINT_PATH = f\"{OUTPUT_DIR}/checkpoints/last\"\n",
+    "\n",
+    "# Derived\n",
+    "DATASET_REPO_ID = f\"{HF_USER}/{DATASET_NAME}\"\n",
+    "DATASET_ROOT = f\"data/{DATASET_NAME}\"\n",
+    "POLICY_REPO_ID = f\"{HF_USER}/{DATASET_NAME}_{POLICY_TYPE}\"\n",
+    "EVAL_REPO_ID = f\"{HF_USER}/eval_{DATASET_NAME}\"\n",
+    "CAMERAS_ARG = _cameras_arg(CAMERAS)\n",
+    "CAMERAS_FLAG = f'--robot.cameras=\"{CAMERAS_ARG}\"' if CAMERAS_ARG else \"\"\n",
+    "\n",
+    "print(f\"Robot  : {ROBOT_TYPE} @ {ROBOT_PORT}\")\n",
+    "print(f\"Teleop : {TELEOP_TYPE} @ {TELEOP_PORT}\")\n",
+    "print(f\"Cameras: {list(CAMERAS) or 'none'}\")\n",
+    "print(f\"Dataset: {DATASET_REPO_ID} ({NUM_EPISODES} episodes) saved to {DATASET_ROOT}\")\n",
+    "print(f\"Policy : {POLICY_TYPE} -> {POLICY_REPO_ID}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## 1. Calibration\n",
+    "\n",
+    "Run once per arm before first use."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Follower\n",
+    "print_cmd(\n",
+    "    \"lerobot-calibrate\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    f\"--robot.id={ROBOT_ID}\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Leader\n",
+    "print_cmd(\n",
+    "    \"lerobot-calibrate\",\n",
+    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
+    "    f\"--teleop.port={TELEOP_PORT}\",\n",
+    "    f\"--teleop.id={TELEOP_ID}\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## 2. Teleoperation\n",
+    "\n",
+    "See the [teleoperation docs](https://huggingface.co/docs/lerobot/il_robots#teleoperate) and the [cameras guide](https://huggingface.co/docs/lerobot/cameras) for more options."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print_cmd(\n",
+    "    \"lerobot-teleoperate\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    f\"--robot.id={ROBOT_ID}\",\n",
+    "    CAMERAS_FLAG,\n",
+    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
+    "    f\"--teleop.port={TELEOP_PORT}\",\n",
+    "    f\"--teleop.id={TELEOP_ID}\",\n",
+    "    \"--display_data=true\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## 3. Record Dataset\n",
+    "\n",
+    "See the [recording docs](https://huggingface.co/docs/lerobot/il_robots#record-a-dataset) for tips on gathering good data."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print_cmd(\n",
+    "    \"lerobot-record\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    f\"--robot.id={ROBOT_ID}\",\n",
+    "    CAMERAS_FLAG,\n",
+    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
+    "    f\"--teleop.port={TELEOP_PORT}\",\n",
+    "    f\"--teleop.id={TELEOP_ID}\",\n",
+    "    f\"--dataset.repo_id={DATASET_REPO_ID}\",\n",
+    "    f\"--dataset.num_episodes={NUM_EPISODES}\",\n",
+    "    f'--dataset.single_task=\"{TASK_DESCRIPTION}\"',\n",
+    "    \"--dataset.streaming_encoding=true\",\n",
+    "    \"--display_data=true\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Resume a previously interrupted recording session\n",
+    "print_cmd(\n",
+    "    \"lerobot-record\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    f\"--robot.id={ROBOT_ID}\",\n",
+    "    CAMERAS_FLAG,\n",
+    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
+    "    f\"--teleop.port={TELEOP_PORT}\",\n",
+    "    f\"--teleop.id={TELEOP_ID}\",\n",
+    "    f\"--dataset.repo_id={DATASET_REPO_ID}\",\n",
+    "    f\"--dataset.root={DATASET_ROOT}\",\n",
+    "    f\"--dataset.num_episodes={NUM_EPISODES}\",\n",
+    "    f'--dataset.single_task=\"{TASK_DESCRIPTION}\"',\n",
+    "    \"--dataset.streaming_encoding=true\",\n",
+    "    \"--display_data=true\",\n",
+    "    \"--resume=true\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## 4. Train Policy\n",
+    "\n",
+    "See the [training docs](https://huggingface.co/docs/lerobot/il_robots#train-a-policy) for configuration options and tips."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print_cmd(\n",
+    "    \"lerobot-train\",\n",
+    "    f\"--dataset.repo_id={DATASET_REPO_ID}\",\n",
+    "    f\"--policy.type={POLICY_TYPE}\",\n",
+    "    f\"--policy.device={POLICY_DEVICE}\",\n",
+    "    f\"--policy.repo_id={POLICY_REPO_ID}\",\n",
+    "    f\"--output_dir={OUTPUT_DIR}\",\n",
+    "    f\"--steps={TRAIN_STEPS}\",\n",
+    "    f\"--save_freq={SAVE_FREQ}\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Resume a previously interrupted training session\n",
+    "print_cmd(\n",
+    "    \"lerobot-train\",\n",
+    "    f\"--config_path={LAST_CHECKPOINT_PATH}/pretrained_model/train_config.json\",\n",
+    "    \"--resume=true\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "---\n",
+    "## 5. Inference\n",
+    "\n",
+    "Uses `POLICY_PATH` from the Configuration cell (defaults to the Hub repo ID). You can also put there the `LAST_CHECKPOINT_PATH`.\n",
+    "\n",
+    "See the [inference docs](https://huggingface.co/docs/lerobot/il_robots#run-inference-and-evaluate-your-policy) for details."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print_cmd(\n",
+    "    \"lerobot-record\",\n",
+    "    f\"--policy.path={POLICY_PATH}\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    f\"--robot.id={ROBOT_ID}\",\n",
+    "    CAMERAS_FLAG,\n",
+    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
+    "    f\"--teleop.port={TELEOP_PORT}\",\n",
+    "    f\"--teleop.id={TELEOP_ID}\",\n",
+    "    f\"--dataset.repo_id={EVAL_REPO_ID}\",\n",
+    "    f\"--dataset.num_episodes={NUM_EPISODES}\",\n",
+    "    f'--dataset.single_task=\"{TASK_DESCRIPTION}\"',\n",
+    "    \"--dataset.streaming_encoding=true\",\n",
+    ")"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "lerobot (3.12.3)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}

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]",
@@ -108,9 +108,9 @@ training = [
     "wandb>=0.24.0,<0.25.0",
 ]
 hardware = [
-    "pynput>=1.7.8,<1.9.0",
-    "pyserial>=3.5,<4.0",
-    "deepdiff>=7.0.1,<9.0.0",
+    "lerobot[pynput-dep]",
+    "lerobot[pyserial-dep]",
+    "lerobot[deepdiff-dep]",
 ]
 viz = [
     "rerun-sdk>=0.24.0,<0.27.0",
@@ -129,6 +129,7 @@ av-dep = ["av>=15.0.0,<16.0.0"]
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.9.17"]
 transformers-dep = ["transformers==5.3.0"] # TODO(Steven): https://github.com/huggingface/lerobot/pull/3249
+sentencepiece-dep = ["sentencepiece>=0.2.0,<0.3.0"] # FAST action tokenizer backend (pi052, pi0_fast)
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
 can-dep = ["python-can>=4.2.0,<5.0.0"]
 peft-dep = ["peft>=0.18.0,<1.0.0"]
@@ -136,10 +137,14 @@ scipy-dep = ["scipy>=1.14.0,<2.0.0"]
 diffusers-dep = ["diffusers>=0.27.2,<0.36.0"]
 qwen-vl-utils-dep = ["qwen-vl-utils>=0.0.11,<0.1.0"]
 matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0", "contourpy>=1.3.0,<2.0.0"] # NOTE: Explicitly listing contourpy helps the resolver converge faster.
+pyserial-dep = ["pyserial>=3.5,<4.0"]
+deepdiff-dep = ["deepdiff>=7.0.1,<9.0.0"]
+pynput-dep = ["pynput>=1.7.8,<1.9.0"]
+pyzmq-dep = ["pyzmq>=26.2.1,<28.0.0"]
 
 # Motors
-feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
-dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
+feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
+dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
 damiao = ["lerobot[can-dep]"]
 robstride = ["lerobot[can-dep]"]
 
@@ -147,10 +152,11 @@ robstride = ["lerobot[can-dep]"]
 openarms = ["lerobot[damiao]"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
-lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
+lekiwi = ["lerobot[feetech]", "lerobot[pyzmq-dep]"]
 unitree_g1 = [
     # "unitree-sdk2==1.0.1",
-    "pyzmq>=26.2.1,<28.0.0",
+    "lerobot[pyzmq-dep]",
+    "lerobot[pyserial-dep]",
     "onnxruntime>=1.16.0,<2.0.0",
     "onnx>=1.16.0,<2.0.0",
     "meshcat>=0.3.0,<0.4.0",
@@ -174,7 +180,7 @@ wallx = [
     "torchdiffeq>=0.2.4,<0.3.0",
     "lerobot[qwen-vl-utils-dep]",
 ]
-pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
+pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]", "lerobot[sentencepiece-dep]"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0"]
 multi_task_dit = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]"]
 groot = [
@@ -195,8 +201,26 @@ hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpci
 async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
 peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
 
+# Annotation pipeline (lerobot-annotate). datatrove is mandatory; vllm is
+# the preferred backend on Linux, with a transformers fallback elsewhere.
+annotations = [
+    "lerobot[dataset]",
+    "lerobot[transformers-dep]",
+    "datatrove>=0.4.0,<2.0.0",
+    "vllm>=0.6.0,<1.0.0; sys_platform == 'linux'",
+]
+
+# Tool implementations under src/lerobot/tools/. Each tool's dependencies
+# are isolated so adding a new tool doesn't bloat the base install.
+# Currently only `say` (Kyutai pocket-tts; CPU-only, ~100M params).
+tools = [
+    "pocket-tts>=1.0.0,<3.0.0",
+    "scipy>=1.11.0,<2.0.0",  # SayTool.output_dir uses scipy.io.wavfile
+]
+
 # Development
-dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1", "ruff>=0.14.1"]
+dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1", "ruff>=0.14.1", "lerobot[notebook]"]
+notebook = ["jupyter>=1.0.0,<2.0.0", "ipykernel>=6.0.0,<7.0.0"]
 test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
@@ -206,6 +230,20 @@ aloha = ["lerobot[dataset]", "gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
 pusht = ["lerobot[dataset]", "gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
 libero = ["lerobot[dataset]", "lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
 metaworld = ["lerobot[dataset]", "metaworld==3.0.0", "lerobot[scipy-dep]"]
+# NOTE: vlabench is NOT exposed as a `lerobot` extra. Its only distribution
+# is the OpenMOSS/VLABench GitHub repo (package name `VLABench`, no PyPI
+# release), so any `vlabench>=X` pip spec is unresolvable. Install it
+# manually alongside MuJoCo / dm-control — see docs/source/vlabench.mdx
+# for the recipe.
+# NOTE: robomme is NOT a pyproject extra — mani-skill hard-pins numpy<2
+# which conflicts with lerobot's numpy>=2 base pin, so the two trees can't
+# resolve into a single env. Install it only in the RoboMME Docker image
+# via `uv pip install --override` (see docker/Dockerfile.benchmark.robomme).
+# NOTE: robocasa is NOT exposed as a `lerobot` extra. Its setup.py pins
+# `lerobot==0.3.3` in install_requires, which cyclically shadows our own
+# workspace `lerobot` and makes the graph unsolvable under any resolver
+# (uv, pip). Install it manually alongside robosuite — see
+# docs/source/robocasa.mdx for the recipe.
 
 # All
 all = [
@@ -269,10 +307,12 @@ lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
 lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
+lerobot-annotate="lerobot.scripts.lerobot_annotate:main"
+lerobot-smolvla2-runtime="lerobot.scripts.lerobot_smolvla2_runtime:main"
 
 # ---------------- Tool Configurations ----------------
 [tool.setuptools.package-data]
-lerobot = ["envs/*.json"]
+lerobot = ["envs/*.json", "annotations/steerable_pipeline/prompts/*.txt"]
 
 [tool.setuptools.packages.find]
 where = ["src"]

scripts/ci/extract_task_descriptions.py

@@ -31,9 +31,23 @@ from __future__ import annotations
 
 import argparse
 import json
+import re
 import sys
 from pathlib import Path
 
+# LIBERO-plus derives task.language by space-joining the perturbation-variant
+# filename (grab_language_from_filename in libero/libero/benchmark/__init__.py),
+# so non-_language_ variants inherit a trailing metadata blob like
+# "view 0 0 100 0 0 initstate 0 noise 45" or "add 16". Strip those tokens so
+# the description matches the base instruction used in the training dataset.
+_LIBERO_PERTURBATION_TAIL_RE = re.compile(
+    r"(?:\s(?:view|initstate|noise|add|tb|table|light|level)(?:\s\d+)+)+$"
+)
+
+
+def _strip_libero_perturbation_tail(instruction: str) -> str:
+    return _LIBERO_PERTURBATION_TAIL_RE.sub("", instruction).strip()
+
 
 def _libero_descriptions(task_suite: str) -> dict[str, str]:
     from libero.libero import benchmark  # type: ignore[import-untyped]
@@ -47,7 +61,10 @@ def _libero_descriptions(task_suite: str) -> dict[str, str]:
         )
         return {}
     suite = suite_dict[task_suite]()
-    return {f"{task_suite}_{i}": suite.get_task(i).language for i in range(suite.n_tasks)}
+    return {
+        f"{task_suite}_{i}": _strip_libero_perturbation_tail(suite.get_task(i).language)
+        for i in range(suite.n_tasks)
+    }
 
 
 def _metaworld_descriptions(task_name: str) -> dict[str, str]:
@@ -57,19 +74,120 @@ def _metaworld_descriptions(task_name: str) -> dict[str, str]:
     return {f"{task_name}_0": label}
 
 
+def _robotwin_descriptions(task_names: str) -> dict[str, str]:
+    """Return descriptions for each requested RoboTwin task. Reads
+    `description/task_instruction/<task>.json` from the RoboTwin clone
+    (cwd is /opt/robotwin in CI). Falls back to the task name if missing."""
+    out: dict[str, str] = {}
+    root = Path("description/task_instruction")
+    for name in (t.strip() for t in task_names.split(",") if t.strip()):
+        desc_file = root / f"{name}.json"
+        desc = name.replace("_", " ")
+        if desc_file.is_file():
+            data = json.loads(desc_file.read_text())
+            full = data.get("full_description") or desc
+            # Strip the schema placeholders ({A}, {a}) — keep the sentence readable.
+            desc = full.replace("<", "").replace(">", "")
+        out[f"{name}_0"] = desc
+    return out
+
+
+def _robocasa_descriptions(task_spec: str) -> dict[str, str]:
+    """For each task in the comma-separated list, emit a cleaned-name label.
+
+    RoboCasa episodes carry their language instruction in the env's
+    `ep_meta['lang']`, populated per reset. Pulling it requires spinning
+    up the full kitchen env per task (~seconds each); we use the task
+    name as the key here and let the eval's episode info carry the
+    actual instruction.
+    """
+    out: dict[str, str] = {}
+    for task in (t.strip() for t in task_spec.split(",") if t.strip()):
+        # Split CamelCase into words: "CloseFridge" → "close fridge".
+        label = "".join(f" {c.lower()}" if c.isupper() else c for c in task).strip()
+        out[f"{task}_0"] = label or task
+    return out
+
+
+_ROBOMME_DESCRIPTIONS = {
+    "BinFill": "Fill the target bin with the correct number of cubes",
+    "PickXtimes": "Pick the indicated cube the specified number of times",
+    "SwingXtimes": "Swing the object the specified number of times",
+    "StopCube": "Grasp and stop the moving cube",
+    "VideoUnmask": "Pick the cube shown in the reference video",
+    "VideoUnmaskSwap": "Pick the cube matching the reference video after a swap",
+    "ButtonUnmask": "Press the button indicated by the reference",
+    "ButtonUnmaskSwap": "Press the correct button after objects are swapped",
+    "PickHighlight": "Pick the highlighted cube",
+    "VideoRepick": "Repick the cube shown in the reference video",
+    "VideoPlaceButton": "Place the cube on the button shown in the video",
+    "VideoPlaceOrder": "Place cubes in the order shown in the video",
+    "MoveCube": "Move the cube to the target location",
+    "InsertPeg": "Insert the peg into the target hole",
+    "PatternLock": "Unlock the pattern by pressing buttons in sequence",
+    "RouteStick": "Route the stick through the required waypoints",
+}
+
+
+def _robomme_descriptions(task_names: str, task_ids: list[int] | None = None) -> dict[str, str]:
+    """Return descriptions for each requested RoboMME task. Keys match the
+    video filename pattern `<task>_<task_id>` used by the eval script."""
+    if task_ids is None:
+        task_ids = [0]
+    out: dict[str, str] = {}
+    for name in (t.strip() for t in task_names.split(",") if t.strip()):
+        desc = _ROBOMME_DESCRIPTIONS.get(name, name)
+        for tid in task_ids:
+            out[f"{name}_{tid}"] = desc
+    return out
+
+
+def _vlabench_descriptions(task_spec: str) -> dict[str, str]:
+    """For each task in the comma-separated list, emit a cleaned-name label.
+
+    VLABench tasks carry language instructions on their dm_control task
+    object, but pulling them requires loading the full env per task
+    (~seconds each). The CI smoke-eval already captures the instruction
+    inside its episode info; this mapping is just enough to key
+    `metrics.json` by `<task>_0`.
+    """
+    out: dict[str, str] = {}
+    for task in (t.strip() for t in task_spec.split(",") if t.strip()):
+        out[f"{task}_0"] = task.replace("_", " ").strip()
+    return out
+
+
 def main() -> int:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--env", required=True, help="Environment family (libero, metaworld, ...)")
     parser.add_argument("--task", required=True, help="Task/suite name (e.g. libero_spatial)")
+    parser.add_argument(
+        "--task-ids",
+        type=str,
+        default=None,
+        help="Comma-separated task IDs (e.g. '0,1,2'). Default: [0]",
+    )
     parser.add_argument("--output", required=True, help="Path to write task_descriptions.json")
     args = parser.parse_args()
 
+    task_ids: list[int] | None = None
+    if args.task_ids:
+        task_ids = [int(x.strip()) for x in args.task_ids.split(",")]
+
     descriptions: dict[str, str] = {}
     try:
-        if args.env == "libero":
+        if args.env == ("libero", "libero_plus"):
             descriptions = _libero_descriptions(args.task)
         elif args.env == "metaworld":
             descriptions = _metaworld_descriptions(args.task)
+        elif args.env == "robotwin":
+            descriptions = _robotwin_descriptions(args.task)
+        elif args.env == "robocasa":
+            descriptions = _robocasa_descriptions(args.task)
+        elif args.env == "robomme":
+            descriptions = _robomme_descriptions(args.task, task_ids=task_ids)
+        elif args.env == "vlabench":
+            descriptions = _vlabench_descriptions(args.task)
         else:
             print(
                 f"[extract_task_descriptions] No description extractor for env '{args.env}'.",

src/lerobot/annotations/__init__.py

@@ -0,0 +1,15 @@
+#!/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.

src/lerobot/annotations/steerable_pipeline/__init__.py

@@ -0,0 +1,36 @@
+#!/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.
+"""Steerable annotation pipeline producing ``language_persistent`` and
+``language_events`` columns for LeRobot datasets.
+
+The pipeline is decomposed into three independently runnable modules whose
+outputs are staged per-episode before a final parquet rewrite:
+
+- :mod:`.modules.plan_subtasks_memory` (Module 1) — persistent styles
+- :mod:`.modules.interjections_and_speech` (Module 2) — event styles + speech
+- :mod:`.modules.general_vqa` (Module 3) — event-style VQA pairs
+"""
+
+from .config import AnnotationPipelineConfig
+from .validator import StagingValidator, ValidationReport
+from .writer import LanguageColumnsWriter
+
+__all__ = [
+    "AnnotationPipelineConfig",
+    "LanguageColumnsWriter",
+    "StagingValidator",
+    "ValidationReport",
+]

src/lerobot/annotations/steerable_pipeline/config.py

@@ -0,0 +1,260 @@
+#!/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, field
+from pathlib import Path
+from typing import Any
+
+
+@dataclass
+class Module1Config:
+    """Module 1 hyperparameters: plan + subtasks + memory + task augmentation.
+
+    Subtask decomposition sees the **whole episode** as one Qwen-VL video
+    block — no keyframe stride or count: the model handles temporal pooling
+    itself and decides where to cut. ``max_video_frames`` only caps the
+    number of frames packed into the video block (a model-capacity bound,
+    not an annotation-logic knob).
+    """
+
+    enabled: bool = True
+    n_task_rephrasings: int = 10
+    """Number of task rephrasings to generate at ``t=0`` as ``task_aug``
+    persistent rows (PR 1 ``CORE_STYLES``). The renderer's ``${task}``
+    binding rotates among them deterministically per ``sample_idx``,
+    realizing Xiao 2022 / CAST-style task-prompt diversity without
+    touching ``meta/tasks.parquet``. Set to 0 to disable."""
+    derive_task_from_video: str = "if_short"
+    """When to bypass the user-provided ``record.episode_task`` and
+    derive a fresh task description from the episode video alone:
+
+    - ``off``       never; always use the canonical task as the basis.
+    - ``if_short``  derive when the canonical task is empty, has fewer
+                    than ``derive_task_min_words`` words, or matches a
+                    placeholder string (``debug``, ``unnamed``, ``tbd``,
+                    ...). Default — fixes noisy / placeholder tasks
+                    without forcing derivation everywhere.
+    - ``always``    ignore the canonical task entirely; always derive
+                    from the video. Useful when the dataset's task
+                    labels are uniformly bad.
+
+    The video-derived task replaces the canonical task as the basis for
+    subtask decomposition, plan, memory, AND the ``task_aug`` rephrasings,
+    so every downstream annotation is grounded in what's actually visible.
+    ``meta/tasks.parquet`` is NOT modified — the Module-1-derived task
+    only lives in ``language_persistent`` rows."""
+    derive_task_min_words: int = 3
+    """Word-count threshold for ``derive_task_from_video=if_short``."""
+    frames_per_second: float = 1.0
+    """Sample one image-frame per ``1/fps`` seconds across the episode for
+    Module 1's subtask-decomposition prompt. ``1.0`` = 1 fps. Capped by
+    ``max_video_frames`` to avoid blowing up the request payload."""
+    max_video_frames: int = 128
+    """Hard cap on the number of frames Module 1 sends. With ``fps=1`` and
+    a 30 s episode this yields 30 frames. Bumped from 32 since each frame
+    is small (~30-100 KB PNG when base64'd)."""
+    min_subtask_seconds: float = 1.5
+    plan_max_steps: int = 8
+    use_video_url: bool = False
+    """When True (and backend supports it, e.g. ``openai``), Module 1
+    sends a ``video_url`` content block pointing at the episode's mp4
+    file instead of pre-decoded frames. Lets the server sample frames at
+    its own ``fps`` — no in-process conv3d cost. The video file is
+    extracted as a per-episode subclip to ``staging/.video_clips/`` so
+    the model sees only this episode's frames."""
+    use_video_url_fps: float = 1.0
+    """Frame-rate hint to send to the server (mm_processor_kwargs.fps).
+    Only used when ``use_video_url=True``. ``1.0`` = sample 1 frame per
+    second, which is plenty for subtask-boundary detection on most
+    manipulation episodes."""
+
+
+@dataclass
+class Module2Config:
+    """Module 2 hyperparameters: interjections + paired speech."""
+
+    enabled: bool = True
+    max_interjections_per_episode: int = 3
+    """Number of mid-episode interjections to generate per episode. Each
+    creates a paired ``(interjection, speech)`` event row plus triggers a
+    ``plan`` refresh at the same timestamp via Module 1. Bumped from the
+    original ``1`` after qwen36moe-10 showed plan/interjection coverage
+    was too sparse for Hi Robot-style training."""
+    interjection_min_t: float = 2.0
+    interjection_window_seconds: float = 2.0
+    """How many seconds of video to attach to the interjection prompt as
+    visual context. Without this the VLM only sees a single frozen frame
+    and writes generic interjections that aren't grounded in the actual
+    motion happening at the chosen timestamp."""
+    interjection_window_frames: int = 4
+    """How many frames to sample over ``interjection_window_seconds``.
+    Default 4 ⇒ ~0.5 fps over the leading 2 seconds — enough for the
+    model to read the ongoing motion, cheap enough to keep prompt size
+    bounded for the 32k context."""
+
+
+@dataclass
+class Module3Config:
+    """Module 3 hyperparameters: general VQA."""
+
+    enabled: bool = True
+    vqa_emission_hz: float = 1.0
+    K: int = 3
+    question_types: tuple[str, ...] = ("bbox", "keypoint", "count", "attribute", "spatial")
+
+
+@dataclass
+class VlmConfig:
+    """Shared Qwen-VL client configuration."""
+
+    backend: str = "openai"
+    """One of ``vllm``, ``transformers``, ``openai``, or ``stub`` (tests only).
+
+    Default ``openai`` talks to a local OpenAI-compatible server (vllm /
+    transformers) which the CLI auto-spawns when ``auto_serve=True``."""
+    model_id: str = "Qwen/Qwen2.5-VL-7B-Instruct"
+    api_base: str = "http://localhost:8000/v1"
+    """Base URL for the ``openai`` backend."""
+    api_key: str = "EMPTY"
+    """API key for the ``openai`` backend; ``EMPTY`` works for local servers."""
+    auto_serve: bool = True
+    """When True with ``backend=openai``, the CLI probes ``api_base``
+    first; if no server answers, it spawns one (default:
+    ``transformers serve``), waits for it to be ready, runs the
+    pipeline, and tears it down on exit. Default ``True`` so a single
+    ``lerobot-annotate`` call can drive the whole flow. Set to ``False``
+    if you want to fail fast when no server is reachable (e.g. you're
+    pointing at a remote endpoint that should already be up)."""
+    serve_port: int = 8000
+    """Port the auto-spawned server binds to. Sets ``api_base`` automatically."""
+    serve_command: str | None = None
+    """Override the auto-serve command (full shell command). When ``None``,
+    we run ``transformers serve <model_id> --port <serve_port> --continuous-batching``.
+
+    When ``parallel_servers > 1``, the literal ``{port}`` placeholder in
+    this command (if present) is substituted per-replica."""
+    parallel_servers: int = 1
+    """When >1, spawn this many independent inference servers (each pinned
+    to a GPU via ``CUDA_VISIBLE_DEVICES`` and listening on
+    ``serve_port + i``) and round-robin client requests across them.
+    Useful when DP/TP NCCL setup is broken on the node — single-GPU
+    replicas don't need cross-GPU communication. When
+    ``parallel_servers > num_gpus``, replicas are round-robin-assigned
+    to GPUs (e.g. 4 replicas on 2 GPUs → 0,1,0,1)."""
+    num_gpus: int = 0
+    """How many physical GPUs are available for round-robin replica
+    placement. ``0`` means ``parallel_servers`` (one GPU per replica,
+    backward-compatible default). Set this to ``2`` with
+    ``parallel_servers=4`` to pack 2 replicas per GPU."""
+    client_concurrency: int = 16
+    """Maximum number of in-flight chat requests the client issues in
+    parallel. vllm batches them internally for free, so bumping this
+    typically gives big throughput wins on a single TP=1 server. Set to
+    ``1`` for strict serial calls."""
+    serve_ready_timeout_s: float = 600.0
+    """Max seconds to wait for the server to start serving requests."""
+    max_new_tokens: int = 512
+    temperature: float = 0.2
+    json_mode: bool = True
+    batch_size: int = 4
+    tensor_parallel_size: int = 1
+    gpu_memory_utilization: float = 0.9
+    """Fraction of GPU memory vllm allocates for weights + KV cache.
+    Lower (e.g. 0.7) when the vision encoder needs cuDNN workspace, or to
+    avoid CUDNN_STATUS_NOT_INITIALIZED on tight VRAM (30B BF16 on 80 GB)."""
+    max_model_len: int | None = None
+    """Cap context length. ``None`` keeps the model's default; on H100 80 GB
+    a 30B BF16 model often needs ``max_model_len=8192`` or smaller to leave
+    room for KV cache."""
+    trust_remote_code: bool = False
+    """Pass ``trust_remote_code`` to HF auto-classes. Default ``False`` —
+    only enable for models that actually ship custom code in their repo
+    (rare for first-class VL releases). On Qwen3-VL it triggers an
+    std::bad_alloc post-load even though the official transformers class
+    is sufficient, so leaving this off is safest."""
+    camera_key: str | None = None
+    """Override the camera stream used for keyframe attachment. ``None`` picks
+    the first ``observation.images.*`` key the dataset declares."""
+    chat_template_kwargs: dict[str, Any] | None = None
+    """Forwarded as ``extra_body.chat_template_kwargs`` on every chat call.
+    Use this to pass model-specific template flags such as
+    ``{"enable_thinking": false}`` for Qwen3.5/Qwen3.6 to suppress the
+    reasoning preamble that otherwise eats the entire ``max_new_tokens``
+    budget before any JSON is emitted."""
+
+
+@dataclass
+class ExecutorConfig:
+    """Executor selection and SLURM hyperparameters."""
+
+    auto_threshold: int = 32
+    force_local: bool = False
+    slurm_partition: str | None = None
+    slurm_gpus: int = 1
+    slurm_time: str = "06:00:00"
+    workers: int = 1
+    episode_parallelism: int = 16
+    """Number of episodes processed concurrently within each module phase.
+    Each in-flight episode sends 3–5 dependent VLM calls; bumping this is
+    how you actually saturate ``parallel_servers`` and ``client_concurrency``
+    — without it, the executor loops one episode at a time and the
+    inference servers sit ~90% idle. Set to ``1`` for strict serial
+    execution."""
+
+
+@dataclass
+class AnnotationPipelineConfig:
+    """Top-level config for ``lerobot-annotate``.
+
+    Mirrors the structure of :class:`lerobot.configs.train.TrainPipelineConfig`:
+    a draccus-parsed dataclass that contains nested per-module sub-configs and
+    leaves the dataset, executor, and VLM choices independently knobbable.
+
+    Output is always in-place: the writer rewrites ``data/chunk-*/file-*.parquet``
+    in place. Multiple revisions of the same dataset live in separate copies.
+    """
+
+    repo_id: str | None = None
+    root: Path | None = None
+
+    staging_dir: Path | None = None
+    """If unset, defaults to ``<root>/.annotate_staging/``."""
+
+    seed: int = 1729
+
+    module_1: Module1Config = field(default_factory=Module1Config)
+    module_2: Module2Config = field(default_factory=Module2Config)
+    module_3: Module3Config = field(default_factory=Module3Config)
+
+    vlm: VlmConfig = field(default_factory=VlmConfig)
+    executor: ExecutorConfig = field(default_factory=ExecutorConfig)
+
+    skip_validation: bool = False
+    only_episodes: tuple[int, ...] | None = None
+
+    push_to_hub: str | None = None
+    """If set, after the pipeline completes, upload the annotated dataset
+    root to the Hugging Face Hub as a dataset repo with this id (e.g.
+    ``pepijn/super_poulain_steerable``). Creates the repo if missing."""
+    push_private: bool = False
+    """When ``push_to_hub`` is set, create the repo as private."""
+    push_commit_message: str | None = None
+    """Override the commit message used for the hub upload."""
+
+    def resolved_staging_dir(self, root: Path) -> Path:
+        return self.staging_dir if self.staging_dir is not None else root / ".annotate_staging"

src/lerobot/annotations/steerable_pipeline/executor.py

@@ -0,0 +1,263 @@
+#!/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.
+"""Executor selection: local vs SLURM via datatrove.
+
+The executor plans **four phases** with the dependency order from the plan:
+
+    phase 1: Module 1 (plan + subtasks + memory)
+    phase 2: Module 2 (interjections + speech)
+    phase 3: Module 1 plan-update pass — re-runs plan emission at every
+             interjection timestamp produced by phase 2
+    phase 4: Module 3 (VQA)
+    phase 5: validator
+    phase 6: writer
+
+Phase 3 is why ``executor.py`` documents the dependency: Module 1 must be
+re-entered after Module 2 to refresh ``plan`` rows at interjection times.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+from .config import AnnotationPipelineConfig, ExecutorConfig
+from .reader import EpisodeRecord, iter_episodes
+from .staging import EpisodeStaging
+from .validator import StagingValidator
+from .writer import LanguageColumnsWriter
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class PhaseResult:
+    """Summary of one pipeline phase across all episodes."""
+
+    name: str
+    episodes_processed: int
+    episodes_skipped: int
+
+
+@dataclass
+class PipelineRunSummary:
+    """Aggregated result returned by :meth:`Executor.run`."""
+
+    phases: list[PhaseResult]
+    written_paths: list[Path]
+    validation_report: Any  # ValidationReport, kept Any to avoid import cycle
+
+
+def select_executor_class(num_episodes: int, config: ExecutorConfig) -> str:
+    """Return ``"local"`` or ``"slurm"`` based on the threshold.
+
+    The plan's "executor selection threshold" lives in
+    :class:`ExecutorConfig.auto_threshold`. ``force_local`` always wins.
+    """
+    if config.force_local:
+        return "local"
+    return "local" if num_episodes <= config.auto_threshold else "slurm"
+
+
+@dataclass
+class Executor:
+    """Run all four phases over a dataset root.
+
+    The executor is intentionally framework-agnostic: by default it runs the
+    phases inline (suitable for tests, small datasets, and the CLI's
+    ``--force-local`` mode). It will optionally hand off to datatrove's
+    :class:`LocalPipelineExecutor` or :class:`SlurmPipelineExecutor` when those
+    are installed and the dataset is large enough to benefit from them.
+
+    Tests construct the executor directly with stub modules.
+    """
+
+    config: AnnotationPipelineConfig
+    module_1: Any  # PlanSubtasksMemoryModule
+    module_2: Any  # InterjectionsAndSpeechModule
+    module_3: Any  # GeneralVqaModule
+    writer: LanguageColumnsWriter
+    validator: StagingValidator
+
+    def run(self, root: Path) -> PipelineRunSummary:
+        records = list(iter_episodes(root, only_episodes=self.config.only_episodes))
+        n = len(records)
+        if n == 0:
+            raise ValueError(f"No episodes found under {root}/data/")
+
+        executor_kind = select_executor_class(n, self.config.executor)
+        print(f"[annotate] {n} episodes total; executor={executor_kind}", flush=True)
+
+        staging_dir = self.config.resolved_staging_dir(root)
+        staging_dir.mkdir(parents=True, exist_ok=True)
+
+        phases: list[PhaseResult] = []
+
+        # Phase 1: Module 1 (plan + subtasks + memory)
+        phases.append(self._run_module_phase("module_1", records, staging_dir, self.module_1))
+        # Phase 2: Module 2 (interjections + speech). Module 2 reads
+        # Module 1's subtask rows from the same staging tree to ground
+        # the interjection prompt in the correct local subtask.
+        phases.append(self._run_module_phase("module_2", records, staging_dir, self.module_2))
+        # Phase 3: Module 1 plan-update pass at interjection timestamps.
+        phases.append(self._run_plan_update_phase(records, staging_dir))
+        # Phase 4: Module 3 (VQA)
+        phases.append(self._run_module_phase("module_3", records, staging_dir, self.module_3))
+
+        print("[annotate] running validator...", flush=True)
+        report = self.validator.validate(records, staging_dir)
+        if not report.ok and not self.config.skip_validation:
+            raise RuntimeError(f"Staging validation failed: {report.summary()}")
+        print(f"[annotate] validator: {report.summary()}", flush=True)
+
+        print(f"[annotate] writing parquet shards into {root}/data/...", flush=True)
+        written = self.writer.write_all(records, staging_dir, root)
+        print(f"[annotate] wrote {len(written)} shard(s); pipeline complete", flush=True)
+
+        # Persist the tool catalog to meta/info.json so chat-template
+        # consumers (PR 3 SmolVLA2 / Pi0.5 / dataset visualizer) can read
+        # it via ``LeRobotDatasetMetadata.tools`` (PR 1). Idempotent and
+        # additive: anything the user pre-populated is preserved; we only
+        # ensure the canonical ``say`` schema is present.
+        self._ensure_tools_in_info(root)
+
+        return PipelineRunSummary(phases=phases, written_paths=written, validation_report=report)
+
+    def _ensure_tools_in_info(self, root: Path) -> None:
+        """Write ``meta/info.json["tools"]`` if missing the canonical ``say``.
+
+        Reads any user-declared tools already in ``info.json`` and merges
+        the canonical ``SAY_TOOL_SCHEMA`` into the list (deduped by
+        ``function.name``). Writes back to disk only if the list
+        changed.
+        """
+        import json  # noqa: PLC0415
+
+        from lerobot.datasets.language import SAY_TOOL_SCHEMA  # noqa: PLC0415
+
+        info_path = root / "meta" / "info.json"
+        if not info_path.exists():
+            return
+        try:
+            info = json.loads(info_path.read_text())
+        except Exception as exc:  # noqa: BLE001
+            print(f"[annotate] could not read {info_path}: {exc}", flush=True)
+            return
+
+        existing = info.get("tools")
+        if not isinstance(existing, list):
+            existing = []
+        names = {
+            (t.get("function") or {}).get("name")
+            for t in existing
+            if isinstance(t, dict)
+        }
+        merged = list(existing)
+        if SAY_TOOL_SCHEMA["function"]["name"] not in names:
+            merged.append(SAY_TOOL_SCHEMA)
+        if merged != existing:
+            info["tools"] = merged
+            info_path.write_text(json.dumps(info, indent=2))
+            print(
+                f"[annotate] meta/info.json: tools={[t['function']['name'] for t in merged]}",
+                flush=True,
+            )
+
+    def _run_module_phase(
+        self,
+        name: str,
+        records: list[EpisodeRecord],
+        staging_dir: Path,
+        module: Any,
+    ) -> PhaseResult:
+        import time as _time  # noqa: PLC0415
+        from concurrent.futures import ThreadPoolExecutor, as_completed  # noqa: PLC0415
+
+        if not module.enabled:
+            print(f"[annotate] phase={name} skipped (module disabled)", flush=True)
+            return PhaseResult(name=name, episodes_processed=0, episodes_skipped=len(records))
+        n = len(records)
+        parallelism = max(1, min(self.config.executor.episode_parallelism, n))
+        print(
+            f"[annotate] phase={name} starting on {n} episode(s) "
+            f"(parallelism={parallelism})",
+            flush=True,
+        )
+        t0 = _time.time()
+
+        def _do(idx_record: tuple[int, EpisodeRecord]) -> tuple[int, int, float]:
+            i, record = idx_record
+            ep_start = _time.time()
+            staging = EpisodeStaging(staging_dir, record.episode_index)
+            module.run_episode(record, staging)
+            return i, record.episode_index, _time.time() - ep_start
+
+        processed = 0
+        if parallelism == 1:
+            for i, record in enumerate(records, 1):
+                _, ep_idx, elapsed = _do((i, record))
+                processed += 1
+                print(
+                    f"[annotate]   {name} episode {i}/{n} "
+                    f"(idx={ep_idx}) done in {elapsed:.1f}s",
+                    flush=True,
+                )
+        else:
+            with ThreadPoolExecutor(max_workers=parallelism) as pool:
+                futures = [pool.submit(_do, (i, r)) for i, r in enumerate(records, 1)]
+                for fut in as_completed(futures):
+                    i, ep_idx, elapsed = fut.result()
+                    processed += 1
+                    print(
+                        f"[annotate]   {name} episode {processed}/{n} "
+                        f"(idx={ep_idx}, submit_order={i}) done in {elapsed:.1f}s",
+                        flush=True,
+                    )
+        total = _time.time() - t0
+        print(f"[annotate] phase={name} complete: {processed}/{n} in {total:.1f}s", flush=True)
+        return PhaseResult(name=name, episodes_processed=processed, episodes_skipped=0)
+
+    def _run_plan_update_phase(  # noqa: PLR0915
+        self, records: list[EpisodeRecord], staging_dir: Path
+    ) -> PhaseResult:
+        """Re-emit ``plan`` rows at each interjection timestamp from Module 2.
+
+        Module 1 owns the prompt; Module 2 produced the timestamps. This phase
+        therefore calls back into Module 1 with the interjection timestamps so
+        Module 1's existing prompt path is reused.
+        """
+        if not self.module_1.enabled or not self.module_2.enabled:
+            return PhaseResult(
+                name="module_1_plan_update", episodes_processed=0, episodes_skipped=len(records)
+            )
+        processed = 0
+        for record in records:
+            staging = EpisodeStaging(staging_dir, record.episode_index)
+            interjection_rows = [
+                row
+                for row in staging.read("module_2")
+                if row.get("style") == "interjection"
+            ]
+            interjection_times = [float(row["timestamp"]) for row in interjection_rows]
+            interjection_texts = [str(row.get("content") or "") for row in interjection_rows]
+            if interjection_times:
+                self.module_1.run_plan_updates(
+                    record, staging, interjection_times, interjection_texts
+                )
+                processed += 1
+        return PhaseResult(name="module_1_plan_update", episodes_processed=processed, episodes_skipped=0)

src/lerobot/annotations/steerable_pipeline/frames.py

@@ -0,0 +1,400 @@
+#!/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.
+"""Keyframe extraction for the annotation pipeline.
+
+Modules attach decoded camera frames to their VLM prompts so the model can
+ground subtask decomposition, interjection scenarios, and VQA in actual
+visual content. The pipeline shares one provider across modules and one
+episode at a time, with a small per-episode cache so multiple modules
+querying the same timestamp pay decode cost once.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Protocol
+
+from .reader import EpisodeRecord
+
+
+class FrameProvider(Protocol):
+    """Decodes camera frames at episode-relative timestamps."""
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """All ``observation.images.*`` feature keys this provider can decode."""
+
+    def frames_at(
+        self,
+        record: EpisodeRecord,
+        timestamps: list[float],
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        """Return one PIL.Image per timestamp from ``camera_key`` (or default).
+
+        Empty list if the camera is unavailable. ``camera_key=None`` falls back
+        to the provider's default camera so existing single-camera callers
+        (Module 1, Module 2) keep working unchanged.
+        """
+
+    def video_for_episode(
+        self,
+        record: EpisodeRecord,
+        max_frames: int,
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        """Return up to ``max_frames`` PIL images covering the whole episode.
+
+        Sampling is uniform across the episode duration. The returned list is
+        intended to be passed as one ``{"type":"video", "video":<list>}``
+        block to a Qwen-VL-compatible model that pools temporally itself.
+        Empty list if no camera available.
+        """
+
+
+@dataclass
+class _NullProvider:
+    """No-op provider used when the dataset has no video keys or in tests."""
+
+    @property
+    def camera_keys(self) -> list[str]:
+        return []
+
+    def frames_at(
+        self,
+        record: EpisodeRecord,
+        timestamps: list[float],
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        return []
+
+    def video_for_episode(
+        self,
+        record: EpisodeRecord,
+        max_frames: int,
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        return []
+
+
+def null_provider() -> FrameProvider:
+    return _NullProvider()
+
+
+@dataclass
+class VideoFrameProvider:
+    """Decodes frames from the dataset's ``observation.images.*`` streams.
+
+    By default the *first* camera key is used for Module 1 (subtask
+    decomposition) and Module 2 (interjection scenarios) — those prompts care
+    about *what is happening*, not which angle. Module 3 (VQA) instead
+    iterates over every camera in :attr:`camera_keys` so each frame's
+    grounded answer (bbox/keypoint/...) is tagged with the camera it was
+    grounded against.
+
+    ``camera_key`` overrides the default-camera choice but does not restrict
+    :attr:`camera_keys`. Pass ``camera_key`` explicitly to ``frames_at`` /
+    ``video_for_episode`` to read a non-default stream.
+
+    Caches up to ``cache_size`` decoded frames per process to keep
+    co-timestamped Module 2 + Module 1 plan-update calls cheap.
+    """
+
+    root: Path
+    camera_key: str | None = None
+    tolerance_s: float = 1e-2
+    cache_size: int = 256
+    _meta: Any = field(default=None, init=False, repr=False)
+    _cache: dict = field(default_factory=dict, init=False, repr=False)
+    _camera_keys: list[str] = field(default_factory=list, init=False, repr=False)
+
+    def __post_init__(self) -> None:
+        from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata  # noqa: PLC0415
+
+        self._meta = LeRobotDatasetMetadata(repo_id="local", root=self.root)
+        # ``camera_keys`` covers both image- and video-stored cameras
+        # (``video_keys`` is video-only). Some datasets declare cameras with
+        # ``dtype=image``, which would otherwise look empty here and silently
+        # disable Module 3 even though the videos are there.
+        keys = list(getattr(self._meta, "camera_keys", None) or self._meta.video_keys or [])
+        # Last-resort fallback: if metadata didn't surface anything but the
+        # caller explicitly named a camera (``--vlm.camera_key=...``), trust
+        # them — the key is by definition known to exist on the dataset.
+        if not keys and self.camera_key:
+            keys = [self.camera_key]
+        self._camera_keys = keys
+        if self.camera_key is None:
+            self.camera_key = keys[0] if keys else None
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """All ``observation.images.*`` keys available on this dataset."""
+        return list(self._camera_keys)
+
+    def frames_at(
+        self,
+        record: EpisodeRecord,
+        timestamps: list[float],
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        target = camera_key if camera_key is not None else self.camera_key
+        if not timestamps or target is None:
+            return []
+
+        out: list[Any] = []
+        misses: list[float] = []
+        miss_indices: list[int] = []
+        for i, ts in enumerate(timestamps):
+            key = (record.episode_index, target, round(float(ts), 6))
+            cached = self._cache.get(key)
+            if cached is not None:
+                out.append(cached)
+            else:
+                out.append(None)
+                misses.append(float(ts))
+                miss_indices.append(i)
+
+        if misses:
+            decoded = self._decode(record.episode_index, misses, target)
+            # decoder may return fewer frames than requested when some
+            # timestamps fall outside the video; pair what we have and
+            # leave the rest as None to be filtered below.
+            for i, img in zip(miss_indices, decoded):
+                out[i] = img
+                key = (record.episode_index, target, round(float(timestamps[i]), 6))
+                if len(self._cache) >= self.cache_size:
+                    self._cache.pop(next(iter(self._cache)))
+                self._cache[key] = img
+        # filter out any None left over from decode failures
+        return [img for img in out if img is not None]
+
+    def _decode(
+        self, episode_index: int, timestamps: list[float], camera_key: str
+    ) -> list[Any]:
+        ep = self._meta.episodes[episode_index]
+        from_timestamp = ep[f"videos/{camera_key}/from_timestamp"]
+        shifted = [from_timestamp + ts for ts in timestamps]
+        video_path = self.root / self._meta.get_video_file_path(episode_index, camera_key)
+
+        try:
+            return _decode_pyav_direct(video_path, shifted, self.tolerance_s)
+        except Exception as exc:
+            # Log loudly the first time decoding fails so silent
+            # Module-3-no-op (every prompt skipped because frames_at returned
+            # []) is debuggable from the job log instead of post-hoc parquet
+            # inspection. Subsequent failures stay quiet.
+            if not getattr(self, "_warned_decode_fail", False):
+                import logging  # noqa: PLC0415
+
+                logging.getLogger(__name__).warning(
+                    "VideoFrameProvider._decode failed for episode=%s camera=%s "
+                    "video_path=%s: %s",
+                    episode_index,
+                    camera_key,
+                    video_path,
+                    exc,
+                    exc_info=True,
+                )
+                self._warned_decode_fail = True
+            return []
+
+
+def _decode_pyav_direct(
+    video_path: Any, timestamps: list[float], tolerance_s: float
+) -> list[Any]:
+    """Decode the requested timestamps from ``video_path`` using PyAV directly.
+
+    Bypasses ``lerobot.datasets.video_utils.decode_video_frames`` entirely
+    because its "pyav" path actually goes through
+    ``decode_video_frames_torchvision`` → ``torchvision.io.VideoReader``,
+    which was removed in torchvision >= 0.22 (the vllm/vllm-openai:latest
+    container ships with torchvision 0.25). The annotation pipeline only
+    needs a handful of PIL images per (episode, ts), so we can decode them
+    with PyAV without any torch dependency at all.
+
+    Returns one ``PIL.Image`` per requested timestamp, in the same order.
+    Any timestamp the decoder couldn't reach is silently dropped (mirrors
+    the previous behaviour); callers filter ``None``/missing entries.
+    """
+    import av  # noqa: PLC0415
+    from PIL import Image  # noqa: PLC0415
+
+    if not timestamps:
+        return []
+
+    targets = sorted(set(timestamps))
+    seek_to = max(0.0, min(targets) - max(0.5, tolerance_s))
+
+    container = av.open(str(video_path))
+    try:
+        stream = container.streams.video[0]
+        # PyAV needs the seek target in stream timebase ticks.
+        if stream.time_base is None:
+            seek_pts = 0
+        else:
+            seek_pts = int(seek_to / float(stream.time_base))
+        try:
+            container.seek(seek_pts, any_frame=False, backward=True, stream=stream)
+        except av.AVError:
+            # Some streams reject the explicit seek; fall back to decoding from start.
+            container.seek(0)
+
+        results: dict[float, Any] = {}
+        target_iter = iter(targets)
+        next_target = next(target_iter, None)
+        for frame in container.decode(stream):
+            if next_target is None:
+                break
+            ts = float(frame.pts * frame.time_base) if frame.pts is not None else None
+            if ts is None:
+                continue
+            # Walk past targets we've already overshot — we keep the closest
+            # frame within tolerance.
+            while next_target is not None and ts >= next_target - tolerance_s:
+                if abs(ts - next_target) <= tolerance_s or ts >= next_target:
+                    img = frame.to_image()  # PIL.Image.Image (RGB)
+                    results.setdefault(next_target, img)
+                    next_target = next(target_iter, None)
+                else:
+                    break
+    finally:
+        container.close()
+
+    return [results[ts] for ts in timestamps if ts in results]
+
+    def video_for_episode(
+        self,
+        record: EpisodeRecord,
+        max_frames: int,
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        """Return up to ``max_frames`` images uniformly sampled across the episode.
+
+        The whole episode duration is covered; the model picks subtask
+        boundaries from the temporal pooling it does internally.
+        """
+        target = camera_key if camera_key is not None else self.camera_key
+        if max_frames <= 0 or target is None or not record.frame_timestamps:
+            return []
+        n_frames = min(max_frames, len(record.frame_timestamps))
+        if n_frames == len(record.frame_timestamps):
+            timestamps = list(record.frame_timestamps)
+        else:
+            t0 = record.frame_timestamps[0]
+            t_last = record.frame_timestamps[-1]
+            if t_last <= t0:
+                timestamps = [float(t0)] * n_frames
+            else:
+                step = (t_last - t0) / (n_frames - 1) if n_frames > 1 else 0.0
+                timestamps = [float(t0 + i * step) for i in range(n_frames)]
+        return self.frames_at(record, timestamps, camera_key=target)
+
+
+def make_frame_provider(root: Path, camera_key: str | None = None) -> FrameProvider:
+    """Build a :class:`VideoFrameProvider` if videos are present, else null."""
+    try:
+        provider = VideoFrameProvider(root=root, camera_key=camera_key)
+    except Exception:
+        return null_provider()
+    if provider.camera_key is None:
+        return null_provider()
+    return provider
+
+
+def to_image_blocks(images: list[Any]) -> list[dict[str, Any]]:
+    """Convert PIL images to Qwen-VL-compatible content blocks."""
+    return [{"type": "image", "image": img} for img in images]
+
+
+def to_video_block(images: list[Any]) -> list[dict[str, Any]]:
+    """Wrap a list of PIL images as one Qwen-VL video block.
+
+    Returns ``[]`` when the list is empty, so the caller can splat the result
+    into a content array without a separate emptiness check.
+    """
+    if not images:
+        return []
+    return [{"type": "video", "video": list(images)}]
+
+
+def to_video_url_block(url: str | None, fps: float = 2.0) -> list[dict[str, Any]]:
+    """Wrap a video file URL as one ``video_url`` block.
+
+    Used by the ``openai`` backend (transformers serve / vllm serve /
+    ktransformers serve), where the server handles frame sampling.
+    Returns ``[]`` when ``url`` is ``None`` so the caller can splat.
+    """
+    if not url:
+        return []
+    return [{"type": "video_url", "video_url": {"url": url}, "fps": fps}]
+
+
+def episode_clip_path(
+    record: EpisodeRecord,
+    provider: "VideoFrameProvider",
+    cache_dir: Path,
+) -> Path | None:
+    """Extract the episode's subclip to ``cache_dir/ep_{idx:06d}.mp4``.
+
+    Returns ``None`` if the dataset has no video tracks. Skips re-extract
+    when the cached clip already exists. Re-encodes to H.264
+    (libx264) so the resulting mp4 is decodable by every downstream
+    video processor — stream-copy would inherit the source codec
+    (often AV1 in modern LeRobot datasets), which vllm's libav build
+    cannot decode.
+    """
+    import subprocess  # noqa: PLC0415
+
+    if provider.camera_key is None:
+        return None
+    cache_dir.mkdir(parents=True, exist_ok=True)
+    out_path = cache_dir / f"ep_{record.episode_index:06d}.mp4"
+    if out_path.exists() and out_path.stat().st_size > 0:
+        return out_path
+    ep = provider._meta.episodes[record.episode_index]
+    from_timestamp = float(ep[f"videos/{provider.camera_key}/from_timestamp"])
+    to_timestamp = float(ep[f"videos/{provider.camera_key}/to_timestamp"])
+    src = provider.root / provider._meta.get_video_file_path(
+        record.episode_index, provider.camera_key
+    )
+    cmd = [
+        "ffmpeg",
+        "-y",
+        "-loglevel",
+        "error",
+        "-ss",
+        f"{from_timestamp:.3f}",
+        "-to",
+        f"{to_timestamp:.3f}",
+        "-i",
+        str(src),
+        "-c:v",
+        "libx264",
+        "-preset",
+        "ultrafast",
+        "-crf",
+        "23",
+        "-pix_fmt",
+        "yuv420p",
+        "-an",
+        str(out_path),
+    ]
+    try:
+        subprocess.run(cmd, check=True, timeout=300)
+    except (subprocess.CalledProcessError, subprocess.TimeoutExpired, FileNotFoundError):
+        return None
+    return out_path if out_path.exists() and out_path.stat().st_size > 0 else None

src/lerobot/annotations/steerable_pipeline/modules/__init__.py

@@ -0,0 +1,25 @@
+#!/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 .general_vqa import GeneralVqaModule
+from .interjections_and_speech import InterjectionsAndSpeechModule
+from .plan_subtasks_memory import PlanSubtasksMemoryModule
+
+__all__ = [
+    "GeneralVqaModule",
+    "InterjectionsAndSpeechModule",
+    "PlanSubtasksMemoryModule",
+]

src/lerobot/annotations/steerable_pipeline/modules/general_vqa.py

@@ -0,0 +1,238 @@
+#!/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.
+"""Module 3: general VQA at a timed cadence.
+
+Anchors ``K`` (question, answer) pairs to ``K`` consecutive frames per
+emission. For datasets with multiple cameras, every emission tick produces
+one ``(vqa, user)`` + ``(vqa, assistant)`` pair *per camera*: each pair is
+generated against that camera's frame and stamped with the matching
+``camera`` field on the emitted rows. The resolver disambiguates via
+``camera=...``; recipes that consume VQA do so through one sub-recipe
+per camera (see ``recipes/pi05_hirobot.yaml``).
+
+Within a single (frame, camera) we still emit at most one ``(vqa, user)``
+and one ``(vqa, assistant)`` row, so the resolver contract stays scalar.
+
+Question types covered (per the plan's Module 3 table): bbox, keypoint,
+count, attribute, spatial. The assistant's ``content`` is a JSON string
+whose schema depends on the question type. Malformed JSON triggers one
+retry inside :meth:`VlmClient.generate_json`.
+"""
+
+from __future__ import annotations
+
+import json
+import random
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+from ..config import Module3Config
+from ..frames import FrameProvider, null_provider, to_image_blocks
+from ..prompts import load as load_prompt
+from ..reader import EpisodeRecord
+from ..staging import EpisodeStaging
+from ..validator import classify_vqa_answer
+from ..vlm_client import VlmClient
+
+
+def _emission_anchor_indices(frame_timestamps: Sequence[float], hz: float, k: int) -> list[int]:
+    """Return the relative frame indices to anchor VQA emissions to.
+
+    For each emission tick (every ``1/hz`` seconds), we anchor ``k``
+    consecutive frames starting at the tick. Ticks fall on the nearest
+    available source frame timestamp.
+    """
+    if hz <= 0 or k <= 0 or not frame_timestamps:
+        return []
+    t0 = frame_timestamps[0]
+    t_last = frame_timestamps[-1]
+    period = 1.0 / hz
+    indices: list[int] = []
+    t = t0
+    while t <= t_last + 1e-9:
+        # find the index of the nearest frame to t
+        nearest_i = min(range(len(frame_timestamps)), key=lambda i: abs(frame_timestamps[i] - t))
+        for offset in range(k):
+            j = nearest_i + offset
+            if j >= len(frame_timestamps):
+                break
+            if not indices or indices[-1] != j:
+                indices.append(j)
+        t += period
+    # dedupe while preserving order
+    seen: set[int] = set()
+    deduped: list[int] = []
+    for i in indices:
+        if i in seen:
+            continue
+        seen.add(i)
+        deduped.append(i)
+    return deduped
+
+
+@dataclass
+class GeneralVqaModule:
+    """Emit grounded VQA pairs at a timed cadence."""
+
+    vlm: VlmClient
+    config: Module3Config
+    seed: int = 1729
+    frame_provider: FrameProvider = field(default_factory=null_provider)
+
+    @property
+    def enabled(self) -> bool:
+        return self.config.enabled
+
+    def run_episode(self, record: EpisodeRecord, staging: EpisodeStaging) -> None:
+        if not record.frame_timestamps:
+            staging.write("module_3", [])
+            return
+        rng = random.Random(f"{self.seed}:{record.episode_index}:vqa")
+        anchor_idx = _emission_anchor_indices(
+            record.frame_timestamps, self.config.vqa_emission_hz, self.config.K
+        )
+        cameras = self._target_cameras()
+        if not cameras:
+            # No camera available — emit nothing rather than producing
+            # untagged rows that would fail validation. Surface a loud one-
+            # time warning so this is never silently a no-op.
+            if not getattr(self, "_warned_no_camera", False):
+                import logging  # noqa: PLC0415
+
+                logging.getLogger(__name__).warning(
+                    "Module 3 (VQA) found no cameras on the frame provider — "
+                    "every episode will emit zero VQA rows. Check that the "
+                    "dataset declares observation.images.* features in "
+                    "meta/info.json; passing --vlm.camera_key=<key> at the "
+                    "CLI now also seeds the cameras list as a fallback."
+                )
+                self._warned_no_camera = True
+            staging.write("module_3", [])
+            return
+
+        # Build all messages first (one per (frame, camera)), then issue them
+        # as a single batched generate_json call so the client can fan them
+        # out concurrently.
+        per_call: list[tuple[float, str, str, list[dict[str, Any]]]] = []
+        for idx in anchor_idx:
+            ts = float(record.frame_timestamps[idx])
+            qtype = rng.choice(self.config.question_types)
+            for camera in cameras:
+                messages = self._build_messages(record, qtype, ts, camera)
+                # Skip cameras that decoded to zero frames at this ts: no point
+                # asking the VLM to ground a bbox without an image.
+                if not _has_image_block(messages):
+                    continue
+                per_call.append((ts, camera, qtype, messages))
+
+        if not per_call:
+            staging.write("module_3", [])
+            return
+
+        results = self.vlm.generate_json([m for _, _, _, m in per_call])
+
+        rows: list[dict[str, Any]] = []
+        for (ts, camera, _qtype, _messages), result in zip(per_call, results):
+            qa = self._postprocess(result)
+            if qa is None:
+                continue
+            question, answer = qa
+            rows.append(
+                {
+                    "role": "user",
+                    "content": question,
+                    "style": "vqa",
+                    "timestamp": ts,
+                    "camera": camera,
+                    "tool_calls": None,
+                }
+            )
+            rows.append(
+                {
+                    "role": "assistant",
+                    "content": json.dumps(answer, sort_keys=True),
+                    "style": "vqa",
+                    "timestamp": ts,
+                    "camera": camera,
+                    "tool_calls": None,
+                }
+            )
+        staging.write("module_3", rows)
+
+    def _target_cameras(self) -> list[str]:
+        """Return the cameras Module 3 should iterate per emission tick.
+
+        Defaults to every camera the provider exposes. Datasets with no
+        cameras (or test/null providers) yield an empty list, which makes
+        ``run_episode`` a no-op.
+        """
+        return list(getattr(self.frame_provider, "camera_keys", []) or [])
+
+    def _build_messages(
+        self,
+        record: EpisodeRecord,
+        question_type: str,
+        frame_timestamp: float,
+        camera_key: str,
+    ) -> list[dict[str, Any]]:
+        prompt = load_prompt("module_3_vqa").format(
+            episode_task=record.episode_task,
+            question_type=question_type,
+        )
+        images = self.frame_provider.frames_at(
+            record, [frame_timestamp], camera_key=camera_key
+        )
+        content = [*to_image_blocks(images), {"type": "text", "text": prompt}]
+        return [{"role": "user", "content": content}]
+
+    def _postprocess(self, result: Any) -> tuple[str, dict[str, Any]] | None:
+        if not isinstance(result, dict):
+            return None
+        question = result.get("question")
+        answer = result.get("answer")
+        if not isinstance(question, str) or not question.strip():
+            return None
+        if not isinstance(answer, dict):
+            return None
+        # The validator will enforce shape; here we just sanity-check that the
+        # answer matches *some* known shape so we can drop garbage early.
+        if classify_vqa_answer(answer) is None:
+            return None
+        return question.strip(), answer
+
+    def _generate_one(
+        self,
+        record: EpisodeRecord,
+        question_type: str,
+        frame_timestamp: float,
+        camera_key: str,
+    ) -> tuple[str, dict[str, Any]] | None:
+        messages = self._build_messages(record, question_type, frame_timestamp, camera_key)
+        result = self.vlm.generate_json([messages])[0]
+        return self._postprocess(result)
+
+
+def _has_image_block(messages: list[dict[str, Any]]) -> bool:
+    """Return True if any user content block is a populated image block."""
+    for msg in messages:
+        content = msg.get("content")
+        if not isinstance(content, list):
+            continue
+        for block in content:
+            if isinstance(block, dict) and block.get("type") == "image":
+                return True
+    return False

src/lerobot/annotations/steerable_pipeline/modules/interjections_and_speech.py

@@ -0,0 +1,231 @@
+#!/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.
+"""Module 2: interjections + paired speech (EVENT styles + speech atoms).
+
+Two sub-passes:
+
+1. At ``t=0``, emit ONLY a speech tool-call atom (acknowledgement of the
+   canonical task). No interjection row — the canonical task is already the
+   user utterance from ``meta/tasks.parquet``.
+
+2. For mid-episode interruptions, emit a co-timestamped pair:
+       {role:user, style:interjection, content:<text>}
+       speech atom (role:assistant, style:None, tool_calls=[say(...)])
+   Both rows go in ``language_events`` at the same timestamp.
+
+Module 1's :meth:`run_plan_updates` reuses Module 2's interjection
+timestamps to refresh the ``plan`` row at the same instant.
+"""
+
+from __future__ import annotations
+
+import random
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+from ..config import Module2Config
+from ..frames import FrameProvider, null_provider, to_image_blocks
+from ..prompts import load as load_prompt
+from ..reader import EpisodeRecord
+from ..staging import EpisodeStaging
+from ..vlm_client import VlmClient
+from ..writer import speech_atom
+
+
+def _snap_to_frame(t: float, frame_timestamps: Sequence[float]) -> float:
+    if not frame_timestamps:
+        return float(t)
+    return float(min(frame_timestamps, key=lambda f: abs(f - t)))
+
+
+@dataclass
+class InterjectionsAndSpeechModule:
+    """Generate task-start speech and mid-episode interjection/speech pairs."""
+
+    vlm: VlmClient
+    config: Module2Config
+    seed: int = 1729
+    frame_provider: FrameProvider = field(default_factory=null_provider)
+
+    @property
+    def enabled(self) -> bool:
+        return self.config.enabled
+
+    def run_episode(self, record: EpisodeRecord, staging: EpisodeStaging) -> None:
+        rows: list[dict[str, Any]] = []
+        if record.frame_timestamps:
+            t0 = float(record.frame_timestamps[0])
+            initial = self._initial_speech(record)
+            if initial:
+                rows.append(speech_atom(t0, initial))
+        # Pull Module 1's subtask spans for this episode so the
+        # interjection prompt can ground itself in the actual current
+        # subtask at each chosen timestamp. Module 1 ran first.
+        subtask_spans = self._read_subtask_spans(staging)
+        rows.extend(self._mid_episode_interjections(record, subtask_spans))
+        staging.write("module_2", rows)
+
+    @staticmethod
+    def _read_subtask_spans(staging: EpisodeStaging) -> list[dict[str, Any]]:
+        rows = [r for r in staging.read("module_1") if r.get("style") == "subtask"]
+        rows.sort(key=lambda r: float(r["timestamp"]))
+        spans: list[dict[str, Any]] = []
+        last_t: float | None = None
+        for r in rows:
+            t = float(r["timestamp"])
+            if last_t is not None and spans:
+                spans[-1]["end"] = t
+            spans.append({"text": r.get("content") or "", "start": t, "end": t})
+            last_t = t
+        return spans
+
+    @staticmethod
+    def _subtask_at(spans: Sequence[dict[str, Any]], t: float) -> str | None:
+        current: str | None = None
+        for span in spans:
+            if float(span["start"]) <= t:
+                current = span.get("text")
+            else:
+                break
+        return current
+
+    def _initial_speech(self, record: EpisodeRecord) -> str | None:
+        prompt = load_prompt("module_2_initial_speech").format(
+            episode_task=record.episode_task,
+        )
+        messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}]
+        result = self.vlm.generate_json([messages])[0]
+        if isinstance(result, dict) and isinstance(result.get("text"), str):
+            text = result["text"].strip()
+            if text:
+                return text
+        return None
+
+    def _mid_episode_interjections(
+        self,
+        record: EpisodeRecord,
+        subtask_spans: Sequence[dict[str, Any]],
+    ) -> list[dict[str, Any]]:
+        """Generate interjections aligned with the actual demo trajectory.
+
+        Teleop data is frozen — the robot already executed every step in
+        the video. A *counterfactual* interjection like "actually skip
+        the wipe" contradicts what then happens in the video, which is
+        what qwen36moe-10/11 surfaced as low-quality interjections.
+
+        Instead, anchor every interjection at a subtask boundary and
+        write it as a natural user request for the *upcoming* subtask.
+        The robot's visible next behavior IS the interjection's effect,
+        so the training signal stays consistent: interjection text →
+        plan refresh → action stream all line up.
+        """
+        if self.config.max_interjections_per_episode <= 0:
+            return []
+        if len(subtask_spans) < 2:
+            # Need at least one transition (subtask 0 → subtask 1).
+            return []
+        # Deterministic per-episode RNG so reruns are stable across SLURM jobs.
+        rng = random.Random(f"{self.seed}:{record.episode_index}:interjection")
+
+        # Boundaries: the start time of every subtask except the first
+        # (which is just t0 and is covered by the initial-task speech atom).
+        boundaries: list[tuple[float, str, str]] = []
+        for i in range(1, len(subtask_spans)):
+            ts = float(subtask_spans[i]["start"])
+            if ts < self.config.interjection_min_t:
+                continue
+            prev_text = (subtask_spans[i - 1].get("text") or "").strip()
+            next_text = (subtask_spans[i].get("text") or "").strip()
+            if not next_text:
+                continue
+            boundaries.append((ts, prev_text, next_text))
+        if not boundaries:
+            return []
+
+        n = min(self.config.max_interjections_per_episode, len(boundaries))
+        chosen = sorted(rng.sample(boundaries, n), key=lambda b: b[0])
+
+        out: list[dict[str, Any]] = []
+        for t, prev_subtask, next_subtask in chosen:
+            t_snap = _snap_to_frame(t, record.frame_timestamps)
+            # Window straddles the boundary so the VLM sees the end of the
+            # previous subtask and the start of the next one — same
+            # conditioning the policy will see at training time.
+            window_ts = self._window_timestamps(t_snap, record.frame_timestamps)
+            prompt = load_prompt("module_2_interjection").format(
+                episode_task=record.episode_task,
+                prev_subtask=prev_subtask or "(starting from initial state)",
+                next_subtask=next_subtask,
+                timestamp=t_snap,
+                window_seconds=self.config.interjection_window_seconds,
+            )
+            images = self.frame_provider.frames_at(record, window_ts)
+            content = [*to_image_blocks(images), {"type": "text", "text": prompt}]
+            messages = [{"role": "user", "content": content}]
+            result = self.vlm.generate_json([messages])[0]
+            if not isinstance(result, dict):
+                continue
+            interjection_text = result.get("interjection")
+            speech_text = result.get("speech")
+            if not isinstance(interjection_text, str) or not interjection_text.strip():
+                continue
+            if not isinstance(speech_text, str) or not speech_text.strip():
+                continue
+            out.append(
+                {
+                    "role": "user",
+                    "content": interjection_text.strip(),
+                    "style": "interjection",
+                    "timestamp": t_snap,
+                    "tool_calls": None,
+                }
+            )
+            out.append(speech_atom(t_snap, speech_text.strip()))
+        return out
+
+    def _window_timestamps(
+        self, t_anchor: float, frame_timestamps: Sequence[float]
+    ) -> list[float]:
+        """Return a small set of frame timestamps centered on ``t_anchor``.
+
+        The window straddles the subtask boundary the interjection sits
+        on: roughly half the frames cover the end of the previous
+        subtask, half cover the start of the next one. The VLM therefore
+        sees BOTH what just finished AND what's about to start, which is
+        the conditioning we need to write a natural "now please do X"
+        request that matches the visible upcoming behavior.
+        """
+        if not frame_timestamps:
+            return [t_anchor]
+        n = max(1, int(self.config.interjection_window_frames))
+        if n == 1:
+            return [t_anchor]
+        window = float(self.config.interjection_window_seconds)
+        step = window / max(1, n - 1)
+        # Center the window on the anchor so half lands before, half after.
+        start_offset = -window / 2.0
+        targets = [t_anchor + start_offset + step * i for i in range(n)]
+        last_ts = float(frame_timestamps[-1])
+        snapped: list[float] = []
+        seen: set[float] = set()
+        for tgt in targets:
+            clamped = min(last_ts, max(0.0, tgt))
+            t = _snap_to_frame(clamped, frame_timestamps)
+            if t not in seen:
+                seen.add(t)
+                snapped.append(t)
+        return snapped or [t_anchor]

src/lerobot/annotations/steerable_pipeline/modules/plan_subtasks_memory.py

@@ -0,0 +1,450 @@
+#!/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.
+"""Module 1: subtask decomposition + plan + memory (PERSISTENT styles)."""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+from pathlib import Path
+
+from ..config import Module1Config
+from ..frames import (
+    FrameProvider,
+    VideoFrameProvider,
+    episode_clip_path,
+    null_provider,
+    to_video_block,
+    to_video_url_block,
+)
+from ..prompts import load as load_prompt
+from ..reader import EpisodeRecord
+from ..staging import EpisodeStaging
+from ..vlm_client import VlmClient
+
+
+def _snap_to_frame(t: float, frame_timestamps: Sequence[float]) -> float:
+    """Snap an arbitrary float to the nearest exact source frame timestamp."""
+    if not frame_timestamps:
+        return float(t)
+    nearest = min(frame_timestamps, key=lambda f: abs(f - t))
+    return float(nearest)
+
+
+@dataclass
+class PlanSubtasksMemoryModule:
+    """Generate subtask spans, plan, and memory rows.
+
+    All output is persistent (lives in ``language_persistent``):
+
+    - ``subtask`` rows: one per span, stamped at the span's *start* timestamp
+      (snapped to an exact frame).
+    - ``plan`` rows: emitted at ``t=0``; refreshed at every interjection
+      timestamp via :meth:`run_plan_updates` (called by the executor after
+      Module 2 completes).
+    - ``memory`` rows: emitted at each subtask boundary (= subtask start
+      timestamp from the second subtask onward).
+    """
+
+    vlm: VlmClient
+    config: Module1Config
+    frame_provider: FrameProvider = field(default_factory=null_provider)
+
+    @property
+    def enabled(self) -> bool:
+        return self.config.enabled
+
+    def run_episode(self, record: EpisodeRecord, staging: EpisodeStaging) -> None:
+        rows: list[dict[str, Any]] = []
+        # Resolve the task that drives every other Module-1 prompt. May be
+        # the canonical ``record.episode_task`` (default), or a fresh
+        # description derived from the video when the canonical task is
+        # empty / placeholder / forced-off (see Module1Config.derive_task_*).
+        effective_task = self._resolve_effective_task(record)
+        # ``task_aug`` rows at t=0 (role=user), one per rephrasing — the
+        # PR 1 renderer rotates ``${task}`` deterministically through them
+        # so the policy sees diverse phrasings during training.
+        t0 = float(record.frame_timestamps[0]) if record.frame_timestamps else 0.0
+        if self.config.n_task_rephrasings > 0 and effective_task:
+            rephrasings = self._generate_task_rephrasings(
+                effective_task, n=self.config.n_task_rephrasings
+            )
+            # Always include the effective task itself as the first variant
+            # so the rotation is guaranteed to cover the source-of-truth
+            # phrasing, not just synthetic alternatives.
+            seen: set[str] = set()
+            ordered = [effective_task, *rephrasings]
+            for phrasing in ordered:
+                key = phrasing.strip()
+                if not key or key in seen:
+                    continue
+                seen.add(key)
+                rows.append(
+                    {
+                        "role": "user",
+                        "content": key,
+                        "style": "task_aug",
+                        "timestamp": t0,
+                        "tool_calls": None,
+                    }
+                )
+
+        subtask_spans = self._generate_subtasks(record, task=effective_task)
+        # subtask rows
+        for span in subtask_spans:
+            rows.append(
+                {
+                    "role": "assistant",
+                    "content": span["text"],
+                    "style": "subtask",
+                    "timestamp": _snap_to_frame(span["start"], record.frame_timestamps),
+                    "tool_calls": None,
+                }
+            )
+        # Plan rows at every subtask boundary — including t=0 (start of
+        # the first subtask). Because the plan is just a numbered list
+        # of *still-todo* subtasks, re-emitting at each boundary makes
+        # the active plan shrink as work progresses: at frame t the
+        # rendered ``${plan}`` is the most recent emission, which
+        # contains exactly the subtasks that started at or after the
+        # current span. Saves the runtime from having to derive
+        # "what's still left" at inference time.
+        for span in subtask_spans:
+            boundary_t = _snap_to_frame(span["start"], record.frame_timestamps)
+            plan_text = self._generate_plan(
+                record, subtask_spans, refresh_t=boundary_t, task=effective_task
+            )
+            if plan_text is not None:
+                rows.append(
+                    {
+                        "role": "assistant",
+                        "content": plan_text,
+                        "style": "plan",
+                        "timestamp": float(boundary_t),
+                        "tool_calls": None,
+                    }
+                )
+        # memory rows at every subtask boundary except the very first start
+        prior_memory = ""
+        for i, span in enumerate(subtask_spans[1:], start=1):
+            completed = subtask_spans[i - 1]["text"]
+            remaining = [s["text"] for s in subtask_spans[i:]]
+            mem_text = self._generate_memory(
+                record, prior_memory, completed, remaining, task=effective_task
+            )
+            if mem_text:
+                ts = _snap_to_frame(span["start"], record.frame_timestamps)
+                rows.append(
+                    {
+                        "role": "assistant",
+                        "content": mem_text,
+                        "style": "memory",
+                        "timestamp": ts,
+                        "tool_calls": None,
+                    }
+                )
+                prior_memory = mem_text
+        staging.write("module_1", rows)
+
+    # ------------------------------------------------------------------
+    # Task derivation + rephrasings
+    # ------------------------------------------------------------------
+
+    _PLACEHOLDER_TASKS: frozenset[str] = frozenset(
+        {
+            "debug",
+            "test",
+            "tbd",
+            "todo",
+            "n/a",
+            "na",
+            "untitled",
+            "unnamed",
+            "default",
+            "placeholder",
+        }
+    )
+
+    def _resolve_effective_task(self, record: EpisodeRecord) -> str:
+        """Decide which task string drives Module 1 for this episode.
+
+        Returns the user-supplied ``record.episode_task`` unless
+        ``derive_task_from_video`` says otherwise (see config docstring).
+        Falls back gracefully to the canonical task if video derivation
+        fails.
+        """
+        canonical = (record.episode_task or "").strip()
+        mode = (self.config.derive_task_from_video or "off").strip().lower()
+        if mode == "always":
+            derived = self._derive_task_from_video(record)
+            return derived or canonical
+        if mode == "if_short" and self._task_seems_bad(canonical):
+            derived = self._derive_task_from_video(record)
+            if derived:
+                return derived
+        return canonical
+
+    def _task_seems_bad(self, task: str) -> bool:
+        if not task:
+            return True
+        if len(task.split()) < int(self.config.derive_task_min_words):
+            return True
+        if task.lower() in self._PLACEHOLDER_TASKS:
+            return True
+        return False
+
+    def _derive_task_from_video(self, record: EpisodeRecord) -> str | None:
+        """Ask the VLM "what is this video about" with no task hint at all."""
+        prompt = load_prompt("module_1_video_task")
+        video_block = self._episode_video_block(record)
+        content = [*video_block, {"type": "text", "text": prompt}]
+        messages = [{"role": "user", "content": content}]
+        result = self.vlm.generate_json([messages])[0]
+        if isinstance(result, dict) and isinstance(result.get("task"), str):
+            text = result["task"].strip()
+            if text:
+                return text
+        return None
+
+    def _generate_task_rephrasings(self, base_task: str, *, n: int) -> list[str]:
+        """Generate ``n`` text-only paraphrases of ``base_task``."""
+        if n <= 0 or not base_task:
+            return []
+        prompt = load_prompt("module_1_task_rephrasings").format(
+            base_task=base_task, n=n
+        )
+        messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}]
+        result = self.vlm.generate_json([messages])[0]
+        if not isinstance(result, dict):
+            return []
+        raw = result.get("rephrasings")
+        if not isinstance(raw, list):
+            return []
+        out: list[str] = []
+        for item in raw:
+            if isinstance(item, str):
+                cleaned = item.strip().strip('"').strip("'")
+                if cleaned:
+                    out.append(cleaned)
+        return out[:n]
+
+    def _episode_video_block(self, record: EpisodeRecord) -> list[dict[str, Any]]:
+        """Same video block ``_generate_subtasks`` builds — extracted helper."""
+        if not record.frame_timestamps:
+            return []
+        if self.config.use_video_url and isinstance(self.frame_provider, VideoFrameProvider):
+            cache_dir = Path(self.frame_provider.root) / ".annotate_staging" / ".video_clips"
+            clip = episode_clip_path(record, self.frame_provider, cache_dir)
+            return (
+                to_video_url_block(f"file://{clip}", fps=self.config.use_video_url_fps)
+                if clip is not None
+                else []
+            )
+        episode_duration = record.frame_timestamps[-1] - record.frame_timestamps[0]
+        target_count = max(
+            1, int(round(episode_duration * self.config.frames_per_second))
+        )
+        target_count = min(target_count, self.config.max_video_frames)
+        video_frames = self.frame_provider.video_for_episode(record, target_count)
+        return to_video_block(video_frames)
+
+    def run_plan_updates(
+        self,
+        record: EpisodeRecord,
+        staging: EpisodeStaging,
+        interjection_times: Sequence[float],
+        interjection_texts: Sequence[str] | None = None,
+    ) -> None:
+        """Append additional ``plan`` rows at every interjection timestamp.
+
+        Plans refresh ONLY on user interjections — subtask generation
+        runs ~1 Hz at inference, but plan re-emission is event-driven.
+        Now also forwards the interjection's own text into the prompt so
+        the refreshed plan can actually reflect the user's correction
+        (the previous version told the model "an interjection happened"
+        without telling it what the user said).
+        """
+        existing = staging.read("module_1")
+        spans = self._reconstruct_subtasks_from_rows(existing)
+        already_planned: set[float] = {
+            float(r["timestamp"]) for r in existing if r.get("style") == "plan"
+        }
+        new_rows = list(existing)
+
+        texts: list[str | None] = (
+            [None] * len(interjection_times)
+            if interjection_texts is None
+            else [str(t) if t else None for t in interjection_texts]
+        )
+        for raw_t, inter_text in zip(interjection_times, texts):
+            t = _snap_to_frame(raw_t, record.frame_timestamps)
+            if t in already_planned:
+                continue
+            already_planned.add(t)
+            plan_text = self._generate_plan(
+                record, spans, refresh_t=t, interjection=inter_text
+            )
+            if plan_text is not None:
+                new_rows.append(
+                    {
+                        "role": "assistant",
+                        "content": plan_text,
+                        "style": "plan",
+                        "timestamp": t,
+                        "tool_calls": None,
+                    }
+                )
+        staging.write("module_1", new_rows)
+
+    @staticmethod
+    def _reconstruct_subtasks_from_rows(rows: Sequence[dict[str, Any]]) -> list[dict[str, Any]]:
+        out = []
+        last_t: float | None = None
+        for row in sorted(
+            (r for r in rows if r.get("style") == "subtask"),
+            key=lambda r: float(r["timestamp"]),
+        ):
+            t = float(row["timestamp"])
+            if last_t is not None:
+                out[-1]["end"] = t
+            out.append({"text": row.get("content") or "", "start": t, "end": t})
+            last_t = t
+        return out
+
+    def _generate_subtasks(
+        self, record: EpisodeRecord, *, task: str | None = None
+    ) -> list[dict[str, Any]]:
+        if record.row_count == 0 or not record.frame_timestamps:
+            return []
+        episode_duration = record.frame_timestamps[-1] - record.frame_timestamps[0]
+        prompt = load_prompt("module_1_subtasks").format(
+            episode_task=(task if task is not None else record.episode_task),
+            min_subtask_seconds=self.config.min_subtask_seconds,
+            max_steps=self.config.plan_max_steps,
+            episode_duration=f"{episode_duration:.3f}",
+        )
+        if self.config.use_video_url and isinstance(self.frame_provider, VideoFrameProvider):
+            cache_dir = Path(self.frame_provider.root) / ".annotate_staging" / ".video_clips"
+            clip = episode_clip_path(record, self.frame_provider, cache_dir)
+            video_block = (
+                to_video_url_block(f"file://{clip}", fps=self.config.use_video_url_fps)
+                if clip is not None
+                else []
+            )
+        else:
+            target_count = max(
+                1,
+                int(round(episode_duration * self.config.frames_per_second)),
+            )
+            target_count = min(target_count, self.config.max_video_frames)
+            video_frames = self.frame_provider.video_for_episode(record, target_count)
+            video_block = to_video_block(video_frames)
+        content = [*video_block, {"type": "text", "text": prompt}]
+        messages = [{"role": "user", "content": content}]
+        result = self.vlm.generate_json([messages])[0]
+        spans = result.get("subtasks") if isinstance(result, dict) else None
+        if not spans:
+            return []
+        # clamp to [t0, t_last] and sort
+        t0 = record.frame_timestamps[0]
+        t_last = record.frame_timestamps[-1]
+        cleaned: list[dict[str, Any]] = []
+        for span in spans:
+            try:
+                start = float(span["start"])
+                end = float(span["end"])
+                text = str(span["text"]).strip()
+            except (KeyError, ValueError, TypeError):
+                continue
+            start = max(t0, min(start, t_last))
+            end = max(t0, min(end, t_last))
+            if end < start:
+                start, end = end, start
+            if not text:
+                continue
+            cleaned.append({"text": text, "start": start, "end": end})
+        cleaned.sort(key=lambda s: s["start"])
+        return cleaned
+
+    def _generate_plan(
+        self,
+        record: EpisodeRecord,  # noqa: ARG002  (kept for signature stability)
+        subtask_spans: Sequence[dict[str, Any]],
+        *,
+        refresh_t: float | None = None,
+        interjection: str | None = None,  # noqa: ARG002
+        task: str | None = None,  # noqa: ARG002
+    ) -> str | None:
+        """Deterministic plan = numbered list of *still-todo* subtasks.
+
+        Previously this called the VLM with a prompt that asked it to
+        compress the subtasks into a "compact hierarchical plan". That
+        produced longer-than-necessary plans, cost an extra VLM round-trip
+        per episode (plus one per interjection on refresh), and could
+        diverge from the actual subtask sequence the model is going to
+        execute. Replacing it with a plain summarisation keeps the plan
+        tightly aligned with the upcoming subtasks and removes the VLM
+        call entirely.
+
+        Layout (matches the v2 plan style — short imperative fragments
+        prefixed by "N. "):
+
+            1. <subtask 1>
+            2. <subtask 2>
+            ...
+
+        On a refresh at ``refresh_t`` (called from ``run_plan_updates``
+        on interjection events), only subtasks whose start is at or
+        after ``refresh_t`` are included — the plan shrinks as work
+        progresses, so it always describes what's left.
+        """
+        if not subtask_spans:
+            return None
+        remaining = [
+            s for s in subtask_spans
+            if refresh_t is None or float(s.get("start", 0.0)) >= float(refresh_t)
+        ]
+        if not remaining:
+            # Past the last subtask boundary on a late refresh — nothing
+            # left to plan; emit None so the caller skips the row.
+            return None
+        return "\n".join(
+            f"{i}. {span.get('text', '').strip()}"
+            for i, span in enumerate(remaining, start=1)
+        )
+
+    def _generate_memory(
+        self,
+        record: EpisodeRecord,
+        prior_memory: str,
+        completed: str,
+        remaining: Sequence[str],
+        *,
+        task: str | None = None,
+    ) -> str:
+        prompt = load_prompt("module_1_memory").format(
+            episode_task=(task if task is not None else record.episode_task),
+            prior_memory=prior_memory or "(none)",
+            completed_subtask=completed,
+            remaining_subtasks=", ".join(remaining) if remaining else "(none)",
+        )
+        messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}]
+        result = self.vlm.generate_json([messages])[0]
+        if isinstance(result, dict) and isinstance(result.get("memory"), str):
+            return result["memory"].strip()
+        return ""

src/lerobot/annotations/steerable_pipeline/prompts/__init__.py

@@ -0,0 +1,33 @@
+#!/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.
+"""Prompt templates loaded as plain text.
+
+One file per use site. Templates use ``str.format(**vars)`` substitution; we
+intentionally avoid jinja2 here so the templates remain inspectable in
+plain editors and roundtrip cleanly through ``ruff format``.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+_DIR = Path(__file__).parent
+
+
+def load(name: str) -> str:
+    """Read prompt template ``name.txt`` from the ``prompts/`` directory."""
+    path = _DIR / f"{name}.txt"
+    return path.read_text(encoding="utf-8")

src/lerobot/annotations/steerable_pipeline/prompts/module_1_memory.txt

@@ -0,0 +1,35 @@
+You are updating the robot's compressed semantic memory at the boundary of
+a completed subtask.
+
+Reference (MEM, Torne 2026):
+"Remove or compress information in the language memory whenever
+appropriate. Keep ONLY the minimal set of relevant information for future
+task execution. Specific object attributes (colors, precise quantities of
+each item) get discarded when their details won't affect subsequent
+actions. Functional outcomes (where items went, how many) are preserved."
+
+Episode task: "{episode_task}"
+Previous memory: {prior_memory}
+Just-completed subtask: "{completed_subtask}"
+Remaining subtasks (for relevance judgement only): {remaining_subtasks}
+
+Write the **shortest possible** state note that future subtasks could
+need. Telegraphic style.
+
+**Hard caps**
+- ≤ 10 words total.
+- No articles. No verbs in past tense ("placed", "moved"). Use
+  comma-separated noun→location fragments.
+- Drop colors/sizes/counts unless a later subtask depends on them.
+- If nothing material changed for downstream subtasks, emit "" (empty
+  string).
+
+Examples
+- Good: "bowl in box, lid open"
+- Good: "3 bowls in cabinet"
+- Good: "cup on tray, drawer closed"
+- Bad:  "The bowl is now in the box and the lid is still open."
+- Bad:  "I placed the green bowl carefully into the cardboard box."
+
+Output strictly valid JSON:
+  {{ "memory": "<≤10-word telegraphic state, or empty>" }}

src/lerobot/annotations/steerable_pipeline/prompts/module_1_subtasks.txt

@@ -0,0 +1,33 @@
+You are labeling a teleoperated robot demonstration.
+
+The user originally asked: "{episode_task}"
+
+You are shown the entire demonstration as a single video. Watch the
+whole clip, then segment it into a list of consecutive atomic subtasks
+the robot performs. Write **ultra-compact** action labels.
+
+Authoring rules — Hi Robot atom granularity, pi0.7-style short prompts:
+
+- Each subtask = one atomic skill the low-level policy can execute.
+- **Hard length cap: ≤ 5 words per subtask.** Drop articles, modifiers,
+  adverbs. Verb + object (+ optional short qualifier) only.
+- Prefer: "pick lettuce", "place bowl in box", "open drawer",
+  "grasp sponge left hand".
+- Avoid: "pick up one piece of lettuce", "carefully place the bowl",
+  "the robot moves its arm to the left".
+- Subtasks are non-overlapping and cover the full episode in order.
+  Choose the cut points yourself based on what you see in the video
+  (gripper open/close events, contact, regrasps, transitions).
+- Each subtask spans at least {min_subtask_seconds} seconds.
+- Do not exceed {max_steps} subtasks total.
+- Every subtask's [start_time, end_time] must lie within
+  [0.0, {episode_duration}] seconds.
+
+Output strictly valid JSON of shape:
+
+  {{
+    "subtasks": [
+      {{"text": "<≤5-word verb phrase>", "start": <float>, "end": <float>}},
+      ...
+    ]
+  }}

src/lerobot/annotations/steerable_pipeline/prompts/module_1_task_rephrasings.txt

@@ -0,0 +1,32 @@
+You are generating training data for a Hi Robot-style policy. We need
+{n} alternative phrasings of the same robot task so the policy sees
+diverse user prompts during training instead of the same canonical
+string repeated every frame.
+
+Original task:
+"{base_task}"
+
+Generate exactly {n} alternative phrasings of the same task. Vary:
+
+- formality (casual / polite / curt)
+- verbosity (mostly short imperative; occasional polite request)
+- word choice (synonyms, different verbs)
+- sentence structure (imperative / question / suggestion)
+
+Hard rules:
+- Each phrasing MUST preserve the exact meaning of the original task.
+  Do not change which object is involved, the destination, or the
+  action. Do not add extra steps. Do not invent new objects.
+- Each phrasing must be a short phrase or sentence, plain prose, no
+  markdown, no quotes, no list numbers.
+- Phrasings must be distinct — no near-duplicates.
+- Output exactly {n} entries.
+
+Output strictly valid JSON:
+  {{
+    "rephrasings": [
+      "<phrasing 1>",
+      "<phrasing 2>",
+      ...
+    ]
+  }}

src/lerobot/annotations/steerable_pipeline/prompts/module_1_video_task.txt

@@ -0,0 +1,17 @@
+The video above shows a robot manipulation episode in full. Look at
+the entire video and describe in ONE concise sentence what the robot
+is doing.
+
+Rules:
+- One sentence, in natural English, like a user instruction.
+- Capture the goal of the demonstration, not low-level motions.
+  Example: "place the yellow cube into the red bin" — not "move the
+  end-effector down 5cm and close the gripper".
+- 4 to 15 words. Plain prose, no markdown, no bullets, no quotes.
+- Do not invent objects or actions that aren't visible.
+- Do not output anything other than the JSON object below.
+
+Output strictly valid JSON:
+  {{
+    "task": "<single concise sentence describing what the robot does in this video>"
+  }}

src/lerobot/annotations/steerable_pipeline/prompts/module_2_initial_speech.txt

@@ -0,0 +1,12 @@
+The user just asked the robot: "{episode_task}".
+
+Generate a short verbal acknowledgement the robot would speak back before
+beginning the task. Style: compact, confident, friendly.
+
+Examples (Hi Robot, Shi 2025): "Sure, I won't put cheese on it.",
+"OK, starting with the sponge.", "Got it.".
+
+Prefer very short replies: "Got it.", "On it.", "OK."
+
+Output strictly valid JSON:
+  {{ "text": "<the spoken acknowledgement>" }}

src/lerobot/annotations/steerable_pipeline/prompts/module_2_interjection.txt

@@ -0,0 +1,46 @@
+You are generating training data for a Hi Robot-style hierarchical
+robot policy. The robot in this demonstration has ALREADY executed
+every step shown in the video — we cannot retroactively change the
+action stream. To keep training data consistent with the video, the
+"interjection" must align with what the robot is *about to do next* in
+the demonstration, framed as a natural mid-task user request.
+
+The episode's overall task: "{episode_task}".
+
+The images above show roughly {window_seconds:.1f} seconds straddling a
+subtask boundary in the demonstration:
+
+- Subtask the robot just finished: "{prev_subtask}"
+- Subtask the robot is about to start: "{next_subtask}"
+- Time into episode: {timestamp:.2f}s
+
+Write ONE compact interjection the user would naturally say at this
+moment to prompt / confirm / encourage the robot to do "{next_subtask}".
+Keep it like a mid-task coaching cue, not a full instruction paragraph.
+Also write the robot's compact verbal acknowledgement.
+
+Hard rules:
+
+- The interjection MUST be consistent with the next subtask. The user
+  cannot ask for something different from what the robot then does in
+  the video. If you're tempted to say "actually skip X" or "do Y
+  instead", DO NOT — those would contradict the demonstration.
+- The interjection must reference an object, location, or action that
+  is plausible given the visible scene and the next subtask text.
+- One short phrase or sentence each. Conversational, not robotic.
+- Prefer direct cues: "{next_subtask}, please."; "Now {next_subtask}."
+- Keep robot speech very short: "OK.", "On it.", "Doing that."
+
+Style examples (vary the phrasing — don't reuse these verbatim):
+  - "Now go ahead and {next_subtask}."
+  - "Great, can you {next_subtask} next?"
+  - "{next_subtask}, please."
+  - "Before you continue, please {next_subtask}."
+  - "Looking good — {next_subtask} now."
+  - "Okay, {next_subtask}."
+
+Output strictly valid JSON:
+  {{
+    "interjection": "<short cue from the user, asking for the next subtask>",
+    "speech":       "<short robot acknowledgement>"
+  }}

src/lerobot/annotations/steerable_pipeline/prompts/module_3_vqa.txt

@@ -0,0 +1,32 @@
+You are generating a frame-grounded visual question/answer pair for
+chain-of-thought training. Reference: ECoT (Zawalski 2024) and Steerable
+Policies — both train policies on grounded features such as bounding box
+pixel coordinates, keypoints, counts, attributes, and spatial relations.
+
+The frame shows a robot working on: "{episode_task}".
+
+Question types and the EXACT answer JSON shape required for each:
+
+  bbox       => {{"detections": [{{"label": "<obj>", "bbox_format": "xyxy",
+                                    "bbox": [x1, y1, x2, y2]}}, ...]}}
+                bbox is in pixel coordinates (x_min, y_min, x_max, y_max).
+                ECoT example: "a white cup [124, 25, 176, 113]".
+
+  keypoint   => {{"label": "<point>", "point_format": "xy",
+                  "point": [x, y]}}
+
+  count      => {{"label": "<obj>", "count": <int>,
+                  "note": "<optional short note>"}}
+
+  attribute  => {{"label": "<obj>", "attribute": "<color|shape|state|...>",
+                  "value": "<observed value>"}}
+
+  spatial    => {{"subject": "<obj>", "relation": "<left_of|right_of|on|in|"
+                  "above|below|near>", "object": "<obj>"}}
+
+Generate a question of type "{question_type}". Output strictly valid JSON:
+
+  {{
+    "question": "<short, frame-grounded question>",
+    "answer":   <object whose shape matches the schema above>
+  }}

src/lerobot/annotations/steerable_pipeline/reader.py

@@ -0,0 +1,219 @@
+#!/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.
+"""Datatrove-shaped reader.
+
+The reader walks ``data/chunk-*/file-*.parquet`` and yields one record per
+episode containing:
+
+- ``episode_index``: int
+- ``frame_timestamps``: tuple[float, ...]
+- ``frame_indices``: tuple[int, ...]
+- ``episode_task``: str (canonical task from ``meta/tasks.parquet``)
+- ``data_path``: pathlib.Path of the source parquet shard
+- ``frames_df``: pandas.DataFrame slice for the episode (only loaded on demand)
+
+This shape lets each module operate per-episode without loading all parquet
+rows into memory at once. It deliberately does not depend on datatrove —
+datatrove integration wraps this generator inside a ``PipelineStep`` in
+:mod:`.executor`.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterator
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+import pyarrow.parquet as pq
+
+from lerobot.datasets.utils import DEFAULT_TASKS_PATH
+
+
+@dataclass
+class EpisodeRecord:
+    """Per-episode record yielded by the reader."""
+
+    episode_index: int
+    episode_task: str
+    frame_timestamps: tuple[float, ...]
+    frame_indices: tuple[int, ...]
+    data_path: Path
+    row_offset: int  # row offset within the parquet file where this episode starts
+    row_count: int  # number of rows for this episode
+
+    def frames_df(self):  # type: ignore[no-untyped-def]
+        """Lazy-load the pandas slice for this episode."""
+        import pandas as pd  # noqa: PLC0415  - deferred for optional dataset extra
+
+        table = pq.read_table(self.data_path)
+        df: pd.DataFrame = table.to_pandas()
+        slice_ = df.iloc[self.row_offset : self.row_offset + self.row_count].reset_index(drop=True)
+        return slice_
+
+
+def _load_tasks_lookup(root: Path) -> dict[int, str]:
+    tasks_path = root / DEFAULT_TASKS_PATH
+    if not tasks_path.exists():
+        return {}
+    table = pq.read_table(tasks_path)
+    cols = {name: table.column(name).to_pylist() for name in table.column_names}
+    if "task_index" in cols and "task" in cols:
+        return dict(zip(cols["task_index"], cols["task"], strict=True))
+    raise ValueError(f"meta/tasks.parquet at {tasks_path} missing 'task_index' or 'task'")
+
+
+def iter_episodes(root: Path, *, only_episodes: tuple[int, ...] | None = None) -> Iterator[EpisodeRecord]:
+    """Yield :class:`EpisodeRecord` for every episode under ``root/data/``.
+
+    Episodes are yielded in ascending ``episode_index`` order. The reader does
+    not assume a specific chunk/file layout: it scans every ``*.parquet``
+    under ``data/`` and groups by ``episode_index``.
+    """
+    tasks = _load_tasks_lookup(root)
+    data_dir = root / "data"
+    parquet_files = sorted(data_dir.rglob("*.parquet"))
+
+    only_set = set(only_episodes) if only_episodes is not None else None
+
+    for path in parquet_files:
+        yield from _iter_one_path(path, tasks, only_set)
+
+
+def _iter_one_path(path: Path, tasks: dict[int, str], only_set: set[int] | None) -> Iterator[EpisodeRecord]:
+    table = pq.read_table(path)
+    names = table.column_names
+    if "episode_index" not in names:
+        return
+    episode_col = table.column("episode_index").to_pylist()
+    timestamp_col = (
+        table.column("timestamp").to_pylist() if "timestamp" in names else [0.0] * len(episode_col)
+    )
+    frame_col = (
+        table.column("frame_index").to_pylist() if "frame_index" in names else list(range(len(episode_col)))
+    )
+    task_col = table.column("task_index").to_pylist() if "task_index" in names else None
+
+    def _build(
+        ep: int,
+        start: int,
+        end: int,
+        task_idx: int | None,
+        ts_buf: list[float],
+        fi_buf: list[int],
+    ) -> EpisodeRecord | None:
+        if only_set is not None and ep not in only_set:
+            return None
+        task = tasks.get(task_idx, "") if task_idx is not None else ""
+        return EpisodeRecord(
+            episode_index=ep,
+            episode_task=task,
+            frame_timestamps=tuple(ts_buf),
+            frame_indices=tuple(fi_buf),
+            data_path=path,
+            row_offset=start,
+            row_count=end - start,
+        )
+
+    cur_ep: int | None = None
+    start_offset = 0
+    ts_buf: list[float] = []
+    fi_buf: list[int] = []
+    cur_task_idx: int | None = None
+
+    for i, ep in enumerate(episode_col):
+        if cur_ep is None:
+            cur_ep = ep
+            start_offset = i
+            ts_buf = [timestamp_col[i]]
+            fi_buf = [frame_col[i]]
+            cur_task_idx = task_col[i] if task_col is not None else None
+            continue
+        if ep != cur_ep:
+            rec = _build(cur_ep, start_offset, i, cur_task_idx, ts_buf, fi_buf)
+            if rec is not None:
+                yield rec
+            cur_ep = ep
+            start_offset = i
+            ts_buf = [timestamp_col[i]]
+            fi_buf = [frame_col[i]]
+            cur_task_idx = task_col[i] if task_col is not None else None
+        else:
+            ts_buf.append(timestamp_col[i])
+            fi_buf.append(frame_col[i])
+
+    if cur_ep is not None:
+        rec = _build(cur_ep, start_offset, len(episode_col), cur_task_idx, ts_buf, fi_buf)
+        if rec is not None:
+            yield rec
+
+
+def gather_data_paths(root: Path) -> list[Path]:
+    """Return every ``data/chunk-*/file-*.parquet`` path under ``root``."""
+    return sorted((root / "data").rglob("*.parquet"))
+
+
+def episode_offsets_per_path(path: Path) -> dict[int, tuple[int, int]]:
+    """Return ``{episode_index: (row_offset, row_count)}`` for one parquet."""
+    table = pq.read_table(path, columns=["episode_index"])
+    episode_col = table.column("episode_index").to_pylist()
+    out: dict[int, tuple[int, int]] = {}
+    cur_ep: int | None = None
+    start = 0
+    for i, ep in enumerate(episode_col):
+        if cur_ep is None:
+            cur_ep = ep
+            start = i
+            continue
+        if ep != cur_ep:
+            out[cur_ep] = (start, i - start)
+            cur_ep = ep
+            start = i
+    if cur_ep is not None:
+        out[cur_ep] = (start, len(episode_col) - start)
+    return out
+
+
+def keyframe_indices(record: EpisodeRecord, k: int) -> list[int]:
+    """Return ``k`` evenly spaced row indices into the episode (relative)."""
+    n = record.row_count
+    if k <= 0 or n == 0:
+        return []
+    if k >= n:
+        return list(range(n))
+    step = (n - 1) / (k - 1) if k > 1 else 0.0
+    return [int(round(i * step)) for i in range(k)] if k > 1 else [n // 2]
+
+
+def lookup_data_path(root: Path, episode_index: int) -> tuple[Path, int, int] | None:
+    """Find the parquet file containing ``episode_index`` and its slice bounds."""
+    for path in gather_data_paths(root):
+        offsets = episode_offsets_per_path(path)
+        if episode_index in offsets:
+            start, count = offsets[episode_index]
+            return path, start, count
+    return None
+
+
+def episode_frame_timestamps(root: Path, episode_index: int) -> tuple[Any, list[float]]:
+    """Return the parquet path and per-frame timestamps for ``episode_index``."""
+    found = lookup_data_path(root, episode_index)
+    if found is None:
+        raise ValueError(f"Episode {episode_index} not found under {root}/data/")
+    path, start, count = found
+    table = pq.read_table(path, columns=["timestamp"])
+    timestamps = table.column("timestamp").to_pylist()[start : start + count]
+    return path, [float(t) for t in timestamps]

src/lerobot/annotations/steerable_pipeline/staging.py

@@ -0,0 +1,98 @@
+#!/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.
+"""Per-episode staging.
+
+Each module writes its raw output as a JSONL file under
+``<staging_dir>/episode_{ep:06d}/<module>.jsonl``. The writer reads back this
+staging tree and partitions rows into the two language columns.
+
+JSONL is preferred over parquet here because the staging artifact is meant to
+be human-inspectable, easy to diff between prompt iterations, and trivially
+appended to. The final dataset format is parquet; staging is just an
+intermediate.
+"""
+
+from __future__ import annotations
+
+import json
+from collections.abc import Iterable, Iterator
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+ModuleName = str
+
+_MODULES: tuple[ModuleName, ...] = (
+    "module_1",
+    "module_2",
+    "module_3",
+)
+
+
+@dataclass
+class EpisodeStaging:
+    """Filesystem layout for a single episode's staged module outputs."""
+
+    root: Path
+    episode_index: int
+
+    @property
+    def episode_dir(self) -> Path:
+        return self.root / f"episode_{self.episode_index:06d}"
+
+    def path_for(self, module: ModuleName) -> Path:
+        if module not in _MODULES:
+            raise ValueError(f"Unknown module {module!r}; expected one of {_MODULES}")
+        return self.episode_dir / f"{module}.jsonl"
+
+    def write(self, module: ModuleName, rows: Iterable[dict[str, Any]]) -> Path:
+        path = self.path_for(module)
+        path.parent.mkdir(parents=True, exist_ok=True)
+        with path.open("w", encoding="utf-8") as f:
+            for row in rows:
+                f.write(json.dumps(row, ensure_ascii=False, sort_keys=True))
+                f.write("\n")
+        return path
+
+    def read(self, module: ModuleName) -> list[dict[str, Any]]:
+        path = self.path_for(module)
+        if not path.exists():
+            return []
+        out: list[dict[str, Any]] = []
+        with path.open(encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if line:
+                    out.append(json.loads(line))
+        return out
+
+    def read_all(self) -> dict[ModuleName, list[dict[str, Any]]]:
+        return {m: self.read(m) for m in _MODULES}
+
+    def has(self, module: ModuleName) -> bool:
+        return self.path_for(module).exists()
+
+
+def iter_staged_episodes(root: Path) -> Iterator[int]:
+    """Yield episode indices for which any staging artifact exists."""
+    if not root.exists():
+        return
+    for child in sorted(root.iterdir()):
+        if child.is_dir() and child.name.startswith("episode_"):
+            try:
+                yield int(child.name.removeprefix("episode_"))
+            except ValueError:
+                continue

src/lerobot/annotations/steerable_pipeline/validator.py

@@ -0,0 +1,334 @@
+#!/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.
+"""Pre-write validation against staged outputs.
+
+Runs after Modules 1–3 have all written their per-episode artifacts but
+*before* the writer rewrites parquet shards. The validator never touches
+parquet; it only inspects the staging tree and the source frame timestamps
+exposed by :class:`EpisodeRecord`.
+
+Checks (per the plan's "Intermediate staging and validation" section):
+
+- exact timestamp alignment against source frame timestamps
+- no orphan speech / interjection pairs
+- plan / memory emission consistency (events have a paired persistent row)
+- VQA assistant ``content`` is valid JSON (one of bbox / keypoint / count /
+  attribute / spatial)
+- every row maps to its correct column under :func:`column_for_style`
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from collections.abc import Iterable, Sequence
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+from lerobot.datasets.language import (
+    LANGUAGE_EVENTS,
+    LANGUAGE_PERSISTENT,
+    column_for_style,
+    is_view_dependent_style,
+    validate_camera_field,
+)
+
+from .reader import EpisodeRecord
+from .staging import EpisodeStaging
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ValidationReport:
+    """Outcome of one validation pass across all episodes."""
+
+    errors: list[str] = field(default_factory=list)
+    warnings: list[str] = field(default_factory=list)
+    episodes_checked: int = 0
+
+    @property
+    def ok(self) -> bool:
+        return not self.errors
+
+    def add_error(self, message: str) -> None:
+        self.errors.append(message)
+
+    def add_warning(self, message: str) -> None:
+        self.warnings.append(message)
+
+    def summary(self) -> str:
+        return f"checked={self.episodes_checked} errors={len(self.errors)} warnings={len(self.warnings)}"
+
+
+VQA_ANSWER_SHAPES: dict[str, set[str]] = {
+    "bbox": {"detections"},
+    "keypoint": {"label", "point_format", "point"},
+    "count": {"label", "count"},
+    "attribute": {"label", "attribute", "value"},
+    "spatial": {"subject", "relation", "object"},
+}
+
+
+def classify_vqa_answer(payload: Any) -> str | None:
+    """Best-effort classification of a VQA answer payload to a question type."""
+    if not isinstance(payload, dict):
+        return None
+    keys = set(payload.keys())
+    for kind, required in VQA_ANSWER_SHAPES.items():
+        if required.issubset(keys):
+            return kind
+    return None
+
+
+@dataclass
+class StagingValidator:
+    """Walks the staging tree and produces a :class:`ValidationReport`."""
+
+    timestamp_atol: float = 0.0  # exact-match by default
+    dataset_camera_keys: tuple[str, ...] | None = None
+    """Known ``observation.images.*`` keys on the dataset. When set, the
+    validator additionally enforces that every view-dependent row's
+    ``camera`` field references one of these keys. Pass ``None`` (default)
+    to skip that cross-check (e.g. in unit tests with no real dataset)."""
+
+    def validate(
+        self,
+        records: Sequence[EpisodeRecord],
+        staging_dir: Path,
+    ) -> ValidationReport:
+        report = ValidationReport()
+        for record in records:
+            self._validate_episode(record, staging_dir, report)
+            report.episodes_checked += 1
+        return report
+
+    def _validate_episode(
+        self,
+        record: EpisodeRecord,
+        staging_dir: Path,
+        report: ValidationReport,
+    ) -> None:
+        staging = EpisodeStaging(staging_dir, record.episode_index)
+        staged = staging.read_all()
+        all_rows: list[dict[str, Any]] = []
+        for module_name, rows in staged.items():
+            for row in rows:
+                row = {**row, "_module": module_name}
+                all_rows.append(row)
+
+        frame_ts = set(record.frame_timestamps)
+
+        events: list[dict[str, Any]] = []
+        persistent: list[dict[str, Any]] = []
+        for row in all_rows:
+            self._check_column_routing(row, report, record.episode_index)
+            self._check_camera_field(
+                row, report, record.episode_index, self.dataset_camera_keys
+            )
+            if column_for_style(row.get("style")) == LANGUAGE_PERSISTENT:
+                persistent.append(row)
+            else:
+                events.append(row)
+
+        for row in events:
+            self._check_event_timestamp_alignment(row, frame_ts, report, record.episode_index)
+
+        self._check_speech_interjection_pairs(events, report, record.episode_index)
+        self._check_plan_memory_consistency(persistent, events, report, record.episode_index)
+        self._check_vqa_json(events, report, record.episode_index)
+        self._check_vqa_uniqueness_per_frame_camera(events, report, record.episode_index)
+
+    def _check_camera_field(
+        self,
+        row: dict[str, Any],
+        report: ValidationReport,
+        episode_index: int,
+        dataset_camera_keys: Sequence[str] | None,
+    ) -> None:
+        """Enforce the camera invariant + that the key matches the dataset's cameras."""
+        style = row.get("style")
+        camera = row.get("camera")
+        try:
+            validate_camera_field(style, camera)
+        except ValueError as exc:
+            report.add_error(
+                f"ep={episode_index} module={row.get('_module')}: {exc}"
+            )
+            return
+        if (
+            is_view_dependent_style(style)
+            and dataset_camera_keys
+            and camera not in dataset_camera_keys
+        ):
+            report.add_error(
+                f"ep={episode_index} module={row.get('_module')}: camera {camera!r} on style "
+                f"{style!r} is not one of the dataset's video keys {sorted(dataset_camera_keys)!r}"
+            )
+
+    def _check_vqa_uniqueness_per_frame_camera(
+        self,
+        events: Iterable[dict[str, Any]],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        """Ensure at most one (vqa, user) and one (vqa, assistant) per (t, camera)."""
+        counts: dict[tuple[float, str, str], int] = {}
+        for row in events:
+            if row.get("style") != "vqa":
+                continue
+            ts = row.get("timestamp")
+            camera = row.get("camera")
+            role = row.get("role")
+            if ts is None or camera is None or role is None:
+                continue  # other validators flag these
+            key = (float(ts), str(camera), str(role))
+            counts[key] = counts.get(key, 0) + 1
+        for (ts, camera, role), n in counts.items():
+            if n > 1:
+                report.add_error(
+                    f"ep={episode_index}: {n} duplicate vqa rows at t={ts} "
+                    f"camera={camera!r} role={role!r}; expected at most one per (t, camera, role)"
+                )
+
+    def _check_column_routing(
+        self,
+        row: dict[str, Any],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        style = row.get("style")
+        module = row.get("_module")
+        try:
+            target_col = column_for_style(style)
+        except ValueError:
+            report.add_error(f"ep={episode_index} module={module}: unknown style {style!r}")
+            return
+        if module == "module_1" and target_col != LANGUAGE_PERSISTENT:
+            report.add_error(
+                f"ep={episode_index} module=module_1 emitted style {style!r} that routes to {target_col} (must be persistent)"
+            )
+        if module in {"module_2", "module_3"} and target_col != LANGUAGE_EVENTS:
+            report.add_error(
+                f"ep={episode_index} module={module} emitted style {style!r} that routes to {target_col} (must be events)"
+            )
+
+    def _check_event_timestamp_alignment(
+        self,
+        row: dict[str, Any],
+        frame_ts: set[float],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        ts = row.get("timestamp")
+        if ts is None:
+            report.add_error(f"ep={episode_index}: event row missing timestamp: {row!r}")
+            return
+        if self.timestamp_atol == 0.0:
+            if float(ts) not in frame_ts:
+                report.add_error(
+                    f"ep={episode_index}: event row timestamp {ts!r} does not match any source frame timestamp"
+                )
+        else:
+            if not any(abs(float(ts) - f) <= self.timestamp_atol for f in frame_ts):
+                report.add_error(
+                    f"ep={episode_index}: event row timestamp {ts!r} not within {self.timestamp_atol}s of any frame"
+                )
+
+    def _check_speech_interjection_pairs(
+        self,
+        events: Iterable[dict[str, Any]],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        speech_ts: dict[float, int] = {}
+        interjection_ts: dict[float, int] = {}
+        for row in events:
+            ts = row.get("timestamp")
+            if ts is None:
+                continue
+            ts_f = float(ts)
+            if row.get("style") is None and row.get("role") == "assistant":
+                speech_ts[ts_f] = speech_ts.get(ts_f, 0) + 1
+            if row.get("style") == "interjection":
+                interjection_ts[ts_f] = interjection_ts.get(ts_f, 0) + 1
+
+        for ts in interjection_ts:
+            if ts not in speech_ts:
+                report.add_error(f"ep={episode_index}: interjection at t={ts} has no paired speech atom")
+
+    def _check_plan_memory_consistency(
+        self,
+        persistent: Sequence[dict[str, Any]],
+        events: Sequence[dict[str, Any]],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        plan_ts = sorted({float(r["timestamp"]) for r in persistent if r.get("style") == "plan"})
+        memory_ts = sorted({float(r["timestamp"]) for r in persistent if r.get("style") == "memory"})
+        subtask_ts = sorted({float(r["timestamp"]) for r in persistent if r.get("style") == "subtask"})
+        interjection_ts = sorted(
+            {
+                float(r["timestamp"])
+                for r in events
+                if r.get("style") == "interjection" and r.get("timestamp") is not None
+            }
+        )
+
+        if persistent and not plan_ts:
+            report.add_warning(f"ep={episode_index}: persistent rows present but no plan emitted")
+        # every interjection should have a same-timestamp plan refresh
+        for ts in interjection_ts:
+            if ts not in set(plan_ts):
+                report.add_error(
+                    f"ep={episode_index}: interjection at t={ts} has no co-timestamped plan update"
+                )
+        # memory should be emitted at subtask boundaries (subset relation)
+        if memory_ts and subtask_ts:
+            mem_set = set(memory_ts)
+            sub_set = set(subtask_ts)
+            stray = sorted(mem_set - sub_set)
+            if stray:
+                report.add_warning(f"ep={episode_index}: memory rows at {stray} not at any subtask boundary")
+
+    def _check_vqa_json(
+        self,
+        events: Iterable[dict[str, Any]],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        for row in events:
+            if row.get("style") != "vqa" or row.get("role") != "assistant":
+                continue
+            content = row.get("content")
+            if content is None:
+                report.add_error(
+                    f"ep={episode_index}: VQA assistant row at t={row.get('timestamp')} has null content"
+                )
+                continue
+            try:
+                payload = json.loads(content)
+            except (TypeError, ValueError) as exc:
+                report.add_error(
+                    f"ep={episode_index}: VQA assistant content not valid JSON at t={row.get('timestamp')}: {exc}"
+                )
+                continue
+            shape = classify_vqa_answer(payload)
+            if shape is None:
+                report.add_error(
+                    f"ep={episode_index}: VQA assistant payload at t={row.get('timestamp')} does not match any known shape: keys={list(payload) if isinstance(payload, dict) else type(payload).__name__}"
+                )

src/lerobot/annotations/steerable_pipeline/vlm_client.py

@@ -0,0 +1,741 @@
+#!/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.
+"""Shared Qwen-VL client.
+
+The pipeline uses a single shared VLM across modules. vLLM is preferred when
+available (high throughput, JSON-guided decoding); transformers is the
+fallback. A ``stub`` backend is used for unit tests so fixtures never call
+into a real model.
+
+The client speaks one method, :meth:`VlmClient.generate_json`, which:
+
+- accepts a list of OpenAI/HF-style multimodal messages,
+- requests JSON output (``json_mode=True`` enables guided decoding when the
+  backend supports it),
+- batches requests transparently,
+- and reprompts once on a JSON parse failure with an inline correction
+  message before raising.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import threading
+from collections.abc import Callable, Sequence
+from dataclasses import dataclass
+from typing import Any, Protocol
+
+from .config import VlmConfig
+
+
+class VlmClient(Protocol):
+    """Protocol every backend must implement."""
+
+    def generate_json(
+        self,
+        messages_batch: Sequence[Sequence[dict[str, Any]]],
+        *,
+        max_new_tokens: int | None = None,
+        temperature: float | None = None,
+    ) -> list[Any]:
+        """Generate one JSON-decoded response per messages list."""
+
+
+@dataclass
+class StubVlmClient:
+    """Deterministic stub used in unit tests.
+
+    A test passes a callable that maps the *last user message text* (or, if
+    that is empty, the full message list) to a JSON-serializable response.
+    """
+
+    responder: Callable[[Sequence[dict[str, Any]]], Any]
+
+    def generate_json(
+        self,
+        messages_batch: Sequence[Sequence[dict[str, Any]]],
+        *,
+        max_new_tokens: int | None = None,
+        temperature: float | None = None,
+    ) -> list[Any]:
+        return [self.responder(list(messages)) for messages in messages_batch]
+
+
+def _strip_to_json(text: str) -> Any:
+    text = text.strip()
+    # Strip <think>...</think> blocks (Qwen3 Thinking style)
+    while "<think>" in text and "</think>" in text:
+        start = text.find("<think>")
+        end = text.find("</think>", start) + len("</think>")
+        text = (text[:start] + text[end:]).strip()
+    # Strip ```json ... ``` fences from chat-tuned backbones
+    if text.startswith("```"):
+        first = text.find("\n")
+        last = text.rfind("```")
+        if first != -1 and last != -1 and last > first:
+            text = text[first + 1 : last].strip()
+    try:
+        return json.loads(text)
+    except (ValueError, json.JSONDecodeError):
+        pass
+    # Fall back to extracting the first balanced {...} block.
+    obj_text = _extract_first_json_object(text)
+    if obj_text is None:
+        raise json.JSONDecodeError("No JSON object found", text, 0)
+    return json.loads(obj_text)
+
+
+def _extract_first_json_object(text: str) -> str | None:
+    """Return the first balanced ``{...}`` substring, ignoring braces in
+    string literals. Returns ``None`` if no balanced block is found."""
+    start = text.find("{")
+    if start < 0:
+        return None
+    depth = 0
+    in_string = False
+    escape = False
+    for i in range(start, len(text)):
+        ch = text[i]
+        if escape:
+            escape = False
+            continue
+        if ch == "\\":
+            escape = True
+            continue
+        if ch == '"' and not escape:
+            in_string = not in_string
+            continue
+        if in_string:
+            continue
+        if ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return text[start : i + 1]
+    return None
+
+
+@dataclass
+class _GenericTextClient:
+    """Wraps any text-generation callable in JSON-mode + one-retry semantics."""
+
+    generate_text: Callable[[Sequence[Sequence[dict[str, Any]]], int, float], list[str]]
+    config: VlmConfig
+
+    def generate_json(
+        self,
+        messages_batch: Sequence[Sequence[dict[str, Any]]],
+        *,
+        max_new_tokens: int | None = None,
+        temperature: float | None = None,
+    ) -> list[Any]:
+        max_tok = max_new_tokens if max_new_tokens is not None else self.config.max_new_tokens
+        temp = temperature if temperature is not None else self.config.temperature
+        raw = self.generate_text(messages_batch, max_tok, temp)
+        out: list[Any] = []
+        for messages, text in zip(messages_batch, raw, strict=True):
+            try:
+                out.append(_strip_to_json(text))
+                continue
+            except (ValueError, json.JSONDecodeError):
+                pass
+            retry = list(messages) + [
+                {"role": "assistant", "content": text},
+                {
+                    "role": "user",
+                    "content": (
+                        "Your previous reply was not valid JSON. "
+                        "Reply with strictly valid JSON, no prose, no fences."
+                    ),
+                },
+            ]
+            retry_text = self.generate_text([retry], max_tok, temp)[0]
+            try:
+                out.append(_strip_to_json(retry_text))
+            except (ValueError, json.JSONDecodeError):
+                # After retry: log preview and return None instead of crashing
+                # the whole pipeline. Modules treat None as "skip".
+                preview = retry_text.strip().replace("\n", " ")[:200]
+                print(
+                    f"[vlm] WARNING: failed to parse JSON after retry; preview: {preview!r}",
+                    flush=True,
+                )
+                out.append(None)
+        return out
+
+
+def make_vlm_client(config: VlmConfig) -> VlmClient:
+    """Build the shared VLM client per the configured backend.
+
+    For ``stub``, callers should construct :class:`StubVlmClient` directly with
+    a responder callable. ``stub`` here is rejected to make accidental misuse
+    obvious.
+    """
+    if config.backend == "stub":
+        raise ValueError(
+            "Use StubVlmClient(...) directly for the stub backend; make_vlm_client builds real clients."
+        )
+    if config.backend == "vllm":
+        return _make_vllm_client(config)
+    if config.backend == "transformers":
+        return _make_transformers_client(config)
+    if config.backend == "openai":
+        return _make_openai_client(config)
+    raise ValueError(f"Unknown VLM backend: {config.backend!r}")
+
+
+def _make_vllm_client(config: VlmConfig) -> VlmClient:
+    try:
+        from vllm import LLM, SamplingParams  # type: ignore[import-not-found]
+    except ImportError as exc:
+        raise ImportError(
+            "vllm is required for backend='vllm'. Install with `pip install lerobot[annotations]`."
+        ) from exc
+    # Workaround for cuDNN 9.x + torch 2.8 conv3d regression that surfaces
+    # as CUDNN_STATUS_NOT_INITIALIZED in Qwen-VL vision-tower patch
+    # embedders. Setting LEROBOT_DISABLE_CUDNN=1 forces native PyTorch
+    # convolution kernels — slower but functional.
+    import os as _os  # noqa: PLC0415
+
+    if _os.environ.get("LEROBOT_DISABLE_CUDNN", "").lower() in {"1", "true", "yes"}:
+        import torch as _torch  # noqa: PLC0415
+
+        _torch.backends.cudnn.enabled = False
+    llm_kwargs: dict[str, Any] = {
+        "model": config.model_id,
+        "tensor_parallel_size": config.tensor_parallel_size,
+        "gpu_memory_utilization": config.gpu_memory_utilization,
+        "trust_remote_code": config.trust_remote_code,
+    }
+    if config.max_model_len is not None:
+        llm_kwargs["max_model_len"] = config.max_model_len
+    llm = LLM(**llm_kwargs)
+
+    def _gen(batch: Sequence[Sequence[dict[str, Any]]], max_tok: int, temp: float) -> list[str]:
+        # ``guided_decoding`` would speed up parsing but its API differs across
+        # vllm releases (dict vs GuidedDecodingParams). The _GenericTextClient
+        # wrapper already has a one-retry JSON-recovery path, so we skip it.
+        params = SamplingParams(max_tokens=max_tok, temperature=temp)
+        # ``llm.chat`` handles chat-template application + multimodal input
+        # extraction (image/video blocks) internally, which ``llm.generate``
+        # does not.
+        outputs = llm.chat([list(m) for m in batch], params)
+        return [o.outputs[0].text for o in outputs]
+
+    return _GenericTextClient(_gen, config)
+
+
+def _make_transformers_client(config: VlmConfig) -> VlmClient:
+    try:
+        import torch  # type: ignore[import-not-found]
+        import transformers  # type: ignore[import-not-found]
+        from transformers import AutoProcessor  # type: ignore[import-not-found]
+    except ImportError as exc:
+        raise ImportError("transformers + torch are required for backend='transformers'.") from exc
+    auto_cls = (
+        getattr(transformers, "AutoModelForImageTextToText", None)
+        or getattr(transformers, "AutoModelForVision2Seq", None)
+    )
+    if auto_cls is None:
+        raise ImportError(
+            "Neither AutoModelForImageTextToText nor AutoModelForVision2Seq is available in this "
+            "transformers version. Install transformers>=4.45 (which has AutoModelForImageTextToText) "
+            "for VL models."
+        )
+    processor = AutoProcessor.from_pretrained(
+        config.model_id, trust_remote_code=config.trust_remote_code
+    )
+    import os as _os  # noqa: PLC0415
+
+    use_accelerate = _os.environ.get("LEROBOT_TRANSFORMERS_DEVICE_MAP", "manual") != "manual"
+    # ``device_map='auto'`` triggers a known std::bad_alloc on the Qwen3-VL
+    # post-load dispatch path (the alloc fails in accelerate's hook setup
+    # even with TBs of host RAM). Default to manual: load on CPU with
+    # ``low_cpu_mem_usage=True``, then ``.to("cuda")``. Set
+    # ``LEROBOT_TRANSFORMERS_DEVICE_MAP=auto`` to opt back into the old path.
+    if use_accelerate:
+        model = auto_cls.from_pretrained(
+            config.model_id,
+            torch_dtype="auto",
+            device_map="auto",
+            low_cpu_mem_usage=True,
+            trust_remote_code=config.trust_remote_code,
+        )
+    else:
+        import torch as _torch  # noqa: PLC0415
+
+        model = auto_cls.from_pretrained(
+            config.model_id,
+            torch_dtype=_torch.bfloat16,
+            low_cpu_mem_usage=True,
+            trust_remote_code=config.trust_remote_code,
+        )
+        model = model.to("cuda")
+    model.eval()
+
+    def _gen(batch: Sequence[Sequence[dict[str, Any]]], max_tok: int, temp: float) -> list[str]:
+        outs: list[str] = []
+        for messages in batch:
+            text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
+            inputs = processor(text=[text], return_tensors="pt").to(model.device)
+            with torch.no_grad():
+                gen = model.generate(
+                    **inputs,
+                    max_new_tokens=max_tok,
+                    temperature=temp,
+                    do_sample=temp > 0.0,
+                )
+            decoded = processor.batch_decode(
+                gen[:, inputs["input_ids"].shape[-1] :], skip_special_tokens=True
+            )[0]
+            outs.append(decoded)
+        return outs
+
+    return _GenericTextClient(_gen, config)
+
+
+def _make_openai_client(config: VlmConfig) -> VlmClient:
+    """Backend that talks to any OpenAI-compatible server.
+
+    Compatible with ``vllm serve``, ``transformers serve``,
+    ``ktransformers serve``, and hosted endpoints. By default the server
+    is expected to be already running. Set ``auto_serve=True`` to have
+    this client spawn one (default: ``transformers serve``), wait until
+    it's ready, and tear it down on process exit.
+
+    Image blocks ``{"type":"image", "image":<PIL.Image>}`` are
+    auto-converted to ``image_url`` data-URLs. Video blocks
+    ``{"type":"video", "video":[<PIL>...]}`` are forwarded as
+    multi-frame ``video_url`` items where supported.
+    """
+    try:
+        from openai import OpenAI  # type: ignore[import-not-found]
+    except ImportError as exc:
+        raise ImportError(
+            "openai package is required for backend='openai'. "
+            "Install with `pip install openai`."
+        ) from exc
+
+    api_base = config.api_base
+    api_key = config.api_key
+    auto_serve = config.auto_serve
+    api_bases: list[str] = [api_base]
+
+    print(
+        f"[lerobot-annotate] backend=openai model={config.model_id} "
+        f"api_base={api_base} auto_serve={auto_serve}",
+        flush=True,
+    )
+    if auto_serve:
+        if config.parallel_servers > 1:
+            print(
+                f"[lerobot-annotate] spawning {config.parallel_servers} parallel servers",
+                flush=True,
+            )
+            api_bases = _spawn_parallel_inference_servers(config)
+        elif _server_is_up(api_base):
+            print(f"[lerobot-annotate] reusing server already up at {api_base}", flush=True)
+        else:
+            print("[lerobot-annotate] no server reachable; spawning one", flush=True)
+            api_base = _spawn_inference_server(config)
+            api_bases = [api_base]
+            print(f"[lerobot-annotate] server ready at {api_base}", flush=True)
+
+    clients = [OpenAI(base_url=base, api_key=api_key) for base in api_bases]
+    client = clients[0]
+    # round-robin counter for parallel mode
+    rr_counter = {"i": 0}
+
+    # ``mm_processor_kwargs`` is a vllm-specific extra; transformers serve
+    # rejects it with HTTP 422. Send it only when explicitly opted in via
+    # an env var (e.g. ``LEROBOT_OPENAI_SEND_MM_KWARGS=1`` for vllm).
+    send_mm_kwargs = os.environ.get(
+        "LEROBOT_OPENAI_SEND_MM_KWARGS", ""
+    ).lower() in {"1", "true", "yes"}
+
+    rr_lock = threading.Lock()
+
+    def _one_call(
+        messages: Sequence[dict[str, Any]], max_tok: int, temp: float
+    ) -> str:
+        api_messages, mm_kwargs = _to_openai_messages(messages)
+        kwargs: dict[str, Any] = {
+            "model": config.model_id,
+            "messages": api_messages,
+            "max_tokens": max_tok,
+            "temperature": temp,
+        }
+        extra_body: dict[str, Any] = {}
+        if send_mm_kwargs and mm_kwargs:
+            extra_body["mm_processor_kwargs"] = {**mm_kwargs, "do_sample_frames": True}
+        if config.chat_template_kwargs:
+            extra_body["chat_template_kwargs"] = config.chat_template_kwargs
+        if extra_body:
+            kwargs["extra_body"] = extra_body
+        with rr_lock:
+            chosen = clients[rr_counter["i"] % len(clients)]
+            rr_counter["i"] += 1
+        response = chosen.chat.completions.create(**kwargs)
+        return response.choices[0].message.content or ""
+
+    def _gen(
+        batch: Sequence[Sequence[dict[str, Any]]], max_tok: int, temp: float
+    ) -> list[str]:
+        if len(batch) <= 1 or config.client_concurrency <= 1:
+            return [_one_call(messages, max_tok, temp) for messages in batch]
+        # Parallel fan-out — vllm batches these on the server side.
+        from concurrent.futures import ThreadPoolExecutor  # noqa: PLC0415
+
+        max_workers = min(config.client_concurrency, len(batch))
+        with ThreadPoolExecutor(max_workers=max_workers) as pool:
+            futures = [
+                pool.submit(_one_call, messages, max_tok, temp) for messages in batch
+            ]
+            return [f.result() for f in futures]
+
+    return _GenericTextClient(_gen, config)
+
+
+def _spawn_parallel_inference_servers(config: VlmConfig) -> list[str]:
+    """Spawn ``config.parallel_servers`` independent vllm replicas.
+
+    Each replica:
+    - is pinned to a single GPU via ``CUDA_VISIBLE_DEVICES``
+    - listens on ``serve_port + i``
+    - is shut down via the same atexit hook as the single-server path
+
+    Returns the list of ``api_base`` URLs the client should round-robin
+    across.
+    """
+    import atexit  # noqa: PLC0415
+    import os as _os  # noqa: PLC0415
+    import shlex  # noqa: PLC0415
+    import signal  # noqa: PLC0415
+    import subprocess  # noqa: PLC0415
+    import sys  # noqa: PLC0415
+    import threading  # noqa: PLC0415
+    import time  # noqa: PLC0415
+
+    n = config.parallel_servers
+    api_bases: list[str] = []
+    procs: list[subprocess.Popen] = []
+    ready_events: list[threading.Event] = []
+    # Multiple readiness signals — uvicorn's own banner is suppressed at
+    # ``--uvicorn-log-level warning``, so we also accept vllm's own
+    # "Starting vLLM API server" line and the route-listing line. The
+    # HTTP probe below is the ultimate fallback.
+    ready_markers = (
+        "Uvicorn running",
+        "Application startup complete",
+        "Starting vLLM API server",
+        "Available routes are",
+    )
+    # Single lock for all server-stream threads so multibyte chars from
+    # different servers don't interleave and tear UTF-8 sequences.
+    print_lock = threading.Lock()
+
+    base_cmd = config.serve_command or (
+        f"vllm serve {shlex.quote(config.model_id)} "
+        f"--tensor-parallel-size 1 "
+        f"--max-model-len {config.max_model_len or 32768} "
+        f"--uvicorn-log-level warning"
+    )
+
+    num_gpus = config.num_gpus if config.num_gpus > 0 else n
+    for i in range(n):
+        port = config.serve_port + i
+        gpu = i % num_gpus
+        env = _os.environ.copy()
+        env["CUDA_VISIBLE_DEVICES"] = str(gpu)
+        cmd = base_cmd
+        if "{port}" in cmd:
+            cmd = cmd.replace("{port}", str(port))
+        else:
+            cmd = f"{cmd} --port {port}"
+        api_base = f"http://localhost:{port}/v1"
+        api_bases.append(api_base)
+        print(f"[server-{i}] launching on GPU {gpu} port {port}: {cmd}", flush=True)
+        proc = subprocess.Popen(
+            shlex.split(cmd),
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+            text=True,
+            bufsize=1,
+            env=env,
+        )
+        procs.append(proc)
+        ready = threading.Event()
+        ready_events.append(ready)
+
+        def _stream(idx: int, p: subprocess.Popen, ev: threading.Event) -> None:
+            # Read whole lines and emit each line atomically under the
+            # shared print_lock so output from N servers stays readable.
+            assert p.stdout is not None
+            for line in iter(p.stdout.readline, ""):
+                with print_lock:
+                    sys.stdout.write(f"[server-{idx}] {line}")
+                    if not line.endswith(("\n", "\r")):
+                        sys.stdout.write("\n")
+                    sys.stdout.flush()
+                if any(m in line for m in ready_markers):
+                    ev.set()
+
+        threading.Thread(target=_stream, args=(i, proc, ready), daemon=True).start()
+
+        def _probe(idx: int, base: str, ev: threading.Event, p: subprocess.Popen) -> None:
+            while not ev.is_set() and p.poll() is None:
+                if _server_is_up(base):
+                    print(f"[server-{idx}] ready (http probe)", flush=True)
+                    ev.set()
+                    return
+                time.sleep(2)
+
+        threading.Thread(target=_probe, args=(i, api_base, ready, proc), daemon=True).start()
+
+    def _shutdown() -> None:
+        for i, p in enumerate(procs):
+            if p.poll() is None:
+                print(f"[server-{i}] stopping pid={p.pid}", flush=True)
+                p.send_signal(signal.SIGINT)
+        for p in procs:
+            try:
+                p.wait(timeout=15)
+            except subprocess.TimeoutExpired:
+                p.kill()
+                p.wait(timeout=5)
+
+    atexit.register(_shutdown)
+
+    deadline = time.monotonic() + config.serve_ready_timeout_s
+    while any(not ev.is_set() for ev in ready_events) and time.monotonic() < deadline:
+        for i, p in enumerate(procs):
+            if p.poll() is not None:
+                raise RuntimeError(
+                    f"[server-{i}] inference server exited unexpectedly with rc={p.returncode}"
+                )
+        time.sleep(2)
+    if any(not ev.is_set() for ev in ready_events):
+        raise RuntimeError(
+            f"[server] not all replicas became ready within {config.serve_ready_timeout_s}s"
+        )
+    print(f"[lerobot-annotate] all {n} servers ready: {api_bases}", flush=True)
+    return api_bases
+
+
+def _server_is_up(api_base: str) -> bool:
+    """Return True if ``api_base/models`` answers 200 within 2 seconds."""
+    import urllib.request  # noqa: PLC0415
+
+    url = api_base.rstrip("/") + "/models"
+    try:
+        with urllib.request.urlopen(url, timeout=2) as resp:
+            return resp.status == 200
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _spawn_inference_server(config: VlmConfig) -> str:
+    """Spawn ``transformers serve`` (or ``serve_command``), wait until it
+    accepts ``/v1/models``, and register a shutdown hook.
+
+    Streams the server's stdout/stderr to the parent terminal in
+    real-time on a background thread so users can see model-load
+    progress and errors as they happen.
+
+    Returns the full ``api_base`` URL the OpenAI client should use.
+    """
+    import atexit  # noqa: PLC0415
+    import shlex  # noqa: PLC0415
+    import signal  # noqa: PLC0415
+    import subprocess  # noqa: PLC0415
+    import sys  # noqa: PLC0415
+    import threading  # noqa: PLC0415
+    import time  # noqa: PLC0415
+    import urllib.request  # noqa: PLC0415
+
+    cmd = config.serve_command
+    if not cmd:
+        cmd = (
+            f"transformers serve {shlex.quote(config.model_id)} "
+            f"--port {config.serve_port} --continuous-batching"
+        )
+    api_base = f"http://localhost:{config.serve_port}/v1"
+    print(f"[server] launching: {cmd}", flush=True)
+    proc = subprocess.Popen(
+        shlex.split(cmd),
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True,
+        bufsize=1,
+    )
+
+    # Watch the server output for the uvicorn readiness banner. This is
+    # more reliable than polling /v1/models because transformers serve
+    # rescans its cache on every model-list request, which can exceed
+    # the urllib timeout and trigger an infinite probe loop.
+    ready_event = threading.Event()
+    # See _spawn_parallel_inference_servers for why we accept these.
+    ready_markers = (
+        "Uvicorn running",
+        "Application startup complete",
+        "Starting vLLM API server",
+        "Available routes are",
+    )
+
+    def _probe() -> None:
+        while not ready_event.is_set() and proc.poll() is None:
+            if _server_is_up(api_base):
+                print("[server] ready (http probe)", flush=True)
+                ready_event.set()
+                return
+            time.sleep(2)
+
+    threading.Thread(target=_probe, daemon=True).start()
+
+    def _stream_output() -> None:
+        # Read raw chunks instead of iterating lines so tqdm progress
+        # bars (which overwrite using \r) flush in real time.
+        assert proc.stdout is not None
+        buf = ""
+        prefix_started = False
+        while True:
+            ch = proc.stdout.read(1)
+            if ch == "":
+                # process exited; flush any tail
+                if buf:
+                    sys.stdout.write(buf)
+                    sys.stdout.flush()
+                return
+            if not prefix_started:
+                sys.stdout.write("[server] ")
+                prefix_started = True
+            sys.stdout.write(ch)
+            sys.stdout.flush()
+            buf += ch
+            if ch in ("\n", "\r"):
+                if any(marker in buf for marker in ready_markers):
+                    ready_event.set()
+                buf = ""
+                prefix_started = False
+
+    threading.Thread(target=_stream_output, daemon=True).start()
+
+    def _shutdown() -> None:
+        if proc.poll() is None:
+            print(f"[server] stopping pid={proc.pid}", flush=True)
+            proc.send_signal(signal.SIGINT)
+            try:
+                proc.wait(timeout=15)
+            except subprocess.TimeoutExpired:
+                proc.kill()
+                proc.wait(timeout=5)
+
+    atexit.register(_shutdown)
+
+    deadline = time.monotonic() + config.serve_ready_timeout_s
+    while time.monotonic() < deadline:
+        if proc.poll() is not None:
+            raise RuntimeError(
+                f"[server] inference server exited unexpectedly with rc={proc.returncode}. "
+                f"See [server] log lines above for the cause."
+            )
+        if ready_event.wait(timeout=2):
+            return api_base
+    proc.terminate()
+    raise RuntimeError(
+        f"[server] did not become ready within {config.serve_ready_timeout_s}s"
+    )
+
+
+def _to_openai_messages(
+    messages: Sequence[dict[str, Any]],
+) -> tuple[list[dict[str, Any]], dict[str, Any]]:
+    """Convert internal messages to OpenAI chat format.
+
+    Returns ``(api_messages, mm_kwargs)``. Multimodal-processor kwargs
+    (``fps`` from ``video_url`` blocks) are extracted out so the caller
+    can pass them via ``extra_body.mm_processor_kwargs`` rather than
+    inside the content blocks (which transformers serve rejects).
+
+    File-URL video blocks are inlined as base64 data URLs.
+    """
+    out_messages: list[dict[str, Any]] = []
+    mm_kwargs: dict[str, Any] = {}
+    for message in messages:
+        content = message.get("content")
+        if not isinstance(content, list):
+            out_messages.append({"role": message["role"], "content": content})
+            continue
+        out_blocks: list[dict[str, Any]] = []
+        for block in content:
+            block_type = block.get("type") if isinstance(block, dict) else None
+            if block_type == "text":
+                out_blocks.append({"type": "text", "text": block.get("text", "")})
+            elif block_type == "image":
+                out_blocks.append(
+                    {"type": "image_url", "image_url": {"url": _pil_to_data_url(block["image"])}}
+                )
+            elif block_type == "video":
+                frames = block.get("video", [])
+                for img in frames:
+                    out_blocks.append(
+                        {"type": "image_url", "image_url": {"url": _pil_to_data_url(img)}}
+                    )
+            elif block_type == "video_url":
+                video_url = dict(block["video_url"])
+                url = video_url.get("url", "")
+                if url.startswith("file://"):
+                    video_url["url"] = _file_to_data_url(url[len("file://") :])
+                out_blocks.append({"type": "video_url", "video_url": video_url})
+                fps = block.get("fps")
+                if fps is not None:
+                    mm_kwargs["fps"] = fps
+            else:
+                out_blocks.append(block)
+        out_messages.append({"role": message["role"], "content": out_blocks})
+    return out_messages, mm_kwargs
+
+
+def _file_to_data_url(path: str) -> str:
+    """Read a local video file and return a base64 ``data:video/mp4`` URL."""
+    import base64  # noqa: PLC0415
+
+    with open(path, "rb") as f:
+        b64 = base64.b64encode(f.read()).decode("ascii")
+    return f"data:video/mp4;base64,{b64}"
+
+
+def _pil_to_data_url(image: Any) -> str:
+    """Encode a PIL.Image as a base64 data URL."""
+    import base64  # noqa: PLC0415
+    import io  # noqa: PLC0415
+
+    buf = io.BytesIO()
+    image.save(buf, format="PNG")
+    b64 = base64.b64encode(buf.getvalue()).decode("ascii")
+    return f"data:image/png;base64,{b64}"
+
+
+def _messages_to_prompt(messages: Sequence[dict[str, Any]]) -> Any:
+    """Pass-through hook used by the vllm backend.
+
+    vllm exposes its own multimodal entry points that vary by version; for the
+    base flow we simply forward the raw message list and let the caller's
+    custom backend handle templating. Real deployments override this.
+    """
+    return list(messages)

src/lerobot/annotations/steerable_pipeline/writer.py

@@ -0,0 +1,341 @@
+#!/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.
+"""Final parquet rewrite.
+
+For every episode the writer:
+
+1. reads the staged module outputs,
+2. partitions them into a persistent slice (PERSISTENT_STYLES) and an event
+   slice (EVENT_ONLY_STYLES + style=None tool-call atoms),
+3. sorts each slice deterministically,
+4. broadcasts the persistent slice across every frame in the episode,
+5. for each frame, materializes the sublist of event rows whose timestamp
+   exactly equals that frame's timestamp,
+6. drops the legacy ``subtask_index`` column,
+7. writes the parquet shard back in place.
+
+The writer does NOT add a dataset-level ``tools`` column. Tool *calls* are
+emitted per-row via the existing ``tool_calls`` field on the v3.1 row
+struct (PR 1) for every speech atom. The tool *schema* (the description
+of the ``say`` function and its parameters) is a fixed code constant —
+``SAY_TOOL_SCHEMA`` below — and downstream chat-template consumers import
+it directly rather than reading a redundant per-row column.
+
+Invariants enforced here (and re-checked by the validator):
+
+- per-episode persistent slice is byte-identical across every frame;
+- ``language_events`` rows on a frame all have ``timestamp == frame_ts``
+  (timestamps come straight from the source parquet — never recomputed);
+- every row passes ``column_for_style(style)``.
+"""
+
+from __future__ import annotations
+
+import logging
+from collections import defaultdict
+from collections.abc import Iterable, Sequence
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+from lerobot.datasets.language import (
+    EVENT_ONLY_STYLES,
+    LANGUAGE_EVENTS,
+    LANGUAGE_PERSISTENT,
+    PERSISTENT_STYLES,
+    column_for_style,
+    validate_camera_field,
+)
+
+from .reader import EpisodeRecord
+from .staging import EpisodeStaging
+
+logger = logging.getLogger(__name__)
+
+
+# Tool schema constants moved to lerobot.datasets.language in PR 1 — single
+# source of truth. Re-exported here so existing imports
+# (``from lerobot.annotations.steerable_pipeline.writer import SAY_TOOL_SCHEMA``)
+# keep working.
+from lerobot.datasets.language import DEFAULT_TOOLS, SAY_TOOL_SCHEMA  # noqa: F401, E402
+
+
+def _row_persistent_sort_key(row: dict[str, Any]) -> tuple:
+    return (float(row["timestamp"]), row.get("style") or "", row.get("role") or "")
+
+
+def _row_event_sort_key(row: dict[str, Any]) -> tuple:
+    # events are bucketed per-frame, but within a frame we still want determinism
+    return (
+        row.get("style") or "",
+        row.get("role") or "",
+        row.get("camera") or "",
+    )
+
+
+def _normalize_persistent_row(row: dict[str, Any]) -> dict[str, Any]:
+    """Coerce a staged row into the persistent column's struct shape."""
+    style = row.get("style")
+    if style not in PERSISTENT_STYLES:
+        raise ValueError(
+            f"persistent slice contains row with non-persistent style {style!r}; "
+            "row would be misrouted under column_for_style()"
+        )
+    if "timestamp" not in row:
+        raise ValueError(f"persistent row missing timestamp: {row!r}")
+    camera = row.get("camera")
+    validate_camera_field(style, camera)
+    return {
+        "role": str(row["role"]),
+        "content": None if row.get("content") is None else str(row["content"]),
+        "style": style,
+        "timestamp": float(row["timestamp"]),
+        "camera": None if camera is None else str(camera),
+        "tool_calls": _normalize_tool_calls(row.get("tool_calls")),
+    }
+
+
+def _normalize_event_row(row: dict[str, Any]) -> dict[str, Any]:
+    """Coerce a staged row into the event column's struct shape (no timestamp)."""
+    style = row.get("style")
+    if style is not None and style not in EVENT_ONLY_STYLES:
+        raise ValueError(
+            f"event slice contains row with style {style!r}; expected None or one of {EVENT_ONLY_STYLES}"
+        )
+    if column_for_style(style) != LANGUAGE_EVENTS:
+        raise ValueError(f"event row with style {style!r} would not route to language_events")
+    camera = row.get("camera")
+    validate_camera_field(style, camera)
+    return {
+        "role": str(row["role"]),
+        "content": None if row.get("content") is None else str(row["content"]),
+        "style": style,
+        "camera": None if camera is None else str(camera),
+        "tool_calls": _normalize_tool_calls(row.get("tool_calls")),
+    }
+
+
+def _normalize_tool_calls(value: Any) -> list[Any] | None:
+    if value is None:
+        return None
+    if not isinstance(value, list):
+        raise ValueError(f"tool_calls must be a list or None, got {type(value).__name__}")
+    return list(value)
+
+
+def _validate_atom_invariants(row: dict[str, Any]) -> None:
+    """At-least-one of content/tool_calls; style=None implies tool_calls."""
+    has_content = row.get("content") is not None
+    has_tools = row.get("tool_calls") is not None
+    if not (has_content or has_tools):
+        raise ValueError(f"row has neither content nor tool_calls: {row!r}")
+    if row.get("style") is None and not has_tools:
+        raise ValueError(f"style=None requires tool_calls: {row!r}")
+
+
+def _validate_speech_atom(row: dict[str, Any]) -> None:
+    """Speech atoms: role=assistant, style=None, content=None, say tool call."""
+    if row.get("style") is not None:
+        return  # not a speech atom
+    if row.get("role") != "assistant":
+        raise ValueError(f"speech atom must have role=assistant: {row!r}")
+    if row.get("content") is not None:
+        raise ValueError(f"speech atom must have content=null: {row!r}")
+    tool_calls = row.get("tool_calls")
+    if not tool_calls or not isinstance(tool_calls, list):
+        raise ValueError(f"speech atom must have non-empty tool_calls list: {row!r}")
+    first = tool_calls[0]
+    if not isinstance(first, dict):
+        raise ValueError(f"speech atom tool_calls[0] must be a dict: {row!r}")
+    if first.get("type") != "function":
+        raise ValueError(f"speech atom tool_calls[0].type must be 'function': {row!r}")
+    fn = first.get("function") or {}
+    if fn.get("name") != "say":
+        raise ValueError(f"speech atom tool_calls[0].function.name must be 'say': {row!r}")
+    args = fn.get("arguments") or {}
+    if not isinstance(args, dict) or "text" not in args or not isinstance(args["text"], str):
+        raise ValueError(f"speech atom must carry 'text' string in arguments: {row!r}")
+
+
+@dataclass
+class LanguageColumnsWriter:
+    """Rewrite ``data/chunk-*/file-*.parquet`` with the two language columns."""
+
+    drop_existing_subtask_index: bool = True
+
+    def write_all(
+        self,
+        records: Sequence[EpisodeRecord],
+        staging_dir: Path,
+        root: Path,
+    ) -> list[Path]:
+        episodes_by_path: dict[Path, list[EpisodeRecord]] = defaultdict(list)
+        for record in records:
+            episodes_by_path[record.data_path].append(record)
+
+        written: list[Path] = []
+        for path, eps in episodes_by_path.items():
+            self._rewrite_one(path, eps, staging_dir, root)
+            written.append(path)
+        return written
+
+    def _rewrite_one(
+        self,
+        path: Path,
+        episodes: Sequence[EpisodeRecord],
+        staging_dir: Path,
+        root: Path,
+    ) -> None:
+        table = pq.read_table(path)
+        n_rows = table.num_rows
+
+        # Ensure we cover every episode in the file. Episodes that don't have
+        # staging artifacts are passed through with empty annotation lists —
+        # this keeps the writer idempotent and safe for partial reruns.
+        staged_per_ep: dict[int, dict[str, list[dict[str, Any]]]] = {}
+        for record in episodes:
+            staging = EpisodeStaging(staging_dir, record.episode_index)
+            staged_per_ep[record.episode_index] = staging.read_all()
+
+        persistent_by_ep: dict[int, list[dict[str, Any]]] = {}
+        events_by_ep_ts: dict[int, dict[float, list[dict[str, Any]]]] = {}
+
+        for ep_index, ep_staged in staged_per_ep.items():
+            persistent_rows: list[dict[str, Any]] = []
+            event_rows: list[dict[str, Any]] = []  # carry timestamp until bucketed
+            for _module_name, rows in ep_staged.items():
+                for row in rows:
+                    style = row.get("style")
+                    if column_for_style(style) == LANGUAGE_PERSISTENT:
+                        persistent_rows.append(row)
+                    else:
+                        event_rows.append(row)
+
+            persistent_rows.sort(key=_row_persistent_sort_key)
+            normalized_persistent = []
+            for r in persistent_rows:
+                _validate_atom_invariants(r)
+                _validate_speech_atom(r)
+                normalized_persistent.append(_normalize_persistent_row(r))
+            persistent_by_ep[ep_index] = normalized_persistent
+
+            buckets: dict[float, list[dict[str, Any]]] = defaultdict(list)
+            for r in event_rows:
+                _validate_atom_invariants(r)
+                _validate_speech_atom(r)
+                ts = float(r["timestamp"])
+                buckets[ts].append(_normalize_event_row(r))
+            for ts in list(buckets.keys()):
+                buckets[ts].sort(key=_row_event_sort_key)
+            events_by_ep_ts[ep_index] = buckets
+
+        episode_col = (
+            table.column("episode_index").to_pylist() if "episode_index" in table.column_names else None
+        )
+        ts_col = table.column("timestamp").to_pylist() if "timestamp" in table.column_names else None
+        if episode_col is None or ts_col is None:
+            raise ValueError(f"{path} is missing 'episode_index' or 'timestamp' — required by the writer.")
+
+        per_row_persistent: list[list[dict[str, Any]]] = []
+        per_row_events: list[list[dict[str, Any]]] = []
+        for i in range(n_rows):
+            ep = episode_col[i]
+            ts = float(ts_col[i])
+            per_row_persistent.append(persistent_by_ep.get(ep, []))
+            buckets = events_by_ep_ts.get(ep, {})
+            per_row_events.append(buckets.get(ts, []))
+
+        new_table = self._materialize_table(
+            table, per_row_persistent, per_row_events, drop_old=self.drop_existing_subtask_index
+        )
+        pq.write_table(new_table, path)
+
+    def _materialize_table(
+        self,
+        table: pa.Table,
+        persistent: list[list[dict[str, Any]]],
+        events: list[list[dict[str, Any]]],
+        *,
+        drop_old: bool,
+    ) -> pa.Table:
+        cols = []
+        names = []
+        for name in table.column_names:
+            if drop_old and name == "subtask_index":
+                continue
+            if name in (LANGUAGE_PERSISTENT, LANGUAGE_EVENTS):
+                continue  # we'll re-add canonical versions
+            # Strip any legacy ``tools`` column previously emitted by older
+            # writers — the schema no longer uses it (constant lives in
+            # SAY_TOOL_SCHEMA / DEFAULT_TOOLS).
+            if name == "tools":
+                continue
+            cols.append(table.column(name))
+            names.append(name)
+
+        # We let pyarrow infer struct/list schema rather than passing the
+        # canonical type from `lerobot.datasets.language` directly: that type
+        # uses `pa.json_()` for the `tool_calls` element type, which
+        # `pa.array(..., type=...)` cannot materialize from Python lists on
+        # current pyarrow versions. The inferred schema round-trips through
+        # parquet and `LeRobotDataset` correctly — see PR 1's
+        # `tests/datasets/test_language.py` which exercises the same flow.
+        persistent_arr = pa.array(persistent)
+        events_arr = pa.array(events)
+
+        cols.extend([persistent_arr, events_arr])
+        names.extend([LANGUAGE_PERSISTENT, LANGUAGE_EVENTS])
+
+        return pa.Table.from_arrays(cols, names=names)
+
+
+def speech_atom(timestamp: float, text: str) -> dict[str, Any]:
+    """Build a canonical speech tool-call atom for the events column."""
+    return {
+        "role": "assistant",
+        "content": None,
+        "style": None,
+        "timestamp": float(timestamp),
+        "camera": None,
+        "tool_calls": [
+            {
+                "type": "function",
+                "function": {
+                    "name": "say",
+                    "arguments": {"text": text},
+                },
+            }
+        ],
+    }
+
+
+def normalize_rows_for_writer(
+    rows: Iterable[dict[str, Any]],
+) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
+    """Helper used by tests/validators to partition a flat row list into
+    (persistent_rows, event_rows) using ``column_for_style``.
+    """
+    persistent: list[dict[str, Any]] = []
+    events: list[dict[str, Any]] = []
+    for row in rows:
+        if column_for_style(row.get("style")) == LANGUAGE_PERSISTENT:
+            persistent.append(row)
+        else:
+            events.append(row)
+    return persistent, events

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

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

src/lerobot/cameras/realsense/camera_realsense.py

@@ -19,16 +19,18 @@ Provides the RealSenseCamera class for capturing frames from Intel RealSense cam
 import logging
 import time
 from threading import Event, Lock, Thread
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
 from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
-try:
-    import pyrealsense2 as rs  # type: ignore  # TODO: add type stubs for pyrealsense2
-except Exception as e:
-    logging.info(f"Could not import realsense: {e}")
+from lerobot.utils.import_utils import _pyrealsense2_available, require_package
+
+if TYPE_CHECKING or _pyrealsense2_available:
+    import pyrealsense2 as rs
+else:
+    rs = None
 
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.errors import DeviceNotConnectedError
@@ -112,7 +114,7 @@ class RealSenseCamera(Camera):
         Args:
             config: The configuration settings for the camera.
         """
-
+        require_package("pyrealsense2", extra="intelrealsense")
         super().__init__(config)
 
         self.config = config

src/lerobot/cameras/zmq/camera_zmq.py

@@ -28,12 +28,19 @@ import json
 import logging
 import time
 from threading import Event, Lock, Thread
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 import cv2
 import numpy as np
 from numpy.typing import NDArray
 
+from lerobot.utils.import_utils import _zmq_available, require_package
+
+if TYPE_CHECKING or _zmq_available:
+    import zmq
+else:
+    zmq = None
+
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.errors import DeviceNotConnectedError
 
@@ -74,8 +81,8 @@ class ZMQCamera(Camera):
     """
 
     def __init__(self, config: ZMQCameraConfig):
+        require_package("pyzmq", extra="pyzmq-dep", import_name="zmq")
         super().__init__(config)
-        import zmq
 
         self.config = config
         self.server_address = config.server_address
@@ -117,8 +124,6 @@ class ZMQCamera(Camera):
         logger.info(f"Connecting to {self}...")
 
         try:
-            import zmq
-
             self.context = zmq.Context()
             self.socket = self.context.socket(zmq.SUB)
             self.socket.setsockopt_string(zmq.SUBSCRIBE, "")
@@ -180,11 +185,8 @@ class ZMQCamera(Camera):
 
         try:
             message = self.socket.recv_string()
-        except Exception as e:
-            # zmq is lazy-imported in connect(), so check by name to avoid a top-level import
-            if type(e).__name__ == "Again":
-                raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
-            raise
+        except zmq.Again as e:
+            raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
 
         # Decode JSON message
         data = json.loads(message)

src/lerobot/common/control_utils.py

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

src/lerobot/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/default.py

@@ -35,6 +35,9 @@ class DatasetConfig:
     revision: str | None = None
     use_imagenet_stats: bool = True
     video_backend: str = field(default_factory=get_safe_default_codec)
+    # When True, video frames are returned as uint8 tensors (0-255) instead of float32 (0.0-1.0).
+    # This reduces memory and speeds up DataLoader IPC. The training pipeline handles the conversion.
+    return_uint8: bool = False
     streaming: bool = False
 
     def __post_init__(self) -> None:

src/lerobot/configs/recipe.py

@@ -0,0 +1,208 @@
+#!/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 the recipe supervises something.
+
+        A recipe is valid if it has at least one of:
+
+        * a ``target: true`` assistant turn (drives text-CE supervision), or
+        * a ``stream: low_level`` turn (drives flow / action supervision via
+          ``predict_actions=True``, even when no assistant turn is targeted —
+          e.g. π0.5-style ``low_level_execution`` where the action expert
+          conditions on a user-only ``${subtask}`` prompt).
+        """
+        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)}")
+
+        has_target = any(turn.target for turn in self.messages)
+        has_low_level = any(turn.stream == "low_level" for turn in self.messages)
+        if not (has_target or has_low_level):
+            raise ValueError(
+                "Message recipes must contain at least one supervised turn — "
+                "either ``target: true`` (text CE) or ``stream: low_level`` "
+                "(flow/action loss)."
+            )
+
+    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/hirobot.yaml

@@ -0,0 +1,77 @@
+# Hi-Robot blend — shared between SmolVLA2 (SmolVLM2 backbone) and
+# PI052 (PaliGemma backbone). π0.5-style split:
+#
+#   The action expert is conditioned on (images, state, subtask) only.
+#   Hierarchical context (task + plan + memory) only flows into the
+#   high-level text head.
+#
+#     high_level_subtask  — predict subtask from (task+plan+memory),
+#                           and the new memory at boundary frames.
+#     low_level_execution — flow loss with [images, subtask, state].
+#     plan_generation     — task → plan.
+#     ask_vqa_{top,wrist} — camera-grounded VQA.
+#
+# Each backbone's text tokenizer renders these messages differently
+# (SmolVLA2 uses the chat template; PI052 concatenates as plain
+# ``Role: content`` text), but the recipe spec is identical.
+
+blend:
+
+  high_level_subtask:
+    weight: 0.50
+    bindings:
+      new_memory: "emitted_at(t, style=memory)"
+    messages:
+      - role: user
+        stream: high_level
+        content: "${task}\nPlan: ${plan}\nMemory: ${memory}"
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+      # Boundary-frame tail: at a subtask transition, also predict
+      # the new memory in the same forward pass.
+      - {role: assistant, content: "${new_memory}", stream: high_level, target: true, if_present: new_memory}
+
+  low_level_execution:
+    weight: 0.30
+    messages:
+      # π0.5-style action conditioning. The action expert sees only
+      # [images, this user turn (= bare subtask), state]. No text-CE
+      # target — subtask prediction is owned by ``high_level_subtask``.
+      # ``stream: low_level`` flips ``predict_actions=True`` so the
+      # flow loss fires.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  plan_generation:
+    weight: 0.10
+    bindings:
+      current_plan: "active_at(t, style=plan)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${current_plan}", stream: high_level, target: true, if_present: current_plan}
+
+  ask_vqa_top:
+    weight: 0.05
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.front)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.front)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.front}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
+
+  ask_vqa_wrist:
+    weight: 0.05
+    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/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/configs/train.py

@@ -56,6 +56,8 @@ class TrainPipelineConfig(HubMixin):
     # Number of workers for the dataloader.
     num_workers: int = 4
     batch_size: int = 8
+    prefetch_factor: int = 4
+    persistent_workers: bool = True
     steps: int = 100_000
     eval_freq: int = 20_000
     log_freq: int = 200

src/lerobot/datasets/__init__.py

@@ -34,9 +34,16 @@ from .dataset_tools import (
     remove_feature,
     split_dataset,
 )
-from .factory import make_dataset, resolve_delta_timestamps
 from .image_writer import safe_stop_image_writer
 from .io_utils import load_episodes, write_stats
+from .language import (
+    EVENT_ONLY_STYLES,
+    LANGUAGE_EVENTS,
+    LANGUAGE_PERSISTENT,
+    PERSISTENT_STYLES,
+    STYLE_REGISTRY,
+    column_for_style,
+)
 from .lerobot_dataset import LeRobotDataset
 from .multi_dataset import MultiLeRobotDataset
 from .pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
@@ -45,6 +52,19 @@ from .streaming_dataset import StreamingLeRobotDataset
 from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card
 from .video_utils import VideoEncodingManager
 
+
+def make_dataset(*args, **kwargs):
+    from .factory import make_dataset as _make_dataset
+
+    return _make_dataset(*args, **kwargs)
+
+
+def resolve_delta_timestamps(*args, **kwargs):
+    from .factory import resolve_delta_timestamps as _resolve_delta_timestamps
+
+    return _resolve_delta_timestamps(*args, **kwargs)
+
+
 # NOTE: Low-level I/O functions (cast_stats_to_numpy, get_parquet_file_size_in_mb, etc.)
 # and legacy migration constants are intentionally NOT re-exported here.
 # Import directly: ``from lerobot.datasets.io_utils import ...``
@@ -53,10 +73,15 @@ __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 +91,7 @@ __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

@@ -16,6 +16,7 @@
 """Private reader component for LeRobotDataset. Handles random-access reading (HF dataset, delta indices, video decoding)."""
 
 from collections.abc import Callable
+from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
 
 import datasets
@@ -49,6 +50,7 @@ class DatasetReader:
         video_backend: str,
         delta_timestamps: dict[str, list[float]] | None,
         image_transforms: Callable | None,
+        return_uint8: bool = False,
     ):
         """Initialize the reader with metadata, filtering, and transform config.
 
@@ -73,6 +75,7 @@ class DatasetReader:
         self._tolerance_s = tolerance_s
         self._video_backend = video_backend
         self._image_transforms = image_transforms
+        self._return_uint8 = return_uint8
 
         self.hf_dataset: datasets.Dataset | None = None
         self._absolute_to_relative_idx: dict[int, int] | None = None
@@ -105,10 +108,8 @@ class DatasetReader:
         """Build absolute-to-relative index mapping from loaded hf_dataset."""
         self._absolute_to_relative_idx = None
         if self.episodes is not None and self.hf_dataset is not None:
-            self._absolute_to_relative_idx = {
-                abs_idx.item() if isinstance(abs_idx, torch.Tensor) else abs_idx: rel_idx
-                for rel_idx, abs_idx in enumerate(self.hf_dataset["index"])
-            }
+            indices = self.hf_dataset.data.column("index").to_numpy()
+            self._absolute_to_relative_idx = dict(zip(indices.tolist(), range(len(indices)), strict=True))
 
     @property
     def num_frames(self) -> int:
@@ -125,10 +126,53 @@ class DatasetReader:
     def _load_hf_dataset(self) -> datasets.Dataset:
         """hf_dataset contains all the observations, states, actions, rewards, etc."""
         features = get_hf_features_from_features(self._meta.features)
+        # Datasets annotated with the PR1 language columns may have been
+        # written without registering those columns in ``meta/info.json``
+        # (e.g. they predate ``CODEBASE_VERSION="v3.1"`` and were
+        # back-filled by ``lerobot-annotate``). Probe a single parquet
+        # shard and graft the column features on so the strict
+        # ``Dataset.from_parquet`` cast doesn't fail with
+        # ``column names don't match``.
+        features = self._extend_features_with_language_columns(features)
         hf_dataset = load_nested_dataset(self.root / "data", features=features, episodes=self.episodes)
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
+    def _extend_features_with_language_columns(
+        self, features: datasets.Features
+    ) -> datasets.Features:
+        """Add ``language_persistent`` / ``language_events`` to ``features``
+        when the underlying parquet shards declare them but the metadata
+        doesn't. No-op when neither column is present or both are
+        already registered.
+        """
+        # Find any one parquet to peek at; bail if there are none yet
+        # (the dataset will fail later for an unrelated reason and we
+        # want that error to surface as-is).
+        try:
+            sample = next((self.root / "data").glob("*/*.parquet"))
+        except StopIteration:
+            return features
+
+        from pyarrow import parquet as _pq  # noqa: PLC0415
+
+        schema_names = set(_pq.read_schema(sample).names)
+        from .language import (  # noqa: PLC0415
+            LANGUAGE_EVENTS,
+            LANGUAGE_PERSISTENT,
+            language_events_column_feature,
+            language_persistent_column_feature,
+        )
+
+        extra: dict[str, object] = {}
+        if LANGUAGE_PERSISTENT in schema_names and LANGUAGE_PERSISTENT not in features:
+            extra[LANGUAGE_PERSISTENT] = language_persistent_column_feature()
+        if LANGUAGE_EVENTS in schema_names and LANGUAGE_EVENTS not in features:
+            extra[LANGUAGE_EVENTS] = language_events_column_feature()
+        if not extra:
+            return features
+        return datasets.Features({**features, **extra})
+
     def _check_cached_episodes_sufficient(self) -> bool:
         """Check if the cached dataset contains all requested episodes and their video files."""
         if self.hf_dataset is None or len(self.hf_dataset) == 0:
@@ -235,16 +279,30 @@ class DatasetReader:
         Segmentation Fault.
         """
         ep = self._meta.episodes[ep_idx]
-        item = {}
-        for vid_key, query_ts in query_timestamps.items():
+
+        def _decode_single(vid_key: str, query_ts: list[float]) -> tuple[str, torch.Tensor]:
             from_timestamp = ep[f"videos/{vid_key}/from_timestamp"]
             shifted_query_ts = [from_timestamp + ts for ts in query_ts]
-
             video_path = self.root / self._meta.get_video_file_path(ep_idx, vid_key)
-            frames = decode_video_frames(video_path, shifted_query_ts, self._tolerance_s, self._video_backend)
-            item[vid_key] = frames.squeeze(0)
+            frames = decode_video_frames(
+                video_path,
+                shifted_query_ts,
+                self._tolerance_s,
+                self._video_backend,
+                return_uint8=self._return_uint8,
+            )
+            return vid_key, frames.squeeze(0)
 
-        return item
+        items = list(query_timestamps.items())
+
+        # Single camera: no threading overhead
+        if len(items) <= 1:
+            return {vid_key: _decode_single(vid_key, query_ts)[1] for vid_key, query_ts in items}
+
+        # Multi-camera: decode in parallel (video decoding releases the GIL)
+        with ThreadPoolExecutor(max_workers=len(items)) as pool:
+            futures = [pool.submit(_decode_single, k, ts) for k, ts in items]
+            return dict(f.result() for f in futures)
 
     def get_item(self, idx) -> dict:
         """Core __getitem__ logic. Assumes hf_dataset is loaded.
@@ -280,9 +338,4 @@ 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/dataset_writer.py

@@ -597,7 +597,7 @@ class DatasetWriter:
 
     def cleanup_interrupted_episode(self, episode_index: int) -> None:
         """Remove temporary image directories for an interrupted episode."""
-        for key in self._meta.video_keys:
+        for key in self._meta.camera_keys:
             img_dir = self._get_image_file_path(
                 episode_index=episode_index, image_key=key, frame_index=0
             ).parent

src/lerobot/datasets/factory.py

@@ -92,6 +92,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 image_transforms=image_transforms,
                 revision=cfg.dataset.revision,
                 video_backend=cfg.dataset.video_backend,
+                return_uint8=True,
                 tolerance_s=cfg.tolerance_s,
             )
         else:
@@ -104,6 +105,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 revision=cfg.dataset.revision,
                 max_num_shards=cfg.num_workers,
                 tolerance_s=cfg.tolerance_s,
+                return_uint8=True,
             )
     else:
         raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")

src/lerobot/datasets/feature_utils.py

@@ -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/image_writer.py

@@ -30,13 +30,13 @@ def safe_stop_image_writer(func):
     def wrapper(*args, **kwargs):
         try:
             return func(*args, **kwargs)
-        except Exception as e:
+        except BaseException:
             dataset = kwargs.get("dataset")
             writer = getattr(dataset, "writer", None) if dataset else None
             if writer is not None and writer.image_writer is not None:
                 logger.warning("Waiting for image writer to terminate...")
                 writer.image_writer.stop()
-            raise e
+            raise
 
     return wrapper
 

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 +1,250 @@
+#!/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 feature used for tool-call payloads.
+
+    Older ``datasets`` versions do not expose ``datasets.Json``. The
+    annotation pipeline currently emits the canonical ``say`` tool call
+    shape, so use that explicit struct instead of falling back to a string
+    that cannot cast structured parquet values.
+    """
+    if hasattr(datasets, "Json"):
+        return datasets.Json()
+    return {
+        "type": datasets.Value("string"),
+        "function": {
+            "name": datasets.Value("string"),
+            "arguments": {"text": 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/lerobot_dataset.py

@@ -56,6 +56,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         force_cache_sync: bool = False,
         download_videos: bool = True,
         video_backend: str | None = None,
+        return_uint8: bool = False,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
         streaming_encoding: bool = False,
@@ -202,6 +203,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.tolerance_s = tolerance_s
         self.revision = revision if revision else CODEBASE_VERSION
         self._video_backend = video_backend if video_backend else get_safe_default_codec()
+        self._return_uint8 = return_uint8
         self._batch_encoding_size = batch_encoding_size
         self._vcodec = resolve_vcodec(vcodec)
         self._encoder_threads = encoder_threads
@@ -225,6 +227,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             video_backend=self._video_backend,
             delta_timestamps=delta_timestamps,
             image_transforms=image_transforms,
+            return_uint8=self._return_uint8,
         )
 
         # Load actual data
@@ -288,6 +291,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 video_backend=self._video_backend,
                 delta_timestamps=self.delta_timestamps,
                 image_transforms=self.image_transforms,
+                return_uint8=self._return_uint8,
             )
         return self.reader
 
@@ -683,6 +687,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.delta_timestamps = None
         obj.episodes = None
         obj._video_backend = video_backend if video_backend is not None else get_safe_default_codec()
+        obj._return_uint8 = False
         obj._batch_encoding_size = batch_encoding_size
         obj._vcodec = vcodec
         obj._encoder_threads = encoder_threads
@@ -775,6 +780,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.delta_timestamps = None
         obj.episodes = None
         obj._video_backend = video_backend if video_backend else get_safe_default_codec()
+        obj._return_uint8 = False
         obj._batch_encoding_size = batch_encoding_size
         obj._vcodec = vcodec
         obj._encoder_threads = encoder_threads

src/lerobot/datasets/streaming_dataset.py

@@ -251,6 +251,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
         seed: int = 42,
         rng: np.random.Generator | None = None,
         shuffle: bool = True,
+        return_uint8: bool = False,
     ):
         """Initialize a StreamingLeRobotDataset.
 
@@ -288,6 +289,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
 
         self.streaming = streaming
         self.buffer_size = buffer_size
+        self._return_uint8 = return_uint8
 
         # We cache the video decoders to avoid re-initializing them at each frame (avoiding a ~10x slowdown)
         self.video_decoder_cache = None
@@ -553,7 +555,11 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
             root = self.meta.url_root if self.streaming and not self.streaming_from_local else self.root
             video_path = f"{root}/{self.meta.get_video_file_path(ep_idx, video_key)}"
             frames = decode_video_frames_torchcodec(
-                video_path, query_ts, self.tolerance_s, decoder_cache=self.video_decoder_cache
+                video_path,
+                query_ts,
+                self.tolerance_s,
+                decoder_cache=self.video_decoder_cache,
+                return_uint8=self._return_uint8,
             )
 
             item[video_key] = frames.squeeze(0) if len(query_ts) == 1 else frames

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"
@@ -238,17 +237,24 @@ def get_safe_version(repo_id: str, version: str | packaging.version.Version) ->
     hub_versions = get_repo_versions(repo_id)
 
     if not hub_versions:
-        raise RevisionNotFoundError(
-            f"""Your dataset must be tagged with a codebase version.
-            Assuming _version_ is the codebase_version value in the info.json, you can run this:
-            ```python
-            from huggingface_hub import HfApi
-
-            hub_api = HfApi()
-            hub_api.create_tag("{repo_id}", tag="_version_", repo_type="dataset")
-            ```
-            """
+        msg = (
+            f"Repo {repo_id!r} has no codebase-version tags. The dataset "
+            f"either doesn't exist on the Hub yet, or it was uploaded "
+            f"without a ``v3.x``-style tag. To tag an existing dataset run:\n"
+            f"  from huggingface_hub import HfApi\n"
+            f"  HfApi().create_tag({repo_id!r}, tag='v3.0', repo_type='dataset', exist_ok=True)"
         )
+        # ``RevisionNotFoundError`` extends ``HfHubHTTPError`` whose
+        # ``__init__`` indexes ``response.headers`` unconditionally on
+        # current ``huggingface_hub`` versions. Constructing it without
+        # a real ``Response`` object crashes with either
+        # ``TypeError: missing 1 required keyword-only argument`` (old
+        # builds) or ``AttributeError: 'NoneType' object has no attribute
+        # 'headers'`` (new builds). Skip that path entirely — this isn't
+        # really an HTTP error, it's a configuration issue — and raise a
+        # plain ``RuntimeError`` so the message actually reaches the
+        # caller.
+        raise RuntimeError(msg)
 
     if target_version in hub_versions:
         return f"v{target_version}"

src/lerobot/datasets/video_utils.py

@@ -123,6 +123,7 @@ def decode_video_frames(
     timestamps: list[float],
     tolerance_s: float,
     backend: str | None = None,
+    return_uint8: bool = False,
 ) -> torch.Tensor:
     """
     Decodes video frames using the specified backend.
@@ -131,19 +132,23 @@ def decode_video_frames(
         video_path (Path): Path to the video file.
         timestamps (list[float]): List of timestamps to extract frames.
         tolerance_s (float): Allowed deviation in seconds for frame retrieval.
-        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available in the platform; otherwise, defaults to "pyav"..
+        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available in the platform; otherwise, defaults to "pyav".
+        return_uint8 (bool): If True, return raw uint8 frames without float32 normalization.
+            This reduces memory for DataLoader IPC; normalization can be done on GPU afterward.
 
     Returns:
-        torch.Tensor: Decoded frames.
+        torch.Tensor: Decoded frames (float32 in [0,1] by default, or uint8 if return_uint8=True).
 
     Currently supports torchcodec on cpu and pyav.
     """
     if backend is None:
         backend = get_safe_default_codec()
     if backend == "torchcodec":
-        return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s)
+        return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s, return_uint8=return_uint8)
     elif backend in ["pyav", "video_reader"]:
-        return decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
+        return decode_video_frames_torchvision(
+            video_path, timestamps, tolerance_s, backend, return_uint8=return_uint8
+        )
     else:
         raise ValueError(f"Unsupported video backend: {backend}")
 
@@ -154,6 +159,7 @@ def decode_video_frames_torchvision(
     tolerance_s: float,
     backend: str = "pyav",
     log_loaded_timestamps: bool = False,
+    return_uint8: bool = False,
 ) -> torch.Tensor:
     """Loads frames associated to the requested timestamps of a video
 
@@ -240,14 +246,17 @@ def decode_video_frames_torchvision(
     if log_loaded_timestamps:
         logger.info(f"{closest_ts=}")
 
-    # convert to the pytorch format which is float32 in [0,1] range (and channel first)
-    closest_frames = closest_frames.type(torch.float32) / 255
-
     if len(timestamps) != len(closest_frames):
         raise FrameTimestampError(
             f"Number of retrieved frames ({len(closest_frames)}) does not match "
             f"number of queried timestamps ({len(timestamps)})"
         )
+
+    if return_uint8:
+        return closest_frames
+
+    # convert to the pytorch format which is float32 in [0,1] range (and channel first)
+    closest_frames = closest_frames.type(torch.float32) / 255
     return closest_frames
 
 
@@ -306,6 +315,7 @@ def decode_video_frames_torchcodec(
     tolerance_s: float,
     log_loaded_timestamps: bool = False,
     decoder_cache: VideoDecoderCache | None = None,
+    return_uint8: bool = False,
 ) -> torch.Tensor:
     """Loads frames associated with the requested timestamps of a video using torchcodec.
 
@@ -373,14 +383,16 @@ def decode_video_frames_torchcodec(
     if log_loaded_timestamps:
         logger.info(f"{closest_ts=}")
 
-    # convert to float32 in [0,1] range
-    closest_frames = (closest_frames / 255.0).type(torch.float32)
-
     if not len(timestamps) == len(closest_frames):
         raise FrameTimestampError(
             f"Retrieved timestamps differ from queried {set(closest_frames) - set(timestamps)}"
         )
 
+    if return_uint8:
+        return closest_frames
+
+    # convert to float32 in [0,1] range
+    closest_frames = (closest_frames / 255.0).type(torch.float32)
     return closest_frames
 
 

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):
@@ -496,6 +498,146 @@ class MetaworldEnv(EnvConfig):
         )
 
 
+@EnvConfig.register_subclass("robocasa")
+@dataclass
+class RoboCasaEnv(EnvConfig):
+    task: str = "CloseFridge"
+    fps: int = 20
+    episode_length: int = 1000
+    obs_type: str = "pixels_agent_pos"
+    render_mode: str = "rgb_array"
+    camera_name: str = "robot0_agentview_left,robot0_eye_in_hand,robot0_agentview_right"
+    observation_height: int = 256
+    observation_width: int = 256
+    visualization_height: int = 512
+    visualization_width: int = 512
+    split: str | None = None
+    # Object-mesh registries to sample from. Upstream default is
+    # ("objaverse", "lightwheel"), but objaverse is ~30GB and the CI image
+    # only ships the lightwheel pack. Override to include objaverse once
+    # you've run `python -m robocasa.scripts.download_kitchen_assets
+    # --type objaverse` locally.
+    obj_registries: list[str] = field(default_factory=lambda: ["lightwheel"])
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(12,))}
+    )
+    features_map: dict[str, str] = field(default_factory=lambda: {ACTION: ACTION, "agent_pos": OBS_STATE})
+
+    def __post_init__(self):
+        if self.obs_type not in ("pixels", "pixels_agent_pos"):
+            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
+
+        # Preserve raw RoboCasa camera names end-to-end (e.g.
+        # `observation.images.robot0_agentview_left`). This matches the
+        # naming convention used by the RoboCasa datasets on the Hub, so
+        # trained policies don't need a `--rename_map` at eval time.
+        cams = [c.strip() for c in self.camera_name.split(",") if c.strip()]
+        for cam in cams:
+            self.features[f"pixels/{cam}"] = PolicyFeature(
+                type=FeatureType.VISUAL,
+                shape=(self.observation_height, self.observation_width, 3),
+            )
+            self.features_map[f"pixels/{cam}"] = f"{OBS_IMAGES}.{cam}"
+
+        if self.obs_type == "pixels_agent_pos":
+            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(16,))
+
+    @property
+    def gym_kwargs(self) -> dict:
+        kwargs: dict[str, Any] = {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+            "observation_height": self.observation_height,
+            "observation_width": self.observation_width,
+            "visualization_height": self.visualization_height,
+            "visualization_width": self.visualization_width,
+        }
+        if self.split is not None:
+            kwargs["split"] = self.split
+        return kwargs
+
+    def create_envs(self, n_envs: int, use_async_envs: bool = False):
+        from .robocasa import create_robocasa_envs
+
+        if self.task is None:
+            raise ValueError("RoboCasaEnv requires a task to be specified")
+        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
+        return create_robocasa_envs(
+            task=self.task,
+            n_envs=n_envs,
+            camera_name=self.camera_name,
+            gym_kwargs=self.gym_kwargs,
+            env_cls=env_cls,
+            episode_length=self.episode_length,
+            obj_registries=tuple(self.obj_registries),
+        )
+
+
+@EnvConfig.register_subclass("vlabench")
+@dataclass
+class VLABenchEnv(EnvConfig):
+    task: str = "select_fruit"
+    fps: int = 10
+    episode_length: int = 500
+    obs_type: str = "pixels_agent_pos"
+    render_mode: str = "rgb_array"
+    render_resolution: tuple[int, int] = (480, 480)
+    robot: str = "franka"
+    action_mode: str = "eef"
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
+        }
+    )
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            ACTION: ACTION,
+            "agent_pos": OBS_STATE,
+            "pixels/image": f"{OBS_IMAGES}.image",
+            "pixels/second_image": f"{OBS_IMAGES}.second_image",
+            "pixels/wrist_image": f"{OBS_IMAGES}.wrist_image",
+        }
+    )
+
+    def __post_init__(self):
+        h, w = self.render_resolution
+        if self.obs_type == "pixels":
+            self.features["pixels/image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+            self.features["pixels/second_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+            self.features["pixels/wrist_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+        elif self.obs_type == "pixels_agent_pos":
+            self.features["pixels/image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+            self.features["pixels/second_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+            self.features["pixels/wrist_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
+            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(7,))
+        else:
+            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+            "render_resolution": self.render_resolution,
+            "robot": self.robot,
+            "max_episode_steps": self.episode_length,
+            "action_mode": self.action_mode,
+        }
+
+    def create_envs(self, n_envs: int, use_async_envs: bool = False):
+        from .vlabench import create_vlabench_envs
+
+        if self.task is None:
+            raise ValueError("VLABenchEnv requires a task to be specified")
+        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
+        return create_vlabench_envs(
+            task=self.task,
+            n_envs=n_envs,
+            gym_kwargs=self.gym_kwargs,
+            env_cls=env_cls,
+        )
+
+
 @EnvConfig.register_subclass("isaaclab_arena")
 @dataclass
 class IsaaclabArenaEnv(HubEnvConfig):
@@ -574,3 +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
@@ -31,20 +32,7 @@ from libero.libero.envs import OffScreenRenderEnv
 
 from lerobot.types import RobotObservation
 
-from .utils import _LazyAsyncVectorEnv
-
-
-def _parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
-    """Normalize camera_name into a non-empty list of strings."""
-    if isinstance(camera_name, str):
-        cams = [c.strip() for c in camera_name.split(",") if c.strip()]
-    elif isinstance(camera_name, (list | tuple)):
-        cams = [str(c).strip() for c in camera_name if str(c).strip()]
-    else:
-        raise TypeError(f"camera_name must be str or sequence[str], got {type(camera_name).__name__}")
-    if not cams:
-        raise ValueError("camera_name resolved to an empty list.")
-    return cams
+from .utils import _LazyAsyncVectorEnv, parse_camera_names
 
 
 def _get_suite(name: str) -> benchmark.Benchmark:
@@ -69,14 +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():
@@ -118,9 +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
@@ -128,7 +138,7 @@ class LiberoEnv(gym.Env):
         self.visualization_width = visualization_width
         self.visualization_height = visualization_height
         self.init_states = init_states
-        self.camera_name = _parse_camera_names(
+        self.camera_name = parse_camera_names(
             camera_name
         )  # agentview_image (main) or robot0_eye_in_hand_image (wrist)
 
@@ -147,7 +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
@@ -380,6 +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)."""
 
@@ -396,6 +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,
         )
 
@@ -418,6 +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.
@@ -437,7 +454,7 @@ def create_libero_envs(
     gym_kwargs = dict(gym_kwargs or {})
     task_ids_filter = gym_kwargs.pop("task_ids", None)  # optional: limit to specific tasks
 
-    camera_names = _parse_camera_names(camera_name)
+    camera_names = parse_camera_names(camera_name)
     suite_names = [s.strip() for s in str(task).split(",") if s.strip()]
     if not suite_names:
         raise ValueError("`task` must contain at least one LIBERO suite name.")
@@ -462,6 +479,7 @@ def create_libero_envs(
         # Probe once and reuse to avoid creating a temp env per task.
         cached_obs_space: spaces.Space | None = None
         cached_act_space: spaces.Space | None = None
+        cached_metadata: dict[str, Any] | None = None
 
         for tid in selected:
             fns = _make_env_fns(
@@ -475,12 +493,14 @@ 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)
+                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[suite_name][tid] = lazy
             else:
                 out[suite_name][tid] = env_cls(fns)

src/lerobot/envs/metaworld.py

@@ -311,6 +311,7 @@ def create_metaworld_envs(
     is_async = env_cls is gym.vector.AsyncVectorEnv
     cached_obs_space = None
     cached_act_space = None
+    cached_metadata = None
     out: dict[str, dict[int, Any]] = defaultdict(dict)
 
     for group in task_groups:
@@ -324,10 +325,11 @@ def create_metaworld_envs(
             fns = [(lambda tn=task_name: MetaworldEnv(task=tn, **gym_kwargs)) for _ in range(n_envs)]
 
             if is_async:
-                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space)
+                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)

src/lerobot/envs/robocasa.py

@@ -0,0 +1,425 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import annotations
+
+import logging
+from collections import defaultdict
+from collections.abc import Callable, Sequence
+from functools import partial
+from typing import Any
+
+import gymnasium as gym
+import numpy as np
+from gymnasium import spaces
+
+from lerobot.types import RobotObservation
+
+from .utils import _LazyAsyncVectorEnv, parse_camera_names
+
+logger = logging.getLogger(__name__)
+
+# Dimensions for the flat action/state vectors used by the LeRobot wrapper.
+# These correspond to the PandaOmron robot in RoboCasa365.
+OBS_STATE_DIM = 16  # base_pos(3) + base_quat(4) + ee_pos_rel(3) + ee_quat_rel(4) + gripper_qpos(2)
+ACTION_DIM = 12  # base_motion(4) + control_mode(1) + ee_pos(3) + ee_rot(3) + gripper(1)
+ACTION_LOW = -1.0
+ACTION_HIGH = 1.0
+
+# Default PandaOmron cameras. We surface these raw names directly as
+# `observation.images.<name>` so the LeRobot dataset/policy keys match
+# RoboCasa's native convention (no implicit renaming).
+DEFAULT_CAMERAS = [
+    "robot0_agentview_left",
+    "robot0_eye_in_hand",
+    "robot0_agentview_right",
+]
+
+# Object-mesh registries to sample from. RoboCasa's upstream default is
+# ("objaverse", "lightwheel"), but the objaverse pack is huge (~30GB) and
+# most users — including our CI image — only download the lightwheel pack
+# (`--type objs_lw` in `download_kitchen_assets`). When a sampled object
+# category has zero candidates in every registry, robocasa crashes with
+# `ValueError: Probabilities contain NaN` (0/0 divide in the probability
+# normalization). Restricting to registries that are actually on disk
+# avoids the NaN and matches what the asset download provides.
+DEFAULT_OBJ_REGISTRIES: tuple[str, ...] = ("lightwheel",)
+
+# Task-group shortcuts accepted as `--env.task`. When the user passes one of
+# these names, we expand it to the upstream RoboCasa task list and auto-set
+# the dataset split. Individual task names (optionally comma-separated) still
+# take precedence; this only triggers on an exact group-name match.
+_TASK_GROUP_SPLITS = {
+    "atomic_seen": "target",
+    "composite_seen": "target",
+    "composite_unseen": "target",
+    "pretrain50": "pretrain",
+    "pretrain100": "pretrain",
+    "pretrain200": "pretrain",
+    "pretrain300": "pretrain",
+}
+
+
+def _resolve_tasks(task: str) -> tuple[list[str], str | None]:
+    """Resolve a `--env.task` value to (task_names, split_override).
+
+    If `task` is a known task-group name (e.g. `atomic_seen`, `pretrain100`),
+    expand it via `robocasa.utils.dataset_registry.{TARGET,PRETRAINING}_TASKS`
+    and return the matching split. Otherwise treat `task` as a single task or
+    comma-separated list and leave the split untouched (None).
+    """
+    key = task.strip()
+    if key in _TASK_GROUP_SPLITS:
+        from robocasa.utils.dataset_registry import PRETRAINING_TASKS, TARGET_TASKS
+
+        combined = {**TARGET_TASKS, **PRETRAINING_TASKS}
+        if key not in combined:
+            raise ValueError(
+                f"Task group '{key}' is not available in this version of robocasa. "
+                f"Known groups: {sorted(combined.keys())}."
+            )
+        return list(combined[key]), _TASK_GROUP_SPLITS[key]
+
+    names = [t.strip() for t in task.split(",") if t.strip()]
+    if not names:
+        raise ValueError("`task` must contain at least one RoboCasa task name.")
+    return names, None
+
+
+def convert_action(flat_action: np.ndarray) -> dict[str, Any]:
+    """Split a flat (12,) action vector into a RoboCasa action dict.
+
+    Layout: base_motion(4) + control_mode(1) + ee_pos(3) + ee_rot(3) + gripper(1)
+    """
+    return {
+        "action.base_motion": flat_action[0:4],
+        "action.control_mode": flat_action[4:5],
+        "action.end_effector_position": flat_action[5:8],
+        "action.end_effector_rotation": flat_action[8:11],
+        "action.gripper_close": flat_action[11:12],
+    }
+
+
+class RoboCasaEnv(gym.Env):
+    """LeRobot gym.Env wrapper for RoboCasa365 kitchen environments.
+
+    Wraps RoboCasaGymEnv from the robocasa package and converts its
+    dict-based observations and actions into the flat arrays LeRobot expects.
+    Raw RoboCasa camera names are preserved verbatim under `pixels/<cam>`.
+    """
+
+    metadata = {"render_modes": ["rgb_array"], "render_fps": 20}
+
+    def __init__(
+        self,
+        task: str,
+        camera_name: str | Sequence[str] = ",".join(DEFAULT_CAMERAS),
+        obs_type: str = "pixels_agent_pos",
+        render_mode: str = "rgb_array",
+        observation_width: int = 256,
+        observation_height: int = 256,
+        visualization_width: int = 512,
+        visualization_height: int = 512,
+        split: str | None = None,
+        episode_length: int | None = None,
+        obj_registries: Sequence[str] = DEFAULT_OBJ_REGISTRIES,
+        episode_index: int = 0,
+    ):
+        super().__init__()
+        self.task = task
+        self.obs_type = obs_type
+        self.render_mode = render_mode
+        self.observation_width = observation_width
+        self.observation_height = observation_height
+        self.visualization_width = visualization_width
+        self.visualization_height = visualization_height
+        self.split = split
+        self.obj_registries = tuple(obj_registries)
+        # Per-worker index (0..n_envs-1) used to spread the user-provided
+        # seed across factories so each sub-env explores a distinct layout
+        # even when the same seed is passed to `reset()`.
+        self.episode_index = int(episode_index)
+
+        self.camera_name = parse_camera_names(camera_name)
+
+        self._max_episode_steps = episode_length if episode_length is not None else 1000
+
+        # Deferred — created on first reset() inside the worker subprocess
+        # to avoid inheriting stale GPU/EGL contexts across fork().
+        self._env: Any = None
+        self.task_description = ""
+
+        images = {
+            cam: spaces.Box(
+                low=0,
+                high=255,
+                shape=(self.observation_height, self.observation_width, 3),
+                dtype=np.uint8,
+            )
+            for cam in self.camera_name
+        }
+
+        if self.obs_type == "pixels":
+            self.observation_space = spaces.Dict({"pixels": spaces.Dict(images)})
+        elif self.obs_type == "pixels_agent_pos":
+            self.observation_space = spaces.Dict(
+                {
+                    "pixels": spaces.Dict(images),
+                    "agent_pos": spaces.Box(
+                        low=-np.inf,
+                        high=np.inf,
+                        shape=(OBS_STATE_DIM,),
+                        dtype=np.float32,
+                    ),
+                }
+            )
+        else:
+            raise ValueError(f"Unsupported obs_type '{self.obs_type}'. Use 'pixels' or 'pixels_agent_pos'.")
+
+        self.action_space = spaces.Box(
+            low=ACTION_LOW,
+            high=ACTION_HIGH,
+            shape=(ACTION_DIM,),
+            dtype=np.float32,
+        )
+
+    def _ensure_env(self) -> None:
+        """Create the underlying RoboCasaGymEnv on first use.
+
+        Called inside the worker subprocess after fork(), so each worker gets
+        its own clean rendering context rather than inheriting a stale one from
+        the parent process (which causes crashes with AsyncVectorEnv).
+        """
+        if self._env is not None:
+            return
+        from robocasa.wrappers.gym_wrapper import RoboCasaGymEnv
+
+        # RoboCasaGymEnv defaults split="test", which create_env rejects
+        # (only None/"all"/"pretrain"/"target" are valid). Always pass a
+        # valid value so we don't hit that default. Extra kwargs are
+        # forwarded to the underlying kitchen env via create_env/robosuite.make.
+        self._env = RoboCasaGymEnv(
+            env_name=self.task,
+            camera_widths=self.observation_width,
+            camera_heights=self.observation_height,
+            split=self.split if self.split is not None else "all",
+            obj_registries=self.obj_registries,
+        )
+
+        ep_meta = self._env.env.get_ep_meta()
+        self.task_description = ep_meta.get("lang", self.task)
+
+    def _format_raw_obs(self, raw_obs: dict) -> RobotObservation:
+        """Convert RoboCasaGymEnv observation dict to LeRobot format."""
+        # RoboCasaGymEnv emits camera frames under "video.<cam>".
+        images = {cam: raw_obs[f"video.{cam}"] for cam in self.camera_name if f"video.{cam}" in raw_obs}
+
+        if self.obs_type == "pixels":
+            return {"pixels": images}
+
+        # `state.*` keys come from PandaOmronKeyConverter inside the wrapper.
+        agent_pos = np.concatenate(
+            [
+                raw_obs.get("state.base_position", np.zeros(3)),
+                raw_obs.get("state.base_rotation", np.zeros(4)),
+                raw_obs.get("state.end_effector_position_relative", np.zeros(3)),
+                raw_obs.get("state.end_effector_rotation_relative", np.zeros(4)),
+                raw_obs.get("state.gripper_qpos", np.zeros(2)),
+            ],
+            axis=-1,
+        ).astype(np.float32)
+
+        return {"pixels": images, "agent_pos": agent_pos}
+
+    def render(self) -> np.ndarray:
+        self._ensure_env()
+        assert self._env is not None
+        return self._env.render()
+
+    def reset(self, seed=None, **kwargs):
+        self._ensure_env()
+        assert self._env is not None
+        super().reset(seed=seed)
+        # Spread the seed across workers so n_envs factories don't all
+        # roll the same scene. With an explicit user seed we shift it by
+        # episode_index; with no seed we fall back to episode_index so
+        # each worker is still distinct rather than inheriting the same
+        # global RNG state.
+        worker_seed = seed + self.episode_index if seed is not None else self.episode_index
+        raw_obs, info = self._env.reset(seed=worker_seed)
+
+        ep_meta = self._env.env.get_ep_meta()
+        self.task_description = ep_meta.get("lang", self.task)
+
+        observation = self._format_raw_obs(raw_obs)
+        info = {"is_success": False}
+        return observation, info
+
+    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
+        self._ensure_env()
+        assert self._env is not None
+        if action.ndim != 1:
+            raise ValueError(
+                f"Expected action to be 1-D (shape (action_dim,)), "
+                f"but got shape {action.shape} with ndim={action.ndim}"
+            )
+
+        action_dict = convert_action(action)
+        raw_obs, reward, done, truncated, info = self._env.step(action_dict)
+
+        is_success = bool(info.get("success", False))
+        terminated = done or is_success
+        info.update({"task": self.task, "done": done, "is_success": is_success})
+
+        observation = self._format_raw_obs(raw_obs)
+        if terminated:
+            info["final_info"] = {
+                "task": self.task,
+                "done": bool(done),
+                "is_success": bool(is_success),
+            }
+            self.reset()
+
+        return observation, reward, terminated, truncated, info
+
+    def close(self):
+        if self._env is not None:
+            self._env.close()
+
+
+def _make_env_fns(
+    *,
+    task: str,
+    n_envs: int,
+    camera_names: list[str],
+    obs_type: str,
+    render_mode: str,
+    observation_width: int,
+    observation_height: int,
+    visualization_width: int,
+    visualization_height: int,
+    split: str | None,
+    episode_length: int | None,
+    obj_registries: Sequence[str],
+) -> list[Callable[[], RoboCasaEnv]]:
+    """Build n_envs factory callables for a single task.
+
+    Each factory carries a distinct ``episode_index`` (``0..n_envs-1``) so
+    ``RoboCasaEnv.reset()`` can derive a per-worker seed series from the
+    user-provided seed.
+    """
+
+    def _make_env(episode_index: int) -> RoboCasaEnv:
+        return RoboCasaEnv(
+            task=task,
+            camera_name=camera_names,
+            obs_type=obs_type,
+            render_mode=render_mode,
+            observation_width=observation_width,
+            observation_height=observation_height,
+            visualization_width=visualization_width,
+            visualization_height=visualization_height,
+            split=split,
+            episode_length=episode_length,
+            obj_registries=obj_registries,
+            episode_index=episode_index,
+        )
+
+    return [partial(_make_env, i) for i in range(n_envs)]
+
+
+def create_robocasa_envs(
+    task: str,
+    n_envs: int,
+    gym_kwargs: dict[str, Any] | None = None,
+    camera_name: str | Sequence[str] = ",".join(DEFAULT_CAMERAS),
+    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
+    episode_length: int | None = None,
+    obj_registries: Sequence[str] = DEFAULT_OBJ_REGISTRIES,
+) -> dict[str, dict[int, Any]]:
+    """Create vectorized RoboCasa365 environments with a consistent return shape.
+
+    Returns:
+        dict[task_name][task_id] -> vec_env (env_cls([...]) with exactly n_envs factories)
+
+    `task` can be:
+      - a single task name (e.g. `CloseFridge`)
+      - a comma-separated list of task names (e.g. `CloseFridge,PickPlaceCoffee`)
+      - a benchmark-group shortcut (`atomic_seen`, `composite_seen`,
+        `composite_unseen`, `pretrain50`, `pretrain100`, `pretrain200`,
+        `pretrain300`), which auto-expands to the upstream task list and
+        auto-sets the dataset `split` ("target" or "pretrain").
+    """
+    if env_cls is None or not callable(env_cls):
+        raise ValueError("env_cls must be a callable that wraps a list of environment factory callables.")
+    if not isinstance(n_envs, int) or n_envs <= 0:
+        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
+
+    gym_kwargs = dict(gym_kwargs or {})
+    obs_type = gym_kwargs.pop("obs_type", "pixels_agent_pos")
+    render_mode = gym_kwargs.pop("render_mode", "rgb_array")
+    observation_width = gym_kwargs.pop("observation_width", 256)
+    observation_height = gym_kwargs.pop("observation_height", 256)
+    visualization_width = gym_kwargs.pop("visualization_width", 512)
+    visualization_height = gym_kwargs.pop("visualization_height", 512)
+    split = gym_kwargs.pop("split", None)
+
+    camera_names = parse_camera_names(camera_name)
+    task_names, group_split = _resolve_tasks(str(task))
+    if group_split is not None and split is None:
+        split = group_split
+
+    logger.info(
+        "Creating RoboCasa envs | tasks=%s | split=%s | n_envs(per task)=%d",
+        task_names,
+        split,
+        n_envs,
+    )
+
+    is_async = env_cls is gym.vector.AsyncVectorEnv
+
+    cached_obs_space: spaces.Space | None = None
+    cached_act_space: spaces.Space | None = None
+    cached_metadata: dict[str, Any] | None = None
+    out: dict[str, dict[int, Any]] = defaultdict(dict)
+
+    for task_name in task_names:
+        fns = _make_env_fns(
+            task=task_name,
+            n_envs=n_envs,
+            camera_names=camera_names,
+            obs_type=obs_type,
+            render_mode=render_mode,
+            observation_width=observation_width,
+            observation_height=observation_height,
+            visualization_width=visualization_width,
+            visualization_height=visualization_height,
+            split=split,
+            episode_length=episode_length,
+            obj_registries=obj_registries,
+        )
+
+        if is_async:
+            lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
+            if cached_obs_space is None:
+                cached_obs_space = lazy.observation_space
+                cached_act_space = lazy.action_space
+                cached_metadata = lazy.metadata
+            out[task_name][0] = lazy
+        else:
+            out[task_name][0] = env_cls(fns)
+        logger.info("Built vec env | task=%s | n_envs=%d", task_name, n_envs)
+
+    return {name: dict(task_map) for name, task_map in out.items()}

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/utils.py

@@ -34,6 +34,25 @@ from lerobot.utils.utils import get_channel_first_image_shape
 from .configs import EnvConfig
 
 
+def parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
+    """Normalize ``camera_name`` into a non-empty list of strings.
+
+    Accepts a comma-separated string (``"cam_a,cam_b"``) or a sequence of
+    strings (tuples/lists). Whitespace is stripped; empty entries are
+    dropped. Raises ``TypeError`` for unsupported input types and
+    ``ValueError`` when the normalized list is empty.
+    """
+    if isinstance(camera_name, str):
+        cams = [c.strip() for c in camera_name.split(",") if c.strip()]
+    elif isinstance(camera_name, (list | tuple)):
+        cams = [str(c).strip() for c in camera_name if str(c).strip()]
+    else:
+        raise TypeError(f"camera_name must be str or sequence[str], got {type(camera_name).__name__}")
+    if not cams:
+        raise ValueError("camera_name resolved to an empty list.")
+    return cams
+
+
 def _convert_nested_dict(d):
     result = {}
     for k, v in d.items():
@@ -153,17 +172,20 @@ class _LazyAsyncVectorEnv:
         env_fns: list[Callable],
         observation_space=None,
         action_space=None,
+        metadata=None,
     ):
         self._env_fns = env_fns
         self._env: gym.vector.AsyncVectorEnv | None = None
         self.num_envs = len(env_fns)
-        if observation_space is not None and action_space is not None:
+        if observation_space is not None and action_space is not None and metadata is not None:
             self.observation_space = observation_space
             self.action_space = action_space
+            self.metadata = metadata
         else:
             tmp = env_fns[0]()
             self.observation_space = tmp.observation_space
             self.action_space = tmp.action_space
+            self.metadata = tmp.metadata
             tmp.close()
         self.single_observation_space = self.observation_space
         self.single_action_space = self.action_space
@@ -172,6 +194,10 @@ class _LazyAsyncVectorEnv:
         if self._env is None:
             self._env = gym.vector.AsyncVectorEnv(self._env_fns, context="forkserver", shared_memory=True)
 
+    @property
+    def unwrapped(self):
+        return self
+
     def reset(self, **kwargs):
         self._ensure()
         return self._env.reset(**kwargs)

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/model/kinematics.py

@@ -12,8 +12,19 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
 import numpy as np
 
+from lerobot.utils.import_utils import _placo_available, require_package
+
+if TYPE_CHECKING or _placo_available:
+    import placo  # type: ignore[import-not-found]
+else:
+    placo = None
+
 
 class RobotKinematics:
     """Robot kinematics using placo library for forward and inverse kinematics."""
@@ -32,13 +43,7 @@ class RobotKinematics:
             target_frame_name (str): Name of the end-effector frame in the URDF
             joint_names (list[str] | None): List of joint names to use for the kinematics solver
         """
-        try:
-            import placo  # type: ignore[import-not-found] # C++ library with Python bindings, no type stubs available. TODO: Create stub file or request upstream typing support.
-        except ImportError as e:
-            raise ImportError(
-                "placo is required for RobotKinematics. "
-                "Please install the optional dependencies of `kinematics` in the package."
-            ) from e
+        require_package("placo", extra="placo-dep")
 
         self.robot = placo.RobotWrapper(urdf_path)
         self.solver = placo.KinematicsSolver(self.robot)

src/lerobot/motors/damiao/damiao.py

@@ -24,7 +24,7 @@ from functools import cached_property
 from typing import TYPE_CHECKING, Any, TypedDict
 
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.import_utils import _can_available
+from lerobot.utils.import_utils import _can_available, require_package
 
 if TYPE_CHECKING or _can_available:
     import can
@@ -111,6 +111,7 @@ class DamiaoMotorsBus(MotorsBusBase):
             bitrate: Nominal bitrate in bps (default: 1000000 = 1 Mbps)
             data_bitrate: Data bitrate for CAN FD in bps (default: 5000000 = 5 Mbps), ignored if use_can_fd is False
         """
+        require_package("python-can", extra="damiao", import_name="can")
         super().__init__(port, motors, calibration)
         self.port = port
         self.can_interface = can_interface

src/lerobot/motors/motors_bus.py

@@ -356,8 +356,8 @@ class SerialMotorsBus(MotorsBusBase):
         motors: dict[str, Motor],
         calibration: dict[str, MotorCalibration] | None = None,
     ):
-        require_package("pyserial", extra="hardware", import_name="serial")
-        require_package("deepdiff", extra="hardware")
+        require_package("pyserial", extra="pyserial-dep", import_name="serial")
+        require_package("deepdiff", extra="deepdiff-dep")
         super().__init__(port, motors, calibration)
 
         self.port_handler: PortHandler

src/lerobot/motors/robstride/robstride.py

@@ -23,12 +23,12 @@ from types import SimpleNamespace
 from typing import TYPE_CHECKING, Any, TypedDict
 
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.import_utils import _can_available
+from lerobot.utils.import_utils import _can_available, require_package
 
 if TYPE_CHECKING or _can_available:
     import can
 else:
-    can = SimpleNamespace(Message=object, interface=None)
+    can = SimpleNamespace(Message=object, interface=None, BusABC=object)
 import numpy as np
 
 from lerobot.utils.errors import DeviceNotConnectedError
@@ -106,6 +106,7 @@ class RobstrideMotorsBus(MotorsBusBase):
             bitrate: Nominal bitrate in bps (default: 1000000 = 1 Mbps)
             data_bitrate: Data bitrate for CAN FD in bps (default: 5000000 = 5 Mbps), ignored if use_can_fd is False
         """
+        require_package("python-can", extra="robstride", import_name="can")
         super().__init__(port, motors, calibration)
         self.port = port
         self.can_interface = can_interface

src/lerobot/optim/schedulers.py

@@ -18,14 +18,21 @@ import logging
 import math
 from dataclasses import asdict, dataclass
 from pathlib import Path
+from typing import TYPE_CHECKING
 
 import draccus
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LambdaLR, LRScheduler
 
 from lerobot.utils.constants import SCHEDULER_STATE
+from lerobot.utils.import_utils import _diffusers_available, require_package
 from lerobot.utils.io_utils import deserialize_json_into_object, write_json
 
+if TYPE_CHECKING or _diffusers_available:
+    from diffusers.optimization import get_scheduler
+else:
+    get_scheduler = None
+
 
 @dataclass
 class LRSchedulerConfig(draccus.ChoiceRegistry, abc.ABC):
@@ -47,10 +54,7 @@ class DiffuserSchedulerConfig(LRSchedulerConfig):
     num_warmup_steps: int | None = None
 
     def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
-        from lerobot.utils.import_utils import require_package
-
         require_package("diffusers", extra="diffusion")
-        from diffusers.optimization import get_scheduler
 
         kwargs = {**asdict(self), "num_training_steps": num_training_steps, "optimizer": optimizer}
         return get_scheduler(**kwargs)

src/lerobot/policies/__init__.py

@@ -20,12 +20,14 @@ from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig as M
 from .pi0.configuration_pi0 import PI0Config as PI0Config
 from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
 from .pi05.configuration_pi05 import PI05Config as PI05Config
+from .pi052.configuration_pi052 import PI052Config as PI052Config
 from .pretrained import PreTrainedPolicy as PreTrainedPolicy
 from .rtc import ActionInterpolator as ActionInterpolator
 from .sac.configuration_sac import SACConfig as SACConfig
 from .sac.reward_model.configuration_classifier import RewardClassifierConfig as RewardClassifierConfig
 from .sarm.configuration_sarm import SARMConfig as SARMConfig
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
+from .smolvla2.configuration_smolvla2 import SmolVLA2Config as SmolVLA2Config
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .utils import make_robot_action, prepare_observation_for_inference
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
@@ -45,10 +47,12 @@ __all__ = [
     "PI0Config",
     "PI0FastConfig",
     "PI05Config",
+    "PI052Config",
     "RewardClassifierConfig",
     "SACConfig",
     "SARMConfig",
     "SmolVLAConfig",
+    "SmolVLA2Config",
     "TDMPCConfig",
     "VQBeTConfig",
     "WallXConfig",

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -23,6 +23,7 @@ TODO(alexander-soare):
 import math
 from collections import deque
 from collections.abc import Callable
+from typing import TYPE_CHECKING
 
 import einops
 import numpy as np
@@ -32,6 +33,14 @@ import torchvision
 from torch import Tensor, nn
 
 from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
+from lerobot.utils.import_utils import _diffusers_available, require_package
+
+if TYPE_CHECKING or _diffusers_available:
+    from diffusers.schedulers.scheduling_ddim import DDIMScheduler
+    from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
+else:
+    DDIMScheduler = None
+    DDPMScheduler = None
 
 from ..pretrained import PreTrainedPolicy
 from ..utils import (
@@ -64,6 +73,7 @@ class DiffusionPolicy(PreTrainedPolicy):
             dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
                 that they will be passed with a call to `load_state_dict` before the policy is used.
         """
+        require_package("diffusers", extra="diffusion")
         super().__init__(config)
         config.validate_features()
         self.config = config
@@ -155,11 +165,7 @@ def _make_noise_scheduler(name: str, **kwargs: dict):
     Factory for noise scheduler instances of the requested type. All kwargs are passed
     to the scheduler.
     """
-    from lerobot.utils.import_utils import require_package
-
     require_package("diffusers", extra="diffusion")
-    from diffusers.schedulers.scheduling_ddim import DDIMScheduler
-    from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
 
     if name == "DDPM":
         return DDPMScheduler(**kwargs)

src/lerobot/policies/factory.py

@@ -128,6 +128,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from .pi05.modeling_pi05 import PI05Policy
 
         return PI05Policy
+    elif name == "pi052":
+        from .pi052.modeling_pi052 import PI052Policy
+
+        return PI052Policy
     elif name == "sac":
         from .sac.modeling_sac import SACPolicy
 
@@ -140,6 +144,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from .smolvla.modeling_smolvla import SmolVLAPolicy
 
         return SmolVLAPolicy
+    elif name == "smolvla2":
+        from .smolvla2.modeling_smolvla2 import SmolVLA2Policy
+
+        return SmolVLA2Policy
     elif name == "sarm":
         from .sarm.modeling_sarm import SARMRewardModel
 
@@ -196,10 +204,18 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return PI0Config(**kwargs)
     elif policy_type == "pi05":
         return PI05Config(**kwargs)
+    elif policy_type == "pi052":
+        from .pi052.configuration_pi052 import PI052Config
+
+        return PI052Config(**kwargs)
     elif policy_type == "sac":
         return SACConfig(**kwargs)
     elif policy_type == "smolvla":
         return SmolVLAConfig(**kwargs)
+    elif policy_type == "smolvla2":
+        from .smolvla2.configuration_smolvla2 import SmolVLA2Config
+
+        return SmolVLA2Config(**kwargs)
     elif policy_type == "reward_classifier":
         return RewardClassifierConfig(**kwargs)
     elif policy_type == "groot":
@@ -236,6 +252,12 @@ class ProcessorConfigKwargs(TypedDict, total=False):
     preprocessor_overrides: dict[str, Any] | None
     postprocessor_overrides: dict[str, Any] | None
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None
+    # Optional: HF Hub repo id of the dataset the policy is being
+    # trained on. Used by policies that auto-fit pieces of their
+    # preprocessing (e.g. pi052's FAST action tokenizer per
+    # Pertsch et al. 2025 [64], π0.5 §III.C). When omitted, those
+    # policies fall back to their universal pre-fitted tokenizers.
+    dataset_repo_id: str | None
 
 
 def make_pre_post_processors(
@@ -362,6 +384,22 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
+    elif policy_cfg.type == "pi052":
+        # NOTE: PI052Config subclasses PI05Config, so this branch MUST
+        # come before the PI05Config isinstance check below (otherwise
+        # pi052 would silently pick up π0.5's processor).
+        from .pi052.processor_pi052 import make_pi052_pre_post_processors
+
+        processors = make_pi052_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            # ``dataset_repo_id`` flows in via kwargs when FAST CE is
+            # enabled — the train loop sets it from ``--dataset.repo_id``.
+            # When ``None``, ``make_pi052_pre_post_processors`` skips
+            # the auto-fit and uses the universal tokenizer.
+            dataset_repo_id=kwargs.get("dataset_repo_id"),
+        )
+
     elif isinstance(policy_cfg, PI05Config):
         from .pi05.processor_pi05 import make_pi05_pre_post_processors
 
@@ -386,6 +424,17 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
+    elif policy_cfg.type == "smolvla2":
+        # NOTE: SmolVLA2Config subclasses SmolVLAConfig, so this branch
+        # MUST come before the SmolVLAConfig isinstance check below
+        # (otherwise SmolVLA2 would silently pick up SmolVLA's processor).
+        from .smolvla2.processor_smolvla2 import make_smolvla2_pre_post_processors
+
+        processors = make_smolvla2_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
     elif isinstance(policy_cfg, SmolVLAConfig):
         from .smolvla.processor_smolvla import make_smolvla_pre_post_processors
 

src/lerobot/policies/groot/modeling_groot.py

@@ -43,6 +43,7 @@ from torch import Tensor
 
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.utils.constants import ACTION, OBS_IMAGES
+from lerobot.utils.import_utils import require_package
 
 from ..pretrained import PreTrainedPolicy
 from .configuration_groot import GrootConfig
@@ -59,6 +60,7 @@ class GrootPolicy(PreTrainedPolicy):
 
     def __init__(self, config: GrootConfig, **kwargs):
         """Initialize Groot policy wrapper."""
+        require_package("transformers", extra="groot")
         super().__init__(config)
         config.validate_features()
         self.config = config

src/lerobot/policies/multi_task_dit/modeling_multi_task_dit.py

@@ -36,7 +36,7 @@ import torch.nn.functional as F  # noqa: N812
 import torchvision
 from torch import Tensor
 
-from lerobot.utils.import_utils import _transformers_available
+from lerobot.utils.import_utils import _diffusers_available, _transformers_available, require_package
 
 from .configuration_multi_task_dit import MultiTaskDiTConfig
 
@@ -46,6 +46,13 @@ if TYPE_CHECKING or _transformers_available:
 else:
     CLIPTextModel = None
     CLIPVisionModel = None
+
+if TYPE_CHECKING or _diffusers_available:
+    from diffusers.schedulers.scheduling_ddim import DDIMScheduler
+    from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
+else:
+    DDIMScheduler = None
+    DDPMScheduler = None
 from lerobot.utils.constants import (
     ACTION,
     OBS_IMAGES,
@@ -65,6 +72,8 @@ class MultiTaskDiTPolicy(PreTrainedPolicy):
     name = "multi_task_dit"
 
     def __init__(self, config: MultiTaskDiTConfig, **kwargs):
+        require_package("transformers", extra="multi_task_dit")
+        require_package("diffusers", extra="multi_task_dit")
         super().__init__(config)
         config.validate_features()
         self.config = config
@@ -643,12 +652,6 @@ class DiffusionObjective(nn.Module):
             "prediction_type": config.prediction_type,
         }
 
-        from lerobot.utils.import_utils import require_package
-
-        require_package("diffusers", extra="multi_task_dit")
-        from diffusers.schedulers.scheduling_ddim import DDIMScheduler
-        from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
-
         if config.noise_scheduler_type == "DDPM":
             self.noise_scheduler: DDPMScheduler | DDIMScheduler = DDPMScheduler(**scheduler_kwargs)
         elif config.noise_scheduler_type == "DDIM":

src/lerobot/policies/pi0/modeling_pi0.py

@@ -26,7 +26,7 @@ import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
 
-from lerobot.utils.import_utils import _transformers_available
+from lerobot.utils.import_utils import _transformers_available, require_package
 
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _transformers_available:
@@ -947,6 +947,7 @@ class PI0Policy(PreTrainedPolicy):
         Args:
             config: Policy configuration class instance.
         """
+        require_package("transformers", extra="pi")
         super().__init__(config)
         config.validate_features()
         self.config = config

src/lerobot/policies/pi05/modeling_pi05.py

@@ -26,7 +26,7 @@ import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
 
-from lerobot.utils.import_utils import _transformers_available
+from lerobot.utils.import_utils import _transformers_available, require_package
 
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _transformers_available:
@@ -918,6 +918,7 @@ class PI05Policy(PreTrainedPolicy):
         Args:
             config: Policy configuration class instance.
         """
+        require_package("transformers", extra="pi")
         super().__init__(config)
         config.validate_features()
         self.config = config

src/lerobot/policies/pi052/__init__.py

@@ -0,0 +1,42 @@
+# 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.
+
+"""π0.5 v2 — full reproduction of the π0.5 paper's hierarchical
+inference recipe on lerobot.
+
+Extends :class:`lerobot.policies.pi05.PI05Policy` with:
+
+* recipe-driven training (PR 1's :class:`RenderMessagesStep`),
+* PaliGemma ``lm_head`` cross-entropy on supervised subtask spans
+  (the "high-level subtask prediction" of the paper, §IV.D),
+* AR text generation at inference (:meth:`PI052Policy.select_message`),
+* per-component prompt dropout (Pi 0.7 §V.E) for regularising the
+  text head against missing context at inference.
+
+See ``src/lerobot/configs/recipes/hirobot.yaml`` for the
+canonical training recipe and
+``examples/training/pi052_hirobot.slurm`` for the launcher.
+"""
+
+from .configuration_pi052 import PI052Config
+from .modeling_pi052 import PI052Policy
+from .processor_pi052 import make_pi052_pre_post_processors
+from .text_processor_pi052 import PI052TextTokenizerStep
+
+__all__ = [
+    "PI052Config",
+    "PI052Policy",
+    "PI052TextTokenizerStep",
+    "make_pi052_pre_post_processors",
+]

src/lerobot/policies/pi052/configuration_pi052.py

@@ -0,0 +1,172 @@
+# 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.
+
+"""π0.5 v2 (with text head) — reproduction of the π0.5 paper's
+hierarchical inference recipe.
+
+Same architecture as the existing ``PI05Policy`` (PaliGemma 2B VLM +
+~300M Gemma action expert, joint training with FAST tokens during
+pre-train and flow matching during post-train), but with the
+PaliGemma ``lm_head`` re-enabled so the same model can be supervised
+to predict both:
+
+  * **subtask strings** at the high level (cross-entropy on the LM
+    head), and
+  * **action chunks** at the low level (flow matching on the
+    action-expert tokens).
+
+This is the dual-head co-training pattern from the paper:
+
+    L = H(x, f_θ_text) + α * ‖ω - a - f_θ_action(a_τ, o, ℓ)‖²
+
+with α = 10.0 per § IV.D of arxiv:2504.16054. The π0.5 model splits
+inference into a text-prediction step followed by an action-prediction
+step, which mirrors what ``SmolVLA2Runtime`` already does on a
+SmolVLM2 backbone.
+"""
+
+from dataclasses import dataclass
+
+from lerobot.configs import PreTrainedConfig
+
+from ..pi05.configuration_pi05 import PI05Config
+
+
+@PreTrainedConfig.register_subclass("pi052")
+@dataclass
+class PI052Config(PI05Config):
+    """π0.5 with the PaliGemma LM head re-enabled for subtask prediction.
+
+    See ``SmolVLA2Config`` for the analogous SmolVLM2-backed dual-head
+    config. Same recipe-driven training surface; the only differences
+    are which backbone the policy uses (PaliGemma here vs SmolVLM2
+    there) and the default loss-weight scale (paper §IV.D uses
+    ``α=10`` between the two heads, which we encode as
+    ``flow_loss_weight=10, text_loss_weight=1``).
+    """
+
+    # Recipe / language stack ---------------------------------------------
+    recipe_path: str | None = "recipes/hirobot.yaml"
+    """Path (absolute or relative to ``src/lerobot/configs/``) to a
+    ``TrainingRecipe`` YAML. Defaults to the canonical Hi-Robot blend
+    shipped alongside this policy. Set to ``None`` to disable recipe
+    rendering and fall back to π0.5's single-task ``Task: ... Action:``
+    prompt path (unannotated datasets keep working that way)."""
+
+    apply_chat_template: bool = False
+    """PaliGemma is *not* chat-pretrained — its tokenizer doesn't ship a
+    chat template. So unlike SmolVLA2 we don't apply one. The recipe
+    renderer's output is concatenated as a plain prefix + assistant
+    suffix instead, mirroring how the π0.5 paper's high-level inference
+    samples text auto-regressively after the prefix."""
+
+    # Loss weights --------------------------------------------------------
+    # Paper §IV.D: total = H(text) + α * MSE(flow), α = 10. We split
+    # the same total into two configurable knobs so individual scaling
+    # is recoverable.
+    text_loss_weight: float = 1.0
+    """Weight on the LM-head cross-entropy term. Set to ``0`` to disable
+    text training entirely (reverts to flow-only / π0.5 behaviour)."""
+
+    flow_loss_weight: float = 10.0
+    """Weight on the action-expert flow-matching term. Default ``10.0``
+    matches the paper's α."""
+
+    # Backbone training ---------------------------------------------------
+    unfreeze_lm_head: bool = True
+    """Whether to keep the PaliGemma ``lm_head`` unfrozen for fine-tuning.
+    The existing ``PI05Policy`` zeroes / freezes the head on load
+    because it never reads from it. Must be ``True`` for π0.5-style
+    hierarchical inference."""
+
+    # Per-component prompt dropout (Pi0.7 §V.E) ---------------------------
+    # Same regulariser surface as SmolVLA2: randomly drop non-target
+    # context messages so the LM head learns to handle missing /
+    # stale plan / memory at inference. Defaults to 0.0 so behaviour
+    # is identical until explicitly enabled.
+    plan_dropout_prob: float = 0.0
+    memory_dropout_prob: float = 0.0
+    subtask_dropout_prob: float = 0.0
+
+    # FAST discrete-action supervision — paper §III.B-C ------------------
+    # When enabled, actions are *also* tokenised via the FAST tokenizer
+    # ("physical-intelligence/fast") and supervised with cross-entropy
+    # on the PaliGemma LM head — exactly as in the paper's pre-training
+    # objective (Eq. 1 mixes FAST CE + flow MSE + subtask CE). The
+    # ActionTokenizerProcessorStep is wired into the preprocessor
+    # pipeline when this flag is set; the loss is computed in
+    # PI052Policy.forward.
+    enable_fast_action_loss: bool = True
+    """If True, tokenise actions with the FAST tokenizer and add a
+    cross-entropy loss on the LM head. On by default to match the
+    π0.5 paper's three-loss objective (text CE + FAST CE + flow MSE,
+    §III.B-C Eq. 1). Set to False if you only want the
+    post-training-style flow + text recipe."""
+
+    action_tokenizer_name: str = "physical-intelligence/fast"
+    """HF identifier for the FAST action tokenizer."""
+
+    max_action_tokens: int = 256
+    """Maximum number of FAST tokens per action chunk."""
+
+    fast_skip_tokens: int = 128
+    """Number of low-vocab tokens the FAST tokenizer skips to avoid
+    collisions with PaliGemma's text vocabulary."""
+
+    fast_action_loss_weight: float = 1.0
+    """Weight on the FAST-action-token CE loss. Paper §III.C uses 1.0."""
+
+    auto_fit_fast_tokenizer: bool = False
+    """If True, the processor factory checks ``fast_tokenizer_cache_dir``
+    for a previously-fitted tokenizer keyed on ``(dataset_repo_id,
+    base_tokenizer_name, fit_samples)``. On cache miss, it loads
+    ``action_tokenizer_name`` as a base, samples
+    ``fast_tokenizer_fit_samples`` action chunks from the dataset, runs
+    ``.fit()``, saves the result, and uses *that* fitted path as the
+    actual tokenizer. Pertsch et al. 2025 (FAST paper [64], π0.5 §III.C)
+    explicitly recommend per-dataset fitting for best compression.
+
+    Off by default because the fit requires a separate pre-training
+    pass over the dataset (~1-2 min on a medium dataset) and depends
+    on the FAST tokenizer snapshot having a ``.fit()`` method. Opt in
+    when you want paper-faithful compression; leave off to fall back
+    on the universal ``physical-intelligence/fast`` codebook."""
+
+    fast_tokenizer_cache_dir: str = "~/.cache/lerobot/fast_tokenizers"
+    """Where fitted FAST tokenizers are stored. ``~`` expands."""
+
+    fast_tokenizer_fit_samples: int = 1024
+    """Number of action chunks to sample for the fit. The FAST paper uses
+    a few thousand; 1024 is a reasonable default for medium datasets."""
+
+    # Knowledge insulation — paper §III.B --------------------------------
+    # When enabled, gradients from the action expert's flow loss are
+    # blocked from flowing back into the VLM's K/V projections. This
+    # prevents the action loss from over-fitting the language backbone
+    # to robot-specific features. Implemented in ``modeling_pi052`` as
+    # a per-instance monkey-patch on ``paligemma_with_expert.forward``
+    # that splits queries into VLM and action halves and ``.detach()``-s
+    # the VLM K/V tensors used in the action-half's attention.
+    knowledge_insulation: bool = False
+    """If True, route every transformer layer through the KI
+    attention path that blocks action→VLM gradient flow on K/V."""
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        # Backbone needs gradients flowing through the text head when
+        # we're training it. Override the π0.5 default
+        # (``train_expert_only=True``) unless the user explicitly opts
+        # out of text training via ``text_loss_weight=0``.
+        if self.text_loss_weight > 0 and self.unfreeze_lm_head:
+            self.train_expert_only = False

src/lerobot/policies/pi052/fit_fast_tokenizer.py

@@ -0,0 +1,204 @@
+# 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.
+
+"""Dataset-specific FAST action tokenizer fitting.
+
+The published ``physical-intelligence/fast`` tokenizer is a *universal*
+codebook fitted on a heterogeneous mix of robot datasets. Per Pertsch
+et al. 2025 (the FAST paper, [64] in the π0.5 paper) and §III.C of
+π0.5 itself, the recommended practice is to **finetune the tokenizer on
+your specific dataset's action distribution** before training the
+policy — same way one would adapt a language tokenizer to a domain
+corpus. Without this finetune step, action sequences from your robot
+may require more tokens per chunk than necessary, lowering effective
+compression and slowing convergence of the action-CE loss.
+
+This module provides a single utility, :func:`fit_fast_tokenizer`,
+that does the finetune. The training entry point invokes it
+automatically when the policy's ``enable_fast_action_loss`` and
+``auto_fit_fast_tokenizer`` flags are both ``True`` and no cached
+fitted tokenizer is found at ``fast_tokenizer_cache_dir``.
+
+The fitted tokenizer is saved to
+``{cache_dir}/{dataset_hash}_{base_hash}/`` so successive training
+runs over the same dataset re-use it.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import logging
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+logger = logging.getLogger(__name__)
+
+
+def _dataset_signature(
+    dataset_repo_id: str,
+    base_tokenizer_name: str,
+    n_samples: int,
+    chunk_size: int,
+) -> str:
+    """Deterministic short hash for naming the cache directory.
+
+    Keys on (dataset, base tokenizer, sample count, chunk size) so any
+    of those changing re-runs the fit. ``chunk_size`` matters because
+    the tokenizer is fit on chunks of that length.
+    """
+    h = hashlib.sha256()
+    h.update(dataset_repo_id.encode("utf-8"))
+    h.update(b"\0")
+    h.update(base_tokenizer_name.encode("utf-8"))
+    h.update(b"\0")
+    h.update(str(n_samples).encode("utf-8"))
+    h.update(b"\0")
+    h.update(str(chunk_size).encode("utf-8"))
+    return h.hexdigest()[:16]
+
+
+def fit_fast_tokenizer(
+    *,
+    dataset_repo_id: str,
+    cache_dir: str | Path,
+    base_tokenizer_name: str = "physical-intelligence/fast",
+    n_samples: int = 1024,
+    chunk_size: int = 50,
+    seed: int = 42,
+) -> str:
+    """Fit a FAST tokenizer on a LeRobot dataset's action distribution.
+
+    Args:
+        dataset_repo_id: HF Hub repo id of the LeRobotDataset to fit on.
+        cache_dir: Directory under which to save (and look up) fitted
+            tokenizers. The actual save path is
+            ``{cache_dir}/{signature}``.
+        base_tokenizer_name: HF identifier for the base FAST tokenizer
+            to finetune from. ``physical-intelligence/fast`` is the
+            universal one.
+        n_samples: Number of action chunks to sample for the fit. The
+            FAST paper uses a few thousand; ``1024`` is a good default
+            for medium datasets.
+        chunk_size: Length of each action chunk (matches
+            ``policy.chunk_size``). The FAST tokenizer is fit on
+            sequences of this length.
+        seed: RNG seed for sample selection.
+
+    Returns:
+        The local path to the fitted tokenizer. Passed directly to
+        ``--policy.action_tokenizer_name`` for the training run.
+
+    Raises:
+        ImportError: If the ``transformers`` library doesn't expose
+            ``AutoProcessor`` or the FAST tokenizer doesn't have a
+            ``.fit()`` method (then you're on an older FAST snapshot —
+            update to the current published model).
+        FileNotFoundError: If the dataset can't be loaded.
+    """
+    cache_dir = Path(cache_dir)
+    sig = _dataset_signature(dataset_repo_id, base_tokenizer_name, n_samples, chunk_size)
+    out_dir = cache_dir / sig
+
+    if out_dir.exists() and (out_dir / "preprocessor_config.json").exists():
+        logger.info(
+            "FAST tokenizer cache hit: %s — re-using fitted tokenizer for "
+            "dataset=%s base=%s n_samples=%d",
+            out_dir, dataset_repo_id, base_tokenizer_name, n_samples,
+        )
+        return str(out_dir)
+
+    logger.info(
+        "FAST tokenizer cache miss — fitting on dataset=%s "
+        "base=%s n_samples=%d chunk_size=%d → %s",
+        dataset_repo_id, base_tokenizer_name, n_samples, chunk_size, out_dir,
+    )
+
+    from transformers import AutoProcessor  # noqa: PLC0415
+
+    from lerobot.datasets.lerobot_dataset import LeRobotDataset  # noqa: PLC0415
+
+    # Stream a single episode's worth of action chunks at a time so
+    # we don't blow memory on huge datasets. Random episode +
+    # random start offset gives a reasonable spread.
+    rng = np.random.default_rng(seed)
+    actions_buf: list[np.ndarray] = []
+
+    # Load just the metadata first to know episode boundaries.
+    ds_meta_only = LeRobotDataset(dataset_repo_id, episodes=[0])
+    num_episodes = ds_meta_only.meta.total_episodes
+    del ds_meta_only
+
+    samples_per_episode = max(1, n_samples // max(num_episodes, 1))
+    collected = 0
+    eps_visited = 0
+    for ep_idx in rng.permutation(num_episodes):
+        if collected >= n_samples:
+            break
+        ep_idx = int(ep_idx)
+        try:
+            ds = LeRobotDataset(dataset_repo_id, episodes=[ep_idx])
+        except Exception as exc:  # noqa: BLE001
+            logger.warning("FAST fit: skipping episode %d: %s", ep_idx, exc)
+            continue
+        if len(ds) < chunk_size:
+            continue
+        # Sample ``samples_per_episode`` start indices uniformly within
+        # the episode.
+        starts = rng.integers(0, len(ds) - chunk_size + 1, size=samples_per_episode)
+        for s in starts:
+            try:
+                chunk_actions = np.stack(
+                    [
+                        np.asarray(ds[int(s) + j]["action"].cpu().numpy())
+                        for j in range(chunk_size)
+                    ],
+                    axis=0,
+                )
+            except Exception as exc:  # noqa: BLE001
+                logger.debug("FAST fit: chunk at ep=%d s=%d failed: %s", ep_idx, s, exc)
+                continue
+            actions_buf.append(chunk_actions)
+            collected += 1
+            if collected >= n_samples:
+                break
+        eps_visited += 1
+
+    if not actions_buf:
+        raise RuntimeError(
+            f"FAST fit collected zero action chunks from {dataset_repo_id!r}. "
+            "Check that the dataset has an ``action`` column and chunks of "
+            f"length ≥ {chunk_size}."
+        )
+
+    actions = np.stack(actions_buf, axis=0)  # (N, H, D)
+    logger.info(
+        "FAST fit: collected %d chunks of shape %s from %d episodes",
+        actions.shape[0], actions.shape[1:], eps_visited,
+    )
+
+    base = AutoProcessor.from_pretrained(base_tokenizer_name, trust_remote_code=True)
+    if not hasattr(base, "fit"):
+        raise ImportError(
+            f"Base FAST tokenizer {base_tokenizer_name!r} has no ``.fit()`` "
+            "method — your transformers / model snapshot is too old. Update "
+            "to the current ``physical-intelligence/fast`` revision."
+        )
+
+    fitted = base.fit(actions)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    fitted.save_pretrained(str(out_dir))
+    logger.info("FAST fit: saved fitted tokenizer to %s", out_dir)
+    return str(out_dir)

src/lerobot/policies/pi052/modeling_pi052.py

@@ -0,0 +1,821 @@
+# 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.
+
+"""π0.5 v2 policy — dual-head training & hierarchical inference.
+
+A thin subclass of :class:`PI05Policy` that:
+
+* keeps the PaliGemma ``lm_head`` unfrozen during fine-tuning
+  (``PI05Policy`` zeroes / freezes it because it never reads from
+  the head; ``PI052Config.unfreeze_lm_head`` flips that),
+* adds a ``text_loss`` term computed via cross-entropy on
+  ``text_labels`` (built by ``PI052TextTokenizerStep``),
+* adds :meth:`select_message` for AR text generation at inference
+  (the high-level step in the π0.5 paper's two-stage inference loop),
+* combines both losses in :meth:`forward` per Eq. (1) of the paper:
+
+      L = H(x, f_θ_text) + α * ‖ω - a - f_θ_action(...)‖²
+
+  with α controllable via ``config.flow_loss_weight``.
+
+The same multi-rate runtime that drives ``SmolVLA2Runtime`` (see
+``lerobot.policies.smolvla2.inference``) can drive this policy too —
+both expose ``predict_action_chunk`` for the action expert and
+``select_message`` for the LM head.
+"""
+
+from __future__ import annotations
+
+import logging
+import math
+import types
+from typing import Any
+
+import torch
+from torch import Tensor
+from torch.nn import functional as F
+
+from lerobot.utils.constants import OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS
+
+from ..pi05.configuration_pi05 import PI05Config
+from ..pi05.modeling_pi05 import PI05Policy
+from .configuration_pi052 import PI052Config
+
+logger = logging.getLogger(__name__)
+
+
+# ----------------------------------------------------------------------
+# Loss helpers (shared between fused and prefix-only paths)
+# ----------------------------------------------------------------------
+
+
+def _mask_per_sample(per_sample: Tensor, predict_actions_t: Tensor | None) -> Tensor:
+    """Mean over samples where ``predict_actions_t`` is True, else over all."""
+    if predict_actions_t is None:
+        return per_sample.mean()
+    mask = predict_actions_t.to(per_sample.dtype)
+    return (per_sample * mask).sum() / mask.sum().clamp(min=1.0)
+
+
+def _shifted_ce(logits: Tensor, labels: Tensor) -> Tensor:
+    """Next-token CE: hidden at t predicts label at t+1, ignore_index=-100.
+
+    Mean over non-ignored positions across the batch. Returns 0 cleanly
+    when no positions are supervised (clamp(min=1) on the denominator).
+    """
+    shift_logits = logits[:, :-1, :].contiguous()
+    shift_labels = labels[:, 1:].contiguous().long()
+    valid = shift_labels != -100
+    loss = F.cross_entropy(
+        shift_logits.reshape(-1, shift_logits.shape[-1]),
+        shift_labels.reshape(-1),
+        ignore_index=-100,
+        reduction="sum",
+    )
+    return loss / valid.sum().clamp(min=1)
+
+
+def _fast_ce(
+    fast_logits: Tensor,
+    action_tokens: Tensor,
+    action_mask: Tensor,
+    predict_actions_t: Tensor | None,
+) -> Tensor:
+    """FAST-CE with both token-pad masking and per-sample action gating.
+
+    ``action_mask`` is the FAST tokenizer's padding mask; samples whose
+    recipe sets ``predict_actions=False`` (e.g. plan_generation,
+    ask_vqa_*) get *all* their FAST positions masked out via the
+    per-sample gate.
+    """
+    shift_logits = fast_logits[:, :-1, :].contiguous()
+    shift_targets = action_tokens[:, 1:].contiguous().long()
+    shift_valid = action_mask[:, 1:].contiguous().bool()
+    if predict_actions_t is not None:
+        sample_mask = predict_actions_t[:, None].expand_as(shift_valid)
+        shift_valid = shift_valid & sample_mask
+    shift_targets = shift_targets.masked_fill(~shift_valid, -100)
+    return F.cross_entropy(
+        shift_logits.reshape(-1, shift_logits.shape[-1]),
+        shift_targets.reshape(-1),
+        ignore_index=-100,
+        reduction="sum",
+    ) / shift_valid.sum().clamp(min=1)
+
+
+# ----------------------------------------------------------------------
+# Knowledge insulation — ported from pi05_full (branch ``feat/add-pi05``)
+# ----------------------------------------------------------------------
+#
+# Per-layer attention that splits the queries into VLM and action
+# parts, computing attention for action queries with .detach()'d VLM
+# K/V so the action loss's gradient cannot flow back into the VLM's K
+# and V projections. Forward output is bit-equivalent to the standard
+# layer; backward differs only on the path action_loss → VLM K/V.
+
+def _compute_layer_ki(
+    layer_idx,
+    inputs_embeds,
+    attention_mask,
+    position_ids,
+    adarms_cond,
+    paligemma,
+    gemma_expert,
+):
+    from transformers.models.gemma import modeling_gemma  # noqa: PLC0415
+
+    models = [paligemma.model.language_model, gemma_expert.model]
+    query_states, key_states, value_states, gates = [], [], [], []
+
+    vlm_len = inputs_embeds[0].shape[1]
+
+    for i, hidden_states in enumerate(inputs_embeds):
+        layer = models[i].layers[layer_idx]
+        hidden_states, gate = layer.input_layernorm(hidden_states, cond=adarms_cond[i])
+        gates.append(gate)
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
+        q = layer.self_attn.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        k = layer.self_attn.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        v = layer.self_attn.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        query_states.append(q)
+        key_states.append(k)
+        value_states.append(v)
+
+    query_states = torch.cat(query_states, dim=2)
+    key_states = torch.cat(key_states, dim=2)
+    value_states = torch.cat(value_states, dim=2)
+
+    dummy = torch.zeros(
+        query_states.shape[0],
+        query_states.shape[2],
+        query_states.shape[-1],
+        device=query_states.device,
+        dtype=query_states.dtype,
+    )
+    cos, sin = paligemma.model.language_model.rotary_emb(dummy, position_ids)
+    query_states, key_states = modeling_gemma.apply_rotary_pos_emb(
+        query_states, key_states, cos, sin, unsqueeze_dim=1
+    )
+
+    batch_size = query_states.shape[0]
+    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
+
+    # Split queries / K / V at the VLM-vs-action boundary.
+    Q_vlm = query_states[:, :, :vlm_len, :]
+    Q_action = query_states[:, :, vlm_len:, :]
+    K_vlm = key_states[:, :, :vlm_len, :]
+    K_action = key_states[:, :, vlm_len:, :]
+    V_vlm = value_states[:, :, :vlm_len, :]
+    V_action = value_states[:, :, vlm_len:, :]
+
+    # Detach VLM K/V *only* on the path the action queries use.
+    K_vlm_det = K_vlm.detach()
+    V_vlm_det = V_vlm.detach()
+    K_for_vlm = key_states  # full (gradients flow)
+    V_for_vlm = value_states
+    K_for_action = torch.cat([K_vlm_det, K_action], dim=2)
+    V_for_action = torch.cat([V_vlm_det, V_action], dim=2)
+
+    mask_for_vlm = attention_mask[:, :, :vlm_len, :]
+    mask_for_action = attention_mask[:, :, vlm_len:, :]
+
+    att_vlm, _ = modeling_gemma.eager_attention_forward(
+        paligemma.model.language_model.layers[layer_idx].self_attn,
+        Q_vlm, K_for_vlm, V_for_vlm, mask_for_vlm, scaling,
+    )
+    att_action, _ = modeling_gemma.eager_attention_forward(
+        paligemma.model.language_model.layers[layer_idx].self_attn,
+        Q_action, K_for_action, V_for_action, mask_for_action, scaling,
+    )
+    att = torch.cat([att_vlm, att_action], dim=1)
+
+    head_dim = paligemma.model.language_model.layers[layer_idx].self_attn.head_dim
+    att = att.reshape(batch_size, -1, 1 * 8 * head_dim)
+
+    outputs_embeds = []
+    start = 0
+    for i, hidden_states in enumerate(inputs_embeds):
+        layer = models[i].layers[layer_idx]
+        end = start + hidden_states.shape[1]
+        if att.dtype != layer.self_attn.o_proj.weight.dtype:
+            att = att.to(layer.self_attn.o_proj.weight.dtype)
+        out_emb = layer.self_attn.o_proj(att[:, start:end])
+        out_emb = modeling_gemma._gated_residual(hidden_states, out_emb, gates[i])  # noqa: SLF001
+        after_first = out_emb.clone()
+        out_emb, gate = layer.post_attention_layernorm(out_emb.clone(), cond=adarms_cond[i])
+        if layer.mlp.up_proj.weight.dtype == torch.bfloat16:
+            out_emb = out_emb.to(dtype=torch.bfloat16)
+        out_emb = layer.mlp(out_emb)
+        out_emb = modeling_gemma._gated_residual(after_first, out_emb, gate)  # noqa: SLF001
+        outputs_embeds.append(out_emb)
+        start = end
+    return outputs_embeds
+
+
+def _paligemma_forward_ki(
+    self,
+    attention_mask=None,
+    position_ids=None,
+    past_key_values=None,
+    inputs_embeds=None,
+    use_cache=None,
+    adarms_cond=None,
+):
+    """Replacement ``PaliGemmaWithExpertModel.forward`` that routes the
+    dual-expert layer pass through :func:`_compute_layer_ki`.
+
+    Bound onto the model instance when ``config.knowledge_insulation``
+    is True (see ``PI052Policy.__init__``). Single-expert branches
+    (VLM-only or action-only) defer back to the original forward —
+    KI only matters when actions and VLM tokens are forwarded together.
+    """
+    from ..pi05.modeling_pi05 import layernorm_forward  # noqa: PLC0415
+
+    if adarms_cond is None:
+        adarms_cond = [None, None]
+
+    # Single-expert paths: defer to the original forward saved in
+    # PI052Policy.__init__.
+    if inputs_embeds[0] is None or inputs_embeds[1] is None:
+        return self._pi052_orig_forward(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            adarms_cond=adarms_cond,
+        )
+
+    models = [self.paligemma.model.language_model, self.gemma_expert.model]
+    num_layers = self.paligemma.config.text_config.num_hidden_layers
+    use_gc = (
+        hasattr(self.gemma_expert.model, "gradient_checkpointing")
+        and self.gemma_expert.model.gradient_checkpointing
+        and self.training
+    ) or (
+        hasattr(self, "gradient_checkpointing") and self.gradient_checkpointing and self.training
+    )
+
+    for layer_idx in range(num_layers):
+        if use_gc:
+            inputs_embeds = torch.utils.checkpoint.checkpoint(
+                _compute_layer_ki,
+                layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond,
+                use_reentrant=False, preserve_rng_state=False,
+                paligemma=self.paligemma, gemma_expert=self.gemma_expert,
+            )
+        else:
+            inputs_embeds = _compute_layer_ki(
+                layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond,
+                paligemma=self.paligemma, gemma_expert=self.gemma_expert,
+            )
+
+    outputs_embeds = []
+    for i, hidden_states in enumerate(inputs_embeds):
+        out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
+        outputs_embeds.append(out_emb)
+    return [outputs_embeds[0], outputs_embeds[1]], None
+
+
+class PI052Policy(PI05Policy):
+    """π0.5 with the PaliGemma LM head re-enabled."""
+
+    config_class = PI052Config
+    name = "pi052"
+
+    def __init__(self, config: PI052Config, **kwargs: Any) -> None:
+        super().__init__(config, **kwargs)
+        # ``PI05Policy.__init__`` zeroes the PaliGemma ``lm_head`` and
+        # freezes a few terminal layers when ``train_expert_only`` is
+        # the (default) True. We re-enable the head if the user
+        # wants text supervision.
+        if config.text_loss_weight > 0 and config.unfreeze_lm_head:
+            self._unfreeze_lm_head()
+
+        # Knowledge insulation: bind a custom ``forward`` on the
+        # PaliGemmaWithExpertModel instance that uses
+        # :func:`_compute_layer_ki` for the dual-expert layer pass.
+        # The bind is per-instance, so this doesn't leak into stock
+        # ``pi05`` policies that share the same class.
+        if getattr(config, "knowledge_insulation", False):
+            backbone = self.model.paligemma_with_expert
+            backbone._pi052_orig_forward = backbone.forward
+            backbone.forward = types.MethodType(_paligemma_forward_ki, backbone)
+            logger.info(
+                "PI052: knowledge insulation enabled — action→VLM K/V "
+                "gradients are blocked in attention."
+            )
+
+    # ------------------------------------------------------------------
+    # Head unfreeze helper
+    # ------------------------------------------------------------------
+
+    def _unfreeze_lm_head(self) -> None:
+        """Walk the PaliGemma submodules and re-enable gradients on
+        ``lm_head`` + the immediately preceding norm / last text-model
+        layer that ``PI05Policy`` typically freezes."""
+        backbone = self.model.paligemma_with_expert.paligemma
+        if hasattr(backbone, "lm_head"):
+            for p in backbone.lm_head.parameters():
+                p.requires_grad_(True)
+        # The text model's final norm and last transformer block —
+        # mirror SmolVLA2's logic, which finds these dynamically by
+        # the trainable=False parameters that point at the head's
+        # neighbourhood.
+        text_model = getattr(backbone, "model", None)
+        text_model = getattr(text_model, "language_model", text_model)
+        if text_model is None:
+            return
+        norm = getattr(text_model, "norm", None)
+        if norm is not None:
+            for p in norm.parameters():
+                p.requires_grad_(True)
+        layers = getattr(text_model, "layers", None)
+        if isinstance(layers, (list, torch.nn.ModuleList)) and len(layers) > 0:
+            for p in layers[-1].parameters():
+                p.requires_grad_(True)
+
+    # ------------------------------------------------------------------
+    # Forward (dual loss: flow + text)
+    # ------------------------------------------------------------------
+
+    def forward(
+        self,
+        batch: dict[str, Tensor],
+        reduction: str = "mean",
+    ) -> tuple[Tensor, dict]:
+        """Dual-head forward: flow-matching loss + text-CE loss.
+
+        When ``text_labels`` isn't present in the batch (e.g. the
+        recipe wasn't applied), we delegate to ``PI05Policy.forward``
+        unchanged. Otherwise we compute both losses and sum them with
+        ``flow_loss_weight`` / ``text_loss_weight``.
+        """
+        text_labels = batch.get("text_labels")
+        predict_actions_t = batch.get("predict_actions")
+
+        # Fall through to PI05Policy only on fully unannotated batches
+        # (no recipe applied → no routing fields). For recipe-applied
+        # batches we keep control of the loss dispatch even if all
+        # samples are text-only — delegating would silently train flow
+        # on text-only frames (PI05Policy.forward ignores
+        # ``predict_actions``).
+        if (
+            text_labels is None
+            and predict_actions_t is None
+            and not getattr(self.config, "enable_fast_action_loss", False)
+        ):
+            return super().forward(batch, reduction=reduction)
+
+        run_flow = (
+            self.config.flow_loss_weight > 0
+            and (predict_actions_t is None or bool(predict_actions_t.any().item()))
+        )
+        run_text = self.config.text_loss_weight > 0 and text_labels is not None
+
+        loss_dict: dict[str, Any] = {}
+        total: Tensor | None = None
+
+        # Decide which losses fire this step.
+        run_fast = (
+            getattr(self.config, "enable_fast_action_loss", False)
+            and self.config.fast_action_loss_weight > 0
+            and (predict_actions_t is None or bool(predict_actions_t.any().item()))
+        )
+        action_tokens = action_mask = None
+        if run_fast:
+            from lerobot.utils.constants import ACTION_TOKEN_MASK, ACTION_TOKENS  # noqa: PLC0415
+
+            action_tokens = batch.get(ACTION_TOKENS)
+            action_mask = batch.get(ACTION_TOKEN_MASK)
+            if action_tokens is None or action_mask is None:
+                run_fast = False
+
+        # ------------------------------------------------------------
+        # Dispatch: full fusion when flow is active, otherwise the
+        # prefix-only text+FAST helper (no suffix forward needed).
+        #
+        # Full fusion (flow ON):
+        #   ONE backbone forward with prefix=[images, lang, FAST] +
+        #   suffix=[noisy_actions], suffix→FAST attention masked out.
+        #   All three losses computed from slices of the single output.
+        #
+        # Prefix-only fusion (flow OFF, e.g. text-only recipes):
+        #   ONE prefix-only forward, both text + FAST losses computed
+        #   from slices. No suffix forward → cheaper.
+        # ------------------------------------------------------------
+        if run_flow:
+            flow_loss, text_loss, fast_loss = self._compute_all_losses_fused(
+                batch,
+                text_labels=text_labels if run_text else None,
+                action_tokens=action_tokens if run_fast else None,
+                action_mask=action_mask if run_fast else None,
+                predict_actions_t=predict_actions_t,
+            )
+            loss_dict["flow_loss"] = float(flow_loss.detach().item())
+            total = self.config.flow_loss_weight * flow_loss
+            if text_loss is not None:
+                loss_dict["text_loss"] = float(text_loss.detach().item())
+                total = total + self.config.text_loss_weight * text_loss
+            if fast_loss is not None:
+                loss_dict["fast_action_loss"] = float(fast_loss.detach().item())
+                total = total + self.config.fast_action_loss_weight * fast_loss
+        elif run_text or run_fast:
+            text_loss, fast_loss = self._compute_text_and_fast_loss(
+                batch,
+                text_labels=text_labels if run_text else None,
+                action_tokens=action_tokens if run_fast else None,
+                action_mask=action_mask if run_fast else None,
+                predict_actions_t=predict_actions_t,
+            )
+            if text_loss is not None:
+                loss_dict["text_loss"] = float(text_loss.detach().item())
+                weighted = self.config.text_loss_weight * text_loss
+                total = weighted if total is None else total + weighted
+            if fast_loss is not None:
+                loss_dict["fast_action_loss"] = float(fast_loss.detach().item())
+                weighted = self.config.fast_action_loss_weight * fast_loss
+                total = weighted if total is None else total + weighted
+
+        if total is None:
+            # Both flow and text disabled — make this an obvious bug
+            # rather than a silent zero loss.
+            raise RuntimeError(
+                "PI052Policy.forward: both flow_loss_weight and "
+                "text_loss_weight are 0 (or text_labels missing) — "
+                "nothing to train."
+            )
+
+        loss_dict["loss"] = float(total.detach().item()) if total.dim() == 0 else float("nan")
+        if reduction == "none":
+            return total.expand(batch[OBS_LANGUAGE_TOKENS].shape[0]), loss_dict
+        return total, loss_dict
+
+    # ------------------------------------------------------------------
+    # Text loss
+    # ------------------------------------------------------------------
+
+    def _compute_all_losses_fused(
+        self,
+        batch: dict[str, Tensor],
+        text_labels: Tensor | None,
+        action_tokens: Tensor | None,
+        action_mask: Tensor | None,
+        predict_actions_t: Tensor | None = None,
+    ) -> tuple[Tensor, Tensor | None, Tensor | None]:
+        """Full fusion: flow + text + FAST in ONE backbone forward.
+
+        Builds:
+
+            prefix = [images, language, FAST (when provided)]
+            suffix = [noisy_actions]   (action expert via gemma_expert)
+
+        Then overrides the unified 2D attention mask to *explicitly*
+        zero out ``suffix → FAST`` attention. Without this override
+        the action expert would attend to the discrete FAST tokens
+        and trivially decode them back to the same continuous
+        actions it's supposed to predict via flow matching — the
+        whole training signal collapses.
+
+        Both prefix_out and suffix_out are captured from the same
+        forward. From prefix_out we slice the language and FAST
+        token positions and compute their CE losses. From suffix_out
+        we run the existing flow path (action_out_proj → MSE).
+
+        Returns ``(flow_loss, text_loss, fast_loss)`` where text/fast
+        can be ``None`` when the caller didn't supply the
+        corresponding inputs.
+        """
+        from lerobot.utils.constants import ACTION  # noqa: PLC0415
+
+        from ..pi05.modeling_pi05 import make_att_2d_masks  # noqa: PLC0415
+
+        # ---- preamble (mirrors PI05Pytorch.forward) ------------------
+        actions = self.model.prepare_action(batch)
+        noise = self.model.sample_noise(actions.shape, actions.device)
+        time = self.model.sample_time(actions.shape[0], actions.device)
+        time_expanded = time[:, None, None]
+        x_t = time_expanded * noise + (1 - time_expanded) * actions
+        u_t = noise - actions
+
+        # ---- prefix: images + language + (optional FAST) -------------
+        images, img_masks = self.model._preprocess_images(batch)
+        lang_tokens = batch[OBS_LANGUAGE_TOKENS]
+        lang_masks = batch[OBS_LANGUAGE_ATTENTION_MASK]
+        prefix_embs, prefix_pad, prefix_att = self.model.embed_prefix(
+            images, img_masks, lang_tokens, lang_masks
+        )
+        non_fast_prefix_len = prefix_embs.shape[1]  # images + language only
+
+        fast_len = 0
+        if action_tokens is not None and action_mask is not None:
+            emb_dim = prefix_embs.shape[-1]
+            fast_emb = self.model.paligemma_with_expert.embed_language_tokens(action_tokens)
+            fast_emb = fast_emb * math.sqrt(emb_dim)
+            fast_len = action_tokens.shape[1]
+            ones_att = torch.ones(
+                (action_tokens.shape[0], fast_len),
+                dtype=torch.bool,
+                device=prefix_embs.device,
+            )
+            prefix_embs = torch.cat([prefix_embs, fast_emb], dim=1)
+            prefix_pad = torch.cat([prefix_pad, action_mask.to(prefix_pad.dtype)], dim=1)
+            prefix_att = torch.cat([prefix_att, ones_att], dim=1)
+
+        # ---- suffix: noisy actions ----------------------------------
+        suffix_embs, suffix_pad, suffix_att, adarms_cond = self.model.embed_suffix(x_t, time)
+
+        # ---- bf16 alignment (mirrors PI05Pytorch.forward) -----------
+        first_layer = (
+            self.model.paligemma_with_expert.paligemma.model.language_model.layers[0]
+        )
+        if first_layer.self_attn.q_proj.weight.dtype == torch.bfloat16:
+            suffix_embs = suffix_embs.to(dtype=torch.bfloat16)
+            prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
+
+        # ---- combined attention -------------------------------------
+        pad_masks = torch.cat([prefix_pad, suffix_pad], dim=1)
+        att_masks = torch.cat([prefix_att, suffix_att], dim=1)
+        att_2d_masks = make_att_2d_masks(pad_masks, att_masks)
+
+        # Critical: zero out suffix → FAST attention. Without this the
+        # action expert reads the FAST tokens and trivially decodes
+        # them back to the same continuous actions it's supposed to
+        # predict from noise. Cumulative-block attention from
+        # ``make_att_2d_masks`` doesn't enforce this on its own
+        # because suffix tokens have a strictly higher cumsum than
+        # FAST tokens and therefore attend to them by default.
+        if fast_len > 0:
+            fast_start = non_fast_prefix_len
+            fast_end = non_fast_prefix_len + fast_len  # = prefix_pad.shape[1]
+            att_2d_masks[:, fast_end:, fast_start:fast_end] = False
+
+        position_ids = torch.cumsum(pad_masks, dim=1) - 1
+        att_2d_masks_4d = self.model._prepare_attention_masks_4d(att_2d_masks)
+
+        # ---- forward (capture BOTH expert outputs) ------------------
+        (prefix_out, suffix_out), _ = self.model.paligemma_with_expert.forward(
+            attention_mask=att_2d_masks_4d,
+            position_ids=position_ids,
+            past_key_values=None,
+            inputs_embeds=[prefix_embs, suffix_embs],
+            use_cache=False,
+            adarms_cond=[None, adarms_cond],
+        )
+
+        # ---- flow loss (mirrors PI05Pytorch.forward) ----------------
+        suffix_out_slice = suffix_out[:, -self.model.config.chunk_size :].to(
+            dtype=torch.float32
+        )
+        v_t = self.model.action_out_proj(suffix_out_slice)
+        flow_per_dim = F.mse_loss(u_t, v_t, reduction="none")
+        # Truncate to the actual action dimensionality (PI05 pads
+        # internally to max_action_dim).
+        original_action_dim = self.config.output_features[ACTION].shape[0]
+        flow_per_dim = flow_per_dim[:, :, :original_action_dim]
+        per_sample_flow = flow_per_dim.mean(dim=(1, 2))
+        flow_loss = _mask_per_sample(per_sample_flow, predict_actions_t)
+
+        # ---- text + FAST CE from prefix_out ------------------------
+        lm_head = self.model.paligemma_with_expert.paligemma.lm_head
+
+        text_loss: Tensor | None = None
+        if text_labels is not None and prefix_out is not None:
+            lang_len = text_labels.shape[1]
+            if fast_len > 0:
+                text_hidden = prefix_out[:, -(fast_len + lang_len) : -fast_len, :]
+            else:
+                text_hidden = prefix_out[:, -lang_len:, :]
+            text_logits = lm_head(text_hidden.to(lm_head.weight.dtype))
+            text_loss = _shifted_ce(text_logits, text_labels)
+
+        fast_loss: Tensor | None = None
+        if fast_len > 0 and prefix_out is not None:
+            fast_hidden = prefix_out[:, -fast_len:, :]
+            fast_logits = lm_head(fast_hidden.to(lm_head.weight.dtype))
+            fast_loss = _fast_ce(fast_logits, action_tokens, action_mask, predict_actions_t)
+
+        return flow_loss, text_loss, fast_loss
+
+    def _compute_text_and_fast_loss(
+        self,
+        batch: dict[str, Tensor],
+        text_labels: Tensor | None,
+        action_tokens: Tensor | None,
+        action_mask: Tensor | None,
+        predict_actions_t: Tensor | None = None,
+    ) -> tuple[Tensor | None, Tensor | None]:
+        """Single prefix forward → text CE + FAST CE.
+
+        Embed [images, language] (and FAST when requested) once, run
+        one backbone forward, then slice the resulting hidden states
+        at the language and FAST positions to compute both CE losses.
+        Bit-equivalent to running the two losses in separate forwards
+        because the segment-aware ``make_att_2d_masks`` keeps FAST
+        tokens invisible to language tokens, so adding FAST to the
+        prefix doesn't perturb the hidden states at language positions.
+
+        Returns ``(text_loss, fast_loss)``. Either can be ``None`` if
+        the caller doesn't want that head.
+        """
+        from ..pi05.modeling_pi05 import make_att_2d_masks  # noqa: PLC0415
+
+        images, img_masks = self.model._preprocess_images(batch)
+        lang_tokens = batch[OBS_LANGUAGE_TOKENS]
+        lang_masks = batch[OBS_LANGUAGE_ATTENTION_MASK]
+
+        prefix_embs, prefix_pad, prefix_att = self.model.embed_prefix(
+            images, img_masks, lang_tokens, lang_masks
+        )
+
+        fast_len = 0
+        if action_tokens is not None and action_mask is not None:
+            emb_dim = prefix_embs.shape[-1]
+            fast_emb = self.model.paligemma_with_expert.embed_language_tokens(action_tokens)
+            fast_emb = fast_emb * math.sqrt(emb_dim)
+
+            fast_len = action_tokens.shape[1]
+            ones_att = torch.ones(
+                (action_tokens.shape[0], fast_len),
+                dtype=torch.bool,
+                device=prefix_embs.device,
+            )
+            full_embs = torch.cat([prefix_embs, fast_emb], dim=1)
+            full_pad = torch.cat([prefix_pad, action_mask.to(prefix_pad.dtype)], dim=1)
+            full_att = torch.cat([prefix_att, ones_att], dim=1)
+        else:
+            full_embs = prefix_embs
+            full_pad = prefix_pad
+            full_att = prefix_att
+
+        att_2d = make_att_2d_masks(full_pad, full_att)
+        position_ids = torch.cumsum(full_pad, dim=1) - 1
+
+        (vlm_out, _), _ = self.model.paligemma_with_expert.forward(
+            attention_mask=att_2d,
+            position_ids=position_ids,
+            past_key_values=None,
+            inputs_embeds=[full_embs, None],
+            use_cache=False,
+        )
+        if vlm_out is None:
+            raise RuntimeError(
+                "PI052 text+fast loss: VLM forward returned no hidden states."
+            )
+
+        lm_head = self.model.paligemma_with_expert.paligemma.lm_head
+
+        text_loss: Tensor | None = None
+        if text_labels is not None:
+            lang_len = text_labels.shape[1]
+            # embed_prefix lays out as [images, language]; with FAST
+            # appended the full sequence is [images, language, FAST].
+            if fast_len > 0:
+                text_hidden = vlm_out[:, -(fast_len + lang_len):-fast_len, :]
+            else:
+                text_hidden = vlm_out[:, -lang_len:, :]
+            text_logits = lm_head(text_hidden.to(lm_head.weight.dtype))
+            text_loss = _shifted_ce(text_logits, text_labels)
+
+        fast_loss: Tensor | None = None
+        if action_tokens is not None and action_mask is not None and fast_len > 0:
+            fast_hidden = vlm_out[:, -fast_len:, :]
+            fast_logits = lm_head(fast_hidden.to(lm_head.weight.dtype))
+            fast_loss = _fast_ce(fast_logits, action_tokens, action_mask, predict_actions_t)
+
+        return text_loss, fast_loss
+
+    # ------------------------------------------------------------------
+    # select_message — AR text generation at inference
+    # ------------------------------------------------------------------
+
+    def select_message(
+        self,
+        batch: dict[str, Tensor],
+        *,
+        max_new_tokens: int = 128,
+        min_new_tokens: int = 0,
+        eos_token_id: int | None = None,
+        temperature: float = 0.0,
+        top_p: float = 1.0,
+        tokenizer: Any = None,
+    ) -> str:
+        """Generate text continuation from a multimodal prefix.
+
+        Mirrors ``SmolVLA2Policy.select_message`` so the same
+        :class:`lerobot.policies.smolvla2.inference.SmolVLA2Runtime`
+        can drive π0.5 v2 unchanged.
+        """
+        self.eval()
+
+        if tokenizer is None:
+            from transformers import AutoTokenizer  # noqa: PLC0415
+
+            tokenizer = AutoTokenizer.from_pretrained("google/paligemma-3b-pt-224")
+        if eos_token_id is None:
+            eos_token_id = tokenizer.eos_token_id
+
+        special_ids: set[int] = set()
+        try:
+            for sid in (tokenizer.all_special_ids or []):
+                if sid is not None:
+                    special_ids.add(int(sid))
+        except Exception:  # noqa: BLE001
+            pass
+        if eos_token_id is not None:
+            special_ids.add(int(eos_token_id))
+
+        images, img_masks = self.model._preprocess_images(batch)
+        tokens = batch[OBS_LANGUAGE_TOKENS]
+        masks = batch[OBS_LANGUAGE_ATTENTION_MASK]
+
+        prefix_embs, prefix_pad_masks, prefix_att_masks = self.model.embed_prefix(
+            images, img_masks, tokens, masks
+        )
+
+        device = prefix_embs.device
+        bsize = prefix_embs.shape[0]
+        emb_dim = prefix_embs.shape[-1]
+        text_emb_scale = math.sqrt(emb_dim)
+        ones_step = torch.ones((bsize, 1), dtype=torch.bool, device=device)
+
+        current_embs = prefix_embs
+        current_pad = prefix_pad_masks
+        current_att = prefix_att_masks
+        generated: list[int] = []
+
+        from ..pi05.modeling_pi05 import make_att_2d_masks  # noqa: PLC0415
+
+        backbone = self.model.paligemma_with_expert
+        lm_head = backbone.paligemma.lm_head
+
+        for _ in range(max_new_tokens):
+            att_2d = make_att_2d_masks(current_pad, current_att)
+            position_ids = torch.cumsum(current_pad, dim=1) - 1
+            (vlm_out, _), _ = backbone.forward(
+                attention_mask=att_2d,
+                position_ids=position_ids,
+                past_key_values=None,
+                inputs_embeds=[current_embs, None],
+                use_cache=False,
+            )
+            if vlm_out is None:
+                break
+            last = vlm_out[:, -1:].to(lm_head.weight.dtype)
+            logits_step = lm_head(last)[:, -1]  # (B, V)
+            if special_ids and len(generated) < min_new_tokens:
+                for sid in special_ids:
+                    logits_step[..., sid] = float("-inf")
+            next_ids = self._sample_next_token(logits_step, temperature, top_p)
+            tok_id = int(next_ids[0].item())
+            generated.append(tok_id)
+            if eos_token_id is not None and tok_id == eos_token_id:
+                break
+
+            new_emb = backbone.embed_language_tokens(next_ids.unsqueeze(0))
+            new_emb = new_emb * text_emb_scale
+            current_embs = torch.cat([current_embs, new_emb], dim=1)
+            current_pad = torch.cat([current_pad, ones_step], dim=1)
+            current_att = torch.cat([current_att, ones_step], dim=1)
+
+        decoded = tokenizer.decode(generated, skip_special_tokens=True).strip()
+        if not decoded and generated:
+            try:
+                self._last_select_message_debug = (
+                    f"raw_ids={generated[:16]} "
+                    f"decoded_w_special={tokenizer.decode(generated, skip_special_tokens=False)!r}"
+                )
+            except Exception:  # noqa: BLE001
+                self._last_select_message_debug = f"raw_ids={generated[:16]}"
+        else:
+            self._last_select_message_debug = ""
+        return decoded
+
+    @staticmethod
+    def _sample_next_token(logits: Tensor, temperature: float, top_p: float) -> Tensor:
+        if temperature <= 0.0:
+            return logits.argmax(dim=-1)
+        scaled = logits / max(temperature, 1e-6)
+        probs = torch.softmax(scaled, dim=-1)
+        if top_p < 1.0:
+            sorted_p, sorted_ix = torch.sort(probs, descending=True, dim=-1)
+            cum = torch.cumsum(sorted_p, dim=-1)
+            mask = cum > top_p
+            mask[..., 0] = False
+            sorted_p = sorted_p.masked_fill(mask, 0.0)
+            sorted_p = sorted_p / sorted_p.sum(dim=-1, keepdim=True).clamp_min(1e-8)
+            choice = torch.multinomial(sorted_p, num_samples=1)
+            return sorted_ix.gather(-1, choice).squeeze(-1)
+        return torch.multinomial(probs, num_samples=1).squeeze(-1)

src/lerobot/policies/pi052/processor_pi052.py

@@ -0,0 +1,196 @@
+# 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.
+
+"""π0.5 v2 pre/post-processor factory.
+
+When ``config.recipe_path`` is set, the pre-processor pipeline becomes:
+
+    rename observations
+    add batch dim
+    relative-action prep      (inherited from π0.5)
+    NormalizerProcessorStep
+    RenderMessagesStep        — recipe → messages, target_message_indices,
+                                message_streams (PR 1 of the steerable
+                                stack)
+    PI052TextTokenizerStep    — messages → input_ids + label mask +
+                                predict_actions
+    DeviceProcessorStep
+
+When ``recipe_path`` is ``None`` we delegate to the plain π0.5 pipeline
+so unannotated datasets keep working.
+
+Post-processor is unchanged from π0.5.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import torch
+
+from lerobot.configs.recipe import TrainingRecipe
+from lerobot.processor import (
+    AbsoluteActionsProcessorStep,
+    ActionTokenizerProcessorStep,
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RelativeActionsProcessorStep,
+    RenameObservationsProcessorStep,
+    RenderMessagesStep,
+    UnnormalizerProcessorStep,
+    policy_action_to_transition,
+    transition_to_policy_action,
+)
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+from ..pi05.processor_pi05 import make_pi05_pre_post_processors
+from .configuration_pi052 import PI052Config
+from .text_processor_pi052 import PI052TextTokenizerStep
+
+
+def make_pi052_pre_post_processors(
+    config: PI052Config,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    dataset_repo_id: str | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Build PI0.5-v2's pre/post-processor pipelines.
+
+    Falls through to π0.5's stock pipeline when ``recipe_path`` is unset.
+    """
+    if not config.recipe_path:
+        return make_pi05_pre_post_processors(config, dataset_stats=dataset_stats)
+
+    recipe = _load_recipe(config.recipe_path)
+
+    relative_step = RelativeActionsProcessorStep(
+        enabled=config.use_relative_actions,
+        exclude_joints=getattr(config, "relative_exclude_joints", []),
+        action_names=getattr(config, "action_feature_names", None),
+    )
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        relative_step,
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        RenderMessagesStep(recipe=recipe),
+        PI052TextTokenizerStep(
+            tokenizer_name="google/paligemma-3b-pt-224",
+            max_length=config.tokenizer_max_length,
+            plan_dropout_prob=getattr(config, "plan_dropout_prob", 0.0),
+            memory_dropout_prob=getattr(config, "memory_dropout_prob", 0.0),
+            subtask_dropout_prob=getattr(config, "subtask_dropout_prob", 0.0),
+        ),
+    ]
+
+    # FAST tokenizer for discrete-action CE supervision (paper §III.C).
+    # Only inserted when explicitly enabled — keeps the post-training-
+    # style recipe (flow + text) as the default. When on, the step
+    # writes ACTION_TOKENS / ACTION_TOKEN_MASK into
+    # ``COMPLEMENTARY_DATA`` and the modeling forward picks them up.
+    if getattr(config, "enable_fast_action_loss", False):
+        # Per Pertsch et al. 2025 (FAST [64], π0.5 §III.C): fit the
+        # tokenizer on this dataset's action distribution rather than
+        # using the universal codebook off the shelf. We do this once
+        # and cache to disk, keyed on (dataset, base, n_samples).
+        action_tokenizer_path = config.action_tokenizer_name
+        if (
+            getattr(config, "auto_fit_fast_tokenizer", False)
+            and dataset_repo_id is not None
+        ):
+            from .fit_fast_tokenizer import fit_fast_tokenizer  # noqa: PLC0415
+
+            cache_dir = Path(config.fast_tokenizer_cache_dir).expanduser()
+            try:
+                action_tokenizer_path = fit_fast_tokenizer(
+                    dataset_repo_id=dataset_repo_id,
+                    cache_dir=cache_dir,
+                    base_tokenizer_name=config.action_tokenizer_name,
+                    n_samples=config.fast_tokenizer_fit_samples,
+                    chunk_size=config.chunk_size,
+                )
+            except Exception as exc:  # noqa: BLE001
+                import logging  # noqa: PLC0415
+
+                logging.getLogger(__name__).warning(
+                    "FAST tokenizer fit failed (%s) — falling back to "
+                    "the universal base tokenizer %r. Train will still "
+                    "work but compression will be suboptimal.",
+                    exc, config.action_tokenizer_name,
+                )
+
+        input_steps.append(
+            ActionTokenizerProcessorStep(
+                action_tokenizer_name=action_tokenizer_path,
+                max_action_tokens=config.max_action_tokens,
+                fast_skip_tokens=config.fast_skip_tokens,
+                paligemma_tokenizer_name="google/paligemma-3b-pt-224",
+            )
+        )
+
+    input_steps.append(DeviceProcessorStep(device=config.device))
+
+    output_steps = [
+        UnnormalizerProcessorStep(
+            features=config.output_features,
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        AbsoluteActionsProcessorStep(
+            enabled=config.use_relative_actions,
+            relative_step=relative_step,
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )
+
+
+def _load_recipe(path_str: str) -> TrainingRecipe:
+    """Resolve ``path_str`` to a ``TrainingRecipe``.
+
+    Accepts an absolute path or a path relative to
+    ``src/lerobot/configs/`` (same lookup rules as
+    ``make_smolvla2_pre_post_processors``).
+    """
+    p = Path(path_str)
+    if not p.is_absolute() and not p.exists():
+        from lerobot.configs import recipe as _recipe_module  # noqa: PLC0415
+
+        configs_dir = Path(_recipe_module.__file__).resolve().parent
+        candidate = configs_dir / path_str
+        if candidate.exists():
+            p = candidate
+    return TrainingRecipe.from_yaml(p)