fix(smolvla2): per-recipe inference prompts to match training shape

The four high-level steps shared one generic
``_control_context_messages`` that jammed task + plan + memory +
completed_subtask into a single user message. The recipes in
``smolvla2_hirobot.yaml`` each have a *specific* multi-message layout
(``memory_update``: ``user(task) → assistant(prev memory) →
user(completed subtask)``; ``high_level_subtask``: ``user(task+plan+
memory) → user(current subtask)``; ``user_interjection_response``:
``user(task) → assistant(prev plan) → user(interjection)``). After
``apply_chat_template`` those layouts produce different prompts than
the runtime's flattened single-user-turn version, and the model fell
back to its dominant training mode (VQA JSON output) — generating
``":":":":":":...`` repetition.

Add four per-recipe prompt builders (``_msgs_for_subtask``,
``_msgs_for_memory``, ``_msgs_for_interjection``, ``_msgs_for_vqa``),
each mirroring its sub-recipe's exact message structure including
the ``if_present`` skips. Wire each high-level step to its matching
builder. Inference prompts now line up with what the model saw in
training, so generation should produce coherent text instead of
repeated tokens.

Generic ``_control_context_messages`` is kept (still used by tests
and the no-recipe fallback path).

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

@@ -0,0 +1,945 @@
+# 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.
+
+# Integration tests: build an isolated Docker image per benchmark and run a
+# 1-episode smoke eval. Each benchmark gets its own image so incompatible
+# dependency trees (e.g. hf-libero vs metaworld==3.0.0) can never collide.
+#
+# To add a new benchmark:
+#   1. Add docker/Dockerfile.benchmark.<name>  (install only lerobot[<name>])
+#   2. Copy one of the jobs below and adjust the image name and eval command.
+name: Benchmark Integration Tests
+
+on:
+  # Run manually from the Actions tab
+  workflow_dispatch:
+
+  # Run every Monday at 02:00 UTC.
+  schedule:
+    - cron: "0 2 * * 1"
+
+  push:
+    branches:
+      - main
+    paths:
+      - "src/lerobot/envs/**"
+      - "src/lerobot/scripts/lerobot_eval.py"
+      - "docker/Dockerfile.benchmark.*"
+      - ".github/workflows/benchmark_tests.yml"
+      - "pyproject.toml"
+
+  pull_request:
+    branches:
+      - main
+    paths:
+      - "src/lerobot/envs/**"
+      - "src/lerobot/scripts/lerobot_eval.py"
+      - "docker/Dockerfile.benchmark.*"
+      - ".github/workflows/benchmark_tests.yml"
+      - "pyproject.toml"
+
+permissions:
+  contents: read
+
+env:
+  UV_VERSION: "0.8.0"
+  PYTHON_VERSION: "3.12"
+
+# Cancel in-flight runs for the same branch/PR.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  # ── LIBERO ────────────────────────────────────────────────────────────────
+  # Isolated image: lerobot[libero] only (hf-libero, dm-control, mujoco chain)
+  libero-integration-test:
+    name: Libero — 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: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+
+      # Build the benchmark-specific image. The Dockerfile separates dep-install
+      # from source-copy, so code-only changes skip the slow uv-sync layer
+      # when the runner has a warm Docker daemon cache.
+      - name: Build Libero benchmark image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: docker/Dockerfile.benchmark.libero
+          push: false
+          load: true
+          tags: lerobot-benchmark-libero:ci
+
+      - name: Run Libero smoke eval (1 episode)
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          # Named container (no --rm) so we can docker cp artifacts out.
+          # Output to /tmp inside the container — /artifacts doesn't exist
+          # and user_lerobot cannot create root-level dirs.
+          docker run --name libero-eval --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            lerobot-benchmark-libero:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              lerobot-eval \
+                --policy.path=lerobot/smolvla_libero \
+                --env.type=libero \
+                --env.task=libero_spatial \
+                --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 --task libero_spatial \
+                --output /tmp/eval-artifacts/task_descriptions.json
+            "
+
+      - name: Copy Libero artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/libero-artifacts
+          docker cp libero-eval:/tmp/eval-artifacts/. /tmp/libero-artifacts/ 2>/dev/null || true
+          docker rm -f libero-eval || true
+
+      - name: Parse Libero eval metrics
+        if: always()
+        run: |
+          python3 scripts/ci/parse_eval_metrics.py \
+            --artifacts-dir /tmp/libero-artifacts \
+            --env libero \
+            --task libero_spatial \
+            --policy lerobot/smolvla_libero
+
+      - name: Upload Libero rollout video
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: libero-rollout-video
+          path: /tmp/libero-artifacts/videos/
+          if-no-files-found: warn
+
+      - name: Upload Libero eval metrics
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: libero-metrics
+          path: /tmp/libero-artifacts/metrics.json
+          if-no-files-found: warn
+
+      # ── LIBERO TRAIN+EVAL SMOKE ──────────────────────────────────────────────
+      # Train SmolVLA for 1 step (batch_size=1, dataset episode 0 only) then
+      # immediately runs eval inside the training loop (eval_freq=1, 1 episode).
+      # Tests the full train→eval-within-training pipeline end-to-end.
+      - name: Run Libero train+eval smoke (1 step, eval_freq=1)
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          docker run --name libero-train-smoke --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            lerobot-benchmark-libero:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              accelerate launch --num_processes=1 \$(which lerobot-train) \
+                --policy.path=lerobot/smolvla_base \
+                --policy.load_vlm_weights=true \
+                --policy.scheduler_decay_steps=25000 \
+                --policy.freeze_vision_encoder=false \
+                --policy.train_expert_only=false \
+                --dataset.repo_id=lerobot/libero \
+                --dataset.episodes=[0] \
+                --dataset.use_imagenet_stats=false \
+                --env.type=libero \
+                --env.task=libero_spatial \
+                '--env.camera_name_mapping={\"agentview_image\": \"camera1\", \"robot0_eye_in_hand_image\": \"camera2\"}' \
+                --policy.empty_cameras=1 \
+                --output_dir=/tmp/train-smoke \
+                --steps=1 \
+                --batch_size=1 \
+                --eval_freq=1 \
+                --eval.n_episodes=1 \
+                --eval.batch_size=1 \
+                --eval.use_async_envs=false \
+                --save_freq=1 \
+                --policy.push_to_hub=false \
+                '--rename_map={\"observation.images.image\": \"observation.images.camera1\", \"observation.images.image2\": \"observation.images.camera2\"}'
+            "
+
+      - name: Copy Libero train-smoke artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/libero-train-smoke-artifacts
+          docker cp libero-train-smoke:/tmp/train-smoke/. /tmp/libero-train-smoke-artifacts/ 2>/dev/null || true
+          docker rm -f libero-train-smoke || true
+
+      - name: Upload Libero train-smoke eval video
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: libero-train-smoke-video
+          path: /tmp/libero-train-smoke-artifacts/eval/
+          if-no-files-found: warn
+
+  # ── METAWORLD ─────────────────────────────────────────────────────────────
+  # Isolated image: lerobot[metaworld] only (metaworld==3.0.0, mujoco>=3 chain)
+  metaworld-integration-test:
+    name: MetaWorld — 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: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+
+      - name: Build MetaWorld benchmark image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: docker/Dockerfile.benchmark.metaworld
+          push: false
+          load: true
+          tags: lerobot-benchmark-metaworld:ci
+
+      - name: Run MetaWorld smoke eval (1 episode)
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          docker run --name metaworld-eval --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            lerobot-benchmark-metaworld:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              lerobot-eval \
+                --policy.path=lerobot/smolvla_metaworld \
+                --env.type=metaworld \
+                --env.task=metaworld-push-v3 \
+                --eval.batch_size=1 \
+                --eval.n_episodes=1 \
+                --eval.use_async_envs=false \
+                --policy.device=cuda \
+                '--rename_map={\"observation.image\": \"observation.images.camera1\"}' \
+                --policy.empty_cameras=2 \
+                --output_dir=/tmp/eval-artifacts
+              python scripts/ci/extract_task_descriptions.py \
+                --env metaworld --task metaworld-push-v3 \
+                --output /tmp/eval-artifacts/task_descriptions.json
+            "
+
+      - name: Copy MetaWorld artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/metaworld-artifacts
+          docker cp metaworld-eval:/tmp/eval-artifacts/. /tmp/metaworld-artifacts/ 2>/dev/null || true
+          docker rm -f metaworld-eval || true
+
+      - name: Parse MetaWorld eval metrics
+        if: always()
+        run: |
+          python3 scripts/ci/parse_eval_metrics.py \
+            --artifacts-dir /tmp/metaworld-artifacts \
+            --env metaworld \
+            --task metaworld-push-v3 \
+            --policy lerobot/smolvla_metaworld
+
+      - name: Upload MetaWorld rollout video
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: metaworld-rollout-video
+          path: /tmp/metaworld-artifacts/videos/
+          if-no-files-found: warn
+
+      - name: Upload MetaWorld eval metrics
+        if: always()
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          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/fast_tests.yml

@@ -12,7 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# This workflow handles fast testing.
+# This workflow validates each optional-dependency tier in isolation.
+# Each tier installs a different extra and runs the full test suite.
+# Tests that require an extra not installed in the current tier are
+# skipped automatically via pytest.importorskip guards.
 name: Fast Tests
 
 on:
@@ -54,8 +57,9 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  # This job runs pytests with the default dependencies.
-  # It runs everytime we commit to a PR or push to main
+  # This job runs pytests in isolated dependency tiers.
+  # Each tier installs a different extra and runs the full suite;
+  # tests gated behind other extras skip automatically.
   fast-pytest-tests:
     name: Fast Pytest Tests
     runs-on: ubuntu-latest
@@ -89,8 +93,9 @@ jobs:
           version: ${{ env.UV_VERSION }}
           python-version: ${{ env.PYTHON_VERSION }}
 
-      - name: Install lerobot with test extras
-        run: uv sync --locked --extra "test"
+      # ── Tier 1: Base ──────────────────────────────────────
+      - name: "Tier 1 — Install: base"
+        run: uv sync --locked --extra test
 
       - name: Login to Hugging Face
         if: env.HF_USER_TOKEN != ''
@@ -98,5 +103,26 @@ jobs:
           uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
           uv run hf auth whoami
 
-      - name: Run pytest
+      - name: "Tier 1 — Test: base"
+        run: uv run pytest tests -vv --maxfail=10
+
+      # ── Tier 2: Dataset ──────────────────────────────────
+      - name: "Tier 2 — Install: dataset"
+        run: uv sync --locked --extra test --extra dataset
+
+      - name: "Tier 2 — Test: dataset"
+        run: uv run pytest tests -vv --maxfail=10
+
+      # ── Tier 3: Hardware ─────────────────────────────────
+      - name: "Tier 3 — Install: hardware"
+        run: uv sync --locked --extra test --extra hardware
+
+      - name: "Tier 3 — Test: hardware"
+        run: uv run pytest tests -vv --maxfail=10
+
+      # ── Tier 4: Viz ──────────────────────────────────────
+      - name: "Tier 4 — Install: viz"
+        run: uv sync --locked --extra test --extra viz
+
+      - name: "Tier 4 — Test: viz"
         run: uv run pytest tests -vv --maxfail=10

.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

@@ -0,0 +1,42 @@
+# 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 LIBERO integration tests.
+# Extends the nightly GPU image (which already has all extras installed)
+# with the PR's source code and LIBERO-specific asset setup.
+#
+# Build:  docker build -f docker/Dockerfile.benchmark.libero -t lerobot-benchmark-libero .
+# Run:    docker run --gpus all --rm lerobot-benchmark-libero 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.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.metaworld

@@ -0,0 +1,27 @@
+# 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 MetaWorld integration tests.
+# Extends the nightly GPU image (which already has all extras installed)
+# with the PR's source code.
+#
+# Build:  docker build -f docker/Dockerfile.benchmark.metaworld -t lerobot-benchmark-metaworld .
+# Run:    docker run --gpus all --rm lerobot-benchmark-metaworld lerobot-eval ...
+
+FROM huggingface/lerobot-gpu:latest
+
+# 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/adding_benchmarks.mdx

@@ -216,7 +216,7 @@ class MyBenchmarkEnvConfig(EnvConfig):
 
     def get_env_processors(self):
         """Override if your benchmark needs observation/action transforms."""
-        from lerobot.processor.pipeline import PolicyProcessorPipeline
+        from lerobot.processor import PolicyProcessorPipeline
         from lerobot.processor.env_processor import MyBenchmarkProcessorStep
         return (
             PolicyProcessorPipeline(steps=[MyBenchmarkProcessorStep()]),

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/async.mdx

@@ -170,7 +170,7 @@ python -m lerobot.async_inference.robot_client \
 ```python
 import threading
 from lerobot.robots.so_follower import SO100FollowerConfig
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.async_inference.configs import RobotClientConfig
 from lerobot.async_inference.robot_client import RobotClient
 from lerobot.async_inference.helpers import visualize_action_queue_size

docs/source/backwardcomp.mdx

@@ -41,7 +41,7 @@ The script:
 
 ```python
 # New usage pattern (after migration)
-from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.policies import make_policy, make_pre_post_processors
 
 # Load model and processors separately
 policy = make_policy(config, ds_meta=dataset.meta)

docs/source/bring_your_own_policies.mdx

@@ -47,9 +47,9 @@ Here is a template to get you started, customize the parameters and methods as n
 ```python
 # configuration_my_custom_policy.py
 from dataclasses import dataclass, field
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.optim.optimizers import AdamWConfig
-from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.configs import PreTrainedConfig
+from lerobot.optim import AdamWConfig
+from lerobot.optim import CosineDecayWithWarmupSchedulerConfig
 
 @PreTrainedConfig.register_subclass("my_custom_policy")
 @dataclass
@@ -120,7 +120,7 @@ import torch
 import torch.nn as nn
 from typing import Any
 
-from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies import PreTrainedPolicy
 from lerobot.utils.constants import ACTION
 from .configuration_my_custom_policy import MyCustomPolicyConfig
 

docs/source/cameras.mdx

@@ -79,9 +79,8 @@ The following examples show how to use the camera API to configure and capture f
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.cameras.opencv.camera_opencv import OpenCVCamera
-from lerobot.cameras.configs import ColorMode, Cv2Rotation
+from lerobot.cameras.opencv import OpenCVCamera, OpenCVCameraConfig
+from lerobot.cameras import ColorMode, Cv2Rotation
 
 # Construct an `OpenCVCameraConfig` with your desired FPS, resolution, color mode, and rotation.
 config = OpenCVCameraConfig(
@@ -126,9 +125,8 @@ with OpenCVCamera(config) as camera:
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig
-from lerobot.cameras.realsense.camera_realsense import RealSenseCamera
-from lerobot.cameras.configs import ColorMode, Cv2Rotation
+from lerobot.cameras.realsense import RealSenseCamera, RealSenseCameraConfig
+from lerobot.cameras import ColorMode, Cv2Rotation
 
 # Create a `RealSenseCameraConfig` specifying your camera’s serial number and enabling depth.
 config = RealSenseCameraConfig(

docs/source/dataset_subtask.mdx

@@ -1,278 +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.lerobot_dataset 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.tokenizer_processor import TokenizerProcessor
-from lerobot.processor.pipeline import ProcessorPipeline
-
-# Create a tokenizer processor
-tokenizer_processor = TokenizerProcessor(
-    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.lerobot_dataset 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.lerobot_dataset 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/earthrover_mini_plus.mdx

@@ -66,10 +66,10 @@ The SDK gives you:
 
 Follow our [Installation Guide](./installation) to install LeRobot.
 
-In addition to the base installation, install the EarthRover Mini dependencies:
+In addition to the base installation, install the EarthRover Mini with hardware dependencies:
 
 ```bash
-pip install -e .
+pip install -e ".[hardware]"
 ```
 
 ## How It Works

docs/source/env_processor.mdx

@@ -173,8 +173,8 @@ observation = {
 The `make_env_pre_post_processors` function follows the same pattern as `make_pre_post_processors` for policies:
 
 ```python
-from lerobot.envs.factory import make_env_pre_post_processors
-from lerobot.envs.configs import LiberoEnv, PushtEnv
+from lerobot.envs import make_env_pre_post_processors, PushtEnv
+from lerobot.envs.configs import LiberoEnv
 
 # For LIBERO: Returns LiberoProcessorStep in preprocessor
 libero_cfg = LiberoEnv(task="libero_spatial", camera_name=["agentview"])
@@ -257,7 +257,7 @@ def eval_main(cfg: EvalPipelineConfig):
 The `LiberoProcessorStep` demonstrates a real-world environment processor:
 
 ```python
-from lerobot.processor.pipeline import ObservationProcessorStep
+from lerobot.processor import ObservationProcessorStep
 
 @dataclass
 @ProcessorStepRegistry.register(name="libero_processor")

docs/source/envhub.mdx

@@ -34,7 +34,7 @@ Finally, your environment must implement the standard `gym.vector.VectorEnv` int
 Loading an environment from the Hub is as simple as:
 
 ```python
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 
 # Load a hub environment (requires explicit consent to run remote code)
 env = make_env("lerobot/cartpole-env", trust_remote_code=True)
@@ -191,7 +191,7 @@ api.upload_folder(
 ### Basic Usage
 
 ```python
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 
 # Load from the hub
 envs_dict = make_env(
@@ -314,7 +314,7 @@ env = make_env("trusted-org/verified-env@a1b2c3d4", trust_remote_code=True)
 Here's a complete example using the reference CartPole environment:
 
 ```python
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 import numpy as np
 
 # Load the environment

docs/source/envhub_isaaclab_arena.mdx

@@ -58,10 +58,10 @@ pip install -e .
 cd ..
 
 
-# 5. Install LeRobot
+# 5. Install LeRobot (evaluation extra for env/policy evaluation)
 git clone https://github.com/huggingface/lerobot.git
 cd lerobot
-pip install -e .
+pip install -e ".[evaluation]"
 cd ..
 
 
@@ -262,7 +262,7 @@ def main(cfg: EvalPipelineConfig):
     """Run random action rollout for IsaacLab Arena environment."""
     logging.info(pformat(asdict(cfg)))
 
-    from lerobot.envs.factory import make_env
+    from lerobot.envs import make_env
 
     env_dict = make_env(
         cfg.env,

docs/source/envhub_leisaac.mdx

@@ -74,7 +74,7 @@ EnvHub exposes every LeIsaac-supported task in a uniform interface. The examples
 # envhub_random_action.py
 
 import torch
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 
 # Load from the hub
 envs_dict = make_env("LightwheelAI/leisaac_env:envs/so101_pick_orange.py", n_envs=1, trust_remote_code=True)
@@ -142,7 +142,7 @@ from lerobot.teleoperators import (  # noqa: F401
 )
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import init_logging
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 
 
 @dataclass
@@ -282,7 +282,7 @@ Note: when working with `bi_so101_fold_cloth`, call `initialize()` immediately a
 
 ```python
 import torch
-from lerobot.envs.factory import make_env
+from lerobot.envs import make_env
 
 # Load from the hub
 envs_dict = make_env("LightwheelAI/leisaac_env:envs/bi_so101_fold_cloth.py", n_envs=1, trust_remote_code=True)

docs/source/hilserl.mdx

@@ -685,6 +685,10 @@ Example configuration for training the [reward classifier](https://huggingface.c
 
 ```json
 {
+  "dataset": {
+    "repo_id": "hf_username/dataset_name",
+    "root": null
+  },
   "policy": {
     "type": "reward_classifier",
     "model_name": "helper2424/resnet10",
@@ -705,8 +709,28 @@ Example configuration for training the [reward classifier](https://huggingface.c
         "type": "VISUAL",
         "shape": [3, 128, 128]
       }
-    }
-  }
+    },
+    "push_to_hub": true,
+    "repo_id": "hf_username/model_repo"
+  },
+  "batch_size": 16,
+  "num_workers": 4,
+  "steps": 5000,
+  "log_freq": 10,
+  "eval_freq": 1000,
+  "save_freq": 1000,
+  "save_checkpoint": true,
+  "seed": 2,
+  "resume": false,
+  "optimizer": {
+    "grad_clip_norm": 10.0
+  },
+  "wandb": {
+    "enable": true,
+    "project": "reward-classifier",
+    "disable_artifact": false
+  },
+  "job_name": "reward-classifier"
 }
 ```
 

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.
@@ -58,8 +64,8 @@ lerobot-teleoperate \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so_leader import SO101LeaderConfig, SO101Leader
-from lerobot.robots.so_follower import SO101FollowerConfig, SO101Follower
+from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
+from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
 
 robot_config = SO101FollowerConfig(
     port="/dev/tty.usbmodem58760431541",
@@ -116,9 +122,9 @@ lerobot-teleoperate \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.teleoperators.koch_leader import KochLeaderConfig, KochLeader
-from lerobot.robots.koch_follower import KochFollowerConfig, KochFollower
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.teleoperators.koch_leader import KochLeader, KochLeaderConfig
+from lerobot.robots.koch_follower import KochFollower, KochFollowerConfig
 
 camera_config = {
     "front": OpenCVCameraConfig(index_or_path=0, width=1920, height=1080, fps=30)
@@ -195,13 +201,12 @@ lerobot-record \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.datasets import LeRobotDataset
+from lerobot.utils.feature_utils import hw_to_dataset_features
 from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
-from lerobot.teleoperators.so_leader.config_so100_leader import SO100LeaderConfig
-from lerobot.teleoperators.so_leader.so100_leader import SO100Leader
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
+from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 from lerobot.scripts.lerobot_record import record_loop
@@ -410,9 +415,8 @@ lerobot-replay \
 ```python
 import time
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.robots.so_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so_follower.so100_follower import SO100Follower
+from lerobot.datasets import LeRobotDataset
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
@@ -532,15 +536,14 @@ lerobot-record  \
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
-from lerobot.robots.so_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so_follower.so100_follower import SO100Follower
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.datasets import LeRobotDataset
+from lerobot.utils.feature_utils import hw_to_dataset_features
+from lerobot.policies.act import ACTPolicy
+from lerobot.policies import make_pre_post_processors
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.scripts.lerobot_record import record_loop
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

docs/source/installation.mdx

@@ -116,6 +116,8 @@ brew install ffmpeg
 
 ## Step 3: Install LeRobot 🤗
 
+The base `lerobot` install is intentionally **lightweight** — it includes only core ML dependencies (PyTorch, torchvision, numpy, opencv, einops, draccus, huggingface-hub, gymnasium, safetensors). Heavier dependencies are gated behind optional extras so you only install what you need.
+
 ### From Source
 
 First, clone the repository and navigate into the directory:
@@ -131,12 +133,16 @@ Then, install the library in editable mode. This is useful if you plan to contri
 <hfoptions id="install_lerobot_src">
 <hfoption id="conda">
 ```bash
-pip install -e .
+pip install -e ".[core_scripts]"  # For robot workflows (recording, replaying, calibrate)
+pip install -e ".[training]"      # For training policies
+pip install -e ".[all]"           # Everything (all policies, envs, hardware, dev tools)
 ```
 </hfoption>
 <hfoption id="uv">
 ```bash
-uv pip install -e .
+uv pip install -e ".[core_scripts]"  # For robot workflows (recording, replaying, calibrate)
+uv pip install -e ".[training]"      # For training policies
+uv pip install -e ".[all]"           # Everything (all policies, envs, hardware, dev tools)
 ```
 </hfoption>
 </hfoptions>
@@ -162,26 +168,48 @@ uv pip install lerobot
 </hfoptions>
 <!-- prettier-ignore-end -->
 
-_This installs only the default dependencies._
+_This installs only the core ML dependencies. You will need to add extras for most workflows._
 
-**Extra Features:**
-To install additional functionality, use one of the following (If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.):
+**Feature Extras:**
+LeRobot provides **feature-scoped extras** that map to common workflows. If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.
+
+| Extra      | What it adds                                | Typical use case                    |
+| ---------- | ------------------------------------------- | ----------------------------------- |
+| `dataset`  | `datasets`, `av`, `torchcodec`, `jsonlines` | Loading & creating datasets         |
+| `training` | `dataset` + `accelerate`, `wandb`           | Training policies                   |
+| `hardware` | `pynput`, `pyserial`, `deepdiff`            | Connecting to real robots           |
+| `viz`      | `rerun-sdk`                                 | Visualization during recording/eval |
+
+**Composite Extras** combine feature extras for common CLI scripts:
+
+| Extra          | Includes                       | Typical use case                                        |
+| -------------- | ------------------------------ | ------------------------------------------------------- |
+| `core_scripts` | `dataset` + `hardware` + `viz` | `lerobot-record`, `lerobot-replay`, `lerobot-calibrate` |
+| `evaluation`   | `av`                           | `lerobot-eval` (add policy + env extras as needed)      |
+| `dataset_viz`  | `dataset` + `viz`              | `lerobot-dataset-viz`, `lerobot-imgtransform-viz`       |
 
 ```bash
-pip install 'lerobot[all]'          # All available features
-pip install 'lerobot[aloha,pusht]'  # Specific features (Aloha & Pusht)
-pip install 'lerobot[feetech]'      # Feetech motor support
+pip install 'lerobot[core_scripts]'          # Record, replay, calibrate
+pip install 'lerobot[training]'              # Train policies
+pip install 'lerobot[core_scripts,training]' # Record + train
+pip install 'lerobot[all]'                   # Everything
 ```
 
-_Replace `[...]` with your desired features._
+**Policy, environment, and hardware extras** are still available for specific dependencies:
 
-**Available Tags:**
-For a full list of optional dependencies, see:
-https://pypi.org/project/lerobot/
+```bash
+pip install 'lerobot[pi]'             # Pi0/Pi0.5/Pi0-FAST policy deps
+pip install 'lerobot[smolvla]'        # SmolVLA policy deps
+pip install 'lerobot[diffusion]'      # Diffusion policy deps (diffusers)
+pip install 'lerobot[aloha,pusht]'    # Simulation environments
+pip install 'lerobot[feetech]'        # Feetech motor support
+```
+
+_Multiple extras can be combined (e.g., `.[core_scripts,pi,pusht]`). For a full list of available extras, refer to `pyproject.toml`._
 
 ### Troubleshooting
 
-If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
+If you encounter build errors, you may need to install additional system dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
 To install these for Linux run:
 
 ```bash
@@ -196,8 +224,8 @@ LeRobot provides optional extras for specific functionalities. Multiple extras c
 
 ### Simulations
 
-Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht))
-Example:
+Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht)).
+These automatically include the `dataset` extra.
 
 ```bash
 pip install -e ".[aloha]" # or "[pusht]" for example
@@ -213,7 +241,7 @@ pip install -e ".[feetech]" # or "[dynamixel]" for example
 
 ### Experiment Tracking
 
-To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
+Weights and Biases is included in the `training` extra. To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with:
 
 ```bash
 wandb login

docs/source/introduction_processors.mdx

@@ -19,10 +19,10 @@ This means that your favorite policy can be used like this:
 ```python
 import torch
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.datasets import LeRobotDataset
+from lerobot.policies import make_pre_post_processors
 from lerobot.policies.your_policy import YourPolicy
-from lerobot.processor.pipeline import RobotProcessorPipeline, PolicyProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline, PolicyProcessorPipeline
 dataset = LeRobotDataset("hf_user/dataset", episodes=[0])
 sample = dataset[10]
 
@@ -260,7 +260,7 @@ Since processor pipelines can add new features (like velocity fields), change te
 These functions work together by starting with robot hardware specifications (`create_initial_features()`) then simulating the entire pipeline transformation (`aggregate_pipeline_dataset_features()`) to compute the final feature dictionary that gets passed to `LeRobotDataset.create()`, ensuring perfect alignment between what processors output and what datasets expect to store.
 
 ```python
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features
+from lerobot.datasets import aggregate_pipeline_dataset_features
 
 # Start with robot's raw features
 initial_features = create_initial_features(

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/lerobot-dataset-v3.mdx

@@ -89,7 +89,7 @@ A core v3 principle is **decoupling storage from the user API**: data is stored
 
 ```python
 import torch
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 
 repo_id = "yaak-ai/L2D-v3"
 
@@ -135,7 +135,7 @@ for batch in data_loader:
 Use `StreamingLeRobotDataset` to iterate directly from the Hub without local copies. This allows to stream large datasets without the need to downloading them onto disk or loading them onto memory, and is a key feature of the new dataset format.
 
 ```python
-from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
+from lerobot.datasets import StreamingLeRobotDataset
 
 repo_id = "yaak-ai/L2D-v3"
 dataset = StreamingLeRobotDataset(repo_id)  # streams directly from the Hub
@@ -167,8 +167,8 @@ Currently, transforms are applied during **training time only**, not during reco
 Use the `image_transforms` parameter when loading a dataset for training:
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.transforms import ImageTransforms, ImageTransformsConfig, ImageTransformConfig
+from lerobot.datasets import LeRobotDataset
+from lerobot.transforms import ImageTransforms, ImageTransformsConfig, ImageTransformConfig
 
 # Option 1: Use default transform configuration (disabled by default)
 transforms_config = ImageTransformsConfig(
@@ -290,7 +290,7 @@ python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DAT
 When creating or recording datasets, you **must** call `dataset.finalize()` to properly close parquet writers. See the [PR #1903](https://github.com/huggingface/lerobot/pull/1903) for more details.
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 
 # Create dataset and record episodes
 dataset = LeRobotDataset.create(...)

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/multi_gpu_training.mdx

@@ -4,10 +4,10 @@ This guide shows you how to train policies on multiple GPUs using [Hugging Face
 
 ## Installation
 
-First, ensure you have accelerate installed:
+`accelerate` is included in the `training` extra. Install it with:
 
 ```bash
-pip install accelerate
+pip install 'lerobot[training]'
 ```
 
 ## Training with Multiple GPUs

docs/source/phone_teleop.mdx

@@ -45,7 +45,8 @@ Modify the examples to use `PhoneOS.IOS` or `PhoneOS.ANDROID` in `PhoneConfig`.
 Teleoperation example:
 
 ```python
-from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone import Phone, PhoneConfig
+from lerobot.teleoperators.phone.config_phone import PhoneOS
 
 teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
 teleop_device = Phone(teleop_config)

docs/source/pi0.mdx

@@ -110,8 +110,7 @@ lerobot-edit-dataset \
 Or equivalently in Python:
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.dataset_tools import recompute_stats
+from lerobot.datasets import LeRobotDataset, recompute_stats
 
 dataset = LeRobotDataset("your_dataset")
 recompute_stats(dataset, relative_action=True, chunk_size=50, relative_exclude_joints=["gripper"])

docs/source/pi05.mdx

@@ -116,8 +116,7 @@ lerobot-edit-dataset \
 Or equivalently in Python:
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.dataset_tools import recompute_stats
+from lerobot.datasets import LeRobotDataset, recompute_stats
 
 dataset = LeRobotDataset("your_dataset")
 recompute_stats(dataset, relative_action=True, chunk_size=50, relative_exclude_joints=["gripper"])

docs/source/policy_pi0_README.md

@@ -60,11 +60,10 @@ When `use_relative_actions=true`, the training script automatically:
 ### Recomputing stats for an existing dataset
 
 If you want to precompute relative action stats offline, use `recompute_stats` from
-`lerobot.datasets.dataset_tools`:
+`lerobot.datasets`:
 
 ```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.dataset_tools import recompute_stats
+from lerobot.datasets import LeRobotDataset, recompute_stats
 
 dataset = LeRobotDataset("your_org/your_dataset")
 dataset = recompute_stats(

docs/source/robocasa.mdx

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

docs/source/robocerebra.mdx

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

docs/source/robomme.mdx

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

docs/source/robotwin.mdx

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

docs/source/rtc.mdx

@@ -39,9 +39,8 @@ The snippet below provides a simplified pseudo-example of how RTC operates with
 
 ```python
 from lerobot.policies.pi0 import PI0Policy, PI0Config
-from lerobot.configs.types import RTCAttentionSchedule
-from lerobot.policies.rtc.configuration_rtc import RTCConfig
-from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.configs import RTCAttentionSchedule
+from lerobot.policies.rtc import RTCConfig, ActionQueue
 
 # Load Pi0 with RTC enabled
 policy_cfg = PI0Config()

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.

docs/source/xvla.mdx

@@ -418,7 +418,7 @@ Create a custom preprocessing pipeline for your environment:
 
 ```python
 from lerobot.processor import PolicyProcessorPipeline
-from lerobot.policies.xvla.processor_xvla import (
+from lerobot.policies.xvla import (
     XVLAImageToFloatProcessorStep,
     XVLAImageNetNormalizeProcessorStep,
     XVLAAddDomainIdProcessorStep,

examples/annotation/run_hf_job.py

@@ -0,0 +1,67 @@
+#!/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=256 "
+    "--vlm.max_new_tokens=512 "
+    "--executor.episode_parallelism=32 "
+    "--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=10 "
+    "--module_3.K=1 "
+    "--module_3.vqa_emission_hz=1.0 "
+    "--push_to_hub=pepijn223/super_poulain_full_tool2"
+)
+
+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/backward_compatibility/replay.py

@@ -35,7 +35,7 @@ from pprint import pformat
 
 import draccus
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,

examples/dataset/load_lerobot_dataset.py

@@ -31,17 +31,11 @@ from pprint import pprint
 import torch
 from huggingface_hub import HfApi
 
-import lerobot
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata
 
 
 def main():
-    # We ported a number of existing datasets ourselves, use this to see the list:
-    print("List of available datasets:")
-    pprint(lerobot.available_datasets)
-
-    # You can also browse through the datasets created/ported by the community on the hub using the hub api:
+    # Browse datasets created/ported by the community on the hub using the hub api:
     hub_api = HfApi()
     repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
     pprint(repo_ids)

examples/dataset/slurm_compute_rabc.py

@@ -231,7 +231,7 @@ class AggregateProgress(PipelineStep):
         import pyarrow as pa
         import pyarrow.parquet as pq
 
-        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+        from lerobot.datasets import LeRobotDataset
         from lerobot.utils.utils import init_logging
 
         init_logging()

examples/dataset/use_dataset_image_transforms.py

@@ -26,8 +26,8 @@ import torch
 from torchvision.transforms import v2
 from torchvision.transforms.functional import to_pil_image
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.transforms import ImageTransformConfig, ImageTransforms, ImageTransformsConfig
+from lerobot.datasets import LeRobotDataset
+from lerobot.transforms import ImageTransformConfig, ImageTransforms, ImageTransformsConfig
 
 
 def save_image(tensor, filename):

examples/dataset/use_dataset_tools.py

@@ -29,7 +29,8 @@ Usage:
 
 import numpy as np
 
-from lerobot.datasets.dataset_tools import (
+from lerobot.datasets import (
+    LeRobotDataset,
     add_features,
     delete_episodes,
     merge_datasets,
@@ -37,7 +38,6 @@ from lerobot.datasets.dataset_tools import (
     remove_feature,
     split_dataset,
 )
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
 
 
 def main():

examples/hil/hil_data_collection.py

@@ -112,17 +112,18 @@ from hil_utils import (
     teleop_smooth_move_to,
 )
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
-from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.configs import parser
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.datasets.feature_utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
-from lerobot.datasets.image_writer import safe_stop_image_writer
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
-from lerobot.datasets.video_utils import VideoEncodingManager
-from lerobot.policies.factory import get_policy_class, make_policy, make_pre_post_processors
-from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.common.control_utils import is_headless, predict_action
+from lerobot.configs import PreTrainedConfig, parser
+from lerobot.datasets import (
+    LeRobotDataset,
+    VideoEncodingManager,
+    aggregate_pipeline_dataset_features,
+    create_initial_features,
+    safe_stop_image_writer,
+)
+from lerobot.policies import PreTrainedPolicy, get_policy_class, make_policy, make_pre_post_processors
 from lerobot.policies.rtc import ActionInterpolator, ActionQueue, LatencyTracker, RTCConfig
 from lerobot.policies.utils import make_robot_action
 from lerobot.processor import (
@@ -131,18 +132,18 @@ from lerobot.processor import (
     RelativeActionsProcessorStep,
     TransitionKey,
     create_transition,
+    rename_stats,
+    to_relative_actions,
 )
-from lerobot.processor.relative_action_processor import to_relative_actions
-from lerobot.processor.rename_processor import rename_stats
 from lerobot.robots import Robot, RobotConfig, make_robot_from_config
-from lerobot.robots.bi_openarm_follower.config_bi_openarm_follower import BiOpenArmFollowerConfig
-from lerobot.robots.so_follower.config_so_follower import SOFollowerRobotConfig  # noqa: F401
+from lerobot.robots.bi_openarm_follower import BiOpenArmFollowerConfig
+from lerobot.robots.so_follower import SOFollowerRobotConfig  # noqa: F401
 from lerobot.teleoperators import Teleoperator, TeleoperatorConfig, make_teleoperator_from_config
-from lerobot.teleoperators.openarm_mini.config_openarm_mini import OpenArmMiniConfig  # noqa: F401
-from lerobot.teleoperators.so_leader.config_so_leader import SOLeaderTeleopConfig  # noqa: F401
+from lerobot.teleoperators.openarm_mini import OpenArmMiniConfig  # noqa: F401
+from lerobot.teleoperators.so_leader import SOLeaderTeleopConfig  # noqa: F401
+from lerobot.utils import get_safe_torch_device
 from lerobot.utils.constants import ACTION, OBS_STATE, OBS_STR
-from lerobot.utils.control_utils import is_headless, predict_action
-from lerobot.utils.device_utils import get_safe_torch_device
+from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import init_logging, log_say
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data

examples/hil/hil_utils.py

@@ -19,13 +19,12 @@ import time
 from dataclasses import dataclass, field
 from pathlib import Path
 
+from lerobot.common.control_utils import is_headless
 from lerobot.processor import (
     IdentityProcessorStep,
     RobotAction,
     RobotObservation,
     RobotProcessorPipeline,
-)
-from lerobot.processor.converters import (
     observation_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
@@ -33,7 +32,6 @@ from lerobot.processor.converters import (
 )
 from lerobot.robots import Robot
 from lerobot.teleoperators import Teleoperator
-from lerobot.utils.control_utils import is_headless
 from lerobot.utils.robot_utils import precise_sleep
 
 logger = logging.getLogger(__name__)

examples/lekiwi/evaluate.py

@@ -14,15 +14,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.datasets.feature_utils import hw_to_dataset_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.datasets import LeRobotDataset
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTPolicy
 from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.constants import ACTION, OBS_STR
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import hw_to_dataset_features
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/lekiwi/record.py

@@ -14,16 +14,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.datasets.feature_utils import hw_to_dataset_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.datasets import LeRobotDataset
 from lerobot.processor import make_default_processors
-from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
-from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient
+from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.constants import ACTION, OBS_STR
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import hw_to_dataset_features
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/lekiwi/replay.py

@@ -16,9 +16,8 @@
 
 import time
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
-from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient
+from lerobot.datasets import LeRobotDataset
+from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say

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
+}

examples/phone_to_so100/evaluate.py

@@ -14,19 +14,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.datasets.feature_utils import combine_feature_dicts
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.configs import FeatureType, PolicyFeature
+from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTPolicy
 from lerobot.processor import (
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
-)
-from lerobot.processor.converters import (
     observation_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
@@ -39,7 +36,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 )
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/phone_to_so100/record.py

@@ -14,13 +14,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import combine_feature_dicts
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     observation_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
@@ -35,11 +34,11 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.scripts.lerobot_record import record_loop
-from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone import Phone, PhoneConfig
+from lerobot.teleoperators.phone.config_phone import PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
-from lerobot.teleoperators.phone.teleop_phone import Phone
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/phone_to_so100/replay.py

@@ -16,10 +16,10 @@
 
 import time
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     robot_action_observation_to_transition,
     transition_to_robot_action,
 )

examples/phone_to_so100/teleoperate.py

@@ -16,8 +16,8 @@
 import time
 
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     robot_action_observation_to_transition,
     transition_to_robot_action,
 )
@@ -28,9 +28,9 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     GripperVelocityToJoint,
     InverseKinematicsEEToJoints,
 )
-from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+from lerobot.teleoperators.phone import Phone, PhoneConfig
+from lerobot.teleoperators.phone.config_phone import PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
-from lerobot.teleoperators.phone.teleop_phone import Phone
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data

examples/port_datasets/port_droid.py

@@ -22,8 +22,7 @@ from pathlib import Path
 import numpy as np
 import tensorflow_datasets as tfds
 
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.utils.utils import get_elapsed_time_in_days_hours_minutes_seconds
 
 DROID_SHARDS = 2048

examples/port_datasets/slurm_aggregate_shards.py

@@ -36,7 +36,7 @@ class AggregateDatasets(PipelineStep):
     def run(self, data=None, rank: int = 0, world_size: int = 1):
         import logging
 
-        from lerobot.datasets.aggregate import aggregate_datasets
+        from lerobot.datasets import aggregate_datasets
         from lerobot.utils.utils import init_logging
 
         init_logging()

examples/port_datasets/slurm_upload.py

@@ -26,8 +26,7 @@ from huggingface_hub import HfApi
 from huggingface_hub.constants import REPOCARD_NAME
 from port_droid import DROID_SHARDS
 
-from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
-from lerobot.datasets.utils import create_lerobot_dataset_card
+from lerobot.datasets import CODEBASE_VERSION, LeRobotDatasetMetadata, create_lerobot_dataset_card
 from lerobot.utils.utils import init_logging
 
 
@@ -155,7 +154,7 @@ class UploadDataset(PipelineStep):
         from datasets.utils.tqdm import disable_progress_bars
         from huggingface_hub import CommitOperationAdd, preupload_lfs_files
 
-        from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+        from lerobot.datasets import LeRobotDatasetMetadata
         from lerobot.utils.utils import init_logging
 
         init_logging()

examples/rtc/eval_dataset.py

@@ -109,15 +109,10 @@ except ImportError:
     MATPLOTLIB_AVAILABLE = False
     plt = None
 
-from lerobot.configs import parser
-from lerobot.configs.default import DatasetConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import RTCAttentionSchedule
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.factory import resolve_delta_timestamps
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.factory import get_policy_class, make_pre_post_processors
-from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.configs import DatasetConfig, PreTrainedConfig, RTCAttentionSchedule, parser
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata, resolve_delta_timestamps
+from lerobot.policies import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc import RTCConfig
 from lerobot.policies.rtc.debug_visualizer import RTCDebugVisualizer
 from lerobot.utils.hub import HubMixin
 from lerobot.utils.utils import init_logging

examples/rtc/eval_with_real_robot.py

@@ -101,26 +101,21 @@ from threading import Event, Lock, Thread
 import torch
 from torch import Tensor
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
-from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.cameras.zmq.configuration_zmq import ZMQCameraConfig  # noqa: F401
-from lerobot.configs import parser
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import RTCAttentionSchedule
-from lerobot.datasets.feature_utils import build_dataset_frame, hw_to_dataset_features
-from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.cameras.zmq import ZMQCameraConfig  # noqa: F401
+from lerobot.configs import PreTrainedConfig, RTCAttentionSchedule, parser
+from lerobot.policies import get_policy_class, make_pre_post_processors
 from lerobot.policies.rtc import ActionInterpolator, ActionQueue, LatencyTracker, RTCConfig
 from lerobot.processor import (
     NormalizerProcessorStep,
     RelativeActionsProcessorStep,
     TransitionKey,
     create_transition,
-)
-from lerobot.processor.factory import (
     make_default_robot_action_processor,
     make_default_robot_observation_processor,
+    to_relative_actions,
 )
-from lerobot.processor.relative_action_processor import to_relative_actions
 from lerobot.rl.process import ProcessSignalHandler
 from lerobot.robots import (  # noqa: F401
     Robot,
@@ -133,6 +128,7 @@ from lerobot.robots import (  # noqa: F401
 )
 from lerobot.robots.utils import make_robot_from_config
 from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
+from lerobot.utils.feature_utils import build_dataset_frame, hw_to_dataset_features
 from lerobot.utils.hub import HubMixin
 from lerobot.utils.utils import init_logging
 

examples/so100_to_so100_EE/evaluate.py

@@ -14,19 +14,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.datasets.feature_utils import combine_feature_dicts
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.configs import FeatureType, PolicyFeature
+from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTPolicy
 from lerobot.processor import (
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
-)
-from lerobot.processor.converters import (
     observation_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
@@ -39,7 +36,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 )
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/so100_to_so100_EE/record.py

@@ -15,13 +15,12 @@
 # limitations under the License.
 
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import combine_feature_dicts
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.common.control_utils import init_keyboard_listener
+from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     observation_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
@@ -36,7 +35,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/so100_to_so100_EE/replay.py

@@ -17,10 +17,10 @@
 
 import time
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     robot_action_observation_to_transition,
     transition_to_robot_action,
 )

examples/so100_to_so100_EE/teleoperate.py

@@ -17,8 +17,8 @@
 import time
 
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
-from lerobot.processor.converters import (
+from lerobot.processor import (
+    RobotProcessorPipeline,
     robot_action_observation_to_transition,
     robot_action_to_transition,
     transition_to_robot_action,

examples/training/train_policy.py

@@ -18,13 +18,11 @@ from pathlib import Path
 
 import torch
 
-from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
-from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs import FeatureType
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.diffusion import DiffusionConfig, DiffusionPolicy
+from lerobot.utils.feature_utils import dataset_to_policy_features
 
 
 def main():

examples/training/train_with_streaming.py

@@ -19,14 +19,12 @@ from pathlib import Path
 
 import torch
 
-from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
-from lerobot.policies.act.configuration_act import ACTConfig
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs import FeatureType
+from lerobot.datasets import LeRobotDatasetMetadata, StreamingLeRobotDataset
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTConfig, ACTPolicy
 from lerobot.utils.constants import ACTION
+from lerobot.utils.feature_utils import dataset_to_policy_features
 
 
 def main():

examples/tutorial/act/act_training_example.py

@@ -4,13 +4,11 @@ from pathlib import Path
 
 import torch
 
-from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.act.configuration_act import ACTConfig
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs import FeatureType
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTConfig, ACTPolicy
+from lerobot.utils.feature_utils import dataset_to_policy_features
 
 
 def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:

examples/tutorial/act/act_using_example.py

@@ -1,9 +1,9 @@
 import torch
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.datasets import LeRobotDatasetMetadata
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.act import ACTPolicy
 from lerobot.policies.utils import build_inference_frame, make_robot_action
 from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 

examples/tutorial/async-inf/robot_client.py

@@ -3,7 +3,7 @@ import threading
 from lerobot.async_inference.configs import RobotClientConfig
 from lerobot.async_inference.helpers import visualize_action_queue_size
 from lerobot.async_inference.robot_client import RobotClient
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.robots.so_follower import SO100FollowerConfig
 
 

examples/tutorial/diffusion/diffusion_training_example.py

@@ -4,13 +4,11 @@ from pathlib import Path
 
 import torch
 
-from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
-from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs import FeatureType
+from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.diffusion import DiffusionConfig, DiffusionPolicy
+from lerobot.utils.feature_utils import dataset_to_policy_features
 
 
 def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:

examples/tutorial/diffusion/diffusion_using_example.py

@@ -1,9 +1,9 @@
 import torch
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
-from lerobot.policies.factory import make_pre_post_processors
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.datasets import LeRobotDatasetMetadata
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.diffusion import DiffusionPolicy
 from lerobot.policies.utils import build_inference_frame, make_robot_action
 from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 

examples/tutorial/pi0/using_pi0_example.py

@@ -1,11 +1,11 @@
 import torch
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import hw_to_dataset_features
-from lerobot.policies.factory import make_pre_post_processors
-from lerobot.policies.pi0.modeling_pi0 import PI0Policy
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.pi0 import PI0Policy
 from lerobot.policies.utils import build_inference_frame, make_robot_action
 from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.utils.feature_utils import hw_to_dataset_features
 
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20

examples/tutorial/rl/hilserl_example.py

@@ -6,17 +6,17 @@ from queue import Empty, Full
 import torch
 import torch.optim as optim
 
-from lerobot.datasets.feature_utils import hw_to_dataset_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets import LeRobotDataset
 from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
-from lerobot.policies.sac.configuration_sac import SACConfig
+from lerobot.policies import SACConfig
 from lerobot.policies.sac.modeling_sac import SACPolicy
 from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
 from lerobot.robots.so_follower import SO100FollowerConfig
+from lerobot.teleoperators import TeleopEvents
 from lerobot.teleoperators.so_leader import SO100LeaderConfig
-from lerobot.teleoperators.utils import TeleopEvents
+from lerobot.utils.feature_utils import hw_to_dataset_features
 
 LOG_EVERY = 10
 SEND_EVERY = 10

examples/tutorial/rl/reward_classifier_example.py

@@ -1,8 +1,7 @@
 import torch
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.factory import make_policy, make_pre_post_processors
-from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
+from lerobot.datasets import LeRobotDataset
+from lerobot.policies import RewardClassifierConfig, make_policy, make_pre_post_processors
 
 
 def main():

examples/tutorial/smolvla/using_smolvla_example.py

@@ -1,11 +1,11 @@
 import torch
 
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import hw_to_dataset_features
-from lerobot.policies.factory import make_pre_post_processors
-from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
+from lerobot.cameras.opencv import OpenCVCameraConfig
+from lerobot.policies import make_pre_post_processors
+from lerobot.policies.smolvla import SmolVLAPolicy
 from lerobot.policies.utils import build_inference_frame, make_robot_action
 from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.utils.feature_utils import hw_to_dataset_features
 
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20

pyproject.toml

@@ -58,45 +58,74 @@ classifiers = [
 keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artificial intelligence"]
 
 dependencies = [
-
-    # Hugging Face dependencies
-    "datasets>=4.0.0,<5.0.0",
-    "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub>=1.0.0,<2.0.0",
-    "accelerate>=1.10.0,<2.0.0",
-
-    # Core dependencies
-    "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
-    "setuptools>=71.0.0,<81.0.0",
-    "cmake>=3.29.0.1,<4.2.0",
-    "packaging>=24.2,<26.0",
-
+    # Core ML
     "torch>=2.7,<2.11.0",
-    "torchcodec>=0.3.0,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # NOTE: Windows support starts at version 0.7 (needs torch==2.8), ffmpeg>=8 support starts at version 0.8.1 (needs torch==2.9), system-wide ffmpeg support starts at version 0.10 (needs torch==2.10).
     "torchvision>=0.22.0,<0.26.0",
-
-    "einops>=0.8.0,<0.9.0",
+    "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
     "opencv-python-headless>=4.9.0,<4.14.0",
-    "av>=15.0.0,<16.0.0",
-    "jsonlines>=4.0.0,<5.0.0",
-    "pynput>=1.7.8,<1.9.0",
-    "pyserial>=3.5,<4.0",
+    "Pillow>=10.0.0,<13.0.0",
+    "einops>=0.8.0,<0.9.0",
 
-    "wandb>=0.24.0,<0.25.0",
+    # Config & Hub
     "draccus==0.10.0", # TODO: Relax version constraint
-    "gymnasium>=1.1.1,<2.0.0",
-    "rerun-sdk>=0.24.0,<0.27.0",
+    "huggingface-hub>=1.0.0,<2.0.0",
+    "requests>=2.32.0,<3.0.0",
 
-    # Support dependencies
-    "deepdiff>=7.0.1,<9.0.0",
-    "imageio[ffmpeg]>=2.34.0,<3.0.0",
+    # Environments
+    # NOTE: gymnasium is used in lerobot.envs (lerobot-train, lerobot-eval), policies/factory,
+    # and robots/unitree. Moving it to an optional extra would require import guards across many
+    # tightly-coupled modules. Candidate for a future refactor to decouple envs from the core.
+    "gymnasium>=1.1.1,<2.0.0",
+
+    # Serialization & checkpointing
+    "safetensors>=0.4.3,<1.0.0",
+
+    # Lightweight utilities
+    "packaging>=24.2,<26.0",
     "termcolor>=2.4.0,<4.0.0",
+    "tqdm>=4.66.0,<5.0.0",
+
+    # Build tools (required by opencv-python-headless on some platforms)
+    "cmake>=3.29.0.1,<4.2.0",
+    "setuptools>=71.0.0,<81.0.0",
 ]
 
 # Optional dependencies
 [project.optional-dependencies]
 
+# ── Feature-scoped extras ──────────────────────────────────
+dataset = [
+    "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]",
+    "torchcodec>=0.3.0,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # NOTE: Windows support starts at version 0.7 (needs torch==2.8), ffmpeg>=8 support starts at version 0.8.1 (needs torch==2.9), system-wide ffmpeg support starts at version 0.10 (needs torch==2.10).
+    "jsonlines>=4.0.0,<5.0.0",
+]
+training = [
+    "lerobot[dataset]",
+    "accelerate>=1.10.0,<2.0.0",
+    "wandb>=0.24.0,<0.25.0",
+]
+hardware = [
+    "lerobot[pynput-dep]",
+    "lerobot[pyserial-dep]",
+    "lerobot[deepdiff-dep]",
+]
+viz = [
+    "rerun-sdk>=0.24.0,<0.27.0",
+]
+# ── User-facing composite extras (map to CLI scripts) ─────
+# lerobot-record, lerobot-replay, lerobot-calibrate, lerobot-teleoperate, etc.
+core_scripts = ["lerobot[dataset]", "lerobot[hardware]", "lerobot[viz]"]
+# lerobot-eval -- base evaluation framework. You also need the policy's extra (e.g., lerobot[pi])
+# and the environment's extra (e.g., lerobot[pusht]) if evaluating in simulation.
+evaluation = ["lerobot[av-dep]"]
+# lerobot-dataset-viz, lerobot-imgtransform-viz
+dataset_viz = ["lerobot[dataset]", "lerobot[viz]"]
+
 # Common
+av-dep = ["av>=15.0.0,<16.0.0"]
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.9.17"]
 transformers-dep = ["transformers==5.3.0"] # TODO(Steven): https://github.com/huggingface/lerobot/pull/3249
@@ -104,12 +133,17 @@ 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"]
 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]"]
 
@@ -117,10 +151,11 @@ 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",
@@ -136,28 +171,28 @@ intelrealsense = [
 phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0", "lerobot[scipy-dep]"]
 
 # Policies
+diffusion = ["lerobot[diffusers-dep]"]
 wallx = [
     "lerobot[transformers-dep]",
-    "lerobot[peft]",
+    "lerobot[peft-dep]",
     "lerobot[scipy-dep]",
     "torchdiffeq>=0.2.4,<0.3.0",
     "lerobot[qwen-vl-utils-dep]",
 ]
 pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
-smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0", "safetensors>=0.4.3,<1.0.0"]
-multi_task_dit = ["lerobot[transformers-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 = [
     "lerobot[transformers-dep]",
-    "lerobot[peft]",
+    "lerobot[peft-dep]",
+    "lerobot[diffusers-dep]",
     "dm-tree>=0.1.8,<1.0.0",
     "timm>=1.0.0,<1.1.0",
-    "safetensors>=0.4.3,<1.0.0",
-    "Pillow>=10.0.0,<13.0.0",
     "decord>=0.6.0,<1.0.0; (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
     "ninja>=1.11.1,<2.0.0",
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
-sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
+sarm = ["lerobot[transformers-dep]", "pydantic>=2.0.0,<3.0.0", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
 xvla = ["lerobot[transformers-dep]"]
 hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
@@ -165,32 +200,75 @@ 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"]
+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"]
 
 # Simulation
 # NOTE: Explicitly listing scipy helps flatten the dependecy tree.
-aloha = ["gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
-pusht = ["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[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
-metaworld = ["metaworld==3.0.0", "lerobot[scipy-dep]"]
+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 = [
+    # Feature-scoped extras
+    "lerobot[dataset]",
+    "lerobot[training]",
+    "lerobot[hardware]",
+    "lerobot[viz]",
     # NOTE(resolver hint): scipy is pulled in transitively via lerobot[scipy-dep] through
     # multiple extras (aloha, metaworld, pi, wallx, phone). Listing it explicitly
     # helps pip's resolver converge by constraining scipy early, before it encounters
     # the loose scipy requirements from transitive deps like dm-control and metaworld.
     "scipy>=1.14.0,<2.0.0",
     "lerobot[dynamixel]",
+    "lerobot[feetech]",
+    "lerobot[damiao]",
+    "lerobot[robstride]",
     "lerobot[gamepad]",
     "lerobot[hopejr]",
     "lerobot[lekiwi]",
+    "lerobot[openarms]",
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
+    "lerobot[diffusion]",
+    "lerobot[multi_task_dit]",
     "lerobot[wallx]",
     "lerobot[pi]",
     "lerobot[smolvla]",
@@ -228,10 +306,12 @@ 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"]
@@ -267,7 +347,9 @@ ignore = [
 ]
 
 [tool.ruff.lint.per-file-ignores]
-"__init__.py" = ["F401", "F403"]
+"__init__.py" = ["F401", "F403", "E402"]
+# E402: conditional-import guards (TYPE_CHECKING / is_package_available) must precede the imports they protect
+"src/lerobot/scripts/convert_dataset_v21_to_v30.py" = ["E402"]
 "src/lerobot/policies/wall_x/**" = ["N801", "N812", "SIM102", "SIM108", "SIM210", "SIM211", "B006", "B007", "SIM118"] # Supprese these as they are coming from original Qwen2_5_vl code TODO(pepijn): refactor original
 
 [tool.ruff.lint.isort]

scripts/ci/extract_task_descriptions.py

@@ -0,0 +1,207 @@
+#!/usr/bin/env python3
+# 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.
+
+"""Extract natural-language task descriptions for a benchmark suite.
+
+Runs inside the benchmark Docker container (where the env library is installed)
+immediately after lerobot-eval, writing a JSON file that parse_eval_metrics.py
+picks up and embeds in metrics.json.
+
+Output format: {"<suite>_<task_idx>": "<nl instruction>", ...}
+
+Usage:
+    python scripts/ci/extract_task_descriptions.py \\
+        --env libero --task libero_spatial \\
+        --output /tmp/eval-artifacts/task_descriptions.json
+"""
+
+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]
+
+    suite_dict = benchmark.get_benchmark_dict()
+    if task_suite not in suite_dict:
+        print(
+            f"[extract_task_descriptions] Unknown LIBERO suite '{task_suite}'. "
+            f"Available: {list(suite_dict.keys())}",
+            file=sys.stderr,
+        )
+        return {}
+    suite = suite_dict[task_suite]()
+    return {
+        f"{task_suite}_{i}": _strip_libero_perturbation_tail(suite.get_task(i).language)
+        for i in range(suite.n_tasks)
+    }
+
+
+def _metaworld_descriptions(task_name: str) -> dict[str, str]:
+    # MetaWorld tasks don't expose a separate NL description attribute;
+    # use a cleaned version of the task name as the description.
+    label = task_name.removeprefix("metaworld-").replace("-", " ").strip()
+    return {f"{task_name}_0": label}
+
+
+def _robotwin_descriptions(task_names: str) -> dict[str, str]:
+    """Return descriptions for each requested RoboTwin task. Reads
+    `description/task_instruction/<task>.json` from the RoboTwin clone
+    (cwd is /opt/robotwin in CI). Falls back to the task name if missing."""
+    out: dict[str, str] = {}
+    root = Path("description/task_instruction")
+    for name in (t.strip() for t in task_names.split(",") if t.strip()):
+        desc_file = root / f"{name}.json"
+        desc = name.replace("_", " ")
+        if desc_file.is_file():
+            data = json.loads(desc_file.read_text())
+            full = data.get("full_description") or desc
+            # Strip the schema placeholders ({A}, {a}) — keep the sentence readable.
+            desc = full.replace("<", "").replace(">", "")
+        out[f"{name}_0"] = desc
+    return out
+
+
+def _robocasa_descriptions(task_spec: str) -> dict[str, str]:
+    """For each task in the comma-separated list, emit a cleaned-name label.
+
+    RoboCasa episodes carry their language instruction in the env's
+    `ep_meta['lang']`, populated per reset. Pulling it requires spinning
+    up the full kitchen env per task (~seconds each); we use the task
+    name as the key here and let the eval's episode info carry the
+    actual instruction.
+    """
+    out: dict[str, str] = {}
+    for task in (t.strip() for t in task_spec.split(",") if t.strip()):
+        # Split CamelCase into words: "CloseFridge" → "close fridge".
+        label = "".join(f" {c.lower()}" if c.isupper() else c for c in task).strip()
+        out[f"{task}_0"] = label or task
+    return out
+
+
+_ROBOMME_DESCRIPTIONS = {
+    "BinFill": "Fill the target bin with the correct number of cubes",
+    "PickXtimes": "Pick the indicated cube the specified number of times",
+    "SwingXtimes": "Swing the object the specified number of times",
+    "StopCube": "Grasp and stop the moving cube",
+    "VideoUnmask": "Pick the cube shown in the reference video",
+    "VideoUnmaskSwap": "Pick the cube matching the reference video after a swap",
+    "ButtonUnmask": "Press the button indicated by the reference",
+    "ButtonUnmaskSwap": "Press the correct button after objects are swapped",
+    "PickHighlight": "Pick the highlighted cube",
+    "VideoRepick": "Repick the cube shown in the reference video",
+    "VideoPlaceButton": "Place the cube on the button shown in the video",
+    "VideoPlaceOrder": "Place cubes in the order shown in the video",
+    "MoveCube": "Move the cube to the target location",
+    "InsertPeg": "Insert the peg into the target hole",
+    "PatternLock": "Unlock the pattern by pressing buttons in sequence",
+    "RouteStick": "Route the stick through the required waypoints",
+}
+
+
+def _robomme_descriptions(task_names: str, task_ids: list[int] | None = None) -> dict[str, str]:
+    """Return descriptions for each requested RoboMME task. Keys match the
+    video filename pattern `<task>_<task_id>` used by the eval script."""
+    if task_ids is None:
+        task_ids = [0]
+    out: dict[str, str] = {}
+    for name in (t.strip() for t in task_names.split(",") if t.strip()):
+        desc = _ROBOMME_DESCRIPTIONS.get(name, name)
+        for tid in task_ids:
+            out[f"{name}_{tid}"] = desc
+    return out
+
+
+def _vlabench_descriptions(task_spec: str) -> dict[str, str]:
+    """For each task in the comma-separated list, emit a cleaned-name label.
+
+    VLABench tasks carry language instructions on their dm_control task
+    object, but pulling them requires loading the full env per task
+    (~seconds each). The CI smoke-eval already captures the instruction
+    inside its episode info; this mapping is just enough to key
+    `metrics.json` by `<task>_0`.
+    """
+    out: dict[str, str] = {}
+    for task in (t.strip() for t in task_spec.split(",") if t.strip()):
+        out[f"{task}_0"] = task.replace("_", " ").strip()
+    return out
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--env", required=True, help="Environment family (libero, metaworld, ...)")
+    parser.add_argument("--task", required=True, help="Task/suite name (e.g. libero_spatial)")
+    parser.add_argument(
+        "--task-ids",
+        type=str,
+        default=None,
+        help="Comma-separated task IDs (e.g. '0,1,2'). Default: [0]",
+    )
+    parser.add_argument("--output", required=True, help="Path to write task_descriptions.json")
+    args = parser.parse_args()
+
+    task_ids: list[int] | None = None
+    if args.task_ids:
+        task_ids = [int(x.strip()) for x in args.task_ids.split(",")]
+
+    descriptions: dict[str, str] = {}
+    try:
+        if args.env == ("libero", "libero_plus"):
+            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}'.",
+                file=sys.stderr,
+            )
+    except Exception as exc:
+        print(f"[extract_task_descriptions] Warning: {exc}", file=sys.stderr)
+
+    out_path = Path(args.output)
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    out_path.write_text(json.dumps(descriptions, indent=2))
+    print(f"[extract_task_descriptions] {len(descriptions)} descriptions → {out_path}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

scripts/ci/parse_eval_metrics.py

@@ -0,0 +1,147 @@
+#!/usr/bin/env python3
+# 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.
+
+"""Parse lerobot-eval output into a small metrics.json artifact.
+
+Reads eval_info.json written by lerobot-eval --output_dir and extracts the
+key metrics needed by the health dashboard. Handles both single-task and
+multi-task eval output formats.
+
+NOTE: This script runs on the bare CI runner (not inside Docker), so it
+must use only Python stdlib modules. Do not add third-party imports.
+
+Usage:
+    python scripts/ci/parse_eval_metrics.py \\
+        --artifacts-dir /tmp/libero-artifacts \\
+        --env libero \\
+        --task libero_spatial \\
+        --policy pepijn223/smolvla_libero
+
+Writes <artifacts-dir>/metrics.json. The CI workflow then uploads this file
+as a GitHub Actions artifact named "<env>-metrics".
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import sys
+from pathlib import Path
+
+
+def _safe_float(v: float | int | None) -> float | None:
+    if v is None:
+        return None
+    f = float(v)
+    return None if math.isnan(f) else f
+
+
+def _safe_int(v: float | int | None) -> int | None:
+    if v is None:
+        return None
+    f = float(v)
+    return None if math.isnan(f) else int(f)
+
+
+def _extract_metrics(info: dict) -> tuple[float | None, int | None, float | None, float | None]:
+    """Extract (pc_success, n_episodes, avg_sum_reward, eval_s) from eval_info.json.
+
+    Handles two output shapes:
+      - Single-task: {"aggregated": {"pc_success": 80.0, ...}}
+      - Multi-task:  {"overall": {"pc_success": 80.0, "n_episodes": 5, ...}}
+    """
+    for key in ("aggregated", "overall"):
+        if key not in info:
+            continue
+        agg = info[key]
+        pc = agg.get("pc_success")
+        n = agg.get("n_episodes")
+        reward = agg.get("avg_sum_reward")
+        eval_s = agg.get("eval_s")
+
+        if pc is not None and not math.isnan(pc):
+            return (
+                float(pc),
+                _safe_int(n),
+                _safe_float(reward),
+                _safe_float(eval_s),
+            )
+
+    return None, None, None, None
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    parser.add_argument("--artifacts-dir", required=True, help="Path to the mounted artifacts volume")
+    parser.add_argument("--env", required=True, help="Environment name (e.g. libero)")
+    parser.add_argument("--task", required=True, help="Task name (e.g. libero_spatial)")
+    parser.add_argument("--policy", required=True, help="Policy hub path (e.g. pepijn223/smolvla_libero)")
+    args = parser.parse_args()
+
+    artifacts_dir = Path(args.artifacts_dir)
+    eval_info_path = artifacts_dir / "eval_info.json"
+
+    pc_success: float | None = None
+    n_episodes: int | None = None
+    avg_sum_reward: float | None = None
+    eval_s: float | None = None
+
+    if eval_info_path.exists():
+        try:
+            info = json.loads(eval_info_path.read_text())
+            pc_success, n_episodes, avg_sum_reward, eval_s = _extract_metrics(info)
+        except (json.JSONDecodeError, KeyError, TypeError) as exc:
+            print(f"[parse_eval_metrics] Warning: could not parse eval_info.json: {exc}", file=sys.stderr)
+    else:
+        print(
+            f"[parse_eval_metrics] Warning: {eval_info_path} not found — eval may have failed.",
+            file=sys.stderr,
+        )
+
+    task_descriptions: dict[str, str] = {}
+    task_desc_path = artifacts_dir / "task_descriptions.json"
+    if task_desc_path.exists():
+        try:
+            task_descriptions = json.loads(task_desc_path.read_text())
+        except json.JSONDecodeError as exc:
+            print(
+                f"[parse_eval_metrics] Warning: could not parse task_descriptions.json: {exc}",
+                file=sys.stderr,
+            )
+
+    metrics = {
+        "env": args.env,
+        "task": args.task,
+        "policy": args.policy,
+        "pc_success": pc_success,
+        "n_episodes": n_episodes,
+        "avg_sum_reward": avg_sum_reward,
+        "eval_s": eval_s,
+        "task_descriptions": task_descriptions,
+    }
+
+    out_path = artifacts_dir / "metrics.json"
+    out_path.write_text(json.dumps(metrics, indent=2))
+    print(f"[parse_eval_metrics] Written: {out_path}")
+    print(json.dumps(metrics, indent=2))
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

src/lerobot/__init__.py

@@ -13,188 +13,39 @@
 # 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.
+
 """
-This file contains lists of available environments, dataset and policies to reflect the current state of LeRobot library.
-We do not want to import all the dependencies, but instead we keep it lightweight to ensure fast access to these variables.
+LeRobot -- PyTorch library for real-world robotics.
 
-Example:
-    ```python
-        import lerobot
-        print(lerobot.available_envs)
-        print(lerobot.available_tasks_per_env)
-        print(lerobot.available_datasets)
-        print(lerobot.available_datasets_per_env)
-        print(lerobot.available_real_world_datasets)
-        print(lerobot.available_policies)
-        print(lerobot.available_policies_per_env)
-        print(lerobot.available_robots)
-        print(lerobot.available_cameras)
-        print(lerobot.available_motors)
-    ```
+Provides datasets, pretrained policies, and tools for training, evaluation,
+data collection, and robot control. Integrates with Hugging Face Hub for
+model and dataset sharing.
 
-When implementing a new dataset loadable with LeRobotDataset follow these steps:
-- Update `available_datasets_per_env` in `lerobot/__init__.py`
+The base install is intentionally lightweight. Feature-specific dependencies
+are gated behind optional extras::
 
-When implementing a new environment (e.g. `gym_aloha`), follow these steps:
-- Update `available_tasks_per_env` and `available_datasets_per_env` in `lerobot/__init__.py`
-
-When implementing a new policy class (e.g. `DiffusionPolicy`) follow these steps:
-- Update `available_policies` and `available_policies_per_env`, in `lerobot/__init__.py`
-- Set the required `name` class attribute.
-- Update variables in `tests/test_available.py` by importing your new Policy class
+    pip install 'lerobot[dataset]'       # dataset loading & creation
+    pip install 'lerobot[training]'      # training loop + wandb
+    pip install 'lerobot[hardware]'      # real robot control
+    pip install 'lerobot[core_scripts]'  # dataset + hardware + viz (record, replay, calibrate, etc.)
+    pip install 'lerobot[all]'           # everything
 """
 
-import itertools
+from lerobot.__version__ import __version__
 
-from lerobot.__version__ import __version__  # noqa: F401
-
-# TODO(rcadene): Improve policies and envs. As of now, an item in `available_policies`
-# refers to a yaml file AND a modeling name. Same for `available_envs` which refers to
-# a yaml file AND a environment name. The difference should be more obvious.
-available_tasks_per_env = {
-    "aloha": [
-        "AlohaInsertion-v0",
-        "AlohaTransferCube-v0",
+# Maps optional extras to the CLI entry-points they unlock.
+available_extras: dict[str, list[str]] = {
+    "dataset": ["lerobot-dataset-viz", "lerobot-imgtransform-viz", "lerobot-edit-dataset"],
+    "training": ["lerobot-train"],
+    "hardware": [
+        "lerobot-calibrate",
+        "lerobot-find-port",
+        "lerobot-find-cameras",
+        "lerobot-find-joint-limits",
+        "lerobot-setup-motors",
     ],
-    "pusht": ["PushT-v0"],
-}
-available_envs = list(available_tasks_per_env.keys())
-
-available_datasets_per_env = {
-    "aloha": [
-        "lerobot/aloha_sim_insertion_human",
-        "lerobot/aloha_sim_insertion_scripted",
-        "lerobot/aloha_sim_transfer_cube_human",
-        "lerobot/aloha_sim_transfer_cube_scripted",
-        "lerobot/aloha_sim_insertion_human_image",
-        "lerobot/aloha_sim_insertion_scripted_image",
-        "lerobot/aloha_sim_transfer_cube_human_image",
-        "lerobot/aloha_sim_transfer_cube_scripted_image",
-    ],
-    # TODO(alexander-soare): Add "lerobot/pusht_keypoints". Right now we can't because this is too tightly
-    # coupled with tests.
-    "pusht": ["lerobot/pusht", "lerobot/pusht_image"],
+    "core_scripts": ["lerobot-record", "lerobot-replay", "lerobot-teleoperate"],
+    "evaluation": ["lerobot-eval"],
 }
 
-available_real_world_datasets = [
-    "lerobot/aloha_mobile_cabinet",
-    "lerobot/aloha_mobile_chair",
-    "lerobot/aloha_mobile_elevator",
-    "lerobot/aloha_mobile_shrimp",
-    "lerobot/aloha_mobile_wash_pan",
-    "lerobot/aloha_mobile_wipe_wine",
-    "lerobot/aloha_static_battery",
-    "lerobot/aloha_static_candy",
-    "lerobot/aloha_static_coffee",
-    "lerobot/aloha_static_coffee_new",
-    "lerobot/aloha_static_cups_open",
-    "lerobot/aloha_static_fork_pick_up",
-    "lerobot/aloha_static_pingpong_test",
-    "lerobot/aloha_static_pro_pencil",
-    "lerobot/aloha_static_screw_driver",
-    "lerobot/aloha_static_tape",
-    "lerobot/aloha_static_thread_velcro",
-    "lerobot/aloha_static_towel",
-    "lerobot/aloha_static_vinh_cup",
-    "lerobot/aloha_static_vinh_cup_left",
-    "lerobot/aloha_static_ziploc_slide",
-    "lerobot/umi_cup_in_the_wild",
-    "lerobot/unitreeh1_fold_clothes",
-    "lerobot/unitreeh1_rearrange_objects",
-    "lerobot/unitreeh1_two_robot_greeting",
-    "lerobot/unitreeh1_warehouse",
-    "lerobot/nyu_rot_dataset",
-    "lerobot/utokyo_saytap",
-    "lerobot/imperialcollege_sawyer_wrist_cam",
-    "lerobot/utokyo_xarm_bimanual",
-    "lerobot/tokyo_u_lsmo",
-    "lerobot/utokyo_pr2_opening_fridge",
-    "lerobot/cmu_franka_exploration_dataset",
-    "lerobot/cmu_stretch",
-    "lerobot/asu_table_top",
-    "lerobot/utokyo_pr2_tabletop_manipulation",
-    "lerobot/utokyo_xarm_pick_and_place",
-    "lerobot/ucsd_kitchen_dataset",
-    "lerobot/austin_buds_dataset",
-    "lerobot/dlr_sara_grid_clamp",
-    "lerobot/conq_hose_manipulation",
-    "lerobot/columbia_cairlab_pusht_real",
-    "lerobot/dlr_sara_pour",
-    "lerobot/dlr_edan_shared_control",
-    "lerobot/ucsd_pick_and_place_dataset",
-    "lerobot/berkeley_cable_routing",
-    "lerobot/nyu_franka_play_dataset",
-    "lerobot/austin_sirius_dataset",
-    "lerobot/cmu_play_fusion",
-    "lerobot/berkeley_gnm_sac_son",
-    "lerobot/nyu_door_opening_surprising_effectiveness",
-    "lerobot/berkeley_fanuc_manipulation",
-    "lerobot/jaco_play",
-    "lerobot/viola",
-    "lerobot/kaist_nonprehensile",
-    "lerobot/berkeley_mvp",
-    "lerobot/uiuc_d3field",
-    "lerobot/berkeley_gnm_recon",
-    "lerobot/austin_sailor_dataset",
-    "lerobot/utaustin_mutex",
-    "lerobot/roboturk",
-    "lerobot/stanford_hydra_dataset",
-    "lerobot/berkeley_autolab_ur5",
-    "lerobot/stanford_robocook",
-    "lerobot/toto",
-    "lerobot/fmb",
-    "lerobot/droid_100",
-    "lerobot/berkeley_rpt",
-    "lerobot/stanford_kuka_multimodal_dataset",
-    "lerobot/iamlab_cmu_pickup_insert",
-    "lerobot/taco_play",
-    "lerobot/berkeley_gnm_cory_hall",
-    "lerobot/usc_cloth_sim",
-]
-
-available_datasets = sorted(
-    set(itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets))
-)
-
-# lists all available policies from `lerobot/policies`
-available_policies = ["act", "diffusion", "tdmpc", "vqbet"]
-
-# lists all available robots from `lerobot/robots`
-available_robots = [
-    "koch",
-    "koch_bimanual",
-    "aloha",
-    "so100",
-    "so101",
-]
-
-# lists all available cameras from `lerobot/cameras`
-available_cameras = [
-    "opencv",
-    "intelrealsense",
-]
-
-# lists all available motors from `lerobot/motors`
-available_motors = [
-    "dynamixel",
-    "feetech",
-]
-
-# keys and values refer to yaml files
-available_policies_per_env = {
-    "aloha": ["act"],
-    "pusht": ["diffusion", "vqbet"],
-    "koch_real": ["act_koch_real"],
-    "aloha_real": ["act_aloha_real"],
-}
-
-env_task_pairs = [(env, task) for env, tasks in available_tasks_per_env.items() for task in tasks]
-env_dataset_pairs = [
-    (env, dataset) for env, datasets in available_datasets_per_env.items() for dataset in datasets
-]
-env_dataset_policy_triplets = [
-    (env, dataset, policy)
-    for env, datasets in available_datasets_per_env.items()
-    for dataset in datasets
-    for policy in available_policies_per_env[env]
-]
+__all__ = ["__version__", "available_extras"]

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 +1,443 @@
+#!/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 row at t=0
+        plan_text = self._generate_plan(record, subtask_spans, task=effective_task)
+        if plan_text is not None:
+            rows.append(
+                {
+                    "role": "assistant",
+                    "content": plan_text,
+                    "style": "plan",
+                    "timestamp": float(t0),
+                    "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,
+        subtask_spans: Sequence[dict[str, Any]],
+        *,
+        refresh_t: float | None = None,
+        interjection: str | None = None,
+        task: str | None = None,
+    ) -> str | None:
+        if not subtask_spans:
+            return None
+        subtasks_text = "\n".join(f"- {s['text']}" for s in subtask_spans)
+        prompt = load_prompt("module_1_plan").format(
+            episode_task=(task if task is not None else record.episode_task),
+            subtasks_text=subtasks_text,
+            plan_max_steps=self.config.plan_max_steps,
+        )
+        if refresh_t is not None:
+            # ``current_subtask`` is the span the refresh time falls into,
+            # so the model knows where in the demonstration the planner is
+            # standing when it re-emits.
+            current_subtask = ""
+            for span in subtask_spans:
+                if float(span["start"]) <= refresh_t and (
+                    "end" not in span or float(span["end"]) > refresh_t
+                ):
+                    current_subtask = span.get("text", "")
+                    break
+            if interjection:
+                prompt += (
+                    f"\n\n(Plan refresh at t={refresh_t:.2f}s after a user "
+                    f"interjection: {interjection!r}. Current subtask just "
+                    f"before the interjection: {current_subtask!r}. Update "
+                    f"the plan so it reflects the interjection — drop or "
+                    f"reorder steps as needed; do not just restate.)\n"
+                )
+            else:
+                # Refresh without an interjection text: still tell the model
+                # where in the episode the plan stands so the re-emission
+                # is grounded. Should be rare — plan refreshes are
+                # interjection-driven by design.
+                prompt += (
+                    f"\n\n(Plan refresh at t={refresh_t:.2f}s. Current "
+                    f"subtask: {current_subtask!r}.)\n"
+                )
+        messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}]
+        result = self.vlm.generate_json([messages])[0]
+        if isinstance(result, dict) and isinstance(result.get("plan"), str):
+            return result["plan"].strip()
+        return None
+
+    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 +1,31 @@
+You are updating the robot's compressed semantic memory at the boundary of
+a completed subtask.
+
+Reference (verbatim from 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."
+
+Concrete example from MEM:
+  Before: "I put a light green bowl, a dark blue bowl and a bright yellow
+           bowl into the top right cabinet"
+  After:  "I placed three bowls in the top right cabinet"
+
+Episode task: "{episode_task}"
+Previous memory: {prior_memory}
+Just-completed subtask: "{completed_subtask}"
+Remaining subtasks (for relevance judgement only): {remaining_subtasks}
+
+Update the memory as a compact state note.
+
+Rules:
+- Keep only facts needed later.
+- Keep WHAT changed; drop HOW it was done.
+- Use fragments when clear.
+- Prefer: "bowl in box; lid still open"
+- Avoid: "The robot placed the bowl into the box and the lid remains open."
+
+Output strictly valid JSON:
+  {{ "memory": "<brief state note>" }}

src/lerobot/annotations/steerable_pipeline/prompts/module_1_plan.txt

@@ -0,0 +1,18 @@
+You are the high-level planner for a robot demonstrating: "{episode_task}".
+
+Given the subtask decomposition below, write a compact hierarchical PLAN.
+Use short imperative fragments, like pi0.7 context prompts.
+
+Subtasks for context:
+{subtasks_text}
+
+Authoring rules:
+- 3 to {plan_max_steps} steps.
+- Each step is one logical chunk, not one motion.
+- Steps must be in execution order.
+- Brief commands, not full sentences.
+- Prefer: "open air fryer"; avoid: "The robot should open the air fryer."
+- Plain text, no markdown headers.
+
+Output strictly valid JSON:
+  {{ "plan": "1. ...\n2. ...\n3. ..." }}