fix(viz): use PyAV for AV1 video decoding, AutoImageProcessor for SigLIP

- Replace cv2.VideoCapture with PyAV (av library) which handles AV1 codec properly. Decode each video once and index by frame number. - Use AutoImageProcessor instead of AutoProcessor to avoid loading the SigLIP tokenizer (which requires sentencepiece). Made-with: Cursor
chore(viz): bump SigLIP batch size to 512 for H100
2026-05-12 07:09:43 +00:00 · 2026-03-23 23:02:50 -07:00 · 2026-03-23 20:31:02 -07:00 · 2026-03-23 20:29:36 -07:00 · 2026-03-23 20:17:59 -07:00 · 2026-03-23 20:15:15 -07:00
388 changed files with 27876 additions and 8139 deletions
@@ -22,20 +22,21 @@ Short, imperative summary (e.g., "fix(robots): handle None in sensor parser"). S
 - Short, concrete bullets of the modifications (files/behaviour).
 - Short note if this introduces breaking changes and migration steps.

-## How was this tested
+## How was this tested (or how to run locally)

 - Tests added: list new tests or test files.
 - Manual checks / dataset runs performed.
+- Instructions for the reviewer

-## How to run locally (reviewer)
+Example:

- Run the relevant tests:
+- Ran the relevant tests:

  ```bash
  pytest -q tests/ -k <keyword>
  ```

- Run a quick example or CLI (if applicable):
+- Reproduce with a quick example or CLI (if applicable):

  ```bash
  lerobot-train --some.option=true
@@ -18,6 +18,11 @@ name: Documentation
 on:
  # Allows running this workflow manually from the Actions tab
  workflow_dispatch:
+    inputs:
+      version:
+        description: 'Version tag (e.g. v0.1.2) - Leave empty for standard main build'
+        required: false
+        type: string

  # Triggers the workflow on push events to main for the docs folder
  push:
@@ -33,6 +38,9 @@ on:
    paths:
      - "docs/**"

+  release:
+    types: [published]
+
 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
 concurrency:
  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
@@ -43,7 +51,7 @@ jobs:
  build_main_docs:
    name: Build Main Docs
    if: >
-      (github.event_name == 'push' || github.event_name == 'workflow_dispatch') &&
+      (github.event_name == 'push' || github.event_name == 'workflow_dispatch' || github.event_name == 'release') &&
      github.repository == 'huggingface/lerobot'
    permissions:
      contents: read
@@ -51,7 +59,13 @@ jobs:
    with:
      commit_sha: ${{ github.sha }}
      package: lerobot
-      additional_args: --not_python_module
+      additional_args: >-
+        --not_python_module
+        ${{
+          (github.event_name == 'release' && format('--version {0}', github.event.release.tag_name)) ||
+          (inputs.version != '' && format('--version {0}', inputs.version)) ||
+          ''
+        }}
    secrets:
      token: ${{ secrets.HUGGINGFACE_PUSH }}
      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}
@@ -44,7 +44,7 @@ permissions:
 # Sets up the environment variables
 env:
  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"

 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
 concurrency:
@@ -61,8 +61,9 @@ jobs:
      MUJOCO_GL: egl
      HF_HOME: /mnt/cache/.cache/huggingface
      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
        with:
          persist-credentials: false
          lfs: true
@@ -89,5 +90,11 @@ jobs:
      - name: Install lerobot with test extras
        run: uv sync --extra "test"

+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami
+
      - name: Run pytest
        run: uv run pytest tests -vv --maxfail=10
@@ -37,7 +37,7 @@ permissions:
 # Sets up the environment variables
 env:
  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu

 # Ensures that only the latest action is built, canceling older runs.
@@ -60,8 +60,9 @@ jobs:
      MUJOCO_GL: egl
      HF_HOME: /mnt/cache/.cache/huggingface
      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
@@ -85,7 +86,13 @@ jobs:
          python-version: ${{ env.PYTHON_VERSION }}

      - name: Install lerobot with all extras
-        run: uv sync --all-extras --no-extra groot --no-extra wallx # TODO(Steven): Make flash-attn optional
+        run: uv sync --extra all # TODO(Steven): Make flash-attn optional
+
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami

      - name: Run pytest (all extras)
        run: uv run pytest tests -vv --maxfail=10
@@ -101,9 +108,11 @@ jobs:
    runs-on:
      group: aws-general-8-plus
    if: |
-      (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
-      github.event_name == 'push' ||
-      github.event_name == 'workflow_dispatch'
+      github.repository == 'huggingface/lerobot' && (
+        (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
+        github.event_name == 'push' ||
+        github.event_name == 'workflow_dispatch'
+      )
    outputs:
      image_tag: ${{ steps.set_tag.outputs.image_tag }}
    env:
@@ -127,7 +136,7 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
@@ -160,6 +169,7 @@ jobs:
      HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
      TORCH_HOME: /home/user_lerobot/.cache/torch
      TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    container:
      image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
      options: --gpus all --shm-size "16gb"
@@ -171,6 +181,13 @@ jobs:
        shell: bash
        working-directory: /lerobot
    steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
+      - name: Fix ptxas permissions
+        run: chmod +x /lerobot/.venv/lib/python3.12/site-packages/triton/backends/nvidia/bin/ptxas
      - name: Run pytest on GPU
        run: pytest tests -vv --maxfail=10
      - name: Run end-to-end tests
@@ -186,15 +203,18 @@ jobs:
    steps:
      - name: Get Docker Hub Token and Delete Image
        # zizmor: ignore[template-injection]
+        env:
+          DOCKERHUB_LEROBOT_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          DOCKERHUB_LEROBOT_PASSWORD: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+          IMAGE_FULL: ${{ needs.build-and-push-docker.outputs.image_tag }}
        run: |
-          IMAGE_NAME=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f1)
-          IMAGE_TAG=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f2)
-
+          IMAGE_NAME=$(echo "$IMAGE_FULL" | cut -d':' -f1)
+          IMAGE_TAG=$(echo "$IMAGE_FULL" | cut -d':' -f2-)
          echo "Attempting to delete image: $IMAGE_NAME:$IMAGE_TAG"

          TOKEN=$(curl -s -H "Content-Type: application/json" \
                       -X POST \
-                       -d '{"username": "${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}", "password": "${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}"}' \
+                       -d "{\"username\": \"$DOCKERHUB_LEROBOT_USERNAME\", \"password\": \"$DOCKERHUB_LEROBOT_PASSWORD\"}" \
                       https://hub.docker.com/v2/users/login/ | jq -r .token)

          if [ "$TOKEN" == "null" ] || [ -z "$TOKEN" ]; then
@@ -205,7 +225,7 @@ jobs:
          HTTP_RESPONSE=$(curl -s -o /dev/null -w "%{http_code}" \
                               -H "Authorization: JWT ${TOKEN}" \
                               -X DELETE \
-                               https://hub.docker.com/v2/repositories/${IMAGE_NAME}/tags/${IMAGE_TAG}/)
+                               https://hub.docker.com/v2/repositories/${IMAGE_NAME}/tags/$IMAGE_TAG)

          if [ "$HTTP_RESPONSE" -eq 204 ]; then
            echo "Successfully deleted Docker image tag: $IMAGE_NAME:$IMAGE_TAG"
@@ -42,38 +42,26 @@ jobs:

            // Keyword Heuristics

-            // Domain Specific
-            if (matches(/\b(bug|error|issue|fault|crash|exception)\b/i)) labelsToAdd.add('bug');
-            if (matches(/\b(feature|enhancement|improvement|support|implement|proposal)\b/i)) labelsToAdd.add('enhancement');
-            if (matches(/\b(question|help|how to||clarify|explain|unclear)\b/i)) labelsToAdd.add('question');
-            if (matches(/\b(maintenance|documentation|docs|readme|tutorial|guide|wiki)\b/i)) labelsToAdd.add('documentation');
-            if (matches(/\b(example|script|sample|demo|notebook)s?\b/i)) labelsToAdd.add('examples');
+            if (matches(/\b(bug|error|crash|exception)\b/i)) labelsToAdd.add('bug');
+            if (matches(/\b(new feature|enhancement|improvement|proposal|feature request)\b/i)) labelsToAdd.add('enhancement');
+            if (matches(/\b(question|how to|clarify|explain|how do i|help me|question about)\b/i)) labelsToAdd.add('question');
+            if (matches(/\b(documentation|docs?|readme|tutorial|wiki|typo|docstring)\b/i)) labelsToAdd.add('documentation');
+            if (matches(/\b(example|sample|demo|notebook)s?\b/i)) labelsToAdd.add('examples');
            if (matches(/\b(datasets?|data loader|data augmentation|data preprocessing)\b/i)) labelsToAdd.add('dataset');
            if (matches(/\b(mujoco|isaac|simulation|sim)\b/i)) labelsToAdd.add('simulation');
-            if (matches(/\b(train|training|loss|optimizer|backward|gradient|wandb|sac)\b/i)) labelsToAdd.add('training');
-            if (matches(/\b(rerun|plot|video|render|visualiz|gif)/i)) labelsToAdd.add('visualization');
-            if (matches(/\b(camera|realsense|lidar|depth|sensor|imu|microphone|rgbd)\b/i)) labelsToAdd.add('sensors');
-            if (matches(/\b(aloha|koch|so-100|so100|mobile|teleop|manipulator|robots?)\b/i)) labelsToAdd.add('robots');
+            if (matches(/\b(train|training|optimizer|gradient|wandb|sac)\b/i)) labelsToAdd.add('training');
+            if (matches(/\b(rerun|plot|render|rendering|visualizer)/i)) labelsToAdd.add('visualization');
+            if (matches(/\b(cameras?|opencv|realsense|lidars?|sensors?|imus?|microphones?|rgbd|encoders?)\b/i)) labelsToAdd.add('sensors');
+            if (matches(/\b(urdf|actuators?|calibration|end-effector|kinematics)\b/i)) labelsToAdd.add('robots');
            if (matches(/\b(teleop|teleoperator|controller|leader|follower|joystick|gamepad)\b/i)) labelsToAdd.add('teleoperators');
-            if (matches(/\b(policy|policies|p0licy)\b/i)) labelsToAdd.add('policies');
-            if (matches(/\b(processors?|pipeline)\b/i)) labelsToAdd.add('processor');
-            if (matches(/\b(eval|evaluate|evaluation|metrics?|score|benchmark)\b/i)) labelsToAdd.add('evaluation');
-
-            // Infrastructure & Code Quality
+            if (matches(/\b(policy|policies|model?)\b/i)) labelsToAdd.add('policies');
+            if (matches(/\b(processor|pipeline|preprocessor|postprocessor)s?\b/i)) labelsToAdd.add('processor');
+            if (matches(/\b(eval|evaluate|evaluation|metrics?|score|benchmarks?)\b/i)) labelsToAdd.add('evaluation');
            if (matches(/\b(tests?|pytest|unittest|failing test)\b/i)) labelsToAdd.add('tests');
-            if (matches(/\b(ci|github actions|workflow|gha|actions?|pipeline)\b/i)) {
-              labelsToAdd.add('CI');
-              labelsToAdd.add('github_actions');
-            }
-            if (matches(/\b(perf|latency|throughput|fps|speed|performance)\b/i)) labelsToAdd.add('performance');
-            if (matches(/\b(dependency|requirements|pip|conda|install error|importerror|package not found)\b/i)) labelsToAdd.add('dependencies');
-            if (matches(/\b(python|pyproject|requirements(\.txt)?|pip install|typing error)\b/i)) labelsToAdd.add('python');
-
-            // Documentation & Meta
-            if (matches(/\b(doc|documentation|docs|readme|typo|how to)\b/i)) labelsToAdd.add('documentation');
-            if (matches(/\b(refactor|cleanup|restructure|rename|modernize code)\b/i)) labelsToAdd.add('refactor');
-            if (matches(/\b(release|changelog|version bump|cut a release|tag v)\b/i)) labelsToAdd.add('release');
-            if (matches(/\b(breaking change|major change)\b/i)) labelsToAdd.add('breaking change');
+            if (matches(/\b(ci|github actions?|github workflows?|gha|docker|pypi)\b/i)) labelsToAdd.add('CI');
+            if (matches(/\b(perf|latency|throughput|fps|speed|performance|slow|fast|slower|faster|memory usage)\b/i)) labelsToAdd.add('performance');
+            if (matches(/\b(dependency|dependencies|pip|install error|importerror|package not found|pyproject)\b/i)) labelsToAdd.add('dependencies');
+            if (matches(/\b(configuration|config|arguments?|input feature|dracuss)\b/i)) labelsToAdd.add('configuration');

            // Apply Labels
            const labels = Array.from(labelsToAdd).filter(Boolean);
@@ -28,7 +28,7 @@ on:
 # Sets up the environment variables
 env:
  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest

@@ -52,7 +52,7 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
@@ -87,7 +87,7 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
@@ -119,6 +119,7 @@ jobs:
      HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
      TORCH_HOME: /home/user_lerobot/.cache/torch
      TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    container:
      image: ${{ needs.build-docker-cpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
      options: --shm-size "16gb"
@@ -130,6 +131,11 @@ jobs:
        shell: bash
        working-directory: /lerobot
    steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
      - name: Run pytest on CPU
        run: pytest tests -vv --maxfail=10
      - name: Run end-to-end tests
@@ -146,6 +152,7 @@ jobs:
      HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
      TORCH_HOME: /home/user_lerobot/.cache/torch
      TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    container:
      image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
      options: --gpus all --shm-size "16gb"
@@ -157,6 +164,11 @@ jobs:
        shell: bash
        working-directory: /lerobot
    steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
      - name: Run pytest on GPU
        run: pytest tests -vv --maxfail=10
      - name: Run end-to-end tests
@@ -174,6 +186,7 @@ jobs:
      TORCH_HOME: /home/user_lerobot/.cache/torch
      TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
      CUDA_VISIBLE_DEVICES: "0,1,2,3"
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    container:
      image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
      options: --gpus all --shm-size "16gb"
@@ -185,12 +198,15 @@ jobs:
        shell: bash
        working-directory: /lerobot
    steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
      - name: Verify GPU availability
        run: |
          nvidia-smi
          python -c "import torch; print(f'PyTorch CUDA available: {torch.cuda.is_available()}'); print(f'Number of GPUs: {torch.cuda.device_count()}')"

      - name: Run multi-GPU training tests
-      # TODO(Steven): Investigate why motors tests are failing in multi-GPU setup
-        run: pytest tests -vv --maxfail=10 --ignore=tests/motors/
-        timeout-minutes: 10
+        run: pytest -vv tests/training/
@@ -43,14 +43,14 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          persist-credentials: false

      - name: Set up Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
        with:
-          python-version: '3.10'
+          python-version: '3.12'

      - name: Run pre-commit hooks
        uses: pre-commit/action@v3.0.1 # zizmor: ignore[unpinned-uses]
@@ -22,7 +22,7 @@ on:
 # Sets up the environment variables
 env:
  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"

 jobs:
  # This job builds the Python package and publishes it to PyPI
@@ -38,14 +38,14 @@ jobs:

    steps:
      - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          persist-credentials: false

      - name: Set up Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
        with:
-          python-version: '3.10'
+          python-version: '3.12'

      - name: Extract Version
        id: extract_info
@@ -83,14 +83,6 @@ jobs:
            exit 1
          fi

-      - name: Remove Tags with Git dependencies
-        # TODO(Steven): Temporary patch to remove pi from PyPi 0.4.0 release due to its reliance on git dependencies.
-        run: |
-          echo "::info:: Checking for Git dependencies to remove from pyproject.toml..."
-          grep -E '@ git\+https|lerobot\[pi\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
-          sed -E -i '/@ git\+https|lerobot\[pi\]/d' pyproject.toml
-          echo "::info:: Git dependencies removed. Proceeding with build."
-
      - name: Install build dependencies
        run: python -m pip install build

@@ -135,7 +127,7 @@ jobs:
    env:
      MUJOCO_GL: egl
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
@@ -177,4 +169,3 @@ jobs:

 # TODO(Steven): Publish draft/pre-release and to test pypi weekly
 # TODO(Steven): Separate build and publish job
-# TODO(Steven): Tag documentation with the same version as the package
@@ -43,7 +43,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout code
-        uses: actions/checkout@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/checkout@v6 # zizmor: ignore[unpinned-uses]
        with:
          fetch-depth: 0
          persist-credentials: false
@@ -20,8 +20,8 @@ on:
  workflow_dispatch:

  # Run on the 1st and 15th of every month at 09:00 UTC
-  schedule:
-    - cron: '0 2 1,15 * *'
+  # schedule:
+  #  - cron: '0 2 1,15 * *'

 permissions:
  contents: read
@@ -29,7 +29,7 @@ permissions:
 # Sets up the environment variables
 env:
  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound

 # Ensures that only the latest action is built, canceling older runs.
@@ -48,8 +48,9 @@ jobs:
      MUJOCO_GL: egl
      HF_HOME: /mnt/cache/.cache/huggingface
      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
@@ -78,8 +79,12 @@ jobs:
          echo "Dependencies unbound:" && cat pyproject.toml

      - name: Install lerobot with all extras
-        run: uv sync --all-extras --no-extra groot --no-extra wallx # TODO(Steven): Make flash-attn optional
-
+        run: uv sync --extra all # TODO(Steven): Make flash-attn optional
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami
      - name: Run pytest (all extras)
        run: uv run pytest tests -vv

@@ -91,6 +96,7 @@ jobs:
    name: Build and Push Docker
    runs-on:
      group: aws-general-8-plus
+    if: github.repository == 'huggingface/lerobot'
    outputs:
      image_tag: ${{ env.DOCKER_IMAGE_NAME }}
    env:
@@ -101,7 +107,7 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
@@ -136,6 +142,7 @@ jobs:
      HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
      TORCH_HOME: /home/user_lerobot/.cache/torch
      TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    container:
      image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
      options: --gpus all --shm-size "16gb"
@@ -147,6 +154,11 @@ jobs:
        shell: bash
        working-directory: /lerobot
    steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
      - name: Run pytest on GPU
        run: pytest tests -vv
      - name: Run end-to-end tests
@@ -162,15 +174,19 @@ jobs:
    steps:
      - name: Get Docker Hub Token and Delete Image
        # zizmor: ignore[template-injection]
+        env:
+          DOCKERHUB_LEROBOT_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          DOCKERHUB_LEROBOT_PASSWORD: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+          IMAGE_FULL: ${{ needs.build-and-push-docker.outputs.image_tag }}
        run: |
-          IMAGE_NAME=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f1)
-          IMAGE_TAG=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f2)
+          IMAGE_NAME=$(echo "$IMAGE_FULL" | cut -d':' -f1)
+          IMAGE_TAG=$(echo "$IMAGE_FULL" | cut -d':' -f2)

          echo "Attempting to delete image: $IMAGE_NAME:$IMAGE_TAG"

          TOKEN=$(curl -s -H "Content-Type: application/json" \
                       -X POST \
-                       -d '{"username": "${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}", "password": "${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}"}' \
+                       -d "{\"username\": \"$DOCKERHUB_LEROBOT_USERNAME\", \"password\": \"$DOCKERHUB_LEROBOT_PASSWORD\"}" \
                       https://hub.docker.com/v2/users/login/ | jq -r .token)

          if [ "$TOKEN" == "null" ] || [ -z "$TOKEN" ]; then
@@ -181,7 +197,7 @@ jobs:
          HTTP_RESPONSE=$(curl -s -o /dev/null -w "%{http_code}" \
                               -H "Authorization: JWT ${TOKEN}" \
                               -X DELETE \
-                               https://hub.docker.com/v2/repositories/${IMAGE_NAME}/tags/${IMAGE_TAG}/)
+                               https://hub.docker.com/v2/repositories/${IMAGE_NAME}/tags/$IMAGE_TAG)

          if [ "$HTTP_RESPONSE" -eq 204 ]; then
            echo "Successfully deleted Docker image tag: $IMAGE_NAME:$IMAGE_TAG"
@@ -13,7 +13,7 @@
 # limitations under the License.

 default_language_version:
-    python: python3.10
+    python: python3.12

 exclude: "tests/artifacts/.*\\.safetensors$"

@@ -55,7 +55,7 @@ repos:
    rev: v3.21.0
    hooks:
    -   id: pyupgrade
-        args: [--py310-plus]
+        args: [--py312-plus]

  ##### Markdown Quality #####
  - repo: https://github.com/rbubley/mirrors-prettier
@@ -0,0 +1,25 @@
+# AI Usage Policy
+
+The LeRobot project welcomes contributions from everyone, and we have a few guidelines regarding AI usage to ensure high code quality, clear communication, and a healthy open-source ecosystem:
+
+- **Please disclose significant AI assistance.** If you used AI tools (e.g., Copilot, Claude, Cursor, ChatGPT) to generate a substantial portion of your code or text, let us know in your PR description. Transparency helps us review your changes more effectively.
+- **Own your code (The Human-in-the-Loop).** You must fully understand all the changes you are proposing. If you cannot explain what your AI-assisted code does or how it interacts with LeRobot's broader architecture, please take the time to learn and test it before submitting.
+- **Keep issues and discussions focused.** You are welcome to use AI to help draft issues or PR descriptions, but please review and edit them carefully before posting. AI can often be overly verbose; trimming the noise and getting straight to the point helps our maintainers address your needs faster.
+
+Our core maintainers also use AI tools to aid their workflows, but they do so while bringing deep contextual knowledge of the LeRobot codebase to validate the output. We ask all contributors to apply that same level of rigor.
+
+## Remember the Human Maintainers
+
+Please remember that LeRobot is maintained by a dedicated team of humans.
+
+Every discussion, issue, and pull request is read and reviewed by real people. While AI tools can generate thousands of lines of code in seconds, reviewing that code still takes human time and energy. Submitting unverified or low-effort AI output puts an unfair burden on our maintainers.
+
+Today, the quality of the AI output still heavily depends on the developer driving the tool. We ask that you respect our maintainers' time by thoroughly vetting, testing, and refining your submissions.
+
+## AI is Welcome Here
+
+LeRobot operates at the cutting edge of AI and robotics, and many of our maintainers actively embrace AI coding assistants as valuable productivity tools. We are a pro-AI project!
+
+Our reason for having an AI policy is not an anti-AI stance. Rather, it exists to ensure that AI is used to enhance human contributions, not replace them with unverified noise. It's about how the tools are used, not the tools themselves.
+
+We value the unique human insight you bring to the LeRobot community. Let AI empower your workflow, but always let your own judgment take the wheel.
@@ -1,323 +1,83 @@
-# How to contribute to 🤗 LeRobot?
+# How to contribute to 🤗 LeRobot

-Everyone is welcome to contribute, and we value everybody's contribution. Code
-is thus not the only way to help the community. Answering questions, helping
-others, reaching out and improving the documentations are immensely valuable to
-the community.
+Everyone is welcome to contribute, and we value everybody's contribution. Code is not the only way to help the community. Answering questions, helping others, reaching out, and improving the documentation are immensely valuable.

-It also helps us if you spread the word: reference the library from blog posts
-on the awesome projects it made possible, shout out on Twitter when it has
-helped you, or simply ⭐️ the repo to say "thank you".
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](https://github.com/huggingface/lerobot/blob/main/CODE_OF_CONDUCT.md) and our [AI policy](https://github.com/huggingface/lerobot/blob/main/AI_POLICY.md).

-Whichever way you choose to contribute, please be mindful to respect our
-[code of conduct](https://github.com/huggingface/lerobot/blob/main/CODE_OF_CONDUCT.md).
+## Ways to Contribute

-## You can contribute in so many ways!
+You can contribute in many ways:

-Some of the ways you can contribute to 🤗 LeRobot:
+- **Fixing issues:** Resolve bugs or improve existing code.
+- **New features:** Develop new features.
+- **Extend:** Implement new models/policies, robots, or simulation environments and upload datasets to the Hugging Face Hub.
+- **Documentation:** Improve examples, guides, and docstrings.
+- **Feedback:** Submit tickets related to bugs or desired new features.

- Fixing outstanding issues with the existing code.
- Implementing new models, datasets or simulation environments.
- Contributing to the examples or to the documentation.
- Submitting issues related to bugs or desired new features.
+If you are unsure where to start, join our [Discord Channel](https://discord.gg/q8Dzzpym3f).

-Following the guides below, feel free to open issues and PRs and to coordinate your efforts with the community on our [Discord Channel](https://discord.gg/VjFz58wn3R). For specific inquiries, reach out to [Remi Cadene](mailto:remi.cadene@huggingface.co).
+## Development Setup

-If you are not sure how to contribute or want to know the next features we working on, look on this project page: [LeRobot TODO](https://github.com/orgs/huggingface/projects/46)
+To contribute code, you need to set up a development environment.

-## Submitting a new issue or feature request
+### 1. Fork and Clone

-Do your best to follow these guidelines when submitting an issue or a feature
-request. It will make it easier for us to come back to you quickly and with good
-feedback.
-
-### Did you find a bug?
-
-The 🤗 LeRobot library is robust and reliable thanks to the users who notify us of
-the problems they encounter. So thank you for reporting an issue.
-
-First, we would really appreciate it if you could **make sure the bug was not
-already reported** (use the search bar on Github under Issues).
-
-Did not find it? :( So we can act quickly on it, please follow these steps:
-
- Include your **OS type and version**, the versions of **Python** and **PyTorch**.
- A short, self-contained, code snippet that allows us to reproduce the bug in
-  less than 30s.
- The full traceback if an exception is raised.
- Attach any other additional information, like screenshots, you think may help.
-
-### Do you want a new feature?
-
-A good feature request addresses the following points:
-
-1. Motivation first:
-
- Is it related to a problem/frustration with the library? If so, please explain
-  why. Providing a code snippet that demonstrates the problem is best.
- Is it related to something you would need for a project? We'd love to hear
-  about it!
- Is it something you worked on and think could benefit the community?
-  Awesome! Tell us what problem it solved for you.
-
-2. Write a _paragraph_ describing the feature.
-3. Provide a **code snippet** that demonstrates its future use.
-4. In case this is related to a paper, please attach a link.
-5. Attach any additional information (drawings, screenshots, etc.) you think may help.
-
-If your issue is well written we're already 80% of the way there by the time you
-post it.
-
-## Adding new policies, datasets or environments
-
-Look at our implementations for [datasets](./src/lerobot/datasets/), [policies](./src/lerobot/policies/),
-environments ([aloha](https://github.com/huggingface/gym-aloha),
-[pusht](https://github.com/huggingface/gym-pusht))
-and follow the same api design.
-
-When implementing a new dataset loadable with LeRobotDataset follow these steps:
-
- Update `available_datasets_per_env` in `lerobot/__init__.py`
-
-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
-
-## Submitting a pull request (PR)
-
-Before writing code, we strongly advise you to search through the existing PRs or
-issues to make sure that nobody is already working on the same thing. If you are
-unsure, it is always a good idea to open an issue to get some feedback.
-
-You will need basic `git` proficiency to be able to contribute to
-🤗 LeRobot. `git` is not the easiest tool to use but it has the greatest
-manual. Type `git --help` in a shell and enjoy. If you prefer books, [Pro
-Git](https://git-scm.com/book/en/v2) is a very good reference.
-
-Follow these steps to start contributing:
-
-1. Fork the [repository](https://github.com/huggingface/lerobot) by
-   clicking on the 'Fork' button on the repository's page. This creates a copy of the code
-   under your GitHub user account.
-
-2. Clone your fork to your local disk, and add the base repository as a remote. The following command
-   assumes you have your public SSH key uploaded to GitHub. See the following guide for more
-   [information](https://docs.github.com/en/repositories/creating-and-managing-repositories/cloning-a-repository).
-
-   ```bash
-   git clone git@github.com:<your Github handle>/lerobot.git
-   cd lerobot
-   git remote add upstream https://github.com/huggingface/lerobot.git
-   ```
-
-3. Create a new branch to hold your development changes, and do this for every new PR you work on.
-
-   Start by synchronizing your `main` branch with the `upstream/main` branch (more details in the [GitHub Docs](https://docs.github.com/en/github/collaborating-with-issues-and-pull-requests/syncing-a-fork)):
-
-   ```bash
-   git checkout main
-   git fetch upstream
-   git rebase upstream/main
-   ```
-
-   Once your `main` branch is synchronized, create a new branch from it:
-
-   ```bash
-   git checkout -b a-descriptive-name-for-my-changes
-   ```
-
-   🚨 **Do not** work on the `main` branch.
-
-4. for development, we advise to use a tool like `poetry` or `uv` instead of just `pip` to easily track our dependencies.
-   Follow the instructions to [install poetry](https://python-poetry.org/docs/#installation) (use a version >=2.1.0) or to [install uv](https://docs.astral.sh/uv/getting-started/installation/#installation-methods) if you don't have one of them already.
-
-   Set up a development environment with conda:
-
-   ```bash
-   conda create -y -n lerobot-dev python=3.10 && conda activate lerobot-dev
-   ```
-
-   If you're using `uv`, it can manage python versions so you can instead do:
-
-   ```bash
-   uv venv --python 3.10 && source .venv/bin/activate
-   ```
-
-   To develop on 🤗 LeRobot, you will at least need to install the `dev` and `test` extras dependencies along with the core library:
-
-   using `poetry`
-
-   ```bash
-   poetry sync --extras "dev test"
-   ```
-
-   using `uv`
-
-   ```bash
-   uv sync --extra dev --extra test
-   ```
-
-   You can also install the project with all its dependencies (including environments):
-
-   using `poetry`
-
-   ```bash
-   poetry sync --all-extras
-   ```
-
-   using `uv`
-
-   ```bash
-   uv sync --all-extras
-   ```
-
-   > **Note:** If you don't install simulation environments with `--all-extras`, the tests that require them will be skipped when running the pytest suite locally. However, they _will_ be tested in the CI. In general, we advise you to install everything and test locally before pushing.
-
-   Whichever command you chose to install the project (e.g. `poetry sync --all-extras`), you should run it again when pulling code with an updated version of `pyproject.toml` and `poetry.lock` in order to synchronize your virtual environment with the new dependencies.
-
-   The equivalent of `pip install some-package`, would just be:
-
-   using `poetry`
-
-   ```bash
-   poetry add some-package
-   ```
-
-   using `uv`
-
-   ```bash
-   uv add some-package
-   ```
-
-   When making changes to the poetry sections of the `pyproject.toml`, you should run the following command to lock dependencies.
-   using `poetry`
-
-   ```bash
-   poetry lock
-   ```
-
-   using `uv`
-
-   ```bash
-   uv lock
-   ```
-
-5. Develop the features on your branch.
-
-   As you work on the features, you should make sure that the test suite
-   passes. You should run the tests impacted by your changes like this (see
-   below an explanation regarding the environment variable):
-
-   ```bash
-   pytest tests/<TEST_TO_RUN>.py
-   ```
-
-6. Follow our style.
-
-   `lerobot` relies on `ruff` to format its source code
-   consistently. Set up [`pre-commit`](https://pre-commit.com/) to run these checks
-   automatically as Git commit hooks.
-
-   Install `pre-commit` hooks:
-
-   ```bash
-   pre-commit install
-   ```
-
-   You can run these hooks whenever you need on staged files with:
-
-   ```bash
-   pre-commit
-   ```
-
-   Once you're happy with your changes, add changed files using `git add` and
-   make a commit with `git commit` to record your changes locally:
-
-   ```bash
-   git add modified_file.py
-   git commit
-   ```
-
-   Note, if you already committed some changes that have a wrong formatting, you can use:
-
-   ```bash
-   pre-commit run --all-files
-   ```
-
-   Please write [good commit messages](https://chris.beams.io/posts/git-commit/).
-
-   It is a good idea to sync your copy of the code with the original
-   repository regularly. This way you can quickly account for changes:
-
-   ```bash
-   git fetch upstream
-   git rebase upstream/main
-   ```
-
-   Push the changes to your account using:
-
-   ```bash
-   git push -u origin a-descriptive-name-for-my-changes
-   ```
-
-7. Once you are satisfied (**and the checklist below is happy too**), go to the
-   webpage of your fork on GitHub. Click on 'Pull request' to send your changes
-   to the project maintainers for review.
-
-8. It's ok if maintainers ask you for changes. It happens to core contributors
-   too! So everyone can see the changes in the Pull request, work in your local
-   branch and push the changes to your fork. They will automatically appear in
-   the pull request.
-
-### Checklist
-
-1. The title of your pull request should be a summary of its contribution;
-2. If your pull request addresses an issue, please mention the issue number in
-   the pull request description to make sure they are linked (and people
-   consulting the issue know you are working on it);
-3. To indicate a work in progress please prefix the title with `[WIP]`, or preferably mark
-   the PR as a draft PR. These are useful to avoid duplicated work, and to differentiate
-   it from PRs ready to be merged;
-4. Make sure existing tests pass;
-
-### Tests
-
-An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the [tests folder](https://github.com/huggingface/lerobot/tree/main/tests).
-
-Install [git lfs](https://git-lfs.com/) to retrieve test artifacts (if you don't have it already).
-
-On Mac:
+Fork the repository on GitHub, then clone your fork:

 ```bash
-brew install git-lfs
-git lfs install
+git clone https://github.com/<your-handle>/lerobot.git
+cd lerobot
+git remote add upstream https://github.com/huggingface/lerobot.git
 ```

-On Ubuntu:
+### 2. Environment Installation
+
+Please follow our [Installation Guide](https://huggingface.co/docs/lerobot/installation) for the environment setup & installation from source.
+
+## Running Tests & Quality Checks
+
+### Code Style (Pre-commit)
+
+Install `pre-commit` hooks to run checks automatically before you commit:

 ```bash
-sudo apt-get install git-lfs
-git lfs install
+pre-commit install
 ```

-Pull artifacts if they're not in [tests/artifacts](tests/artifacts)
+To run checks manually on all files:

 ```bash
+pre-commit run --all-files
+```
+
+### Running Tests
+
+We use `pytest`. First, ensure you have test artifacts by installing **git-lfs**:
+
+```bash
+git lfs install
 git lfs pull
 ```

-We use `pytest` in order to run the tests. From the root of the
-repository, here's how to run tests with `pytest` for the library:
+Run the full suite (this may require extras installed):

 ```bash
-python -m pytest -sv ./tests
+pytest -sv ./tests
 ```

-You can specify a smaller set of tests in order to test only the feature
-you're working on.
+Or run a specific test file during development:
+
+```bash
+pytest -sv tests/test_specific_feature.py
+```
+
+## Submitting Issues & Pull Requests
+
+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.
+
+Thank you for contributing to LeRobot!
@@ -1,2 +1,3 @@
 include src/lerobot/templates/lerobot_modelcard_template.md
 include src/lerobot/datasets/card_template.md
+include src/lerobot/envs/metaworld_config.json
@@ -10,6 +10,7 @@
 [![Status](https://img.shields.io/pypi/status/lerobot)](https://pypi.org/project/lerobot/)
 [![Version](https://img.shields.io/pypi/v/lerobot)](https://pypi.org/project/lerobot/)
 [![Contributor Covenant](https://img.shields.io/badge/Contributor%20Covenant-v2.1-ff69b4.svg)](https://github.com/huggingface/lerobot/blob/main/CODE_OF_CONDUCT.md)
+[![Discord](https://img.shields.io/badge/Discord-Join_Us-5865F2?style=flat&logo=discord&logoColor=white)](https://discord.gg/q8Dzzpym3f)

 </div>

@@ -99,11 +100,11 @@ lerobot-train \
  --dataset.repo_id=lerobot/aloha_mobile_cabinet
 ```

-| Category                   | Models                                                                                                                                                                 |
-| -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md)                       |
-| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                            |
-| **VLAs Models**            | [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx) |
+| Category                   | Models                                                                                                                                                                                                       |
+| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md)                                                             |
+| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                  |
+| **VLAs Models**            | [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx) |

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

@@ -127,13 +128,14 @@ Learn how to implement your own simulation environment or benchmark and distribu
 ## Resources

 - **[Documentation](https://huggingface.co/docs/lerobot/index):** The complete guide to tutorials & API.
- **[Discord](https://discord.gg/3gxM6Avj):** Join the `LeRobot` server to discuss with the community.
+- **[Chinese Tutorials: LeRobot+SO-ARM101中文教程-同济子豪兄](https://zihao-ai.feishu.cn/wiki/space/7589642043471924447)** Detailed doc for assembling, teleoperate, dataset, train, deploy. Verified by Seed Studio and 5 global hackathon players.
+- **[Discord](https://discord.gg/q8Dzzpym3f):** Join the `LeRobot` server to discuss with the community.
 - **[X](https://x.com/LeRobotHF):** Follow us on X to stay up-to-date with the latest developments.
 - **[Robot Learning Tutorial](https://huggingface.co/spaces/lerobot/robot-learning-tutorial):** A free, hands-on course to learn robot learning using LeRobot.

 ## Citation

-If you use LeRobot in your research, please cite:
+If you use LeRobot in your project, please cite the GitHub repository to acknowledge the ongoing development and contributors:

 ```bibtex
@misc{cadene2024lerobot,
@@ -144,9 +146,26 @@ If you use LeRobot in your research, please cite:
 }
 ```

+If you are referencing our research or the academic paper, please also cite our ICLR publication:
+
+<details>
+<summary><b>ICLR 2026 Paper</b></summary>
+
+```bibtex
+@inproceedings{cadenelerobot,
+  title={LeRobot: An Open-Source Library for End-to-End Robot Learning},
+  author={Cadene, Remi and Alibert, Simon and Capuano, Francesco and Aractingi, Michel and Zouitine, Adil and Kooijmans, Pepijn and Choghari, Jade and Russi, Martino and Pascal, Caroline and Palma, Steven and Shukor, Mustafa and Moss, Jess and Soare, Alexander and Aubakirova, Dana and Lhoest, Quentin and Gallou\'edec, Quentin and Wolf, Thomas},
+  booktitle={The Fourteenth International Conference on Learning Representations},
+  year={2026},
+  url={https://arxiv.org/abs/2602.22818}
+}
+```
+
+</details>
+
 ## Contribute

-We welcome contributions from everyone in the community! To get started, please read our [CONTRIBUTING.md](./CONTRIBUTING.md) guide. Whether you're adding a new feature, improving documentation, or fixing a bug, your help and feedback are invaluable. We're incredibly excited about the future of open-source robotics and can't wait to work with you on what's next—thank you for your support!
+We welcome contributions from everyone in the community! To get started, please read our [CONTRIBUTING.md](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md) guide. Whether you're adding a new feature, improving documentation, or fixing a bug, your help and feedback are invaluable. We're incredibly excited about the future of open-source robotics and can't wait to work with you on what's next—thank you for your support!

 <p align="center">
  <img alt="SO101 Video" src="./media/readme/so100_video.webp" width="640px">
@@ -0,0 +1,48 @@
+# Security Policy
+
+## Project Status & Philosophy
+
+`lerobot` has so far been primarily a research and prototyping tool, which is why deployment security hasn’t been a strong focus until now. As `lerobot` continues to be adopted and deployed in production, we are paying much closer attention to these kinds of issues.
+
+Fortunately, being an open-source project, the community can also help by reporting and fixing vulnerabilities. We appreciate your efforts to responsibly disclose your findings and will make every effort to acknowledge your contributions.
+
+## Reporting a Vulnerability
+
+To report a security issue, please use the GitHub Security Advisory ["Report a Vulnerability"](https://github.com/huggingface/lerobot/security/advisories/new) tab.
+
+The `lerobot` team will send a response indicating the next steps in handling your report. After the initial reply to your report, the security team will keep you informed of the progress towards a fix and full announcement, and may ask for additional information or guidance.
+
+#### Hugging Face Security Team
+
+Since this project is part of the Hugging Face ecosystem, feel free to submit vulnerability reports directly to: **[security@huggingface.co](mailto:security@huggingface.co)**. Someone from the HF security team will review the report and recommend next steps.
+
+#### Open Source Disclosures
+
+If reporting a vulnerability specific to the open-source codebase (and not the underlying Hub infrastructure), you may also use [Huntr](https://huntr.com), a vulnerability disclosure program for open source software.
+
+## Supported Versions
+
+Currently, we treat `lerobot` as a rolling release. We prioritize security updates for the latest available version (`main` branch).
+
+| Version  | Supported |
+| -------- | --------- |
+| Latest   | ✅        |
+| < Latest | ❌        |
+
+## Secure Usage Guidelines
+
+`lerobot` is tightly coupled to the Hugging Face Hub for sharing data and pretrained policies. When downloading artifacts uploaded by others, you expose yourself to risks. Please read below for recommendations to keep your runtime and robot environment safe.
+
+### Remote Artefacts (Weights & Policies)
+
+Models and policies uploaded to the Hugging Face Hub come in different formats. We heavily recommend uploading and downloading models in the [`safetensors`](https://github.com/huggingface/safetensors) format.
+
+`safetensors` was developed specifically to prevent arbitrary code execution on your system, which is critical when running software on physical hardware/robots.
+
+To avoid loading models from unsafe formats (e.g., `pickle`), you should ensure you are prioritizing `safetensors` files.
+
+### Remote Code
+
+Some models or environments on the Hub may require `trust_remote_code=True` to run custom architecture code.
+
+Please **always** verify the content of the modeling files when using this argument. We recommend setting a specific `revision` (commit hash) when loading remote code to ensure you protect yourself from unverified updates to the repository.
@@ -28,9 +28,9 @@ We don't expect the same optimal settings for a dataset of images from a simulat
 For these reasons, we run this benchmark on four representative datasets:

 - `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
- `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
- `aliberts/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
- `aliberts/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+- `lerobot/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
+- `lerobot/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
+- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.

 Note: The datasets used for this benchmark need to be image datasets, not video datasets.

@@ -179,7 +179,7 @@ python benchmark/video/run_video_benchmark.py \
    --output-dir outputs/video_benchmark \
    --repo-ids \
        lerobot/pusht_image \
-        aliberts/aloha_mobile_shrimp_image \
+        lerobot/aloha_mobile_shrimp_image \
    --vcodec libx264 libx265 \
    --pix-fmt yuv444p yuv420p \
    --g 2 20 None \
@@ -203,9 +203,9 @@ python benchmark/video/run_video_benchmark.py \
    --output-dir outputs/video_benchmark \
    --repo-ids \
        lerobot/pusht_image \
-        aliberts/aloha_mobile_shrimp_image \
-        aliberts/paris_street \
-        aliberts/kitchen \
+        lerobot/aloha_mobile_shrimp_image \
+        lerobot/paris_street \
+        lerobot/kitchen \
    --vcodec libx264 libx265 \
    --pix-fmt yuv444p yuv420p \
    --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -221,9 +221,9 @@ python benchmark/video/run_video_benchmark.py \
    --output-dir outputs/video_benchmark \
    --repo-ids \
        lerobot/pusht_image \
-        aliberts/aloha_mobile_shrimp_image \
-        aliberts/paris_street \
-        aliberts/kitchen \
+        lerobot/aloha_mobile_shrimp_image \
+        lerobot/paris_street \
+        lerobot/kitchen \
    --vcodec libsvtav1 \
    --pix-fmt yuv420p \
    --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -252,37 +252,37 @@ Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_read

 These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`

-| video_images_size_ratio            | vcodec     | pix_fmt |           |           |           |
-| ---------------------------------- | ---------- | ------- | --------- | --------- | --------- |
-|                                    | libx264    |         | libx265   |           | libsvtav1 |
-| repo_id                            | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
-| lerobot/pusht_image                | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
-| aliberts/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
-| aliberts/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
-| aliberts/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+| video_images_size_ratio           | vcodec     | pix_fmt |           |           |           |
+| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
+|                                   | libx264    |         | libx265   |           | libsvtav1 |
+| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| lerobot/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
+| lerobot/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
+| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |

-| video_images_load_time_ratio       | vcodec  | pix_fmt |          |         |           |
-| ---------------------------------- | ------- | ------- | -------- | ------- | --------- |
-|                                    | libx264 |         | libx265  |         | libsvtav1 |
-| repo_id                            | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
-| lerobot/pusht_image                | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
-| aliberts/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
-| aliberts/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
-| aliberts/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
+| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
+|                                   | libx264 |         | libx265  |         | libsvtav1 |
+| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| lerobot/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
+| lerobot/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
+| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |

-|                                    |          | vcodec   | pix_fmt      |          |           |              |
-| ---------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
-|                                    |          | libx264  |              | libx265  |           | libsvtav1    |
-| repo_id                            | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
-| lerobot/pusht_image                | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
-|                                    | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
-|                                    | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
-| aliberts/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
-|                                    | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
-|                                    | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
-| aliberts/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
-|                                    | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
-|                                    | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
-| aliberts/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
-|                                    | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
-|                                    | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
+|                                   |          | vcodec   | pix_fmt      |          |           |              |
+| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
+|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
+| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| lerobot/pusht_image               | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
+|                                   | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
+|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+| lerobot/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
+|                                   | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
+|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+| lerobot/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
+|                                   | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
+|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+| lerobot/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
+|                                   | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
+|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
@@ -24,7 +24,7 @@ ARG OS_VERSION=22.04
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}

 # Define Python version argument
-ARG PYTHON_VERSION=3.10
+ARG PYTHON_VERSION=3.12

 # Configure environment variables
 ENV DEBIAN_FRONTEND=noninteractive \
@@ -73,7 +73,7 @@ ENV HOME=/home/user_lerobot \
 RUN uv venv --python python${PYTHON_VERSION}

 # Install Python dependencies for caching
-COPY --chown=user_lerobot:user_lerobot pyproject.toml README.md MANIFEST.in ./
+COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml README.md MANIFEST.in ./
 COPY --chown=user_lerobot:user_lerobot src/ src/

 ARG UNBOUND_DEPS=false
@@ -85,6 +85,8 @@ RUN if [ "$UNBOUND_DEPS" = "true" ]; then \

 RUN uv pip install --no-cache ".[all]"

+RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
+
 # Copy the rest of the application source code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .
@@ -18,8 +18,10 @@
 # docker build -f docker/Dockerfile.user -t lerobot-user .
 # docker run -it --rm lerobot-user

+# With USB physical access : docker run -it --device=/dev/ -v /dev/:/dev/ --rm lerobot-user
+
 # Configure the base image
-ARG PYTHON_VERSION=3.10
+ARG PYTHON_VERSION=3.12
 FROM python:${PYTHON_VERSION}-slim

 # Configure environment variables
@@ -59,7 +61,7 @@ ENV HOME=/home/user_lerobot \
 RUN uv venv

 # Install Python dependencies for caching
-COPY --chown=user_lerobot:user_lerobot pyproject.toml README.md MANIFEST.in ./
+COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml README.md MANIFEST.in ./
 COPY --chown=user_lerobot:user_lerobot src/ src/

 ARG UNBOUND_DEPS=false
@@ -7,8 +7,6 @@
 - sections:
  - local: il_robots
    title: Imitation Learning for Robots
-  - local: cameras
-    title: Cameras
  - local: bring_your_own_policies
    title: Bring Your Own Policies
  - local: integrate_hardware
@@ -29,6 +27,10 @@
    title: Porting Large Datasets
  - local: using_dataset_tools
    title: Using the Dataset Tools
+  - local: dataset_subtask
+    title: Using Subtasks in the Dataset
+  - local: streaming_video_encoding
+    title: Streaming Video Encoding
  title: "Datasets"
 - sections:
  - local: act
@@ -37,12 +39,16 @@
    title: SmolVLA
  - local: pi0
    title: π₀ (Pi0)
+  - local: pi0fast
+    title: π₀-FAST (Pi0Fast)
  - local: pi05
    title: π₀.₅ (Pi05)
  - local: groot
    title: NVIDIA GR00T N1.5
  - local: xvla
    title: X-VLA
+  - local: walloss
+    title: WALL-OSS
  title: "Policies"
 - sections:
  - local: sarm
@@ -59,6 +65,8 @@
    title: Environments from the Hub
  - local: envhub_leisaac
    title: Control & Train Robots in Sim (LeIsaac)
+  - local: envhub_isaaclab_arena
+    title: NVIDIA IsaacLab Arena Environments
  - local: libero
    title: Using Libero
  - local: metaworld
@@ -93,11 +101,19 @@
    title: Unitree G1
  - local: earthrover_mini_plus
    title: Earth Rover Mini
+  - local: omx
+    title: OMX
+  - local: openarm
+    title: OpenArm
  title: "Robots"
 - sections:
  - local: phone_teleop
    title: Phone
  title: "Teleoperators"
+- sections:
+  - local: cameras
+    title: Cameras
+  title: "Sensors"
 - sections:
  - local: torch_accelerators
    title: PyTorch accelerators
@@ -107,6 +123,8 @@
    title: Notebooks
  - local: feetech
    title: Updating Feetech Firmware
+  - local: damiao
+    title: Damiao Motors and CAN Bus
  title: "Resources"
 - sections:
  - local: contributing
@@ -88,5 +88,8 @@ lerobot-record \
  --dataset.repo_id=${HF_USER}/eval_act_your_dataset \
  --dataset.num_episodes=10 \
  --dataset.single_task="Your task description" \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
  --policy.path=${HF_USER}/act_policy
 ```
@@ -48,7 +48,7 @@ python -m lerobot.async_inference.robot_client \
    --task="dummy" \ # POLICY: The task to run the policy on (`Fold my t-shirt`). Not necessarily defined for all policies, such as `act`
    --policy_type=your_policy_type \ # POLICY: the type of policy to run (smolvla, act, etc)
    --pretrained_name_or_path=user/model \ # POLICY: the model name/path on server to the checkpoint to run (e.g., lerobot/smolvla_base)
-    --policy_device=mps \ # POLICY: the device to run the policy on, on the server
+    --policy_device=mps \ # POLICY: the device to run the policy on, on the server (cuda, mps, xpu, cpu)
    --actions_per_chunk=50 \ # POLICY: the number of actions to output at once
    --chunk_size_threshold=0.5 \ # CLIENT: the threshold for the chunk size before sending a new observation to the server
    --aggregate_fn_name=weighted_average \ # CLIENT: the function to aggregate actions on overlapping portions
@@ -169,7 +169,7 @@ python -m lerobot.async_inference.robot_client \
 <!-- prettier-ignore-start -->
 ```python
 import threading
-from lerobot.robots.so100_follower import SO100FollowerConfig
+from lerobot.robots.so_follower import SO100FollowerConfig
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
 from lerobot.async_inference.configs import RobotClientConfig
 from lerobot.async_inference.robot_client import RobotClient
@@ -195,6 +195,7 @@ client_cfg = RobotClientConfig(
    robot=robot_cfg,
    server_address="localhost:8080",
    policy_device="mps",
+    client_device="cpu",
    policy_type="smolvla",
    pretrained_name_or_path="<user>/smolvla_async",
    chunk_size_threshold=0.5,
@@ -32,7 +32,7 @@ version = "0.1.0"
 dependencies = [
    # your policy-specific dependencies
 ]
-requires-python = ">= 3.11"
+requires-python = ">= 3.12"

 [build-system]
 build-backend = # your-build-backend
@@ -82,7 +82,7 @@ Create your policy implementation by inheriting from LeRobot's base `PreTrainedP
 # modeling_my_custom_policy.py
 import torch
 import torch.nn as nn
-from typing import Dict, Any
+from typing import Any

 from lerobot.policies.pretrained import PreTrainedPolicy
 from .configuration_my_custom_policy import MyCustomPolicyConfig
@@ -91,7 +91,7 @@ class MyCustomPolicy(PreTrainedPolicy):
    config_class = MyCustomPolicyConfig
    name = "my_custom_policy"

-    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: Dict[str, Any] = None):
+    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: dict[str, Any] = None):
        super().__init__(config, dataset_stats)
        ...
 ```
@@ -102,7 +102,7 @@ Create processor functions:

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


@@ -1,12 +1,22 @@
 # Cameras

-LeRobot offers multiple options for video capture, including phone cameras, built-in laptop cameras, external webcams, and Intel RealSense cameras. To efficiently record frames from most cameras, you can use either the `OpenCVCamera` or `RealSenseCamera` class. For additional compatibility details on the `OpenCVCamera` class, refer to the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
+LeRobot offers multiple options for video capture:

-### Finding your camera
+| Class             | Supported Cameras                   |
+| ----------------- | ----------------------------------- |
+| `OpenCVCamera`    | Phone, built-in laptop, USB webcams |
+| `ZMQCamera`       | Network-connected cameras           |
+| `RealSenseCamera` | Intel RealSense (with depth)        |
+| `Reachy2Camera`   | Reachy 2 robot cameras              |

-To instantiate a camera, you need a camera identifier. This identifier might change if you reboot your computer or re-plug your camera, a behavior mostly dependant on your operating system.
+> [!TIP]
+> For `OpenCVCamera` compatibility details, see the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).

-To find the camera indices of the cameras plugged into your system, run the following script:
+### Find your camera
+
+Every camera requires a unique identifier to be instantiated, allowing you to distinguish between multiple connected devices.
+
+`OpenCVCamera` and `RealSenseCamera` support auto-discovery. Run the command below to list available devices and their identifiers. Note that these identifiers may change after rebooting your computer or re-plugging the camera, depending on your operating system.

 ```bash
 lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
@@ -14,7 +24,7 @@ lerobot-find-cameras opencv # or realsense for Intel Realsense cameras

 The output will look something like this if you have two cameras connected:

-```
+```bash
 --- Detected Cameras ---
 Camera #0:
  Name: OpenCV Camera @ 0
@@ -33,13 +43,37 @@ Camera #0:
 > [!WARNING]
 > When using Intel RealSense cameras in `macOS`, you could get this [error](https://github.com/IntelRealSense/librealsense/issues/12307): `Error finding RealSense cameras: failed to set power state`, this can be solved by running the same command with `sudo` permissions. Note that using RealSense cameras in `macOS` is unstable.

-## Use Cameras
+`ZMQCamera` and `Reachy2Camera` do not support auto-discovery. They must be configured manually by providing their network address and port or robot SDK settings.

-Below are two examples, demonstrating how to work with the API.
+## Use cameras

- **Asynchronous frame capture** using an OpenCV-based camera
+### Frame access modes
+
+All camera classes implement three access modes for capturing frames:
+
+| Method                    | Behavior                                                                                                                                                   | Blocks?        | Best For                                 |
+| ------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------- | ---------------------------------------- |
+| `read()`                  | Waits for the camera hardware to return a frame. May block for a long time depending on the camera and SDK.                                                | Yes            | Simple scripts, sequential capture       |
+| `async_read(timeout_ms)`  | Returns the latest unconsumed frame from background thread. Blocks only if buffer is empty, up to `timeout_ms`. Raises `TimeoutError` if no frame arrives. | With a timeout | Control loops synchronized to camera FPS |
+| `read_latest(max_age_ms)` | Peeks at the most recent frame in buffer (may be stale). Raises `TimeoutError` if frame is older than `max_age_ms`.                                        | No             | UI visualization, logging, monitoring    |
+
+### Usage examples
+
+The following examples show how to use the camera API to configure and capture frames from different camera types.
+
+- **Blocking and non-blocking frame capture** using an OpenCV-based camera
 - **Color and depth capture** using an Intel RealSense camera

+> [!WARNING]
+> Failing to cleanly disconnect cameras can cause resource leaks. Use the context manager protocol to ensure automatic cleanup:
+>
+> ```python
+> with OpenCVCamera(config) as camera:
+>     ...
+> ```
+>
+> You can also call `connect()` and `disconnect()` manually, but always use a `finally` block for the latter.
+
 <hfoptions id="shell_restart">
 <hfoption id="Open CV Camera">

@@ -60,16 +94,30 @@ config = OpenCVCameraConfig(
 )

 # Instantiate and connect an `OpenCVCamera`, performing a warm-up read (default).
-camera = OpenCVCamera(config)
-camera.connect()
+with OpenCVCamera(config) as camera:
+
+    # Read a frame synchronously — blocks until hardware delivers a new frame
+    frame = camera.read()
+    print(f"read() call returned frame with shape:", frame.shape)
+
+    # Read a frame asynchronously with a timeout — returns the latest unconsumed frame or waits up to timeout_ms for a new one
+    try:
+        for i in range(10):
+            frame = camera.async_read(timeout_ms=200)
+            print(f"async_read call returned frame {i} with shape:", frame.shape)
+    except TimeoutError as e:
+        print(f"No frame received within timeout: {e}")
+
+    # Instantly return a frame - returns the most recent frame captured by the camera
+    try:
+        initial_frame = camera.read_latest(max_age_ms=1000)
+        for i in range(10):
+            frame = camera.read_latest(max_age_ms=1000)
+            print(f"read_latest call returned frame {i} with shape:", frame.shape)
+            print(f"Was a new frame received by the camera? {not (initial_frame == frame).any()}")
+    except TimeoutError as e:
+        print(f"Frame too old: {e}")

-# Read frames asynchronously in a loop via `async_read(timeout_ms)`
-try:
-    for i in range(10):
-        frame = camera.async_read(timeout_ms=200)
-        print(f"Async frame {i} shape:", frame.shape)
-finally:
-    camera.disconnect()
 ```
 <!-- prettier-ignore-end -->

@@ -111,10 +159,10 @@ finally:
 </hfoption>
 </hfoptions>

-## Use your phone
+## Use your phone's camera

 <hfoptions id="use phone">
-<hfoption id="Mac">
+<hfoption id="iPhone & macOS">

 To use your iPhone as a camera on macOS, enable the Continuity Camera feature:

@@ -124,83 +172,49 @@ To use your iPhone as a camera on macOS, enable the Continuity Camera feature:

 For more details, visit [Apple support](https://support.apple.com/en-gb/guide/mac-help/mchl77879b8a/mac).

-Your iPhone should be detected automatically when running the camera setup script in the next section.
-
 </hfoption>
-<hfoption id="Linux">
+<hfoption id="OBS virtual camera">

-If you want to use your phone as a camera on Linux, follow these steps to set up a virtual camera
+If you want to use your phone as a camera using OBS, follow these steps to set up a virtual camera.

-1. _Install `v4l2loopback-dkms` and `v4l-utils`_. Those packages are required to create virtual camera devices (`v4l2loopback`) and verify their settings with the `v4l2-ctl` utility from `v4l-utils`. Install them using:
+1. _(Linux only) Install `v4l2loopback-dkms` and `v4l-utils`_. These packages create virtual camera devices and verify their settings. Install with:

-<!-- prettier-ignore-start -->
-```python
+```bash
 sudo apt install v4l2loopback-dkms v4l-utils
 ```
-<!-- prettier-ignore-end -->

-2. _Install [DroidCam](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
-3. _Install [OBS Studio](https://obsproject.com)_. This software will help you manage the camera feed. Install it using [Flatpak](https://flatpak.org):
+2. _Install the [DroidCam app](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
+3. _Download and install [OBS Studio](https://obsproject.com)_.
+4. _Download and install the [DroidCam OBS plugin](https://droidcam.app/obs)_.
+5. _Start OBS Studio_.

-<!-- prettier-ignore-start -->
-```python
-flatpak install flathub com.obsproject.Studio
-```
-<!-- prettier-ignore-end -->
-
-4. _Install the DroidCam OBS plugin_. This plugin integrates DroidCam with OBS Studio. Install it with:
-
-<!-- prettier-ignore-start -->
-```python
-flatpak install flathub com.obsproject.Studio.Plugin.DroidCam
-```
-<!-- prettier-ignore-end -->
-
-5. _Start OBS Studio_. Launch with:
-
-<!-- prettier-ignore-start -->
-```python
-flatpak run com.obsproject.Studio
-```
-<!-- prettier-ignore-end -->
-
-6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480`.
-7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it in.
+6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480` to avoid the watermarks.
+7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video` or `OBS > Preferences... > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it.
 8. _Start virtual camera_. In OBS Studio, follow the instructions [here](https://obsproject.com/kb/virtual-camera-guide).
-9. _Verify the virtual camera setup_. Use `v4l2-ctl` to list the devices:
+9. _Verify the virtual camera setup and resolution_.
+   - **Linux**: Use `v4l2-ctl` to list devices and check resolution:
+     ```bash
+     v4l2-ctl --list-devices  # find VirtualCam and note its /dev/videoX path
+     v4l2-ctl -d /dev/videoX --get-fmt-video  # replace with your VirtualCam path
+     ```
+     You should see `VirtualCam` listed and resolution `640x480`.
+   - **macOS**: Open Photo Booth or FaceTime and select "OBS Virtual Camera" as the input.
+   - **Windows**: The native Camera app doesn't support virtual cameras. Use a video conferencing app (Zoom, Teams) or run `lerobot-find-cameras opencv` directly to verify.

-<!-- prettier-ignore-start -->
-```python
-v4l2-ctl --list-devices
-```
-<!-- prettier-ignore-end -->
+<details>
+<summary><strong>Troubleshooting</strong></summary>

-You should see an entry like:
+> The virtual camera resolution is incorrect.

-```
-VirtualCam (platform:v4l2loopback-000):
-/dev/video1
-```
+Delete the virtual camera source and recreate it. The resolution cannot be changed after creation.

-10. _Check the camera resolution_. Use `v4l2-ctl` to ensure that the virtual camera output resolution is `640x480`. Change `/dev/video1` to the port of your virtual camera from the output of `v4l2-ctl --list-devices`.
+> Error reading frame in background thread for OpenCVCamera(X): OpenCVCamera(X) frame width=640 or height=480 do not match configured width=1920 or height=1080.

-<!-- prettier-ignore-start -->
-```python
-v4l2-ctl -d /dev/video1 --get-fmt-video
-```
-<!-- prettier-ignore-end -->
+This error is caused by OBS Virtual Camera advertising a `1920x1080` resolution despite rescaling. The only fix for now is to comment out the width and height check in `_postprocess_image()`.

-You should see an entry like:
-
-```
->>> Format Video Capture:
->>>	Width/Height      : 640/480
->>>	Pixel Format      : 'YUYV' (YUYV 4:2:2)
-```
-
-Troubleshooting: If the resolution is not correct you will have to delete the Virtual Camera port and try again as it cannot be changed.
-
-If everything is set up correctly, you can proceed with the rest of the tutorial.
+</details>

 </hfoption>
 </hfoptions>
+
+If everything is set up correctly, your phone will appear as a standard OpenCV camera and can be used with `OpenCVCamera`.
@@ -0,0 +1,165 @@
+# Damiao Motors and CAN Bus
+
+This guide covers setup and usage of Damiao motors with LeRobot via CAN bus communication.
+
+Currently, only Linux is supported, as the OpenArms CAN adapter only has drivers for Linux.
+
+## Linux CAN Setup
+
+Before using Damiao motors, you need to set up the CAN interface on your Linux system.
+
+### Install CAN Utilities
+
+```bash
+sudo apt-get install can-utils
+```
+
+### Configure CAN Interface (Manual)
+
+For standard CAN FD (recommended for OpenArms):
+
+```bash
+sudo ip link set can0 down
+sudo ip link set can0 type can bitrate 1000000 dbitrate 5000000 fd on
+sudo ip link set can0 up
+```
+
+For standard CAN (without FD):
+
+```bash
+sudo ip link set can0 down
+sudo ip link set can0 type can bitrate 1000000
+sudo ip link set can0 up
+```
+
+### Configure CAN Interface (Using LeRobot)
+
+LeRobot provides a utility script to setup and test CAN interfaces:
+
+```bash
+# Setup multiple interfaces (e.g., OpenArms Followers with 2 CAN buses)
+lerobot-setup-can --mode=setup --interfaces=can0,can1
+```
+
+## Debugging CAN Communication
+
+Use the built-in debug tools to test motor communication:
+
+```bash
+# Test motors on all interfaces
+lerobot-setup-can --mode=test --interfaces=can0,can1
+
+# Run speed/latency test
+lerobot-setup-can --mode=speed --interfaces=can0
+```
+
+The test mode will scan for motors (IDs 0x01-0x08) and report which ones respond. Example output:
+
+```
+can0: UP (CAN FD)
+  Motor 0x01 (joint_1): ✓ FOUND
+    → Response 0x11 [FD]: 00112233...
+  Motor 0x02 (joint_2): ✓ FOUND
+  Motor 0x03 (joint_3): ✗ No response
+  ...
+  Summary: 2/8 motors found
+```
+
+## Usage
+
+### Basic Setup
+
+```python
+from lerobot.motors import Motor
+from lerobot.motors.damiao import DamiaoMotorsBus
+
+# Define your motors with send/receive CAN IDs
+motors = {
+    "joint_1": Motor(id=0x01, motor_type_str="dm8009", recv_id=0x11),
+    "joint_2": Motor(id=0x02, motor_type_str="dm4340", recv_id=0x12),
+    "joint_3": Motor(id=0x03, motor_type_str="dm4310", recv_id=0x13),
+}
+
+# Create the bus
+bus = DamiaoMotorsBus(
+    port="can0",  # Linux socketcan interface
+    motors=motors,
+)
+
+# Connect
+bus.connect()
+```
+
+### Reading Motor States
+
+```python
+# Read single motor position (degrees)
+position = bus.read("Present_Position", "joint_1")
+
+# Read from multiple motors
+positions = bus.sync_read("Present_Position")  # All motors
+positions = bus.sync_read("Present_Position", ["joint_1", "joint_2"])
+
+# Read all states at once (position, velocity, torque)
+states = bus.sync_read_all_states()
+# Returns: {'joint_1': {'position': 45.2, 'velocity': 1.3, 'torque': 0.5}, ...}
+```
+
+### Writing Motor Commands
+
+```python
+# Enable torque
+bus.enable_torque()
+
+# Set goal position (degrees)
+bus.write("Goal_Position", "joint_1", 45.0)
+
+# Set positions for multiple motors
+bus.sync_write("Goal_Position", {
+    "joint_1": 45.0,
+    "joint_2": -30.0,
+    "joint_3": 90.0,
+})
+
+# Disable torque
+bus.disable_torque()
+```
+
+## Configuration Options
+
+| Parameter      | Default   | Description                                                 |
+| -------------- | --------- | ----------------------------------------------------------- |
+| `port`         | -         | CAN interface (`can0`) or serial port (`/dev/cu.usbmodem*`) |
+| `use_can_fd`   | `True`    | Enable CAN FD for higher data rates                         |
+| `bitrate`      | `1000000` | Nominal bitrate (1 Mbps)                                    |
+| `data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps)                                |
+
+## Motor Configuration
+
+Each motor requires:
+
+- `id`: CAN ID for sending commands
+- `motor_type`: One of the supported motor types (e.g., `"dm8009"`, `"dm4340"`)
+- `recv_id`: CAN ID for receiving responses
+
+OpenArms default IDs follow the pattern: send ID `0x0N`, receive ID `0x1N` where N is the joint number.
+
+## Troubleshooting
+
+### No Response from Motors
+
+1. **Check power**
+2. **Verify CAN wiring**: Check CAN-H, CAN-L, and GND connections
+3. **Check motor IDs**: Use Damiao Debugging Tools to verify/configure IDs
+4. **Test CAN interface**: Run `candump can0` to see if messages are being received
+5. **Run diagnostics**: `lerobot-setup-can --mode=test --interfaces=can0`
+
+### Motor Timeout Parameter
+
+If motors were configured with timeout=0, they won't respond to commands. Use Damiao Debugging Tools to set a non-zero timeout value.
+
+### Verify CAN FD Status
+
+```bash
+ip -d link show can0 | grep fd
+```
@@ -0,0 +1,278 @@
+# 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
@@ -1,5 +1,11 @@
 # EarthRover Mini Plus

+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Earth_Rover_Mini_5_240c9adc-4f9e-44b7-982f-5d1dc24af1d8.png.webp"
+  alt="EarthRover Mini Plus"
+  width="70%"
+/>
+
 The EarthRover Mini Plus is a fully open source mobile robot that connects through the cloud using the Frodobots SDK. This lets you control the robot and record datasets for training AI models.

 ## What You Need
@@ -7,28 +13,47 @@ The EarthRover Mini Plus is a fully open source mobile robot that connects throu
 ### Hardware

 - EarthRover Mini robot
- Computer with Python 3.10 or newer
+- Computer with Python 3.12 or newer
 - Internet connection

 ### Setting Up the Frodobots SDK

-The robot needs the [Frodobots SDK](https://github.com/Frodobots/earth-rovers-sdk) running on your computer. Here's how:
+The robot needs the [Frodobots SDK](https://github.com/frodobots-org/earth-rovers-sdk) running on your computer. Here's how:

 1. Download and install the SDK:

 ```bash
-git clone https://github.com/Frodobots/earth-rovers-sdk.git
+git clone https://github.com/frodobots-org/earth-rovers-sdk.git
 cd earth-rovers-sdk
 pip install -r requirements.txt
 ```

-2. Start the SDK:
+2. Save Credentials:
+
+Write your .env variables with the SDK API key and bot name provided by the Frodobots team.
+
+```bash
+SDK_API_TOKEN=your_sdk_api_token_here
+BOT_SLUG=your_bot_slug_here
+CHROME_EXECUTABLE_PATH=/path/to/chrome_or_chromium
+# Default value is MAP_ZOOM_LEVEL=18 https://wiki.openstreetmap.org/wiki/Zoom_levels
+MAP_ZOOM_LEVEL=18
+MISSION_SLUG=your_mission_slug_here
+# Image quality between 0.1 and 1.0 (default: 0.8)
+# Recommended: 0.8 for better performance
+IMAGE_QUALITY=0.8
+# Image format: jpeg, png or webp (default: png)
+# Recommended: jpeg for better performance and lower bandwidth usage
+IMAGE_FORMAT=jpeg
+```
+
+3. Start the SDK:

 ```bash
 hypercorn main:app --reload
 ```

-3. Open your web browser and go to `http://localhost:8000`, then click "Join"
+4. Open your web browser and go to `http://localhost:8000`, then click "Join"

 The SDK gives you:

@@ -145,13 +170,13 @@ Once you can drive the robot well, you can start recording data to train AI mode
 We use Hugging Face to store your data online. First, log in with your token from [Hugging Face settings](https://huggingface.co/settings/tokens):

 ```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```

 Store your Hugging Face username:

 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```

@@ -160,13 +185,16 @@ echo $HF_USER
 Use the standard recording command:

 ```bash
-python src/lerobot/scripts/lerobot_record.py \
+lerobot-record \
    --robot.type=earthrover_mini_plus \
    --teleop.type=keyboard_rover \
    --dataset.repo_id=your_username/dataset_name \
    --dataset.num_episodes=2 \
    --dataset.fps=10 \
    --dataset.single_task="Navigate around obstacles" \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    # --dataset.vcodec=auto \
    --display_data=true
 ```

@@ -176,22 +204,26 @@ Replace `your_username/dataset_name` with your Hugging Face username and a name

 Your dataset includes:

-**Your Actions (2 things)**:
+**Your Actions (2 features)**:

- How much you moved forward/backward
- How much you turned left/right
+- `linear_velocity`: How much you moved forward/backward
+- `angular_velocity`: How much you turned left/right

-**Robot Observations (12 things)**:
+**Robot Observations (24 features)**:

 - Front camera video
 - Rear camera video
 - Current speed
 - Battery level
- Which way the robot is facing
- GPS location (latitude, longitude, signal strength)
+- Orientation
+- GPS (latitude, longitude, signal strength)
 - Network signal strength
 - Vibration level
- Lamp status (on/off)
+- Lamp state (on/off)
+- Accelerometer (x, y, z)
+- Gyroscope (x, y, z)
+- Magnetometer (x, y, z)
+- Wheel RPMs (4 wheels)

 ### Where Your Data Goes

@@ -2,14 +2,32 @@

 The **EnvHub** feature allows you to load simulation environments directly from the Hugging Face Hub with a single line of code. This unlocks a powerful new model for collaboration: instead of environments being locked away inside monolithic libraries, anyone can publish custom environments and share them with the community.

-## Overview
+## What is EnvHub?

-With EnvHub, you can:
+EnvHub lets you create custom robotics simulation environments with your own robot models and scenarios, and make them easily usable by anyone through the LeRobot framework.

- Load environments from the Hub instantly
- Share your custom simulation tasks with the community
- Version control your environments using Git
- Distribute complex physics simulations without packaging hassles
+EnvHub packages are stored on the Hugging Face Hub, and can be seamlessly pulled and used in your AI robotics projects through LeRobot with a single line of code.
+
+Thanks to EnvHub, you can:
+
+1. **Create and publish environments** to the Hugging Face Hub as Git repositories, and distribute complex physics simulations without packaging hassles
+2. **Load environments** dynamically, without installing them as packages
+3. **Version and track** environment changes using Git semantics
+4. **Discover** new simulation tasks shared by the community
+
+This design means you can go from discovering an interesting environment on the Hub to running experiments in seconds, or create your own custom robot and environment without worrying about dependency conflicts or complex installation procedures.
+
+When you create an EnvHub package, you can build anything you want inside it and use any simulation tool you like: this is your own space to play with. The only requirement is that the package contains an `env.py` file that defines the environment and allows LeRobot to load and use your EnvHub package.
+
+This `env.py` file needs to expose a small API so LeRobot can load and run it. In particular, you must provide a `make_env(n_envs: int = 1, use_async_envs: bool = False)` or `make_env(n_envs: int = 1, use_async_envs: bool = False, cfg: EnvConfig)` function, which is the main entry point for LeRobot. It should return one of:
+
+- A `gym.vector.VectorEnv` (most common)
+- A single `gym.Env` (will be automatically wrapped)
+- A dict mapping `{suite_name: {task_id: VectorEnv}}` (for multi-task benchmarks)
+
+You can also pass an `EnvConfig` object to `make_env` to configure the environment (e.g. the number of environments, task, camera name, initial states, control mode, episode length, etc.).
+
+Finally, your environment must implement the standard `gym.vector.VectorEnv` interface so it works with LeRobot, including methods like `reset` and `step`.

 ## Quick Start

@@ -29,17 +47,6 @@ env = make_env("lerobot/cartpole-env", trust_remote_code=True)
  hash for reproducibility and security.
 </Tip>

-## What is EnvHub?
-
-EnvHub is a framework that allows researchers and developers to:
-
-1. **Publish environments** to the Hugging Face Hub as Git repositories
-2. **Load environments** dynamically without installing them as packages
-3. **Version and track** environment changes using Git semantics
-4. **Discover** new simulation tasks shared by the community
-
-This design means you can go from discovering an interesting environment on the Hub to running experiments in seconds, without worrying about dependency conflicts or complex installation procedures.
-
 ## Repository Structure

 To make your environment loadable from the Hub, your repository must contain at minimum:
@@ -148,10 +155,10 @@ Upload your repository to Hugging Face:
 pip install huggingface_hub

 # Login to Hugging Face
-huggingface-cli login
+hf auth login

 # Create a new repository
-huggingface-cli repo create my-custom-env --type space --org my-org
+hf repo create my-org/my-custom-env

 # Initialize git and push
 git init
@@ -0,0 +1,510 @@
+# NVIDIA IsaacLab Arena & LeRobot
+
+LeRobot EnvHub now supports **GPU-accelerated simulation** with IsaacLab Arena for policy evaluation at scale.
+Train and evaluate imitation learning policies with high-fidelity simulation — all integrated into the LeRobot ecosystem.
+
+<img
+  src="https://huggingface.co/nvidia/isaaclab-arena-envs/resolve/main/assets/Gr1OpenMicrowaveEnvironment.png"
+  alt="IsaacLab Arena - GR1 Microwave Environment"
+  style={{ maxWidth: "100%", borderRadius: "8px", marginBottom: "1rem" }}
+/>
+
+[IsaacLab Arena](https://github.com/isaac-sim/IsaacLab-Arena) integrates with NVIDIA IsaacLab to provide:
+
+- 🤖 **Humanoid embodiments**: GR1, G1, Galileo with various configurations
+- 🎯 **Manipulation & loco-manipulation tasks**: Door opening, pick-and-place, button pressing, and more
+- ⚡ **GPU-accelerated rollouts**: Parallel environment execution on NVIDIA GPUs
+- 🖼️ **RTX Rendering**: Evaluate vision-based policies with realistic rendering, reflections and refractions
+- 📦 **LeRobot-compatible datasets**: Ready for training with GR00T N1x, PI0, SmolVLA, ACT, and Diffusion policies
+- 🔄 **EnvHub integration**: Load environments from HuggingFace EnvHub with one line
+
+## Installation
+
+### Prerequisites
+
+Hardware requirements are shared with Isaac Sim, and are detailed in [Isaac Sim Requirements](https://docs.isaacsim.omniverse.nvidia.com/5.1.0/installation/requirements.html).
+
+- NVIDIA GPU with CUDA support
+- NVIDIA driver compatible with IsaacSim 5.1.0
+- Linux (Ubuntu 22.04 / 24.04)
+
+### Setup
+
+```bash
+# 1. Create conda environment
+conda create -y -n lerobot-arena python=3.11
+conda activate lerobot-arena
+conda install -y -c conda-forge ffmpeg=7.1.1
+
+# 2. Install Isaac Sim 5.1.0
+pip install "isaacsim[all,extscache]==5.1.0" --extra-index-url https://pypi.nvidia.com
+
+# Accept NVIDIA EULA (required)
+export ACCEPT_EULA=Y
+export PRIVACY_CONSENT=Y
+
+# 3. Install IsaacLab 2.3.0
+git clone https://github.com/isaac-sim/IsaacLab.git
+cd IsaacLab
+git checkout v2.3.0
+./isaaclab.sh -i
+cd ..
+
+# 4. Install IsaacLab Arena
+git clone https://github.com/isaac-sim/IsaacLab-Arena.git
+cd IsaacLab-Arena
+git checkout release/0.1.1
+pip install -e .
+cd ..
+
+
+# 5. Install LeRobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e .
+cd ..
+
+
+# 6. Install additional dependencies
+pip install onnxruntime==1.23.2 lightwheel-sdk==1.0.1 vuer[all]==0.0.70 qpsolvers==4.8.1
+pip install numpy==1.26.0 # Isaac Sim 5.1 depends on numpy==1.26.0, this will be fixed in next release
+```
+
+## Evaluating Policies
+
+### Pre-trained Policies
+
+The following trained policies are available:
+
+| Policy                      | Architecture | Task          | Link                                                                     |
+| :-------------------------- | :----------- | :------------ | :----------------------------------------------------------------------- |
+| pi05-arena-gr1-microwave    | PI0.5        | GR1 Microwave | [HuggingFace](https://huggingface.co/nvidia/pi05-arena-gr1-microwave)    |
+| smolvla-arena-gr1-microwave | SmolVLA      | GR1 Microwave | [HuggingFace](https://huggingface.co/nvidia/smolvla-arena-gr1-microwave) |
+
+### Evaluate SmolVLA
+
+```bash
+pip install -e ".[smolvla]"
+pip install numpy==1.26.0 # revert numpy to version 1.26
+```
+
+```bash
+lerobot-eval \
+    --policy.path=nvidia/smolvla-arena-gr1-microwave \
+    --env.type=isaaclab_arena \
+    --env.hub_path=nvidia/isaaclab-arena-envs \
+    --rename_map='{"observation.images.robot_pov_cam_rgb": "observation.images.robot_pov_cam"}' \
+    --policy.device=cuda \
+    --env.environment=gr1_microwave \
+    --env.embodiment=gr1_pink \
+    --env.object=mustard_bottle \
+    --env.headless=false \
+    --env.enable_cameras=true \
+    --env.video=true \
+    --env.video_length=10 \
+    --env.video_interval=15 \
+    --env.state_keys=robot_joint_pos \
+    --env.camera_keys=robot_pov_cam_rgb \
+    --trust_remote_code=True \
+    --eval.batch_size=1
+```
+
+### Evaluate PI0.5
+
+```bash
+pip install -e ".[pi]"
+pip install numpy==1.26.0 # revert numpy to version 1.26
+```
+
+<Tip>PI0.5 requires disabling torch compile for evaluation:</Tip>
+
+```bash
+TORCH_COMPILE_DISABLE=1 TORCHINDUCTOR_DISABLE=1 lerobot-eval \
+    --policy.path=nvidia/pi05-arena-gr1-microwave \
+    --env.type=isaaclab_arena \
+    --env.hub_path=nvidia/isaaclab-arena-envs \
+    --rename_map='{"observation.images.robot_pov_cam_rgb": "observation.images.robot_pov_cam"}' \
+    --policy.device=cuda \
+    --env.environment=gr1_microwave \
+    --env.embodiment=gr1_pink \
+    --env.object=mustard_bottle \
+    --env.headless=false \
+    --env.enable_cameras=true \
+    --env.video=true \
+    --env.video_length=15 \
+    --env.video_interval=15 \
+    --env.state_keys=robot_joint_pos \
+    --env.camera_keys=robot_pov_cam_rgb \
+    --trust_remote_code=True \
+    --eval.batch_size=1
+```
+
+<Tip>
+  To change the number of parallel environments, use the ```--eval.batch_size```
+  flag.
+</Tip>
+
+### What to Expect
+
+During evaluation, you will see a progress bar showing the running success rate:
+
+```
+Stepping through eval batches:   8%|██████▍    | 4/50 [00:45<08:06, 10.58s/it, running_success_rate=25.0%]
+```
+
+### Video Recording
+
+To enable video recording during evaluation, add the following flags to your command:
+
+```bash
+--env.video=true \
+--env.video_length=15 \
+--env.video_interval=15
+```
+
+For more details on video recording, see the [IsaacLab Recording Documentation](https://isaac-sim.github.io/IsaacLab/main/source/how-to/record_video.html).
+
+<Tip>
+When running headless with `--env.headless=true`, you must also enable cameras explicitly for camera enabled environments:
+
+```bash
+--env.headless=true --env.enable_cameras=true
+```
+
+</Tip>
+
+### Output Directory
+
+Evaluation videos are saved to the output directory with the following structure:
+
+```
+outputs/eval/<date>/<timestamp>_<env>_<policy>/videos/<task>_<env_id>/eval_episode_<n>.mp4
+```
+
+For example:
+
+```
+outputs/eval/2026-01-02/14-38-01_isaaclab_arena_smolvla/videos/gr1_microwave_0/eval_episode_0.mp4
+```
+
+## Training Policies
+
+To learn more about training policies with LeRobot, please refer to the training documentation:
+
+- [SmolVLA](./smolvla)
+- [Pi0.5](./pi05)
+- [GR00T N1.5](./groot)
+
+Sample IsaacLab Arena datasets are available on HuggingFace Hub for experimentation:
+
+| Dataset                                                                                                   | Description                | Frames |
+| :-------------------------------------------------------------------------------------------------------- | :------------------------- | :----- |
+| [Arena-GR1-Manipulation-Task](https://huggingface.co/datasets/nvidia/Arena-GR1-Manipulation-Task-v3)      | GR1 microwave manipulation | ~4K    |
+| [Arena-G1-Loco-Manipulation-Task](https://huggingface.co/datasets/nvidia/Arena-G1-Loco-Manipulation-Task) | G1 loco-manipulation       | ~4K    |
+
+## Environment Configuration
+
+### Full Configuration Options
+
+```python
+from lerobot.envs.configs import IsaaclabArenaEnv
+
+config = IsaaclabArenaEnv(
+    # Environment selection
+    environment="gr1_microwave",      # Task environment
+    embodiment="gr1_pink",            # Robot embodiment
+    object="power_drill",             # Object to manipulate
+
+    # Simulation settings
+    episode_length=300,               # Max steps per episode
+    headless=True,                    # Run without GUI
+    device="cuda:0",                  # GPU device
+    seed=42,                          # Random seed
+
+    # Observation configuration
+    state_keys="robot_joint_pos",     # State observation keys (comma-separated)
+    camera_keys="robot_pov_cam_rgb",  # Camera observation keys (comma-separated)
+    state_dim=54,                     # Expected state dimension
+    action_dim=36,                    # Expected action dimension
+    camera_height=512,                # Camera image height
+    camera_width=512,                 # Camera image width
+    enable_cameras=True,              # Enable camera observations
+
+    # Video recording
+    video=False,                      # Enable video recording
+    video_length=100,                 # Frames per video
+    video_interval=200,               # Steps between recordings
+
+    # Advanced
+    mimic=False,                      # Enable mimic mode
+    teleop_device=None,               # Teleoperation device
+    disable_fabric=False,             # Disable fabric optimization
+    enable_pinocchio=True,            # Enable Pinocchio for IK
+)
+```
+
+### Using Environment Hub directly for advanced usage
+
+Create a file called `test_env_load_arena.py` or [download from the EnvHub](https://huggingface.co/nvidia/isaaclab-arena-envs/blob/main/tests/test_env_load_arena.py):
+
+```python
+import logging
+from dataclasses import asdict
+from pprint import pformat
+import torch
+import tqdm
+from lerobot.configs import parser
+from lerobot.configs.eval import EvalPipelineConfig
+
+
+@parser.wrap()
+def main(cfg: EvalPipelineConfig):
+    """Run random action rollout for IsaacLab Arena environment."""
+    logging.info(pformat(asdict(cfg)))
+
+    from lerobot.envs.factory import make_env
+
+    env_dict = make_env(
+        cfg.env,
+        n_envs=cfg.env.num_envs,
+        trust_remote_code=True,
+    )
+    env = next(iter(env_dict.values()))[0]
+    env.reset()
+    for _ in tqdm.tqdm(range(cfg.env.episode_length)):
+        with torch.inference_mode():
+            actions = env.action_space.sample()
+            obs, rewards, terminated, truncated, info = env.step(actions)
+            if terminated.any() or truncated.any():
+                obs, info = env.reset()
+    env.close()
+
+
+if __name__ == "__main__":
+    main()
+```
+
+Run with:
+
+```bash
+python test_env_load_arena.py \
+    --env.environment=g1_locomanip_pnp \
+    --env.embodiment=gr1_pink \
+    --env.object=cracker_box \
+    --env.num_envs=4 \
+    --env.enable_cameras=true \
+    --env.seed=1000 \
+    --env.video=true \
+    --env.video_length=10 \
+    --env.video_interval=15 \
+    --env.headless=false \
+    --env.hub_path=nvidia/isaaclab-arena-envs \
+    --env.type=isaaclab_arena
+```
+
+## Creating New Environments
+
+First create a new IsaacLab Arena environment by following the [IsaacLab Arena Documentation](https://isaac-sim.github.io/IsaacLab-Arena/release/0.1.1/index.html).
+
+Clone our EnvHub repo:
+
+```bash
+git clone https://huggingface.co/nvidia/isaaclab-arena-envs
+```
+
+Modify the `example_envs.yaml` file based on your new environment.
+[Upload](./envhub#step-3-upload-to-the-hub) your modified repo to HuggingFace EnvHub.
+
+<Tip>
+  Your IsaacLab Arena environment code must be locally available during
+  evaluation. Users can clone your environment repository separately, or you can
+  bundle the environment code and assets directly in your EnvHub repo.
+</Tip>
+
+Then, when evaluating, use your new environment:
+
+```bash
+lerobot-eval \
+    --env.hub_path=<your-env-hub-path>/isaaclab-arena-envs \
+    --env.environment=<your new environment> \
+    ...other flags...
+```
+
+We look forward to your contributions!
+
+## Troubleshooting
+
+### CUDA out of memory
+
+Reduce `batch_size` or use a GPU with more VRAM:
+
+```bash
+--eval.batch_size=1
+```
+
+### EULA not accepted
+
+Set environment variables before running:
+
+```bash
+export ACCEPT_EULA=Y
+export PRIVACY_CONSENT=Y
+```
+
+### Video recording not working
+
+Enable cameras when running headless:
+
+```bash
+--env.video=true --env.enable_cameras=true --env.headless=true
+```
+
+### Policy output dimension mismatch
+
+Ensure `action_dim` matches your policy:
+
+```bash
+--env.action_dim=36
+```
+
+### libGLU.so.1 Errors during Isaac Sim initialization
+
+Ensure you have the following dependencies installed, this is likely to happen on headless machines.
+
+```bash
+sudo apt update && sudo apt install -y libglu1-mesa libxt6
+```
+
+## See Also
+
+- [EnvHub Documentation](./envhub.mdx) - General EnvHub usage
+- [IsaacLab Arena GitHub](https://github.com/isaac-sim/IsaacLab-Arena)
+- [IsaacLab Documentation](https://isaac-sim.github.io/IsaacLab/)
+
+## Lightwheel LW-BenchHub
+
+[Lightwheel](https://www.lightwheel.ai) is bringing `Lightwheel-Libero-Tasks` and `Lightwheel-RoboCasa-Tasks` with 268 tasks to the LeRobot ecosystem.
+LW-BenchHub collects and generates large-scale datasets via teleoperation that comply with the LeRobot specification, enabling out-of-the-box training and evaluation workflows.
+With the unified interface provided by EnvHub, developers can quickly build end-to-end experimental pipelines.
+
+### Install
+
+Assuming you followed the [Installation](#installation) steps, you can install LW-BenchHub with:
+
+```bash
+conda install pinocchio -c conda-forge -y
+pip install numpy==1.26.0 # revert numpy to version 1.26
+
+sudo apt-get install git-lfs && git lfs install
+
+git clone https://github.com/LightwheelAI/lw_benchhub
+git lfs pull # Ensure LFS files (e.g., .usd assets) are downloaded
+
+cd lw_benchhub
+pip install -e .
+```
+
+For more detailed instructions, please refer to the [LW-BenchHub Documentation](https://docs.lightwheel.net/lw_benchhub/usage/Installation).
+
+### Lightwheel Tasks Dataset
+
+LW-BenchHub datasets are available on HuggingFace Hub:
+
+| Dataset                                                                                                       | Description             | Tasks | Frames |
+| :------------------------------------------------------------------------------------------------------------ | :---------------------- | :---- | :----- |
+| [Lightwheel-Tasks-X7S](https://huggingface.co/datasets/LightwheelAI/Lightwheel-Tasks-X7S)                     | X7S LIBERO and RoboCasa | 117   | ~10.3M |
+| [Lightwheel-Tasks-Double-Piper](https://huggingface.co/datasets/LightwheelAI/Lightwheel-Tasks-Double-Piper)   | Double-Piper LIBERO     | 130   | ~6.0M  |
+| [Lightwheel-Tasks-G1-Controller](https://huggingface.co/datasets/LightwheelAI/Lightwheel-Tasks-G1-Controller) | G1-Controller LIBERO    | 62    | ~2.7M  |
+| [Lightwheel-Tasks-G1-WBC](https://huggingface.co/datasets/LightwheelAI/Lightwheel-Tasks-G1-WBC)               | G1-WBC RoboCasa         | 32    | ~1.5M  |
+
+For training policies, refer to the [Training Policies](#training-policies) section.
+
+### Evaluating Policies
+
+#### Pre-trained Policies
+
+The following trained policies are available:
+
+| Policy                   | Architecture | Task                           | Layout     | Robot           | Link                                                                                  |
+| :----------------------- | :----------- | :----------------------------- | :--------- | :-------------- | :------------------------------------------------------------------------------------ |
+| smolvla-double-piper-pnp | SmolVLA      | L90K1PutTheBlackBowlOnThePlate | libero-1-1 | DoublePiper-Abs | [HuggingFace](https://huggingface.co/LightwheelAI/smolvla-double-piper-pnp/tree/main) |
+
+#### Evaluate SmolVLA
+
+```bash
+lerobot-eval \
+  --policy.path=LightwheelAI/smolvla-double-piper-pnp \
+  --env.type=isaaclab_arena \
+  --rename_map='{"observation.images.left_hand_camera_rgb": "observation.images.left_hand", "observation.images.right_hand_camera_rgb": "observation.images.right_hand", "observation.images.first_person_camera_rgb": "observation.images.first_person"}' \
+  --env.hub_path=LightwheelAI/lw_benchhub_env \
+  --env.kwargs='{"config_path": "configs/envhub/example.yml"}' \
+  --trust_remote_code=true \
+  --env.state_keys=joint_pos \
+  --env.action_dim=12 \
+  --env.camera_keys=left_hand_camera_rgb,right_hand_camera_rgb,first_person_camera_rgb \
+  --policy.device=cuda \
+  --eval.batch_size=10 \
+  --eval.n_episodes=100
+```
+
+### Environment Configuration
+
+Evaluation can be quickly launched by modifying the `robot`, `task`, and `layout` settings in the configuration file.
+
+#### Full Configuration Options
+
+```yml
+# =========================
+# Basic Settings
+# =========================
+disable_fabric: false
+device: cuda:0
+sensitivity: 1.0
+step_hz: 50
+enable_cameras: true
+execute_mode: eval
+episode_length_s: 20.0 # Episode length in seconds, increase if episodes timeout during eval
+
+# =========================
+# Robot Settings
+# =========================
+robot: DoublePiper-Abs # Robot type, DoublePiper-Abs, X7S-Abs, G1-Controller or G1-Controller-DecoupledWBC
+robot_scale: 1.0
+
+# =========================
+# Task & Scene Settings
+# =========================
+task: L90K1PutTheBlackBowlOnThePlate # Task name
+scene_backend: robocasa
+task_backend: robocasa
+debug_assets: null
+layout: libero-1-1 # Layout and style ID
+sources:
+  - objaverse
+  - lightwheel
+  - aigen_objs
+object_projects: []
+usd_simplify: false
+seed: 42
+
+# =========================
+# Object Placement Retry Settings
+# =========================
+max_scene_retry: 4
+max_object_placement_retry: 3
+
+resample_objects_placement_on_reset: true
+resample_robot_placement_on_reset: true
+
+# =========================
+# Replay Configuration Settings
+# =========================
+replay_cfgs:
+  add_camera_to_observation: true
+  render_resolution: [640, 480]
+```
+
+### See Also
+
+- [LW-BenchHub GitHub](https://github.com/LightwheelAI/LW-BenchHub)
+- [LW-BenchHub Documentation](https://docs.lightwheel.net/lw_benchhub/)
@@ -137,7 +137,8 @@ from lerobot.teleoperators import (  # noqa: F401
    Teleoperator,
    TeleoperatorConfig,
    make_teleoperator_from_config,
-    so101_leader,
+    so_leader,
+    bi_so_leader,
 )
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import init_logging
@@ -196,7 +197,7 @@ def teleop_loop(teleop: Teleoperator, env: gym.Env, fps: int):
            obs, info = env.reset()

        dt_s = time.perf_counter() - loop_start
-        precise_sleep(1 / fps - dt_s)
+        precise_sleep(max(1 / fps - dt_s, 0.0))
        loop_s = time.perf_counter() - loop_start
        print(f"\ntime: {loop_s * 1e3:.2f}ms ({1 / loop_s:.0f} Hz)")

@@ -222,7 +223,7 @@ def teleoperate(cfg: TeleoperateConfig):

 def main():
    teleoperate(TeleoperateConfig(
-        teleop=so101_leader.SO101LeaderConfig(
+        teleop=so_leader.SO101LeaderConfig(
            port="/dev/ttyACM0",
            id='leader',
            use_degrees=False,
@@ -12,6 +12,12 @@ Developers and researchers can post-train GR00T N1.5 with their own real or synt

 GR00T N1.5 (specifically the GR00T-N1.5-3B model) is built using pre-trained vision and language encoders. It utilizes a flow matching action transformer to model a chunk of actions, conditioned on vision, language, and proprioception.

+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-groot-paper1%20(1).png"
+  alt="An overview of GR00T"
+  width="80%"
+/>
+
 Its strong performance comes from being trained on an expansive and diverse humanoid dataset, which includes:

 - Real captured data from robots.
@@ -103,7 +109,7 @@ Once you have trained your model using your parameters you can run inference in

 ```bash
 lerobot-record \
-  --robot.type=bi_so100_follower \
+  --robot.type=bi_so_follower \
  --robot.left_arm_port=/dev/ttyACM1 \
  --robot.right_arm_port=/dev/ttyACM0 \
  --robot.id=bimanual_follower \
@@ -114,9 +120,12 @@ lerobot-record \
  --display_data=true \
  --dataset.repo_id=<user>/eval_groot-bimanual  \
  --dataset.num_episodes=10 \
-  --dataset.single_task="Grab and handover the red cube to the other arm"
-  --policy.path=<user>/groot-bimanual # your trained model
-  --dataset.episode_time_s=30
+  --dataset.single_task="Grab and handover the red cube to the other arm" \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
+  --policy.path=<user>/groot-bimanual \ # your trained model
+  --dataset.episode_time_s=30 \
  --dataset.reset_time_s=10
 ```

@@ -224,12 +224,15 @@ lerobot-record \
    --teleop.port=/dev/tty.usbmodem1201 \
    --teleop.id=right \
    --teleop.side=right \
-    --dataset.repo_id=nepyope/hand_record_test_with_video_data \
+    --dataset.repo_id=<USER>/hand_record_test_with_video_data \
    --dataset.single_task="Hand recording test with video data" \
    --dataset.num_episodes=1 \
    --dataset.episode_time_s=5 \
    --dataset.push_to_hub=true \
    --dataset.private=true \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    # --dataset.vcodec=auto \
    --display_data=true
 ```

@@ -241,7 +244,7 @@ lerobot-replay \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=right \
    --robot.side=right \
-    --dataset.repo_id=nepyope/hand_record_test_with_camera \
+    --dataset.repo_id=<USER>/hand_record_test_with_camera \
    --dataset.episode=0
 ```

@@ -249,13 +252,13 @@ lerobot-replay \

 ```bash
 lerobot-train \
-  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
+  --dataset.repo_id=<USER>/hand_record_test_with_video_data \
  --policy.type=act \
  --output_dir=outputs/train/hopejr_hand \
  --job_name=hopejr \
  --policy.device=mps \
  --wandb.enable=true \
-  --policy.repo_id=nepyope/hand_test_policy
+  --policy.repo_id=<USER>/hand_test_policy
 ```

 ### Evaluate
@@ -270,8 +273,11 @@ lerobot-record \
  --robot.side=right \
  --robot.cameras='{"main": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}}' \
  --display_data=false \
-  --dataset.repo_id=nepyope/eval_hopejr \
+  --dataset.repo_id=<USER>/eval_hopejr \
  --dataset.single_task="Evaluate hopejr hand policy" \
  --dataset.num_episodes=10 \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
  --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```
@@ -58,8 +58,8 @@ lerobot-teleoperate \

 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so101_leader import SO101LeaderConfig, SO101Leader
-from lerobot.robots.so101_follower import SO101FollowerConfig, SO101Follower
+from lerobot.teleoperators.so_leader import SO101LeaderConfig, SO101Leader
+from lerobot.robots.so_follower import SO101FollowerConfig, SO101Follower

 robot_config = SO101FollowerConfig(
    port="/dev/tty.usbmodem58760431541",
@@ -159,13 +159,13 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:

 ```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```

 Then store your Hugging Face repository name in a variable:

 ```bash
-HF_USER=$(hf auth whoami | head -n 1)
+HF_USER=$(NO_COLOR=1 hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```

@@ -185,7 +185,10 @@ lerobot-record \
    --display_data=true \
    --dataset.repo_id=${HF_USER}/record-test \
    --dataset.num_episodes=5 \
-    --dataset.single_task="Grab the black cube"
+    --dataset.single_task="Grab the black cube" \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
 ```
 </hfoption>
 <hfoption id="API example">
@@ -195,9 +198,9 @@ lerobot-record \
 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.robots.so100_follower import SO100Follower, SO100FollowerConfig
-from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
-from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
+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.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
@@ -324,7 +327,7 @@ You can look for other LeRobot datasets on the hub by searching for `LeRobot` [t
 You can also push your local dataset to the Hub manually, running:

 ```bash
-huggingface-cli upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
+hf upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
 ```

 #### Record function
@@ -408,8 +411,8 @@ lerobot-replay \
 import time

 from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so_follower.so100_follower import SO100Follower
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say

@@ -432,7 +435,7 @@ for idx in range(dataset.num_frames):
    }
    robot.send_action(action)

-    precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
+    precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))

 robot.disconnect()
 ```
@@ -488,7 +491,7 @@ If your local computer doesn't have a powerful GPU you could utilize Google Cola
 Once training is done, upload the latest checkpoint with:

 ```bash
-huggingface-cli upload ${HF_USER}/act_so101_test \
+hf upload ${HF_USER}/act_so101_test \
  outputs/train/act_so101_test/checkpoints/last/pretrained_model
 ```

@@ -496,7 +499,7 @@ You can also upload intermediate checkpoints with:

 ```bash
 CKPT=010000
-huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
+hf upload ${HF_USER}/act_so101_test${CKPT} \
  outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
 ```

@@ -515,6 +518,9 @@ lerobot-record  \
  --display_data=false \
  --dataset.repo_id=${HF_USER}/eval_so100 \
  --dataset.single_task="Put lego brick into the transparent box" \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
  # <- Teleop optional if you want to teleoperate in between episodes \
  # --teleop.type=so100_leader \
  # --teleop.port=/dev/ttyACM0 \
@@ -531,8 +537,8 @@ 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.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so_follower.so100_follower import SO100Follower
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
@@ -1,30 +1,57 @@
 # Installation

-## Install [`miniforge`](https://conda-forge.org/download/)
+This guide uses `conda` (via miniforge) to manage environments (recommended). If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.12 and `ffmpeg` installed with the `libsvtav1` encoder, then skip ahead to [Environment Setup](#step-2-environment-setup).
+
+## Step 1 (`conda` only): Install [`miniforge`](https://conda-forge.org/download/)

 ```bash
 wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
 bash Miniforge3-$(uname)-$(uname -m).sh
 ```

-## Environment Setup
+## Step 2: Environment Setup

-Create a virtual environment with Python 3.10, using conda:
+Create a virtual environment with Python 3.12:

+<!-- prettier-ignore-start -->
+<hfoptions id="create_venv">
+<hfoption id="conda">
 ```bash
-conda create -y -n lerobot python=3.10
+conda create -y -n lerobot python=3.12
 ```
-
-Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
-
+</hfoption>
+<hfoption id="uv">
 ```bash
+uv python install 3.12
+uv venv --python 3.12
+```
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->
+
+Then activate your virtual environment, you have to do this each time you open a shell to use lerobot:
+
+<!-- prettier-ignore-start -->
+<hfoptions id="activate_venv">
+<hfoption id="conda">```bash
 conda activate lerobot
+```</hfoption>
+<hfoption id="uv">
+```bash
+# Linux/macOSsource
+source .venv/bin/activate
+# Windows PowerShell
+source .venv\Scripts\Activate.ps1
 ```
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->

 When using `conda`, install `ffmpeg` in your environment:

 ```bash
 conda install ffmpeg -c conda-forge
+ffmpeg -version  # ffmpeg 8.X is not yet supported !
 ```

 > [!TIP]
@@ -38,7 +65,17 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.

-## Install LeRobot 🤗
+> [!NOTE]
+> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to install `evdev` with the following command:
+>
+> ```bash
+> conda install evdev -c conda-forge
+> ```
+
+> [!IMPORTANT]
+> If you are using `uv` you will have to install `ffmpeg` system-wide (outside of the virtual environment). You rely on `uv` and `torchcodec` ability to dynamically link to the system `ffmpeg`.
+
+## Step 3: Install LeRobot 🤗

 ### From Source

@@ -51,23 +88,45 @@ cd lerobot

 Then, install the library in editable mode. This is useful if you plan to contribute to the code.

+<!-- prettier-ignore-start -->
+<hfoptions id="install_lerobot_src">
+<hfoption id="conda">
 ```bash
 pip install -e .
 ```
+</hfoption>
+<hfoption id="uv">
+```bash
+uv pip install -e .
+```
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->

 ### Installation from PyPI

 **Core Library:**
 Install the base package with:

+<!-- prettier-ignore-start -->
+<hfoptions id="install_lerobot_pypi">
+<hfoption id="conda">
 ```bash
 pip install lerobot
 ```
+</hfoption>
+<hfoption id="uv">
+```bash
+uv pip install lerobot
+```
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->

 _This installs only the default dependencies._

 **Extra Features:**
-To install additional functionality, use one of the following:
+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.):

 ```bash
 pip install 'lerobot[all]'          # All available features
@@ -81,13 +140,10 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/

-> [!NOTE]
-> For lerobot 0.4.0, if you want to install pi, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`
-
 ### Troubleshooting

 If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
-To install these for linux run:
+To install these for Linux run:

 ```bash
 sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev
@@ -97,7 +153,7 @@ For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/

 ## Optional dependencies

-LeRobot provides optional extras for specific functionalities. Multiple extras can be combined (e.g., `.[aloha,feetech]`). For all available extras, refer to `pyproject.toml`.
+LeRobot provides optional extras for specific functionalities. Multiple extras can be combined (e.g., `.[aloha,feetech]`). For all available extras, refer to `pyproject.toml`. If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.

 ### Simulations

@@ -18,7 +18,7 @@ If you're using Feetech or Dynamixel motors, LeRobot provides built-in bus inter
 - [`DynamixelMotorsBus`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/dynamixel/dynamixel.py) – for controlling Dynamixel servos

 Please refer to the [`MotorsBus`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/motors_bus.py) abstract class to learn about its API.
-For a good example of how it can be used, you can have a look at our own [SO101 follower implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/so101_follower/so101_follower.py)
+For a good example of how it can be used, you can have a look at our own [SO101 follower implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/so_follower/so101_follower/so101_follower.py)

 Use these if compatible. Otherwise, you'll need to find or write a Python interface (not covered in this tutorial):

@@ -1,5 +1,11 @@
 # LeKiwi

+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/1740517739083.jpeg"
+  alt="LeKiwi"
+  width="70%"
+/>
+
 In the steps below, we explain how to assemble the LeKiwi mobile robot.

 ## Source the parts
@@ -204,7 +210,7 @@ lerobot-calibrate \

 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so100_leader import SO100LeaderConfig, SO100Leader
+from lerobot.teleoperators.so_leader import SO100LeaderConfig, SO100Leader

 config = SO100LeaderConfig(
    port="/dev/tty.usbmodem58760431551",
@@ -273,13 +279,13 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:

 ```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```

 Then store your Hugging Face repository name in a variable:

 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```

@@ -41,7 +41,10 @@ lerobot-record \
  --display_data=true \
  --dataset.repo_id=${HF_USER}/record-test \
  --dataset.num_episodes=5 \
-  --dataset.single_task="Grab the black cube"
+  --dataset.single_task="Grab the black cube" \
+  --dataset.streaming_encoding=true \
+  # --dataset.vcodec=auto \
+  --dataset.encoder_threads=2
 ```

 See the [recording guide](./il_robots#record-a-dataset) for more details.
@@ -42,6 +42,7 @@ lerobot-eval \
 ```

 - `--env.task` picks the suite (`libero_object`, `libero_spatial`, etc.).
+- `--env.task_ids` picks task ids to run (`[0]`, `[1,2,3]`, etc.). Omit this flag (or set it to `null`) 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 in total.

@@ -0,0 +1,197 @@
+## Order and Assemble the parts
+
+First, assemble the OMX hardware following the official assembly guide.
+
+OMX Assembly Guide: https://ai.robotis.com/omx/assembly_guide_omx.html
+
+OMX robots are shipped preconfigured from the factory. Motor IDs, communication parameters, and joint offsets are already set, so no additional motor setup or calibration is required before using LeRobot.
+
+## Install LeRobot 🤗
+
+To install LeRobot, follow our [Installation Guide](./installation)
+
+In addition to these instructions, you need to install the Dynamixel SDK:
+
+```bash
+pip install -e ".[dynamixel]"
+```
+
+## Connect the robot
+
+To find the port for each bus servo adapter, run this script:
+
+```bash
+lerobot-find-port
+```
+
+This command runs and when prompted, disconnect the USB cable from either the leader or follower arm and press Enter. The output will show 'The port of this MotorsBus is [port]'. This identifies the port for the disconnected arm. Repeat for the other arm to identify both ports.
+
+<hfoptions id="find_port">
+<hfoption id="Mac">
+
+Example output on macOS:
+
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the USB cable from your MotorsBus and press Enter when done.
+
+[...Disconnect corresponding leader or follower arm and press Enter...]
+
+The port of this MotorsBus is /dev/tty.usbmodem575E0032081
+Reconnect the USB cable.
+```
+
+Where the found port is: `/dev/tty.usbmodem575E0032081` corresponding to your leader or follower arm.
+
+</hfoption>
+<hfoption id="Linux">
+
+On Linux, we strongly recommend using udev rules to assign persistent and human-readable device names to the OMX leader and follower arms. This avoids issues where device names such as ttyACM0 and ttyACM1 change when the robot is unplugged, replugged, or when the system is rebooted.
+
+#### 1. Find your device serial numbers
+
+You should have obtained the port numbers like ../../ttyACM? for the leader and follower using `lerobot-find-port`. You can match those results with the serial numbers using the `ls -l /dev/serial/by-id/` command.
+To create udev rules, you need the unique serial number for each OMX device. The easiest way is to list devices under:
+
+```bash
+ls -l /dev/serial/by-id/
+```
+
+You will see output similar to:
+
+```bash
+usb-ROBOTIS_OpenRB-150_228BDD7B503059384C2E3120FF0A2B19-if00 -> ../../ttyACM0
+usb-ROBOTIS_OpenRB-150_67E1ED68503059384C2E3120FF092234-if00 -> ../../ttyACM1
+```
+
+In each line, the serial number is the long string after `usb-ROBOTIS_OpenRB-150_` and before `-if00`.
+
+Follower serial: `228BDD7B503059384C2E3120FF0A2B19`
+
+Leader serial: `67E1ED68503059384C2E3120FF092234`
+
+#### 2. Create the udev rule
+
+Create a new udev rule file:
+
+```bash
+sudo nano /etc/udev/rules.d/99-omx.rules
+```
+
+Paste the following lines, replacing the serial numbers with the values you found above:
+
+```bash
+SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{serial}=="228BDD7B503059384C2E3120FF0A2B19", SYMLINK+="omx_follower"
+SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{serial}=="67E1ED68503059384C2E3120FF092234", SYMLINK+="omx_leader"
+```
+
+Save the file and reload udev rules:
+
+```bash
+sudo udevadm control --reload-rules
+sudo udevadm trigger
+```
+
+Now unplug and replug both devices once.
+
+#### 3. Verify the symlinks
+
+Check that the persistent device names exist:
+
+```bash
+ls -l /dev/omx_follower /dev/omx_leader
+```
+
+You should see them pointing to ttyACM\* devices:
+
+```bash
+/dev/omx_follower -> ttyACM*
+/dev/omx_leader   -> ttyACM*
+```
+
+These names remain stable across reboots and reconnections.
+
+</hfoption>
+</hfoptions>
+
+## Teleoperate
+
+After identifying the correct ports, you can directly teleoperate the follower arm using the leader arm.
+
+<hfoptions id="teleoperate">
+<hfoption id="Mac">
+
+### Teleoperate without camera
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=<your_follower_port> \
+  --robot.id=omx_follower_arm \
+  --teleop.type=omx_leader \
+  --teleop.port=<your_leader_port> \
+  --teleop.id=omx_leader_arm
+```
+
+During teleoperation, motions of the leader arm are mirrored in real time by the follower arm. OMX is already preconfigured, teleoperation can begin immediately without any calibration steps.
+
+### Teleoperate with camera
+
+You can also enable camera input during teleoperation by providing a camera configuration for the follower arm.
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=<your_follower_port> \
+  --robot.id=omx_follower_arm \
+  --robot.cameras="{front: {type: opencv, index_or_path: '/dev/video0', width: 640, height: 480, fps: 30}}" \
+  --teleop.type=omx_leader \
+  --teleop.port=<your_leader_port> \
+  --teleop.id=omx_leader_arm \
+  --display_data=true
+```
+
+When the camera is enabled, the camera stream is displayed in real time and synchronized with the robot state. This setup is useful for visual monitoring and can be reused later for demonstration recording and imitation learning.
+
+</hfoption>
+<hfoption id="Linux">
+
+### Teleoperate without camera
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=/dev/omx_follower \
+  --robot.id=omx_follower_arm \
+  --teleop.type=omx_leader \
+  --teleop.port=/dev/omx_leader \
+  --teleop.id=omx_leader_arm
+```
+
+During teleoperation, motions of the leader arm are mirrored in real time by the follower arm. OMX is already preconfigured, teleoperation can begin immediately without any calibration steps.
+
+### Teleoperate with camera
+
+You can also enable camera input during teleoperation by providing a camera configuration for the follower arm.
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=/dev/omx_follower \
+  --robot.id=omx_follower_arm \
+  --robot.cameras="{front: {type: opencv, index_or_path: '/dev/video0', width: 640, height: 480, fps: 30}}" \
+  --teleop.type=omx_leader \
+  --teleop.port=/dev/omx_leader \
+  --teleop.id=omx_leader_arm \
+  --display_data=true
+```
+
+When the camera is enabled, the camera stream is displayed in real time and synchronized with the robot state. This setup is useful for visual monitoring and can be reused later for demonstration recording and imitation learning.
+
+</hfoption>
+</hfoptions>
+
+Congrats 🎉, your robot is all set to learn a task on its own.
+
+> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/robotis).
@@ -0,0 +1,276 @@
+# OpenArm
+
+[OpenArm](https://openarm.dev) is an open-source 7DOF humanoid arm designed for physical AI research and deployment.
+
+To get your OpenArm, assembled or DIY, and join the global community, browse verified and certified manufacturers worldwide at [openarm.dev](https://openarm.dev).
+
+## What's Unique?
+
+- **Human-Scale Design**: OpenArm is designed with human-like proportions, scaled for a person around 160-165cm tall. This provides an optimal balance between practical reach and manageable inertia for safe, responsive operation.
+
+- **Safety-First Architecture**: Built with QDD backdrivable motors and high compliance, OpenArm prioritizes safe human-robot interaction while maintaining practical payload capabilities (6.0kg peak / 4.1kg nominal) for real-world tasks.
+
+- **Built for Durability**: Critical structural components use aluminum and stainless steel construction, ensuring robust performance for repetitive data collection and continuous research use.
+
+- **Fully Accessible & Buildable**: Every component, from CNC parts and 3D-printed casings to electrical wiring is designed to be purchasable and buildable by individual researchers and labs, with complete fabrication data provided.
+
+- **Practical & Affordable**: At $6,500 USD for a complete bimanual system, OpenArm delivers research-grade capabilities at a fraction of traditional humanoid robot costs.
+
+## Platform Requirements
+
+<Tip warning={true}>
+  **Linux Only**: OpenArm currently only works on Linux. The CAN bus USB adapter
+  does not have macOS drivers and has not been tested on Windows.
+</Tip>
+
+## Safety Guide
+
+Before operating OpenArm, please read the [official safety guide](https://docs.openarm.dev/getting-started/safety-guide). Key points:
+
+- **Secure installation**: Fasten the arm to a flat, stable surface with screws or clamps
+- **Safe distance**: Keep body parts and objects outside the range of motion during operation
+- **Protective equipment**: Always wear safety goggles; use additional PPE as needed
+- **Payload limits**: Do not exceed specified payload limits (6.0kg peak / 4.1kg nominal per arm)
+- **Emergency stop**: Know the location and operation of the emergency stop device
+- **Regular inspection**: Check for loose screws, damaged mechanical limits, unusual noises, and wiring damage
+
+## Hardware Setup
+
+Follow the official [OpenArm hardware documentation](https://docs.openarm.dev) for:
+
+- Bill of materials and sourcing
+- 3D printing instructions
+- Mechanical assembly
+- Electrical wiring
+
+The hardware repositories are available at [github.com/enactic/openarm](https://github.com/enactic/openarm).
+
+## CAN Bus Setup
+
+OpenArm uses CAN bus communication with Damiao motors. Once you have the CAN bus USB adapter plugged into your Linux PC, follow the [Damiao Motors and CAN Bus guide](./damiao) to configure the interface.
+
+Quick setup:
+
+```bash
+# Setup CAN interfaces
+lerobot-setup-can --mode=setup --interfaces=can0,can1
+
+# Test motor communication
+lerobot-setup-can --mode=test --interfaces=can0,can1
+```
+
+## Install LeRobot 🤗
+
+Follow our [Installation Guide](./installation), then install the Damiao motor support:
+
+```bash
+pip install -e ".[damiao]"
+```
+
+## Usage
+
+### Follower Arm (Robot)
+
+<hfoptions id="follower">
+<hfoption id="Command">
+
+```bash
+lerobot-calibrate \
+    --robot.type=openarm_follower \
+    --robot.port=can0 \
+    --robot.side=right \
+    --robot.id=my_openarm_follower
+```
+
+</hfoption>
+<hfoption id="API example">
+
+```python
+from lerobot.robots.openarm_follower import OpenArmFollower, OpenArmFollowerConfig
+
+config = OpenArmFollowerConfig(
+    port="can0",
+    side="right",  # or "left" for left arm
+    id="my_openarm_follower",
+)
+
+follower = OpenArmFollower(config)
+follower.connect()
+
+# Read current state
+obs = follower.get_observation()
+print(obs)
+
+# Send action (position in degrees)
+action = {
+    "joint_1.pos": 0.0,
+    "joint_2.pos": 0.0,
+    "joint_3.pos": 0.0,
+    "joint_4.pos": 45.0,
+    "joint_5.pos": 0.0,
+    "joint_6.pos": 0.0,
+    "joint_7.pos": 0.0,
+    "gripper.pos": 0.0,
+}
+follower.send_action(action)
+
+follower.disconnect()
+```
+
+</hfoption>
+</hfoptions>
+
+### Leader Arm (Teleoperator)
+
+The leader arm is used for teleoperation - manually moving it to control the follower arm.
+
+<hfoptions id="leader">
+<hfoption id="Command">
+
+```bash
+lerobot-calibrate \
+    --teleop.type=openarm_leader \
+    --teleop.port=can1 \
+    --teleop.id=my_openarm_leader
+```
+
+</hfoption>
+<hfoption id="API example">
+
+```python
+from lerobot.teleoperators.openarm_leader import OpenArmLeader, OpenArmLeaderConfig
+
+config = OpenArmLeaderConfig(
+    port="can1",
+    id="my_openarm_leader",
+    manual_control=True,  # Disable torque for manual movement
+)
+
+leader = OpenArmLeader(config)
+leader.connect()
+
+# Read current position (as action to send to follower)
+action = leader.get_action()
+print(action)
+
+leader.disconnect()
+```
+
+</hfoption>
+</hfoptions>
+
+### Teleoperation
+
+To teleoperate OpenArm with leader-follower control:
+
+```bash
+lerobot-teleoperate \
+    --robot.type=openarm_follower \
+    --robot.port=can0 \
+    --robot.side=right \
+    --robot.id=my_follower \
+    --teleop.type=openarm_leader \
+    --teleop.port=can1 \
+    --teleop.id=my_leader
+```
+
+### Bimanual Teleoperation
+
+To teleoperate a bimanual OpenArm setup with two leader and two follower arms:
+
+```bash
+lerobot-teleoperate \
+    --robot.type=bi_openarm_follower \
+    --robot.left_arm_config.port=can0 \
+    --robot.left_arm_config.side=left \
+    --robot.right_arm_config.port=can1 \
+    --robot.right_arm_config.side=right \
+    --robot.id=my_bimanual_follower \
+    --teleop.type=bi_openarm_leader \
+    --teleop.left_arm_config.port=can2 \
+    --teleop.right_arm_config.port=can3 \
+    --teleop.id=my_bimanual_leader
+```
+
+### Recording Data
+
+To record a dataset during teleoperation:
+
+```bash
+lerobot-record \
+    --robot.type=openarm_follower \
+    --robot.port=can0 \
+    --robot.side=right \
+    --robot.id=my_follower \
+    --teleop.type=openarm_leader \
+    --teleop.port=can1 \
+    --teleop.id=my_leader \
+    --repo-id=my_hf_username/my_openarm_dataset \
+    --fps=30 \
+    --num-episodes=10
+```
+
+## Configuration Options
+
+### Follower Configuration
+
+| Parameter             | Default   | Description                                                |
+| --------------------- | --------- | ---------------------------------------------------------- |
+| `port`                | -         | CAN interface (e.g., `can0`)                               |
+| `side`                | `None`    | Arm side: `"left"`, `"right"`, or `None` for custom limits |
+| `use_can_fd`          | `True`    | Enable CAN FD for higher data rates                        |
+| `can_bitrate`         | `1000000` | Nominal bitrate (1 Mbps)                                   |
+| `can_data_bitrate`    | `5000000` | CAN FD data bitrate (5 Mbps)                               |
+| `max_relative_target` | `None`    | Safety limit for relative target positions                 |
+| `position_kp`         | Per-joint | Position control proportional gains                        |
+| `position_kd`         | Per-joint | Position control derivative gains                          |
+
+### Leader Configuration
+
+| Parameter          | Default   | Description                         |
+| ------------------ | --------- | ----------------------------------- |
+| `port`             | -         | CAN interface (e.g., `can1`)        |
+| `manual_control`   | `True`    | Disable torque for manual movement  |
+| `use_can_fd`       | `True`    | Enable CAN FD for higher data rates |
+| `can_bitrate`      | `1000000` | Nominal bitrate (1 Mbps)            |
+| `can_data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps)        |
+
+## Motor Configuration
+
+OpenArm uses Damiao motors with the following default configuration:
+
+| Joint                       | Motor Type | Send ID | Recv ID |
+| --------------------------- | ---------- | ------- | ------- |
+| joint_1 (Shoulder pan)      | DM8009     | 0x01    | 0x11    |
+| joint_2 (Shoulder lift)     | DM8009     | 0x02    | 0x12    |
+| joint_3 (Shoulder rotation) | DM4340     | 0x03    | 0x13    |
+| joint_4 (Elbow flex)        | DM4340     | 0x04    | 0x14    |
+| joint_5 (Wrist roll)        | DM4310     | 0x05    | 0x15    |
+| joint_6 (Wrist pitch)       | DM4310     | 0x06    | 0x16    |
+| joint_7 (Wrist rotation)    | DM4310     | 0x07    | 0x17    |
+| gripper                     | DM4310     | 0x08    | 0x18    |
+
+## Troubleshooting
+
+### No Response from Motors
+
+1. Check power supply connections
+2. Verify CAN wiring (CAN-H, CAN-L, GND)
+3. Run diagnostics: `lerobot-setup-can --mode=test --interfaces=can0`
+4. See the [Damiao troubleshooting guide](./damiao#troubleshooting) for more details
+
+### CAN Interface Not Found
+
+Ensure the CAN interface is configured:
+
+```bash
+ip link show can0
+```
+
+## Resources
+
+- [OpenArm Website](https://openarm.dev)
+- [OpenArm Documentation](https://docs.openarm.dev)
+- [OpenArm GitHub](https://github.com/enactic/openarm)
+- [Safety Guide](https://docs.openarm.dev/getting-started/safety-guide)
+- [Damiao Motors and CAN Bus](./damiao)
@@ -44,7 +44,7 @@ Modify the examples to use `PhoneOS.IOS` or `PhoneOS.ANDROID` in `PhoneConfig`.

 Teleoperation example:

-```36:43:examples/phone_so100_teleop.py
+```python
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS

 teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
@@ -66,12 +66,13 @@ Run on of the examples scripts to teleoperate, record a dataset, replay a datase

 All scripts assume you configured your robot (e.g., SO-100 follower) and set the correct serial port.

-Additionally you need to **copy the urdf of the robot to the examples folder**. For the examples in this tutorial (Using SO100/SO101) it is highly recommended to use the urdf in the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf)
+Additionally you need to **copy the URDF of the robot into the examples folder**. For the examples in this tutorial (using SO100/SO101), copy the `SO101` folder from the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101) into the `examples/phone_to_so100/` directory, so that the URDF file path becomes `examples/phone_to_so100/SO101/so101_new_calib.urdf`.

 - Run this example to teleoperate:

  ```bash
-  python examples/phone_to_so100/teleoperate.py
+  cd examples/phone_to_so100
+  python teleoperate.py
  ```

 After running the example:
@@ -84,26 +85,29 @@ Additionally you can customize mapping or safety limits by editing the processor
 - Run this example to record a dataset, which saves absolute end effector observations and actions:

  ```bash
-  python examples/phone_to_so100/record.py
+  cd examples/phone_to_so100
+  python record.py
  ```

 - Run this example to replay recorded episodes:

  ```bash
-  python examples/phone_to_so100/replay.py
+  cd examples/phone_to_so100
+  python replay.py
  ```

 - Run this example to evaluate a pretrained policy:

  ```bash
-  python examples/phone_to_so100/evaluate.py
+  cd examples/phone_to_so100
+  python evaluate.py
  ```

 ### Important pipeline steps and options

 - Kinematics are used in multiple steps. We use [Placo](https://github.com/Rhoban/placo) which is a wrapper around Pinocchio for handling our kinematics. We construct the kinematics object by passing the robot's URDF and target frame. We set `target_frame_name` to the gripper frame.

-  ```examples/phone_to_so100/teleoperate.py
+  ```python
  kinematics_solver = RobotKinematics(
    urdf_path="./SO101/so101_new_calib.urdf",
    target_frame_name="gripper_frame_link",
@@ -114,7 +118,7 @@ Additionally you can customize mapping or safety limits by editing the processor

 - The `MapPhoneActionToRobotAction` step converts the calibrated phone pose and inputs into target deltas and gripper commands, below is shown what the step outputs.

-  ```src/lerobot/teleoperators/phone/phone_processor.py
+  ```python
  action["enabled"] = enabled
        action["target_x"] = -pos[1] if enabled else 0.0
        action["target_y"] = pos[0] if enabled else 0.0
@@ -127,7 +131,7 @@ Additionally you can customize mapping or safety limits by editing the processor

 - The `EEReferenceAndDelta` step converts target deltas to an absolute desired EE pose, storing a reference on enable, the `end_effector_step_sizes` are the step sizes for the EE pose and can be modified to change the motion speed.

-  ```examples/phone_to_so100/teleoperate.py
+  ```python
  EEReferenceAndDelta(
      kinematics=kinematics_solver,
      end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
@@ -138,7 +142,7 @@ Additionally you can customize mapping or safety limits by editing the processor

 - The `EEBoundsAndSafety` step clamps EE motion to a workspace and checks for large ee step jumps to ensure safety. The `end_effector_bounds` are the bounds for the EE pose and can be modified to change the workspace. The `max_ee_step_m` are the step limits for the EE pose and can be modified to change the safety limits.

-  ```examples/phone_to_so100/teleoperate.py
+  ```python
  EEBoundsAndSafety(
      end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
      max_ee_step_m=0.10,
@@ -147,7 +151,7 @@ Additionally you can customize mapping or safety limits by editing the processor

 - The `GripperVelocityToJoint` step turns a velocity‑like gripper input into absolute gripper position using the current measured state. The `speed_factor` is the factor by which the velocity is multiplied.

-  ```examples/phone_to_so100/teleoperate.py
+  ```python
  GripperVelocityToJoint(speed_factor=20.0)
  ```

@@ -157,7 +161,7 @@ We use different IK initial guesses in the kinematic steps. As initial guess eit

 - Closed loop (used in record/eval): sets `initial_guess_current_joints=True` so IK starts from the measured joints each frame.

-  ```examples/phone_to_so100/record.py
+  ```python
  InverseKinematicsEEToJoints(
      kinematics=kinematics_solver,
      motor_names=list(robot.bus.motors.keys()),
@@ -167,7 +171,7 @@ We use different IK initial guesses in the kinematic steps. As initial guess eit

 - Open loop (used in replay): sets `initial_guess_current_joints=False` so IK continues from the previous IK solution rather than the measured state. This preserves action stability when we replay without feedback.

-  ```examples/phone_to_so100/replay.py
+  ```python
  InverseKinematicsEEToJoints(
      kinematics=kinematics_solver,
      motor_names=list(robot.bus.motors.keys()),
@@ -6,6 +6,12 @@

 π₀ represents a breakthrough in robotics as the first general-purpose robot foundation model developed by [Physical Intelligence](https://www.physicalintelligence.company/blog/pi0). Unlike traditional robot programs that are narrow specialists programmed for repetitive motions, π₀ is designed to be a generalist policy that can understand visual inputs, interpret natural language instructions, and control a variety of different robots across diverse tasks.

+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-pi0%20(1).png"
+  alt="An overview of Pi0"
+  width="85%"
+/>
+
 ### The Vision for Physical Intelligence

 As described by Physical Intelligence, while AI has achieved remarkable success in digital domains, from chess-playing to drug discovery, human intelligence still dramatically outpaces AI in the physical world. To paraphrase Moravec's paradox, winning a game of chess represents an "easy" problem for AI, but folding a shirt or cleaning up a table requires solving some of the most difficult engineering problems ever conceived. π₀ represents a first step toward developing artificial physical intelligence that enables users to simply ask robots to perform any task they want, just like they can with large language models.
@@ -28,11 +34,6 @@ As described by Physical Intelligence, while AI has achieved remarkable success
   pip install -e ".[pi]"
   ```

-   > [!NOTE]
-   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
-   >
-   > This will be solved in the next patch release
-
 ## Training Data and Capabilities

 π₀ is trained on the largest robot interaction dataset to date, combining three key data sources:
@@ -54,7 +55,7 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
    --dataset.repo_id=your_dataset \
    --policy.type=pi0 \
    --output_dir=./outputs/pi0_training \
@@ -64,6 +65,8 @@ python src/lerobot/scripts/lerobot_train.py \
    --policy.compile_model=true \
    --policy.gradient_checkpointing=true \
    --policy.dtype=bfloat16 \
+    --policy.freeze_vision_encoder=false \
+    --policy.train_expert_only=false \
    --steps=3000 \
    --policy.device=cuda \
    --batch_size=32
@@ -79,6 +82,15 @@ python src/lerobot/scripts/lerobot_train.py \
  - [lerobot/pi0_base](https://huggingface.co/lerobot/pi0_base)
  - [lerobot/pi0_libero](https://huggingface.co/lerobot/pi0_libero) (specifically trained on the Libero dataset)

+### Training Parameters Explained
+
+| Parameter               | Default | Description                                 |
+| ----------------------- | ------- | ------------------------------------------- |
+| `freeze_vision_encoder` | `false` | Do not freeze the vision encoder            |
+| `train_expert_only`     | `false` | Do not freeze the VLM, train all parameters |
+
+**💡 Tip**: Setting `train_expert_only=true` freezes the VLM and trains only the action expert and projections, allowing finetuning with reduced memory usage.
+
 ## License

 This model follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).
@@ -36,11 +36,6 @@ This diverse training mixture creates a "curriculum" that enables generalization
   pip install -e ".[pi]"
   ```

-   > [!NOTE]
-   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
-   >
-   > This will be solved in the next patch release
-
 ## Usage

 To use π₀.₅ in your LeRobot configuration, specify the policy type as:
@@ -56,7 +51,7 @@ policy.type=pi05
 Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:

 ```bash
-python src/lerobot/scripts/lerobot_train.py\
+lerobot-train \
    --dataset.repo_id=your_dataset \
    --policy.type=pi05 \
    --output_dir=./outputs/pi05_training \
@@ -67,6 +62,8 @@ python src/lerobot/scripts/lerobot_train.py\
    --policy.gradient_checkpointing=true \
    --wandb.enable=true \
    --policy.dtype=bfloat16 \
+    --policy.freeze_vision_encoder=false \
+    --policy.train_expert_only=false \
    --steps=3000 \
    --policy.device=cuda \
    --batch_size=32
@@ -82,6 +79,15 @@ python src/lerobot/scripts/lerobot_train.py\
  - [lerobot/pi05_base](https://huggingface.co/lerobot/pi05_base)
  - [lerobot/pi05_libero](https://huggingface.co/lerobot/pi05_libero) (specifically trained on the Libero dataset)

+### Training Parameters Explained
+
+| Parameter               | Default | Description                                 |
+| ----------------------- | ------- | ------------------------------------------- |
+| `freeze_vision_encoder` | `false` | Do not freeze the vision encoder            |
+| `train_expert_only`     | `false` | Do not freeze the VLM, train all parameters |
+
+**💡 Tip**: Setting `train_expert_only=true` freezes the VLM and trains only the action expert and projections, allowing finetuning with reduced memory usage.
+
 If your dataset is not converted with `quantiles`, you can convert it with the following command:

 ```bash
@@ -0,0 +1,241 @@
+# π₀-FAST (Pi0-FAST)
+
+π₀-FAST is a **Vision-Language-Action model for general robot control** that uses autoregressive next-token prediction to model continuous robot actions.
+
+## Model Overview
+
+π₀-FAST combines the power of Vision-Language Models with a novel action tokenization approach called **FAST (Frequency-space Action Sequence Tokenization)**. This enables training autoregressive VLAs on highly dexterous tasks that are impossible with standard binning-based discretization, while training **up to 5x faster** than diffusion-based approaches like π₀.
+
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-pifast.png"
+  alt="An overview of Pi0-FAST"
+  width="85%"
+/>
+
+### Why FAST?
+
+Standard approaches for robot action tokenization use simple per-dimension, per-timestep binning schemes. While passable for simple behaviors, this rapidly breaks down for complex and dexterous skills that require precision and high-frequency control.
+
+FAST solves this by compressing action sequences using signal processing techniques, resulting in a dense sequence of action tokens that can be predicted autoregressively—just like language tokens.
+
+### How FAST Tokenization Works
+
+The FAST tokenizer compresses action sequences through the following steps:
+
+1. **Normalize**: Take a continuous action chunk of shape `(H, D)` where `H` is the horizon and `D` is the action dimension. Normalize using one of the supported normalization methods (Quantiles recommended to handle outliers).
+
+2. **Discrete Cosine Transform (DCT)**: Apply DCT (via scipy) to each action dimension separately. DCT is a compression algorithm commonly used in image and audio codecs (JPEG, MP3).
+
+3. **Quantization**: Round and remove insignificant coefficients for each action dimension, producing a sparse frequency matrix.
+
+4. **Flatten**: Flatten the matrix into a 1D vector, with low-frequency components first.
+
+5. **Byte Pair Encoding (BPE)**: Train a BPE tokenizer to compress the DCT coefficients into dense action tokens, typically achieving **10x compression** over prior tokenization approaches.
+
+This approach can transform **any existing VLM** into a VLA by training it to predict these FAST tokens.
+
+## Installation Requirements
+
+1. Install LeRobot by following our [Installation Guide](./installation).
+2. Install π₀-FAST dependencies by running:
+
+   ```bash
+   pip install -e ".[pi]"
+   ```
+
+## Training a Custom FAST Tokenizer
+
+You have two options for the FAST tokenizer:
+
+1. **Use the pre-trained tokenizer**: The `lerobot/fast-action-tokenizer` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.
+
+2. **Train your own tokenizer**: For maximum performance on your specific dataset, you can finetune the tokenizer on your own data.
+
+### Training Your Own Tokenizer
+
+```bash
+lerobot-train-tokenizer \
+    --repo_id "user/my-lerobot-dataset" \
+    --action_horizon 10 \
+    --encoded_dims "0:6" \
+    --vocab_size 1024 \
+    --scale 10.0 \
+    --normalization_mode QUANTILES \
+    --output_dir "./my_fast_tokenizer" \
+    --push_to_hub \
+    --hub_repo_id "username/my-action-tokenizer"
+```
+
+### Key Tokenizer Parameters
+
+| Parameter              | Description                                                                       | Default      |
+| ---------------------- | --------------------------------------------------------------------------------- | ------------ |
+| `--repo_id`            | LeRobot dataset repository ID                                                     | Required     |
+| `--action_horizon`     | Number of future actions in each chunk                                            | `10`         |
+| `--encoded_dims`       | Comma-separated dimension ranges to encode (e.g., `"0:6,7:23"`)                   | `"0:6,7:23"` |
+| `--vocab_size`         | BPE vocabulary size                                                               | `1024`       |
+| `--scale`              | DCT scaling factor for quantization                                               | `10.0`       |
+| `--normalization_mode` | Normalization mode (`MEAN_STD`, `MIN_MAX`, `QUANTILES`, `QUANTILE10`, `IDENTITY`) | `QUANTILES`  |
+| `--sample_fraction`    | Fraction of chunks to sample per episode                                          | `0.1`        |
+
+## Usage
+
+To use π₀-FAST in LeRobot, specify the policy type as:
+
+```python
+policy.type=pi0_fast
+```
+
+## Training
+
+For training π₀-FAST, you can use the LeRobot training script:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=your_dataset \
+    --policy.type=pi0_fast \
+    --output_dir=./outputs/pi0fast_training \
+    --job_name=pi0fast_training \
+    --policy.pretrained_path=lerobot/pi0_fast_base \
+    --policy.dtype=bfloat16 \
+    --policy.gradient_checkpointing=true \
+    --policy.chunk_size=10 \
+    --policy.n_action_steps=10 \
+    --policy.max_action_tokens=256 \
+    --steps=100000 \
+    --batch_size=4 \
+    --policy.device=cuda
+```
+
+### Key Training Parameters
+
+| Parameter                              | Description                                        | Default                         |
+| -------------------------------------- | -------------------------------------------------- | ------------------------------- |
+| `--policy.gradient_checkpointing=true` | Reduces memory usage significantly during training | `false`                         |
+| `--policy.dtype=bfloat16`              | Use mixed precision training for efficiency        | `float32`                       |
+| `--policy.chunk_size`                  | Number of action steps to predict (action horizon) | `50`                            |
+| `--policy.n_action_steps`              | Number of action steps to execute                  | `50`                            |
+| `--policy.max_action_tokens`           | Maximum number of FAST tokens per action chunk     | `256`                           |
+| `--policy.action_tokenizer_name`       | FAST tokenizer to use                              | `lerobot/fast-action-tokenizer` |
+| `--policy.compile_model=true`          | Enable torch.compile for faster training           | `false`                         |
+
+## Inference
+
+### KV-Caching for Fast Inference
+
+π₀-FAST supports **KV-caching**, a widely used optimization in LLM inference. This caches the key-value pairs from the attention mechanism, avoiding redundant computation during autoregressive decoding.
+
+```python
+# KV-caching is enabled by default
+policy.use_kv_cache=true
+```
+
+### Inference Example
+
+```python
+from lerobot.policies.pi0_fast import PI0FastPolicy, PI0FastConfig
+
+# Load the policy
+policy = PI0FastPolicy.from_pretrained("your-model-path")
+
+# During inference
+actions = policy.predict_action_chunk(batch)
+```
+
+## Model Architecture
+
+π₀-FAST uses a PaliGemma-based architecture:
+
+- **Vision Encoder**: SigLIP vision tower for image understanding
+- **Language Model**: Gemma 2B for processing language instructions and predicting action tokens
+
+The model takes images, text instructions, and robot state as input, and outputs discrete FAST tokens that are decoded back to continuous actions.
+
+## Configuration Options
+
+| Parameter            | Description                                     | Default    |
+| -------------------- | ----------------------------------------------- | ---------- |
+| `paligemma_variant`  | VLM backbone variant (`gemma_300m`, `gemma_2b`) | `gemma_2b` |
+| `max_state_dim`      | Maximum state vector dimension (padded)         | `32`       |
+| `max_action_dim`     | Maximum action vector dimension (padded)        | `32`       |
+| `temperature`        | Sampling temperature (0.0 for greedy)           | `0.0`      |
+| `max_decoding_steps` | Maximum decoding steps                          | `256`      |
+| `use_kv_cache`       | Enable KV caching for faster inference          | `true`     |
+
+## Comparison with π₀
+
+| Feature               | π₀                        | π₀-FAST                      |
+| --------------------- | ------------------------- | ---------------------------- |
+| Action Representation | Flow Matching (Diffusion) | Autoregressive Tokens (FAST) |
+| Training Speed        | 1x                        | **5x faster**                |
+| Dexterity             | High                      | High                         |
+| Inference Method      | Iterative Denoising       | Autoregressive Decoding      |
+| KV-Caching            | N/A                       | Supported                    |
+
+## Reproducing π₀Fast results
+
+We reproduce the results of π₀Fast on the LIBERO benchmark using the LeRobot implementation. We take the LeRobot PiFast base model [lerobot/pi0fast-base](https://huggingface.co/lerobot/pi0fast-base) and finetune for an additional 40kk steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
+
+The finetuned model can be found here:
+
+- **π₀Fast LIBERO**: [lerobot/pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero)
+
+With the following training command:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=lerobot/libero \
+  --output_dir=outputs/libero_pi0fast \
+  --job_name=libero_pi0fast \
+  --policy.path=lerobot/pi0fast_base \
+  --policy.dtype=bfloat16 \
+  --steps=100000 \
+  --save_freq=20000 \
+  --batch_size=4 \
+  --policy.device=cuda \
+  --policy.scheduler_warmup_steps=4000 \
+  --policy.scheduler_decay_steps=100000 \
+  --policy.scheduler_decay_lr=1e-5 \
+  --policy.gradient_checkpointing=true \
+  --policy.chunk_size=10 \
+  --policy.n_action_steps=10 \
+  --policy.max_action_tokens=256 \
+  --policy.empty_cameras=1 \
+```
+
+We then evaluate the finetuned model using the LeRobot LIBERO implementation, by running the following command:
+
+```bash
+tasks="libero_object,libero_spatial,libero_goal,libero_10"
+lerobot-eval \
+  --policy.path=lerobot/pi0fast-libero \
+  --policy.max_action_tokens=256 \
+  --env.type=libero \
+  --policy.gradient_checkpointing=false \
+  --env.task=${tasks} \
+  --eval.batch_size=1 \
+  --eval.n_episodes=1 \
+  --rename_map='{"observation.images.image":"observation.images.base_0_rgb","observation.images.image2":"observation.images.left_wrist_0_rgb"}'
+```
+
+**Note:** We set `n_action_steps=10`, similar to the original OpenPI implementation.
+
+### Results
+
+We obtain the following results on the LIBERO benchmark:
+
+| Model       | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average  |
+| ----------- | -------------- | ------------- | ----------- | --------- | -------- |
+| **π₀-fast** | 70.0           | 100.0         | 100.0       | 60.0      | **82.5** |
+
+The full evaluation output folder, including videos, is available [here](https://drive.google.com/drive/folders/1HXpwPTRm4hx6g1sF2P7OOqGG0TwPU7LQ?usp=sharing)
+
+## License
+
+This model follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).
+
+## References
+
+- [FAST: Efficient Robot Action Tokenization](https://www.physicalintelligence.company/research/fast) - Physical Intelligence Blog
+- [OpenPI Repository](https://github.com/Physical-Intelligence/openpi) - Original implementation
+- [FAST Tokenizer on Hugging Face](https://huggingface.co/physical-intelligence/fast) - Pre-trained tokenizer
@@ -0,0 +1,45 @@
+# WALL-OSS
+
+This repository contains the Hugging Face port of [**WALL-OSS**](https://x2robot.com/en/research/68bc2cde8497d7f238dde690), a Vision-Language-Action model for cross-embodiment robotic control based on Qwen2.5-VL with flow matching/FAST action prediction.
+
+---
+
+## Model Overview
+
+| Feature            | Description                                           |
+| ------------------ | ----------------------------------------------------- |
+| Base Model         | Qwen2.5-VL (Vision-Language Model)                    |
+| Action Prediction  | Flow Matching (diffusion) or FAST (discrete tokens)   |
+| Architecture       | Mixture of Experts (MoE) with action-specific routing |
+| Multi-Modal Inputs | Vision (images/videos), Language, Proprioception      |
+
+---
+
+## Additional Resources
+
+Paper: https://arxiv.org/pdf/2509.11766
+
+Official Repository: https://github.com/X-Square-Robot/wall-x
+
+Hugging Face: https://huggingface.co/x-square-robot
+
+---
+
+## Citation
+
+If you use this work, please cite:
+
+```bibtex
+@article{zhai2025igniting,
+    title   = {Igniting VLMs Toward the Embodied Space},
+    author  = {Zhai, Andy and Liu, Brae and Fang, Bruno and Cai, Chalse and Ma, Ellie and Yin, Ethan and Wang, Hao and Zhou, Hugo and Wang, James and Shi, Lights and Liang, Lucy and Wang, Make and Wang, Qian and Gan, Roy and Yu, Ryan and Li, Shalfun and Liu, Starrick and Chen, Sylas and Chen, Vincent and Xu, Zach},
+    journal = {arXiv preprint arXiv:2509.11766},
+    year    = {2025}
+}
+```
+
+---
+
+## License
+
+This model follows the **Apache 2.0 License**, consistent with the original [WallX repository](https://github.com/X-Square-Robot/wall-x).
@@ -30,7 +30,7 @@ Each of these pipelines handle different conversions between different action an

 Below is an example of the three pipelines that we use in the phone to SO-100 follower examples:

-```69:90:examples/phone_so100_record.py
+```python
 phone_to_robot_ee_pose_processor = RobotProcessorPipeline[RobotAction, RobotAction]( # teleop -> dataset action
    steps=[
        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
@@ -84,7 +84,7 @@ Dataset features are determined by the keys saved in the dataset. Each step can

 Below is and example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:

-```src/lerobot/robots/so100_follower/robot_kinematic_processor.py
+```python
    def transform_features(
        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
@@ -103,7 +103,7 @@ Here we declare what PolicyFeatures we modify in this step, so we know what feat

 Below is an example of how we aggregate and merge features in the phone to SO-100 record example:

-```121:145:examples/phone_so100_record.py
+```python
 features=combine_feature_dicts(
        # Run the feature contract of the pipelines
        # This tells you how the features would look like after the pipeline steps
@@ -38,6 +38,7 @@ docker run --rm -it \
  start_rviz:=true start_sdk_server:=true mujoco:=true
 ```

+> [!NOTE]
 > If MuJoCo runs slowly (low simulation frequency), append `-e LD_LIBRARY_PATH="/opt/host-libs:$LD_LIBRARY_PATH" \` to the previous command to improve performance:
 >
 > ```
@@ -141,7 +142,7 @@ If you choose this option but still want to use the VR teleoperation application
 First add reachy2 and reachy2_teleoperator to the imports of the record script. Then you can use the following command:

 ```bash
-python -m lerobot.record \
+lerobot-record \
    --robot.type=reachy2 \
    --robot.ip_address=192.168.0.200 \
    --robot.id=r2-0000 \
@@ -150,6 +151,7 @@ python -m lerobot.record \
    --teleop.type=reachy2_teleoperator \
    --teleop.ip_address=192.168.0.200 \
    --teleop.with_mobile_base=false \
+    --robot.with_torso_camera=true \
    --dataset.repo_id=pollen_robotics/record_test \
    --dataset.single_task="Reachy 2 recording test" \
    --dataset.num_episodes=1 \
@@ -157,6 +159,9 @@ python -m lerobot.record \
    --dataset.fps=15 \
    --dataset.push_to_hub=true \
    --dataset.private=true \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    # --dataset.vcodec=auto \
    --display_data=true
 ```

@@ -165,7 +170,7 @@ python -m lerobot.record \
 **Extended setup overview (all options included):**

 ```bash
-python -m lerobot.record \
+lerobot-record \
    --robot.type=reachy2 \
    --robot.ip_address=192.168.0.200 \
    --robot.use_external_commands=true \
@@ -177,6 +182,8 @@ python -m lerobot.record \
    --robot.with_left_teleop_camera=true \
    --robot.with_right_teleop_camera=true \
    --robot.with_torso_camera=false \
+    --robot.camera_width=640 \
+    --robot.camera_height=480 \
    --robot.disable_torque_on_disconnect=false \
    --robot.max_relative_target=5.0 \
    --teleop.type=reachy2_teleoperator \
@@ -194,6 +201,9 @@ python -m lerobot.record \
    --dataset.fps=15 \
    --dataset.push_to_hub=true \
    --dataset.private=true \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    # --dataset.vcodec=auto \
    --display_data=true
 ```

@@ -212,9 +222,10 @@ Must be set to true if a compliant Reachy 2 is used to control another one.
 From our initial tests, recording **all** joints when only some are moving can reduce model quality with certain policies.
 To avoid this, you can exclude specific parts from recording and replay using:

-````
+```bash
 --robot.with_<part>=false
-```,
+```
+
 with `<part>` being one of : `mobile_base`, `l_arm`, `r_arm", `neck`, `antennas`.
 It determine whether the corresponding part is recorded in the observations. True if not set.

@@ -222,49 +233,60 @@ By default, **all parts are recorded**.

 The same per-part mechanism is available in `reachy2_teleoperator` as well.

-````
-
+```bash
 --teleop.with\_<part>
-
 ```
+
 with `<part>` being one of : `mobile_base`, `l_arm`, `r_arm", `neck`, `antennas`.
 Determine whether the corresponding part is recorded in the actions. True if not set.

 > **Important:** In a given session, the **enabled parts must match** on both the robot and the teleoperator.
-For example, if the robot runs with `--robot.with_mobile_base=false`, the teleoperator must disable the same part `--teleoperator.with_mobile_base=false`.
+> For example, if the robot runs with `--robot.with_mobile_base=false`, the teleoperator must disable the same part `--teleoperator.with_mobile_base=false`.

 ##### Use the relevant cameras

-You can do the same for **cameras**. By default, only the **teleoperation cameras** are recorded (both `left_teleop_camera` and `right_teleop_camera`). Enable or disable each camera with:
+You can do the same for **cameras**. Enable or disable each camera with default parameters using:

+```bash
+--robot.with_left_teleop_camera=<true|false> \
+--robot.with_right_teleop_camera=<true|false> \
+--robot.with_torso_camera=<true|false>
 ```

--robot.with_left_teleop_camera=<true|false>
--robot.with_right_teleop_camera=<true|false>
--robot.with_torso_camera=<true|false>
+By default, no camera is recorded, all camera arguments are set to `false`.
+If you want to, you can use custom `width` and `height` parameters for Reachy 2's cameras using the `--robot.camera_width` & `--robot.camera_height` argument:

-````
+```bash
+--robot.camera_width=1920 \
+--robot.camera_height=1080
+```

+This will change the resolution of all 3 default robot cameras (enabled by the above bool arguments).
+
+If you want, you can add additional cameras other than the ones in the robot as usual with:
+
+```bash
+--robot.cameras="{ extra: {type: opencv, index_or_path: 42, width: 640, height: 480, fps: 30}}" \
+```

 ## Step 2: Replay

 Make sure the robot is configured with the same parts as the dataset:

 ```bash
-python -m lerobot.replay \
+lerobot-replay \
    --robot.type=reachy2 \
    --robot.ip_address=192.168.0.200 \
    --robot.use_external_commands=false \
    --robot.with_mobile_base=false \
    --dataset.repo_id=pollen_robotics/record_test \
    --dataset.episode=0
-    --display_data=true
-````
+```

 ## Step 3: Train

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --dataset.repo_id=pollen_robotics/record_test \
  --policy.type=act \
  --output_dir=outputs/train/reachy2_test \
@@ -277,10 +299,9 @@ python -m lerobot.scripts.train \
 ## Step 4: Evaluate

 ```bash
-python -m lerobot.record \
+lerobot-eval \
  --robot.type=reachy2 \
  --robot.ip_address=192.168.0.200 \
-  --display_data=false \
  --dataset.repo_id=pollen_robotics/eval_record_test \
  --dataset.single_task="Evaluate reachy2 policy" \
  --dataset.num_episodes=10 \
@@ -4,6 +4,12 @@ SARM (Stage-Aware Reward Modeling) is a video-based reward modeling framework fo

 **Paper**: [SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation](https://arxiv.org/abs/2509.25358)

+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-sarm.png"
+  alt="An overview of SARM"
+  width="80%"
+/>
+
 ## Why Reward Models?

 Standard behavior cloning treats all demonstration frames equally, but real-world robot datasets are messy. They contain hesitations, corrections, and variable-quality trajectories. Reward models solve this by learning a generalizable notion of **task progress** from demonstrations: given video frames and a task description, they predict how close the robot is to completing the task (0→1). This learned "progress signal" can be used in multiple ways, two promising applications are: (1) **weighted imitation learning** (RA-BC), where high-progress frames receive more weight during policy training, and (2) **reinforcement learning**, where the reward model provides dense rewards for online or offline policy improvement.
@@ -263,7 +269,7 @@ This generates visualizations showing video frames with subtask boundaries overl
 Train with **no annotations** - uses linear progress from 0 to 1:

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=single_stage \
@@ -282,7 +288,7 @@ python src/lerobot/scripts/lerobot_train.py \
 Train with **dense annotations only** (sparse auto-generated):

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=dense_only \
@@ -301,7 +307,7 @@ python src/lerobot/scripts/lerobot_train.py \
 Train with **both sparse and dense annotations**:

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=sarm \
  --policy.annotation_mode=dual \
@@ -462,7 +468,7 @@ This script:
 Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`). Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:

 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
  --dataset.repo_id=your-username/your-dataset \
  --policy.type=pi0 \
  --use_rabc=true \
@@ -106,6 +106,9 @@ lerobot-record \
  --dataset.repo_id=${HF_USER}/eval_DATASET_NAME_test \  # <- This will be the dataset name on HF Hub
  --dataset.episode_time_s=50 \
  --dataset.num_episodes=10 \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
  # <- Teleop optional if you want to teleoperate in between episodes \
  # --teleop.type=so100_leader \
  # --teleop.port=/dev/ttyACM0 \
@@ -103,7 +103,7 @@ lerobot-setup-motors \

 <!-- prettier-ignore-start -->
 ```python
-from lerobot.robots.so100_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig

 config = SO100FollowerConfig(
    port="/dev/tty.usbmodem585A0076841",
@@ -177,7 +177,7 @@ lerobot-setup-motors \

 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig

 config = SO100LeaderConfig(
    port="/dev/tty.usbmodem585A0076841",
@@ -579,7 +579,7 @@ lerobot-calibrate \

 <!-- prettier-ignore-start -->
 ```python
-from lerobot.robots.so100_follower import SO100FollowerConfig, SO100Follower
+from lerobot.robots.so_follower import SO100FollowerConfig, SO100Follower

 config = SO100FollowerConfig(
    port="/dev/tty.usbmodem585A0076891",
@@ -617,7 +617,7 @@ lerobot-calibrate \

 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so100_leader import SO100LeaderConfig, SO100Leader
+from lerobot.teleoperators.so_leader import SO100LeaderConfig, SO100Leader

 config = SO100LeaderConfig(
    port="/dev/tty.usbmodem58760431551",
@@ -1,5 +1,18 @@
 # SO-101

+<div style="display: flex; align-items: center; gap: 10px;">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Follower.webp"
+    alt="SO-101"
+    width="60%"
+  />
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Leader.webp"
+    alt="SO-101"
+    width="60%"
+  />
+</div>
+
 In the steps below, we explain how to assemble our flagship robot, the SO-101.

 ## Source the parts
@@ -125,7 +138,7 @@ lerobot-setup-motors \

 <!-- prettier-ignore-start -->
 ```python
-from lerobot.robots.so101_follower import SO101Follower, SO101FollowerConfig
+from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig

 config = SO101FollowerConfig(
    port="/dev/tty.usbmodem585A0076841",
@@ -201,7 +214,7 @@ lerobot-setup-motors \

 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so101_leader import SO101Leader, SO101LeaderConfig
+from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig

 config = SO101LeaderConfig(
    port="/dev/tty.usbmodem585A0076841",
@@ -364,7 +377,7 @@ lerobot-calibrate \

 <!-- prettier-ignore-start -->
 ```python
-from lerobot.robots.so101_follower import SO101FollowerConfig, SO101Follower
+from lerobot.robots.so_follower import SO101FollowerConfig, SO101Follower

 config = SO101FollowerConfig(
    port="/dev/tty.usbmodem585A0076891",
@@ -413,7 +426,7 @@ lerobot-calibrate \

 <!-- prettier-ignore-start -->
 ```python
-from lerobot.teleoperators.so101_leader import SO101LeaderConfig, SO101Leader
+from lerobot.teleoperators.so_leader import SO101LeaderConfig, SO101Leader

 config = SO101LeaderConfig(
    port="/dev/tty.usbmodem58760431551",
@@ -0,0 +1,155 @@
+# Streaming Video Encoding Guide
+
+## 1. Overview
+
+Streaming video encoding eliminates the traditional PNG round-trip during video dataset recording. Instead of:
+
+1. Capture frame -> write PNG to disk -> (at episode end) read PNG's -> encode to MP4 -> delete PNG's
+
+Frames can be encoded in real-time during capture:
+
+1. Capture frame -> queue to encoder thread -> encode to MP4 directly
+
+This makes `save_episode()` near-instant (the video is already encoded by the time the episode ends) and removes the blocking wait that previously occurred between episodes, especially with multiple cameras in long episodes.
+
+## 2. Tuning Parameters
+
+| Parameter               | CLI Flag                          | Type          | Default       | Description                                                       |
+| ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
+| `streaming_encoding`    | `--dataset.streaming_encoding`    | `bool`        | `True`        | Enable real-time encoding during capture                          |
+| `vcodec`                | `--dataset.vcodec`                | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
+| `encoder_threads`       | `--dataset.encoder_threads`       | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
+| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `60`          | Max buffered frames per camera (~2s at 30fps). Consumes RAM       |
+
+## 3. Performance Considerations
+
+Streaming encoding means the CPU is encoding video **during** the capture loop, not after. This creates a CPU budget that must be shared between:
+
+- **Control loop** (reading cameras, control the robot, writing non-video data)
+- **Encoder threads** (one pool per camera)
+- **Rerun visualization** (if enabled)
+- **OS and other processes**
+
+### Resolution & Number of Cameras Impact
+
+| Setup                     | Throughput (px/sec) | CPU Encoding Load | Notes                          |
+| ------------------------- | ------------------- | ----------------- | ------------------------------ |
+| 2camsx 640x480x3 @30fps   | 55M                 | Low               | Works on most systems          |
+| 2camsx 1280x720x3 @30fps  | 165M                | Moderate          | Comfortable on modern systems  |
+| 2camsx 1920x1080x3 @30fps | 373M                | High              | Requires powerful high-end CPU |
+
+### `encoder_threads` Tuning
+
+This parameter controls how many threads each encoder instance uses internally:
+
+- **Higher values** (e.g., 4-5): Faster encoding, but uses more CPU cores per camera. Good for high-end systems with many cores.
+- **Lower values** (e.g., 1-2): Less CPU per camera, freeing cores for capture and visualization. Good for low-res images and capable CPUs.
+- **`None` (default)**: Lets the codec decide. Information available in the codec logs.
+
+### Backpressure and Frame Dropping
+
+Each camera has a bounded queue (`encoder_queue_maxsize`, default 60 frames). When the encoder can't keep up:
+
+1. The queue fills up (consuming RAM)
+2. New frames are **dropped** (not blocked) — the capture loop continues uninterrupted
+3. A warning is logged: `"Encoder queue full for {camera}, dropped N frame(s)"`
+4. At episode end, total dropped frames per camera are reported
+
+### Symptoms of Encoder Falling Behind
+
+- **System feels laggy and freezes**: all CPUs are at 100%
+- **Dropped frame warnings** in the log or lower frames/FPS than expected in the recorded dataset
+- **Choppy robot movement**: If CPU is severely overloaded, even the capture loop may be affected
+- **Accumulated rerun lag**: Visualization falls behind real-time
+
+## 4. Hardware-Accelerated Encoding
+
+### When to Use
+
+Use HW encoding when:
+
+- CPU is the bottleneck (dropped frames, choppy robot, rerun lag)
+- You have compatible hardware (GPU or dedicated encoder)
+- You're recording at high throughput (high resolution or with many cameras)
+
+### Choosing a Codec
+
+| Codec                 | CPU Usage | File Size      | Quality | Notes                                                            |
+| --------------------- | --------- | -------------- | ------- | ---------------------------------------------------------------- |
+| `libsvtav1` (default) | High      | Smallest       | Best    | Default. Best compression but most CPU-intensive                 |
+| `h264`                | Medium    | ~30-50% larger | Good    | Software H.264. Lower CPU                                        |
+| HW encoders           | Very Low  | Largest        | Good    | Offloads to dedicated hardware. Best for CPU-constrained systems |
+
+### Available HW Encoders
+
+| Encoder             | Platform      | Hardware                                                                                         | CLI Value                            |
+| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------ |
+| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=h264_videotoolbox` |
+| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=hevc_videotoolbox` |
+| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=h264_nvenc`        |
+| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=hevc_nvenc`        |
+| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.vcodec=h264_vaapi`        |
+| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.vcodec=h264_qsv`          |
+| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.vcodec=auto`              |
+
+> [!NOTE]
+> In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
+
+> [!NOTE]
+> `libsvtav1` is the default because it provides the best training performance; other vcodecs can reduce CPU usage and be faster, but they typically produce larger files and may affect training time.
+
+## 5. Troubleshooting
+
+| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                  |
+| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.vcodec=auto`) |
+| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.vcodec=auto`).                                                                                                                                                    |
+| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                     |
+| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                              |
+| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                              |
+| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.vcodec=auto`                                                                                                                                                                                                                             |
+| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                   |
+
+## 6. Recommended Configurations
+
+These estimates are conservative; we recommend testing them on your setup—start with a low load and increase it gradually.
+
+### High-End Systems: modern 12+ cores (24+ threads)
+
+A throughput between ~250-500M px/sec should be comfortable in CPU. For even better results try HW encoding if available.
+
+```bash
+# 3camsx 1280x720x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
+# 2camsx 1920x1080x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
+lerobot-record --dataset.encoder_threads=5 ...
+
+# 3camsx 1920x1080x3 @30fps: Might require some tuning.
+```
+
+### Mid-Range Systems: modern 8+ cores (16+ threads) or Apple Silicon
+
+A throughput between ~80-300M px/sec should be possible in CPU.
+
+```bash
+# 3camsx 640x480x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
+# 2camsx 1280x720x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
+lerobot-record --dataset.encoder_threads=2 ...
+
+# 2camsx 1920x1080x3 @30fps: Might require some tuning.
+```
+
+### Low-Resource Systems: modern 4+ cores (8+ threads) or Raspberry Pi 5
+
+On very constrained systems, streaming encoding may compete too heavily with the capture loop. Disabling it falls back to the PNG-based approach where encoding happens between episodes (blocking, but doesn't interfere with capture). Alternatively, record at a lower throughput to reduce both capture and encoding load. Consider also changing codec to `h264` and using batch encoding.
+
+```bash
+# 2camsx 640x480x3 @30fps: Requires some tuning.
+
+# Use H.264, disable streaming, consider batching encoding
+lerobot-record --dataset.vcodec=h264 --dataset.streaming_encoding=false ...
+```
+
+## 7. Closing note
+
+Performance ultimately depends on your exact setup — frames-per-second, resolution, CPU cores and load, available memory, episode length, and the encoder you choose. Always test with your target workload, be mindful about your CPU & system capabilities and tune `encoder_threads`, `encoder_queue_maxsize`, and
+`vcodec` reasonably. That said, a common practical configuration (for many applications) is three cameras at 640×480x3 @30fps; this usually runs fine with the default streaming video encoding settings in modern systems. Always verify your recorded dataset is healthy by comparing the video duration to the CLI episode duration and confirming the row count equals FPS × CLI duration.
@@ -1,23 +1,72 @@
-# Unitree G1 Robot Setup and Control
+# Unitree G1

-This guide covers the complete setup process for the Unitree G1 humanoid, from initial connection to running gr00t_wbc locomotion.
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/unitree_thumbnail.jpg"
+  alt="Unitree G1 locomanipulation demo"
+  style={{ width: "100%" }}
+/>

-## About the Unitree G1
-
-We offer support for both 29 and 23 DOF G1. We introduce:
-
- **`unitree g1` robot class, handling low level communication with the humanoid**
- **ZMQ socket bridge** for remote communication over WiFi, allowing one to deploy policies remotely instead of over ethernet or directly on the Orin
- **GR00T locomotion policy** for bipedal walking and balance
- **MuJoCo simulation mode** for testing policies without the physical robot
+The Unitree G1 humanoid is now supported in LeRobot! You can teleoperate, train locomanipulation policies, test in sim, and more. Both 29 and 23 DoF variants are supported.

 ---

-## Part 1: Connect to Robot over Ethernet
+## Part 1: Getting Started

-### Step 1: Configure Your Computer's Ethernet Interface
+### Install the Unitree SDK

-Set a static IP on the same subnet as the robot:
+Follow the [unitree_sdk2_python installation guide](https://github.com/unitreerobotics/unitree_sdk2_python#installation). Tested with `unitree_sdk2py==1.0.1` and `cyclonedds==0.10.2`:
+
+```bash
+conda create -y -n lerobot python=3.12
+conda activate lerobot
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python
+pip install -e .
+cd ..
+```
+
+### Install LeRobot
+
+```bash
+conda install ffmpeg -c conda-forge
+conda install -c conda-forge "pinocchio>=3.0.0,<4.0.0"
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+```
+
+<Tip>
+  For now, pinocchio must be installed from conda-forge (not pip) to include the
+  CasADi bindings needed for arm IK.
+</Tip>
+
+### Test the Installation (Simulation)
+
+The simulation environment has its own dependencies. Check the Simulation environment dependencies: [Unitree G1 Mujoco EnvHub](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main).
+
+```bash
+pip install mujoco loguru msgpack msgpack-numpy
+```
+
+```bash
+lerobot-teleoperate \
+  --robot.type=unitree_g1 \
+  --robot.is_simulation=true \
+  --teleop.type=unitree_g1 \
+  --teleop.id=wbc_unitree \
+  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30, "warmup_s": 5}}' \
+  --display_data=true \
+  --robot.controller=GrootLocomotionController
+```
+
+This will launch a [MuJoCo sim instance](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main) for the G1. You can connect a gamepad to your machine before launching in order to control the robot's locomotion in sim. We support both [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) and [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) via `--robot.controller`.
+
+- Press `9` to release the robot
+- Press `7` / `8` to increase / decrease waist height
+
+### Connect to the Physical Robot
+
+The G1's Ethernet IP is fixed at `192.168.123.164`. Your machine must have a static IP on the same subnet: `192.168.123.x` where `x ≠ 164`.

 ```bash
 # Replace 'enp131s0' with your ethernet interface name (check with `ip a`)
@@ -26,183 +75,228 @@ sudo ip addr add 192.168.123.200/24 dev enp131s0
 sudo ip link set enp131s0 up
 ```

-**Note**: The robot's Ethernet IP is fixed at `192.168.123.164`. Your computer must use `192.168.123.x` where x ≠ 164.
-
-### Step 2: SSH into the Robot
+### SSH into the Robot

 ```bash
 ssh unitree@192.168.123.164
 # Password: 123
 ```

-You should now be connected to the robot's onboard computer.
+### Share Internet via Ethernet

---
-
-## Part 2: Enable WiFi on the Robot
-
-Once connected via Ethernet, follow these steps to enable WiFi:
-
-### Step 1: Enable WiFi Hardware
-
-```bash
-# Unblock WiFi radio
-sudo rfkill unblock wifi
-sudo rfkill unblock all
-
-# Bring up WiFi interface
-sudo ip link set wlan0 up
-
-# Enable NetworkManager control
-sudo nmcli radio wifi on
-sudo nmcli device set wlan0 managed yes
-sudo systemctl restart NetworkManager
-```
-
-### Step 2: Enable Internet Forwarding
+The G1 needs internet access to clone repos and install packages. Share your laptop's connection over Ethernet:

 **On your laptop:**

 ```bash
-# Enable IP forwarding
 sudo sysctl -w net.ipv4.ip_forward=1

-# Set up NAT (replace wlp132s0f0 with your WiFi interface)
+# Replace wlp132s0f0 with your WiFi interface name
 sudo iptables -t nat -A POSTROUTING -o wlp132s0f0 -s 192.168.123.0/24 -j MASQUERADE
 sudo iptables -A FORWARD -i wlp132s0f0 -o enp131s0 -m state --state RELATED,ESTABLISHED -j ACCEPT
 sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
 ```

-**On the robot:**
+**On the G1:**

 ```bash
-# Add laptop as default gateway
 sudo ip route del default 2>/dev/null || true
 sudo ip route add default via 192.168.123.200 dev eth0
 echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf

-# Test connection
+# Verify
 ping -c 3 8.8.8.8
 ```

-### Step 3: Connect to WiFi Network
+### Install the Unitree SDK on the G1
+
+Follow the [unitree_sdk2_python installation guide](https://github.com/unitreerobotics/unitree_sdk2_python#installation):
+
+```bash
+conda create -y -n lerobot python=3.12
+conda activate lerobot
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python
+python -m pip install -e .
+cd ..
+```
+
+### Install LeRobot on the G1
+
+```bash
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+conda install -c conda-forge "pinocchio>=3.0.0,<4.0.0"
+python -m pip install -e '.[unitree_g1]'
+```
+
+<Tip>
+  For now, pinocchio must be installed from conda-forge (not pip) to include the
+  CasADi bindings needed for arm IK.
+</Tip>
+
+### (Optional) Enable WiFi on the Robot
+
+For wireless SSH access, you can enable WiFi on the G1 (it's blocked by default):
+
+```bash
+sudo rfkill unblock all
+sudo ip link set wlan0 up
+sudo nmcli radio wifi on
+sudo nmcli device set wlan0 managed yes
+sudo systemctl restart NetworkManager
+```
+
+**Connect to a WiFi network:**

 ```bash
-# List available networks
 nmcli device wifi list

-# Connect to your WiFi (example)
 sudo nmcli connection add type wifi ifname wlan0 con-name "YourNetwork" ssid "YourNetwork"
 sudo nmcli connection modify "YourNetwork" wifi-sec.key-mgmt wpa-psk
 sudo nmcli connection modify "YourNetwork" wifi-sec.psk "YourPassword"
 sudo nmcli connection modify "YourNetwork" connection.autoconnect yes
 sudo nmcli connection up "YourNetwork"

-# Check WiFi IP address
 ip a show wlan0
 ```

-### Step 4: SSH Over WiFi
-
-Once connected to WiFi, note the robot's IP address and disconnect the Ethernet cable. You can now SSH over WiFi:
+You can then SSH over WiFi instead of Ethernet:

 ```bash
-ssh unitree@<YOUR_ROBOT_IP>
+ssh unitree@<ROBOT_WIFI_IP>
 # Password: 123
 ```

-Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address (e.g., `172.18.129.215`).
+---
+
+## Part 2: Teleoperation & Locomotion
+
+### Run the Robot Server
+
+On the robot (from `~/lerobot`):
+
+```bash
+cd ~/lerobot
+python src/lerobot/robots/unitree_g1/run_g1_server.py --camera
+```
+
+### Run the Locomotion Policy
+
+You can run the teleoperation client from your laptop over Ethernet, over WiFi (experimental), or directly on the robot itself. Mind potential latency introduced by your network.
+
+**From your laptop:**
+
+```bash
+lerobot-teleoperate \
+  --robot.type=unitree_g1 \
+  --robot.is_simulation=false \
+  --robot.robot_ip=<ROBOT_IP> \
+  --teleop.type=unitree_g1 \
+  --teleop.id=wbc_unitree \
+  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+  --display_data=true \
+  --robot.controller=HolosomaLocomotionController
+```
+
+We support both [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) and [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) via `--robot.controller`.

 ---

-## Part 3: Robot Server Setup
+## Part 3: Loco-Manipulation with the Homunculus Exoskeleton

-### Step 1: Install LeRobot on the Orin
+We provide a loco-manipulation solution via the Homunculus Exoskeleton — an open-source 7 DoF exoskeleton for whole-body control. Check it out [here](https://github.com/nepyope/hmc_exo).

-SSH into the robot and install LeRobot:
+### Calibrate

 ```bash
-ssh unitree@<YOUR_ROBOT_IP>
-
-conda create -y -n lerobot python=3.10
-conda activate lerobot
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-pip install -e '.[unitree_g1]'
-git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
-cd unitree_sdk2_python  && pip install -e .
+lerobot-calibrate \
+  --teleop.type=unitree_g1 \
+  --teleop.left_arm_config.port=/dev/ttyACM1 \
+  --teleop.right_arm_config.port=/dev/ttyACM0 \
+  --teleop.id=exo
 ```

-**Note**: The Unitree SDK requires CycloneDDS v0.10.2 to be installed. See the [Unitree SDK documentation](https://github.com/unitreerobotics/unitree_sdk2_python) for details.
+During calibration move each joint through its entire range. After fitting, move the joint in a neutral position and press `n` to advance.

-### Step 2: Run the Robot Server
-
-On the robot:
+### Record a Dataset

 ```bash
-python src/lerobot/robots/unitree_g1/run_g1_server.py
+lerobot-record \
+  --robot.type=unitree_g1 \
+  --robot.is_simulation=true \
+  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+  --teleop.type=unitree_g1 \
+  --teleop.left_arm_config.port=/dev/ttyACM1 \
+  --teleop.right_arm_config.port=/dev/ttyACM0 \
+  --teleop.id=exo \
+  --dataset.repo_id=your-username/dataset-name \
+  --dataset.single_task="Test" \
+  --dataset.num_episodes=2 \
+  --dataset.episode_time_s=5 \
+  --dataset.reset_time_s=5 \
+  --dataset.push_to_hub=true \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2
 ```

-**Important**: Keep this terminal running. The server must be active for remote control.
+> **Note:** Omit `--teleop.left_arm_config.port` and `--teleop.right_arm_config.port` if you're only using the joystick.
+
+Example dataset: [nepyope/unitree_box_move_blue_full](https://huggingface.co/datasets/nepyope/unitree_box_move_blue_full)

 ---

-## Part 4: Running GR00T Locomotion
+## Part 4: Training & Inference

-With the robot server running, you can now control the robot from your laptop.
-
-### Step 1: Install LeRobot on your machine
+### Train

 ```bash
-conda create -y -n lerobot python=3.10
-conda activate lerobot
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-pip install -e '.[unitree_g1]'
-git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
-cd unitree_sdk2_python  && pip install -e .
+python src/lerobot/scripts/lerobot_train.py \
+  --dataset.repo_id=your-username/dataset-name  \
+  --policy.type=pi05 \
+  --output_dir=./outputs/pi05_training \
+  --job_name=pi05_training \
+  --policy.repo_id=your-username/your-repo-id \
+  --policy.pretrained_path=lerobot/pi05_base \
+  --policy.compile_model=true \
+  --policy.gradient_checkpointing=true \
+  --wandb.enable=true \
+  --policy.dtype=bfloat16 \
+  --policy.freeze_vision_encoder=false \
+  --policy.train_expert_only=false \
+  --steps=3000 \
+  --policy.device=cuda \
+  --batch_size=32
 ```

-### Step 2: Update Robot IP in Config
+### Inference with RTC

-Edit the config file to match your robot's WiFi IP:
-
-```python
-# In src/lerobot/robots/unitree_g1/config_unitree_g1.py
-robot_ip: str = "<YOUR_ROBOT_IP>"  # Replace with your robot's WiFi IP.
-```
-
-**Note**: When running directly on the G1 (not remotely), set `robot_ip: str = "127.0.0.1"` instead.
-
-### Step 3: Run the Locomotion Policy
+Once trained, we recommend deploying policies using inference-time RTC:

 ```bash
-# Run GR00T locomotion controller
-python examples/unitree_g1/gr00t_locomotion.py --repo-id "nepyope/GR00T-WholeBodyControl_g1"
+python examples/rtc/eval_with_real_robot.py \
+  --policy.path=your-username/your-repo-id \
+  --policy.device=cuda \
+  --robot.type=unitree_g1 \
+  --robot.is_simulation=false \
+  --robot.controller=HolosomaLocomotionController \
+  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+  --task="task_description" \
+  --duration=1000 \
+  --fps=30 \
+  --rtc.enabled=true
 ```

-### Step 4: Control with Remote
-
- **Left stick**: Forward/backward and left/right movement
- **Right stick**: Rotation
- **R1 button**: Raise waist height
- **R2 button**: Lower waist height
-
-Press `Ctrl+C` to stop the policy.
-
 ---

-## Extra: Running in Simulation Mode (MuJoCo)
-
-You can now test and develop policies without a physical robot using MuJoCo. to do so set `is_simulation=True` in config.
-
 ## Additional Resources

 - [Unitree SDK Documentation](https://github.com/unitreerobotics/unitree_sdk2_python)
- [GR00T Policy Repository](https://huggingface.co/nepyope/GR00T-WholeBodyControl_g1)
+- [GR00T-WholeBodyControl](https://github.com/NVlabs/GR00T-WholeBodyControl)
+- [Holosoma](https://github.com/amazon-far/holosoma)
 - [LeRobot Documentation](https://github.com/huggingface/lerobot)
- [Unitree_IL_Lerobot](https://github.com/unitreerobotics/unitree_IL_lerobot)
+- [Unitree IL LeRobot](https://github.com/unitreerobotics/unitree_IL_lerobot)

 ---

-_Last updated: December 2025_
+_Last updated: March 2026_
@@ -12,6 +12,7 @@ LeRobot provides several utilities for manipulating datasets:
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
 6. **Convert to Video** - Convert image-based datasets to video format for efficient storage
+7. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.

 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
@@ -95,26 +96,26 @@ Convert an image-based dataset to video format, creating a new LeRobotDataset wh
 # Local-only: Save to a custom output directory (no hub push)
 lerobot-edit-dataset \
    --repo_id lerobot/pusht_image \
-    --operation.type convert_to_video \
+    --operation.type convert_image_to_video \
    --operation.output_dir /path/to/output/pusht_video

 # Save with new repo_id (local storage)
 lerobot-edit-dataset \
    --repo_id lerobot/pusht_image \
    --new_repo_id lerobot/pusht_video \
-    --operation.type convert_to_video
+    --operation.type convert_image_to_video

 # Convert and push to Hugging Face Hub
 lerobot-edit-dataset \
    --repo_id lerobot/pusht_image \
    --new_repo_id lerobot/pusht_video \
-    --operation.type convert_to_video \
+    --operation.type convert_image_to_video \
    --push_to_hub true

 # Convert with custom video codec and quality settings
 lerobot-edit-dataset \
    --repo_id lerobot/pusht_image \
-    --operation.type convert_to_video \
+    --operation.type convert_image_to_video \
    --operation.output_dir outputs/pusht_video \
    --operation.vcodec libsvtav1 \
    --operation.pix_fmt yuv420p \
@@ -124,16 +125,23 @@ lerobot-edit-dataset \
 # Convert only specific episodes
 lerobot-edit-dataset \
    --repo_id lerobot/pusht_image \
-    --operation.type convert_to_video \
+    --operation.type convert_image_to_video \
    --operation.output_dir outputs/pusht_video \
    --operation.episode_indices "[0, 1, 2, 5, 10]"

 # Convert with multiple workers for parallel processing
 lerobot-edit-dataset \
    --repo_id lerobot/pusht_image \
-    --operation.type convert_to_video \
+    --operation.type convert_image_to_video \
    --operation.output_dir outputs/pusht_video \
    --operation.num_workers 8
+
+# For memory-constrained systems, users can now specify limits:
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_to_video \
+    --operation.max_episodes_per_batch 50 \
+    --operation.max_frames_per_batch 10000
 ```

 **Parameters:**
@@ -149,6 +157,30 @@ lerobot-edit-dataset \

 **Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved.

+### Show the information of datasets
+
+Show the information of datasets such as number of episode, number of frame, File size and so on.
+No change will be made to the dataset
+
+```bash
+
+# Show dataset information without feature details
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type info \
+
+# Show dataset information with feature details
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type info \
+    --operation.show_features true
+
+```
+
+**Parameters:**
+
+- `parameters`: The flag to control show or no show dataset information with feature details.(default=false)
+
 ### Push to Hub

 Add the `--push_to_hub true` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:
@@ -0,0 +1,80 @@
+# WALL-OSS
+
+WALL-OSS is an open-source foundation model for embodied intelligence, proposed by the [XSquare Robot](https://x2robot.com/en/research/68bc2cde8497d7f238dde690) team in 2025. The LeRobot implementation is adapted from their open-source [WallX](https://github.com/X-Square-Robot/wall-x) repository.
+
+X Square Robot’s WALL-OSS is now integrated into Hugging Face’s LeRobot ecosystem. This is an exciting collaborative project between the LeRobot and X Square Robot teams. You can now post-train, evaluate, and deploy WALL-OSS directly through LeRobot. With this, we’re aiming to make it easier for the open-source robotics community to customize and deploy WALL-OSS foundation models. Read and explore WALL-OSS [paper](https://arxiv.org/pdf/2509.11766) and [code](https://github.com/X-Square-Robot/wall-x).
+
+## Model Overview
+
+The WALL-OSS team is building the embodied foundation model to capture and compress the world's most valuable data: the continuous, high-fidelity stream of physical interaction. By creating a direct feedback loop between the model's decisions and the body's lived experience, the emergence of a truly generalizable intelligence is enabled—one that understands not just how the world works, but how to act effectively within it.
+
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/walloss-lerobot-paper.png"
+  alt="An overview of WALL-OSS"
+  width="85%"
+/>
+
+Technically, WALL-OSS introduces a tightly coupled multimodal architecture (tightly-coupled MoE structure) that integrates both discrete and continuous action modeling strategies. Through a two-stage training pipeline (Inspiration → Integration), the model gradually unifies semantic reasoning and high-frequency action generation. Its core innovations include:
+
+- **Embodied perception–enhanced multimodal pretraining**: Large-scale training on unified vision–language–action data to strengthen spatial, causal, and manipulation understanding.
+- **Unified Cross-Level Chain-of-Thought (Uni-CoT)**: A single differentiable framework that unifies high-level instruction reasoning, sub-task decomposition, and fine-grained action synthesis, forming a continuous chain from “understanding” to “execution.”
+- **Mixture-of-Experts (MoE) action heads**: Dynamically activating experts depending on the task phase and modeling actions in discrete or continuous space to maintain stable VLM priors.
+- **Two-stage training paradigm**:
+  - **Inspiration stage**: Injecting discrete action priors to strengthen spatial understanding and semantic-action alignment.
+  - **Integration stage**: Using flow matching to achieve high-frequency continuous control.
+
+## Installation Requirements
+
+1. Install LeRobot by following our [Installation Guide](./installation).
+2. Install WallX dependencies by running:
+
+   ```bash
+   pip install -e ".[wallx]"
+   ```
+
+## Usage
+
+To use WallX in LeRobot, specify the policy type as:
+
+```python
+policy.type=wall_x
+```
+
+## Training
+
+For training WallX, you can use the standard LeRobot training script with the appropriate configuration:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=your_dataset \
+    --policy.type=wall_x \
+    --output_dir=./outputs/wallx_training \
+    --job_name=wallx_training \
+    --policy.repo_id=your_repo_id \
+    --policy.pretrained_name_or_path=x-square-robot/wall-oss-flow \
+    --policy.prediction_mode=diffusion \
+    --policy.attn_implementation=eager \
+    --steps=3000 \
+    --policy.device=cuda \
+    --batch_size=32
+```
+
+### Training Arguments
+
+| Argument                       | Description                                                                                                                                                   |
+| ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `--dataset.repo_id`            | The Hugging Face Hub repository ID for your training dataset (e.g., `lerobot/aloha_sim_insertion_human`)                                                      |
+| `--policy.type`                | Specifies using the WallX policy architecture                                                                                                                 |
+| `--output_dir`                 | Local directory where training checkpoints and logs will be saved                                                                                             |
+| `--job_name`                   | A name identifier for this training run (used in logging/tracking)                                                                                            |
+| `--policy.repo_id`             | Your Hugging Face Hub repo ID where the trained model will be pushed                                                                                          |
+| `--policy.pretrained_path`     | Path to pretrained WallX weights to initialize from (the official WALL-OSS checkpoint)                                                                        |
+| `--policy.prediction_mode`     | The action prediction strategy: `diffusion` or `fast` - `diffusion` uses iterative denoising for action generation, `fast` uses next token prediction instead |
+| `--policy.attn_implementation` | Attention implementation backend - `eager` uses standard PyTorch attention (alternatives include `flash_attention_2` or `sdpa`)                               |
+| `--steps`                      | Total number of training steps to run                                                                                                                         |
+| `--policy.device`              | Device to train on (`cuda` for GPU, `cpu` for CPU)                                                                                                            |
+| `--batch_size`                 | Number of samples per training batch                                                                                                                          |
+
+## License
+
+This model follows the **Apache 2.0 License**, consistent with the original [WallX repository](https://github.com/X-Square-Robot/wall-x).
@@ -154,7 +154,7 @@ lerobot-train \

 ```bash
 lerobot-train \
-  --dataset.repo_id=pepijn223/bimanual-so100-handover-cube \
+  --dataset.repo_id=<USER>/bimanual-so100-handover-cube \
  --output_dir=./outputs/xvla_bimanual \
  --job_name=xvla_so101_training \
  --policy.path="lerobot/xvla-base" \
@@ -22,7 +22,7 @@ lerobot-replay \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=black \
-    --dataset.repo_id=aliberts/record-test \
+    --dataset.repo_id=<USER>/record-test \
    --dataset.episode=2
 ```
 """
@@ -41,8 +41,7 @@ from lerobot.robots import (  # noqa: F401
    RobotConfig,
    koch_follower,
    make_robot_from_config,
-    so100_follower,
-    so101_follower,
+    so_follower,
 )
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
@@ -58,7 +57,7 @@ class DatasetReplayConfig:
    repo_id: str
    # Episode to replay.
    episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
    root: str | Path | None = None
    # Limit the frames per second. By default, uses the policy fps.
    fps: int = 30
@@ -82,24 +81,25 @@ def replay(cfg: ReplayConfig):
    actions = dataset.hf_dataset.select_columns(ACTION)
    robot.connect()

-    log_say("Replaying episode", cfg.play_sounds, blocking=True)
-    for idx in range(dataset.num_frames):
-        start_episode_t = time.perf_counter()
+    try:
+        log_say("Replaying episode", cfg.play_sounds, blocking=True)
+        for idx in range(dataset.num_frames):
+            start_episode_t = time.perf_counter()

-        action_array = actions[idx][ACTION]
-        action = {}
-        for i, name in enumerate(dataset.features[ACTION]["names"]):
-            key = f"{name.removeprefix('main_')}.pos"
-            action[key] = action_array[i].item()
+            action_array = actions[idx][ACTION]
+            action = {}
+            for i, name in enumerate(dataset.features[ACTION]["names"]):
+                key = f"{name.removeprefix('main_')}.pos"
+                action[key] = action_array[i].item()

-        action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
-        action["elbow_flex.pos"] -= 90
-        robot.send_action(action)
+            action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
+            action["elbow_flex.pos"] -= 90
+            robot.send_action(action)

-        dt_s = time.perf_counter() - start_episode_t
-        precise_sleep(1 / dataset.fps - dt_s)
-
-    robot.disconnect()
+            dt_s = time.perf_counter() - start_episode_t
+            precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
+    finally:
+        robot.disconnect()


 if __name__ == "__main__":
@@ -32,7 +32,8 @@ import torch
 from huggingface_hub import HfApi

 import lerobot
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.lerobot_dataset import LeRobotDataset


 def main():
@@ -0,0 +1,490 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+SLURM-distributed SARM RA-BC annotation pipeline.
+
+Computes SARM progress values for all frames in a dataset, distributed across
+SLURM workers, then merges the shards into a single sarm_progress.parquet.
+
+Two subcommands, each a separate SLURM submission:
+
+  compute    – N workers, each computes progress for a subset of episodes
+  aggregate  – 1 worker, merges N shards into sarm_progress.parquet, pushes to hub
+
+Usage:
+    python slurm_compute_rabc.py compute \\
+        --repo-id user/dataset --reward-model-path user/sarm_model \\
+        --stride 10 --device cpu --workers 50 --partition cpu
+
+    python slurm_compute_rabc.py aggregate \\
+        --repo-id user/dataset --reward-model-path user/sarm_model \\
+        --partition cpu --push-to-hub
+"""
+
+import argparse
+from pathlib import Path
+
+from datatrove.executor import LocalPipelineExecutor
+from datatrove.executor.slurm import SlurmPipelineExecutor
+from datatrove.pipeline.base import PipelineStep
+
+
+class ComputeProgressShards(PipelineStep):
+    """Each worker computes SARM progress for its assigned episodes."""
+
+    def __init__(
+        self, repo_id, reward_model_path, stride=1, head_mode="sparse", device="cpu", shard_dir="rabc_shards"
+    ):
+        super().__init__()
+        if stride < 1:
+            raise ValueError(f"stride must be >= 1, got {stride}")
+        self.repo_id = repo_id
+        self.reward_model_path = reward_model_path
+        self.stride = stride
+        self.head_mode = head_mode
+        self.device = device
+        self.shard_dir = shard_dir
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import logging
+        from pathlib import Path
+
+        import numpy as np
+        import pyarrow as pa
+        import pyarrow.parquet as pq
+        import torch
+        from tqdm import tqdm
+
+        from lerobot.policies.sarm.compute_rabc_weights import (
+            generate_all_frame_indices,
+            interpolate_progress,
+            load_sarm_resources,
+        )
+        from lerobot.utils.utils import init_logging
+
+        init_logging()
+
+        dataset, reward_model, preprocess = load_sarm_resources(
+            self.repo_id,
+            self.reward_model_path,
+            self.device,
+        )
+
+        if hasattr(preprocess, "eval"):
+            preprocess.eval()
+        for step in preprocess.steps:
+            if hasattr(step, "eval"):
+                step.eval()
+
+        image_key = reward_model.config.image_key
+        state_key = reward_model.config.state_key
+        frame_gap = reward_model.config.frame_gap
+        center_idx = reward_model.config.n_obs_steps // 2
+
+        dual_mode = reward_model.config.uses_dual_heads
+        compute_sparse = self.head_mode in ("sparse", "both") or not dual_mode
+        compute_dense = self.head_mode in ("dense", "both") and dual_mode
+
+        my_episodes = list(range(dataset.num_episodes))[rank::world_size]
+        if not my_episodes:
+            logging.info(f"Rank {rank}: no episodes assigned")
+            return
+        logging.info(f"Rank {rank}: {len(my_episodes)} / {dataset.num_episodes} episodes")
+
+        all_rows = []
+
+        for ep_idx in tqdm(my_episodes, desc=f"Rank {rank}"):
+            ep = dataset.meta.episodes[ep_idx]
+            ep_start, ep_end = ep["dataset_from_index"], ep["dataset_to_index"]
+            task = dataset[ep_start].get("task", "perform the task")
+
+            all_ep_indices = generate_all_frame_indices(ep_start, ep_end, frame_gap)
+            if self.stride > 1:
+                compute_indices = [i for i in all_ep_indices if (i - ep_start) % self.stride == 0]
+                if (ep_end - 1) not in compute_indices:
+                    compute_indices.append(ep_end - 1)
+                compute_indices = sorted(set(compute_indices))
+            else:
+                compute_indices = all_ep_indices
+
+            frame_results = {}
+            for qi in tqdm(compute_indices, desc=f"  Ep {ep_idx}", leave=False):
+                try:
+                    sample = dataset[qi]
+                    batch = {
+                        image_key: sample[image_key],
+                        "task": task,
+                        "index": qi,
+                        "episode_index": ep_idx,
+                    }
+                    if state_key in sample:
+                        batch[state_key] = sample[state_key]
+
+                    with torch.no_grad():
+                        processed = preprocess(batch)
+                        vf = processed["video_features"].to(self.device)
+                        tf = processed["text_features"].to(self.device)
+                        sf = processed.get("state_features")
+                        if sf is not None:
+                            sf = sf.to(self.device)
+                        lengths = processed.get("lengths")
+
+                        sparse_val = dense_val = np.nan
+                        if compute_sparse:
+                            r = reward_model.calculate_rewards(
+                                text_embeddings=tf,
+                                video_embeddings=vf,
+                                state_features=sf,
+                                lengths=lengths,
+                                return_all_frames=True,
+                                head_mode="sparse",
+                            )
+                            sparse_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
+                        if compute_dense:
+                            r = reward_model.calculate_rewards(
+                                text_embeddings=tf,
+                                video_embeddings=vf,
+                                state_features=sf,
+                                lengths=lengths,
+                                return_all_frames=True,
+                                head_mode="dense",
+                            )
+                            dense_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
+
+                        frame_results[qi] = (sparse_val, dense_val)
+                except Exception as e:
+                    logging.warning(f"Failed frame {qi}: {e}")
+
+            if not frame_results:
+                logging.warning(f"Episode {ep_idx}: all frames failed, skipping")
+                continue
+
+            # Interpolate to all frames in this episode
+            computed_idx = np.array(sorted(frame_results.keys()))
+            all_frame_arr = np.arange(ep_start, ep_end)
+
+            sparse_vals = np.array([frame_results[i][0] for i in computed_idx]) if compute_sparse else None
+            dense_vals = np.array([frame_results[i][1] for i in computed_idx]) if compute_dense else None
+
+            if self.stride > 1 and len(computed_idx) > 1:
+                if compute_sparse:
+                    sparse_vals = interpolate_progress(computed_idx, sparse_vals, all_frame_arr)
+                if compute_dense:
+                    dense_vals = interpolate_progress(computed_idx, dense_vals, all_frame_arr)
+                output_frames = all_frame_arr
+            else:
+                # Use only successfully computed frames to avoid indexing mismatch on failures
+                output_frames = computed_idx
+
+            for i, fi in enumerate(output_frames):
+                row = {"index": int(fi), "episode_index": ep_idx, "frame_index": int(fi - ep_start)}
+                if compute_sparse:
+                    row["progress_sparse"] = float(sparse_vals[i])
+                if compute_dense:
+                    row["progress_dense"] = float(dense_vals[i])
+                all_rows.append(row)
+
+        if all_rows:
+            import pandas as pd
+
+            df = pd.DataFrame(all_rows).sort_values("index").reset_index(drop=True)
+            table = pa.Table.from_pandas(df, preserve_index=False)
+            table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
+            shard_dir = Path(self.shard_dir)
+            shard_dir.mkdir(parents=True, exist_ok=True)
+            out = shard_dir / f"shard_{rank:05d}.parquet"
+            pq.write_table(table, out)
+            logging.info(f"Rank {rank}: saved {len(df)} rows to {out}")
+
+
+class AggregateProgress(PipelineStep):
+    """Merge all shard parquets into final sarm_progress.parquet."""
+
+    def __init__(self, repo_id, reward_model_path, shard_dir="rabc_shards", push_to_hub=False):
+        super().__init__()
+        self.repo_id = repo_id
+        self.reward_model_path = reward_model_path
+        self.shard_dir = shard_dir
+        self.push_to_hub = push_to_hub
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import datetime
+        import logging
+        import os
+        from pathlib import Path
+
+        import pandas as pd
+        import pyarrow as pa
+        import pyarrow.parquet as pq
+
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+        from lerobot.utils.utils import init_logging
+
+        init_logging()
+        if rank != 0:
+            return
+
+        shard_dir = Path(self.shard_dir)
+        shards = sorted(shard_dir.glob("shard_*.parquet"))
+        if not shards:
+            raise FileNotFoundError(f"No shards found in {shard_dir}")
+
+        # Log shard modification time range to help detect stale files
+        mtimes = [os.path.getmtime(s) for s in shards]
+        oldest = datetime.datetime.fromtimestamp(min(mtimes)).isoformat(timespec="seconds")
+        newest = datetime.datetime.fromtimestamp(max(mtimes)).isoformat(timespec="seconds")
+        logging.info(f"Aggregating {len(shards)} shards (oldest: {oldest}, newest: {newest})")
+
+        df = pd.concat([pd.read_parquet(s) for s in shards], ignore_index=True)
+        df = df.sort_values("index").reset_index(drop=True)
+
+        table = pa.Table.from_pandas(df, preserve_index=False)
+        table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
+
+        temp_ds = LeRobotDataset(self.repo_id, download_videos=False)
+        out_path = Path(temp_ds.root) / "sarm_progress.parquet"
+        out_path.parent.mkdir(parents=True, exist_ok=True)
+        pq.write_table(table, out_path)
+        logging.info(f"Saved {len(df)} rows to {out_path}")
+
+        for col in ["progress_sparse", "progress_dense"]:
+            if col in df.columns:
+                v = df[col].dropna()
+                logging.info(
+                    f"{col}: mean={v.mean():.4f} std={v.std():.4f} min={v.min():.4f} max={v.max():.4f}"
+                )
+
+        if self.push_to_hub:
+            from huggingface_hub import HfApi
+
+            api = HfApi()
+            hub_path = "sarm_progress.parquet"
+            logging.info(f"Uploading to {self.repo_id}/{hub_path}")
+            api.upload_file(
+                path_or_fileobj=str(out_path),
+                path_in_repo=hub_path,
+                repo_id=self.repo_id,
+                repo_type="dataset",
+            )
+            logging.info(f"Uploaded: https://huggingface.co/datasets/{self.repo_id}/blob/main/{hub_path}")
+
+
+def make_compute_executor(
+    repo_id,
+    reward_model_path,
+    stride,
+    head_mode,
+    device,
+    shard_dir,
+    logs_dir,
+    job_name,
+    slurm,
+    workers,
+    partition,
+    cpus_per_task,
+    mem_per_cpu,
+):
+    kwargs = {
+        "pipeline": [
+            ComputeProgressShards(repo_id, reward_model_path, stride, head_mode, device, str(shard_dir)),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": workers,
+                "workers": workers,
+                "time": "24:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        return SlurmPipelineExecutor(**kwargs)
+
+    kwargs.update({"tasks": workers, "workers": 1})
+    return LocalPipelineExecutor(**kwargs)
+
+
+def make_aggregate_executor(
+    repo_id,
+    reward_model_path,
+    shard_dir,
+    logs_dir,
+    job_name,
+    slurm,
+    partition,
+    cpus_per_task,
+    mem_per_cpu,
+    push_to_hub,
+):
+    kwargs = {
+        "pipeline": [
+            AggregateProgress(repo_id, reward_model_path, str(shard_dir), push_to_hub),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": 1,
+                "workers": 1,
+                "time": "02:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        return SlurmPipelineExecutor(**kwargs)
+
+    kwargs.update({"tasks": 1, "workers": 1})
+    return LocalPipelineExecutor(**kwargs)
+
+
+def _add_shared_args(p):
+    p.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Hugging Face repository identifier, e.g. 'user/dataset'.",
+    )
+    p.add_argument(
+        "--shard-dir",
+        type=Path,
+        default=Path("rabc_shards"),
+        help="Directory to read/write per-rank parquet shards.",
+    )
+    p.add_argument(
+        "--logs-dir",
+        type=Path,
+        default=Path("logs"),
+        help="Directory for datatrove logs.",
+    )
+    p.add_argument(
+        "--job-name",
+        type=str,
+        default=None,
+        help="SLURM job name (defaults to rabc_<subcommand>).",
+    )
+    p.add_argument(
+        "--slurm",
+        type=int,
+        default=1,
+        help="1 = submit via SLURM; 0 = run locally (useful for debugging).",
+    )
+    p.add_argument(
+        "--partition",
+        type=str,
+        default=None,
+        help="SLURM partition to submit to.",
+    )
+    p.add_argument(
+        "--cpus-per-task",
+        type=int,
+        default=4,
+        help="Number of CPUs per SLURM task.",
+    )
+    p.add_argument(
+        "--mem-per-cpu",
+        type=str,
+        default="4G",
+        help="Memory per CPU, e.g. '4G' or '1950M'.",
+    )
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="SLURM-distributed SARM RA-BC annotation pipeline",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    sub = parser.add_subparsers(dest="command", required=True)
+
+    # compute subcommand
+    cp = sub.add_parser(
+        "compute",
+        help="Distribute progress computation across SLURM workers.",
+    )
+    _add_shared_args(cp)
+    cp.add_argument(
+        "--reward-model-path",
+        type=str,
+        required=True,
+        help="Path or HF repo id of the SARM reward model.",
+    )
+    cp.add_argument(
+        "--stride",
+        type=int,
+        default=1,
+        help="Compute every Nth frame; intermediate frames are interpolated (must be >= 1).",
+    )
+    cp.add_argument(
+        "--head-mode",
+        type=str,
+        default="sparse",
+        choices=["sparse", "dense", "both"],
+        help="Which reward head(s) to compute.",
+    )
+    cp.add_argument(
+        "--device",
+        type=str,
+        default="cpu",
+        help="Device for reward model inference, e.g. 'cpu' or 'cuda'.",
+    )
+    cp.add_argument(
+        "--workers",
+        type=int,
+        default=50,
+        help="Number of parallel SLURM tasks (one shard per worker).",
+    )
+
+    # aggregate subcommand
+    ap = sub.add_parser(
+        "aggregate",
+        help="Merge per-rank shards into a single sarm_progress.parquet.",
+    )
+    _add_shared_args(ap)
+    ap.add_argument(
+        "--reward-model-path",
+        type=str,
+        required=True,
+        help="Path or HF repo id of the SARM reward model (stored in parquet metadata).",
+    )
+    ap.add_argument(
+        "--push-to-hub",
+        action="store_true",
+        help="Upload sarm_progress.parquet to the Hugging Face Hub after aggregation.",
+    )
+
+    args = parser.parse_args()
+    job_name = args.job_name or f"rabc_{args.command}"
+    kwargs = vars(args)
+    kwargs["slurm"] = kwargs.pop("slurm") == 1
+    kwargs["job_name"] = job_name
+    command = kwargs.pop("command")
+
+    executor = make_compute_executor(**kwargs) if command == "compute" else make_aggregate_executor(**kwargs)
+
+    executor.run()
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,717 @@
+"""
+Action consistency analysis for imitation learning datasets.
+
+Two parallel analyses per dataset:
+  1. State-based: KNN in joint-state space → action chunk variance
+  2. Image-based: KNN in SigLIP embedding space → action chunk variance
+
+Comparing them reveals whether visual similarity and proprioceptive similarity
+agree on where the data is inconsistent — and images are what the policy
+primarily sees.
+"""
+
+import json
+from pathlib import Path
+
+import av
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+import torch
+from huggingface_hub import snapshot_download
+from matplotlib.colors import LinearSegmentedColormap
+from PIL import Image
+from scipy.spatial import cKDTree
+from transformers import AutoImageProcessor, AutoModel
+
+DATASETS = [
+    {"repo_id": "lerobot-data-collection/level2_final_quality3", "label": "HQ curated"},
+    {"repo_id": "lerobot-data-collection/level12_rac_2_2026-02-08_1", "label": "Full collection"},
+]
+OUTPUT_DIR = Path(__file__).resolve().parent / "outputs"
+OUTPUT_DIR.mkdir(exist_ok=True)
+
+MAX_FRAMES = 100_000
+K_NEIGHBORS = 50
+ACTION_CHUNK_SIZE = 30
+CAMERA_KEY = "observation.images.base"
+ENCODER_MODEL = "google/siglip-base-patch16-224"
+ENCODE_BATCH_SIZE = 512
+SEED = 42
+DPI = 150
+
+CONSISTENCY_CMAP = LinearSegmentedColormap.from_list(
+    "consistency", ["#0a2e0a", "#1a8e1a", "#88cc22", "#ffaa22", "#ff2222"]
+)
+
+# FK chains from OpenArm bimanual URDF (same as workspace_density.py).
+LEFT_CHAIN = [
+    ((-np.pi / 2, 0, 0), (0, 0.031, 0.698), None),
+    ((0, 0, 0), (0, 0, 0.0625), (0, 0, 1)),
+    ((-np.pi / 2, 0, 0), (-0.0301, 0, 0.06), (-1, 0, 0)),
+    ((0, 0, 0), (0.0301, 0, 0.06625), (0, 0, 1)),
+    ((0, 0, 0), (0, 0.0315, 0.15375), (0, 1, 0)),
+    ((0, 0, 0), (0, -0.0315, 0.0955), (0, 0, 1)),
+    ((0, 0, 0), (0.0375, 0, 0.1205), (1, 0, 0)),
+    ((0, 0, 0), (-0.0375, 0, 0), (0, -1, 0)),
+    ((0, 0, 0), (0, 0, 0.1001), None),
+    ((0, 0, 0), (0, 0, 0.08), None),
+]
+RIGHT_CHAIN = [
+    ((np.pi / 2, 0, 0), (0, -0.031, 0.698), None),
+    ((0, 0, 0), (0, 0, 0.0625), (0, 0, 1)),
+    ((np.pi / 2, 0, 0), (-0.0301, 0, 0.06), (-1, 0, 0)),
+    ((0, 0, 0), (0.0301, 0, 0.06625), (0, 0, 1)),
+    ((0, 0, 0), (0, 0.0315, 0.15375), (0, 1, 0)),
+    ((0, 0, 0), (0, -0.0315, 0.0955), (0, 0, 1)),
+    ((0, 0, 0), (0.0375, 0, 0.1205), (1, 0, 0)),
+    ((0, 0, 0), (-0.0375, 0, 0), (0, 1, 0)),
+    ((0, 0, 0), (0, 0, 0.1001), None),
+    ((0, 0, 0), (0, 0, 0.08), None),
+]
+
+
+# ── FK math ─────────────────────────────────────────────
+
+
+def _rot_x(a: float) -> np.ndarray:
+    c, s = np.cos(a), np.sin(a)
+    return np.array([[1, 0, 0], [0, c, -s], [0, s, c]])
+
+
+def _rot_y(a: float) -> np.ndarray:
+    c, s = np.cos(a), np.sin(a)
+    return np.array([[c, 0, s], [0, 1, 0], [-s, 0, c]])
+
+
+def _rot_z(a: float) -> np.ndarray:
+    c, s = np.cos(a), np.sin(a)
+    return np.array([[c, -s, 0], [s, c, 0], [0, 0, 1]])
+
+
+def _tf(rpy: tuple, xyz: tuple) -> np.ndarray:
+    r, p, y = rpy
+    mat = np.eye(4)
+    mat[:3, :3] = _rot_z(y) @ _rot_y(p) @ _rot_x(r)
+    mat[:3, 3] = xyz
+    return mat
+
+
+def _batch_axis_rot(axis: tuple, angles: np.ndarray) -> np.ndarray:
+    n = len(angles)
+    ax = np.asarray(axis, dtype=np.float64)
+    ax = ax / np.linalg.norm(ax)
+    x, y, z = ax
+    c = np.cos(angles)
+    s = np.sin(angles)
+    t = 1 - c
+    rot = np.zeros((n, 4, 4))
+    rot[:, 0, 0] = t * x * x + c
+    rot[:, 0, 1] = t * x * y - s * z
+    rot[:, 0, 2] = t * x * z + s * y
+    rot[:, 1, 0] = t * x * y + s * z
+    rot[:, 1, 1] = t * y * y + c
+    rot[:, 1, 2] = t * y * z - s * x
+    rot[:, 2, 0] = t * x * z - s * y
+    rot[:, 2, 1] = t * y * z + s * x
+    rot[:, 2, 2] = t * z * z + c
+    rot[:, 3, 3] = 1.0
+    return rot
+
+
+def batch_fk(chain: list, joint_angles: np.ndarray) -> np.ndarray:
+    n = joint_angles.shape[0]
+    tf_batch = np.tile(np.eye(4), (n, 1, 1))
+    qi = 0
+    for rpy, xyz, axis in chain:
+        tf_batch = tf_batch @ _tf(rpy, xyz)
+        if axis is not None:
+            rot = _batch_axis_rot(axis, joint_angles[:, qi])
+            tf_batch = np.einsum("nij,njk->nik", tf_batch, rot)
+            qi += 1
+    return tf_batch[:, :3, 3]
+
+
+# ── Data helpers ────────────────────────────────────────
+
+
+def _flatten_names(obj: object) -> list[str]:
+    if isinstance(obj, dict):
+        out: list[str] = []
+        for v in obj.values():
+            out.extend(_flatten_names(v))
+        return out
+    if isinstance(obj, (list, tuple)):
+        out = []
+        for item in obj:
+            if isinstance(item, (list, tuple, dict)):
+                out.extend(_flatten_names(item))
+            else:
+                out.append(str(item))
+        return out
+    return [str(obj)]
+
+
+def _detect_and_convert(vals: np.ndarray) -> np.ndarray:
+    mx = np.max(np.abs(vals))
+    if mx > 360:
+        print(f"    Unit detection: servo ticks (max={mx:.0f})")
+        return (vals - 2048) / 2048 * np.pi
+    if mx > 6.3:
+        print(f"    Unit detection: degrees (max={mx:.1f})")
+        return np.deg2rad(vals)
+    print(f"    Unit detection: radians (max={mx:.3f})")
+    return vals.astype(np.float64)
+
+
+def _find_joint_indices(features: dict, state_col: str, n_dim: int) -> tuple[list[int], list[int]]:
+    feat = features.get("observation.state", features.get(state_col, {}))
+    names = _flatten_names(feat.get("names", []))
+    left_idx: list[int] = []
+    right_idx: list[int] = []
+    if names and len(names) == n_dim:
+        names_l = [n.lower() for n in names]
+        print(f"  Feature names: {names[:4]}…{names[-4:]}")
+        for j in range(1, 8):
+            for i, nm in enumerate(names_l):
+                if f"left_joint_{j}" in nm and i not in left_idx:
+                    left_idx.append(i)
+                    break
+            for i, nm in enumerate(names_l):
+                if f"right_joint_{j}" in nm and i not in right_idx:
+                    right_idx.append(i)
+                    break
+    if len(left_idx) == 7 and len(right_idx) == 7:
+        print(f"  Matched by name: left={left_idx} right={right_idx}")
+        return left_idx, right_idx
+    if n_dim >= 16:
+        print("  Falling back to positional: [0:7]=left, [8:15]=right")
+        return list(range(7)), list(range(8, 15))
+    if n_dim >= 14:
+        print("  Falling back to positional: [0:7]=left, [7:14]=right")
+        return list(range(7)), list(range(7, 14))
+    raise RuntimeError(f"State dim {n_dim} too small for bimanual 7-DOF robot")
+
+
+def download_data(repo_id: str, camera_key: str) -> Path:
+    print(f"  Downloading {repo_id} (parquet + {camera_key} videos) …")
+    return Path(
+        snapshot_download(
+            repo_id=repo_id,
+            repo_type="dataset",
+            allow_patterns=[
+                "meta/**",
+                "data/**",
+                f"videos/{camera_key}/**",
+            ],
+        )
+    )
+
+
+# ── Data loading ────────────────────────────────────────
+
+
+def _build_action_chunks(
+    actions: np.ndarray, episode_ids: np.ndarray, chunk_size: int
+) -> tuple[np.ndarray, np.ndarray]:
+    """
+    For each frame, concatenate the next chunk_size actions from the same episode.
+    Returns (action_chunks, valid_mask).
+    """
+    n = len(actions)
+    act_dim = actions.shape[1]
+    chunks = np.zeros((n, chunk_size * act_dim), dtype=np.float64)
+    valid = np.zeros(n, dtype=bool)
+
+    for i in range(n):
+        end = i + chunk_size
+        if end > n:
+            continue
+        if episode_ids[i] != episode_ids[end - 1]:
+            continue
+        chunks[i] = actions[i:end].ravel()
+        valid[i] = True
+
+    return chunks, valid
+
+
+def load_state_action_data(local: Path, max_frames: int, chunk_size: int, rng: np.random.Generator) -> dict:
+    """
+    Load observation.state and action, build action chunks, subsample, normalize.
+    Also returns the original row indices (`chosen_idx`) for video frame mapping.
+    """
+    info = json.loads((local / "meta" / "info.json").read_text())
+    features = info.get("features", {})
+
+    dfs = [pd.read_parquet(pq) for pq in sorted((local / "data").glob("**/*.parquet"))]
+    df = pd.concat(dfs, ignore_index=True)
+    n_total = len(df)
+    print(f"  Total frames: {n_total:,}")
+
+    state_col = next((c for c in df.columns if "observation.state" in c), None)
+    action_col = next((c for c in df.columns if c == "action"), None)
+    if state_col is None:
+        raise RuntimeError(f"No observation.state column. Available: {list(df.columns)}")
+    if action_col is None:
+        raise RuntimeError(f"No action column. Available: {list(df.columns)}")
+
+    ep_col = next((c for c in df.columns if c == "episode_index"), None)
+    if ep_col is None:
+        raise RuntimeError(f"No episode_index column. Available: {list(df.columns)}")
+
+    state_all = np.stack(df[state_col].values).astype(np.float64)
+    action_all = np.stack(df[action_col].values).astype(np.float64)
+    episode_all = df[ep_col].values.astype(np.int64)
+
+    n_dim = state_all.shape[1]
+    act_dim = action_all.shape[1]
+    print(f"  State dim: {n_dim}  Action dim: {act_dim}  Chunk size: {chunk_size}")
+    print(f"  Action chunk dim: {chunk_size * act_dim}")
+
+    left_idx, right_idx = _find_joint_indices(features, state_col, n_dim)
+
+    print("  Building action chunks …")
+    action_chunks, valid = _build_action_chunks(action_all, episode_all, chunk_size)
+    valid_idx = np.where(valid)[0]
+    print(f"  Valid frames (with full action chunk): {len(valid_idx):,} / {n_total:,}")
+
+    if len(valid_idx) > max_frames:
+        chosen = np.sort(rng.choice(valid_idx, max_frames, replace=False))
+    else:
+        chosen = valid_idx
+    print(f"  Using {len(chosen):,} frames")
+
+    state_raw = state_all[chosen]
+    action_raw = action_chunks[chosen]
+    episode_ids = episode_all[chosen]
+
+    state_mean = state_raw.mean(axis=0)
+    state_std = state_raw.std(axis=0)
+    state_std[state_std < 1e-8] = 1.0
+    state_norm = (state_raw - state_mean) / state_std
+
+    action_mean = action_raw.mean(axis=0)
+    action_std = action_raw.std(axis=0)
+    action_std[action_std < 1e-8] = 1.0
+    action_norm = (action_raw - action_mean) / action_std
+
+    return {
+        "state_raw": state_raw,
+        "state_norm": state_norm,
+        "action_raw": action_raw,
+        "action_norm": action_norm,
+        "episode_ids": episode_ids,
+        "episode_all": episode_all,
+        "left_joint_idx": left_idx,
+        "right_joint_idx": right_idx,
+        "n_total": n_total,
+        "chosen_idx": chosen,
+        "df": df,
+    }
+
+
+# ── Video → frame extraction ──────────────────────────────
+
+
+def build_video_lookup(local: Path, camera_key: str) -> dict:
+    """
+    Build a mapping from episode_index → {video_path, fps, from_ts}.
+    """
+    info = json.loads((local / "meta" / "info.json").read_text())
+    fps = info["fps"]
+    video_template = info.get(
+        "video_path",
+        "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
+    )
+
+    ep_rows = []
+    for pq in sorted((local / "meta" / "episodes").glob("**/*.parquet")):
+        ep_rows.append(pd.read_parquet(pq))
+    ep_df = pd.concat(ep_rows, ignore_index=True)
+
+    chunk_col = f"videos/{camera_key}/chunk_index"
+    file_col = f"videos/{camera_key}/file_index"
+    ts_from = f"videos/{camera_key}/from_timestamp"
+    if chunk_col not in ep_df.columns:
+        chunk_col = f"{camera_key}/chunk_index"
+        file_col = f"{camera_key}/file_index"
+        ts_from = f"{camera_key}/from_timestamp"
+
+    lookup: dict[int, dict] = {}
+    for _, row in ep_df.iterrows():
+        ci = int(row[chunk_col])
+        fi = int(row[file_col])
+        video_rel = video_template.format(video_key=camera_key, chunk_index=ci, file_index=fi)
+        lookup[int(row["episode_index"])] = {
+            "video_path": local / video_rel,
+            "from_ts": float(row[ts_from]),
+            "fps": fps,
+        }
+    return lookup
+
+
+def _decode_video_frames(video_path: str) -> list[np.ndarray]:
+    """Decode all frames from a video file using PyAV. Returns list of RGB arrays."""
+    container = av.open(video_path)
+    stream = container.streams.video[0]
+    stream.thread_type = "AUTO"
+    decoded = []
+    for frame in container.decode(stream):
+        decoded.append(frame.to_ndarray(format="rgb24"))
+    container.close()
+    return decoded
+
+
+def extract_frames(
+    chosen_idx: np.ndarray,
+    episode_all: np.ndarray,
+    video_lookup: dict,
+) -> list[np.ndarray | None]:
+    """
+    Extract RGB frames for each chosen global index using PyAV.
+    Returns list of (H, W, 3) RGB arrays (or None on failure).
+    """
+    unique_eps = np.unique(episode_all)
+    ep_start: dict[int, int] = {}
+    for ep in unique_eps:
+        ep_start[int(ep)] = int(np.where(episode_all == ep)[0][0])
+
+    # Build jobs: (output_index, video_path, local_frame_number)
+    jobs: list[tuple[int, str, int]] = []
+    for out_i, global_i in enumerate(chosen_idx):
+        ep = int(episode_all[global_i])
+        info = video_lookup.get(ep)
+        if info is None:
+            continue
+        local_frame = global_i - ep_start[ep]
+        jobs.append((out_i, str(info["video_path"]), local_frame))
+
+    # Group by video file, decode each video once
+    from collections import defaultdict
+
+    video_jobs: dict[str, list[tuple[int, int]]] = defaultdict(list)
+    for out_i, vpath, local_frame in jobs:
+        video_jobs[vpath].append((out_i, local_frame))
+
+    frames: list[np.ndarray | None] = [None] * len(chosen_idx)
+    extracted = 0
+    n_videos = len(video_jobs)
+    for vi, (vpath, frame_requests) in enumerate(video_jobs.items()):
+        if not Path(vpath).exists():
+            continue
+        try:
+            decoded = _decode_video_frames(vpath)
+        except Exception as exc:
+            print(f"    Warning: failed to decode {Path(vpath).name}: {exc}")
+            continue
+        for out_i, local_frame in frame_requests:
+            if 0 <= local_frame < len(decoded):
+                frames[out_i] = decoded[local_frame]
+                extracted += 1
+        if (vi + 1) % 50 == 0 or (vi + 1) == n_videos:
+            print(f"    Decoded {vi + 1}/{n_videos} videos ({extracted:,} frames so far)")
+        del decoded
+
+    print(f"  Extracted {extracted:,} / {len(chosen_idx):,} frames from video")
+    return frames
+
+
+# ── SigLIP encoding ─────────────────────────────────────
+
+
+def encode_frames_siglip(
+    frames: list[np.ndarray | None],
+    model_name: str,
+    batch_size: int,
+    device: torch.device,
+) -> np.ndarray:
+    """
+    Encode RGB frames through SigLIP vision encoder.
+    Returns (N, embed_dim) float32 array. Frames that are None get a zero vector.
+    """
+    print(f"  Loading SigLIP model: {model_name} …")
+    processor = AutoImageProcessor.from_pretrained(model_name)
+    model = AutoModel.from_pretrained(model_name).to(device).eval()
+    embed_dim = model.config.vision_config.hidden_size
+
+    n = len(frames)
+    embeddings = np.zeros((n, embed_dim), dtype=np.float32)
+
+    valid_indices = [i for i, f in enumerate(frames) if f is not None]
+    print(f"  Encoding {len(valid_indices):,} valid frames in batches of {batch_size} …")
+
+    for batch_start in range(0, len(valid_indices), batch_size):
+        batch_idx = valid_indices[batch_start : batch_start + batch_size]
+        pil_images = [Image.fromarray(frames[i]) for i in batch_idx]
+
+        inputs = processor(images=pil_images, return_tensors="pt").to(device)
+        with torch.no_grad():
+            image_features = model.get_image_features(**inputs)
+        image_features = torch.nn.functional.normalize(image_features, dim=-1)
+        embeddings[batch_idx] = image_features.cpu().numpy()
+
+        done = min(batch_start + batch_size, len(valid_indices))
+        if done % (batch_size * 10) == 0 or done == len(valid_indices):
+            print(f"    {done:,} / {len(valid_indices):,} encoded")
+
+    del model, processor
+    torch.cuda.empty_cache()
+    return embeddings
+
+
+# ── KNN consistency ─────────────────────────────────────
+
+
+def compute_consistency(
+    features: np.ndarray,
+    action_norm: np.ndarray,
+    episode_ids: np.ndarray,
+    k: int,
+    label: str = "",
+) -> np.ndarray:
+    """
+    For each frame, find K nearest neighbors in feature space from other episodes.
+    Return per-frame action variance (mean across action dims).
+    """
+    n = len(features)
+    print(f"  Building KD-tree on {n:,} vectors ({label}) …")
+    tree = cKDTree(features)
+
+    k_query = min(k * 3, n - 1)
+    print(f"  Querying {k_query} neighbors per frame …")
+    _dists, indices = tree.query(features, k=k_query + 1)
+    indices = indices[:, 1:]
+
+    print(f"  Computing cross-episode action variance ({label}) …")
+    variance = np.zeros(n)
+    for i in range(n):
+        ep_i = episode_ids[i]
+        neighbors = indices[i]
+        cross_ep = neighbors[episode_ids[neighbors] != ep_i][:k]
+        if len(cross_ep) < 2:
+            variance[i] = 0.0
+            continue
+        neighbor_actions = action_norm[cross_ep]
+        variance[i] = np.mean(np.var(neighbor_actions, axis=0))
+
+    return variance
+
+
+# ── Visualization ───────────────────────────────────────
+
+
+def _style_ax(ax: plt.Axes) -> None:
+    ax.set_facecolor("#0d1117")
+    ax.tick_params(colors="#555", labelsize=8)
+    for spine in ax.spines.values():
+        spine.set_color("#333")
+
+
+def _plot_histogram(ax: plt.Axes, variance: np.ndarray, title: str, color: str) -> None:
+    _style_ax(ax)
+    median_var = np.median(variance)
+    mean_var = np.mean(variance)
+    nonzero = variance[variance > 0]
+    if len(nonzero) > 0:
+        bins = np.logspace(np.log10(nonzero.min().clip(1e-6)), np.log10(nonzero.max()), 60)
+        ax.hist(nonzero, bins=bins, color=color, alpha=0.8, edgecolor="#222")
+    ax.set_xscale("log")
+    ax.axvline(median_var, color="#ff6600", linewidth=2, label=f"median={median_var:.3f}")
+    ax.axvline(mean_var, color="#ff2222", linewidth=2, linestyle="--", label=f"mean={mean_var:.3f}")
+    ax.set_xlabel("Action variance (log scale)", color="#888", fontsize=10)
+    ax.set_ylabel("Frame count", color="#888", fontsize=10)
+    ax.set_title(title, color="white", fontsize=11, pad=10)
+    ax.legend(fontsize=8, facecolor="#1a1a2e", edgecolor="#333", labelcolor="white")
+
+
+def _plot_episode_curves(
+    ax: plt.Axes,
+    var_state: np.ndarray,
+    var_image: np.ndarray,
+    episode_ids: np.ndarray,
+    title: str,
+) -> None:
+    _style_ax(ax)
+    unique_eps = np.unique(episode_ids)
+
+    ep_means_s = np.array([var_state[episode_ids == ep].mean() for ep in unique_eps])
+    ep_means_i = np.array([var_image[episode_ids == ep].mean() for ep in unique_eps])
+
+    sorted_s = np.sort(ep_means_s)[::-1]
+    sorted_i = np.sort(ep_means_i)[::-1]
+    ep_x = np.arange(len(unique_eps))
+
+    ax.fill_between(ep_x, sorted_s, alpha=0.2, color="#4363d8")
+    ax.plot(ep_x, sorted_s, color="#4363d8", linewidth=1.2, label=f"State (med={np.median(ep_means_s):.3f})")
+    ax.fill_between(ep_x, sorted_i, alpha=0.2, color="#e6194b")
+    ax.plot(ep_x, sorted_i, color="#e6194b", linewidth=1.2, label=f"Image (med={np.median(ep_means_i):.3f})")
+
+    ax.set_xlabel("Episode rank (worst → best)", color="#888", fontsize=10)
+    ax.set_ylabel("Mean action variance", color="#888", fontsize=10)
+    ax.set_title(title, color="white", fontsize=11, pad=10)
+    ax.legend(fontsize=8, facecolor="#1a1a2e", edgecolor="#333", labelcolor="white")
+
+
+def _plot_heatmap(
+    ax: plt.Axes, fig: plt.Figure, tcp_xz: np.ndarray, variance: np.ndarray, title: str
+) -> None:
+    _style_ax(ax)
+    order = np.argsort(variance)
+    pts = tcp_xz[order]
+    var_sorted = variance[order]
+    vmin = np.percentile(variance[variance > 0], 5) if np.any(variance > 0) else 0
+    vmax = np.percentile(variance[variance > 0], 95) if np.any(variance > 0) else 1
+    sc = ax.scatter(
+        pts[:, 0],
+        pts[:, 1],
+        c=var_sorted,
+        cmap=CONSISTENCY_CMAP,
+        s=0.5,
+        alpha=0.6,
+        vmin=vmin,
+        vmax=vmax,
+        rasterized=True,
+    )
+    ax.set_xlabel("X (m)", color="#888", fontsize=10)
+    ax.set_ylabel("Z (m)", color="#888", fontsize=10)
+    ax.set_title(title, color="white", fontsize=11, pad=10)
+    ax.set_aspect("equal")
+    cbar = fig.colorbar(sc, ax=ax, shrink=0.8, pad=0.02)
+    cbar.set_label("Action variance", color="white", fontsize=9)
+    cbar.ax.tick_params(colors="#aaa", labelsize=7)
+
+
+def render(results: list[dict], out_path: Path) -> None:
+    """
+    4-row x N-column figure:
+      Row 0: State-based variance histogram
+      Row 1: Image-based variance histogram
+      Row 2: Per-episode curves (both overlaid)
+      Row 3: Spatial heatmap (image-based variance)
+    """
+    n_ds = len(results)
+    fig, axes = plt.subplots(4, n_ds, figsize=(9 * n_ds, 24), facecolor="#0d1117")
+    if n_ds == 1:
+        axes = axes[:, np.newaxis]
+
+    headline_parts = []
+    for col, r in enumerate(results):
+        label = r["label"]
+        var_s = r["var_state"]
+        var_i = r["var_image"]
+        tcp_xz = r["tcp_xz"]
+        episode_ids = r["episode_ids"]
+
+        med_s = np.median(var_s)
+        med_i = np.median(var_i)
+        headline_parts.append(f"{label}: state={med_s:.3f}, image={med_i:.3f}")
+
+        _plot_histogram(axes[0, col], var_s, f"{label}\nState-based variance (K={K_NEIGHBORS})", "#4363d8")
+        _plot_histogram(
+            axes[1, col], var_i, f"{label}\nImage-based variance (SigLIP, K={K_NEIGHBORS})", "#e6194b"
+        )
+        _plot_episode_curves(
+            axes[2, col],
+            var_s,
+            var_i,
+            episode_ids,
+            f"{label}\nPer-episode inconsistency ({len(np.unique(episode_ids)):,} episodes)",
+        )
+        _plot_heatmap(
+            axes[3, col],
+            fig,
+            tcp_xz,
+            var_i,
+            f"{label}\nImage-based variance by TCP position (XZ)",
+        )
+
+    fig.suptitle(
+        f"Action Consistency: State vs Image  (chunk={ACTION_CHUNK_SIZE}, K={K_NEIGHBORS})\n"
+        + "  |  ".join(headline_parts),
+        color="white",
+        fontsize=15,
+        y=0.99,
+    )
+    plt.tight_layout(rect=[0, 0, 1, 0.96])
+    plt.savefig(out_path, dpi=DPI, bbox_inches="tight", facecolor=fig.get_facecolor())
+    plt.close()
+    print(f"\n✓ Saved: {out_path}")
+
+
+# ── Main ────────────────────────────────────────────────
+
+
+def main() -> None:
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    print(f"Device: {device}")
+    rng = np.random.default_rng(SEED)
+    results = []
+
+    for ds in DATASETS:
+        repo_id, label = ds["repo_id"], ds["label"]
+        print(f"\n{'=' * 60}")
+        print(f"  {label}: {repo_id}")
+        print(f"{'=' * 60}")
+
+        local = download_data(repo_id, CAMERA_KEY)
+        data = load_state_action_data(local, MAX_FRAMES, ACTION_CHUNK_SIZE, rng)
+
+        # --- State-based KNN ---
+        var_state = compute_consistency(
+            data["state_norm"], data["action_norm"], data["episode_ids"], K_NEIGHBORS, "state"
+        )
+        print(
+            f"  State variance: median={np.median(var_state):.4f}  "
+            f"mean={np.mean(var_state):.4f}  p90={np.percentile(var_state, 90):.4f}"
+        )
+
+        # --- Image-based KNN ---
+        print("\n  Preparing image embeddings …")
+        video_lookup = build_video_lookup(local, CAMERA_KEY)
+        frames = extract_frames(data["chosen_idx"], data["episode_all"], video_lookup)
+        embeddings = encode_frames_siglip(frames, ENCODER_MODEL, ENCODE_BATCH_SIZE, device)
+        del frames  # free memory
+
+        var_image = compute_consistency(
+            embeddings, data["action_norm"], data["episode_ids"], K_NEIGHBORS, "image"
+        )
+        print(
+            f"  Image variance: median={np.median(var_image):.4f}  "
+            f"mean={np.mean(var_image):.4f}  p90={np.percentile(var_image, 90):.4f}"
+        )
+
+        # FK for spatial heatmap
+        print("  Computing FK for spatial heatmap …")
+        left_raw = data["state_raw"][:, data["left_joint_idx"]]
+        left_rad = _detect_and_convert(left_raw)
+        left_tcp = batch_fk(LEFT_CHAIN, left_rad)
+        tcp_xz = left_tcp[:, [0, 2]]
+
+        results.append(
+            {
+                "label": label,
+                "var_state": var_state,
+                "var_image": var_image,
+                "episode_ids": data["episode_ids"],
+                "tcp_xz": tcp_xz,
+                "n_total": data["n_total"],
+            }
+        )
+
+    out = OUTPUT_DIR / "action_consistency_comparison.jpg"
+    render(results, out)
+
+    # Save worst-episodes summary (image-based, since that's the stronger signal)
+    worst_summary = {}
+    for r in results:
+        unique_eps = np.unique(r["episode_ids"])
+        ep_means = {int(ep): float(r["var_image"][r["episode_ids"] == ep].mean()) for ep in unique_eps}
+        ranked = sorted(ep_means.items(), key=lambda x: x[1], reverse=True)[:50]
+        worst_summary[r["label"]] = [{"episode": ep, "mean_variance": v} for ep, v in ranked]
+    worst_path = OUTPUT_DIR / "action_consistency_worst_episodes.json"
+    worst_path.write_text(json.dumps(worst_summary, indent=2))
+    print(f"✓ Saved worst episodes: {worst_path}")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,178 @@
+"""
+Create a JPG grid of random frames sampled from a LeRobot video dataset.
+Downloads metadata + video chunks from HuggingFace, picks random frames,
+decodes them, and tiles into a single image.
+"""
+
+import json
+import random
+from pathlib import Path
+
+import cv2
+import numpy as np
+import pandas as pd
+from huggingface_hub import snapshot_download
+
+REPO_ID = "lerobot-data-collection/level2_final_quality3"
+CAMERA_KEY = "observation.images.base"
+GRID_COLS = 15
+GRID_ROWS = 10
+THUMB_WIDTH = 160
+OUTPUT_DIR = Path(__file__).resolve().parent / "outputs"
+OUTPUT_DIR.mkdir(exist_ok=True)
+SEED = 1
+
+
+def download_metadata(repo_id: str) -> Path:
+    """Download only metadata (no videos yet)."""
+    print(f"[1/3] Downloading metadata for {repo_id} …")
+    return Path(
+        snapshot_download(
+            repo_id=repo_id,
+            repo_type="dataset",
+            allow_patterns=["meta/**"],
+            ignore_patterns=["*.mp4"],
+        )
+    )
+
+
+def load_video_info(local: Path) -> tuple[str, list[dict], int]:
+    """Parse info.json and episode parquets. Returns (camera_key, episode_rows, fps)."""
+    info = json.loads((local / "meta" / "info.json").read_text())
+    fps = info["fps"]
+    features = info["features"]
+
+    video_keys = [k for k, v in features.items() if v.get("dtype") == "video"]
+    if not video_keys:
+        raise RuntimeError("No video keys found in dataset features")
+
+    if CAMERA_KEY is not None:
+        if CAMERA_KEY not in video_keys:
+            raise RuntimeError(f"CAMERA_KEY='{CAMERA_KEY}' not found. Available: {video_keys}")
+        cam = CAMERA_KEY
+    else:
+        cam = video_keys[0]
+    print(f"   camera='{cam}'  all_cams={video_keys}  fps={fps}")
+
+    ep_rows = []
+    for pq in sorted((local / "meta" / "episodes").glob("**/*.parquet")):
+        ep_rows.append(pd.read_parquet(pq))
+    ep_df = pd.concat(ep_rows, ignore_index=True)
+
+    video_template = info.get(
+        "video_path",
+        "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
+    )
+
+    chunk_col = f"videos/{cam}/chunk_index"
+    file_col = f"videos/{cam}/file_index"
+    ts_from = f"videos/{cam}/from_timestamp"
+    ts_to = f"videos/{cam}/to_timestamp"
+    if chunk_col not in ep_df.columns:
+        chunk_col = f"{cam}/chunk_index"
+        file_col = f"{cam}/file_index"
+        ts_from = f"{cam}/from_timestamp"
+        ts_to = f"{cam}/to_timestamp"
+
+    episodes = []
+    for _, row in ep_df.iterrows():
+        ci = int(row[chunk_col])
+        fi = int(row[file_col])
+        episodes.append(
+            {
+                "episode_index": int(row["episode_index"]),
+                "chunk_index": ci,
+                "file_index": fi,
+                "from_ts": float(row[ts_from]),
+                "to_ts": float(row[ts_to]),
+                "video_rel": video_template.format(video_key=cam, chunk_index=ci, file_index=fi),
+            }
+        )
+    return cam, episodes, fps
+
+
+def pick_random_frames(episodes: list[dict], fps: int, n: int, rng: random.Random) -> list[dict]:
+    """Pick n random (episode, timestamp) pairs, return sorted by video file for efficient access."""
+    picks = []
+    for _ in range(n):
+        ep = rng.choice(episodes)
+        duration = ep["to_ts"] - ep["from_ts"]
+        if duration <= 0:
+            continue
+        t = ep["from_ts"] + rng.random() * duration
+        picks.append({**ep, "seek_ts": t})
+    picks.sort(key=lambda p: (p["video_rel"], p["seek_ts"]))
+    return picks
+
+
+def download_video_files(repo_id: str, local: Path, picks: list[dict]) -> None:
+    """Download only the video files we need."""
+    needed = sorted({p["video_rel"] for p in picks})
+    print(f"[2/3] Downloading {len(needed)} video file(s) …")
+    snapshot_download(
+        repo_id=repo_id,
+        repo_type="dataset",
+        local_dir=str(local),
+        allow_patterns=needed,
+    )
+
+
+def extract_frame(video_path: Path, seek_ts: float) -> np.ndarray | None:
+    """Decode a single frame at the given timestamp."""
+    cap = cv2.VideoCapture(str(video_path))
+    cap.set(cv2.CAP_PROP_POS_MSEC, seek_ts * 1000.0)
+    ret, frame = cap.read()
+    cap.release()
+    return frame if ret else None
+
+
+def build_grid(frames: list[np.ndarray], cols: int, thumb_w: int) -> np.ndarray:
+    """Resize frames to uniform thumbnails and tile into a grid."""
+    if not frames:
+        raise RuntimeError("No frames decoded")
+
+    h0, w0 = frames[0].shape[:2]
+    thumb_h = int(thumb_w * h0 / w0)
+
+    thumbs = [cv2.resize(f, (thumb_w, thumb_h), interpolation=cv2.INTER_AREA) for f in frames]
+
+    rows = []
+    for i in range(0, len(thumbs), cols):
+        row_thumbs = thumbs[i : i + cols]
+        while len(row_thumbs) < cols:
+            row_thumbs.append(np.zeros_like(row_thumbs[0]))
+        rows.append(np.hstack(row_thumbs))
+    return np.vstack(rows)
+
+
+def main() -> None:
+    rng = random.Random(SEED)
+    n_frames = GRID_COLS * GRID_ROWS
+
+    local = download_metadata(REPO_ID)
+    cam, episodes, fps = load_video_info(local)
+    picks = pick_random_frames(episodes, fps, n_frames, rng)
+    download_video_files(REPO_ID, local, picks)
+
+    print(f"[3/3] Decoding {n_frames} frames …")
+    frames: list[np.ndarray] = []
+    for p in picks:
+        vp = local / p["video_rel"]
+        if not vp.exists():
+            print(f"   SKIP: {p['video_rel']} not found")
+            continue
+        frame = extract_frame(vp, p["seek_ts"])
+        if frame is not None:
+            frames.append(frame)
+
+    print(f"   Decoded {len(frames)}/{n_frames} frames")
+    grid = build_grid(frames, GRID_COLS, THUMB_WIDTH)
+
+    safe_name = REPO_ID.replace("/", "_")
+    out_path = OUTPUT_DIR / f"{safe_name}_grid_{GRID_COLS}x{GRID_ROWS}.jpg"
+    cv2.imwrite(str(out_path), grid, [cv2.IMWRITE_JPEG_QUALITY, 92])
+    print(f"\n✓ Saved: {out_path}  ({grid.shape[1]}×{grid.shape[0]})")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,526 @@
+"""
+Create MP4 videos with sarm_progress overlay for specified episodes.
+Downloads datasets from HuggingFace, extracts episode video + progress data,
+and draws the progress line directly on each frame (no panel, no axes).
+"""
+
+import json
+import subprocess
+from pathlib import Path
+
+import cv2
+import numpy as np
+import pandas as pd
+from huggingface_hub import snapshot_download
+
+DATASETS = [
+    {"repo_id": "lerobot-data-collection/level2_final_quality3", "episode": 250},
+]
+CAMERA_KEY = (
+    "observation.images.base"  # None = auto-select first camera, or set e.g. "observation.images.top"
+)
+OUTPUT_DIR = Path(__file__).resolve().parent / "outputs"
+OUTPUT_DIR.mkdir(exist_ok=True)
+
+# Progress line spans the full video height
+GRAPH_Y_TOP_FRAC = 0.01
+GRAPH_Y_BOT_FRAC = 0.99
+LINE_THICKNESS = 3
+SHADOW_THICKNESS = 6  # white edge thickness
+REF_ALPHA = 0.45  # opacity of the 1.0 reference line
+FILL_ALPHA = 0.55  # opacity of the grey fill under the line
+SCORE_FONT_SCALE = 0.8
+TASK_FONT_SCALE = 0.55
+
+
+def download_episode(repo_id: str, episode: int) -> Path:
+    """Download only the files needed for this episode."""
+    # We need: meta/, sarm_progress.parquet, and the relevant video/data chunks.
+    # We'll download meta + sarm first, then figure out chunks.
+    print(f"\n[1/5] Downloading metadata for {repo_id} …")
+    local = Path(
+        snapshot_download(
+            repo_id=repo_id,
+            repo_type="dataset",
+            allow_patterns=["meta/**", "sarm_progress.parquet"],
+            ignore_patterns=["*.mp4"],
+        )
+    )
+    return local
+
+
+def load_episode_meta(local: Path, episode: int) -> dict:
+    """Read info.json + episode-level parquet to get fps, video paths, timestamps."""
+    info = json.loads((local / "meta" / "info.json").read_text())
+    fps = info["fps"]
+    features = info["features"]
+
+    # Find video keys (keys whose dtype=="video")
+    video_keys = [k for k, v in features.items() if v.get("dtype") == "video"]
+    if not video_keys:
+        raise RuntimeError("No video keys found in dataset features")
+    if CAMERA_KEY is not None:
+        if CAMERA_KEY not in video_keys:
+            raise RuntimeError(f"CAMERA_KEY='{CAMERA_KEY}' not found. Available: {video_keys}")
+        first_cam = CAMERA_KEY
+    else:
+        first_cam = video_keys[0]
+    print(f"   fps={fps}  camera='{first_cam}'  all_cams={video_keys}")
+
+    # Load all episode-meta parquet files and find our episode
+    ep_rows = []
+    for pq in sorted((local / "meta" / "episodes").glob("**/*.parquet")):
+        df = pd.read_parquet(pq)
+        ep_rows.append(df)
+    ep_df = pd.concat(ep_rows, ignore_index=True)
+    row = ep_df[ep_df["episode_index"] == episode]
+    if row.empty:
+        raise RuntimeError(f"Episode {episode} not found in episode metadata")
+    row = row.iloc[0]
+
+    # Extract video chunk/file index for first camera
+    # Try both dot and slash variants of the key
+    chunk_col = f"videos/{first_cam}/chunk_index"
+    file_col = f"videos/{first_cam}/file_index"
+    ts_col = f"videos/{first_cam}/from_timestamp"
+    to_col = f"videos/{first_cam}/to_timestamp"
+
+    # Some datasets use different column naming
+    if chunk_col not in row.index:
+        # Try without the 'videos/' prefix
+        chunk_col = f"{first_cam}/chunk_index"
+        file_col = f"{first_cam}/file_index"
+        ts_col = f"{first_cam}/from_timestamp"
+        to_col = f"{first_cam}/to_timestamp"
+    if chunk_col not in row.index:
+        raise RuntimeError(
+            f"Cannot find video metadata columns for {first_cam}.\nAvailable: {list(row.index)}"
+        )
+
+    chunk_idx = int(row[chunk_col])
+    file_idx = int(row[file_col])
+    from_ts = float(row[ts_col])
+    to_ts = float(row[to_col])
+
+    video_template = info.get(
+        "video_path", "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4"
+    )
+    video_rel = video_template.format(
+        video_key=first_cam,
+        chunk_index=chunk_idx,
+        file_index=file_idx,
+    )
+
+    # Load task name for this episode
+    # tasks.parquet uses the task string as the row index; task_index column holds the int id
+    task_name = ""
+    try:
+        # Prefer the 'tasks' list directly on the episode row
+        if "tasks" in row.index and row["tasks"] is not None:
+            tasks_val = row["tasks"]
+            if isinstance(tasks_val, (list, tuple, np.ndarray)) and len(tasks_val) > 0:
+                task_name = str(tasks_val[0])
+            else:
+                task_name = str(tasks_val).strip("[]'")
+        else:
+            tasks_pq = local / "meta" / "tasks.parquet"
+            if tasks_pq.exists():
+                tasks_df = pd.read_parquet(tasks_pq)
+                # Row index is the task string; task_index column is the int
+                task_idx = int(row.get("task_index", 0)) if "task_index" in row.index else 0
+                match = tasks_df[tasks_df["task_index"] == task_idx]
+                if not match.empty:
+                    task_name = str(match.index[0])
+        print(f"   Task name: '{task_name}'")
+    except Exception as e:
+        print(f"   WARNING: could not load task name: {e}")
+
+    return {
+        "fps": fps,
+        "first_cam": first_cam,
+        "video_rel": video_rel,
+        "chunk_index": chunk_idx,
+        "file_index": file_idx,
+        "from_ts": from_ts,
+        "to_ts": to_ts,
+        "task_name": task_name,
+    }
+
+
+def download_video(repo_id: str, local: Path, video_rel: str) -> Path:
+    """Download the specific video file if not already present."""
+    video_path = local / video_rel
+    if video_path.exists():
+        print(f"   Video already cached: {video_path}")
+        return video_path
+    print(f"[2/5] Downloading video file {video_rel} …")
+    snapshot_download(
+        repo_id=repo_id,
+        repo_type="dataset",
+        local_dir=str(local),
+        allow_patterns=[video_rel],
+    )
+    if not video_path.exists():
+        raise RuntimeError(f"Video not found after download: {video_path}")
+    return video_path
+
+
+def load_progress(local: Path, episode: int) -> np.ndarray | None:
+    """Load sarm_progress values for this episode. Returns sorted array of (frame_index, progress)."""
+    pq_path = local / "sarm_progress.parquet"
+    if not pq_path.exists():
+        print("   WARNING: sarm_progress.parquet not found, trying data parquet …")
+        return None
+    df = pd.read_parquet(pq_path)
+    print(f"   sarm_progress.parquet columns: {list(df.columns)}")
+    ep_df = df[df["episode_index"] == episode].copy()
+    if ep_df.empty:
+        print(f"   WARNING: No sarm_progress rows for episode {episode}")
+        return None
+    ep_df = ep_df.sort_values("frame_index")
+
+    # Prefer dense, fall back to sparse
+    if "progress_dense" in ep_df.columns and ep_df["progress_dense"].notna().any():
+        prog_col = "progress_dense"
+    elif "progress_sparse" in ep_df.columns:
+        prog_col = "progress_sparse"
+    else:
+        # Last resort: any column with 'progress' in the name
+        prog_cols = [c for c in ep_df.columns if "progress" in c.lower()]
+        if not prog_cols:
+            return None
+        prog_col = prog_cols[0]
+
+    print(f"   Using progress column: '{prog_col}'")
+    return ep_df[["frame_index", prog_col]].rename(columns={prog_col: "progress"}).values
+
+
+def extract_episode_clip(video_path: Path, from_ts: float, to_ts: float, out_path: Path) -> Path:
+    """Use ffmpeg to cut the episode segment from the combined video file."""
+    duration = to_ts - from_ts
+    print(f"[3/5] Extracting clip [{from_ts:.3f}s → {to_ts:.3f}s] ({duration:.2f}s) …")
+    cmd = [
+        "ffmpeg",
+        "-y",
+        "-ss",
+        str(from_ts),
+        "-i",
+        str(video_path),
+        "-t",
+        str(duration),
+        "-c:v",
+        "libx264",
+        "-preset",
+        "fast",
+        "-crf",
+        "18",
+        "-an",
+        str(out_path),
+    ]
+    result = subprocess.run(cmd, capture_output=True, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"ffmpeg clip extraction failed:\n{result.stderr}")
+    return out_path
+
+
+def precompute_pixels(
+    progress_data: np.ndarray,
+    n_frames: int,
+    frame_w: int,
+    frame_h: int,
+) -> np.ndarray:
+    """
+    Map each progress sample to pixel coordinates.
+    Returns array of shape (N, 2) with (x, y) in pixel space.
+    x spans full video width; y maps progress [0,1] to graph band.
+    """
+    frame_indices = progress_data[:, 0].astype(float)
+    progress_vals = np.clip(progress_data[:, 1].astype(float), 0.0, 1.0)
+
+    y_top = int(frame_h * GRAPH_Y_TOP_FRAC)
+    y_bot = int(frame_h * GRAPH_Y_BOT_FRAC)
+    graph_h = y_bot - y_top
+
+    xs = (frame_indices / (n_frames - 1) * (frame_w - 1)).astype(int)
+    # progress=1 → y_top, progress=0 → y_bot
+    ys = (y_bot - progress_vals * graph_h).astype(int)
+
+    return np.stack([xs, ys], axis=1)  # (N, 2)
+
+
+def progress_color(t: float) -> tuple[int, int, int]:
+    """Interpolate BGR color red→green based on normalised position t in [0,1]."""
+    r = int(255 * (1.0 - t))
+    g = int(255 * t)
+    return (0, g, r)  # BGR
+
+
+def prerender_fill(
+    pixels: np.ndarray,
+    frame_w: int,
+    frame_h: int,
+) -> np.ndarray:
+    """Pre-render the full grey fill polygon under the curve as a BGRA image."""
+    y_bot = int(frame_h * GRAPH_Y_BOT_FRAC)
+    fill_img = np.zeros((frame_h, frame_w, 4), dtype=np.uint8)
+    poly = np.concatenate(
+        [
+            pixels,
+            [[pixels[-1][0], y_bot], [pixels[0][0], y_bot]],
+        ],
+        axis=0,
+    ).astype(np.int32)
+    cv2.fillPoly(fill_img, [poly], color=(128, 128, 128, int(255 * FILL_ALPHA)))
+    return fill_img
+
+
+def alpha_composite(base: np.ndarray, overlay_bgra: np.ndarray, x_max: int) -> None:
+    """Blend overlay onto base in-place, but only for x < x_max."""
+    if x_max <= 0:
+        return
+    roi_b = base[:, :x_max]
+    roi_o = overlay_bgra[:, :x_max]
+    alpha = roi_o[:, :, 3:4].astype(np.float32) / 255.0
+    roi_b[:] = np.clip(
+        roi_o[:, :, :3].astype(np.float32) * alpha + roi_b.astype(np.float32) * (1.0 - alpha),
+        0,
+        255,
+    ).astype(np.uint8)
+
+
+def draw_text_outlined(
+    frame: np.ndarray,
+    text: str,
+    pos: tuple[int, int],
+    font_scale: float,
+    thickness: int = 1,
+) -> None:
+    """Draw text with a dark outline for readability on any background."""
+    font = cv2.FONT_HERSHEY_SIMPLEX
+    cv2.putText(frame, text, pos, font, font_scale, (0, 0, 0), thickness + 2, cv2.LINE_AA)
+    cv2.putText(frame, text, pos, font, font_scale, (255, 255, 255), thickness, cv2.LINE_AA)
+
+
+def composite_video(
+    clip_path: Path,
+    progress_data: np.ndarray,
+    out_path: Path,
+    fps: float,
+    frame_h: int,
+    frame_w: int,
+    task_name: str = "",
+) -> Path:
+    """Read clip frames, draw gradient progress line with fill + labels, export as GIF."""
+    n_total = int(cv2.VideoCapture(str(clip_path)).get(cv2.CAP_PROP_FRAME_COUNT))
+    pixels = precompute_pixels(progress_data, n_total, frame_w, frame_h)
+
+    y_ref = int(frame_h * GRAPH_Y_TOP_FRAC)
+
+    # Pre-render fill polygon (line is drawn per-frame with live color)
+    fill_img = prerender_fill(pixels, frame_w, frame_h)
+
+    # 1.0 reference line overlay (full width, drawn once)
+    ref_img = np.zeros((frame_h, frame_w, 4), dtype=np.uint8)
+    cv2.line(ref_img, (0, y_ref), (frame_w - 1, y_ref), (200, 200, 200, int(255 * REF_ALPHA)), 1, cv2.LINE_AA)
+
+    frame_indices = progress_data[:, 0].astype(int)
+    progress_vals = progress_data[:, 1].astype(float)
+
+    print(f"[4/4] Compositing {n_total} frames …")
+    cap = cv2.VideoCapture(str(clip_path))
+    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+    tmp_path = out_path.parent / (out_path.stem + "_tmp.mp4")
+    writer = cv2.VideoWriter(str(tmp_path), fourcc, fps, (frame_w, frame_h))
+
+    fi = 0
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break
+
+        n_drawn = int(np.searchsorted(frame_indices, fi, side="right"))
+        x_cur = int(pixels[min(n_drawn, len(pixels)) - 1][0]) + 1 if n_drawn > 0 else 0
+
+        # 1. reference line (full width, always)
+        alpha_composite(frame, ref_img, frame_w)
+
+        # 2. grey fill under curve up to current x
+        alpha_composite(frame, fill_img, x_cur)
+
+        # 3. progress line — single color that transitions red→green over time
+        if n_drawn >= 2:
+            t_cur = (n_drawn - 1) / max(len(progress_vals) - 1, 1)
+            line_col = progress_color(t_cur)
+            pts = pixels[:n_drawn].reshape(-1, 1, 2).astype(np.int32)
+            cv2.polylines(
+                frame,
+                [pts],
+                isClosed=False,
+                color=(255, 255, 255),
+                thickness=SHADOW_THICKNESS,
+                lineType=cv2.LINE_AA,
+            )
+            cv2.polylines(
+                frame, [pts], isClosed=False, color=line_col, thickness=LINE_THICKNESS, lineType=cv2.LINE_AA
+            )
+
+        # 4. score — bottom right
+        if n_drawn > 0:
+            score = float(progress_vals[min(n_drawn, len(progress_vals)) - 1])
+            score_text = f"{score:.2f}"
+            (tw, th), _ = cv2.getTextSize(score_text, cv2.FONT_HERSHEY_SIMPLEX, SCORE_FONT_SCALE, 2)
+            sx = frame_w - tw - 12
+            sy = frame_h - 12
+            # coloured score matching current gradient position
+            t_cur = (n_drawn - 1) / max(len(progress_vals) - 1, 1)
+            score_col = progress_color(t_cur)
+            cv2.putText(
+                frame,
+                score_text,
+                (sx, sy),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                SCORE_FONT_SCALE,
+                (0, 0, 0),
+                4,
+                cv2.LINE_AA,
+            )
+            cv2.putText(
+                frame,
+                score_text,
+                (sx, sy),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                SCORE_FONT_SCALE,
+                score_col,
+                2,
+                cv2.LINE_AA,
+            )
+
+        # 5. task name — top centre
+        if task_name:
+            (tw, _), _ = cv2.getTextSize(task_name, cv2.FONT_HERSHEY_SIMPLEX, TASK_FONT_SCALE, 1)
+            tx = max((frame_w - tw) // 2, 4)
+            draw_text_outlined(frame, task_name, (tx, 22), TASK_FONT_SCALE)
+
+        writer.write(frame)
+        fi += 1
+        if fi % 100 == 0:
+            print(f"   Frame {fi}/{n_total} …", end="\r")
+
+    cap.release()
+    writer.release()
+    print()
+
+    # Convert to GIF: full resolution, 12fps, 128-color diff palette (<40MB)
+    gif_path = out_path.with_suffix(".gif")
+    palette = out_path.parent / "_palette.png"
+    r1 = subprocess.run(  # nosec B607
+        [
+            "ffmpeg",
+            "-y",
+            "-i",
+            str(tmp_path),
+            "-vf",
+            f"fps=10,scale={frame_w}:-1:flags=lanczos,palettegen=max_colors=128:stats_mode=diff",
+            "-update",
+            "1",
+            str(palette),
+        ],
+        capture_output=True,
+        text=True,
+    )
+    if r1.returncode != 0:
+        print(f"   WARNING: palettegen failed:\n{r1.stderr[-500:]}")
+    r2 = subprocess.run(  # nosec B607
+        [
+            "ffmpeg",
+            "-y",
+            "-i",
+            str(tmp_path),
+            "-i",
+            str(palette),
+            "-filter_complex",
+            f"fps=10,scale={frame_w}:-1:flags=lanczos[v];[v][1:v]paletteuse=dither=bayer:bayer_scale=3",
+            str(gif_path),
+        ],
+        capture_output=True,
+        text=True,
+    )
+    if r2.returncode != 0:
+        print(f"   WARNING: gif encode failed:\n{r2.stderr[-500:]}")
+    tmp_path.unlink(missing_ok=True)
+    palette.unlink(missing_ok=True)
+    return gif_path
+
+
+def process_dataset(repo_id: str, episode: int):
+    safe_name = repo_id.replace("/", "_")
+    print(f"\n{'=' * 60}")
+    print(f"Processing: {repo_id}  |  episode {episode}")
+    print(f"{'=' * 60}")
+
+    # 1. Download metadata
+    local = download_episode(repo_id, episode)
+    print(f"   Local cache: {local}")
+
+    # 2. Read episode metadata
+    ep_meta = load_episode_meta(local, episode)
+    print(f"   Episode meta: {ep_meta}")
+
+    # 3. Download video file
+    video_path = download_video(repo_id, local, ep_meta["video_rel"])
+
+    # 4. Extract clip
+    clip_path = OUTPUT_DIR / f"{safe_name}_ep{episode}_clip.mp4"
+    extract_episode_clip(video_path, ep_meta["from_ts"], ep_meta["to_ts"], clip_path)
+
+    # 5. Load progress data
+    progress_data = load_progress(local, episode)
+    if progress_data is None:
+        print("   ERROR: Could not load sarm_progress data. Skipping overlay.")
+        return
+
+    n_progress = len(progress_data)
+    print(f"   Progress frames: {n_progress}")
+
+    # 6. Get clip dimensions
+    cap = cv2.VideoCapture(str(clip_path))
+    frame_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    frame_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    n_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    actual_fps = cap.get(cv2.CAP_PROP_FPS) or ep_meta["fps"]
+    cap.release()
+    print(f"   Clip: {frame_w}×{frame_h}  {n_frames} frames @ {actual_fps:.1f}fps")
+
+    # 7. Composite (draw line directly on frames)
+    out_path = OUTPUT_DIR / f"{safe_name}_ep{episode}_progress.mp4"
+    final = composite_video(
+        clip_path,
+        progress_data,
+        out_path,
+        actual_fps,
+        frame_h,
+        frame_w,
+        task_name=ep_meta.get("task_name", ""),
+    )
+    clip_path.unlink(missing_ok=True)
+    print(f"\n✓ Done: {final}")
+    return final
+
+
+if __name__ == "__main__":
+    results = []
+    for cfg in DATASETS:
+        try:
+            out = process_dataset(cfg["repo_id"], cfg["episode"])
+            if out:
+                results.append(out)
+        except Exception as e:
+            print(f"\nERROR processing {cfg['repo_id']}: {e}")
+            import traceback
+
+            traceback.print_exc()
+
+    print("\n" + "=" * 60)
+    print("Output files:")
+    for r in results:
+        print(f"  {r}")
@@ -0,0 +1,496 @@
+"""
+Visualize end-effector workspace density and trajectory clusters for OpenArm datasets.
+Downloads joint position data (no videos) from HuggingFace, computes forward
+kinematics per episode, clusters trajectories with K-means, and renders
+2D projections comparing dataset coverage and multimodality.
+"""
+
+import json
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+from huggingface_hub import snapshot_download
+from sklearn.cluster import KMeans
+
+DATASETS = [
+    {"repo_id": "lerobot-data-collection/level2_final_quality3", "label": "HQ curated"},
+    {"repo_id": "lerobot-data-collection/level12_rac_2_2026-02-08_1", "label": "Full collection"},
+]
+OUTPUT_DIR = Path(__file__).resolve().parent / "outputs"
+OUTPUT_DIR.mkdir(exist_ok=True)
+
+N_CLUSTERS = 10
+WAYPOINTS = 50
+SEED = 42
+DPI = 180
+
+CLUSTER_COLORS = [
+    "#e6194b",
+    "#3cb44b",
+    "#4363d8",
+    "#f58231",
+    "#911eb4",
+    "#42d4f4",
+    "#f032e6",
+    "#bfef45",
+    "#fabed4",
+    "#dcbeff",
+    "#9a6324",
+    "#fffac8",
+    "#800000",
+    "#aaffc3",
+    "#808000",
+    "#ffd8b1",
+    "#000075",
+    "#a9a9a9",
+]
+
+# FK chains extracted from OpenArm bimanual URDF.
+# Each entry: (rpy, xyz, revolute_axis_or_None).
+LEFT_CHAIN = [
+    ((-np.pi / 2, 0, 0), (0, 0.031, 0.698), None),
+    ((0, 0, 0), (0, 0, 0.0625), (0, 0, 1)),
+    ((-np.pi / 2, 0, 0), (-0.0301, 0, 0.06), (-1, 0, 0)),
+    ((0, 0, 0), (0.0301, 0, 0.06625), (0, 0, 1)),
+    ((0, 0, 0), (0, 0.0315, 0.15375), (0, 1, 0)),
+    ((0, 0, 0), (0, -0.0315, 0.0955), (0, 0, 1)),
+    ((0, 0, 0), (0.0375, 0, 0.1205), (1, 0, 0)),
+    ((0, 0, 0), (-0.0375, 0, 0), (0, -1, 0)),
+    ((0, 0, 0), (0, 0, 0.1001), None),
+    ((0, 0, 0), (0, 0, 0.08), None),
+]
+RIGHT_CHAIN = [
+    ((np.pi / 2, 0, 0), (0, -0.031, 0.698), None),
+    ((0, 0, 0), (0, 0, 0.0625), (0, 0, 1)),
+    ((np.pi / 2, 0, 0), (-0.0301, 0, 0.06), (-1, 0, 0)),
+    ((0, 0, 0), (0.0301, 0, 0.06625), (0, 0, 1)),
+    ((0, 0, 0), (0, 0.0315, 0.15375), (0, 1, 0)),
+    ((0, 0, 0), (0, -0.0315, 0.0955), (0, 0, 1)),
+    ((0, 0, 0), (0.0375, 0, 0.1205), (1, 0, 0)),
+    ((0, 0, 0), (-0.0375, 0, 0), (0, 1, 0)),
+    ((0, 0, 0), (0, 0, 0.1001), None),
+    ((0, 0, 0), (0, 0, 0.08), None),
+]
+
+
+# ── FK math ─────────────────────────────────────────────
+
+
+def _rot_x(a: float) -> np.ndarray:
+    c, s = np.cos(a), np.sin(a)
+    return np.array([[1, 0, 0], [0, c, -s], [0, s, c]])
+
+
+def _rot_y(a: float) -> np.ndarray:
+    c, s = np.cos(a), np.sin(a)
+    return np.array([[c, 0, s], [0, 1, 0], [-s, 0, c]])
+
+
+def _rot_z(a: float) -> np.ndarray:
+    c, s = np.cos(a), np.sin(a)
+    return np.array([[c, -s, 0], [s, c, 0], [0, 0, 1]])
+
+
+def _tf(rpy: tuple, xyz: tuple) -> np.ndarray:
+    """Build a 4x4 homogeneous transform from URDF rpy + xyz."""
+    r, p, y = rpy
+    mat = np.eye(4)
+    mat[:3, :3] = _rot_z(y) @ _rot_y(p) @ _rot_x(r)
+    mat[:3, 3] = xyz
+    return mat
+
+
+def _batch_axis_rot(axis: tuple, angles: np.ndarray) -> np.ndarray:
+    """Batched Rodrigues rotation: (n,) angles around a fixed axis → (n, 4, 4)."""
+    n = len(angles)
+    ax = np.asarray(axis, dtype=np.float64)
+    ax = ax / np.linalg.norm(ax)
+    x, y, z = ax
+    c = np.cos(angles)
+    s = np.sin(angles)
+    t = 1 - c
+    rot = np.zeros((n, 4, 4))
+    rot[:, 0, 0] = t * x * x + c
+    rot[:, 0, 1] = t * x * y - s * z
+    rot[:, 0, 2] = t * x * z + s * y
+    rot[:, 1, 0] = t * x * y + s * z
+    rot[:, 1, 1] = t * y * y + c
+    rot[:, 1, 2] = t * y * z - s * x
+    rot[:, 2, 0] = t * x * z - s * y
+    rot[:, 2, 1] = t * y * z + s * x
+    rot[:, 2, 2] = t * z * z + c
+    rot[:, 3, 3] = 1.0
+    return rot
+
+
+def batch_fk(chain: list, joint_angles: np.ndarray) -> np.ndarray:
+    """Vectorized FK: (n, 7) radians → (n, 3) TCP positions in world frame."""
+    n = joint_angles.shape[0]
+    tf_batch = np.tile(np.eye(4), (n, 1, 1))
+    qi = 0
+    for rpy, xyz, axis in chain:
+        tf_batch = tf_batch @ _tf(rpy, xyz)
+        if axis is not None:
+            rot = _batch_axis_rot(axis, joint_angles[:, qi])
+            tf_batch = np.einsum("nij,njk->nik", tf_batch, rot)
+            qi += 1
+    return tf_batch[:, :3, 3]
+
+
+# ── Data loading ────────────────────────────────────────
+
+
+def _flatten_names(obj: object) -> list[str]:
+    """Recursively flatten a names structure (list, dict, or nested) into a flat string list."""
+    if isinstance(obj, dict):
+        out: list[str] = []
+        for v in obj.values():
+            out.extend(_flatten_names(v))
+        return out
+    if isinstance(obj, (list, tuple)):
+        out = []
+        for item in obj:
+            if isinstance(item, (list, tuple, dict)):
+                out.extend(_flatten_names(item))
+            else:
+                out.append(str(item))
+        return out
+    return [str(obj)]
+
+
+def _detect_and_convert(vals: np.ndarray) -> np.ndarray:
+    """Auto-detect servo ticks / degrees / radians and convert to radians."""
+    mx = np.max(np.abs(vals))
+    if mx > 360:
+        print(f"    Unit detection: servo ticks (max={mx:.0f})")
+        return (vals - 2048) / 2048 * np.pi
+    if mx > 6.3:
+        print(f"    Unit detection: degrees (max={mx:.1f})")
+        return np.deg2rad(vals)
+    print(f"    Unit detection: radians (max={mx:.3f})")
+    return vals.astype(np.float64)
+
+
+def _find_joint_indices(features: dict, state_col: str, n_dim: int) -> tuple[list[int], list[int]]:
+    """Try to find left/right joint indices from info.json feature names."""
+    feat = features.get("observation.state", features.get(state_col, {}))
+    names = _flatten_names(feat.get("names", []))
+
+    left_idx: list[int] = []
+    right_idx: list[int] = []
+    if names and len(names) == n_dim:
+        names_l = [n.lower() for n in names]
+        print(f"  Feature names: {names[:4]}…{names[-4:]}")
+        for j in range(1, 8):
+            for i, nm in enumerate(names_l):
+                if f"left_joint_{j}" in nm and i not in left_idx:
+                    left_idx.append(i)
+                    break
+            for i, nm in enumerate(names_l):
+                if f"right_joint_{j}" in nm and i not in right_idx:
+                    right_idx.append(i)
+                    break
+
+    if len(left_idx) == 7 and len(right_idx) == 7:
+        print(f"  Matched by name: left={left_idx} right={right_idx}")
+        return left_idx, right_idx
+    if n_dim >= 16:
+        print("  Falling back to positional: [0:7]=left, [8:15]=right")
+        return list(range(7)), list(range(8, 15))
+    if n_dim >= 14:
+        print("  Falling back to positional: [0:7]=left, [7:14]=right")
+        return list(range(7)), list(range(7, 14))
+    raise RuntimeError(f"State dim {n_dim} too small for bimanual 7-DOF robot")
+
+
+def download_data(repo_id: str) -> Path:
+    print(f"  Downloading {repo_id} (parquet only) …")
+    return Path(
+        snapshot_download(
+            repo_id=repo_id,
+            repo_type="dataset",
+            allow_patterns=["meta/**", "data/**"],
+            ignore_patterns=["*.mp4", "videos/**"],
+        )
+    )
+
+
+def resample_trajectory(traj: np.ndarray, n_waypoints: int) -> np.ndarray:
+    """Resample a (F, 3) trajectory to exactly n_waypoints via linear interpolation."""
+    f = traj.shape[0]
+    if f == n_waypoints:
+        return traj
+    old_t = np.linspace(0, 1, f)
+    new_t = np.linspace(0, 1, n_waypoints)
+    return np.column_stack([np.interp(new_t, old_t, traj[:, d]) for d in range(3)])
+
+
+def load_episode_trajectories(local: Path) -> list[dict]:
+    """
+    Load per-episode joint data, compute FK, return list of trajectory dicts.
+    Each dict: {"left_tcp": (F,3), "right_tcp": (F,3), "episode_index": int}.
+    Uses all episodes in the dataset for a fair comparison.
+    """
+    info = json.loads((local / "meta" / "info.json").read_text())
+    features = info.get("features", {})
+
+    dfs = [pd.read_parquet(pq) for pq in sorted((local / "data").glob("**/*.parquet"))]
+    df = pd.concat(dfs, ignore_index=True)
+    print(f"  Total frames: {len(df):,}")
+
+    state_col = next((c for c in df.columns if "observation.state" in c), None)
+    if state_col is None:
+        raise RuntimeError(f"No observation.state column. Available: {list(df.columns)}")
+
+    first = df[state_col].iloc[0]
+    if not hasattr(first, "__len__"):
+        raise RuntimeError(f"observation.state is scalar ({type(first)}), expected array")
+
+    state = np.stack(df[state_col].values).astype(np.float64)
+    n_dim = state.shape[1]
+    print(f"  State dim: {n_dim}  max|val|: {np.max(np.abs(state)):.1f}")
+
+    left_idx, right_idx = _find_joint_indices(features, state_col, n_dim)
+
+    ep_col = next((c for c in df.columns if c == "episode_index"), None)
+    if ep_col is None:
+        raise RuntimeError(f"No episode_index column. Available: {list(df.columns)}")
+
+    episode_ids = df[ep_col].values
+    unique_eps = np.unique(episode_ids)
+    print(f"  Episodes: {len(unique_eps):,}")
+
+    left_raw = state[:, left_idx]
+    right_raw = state[:, right_idx]
+    left_all = _detect_and_convert(left_raw)
+    right_all = _detect_and_convert(right_raw)
+
+    print("  Computing FK per episode …")
+    trajectories = []
+    for ep_id in unique_eps:
+        mask = episode_ids == ep_id
+        left_tcp = batch_fk(LEFT_CHAIN, left_all[mask])
+        right_tcp = batch_fk(RIGHT_CHAIN, right_all[mask])
+        if len(left_tcp) < 3:
+            continue
+        trajectories.append({"left_tcp": left_tcp, "right_tcp": right_tcp, "episode_index": int(ep_id)})
+
+    print(f"  Valid trajectories: {len(trajectories):,}")
+    return trajectories
+
+
+# ── Clustering ──────────────────────────────────────────
+
+
+def cluster_trajectories(
+    trajectories: list[dict], n_clusters: int, n_waypoints: int
+) -> tuple[np.ndarray, np.ndarray]:
+    """
+    K-means on resampled trajectory features.
+    Combines left+right TCP into a single feature vector per episode.
+    Returns (labels, centroid_trajs (k, waypoints, 6), spread_per_cluster (k,) in metres).
+    Spread = mean per-waypoint Euclidean distance from each trajectory to its centroid.
+    """
+    feat_vecs = []
+    for t in trajectories:
+        left_rs = resample_trajectory(t["left_tcp"], n_waypoints)
+        right_rs = resample_trajectory(t["right_tcp"], n_waypoints)
+        feat_vecs.append(np.concatenate([left_rs.ravel(), right_rs.ravel()]))
+    feat_matrix = np.array(feat_vecs)
+
+    k = min(n_clusters, len(feat_vecs))
+    km = KMeans(n_clusters=k, n_init=10, random_state=SEED)
+    labels = km.fit_predict(feat_matrix)
+
+    centroids_flat = km.cluster_centers_
+    centroid_trajs = np.zeros((k, n_waypoints, 6))
+    for ci in range(k):
+        left_flat = centroids_flat[ci, : n_waypoints * 3]
+        right_flat = centroids_flat[ci, n_waypoints * 3 :]
+        centroid_trajs[ci, :, :3] = left_flat.reshape(n_waypoints, 3)
+        centroid_trajs[ci, :, 3:] = right_flat.reshape(n_waypoints, 3)
+
+    # Mean per-waypoint distance to centroid (in metres) for each cluster
+    spread = np.zeros(k)
+    for ci in range(k):
+        members = np.where(labels == ci)[0]
+        if len(members) == 0:
+            continue
+        centroid_left = centroid_trajs[ci, :, :3]
+        centroid_right = centroid_trajs[ci, :, 3:]
+        dists = []
+        for mi in members:
+            t = trajectories[mi]
+            left_rs = resample_trajectory(t["left_tcp"], n_waypoints)
+            right_rs = resample_trajectory(t["right_tcp"], n_waypoints)
+            d_left = np.linalg.norm(left_rs - centroid_left, axis=1).mean()
+            d_right = np.linalg.norm(right_rs - centroid_right, axis=1).mean()
+            dists.append((d_left + d_right) / 2)
+        spread[ci] = np.mean(dists)
+
+    return labels, centroid_trajs, spread
+
+
+# ── Visualization ───────────────────────────────────────
+
+PROJ_VIEWS = [
+    ("XZ (side)", 0, 2, "X (m)", "Z (m)"),
+    ("XY (top)", 0, 1, "X (m)", "Y (m)"),
+    ("YZ (front)", 1, 2, "Y (m)", "Z (m)"),
+]
+
+
+def render(results: list[dict], out_path: Path) -> None:
+    """
+    2-row × 3-col grid per dataset (3 projections × 2 datasets).
+    Trajectory lines colored by cluster, centroid trajectories drawn thick.
+    """
+    n_ds = len(results)
+    n_proj = len(PROJ_VIEWS)
+    fig, axes = plt.subplots(n_ds, n_proj, figsize=(7 * n_proj, 7 * n_ds), facecolor="#0d1117")
+    if n_ds == 1:
+        axes = axes[np.newaxis, :]
+
+    for row, r in enumerate(results):
+        trajectories = r["trajectories"]
+        labels = r["labels"]
+        centroids = r["centroids"]
+        k = centroids.shape[0]
+
+        cluster_sizes = np.bincount(labels, minlength=k)
+        size_order = np.argsort(-cluster_sizes)
+        pcts = cluster_sizes / len(labels) * 100
+        spread = r["spread"]
+
+        for col, (view_name, dim_a, dim_b, xlabel, ylabel) in enumerate(PROJ_VIEWS):
+            ax = axes[row, col]
+            ax.set_facecolor("#0d1117")
+
+            for ti, traj in enumerate(trajectories):
+                color = CLUSTER_COLORS[labels[ti] % len(CLUSTER_COLORS)]
+                for tcp_key in ("left_tcp", "right_tcp"):
+                    pts = traj[tcp_key]
+                    ax.plot(pts[:, dim_a], pts[:, dim_b], color=color, alpha=0.12, linewidth=0.4)
+
+            for ci in range(k):
+                color = CLUSTER_COLORS[ci % len(CLUSTER_COLORS)]
+                left_c = centroids[ci, :, :3]
+                right_c = centroids[ci, :, 3:]
+                lw = 1.5 + 2.0 * cluster_sizes[ci] / cluster_sizes.max()
+                for c_pts in (left_c, right_c):
+                    ax.plot(
+                        c_pts[:, dim_a],
+                        c_pts[:, dim_b],
+                        color=color,
+                        linewidth=lw,
+                        alpha=0.95,
+                        zorder=10,
+                    )
+                    ax.plot(
+                        c_pts[0, dim_a],
+                        c_pts[0, dim_b],
+                        "o",
+                        color=color,
+                        markersize=4,
+                        zorder=11,
+                    )
+                    ax.plot(
+                        c_pts[-1, dim_a],
+                        c_pts[-1, dim_b],
+                        "s",
+                        color=color,
+                        markersize=4,
+                        zorder=11,
+                    )
+
+            ax.set_xlabel(xlabel, color="#888", fontsize=9)
+            ax.set_ylabel(ylabel, color="#888", fontsize=9)
+            ax.tick_params(colors="#555", labelsize=7)
+            for spine in ax.spines.values():
+                spine.set_color("#333")
+            ax.set_aspect("equal")
+
+            mean_spread_cm = np.average(spread, weights=cluster_sizes) * 100
+            if col == 0:
+                ax.set_title(
+                    f"{r['label']}  ({r['n_episodes']:,} episodes, {k} clusters, "
+                    f"avg spread {mean_spread_cm:.1f}cm)",
+                    color="white",
+                    fontsize=11,
+                    pad=10,
+                )
+            else:
+                ax.set_title(view_name, color="#aaa", fontsize=10, pad=8)
+
+        # Cluster size + spread legend on the rightmost panel
+        legend_ax = axes[row, -1]
+        for ci in size_order:
+            color = CLUSTER_COLORS[ci % len(CLUSTER_COLORS)]
+            spread_cm = spread[ci] * 100
+            label = f"C{ci}: {cluster_sizes[ci]} eps ({pcts[ci]:.0f}%) ±{spread_cm:.1f}cm"
+            legend_ax.plot([], [], color=color, linewidth=3, label=label)
+        legend_ax.legend(
+            loc="upper right",
+            fontsize=7,
+            frameon=True,
+            facecolor="#1a1a2e",
+            edgecolor="#333",
+            labelcolor="white",
+            handlelength=1.5,
+        )
+
+    fig.suptitle(
+        "End-Effector Trajectory Clusters (FK · K-means)",
+        color="white",
+        fontsize=16,
+        y=0.98,
+    )
+    plt.tight_layout(rect=[0, 0, 1, 0.95])
+    plt.savefig(out_path, dpi=DPI, bbox_inches="tight", facecolor=fig.get_facecolor())
+    plt.close()
+    print(f"\n✓ Saved: {out_path}")
+
+
+# ── Main ────────────────────────────────────────────────
+
+
+def main() -> None:
+    results = []
+
+    for ds in DATASETS:
+        repo_id, label = ds["repo_id"], ds["label"]
+        print(f"\n{'=' * 60}")
+        print(f"  {label}: {repo_id}")
+        print(f"{'=' * 60}")
+
+        local = download_data(repo_id)
+        trajectories = load_episode_trajectories(local)
+        labels, centroids, spread = cluster_trajectories(trajectories, N_CLUSTERS, WAYPOINTS)
+
+        cluster_sizes = np.bincount(labels, minlength=centroids.shape[0])
+        print(f"  Cluster sizes: {sorted(cluster_sizes, reverse=True)}")
+        for ci in np.argsort(-cluster_sizes):
+            print(
+                f"    C{ci}: {cluster_sizes[ci]} eps ({cluster_sizes[ci] / len(labels) * 100:.0f}%) "
+                f"spread ±{spread[ci] * 100:.1f}cm"
+            )
+
+        results.append(
+            {
+                "label": label,
+                "trajectories": trajectories,
+                "labels": labels,
+                "centroids": centroids,
+                "spread": spread,
+                "n_episodes": len(trajectories),
+            }
+        )
+
+    out = OUTPUT_DIR / "workspace_trajectory_clusters.jpg"
+    render(results, out)
+
+
+if __name__ == "__main__":
+    main()
@@ -14,8 +14,8 @@
 # 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.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.processor import make_default_processors
@@ -78,40 +78,24 @@ def main():
    listener, events = init_keyboard_listener()
    init_rerun(session_name="lekiwi_evaluate")

-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")

-    print("Starting evaluate loop...")
-    recorded_episodes = 0
-    while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
+        print("Starting evaluate loop...")
+        recorded_episodes = 0
+        while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")

-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=teleop_action_processor,
-            robot_action_processor=robot_action_processor,
-            robot_observation_processor=robot_observation_processor,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (
-            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-        ):
-            log_say("Reset the environment")
+            # Main record loop
            record_loop(
                robot=robot,
                events=events,
                fps=FPS,
+                policy=policy,
+                preprocessor=preprocessor,  # Pass the pre and post policy processors
+                postprocessor=postprocessor,
+                dataset=dataset,
                control_time_s=EPISODE_TIME_SEC,
                single_task=TASK_DESCRIPTION,
                display_data=True,
@@ -120,24 +104,42 @@ def main():
                robot_observation_processor=robot_observation_processor,
            )

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

-        # Save episode
-        dataset.save_episode()
-        recorded_episodes += 1
+            if events["rerecord_episode"]:
+                log_say("Re-record episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue

-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            recorded_episodes += 1

-    dataset.finalize()
-    dataset.push_to_hub()
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        listener.stop()
+
+        dataset.finalize()
+        dataset.push_to_hub()


 if __name__ == "__main__":
@@ -14,14 +14,14 @@
 # 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.datasets.utils import hw_to_dataset_features
 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.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig
-from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
+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.utils import log_say
@@ -74,40 +74,23 @@ def main():
    listener, events = init_keyboard_listener()
    init_rerun(session_name="lekiwi_record")

-    if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+    try:
+        if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
+            raise ValueError("Robot or teleop is not connected!")

-    print("Starting record loop...")
-    recorded_episodes = 0
-    while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Recording episode {recorded_episodes}")
+        print("Starting record loop...")
+        recorded_episodes = 0
+        while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {recorded_episodes}")

-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            dataset=dataset,
-            teleop=[leader_arm, keyboard],
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=teleop_action_processor,
-            robot_action_processor=robot_action_processor,
-            robot_observation_processor=robot_observation_processor,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (
-            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-        ):
-            log_say("Reset the environment")
+            # Main record loop
            record_loop(
                robot=robot,
                events=events,
                fps=FPS,
+                dataset=dataset,
                teleop=[leader_arm, keyboard],
-                control_time_s=RESET_TIME_SEC,
+                control_time_s=EPISODE_TIME_SEC,
                single_task=TASK_DESCRIPTION,
                display_data=True,
                teleop_action_processor=teleop_action_processor,
@@ -115,26 +98,44 @@ def main():
                robot_observation_processor=robot_observation_processor,
            )

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

-        # Save episode
-        dataset.save_episode()
-        recorded_episodes += 1
+            if events["rerecord_episode"]:
+                log_say("Re-record episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue

-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    leader_arm.disconnect()
-    keyboard.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            recorded_episodes += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        leader_arm.disconnect()
+        keyboard.disconnect()
+        listener.stop()

-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()


 if __name__ == "__main__":
@@ -42,25 +42,27 @@ def main():
    # Connect to the robot
    robot.connect()

-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")

-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+        print("Starting replay loop...")
+        log_say(f"Replaying episode {EPISODE_IDX}")
+        for idx in range(len(episode_frames)):
+            t0 = time.perf_counter()

-        # Get recorded action from dataset
-        action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+            # Get recorded action from dataset
+            action = {
+                name: float(actions[idx][ACTION][i])
+                for i, name in enumerate(dataset.features[ACTION]["names"])
+            }

-        # Send action to robot
-        _ = robot.send_action(action)
+            # Send action to robot
+            _ = robot.send_action(action)

-        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
-
-    robot.disconnect()
+            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+    finally:
+        robot.disconnect()


 if __name__ == "__main__":
@@ -18,7 +18,7 @@ import time

 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.teleoperators.keyboard.teleop_keyboard import KeyboardTeleop, KeyboardTeleopConfig
-from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data

@@ -16,15 +16,13 @@

 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.datasets.utils import combine_feature_dicts
 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.processor import (
-    RobotAction,
-    RobotObservation,
    RobotProcessorPipeline,
    make_default_teleop_action_processor,
 )
@@ -34,13 +32,13 @@ from lerobot.processor.converters import (
    transition_to_observation,
    transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
    ForwardKinematicsJointsToEE,
    InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 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.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
@@ -143,38 +141,24 @@ def main():
    listener, events = init_keyboard_listener()
    init_rerun(session_name="phone_so100_evaluate")

-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")

-    print("Starting evaluate loop...")
-    episode_idx = 0
-    for episode_idx in range(NUM_EPISODES):
-        log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting evaluate loop...")
+        episode_idx = 0
+        for episode_idx in range(NUM_EPISODES):
+            log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")

-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=make_default_teleop_action_processor(),
-            robot_action_processor=robot_ee_to_joints_processor,
-            robot_observation_processor=robot_joints_to_ee_pose_processor,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
-            log_say("Reset the environment")
+            # Main record loop
            record_loop(
                robot=robot,
                events=events,
                fps=FPS,
+                policy=policy,
+                preprocessor=preprocessor,  # Pass the pre and post policy processors
+                postprocessor=postprocessor,
+                dataset=dataset,
                control_time_s=EPISODE_TIME_SEC,
                single_task=TASK_DESCRIPTION,
                display_data=True,
@@ -183,24 +167,41 @@ def main():
                robot_observation_processor=robot_joints_to_ee_pose_processor,
            )

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

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

-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        listener.stop()

-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()


 if __name__ == "__main__":
@@ -15,30 +15,30 @@
 # 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.datasets.utils import combine_feature_dicts
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
    observation_to_transition,
    robot_action_observation_to_transition,
    transition_to_observation,
    transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
    EEBoundsAndSafety,
    EEReferenceAndDelta,
    ForwardKinematicsJointsToEE,
    GripperVelocityToJoint,
    InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, 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.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
@@ -150,38 +150,23 @@ def main():
    listener, events = init_keyboard_listener()
    init_rerun(session_name="phone_so100_record")

-    if not robot.is_connected or not phone.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+    try:
+        if not robot.is_connected or not phone.is_connected:
+            raise ValueError("Robot or teleop is not connected!")

-    print("Starting record loop. Move your phone to teleoperate the robot...")
-    episode_idx = 0
-    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting record loop. Move your phone to teleoperate the robot...")
+        episode_idx = 0
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")

-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            teleop=phone,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=phone_to_robot_ee_pose_processor,
-            robot_action_processor=robot_ee_to_joints_processor,
-            robot_observation_processor=robot_joints_to_ee_pose,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-            log_say("Reset the environment")
+            # Main record loop
            record_loop(
                robot=robot,
                events=events,
                fps=FPS,
                teleop=phone,
-                control_time_s=RESET_TIME_SEC,
+                dataset=dataset,
+                control_time_s=EPISODE_TIME_SEC,
                single_task=TASK_DESCRIPTION,
                display_data=True,
                teleop_action_processor=phone_to_robot_ee_pose_processor,
@@ -189,25 +174,43 @@ def main():
                robot_observation_processor=robot_joints_to_ee_pose,
            )

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

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

-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    phone.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        phone.disconnect()
+        listener.stop()

-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()


 if __name__ == "__main__":
@@ -18,16 +18,16 @@ import time

 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
    robot_action_observation_to_transition,
    transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
    InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
@@ -74,32 +74,34 @@ def main():
    # Connect to the robot
    robot.connect()

-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")

-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+        print("Starting replay loop...")
+        log_say(f"Replaying episode {EPISODE_IDX}")
+        for idx in range(len(episode_frames)):
+            t0 = time.perf_counter()

-        # Get recorded action from dataset
-        ee_action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+            # Get recorded action from dataset
+            ee_action = {
+                name: float(actions[idx][ACTION][i])
+                for i, name in enumerate(dataset.features[ACTION]["names"])
+            }

-        # Get robot observation
-        robot_obs = robot.get_observation()
+            # Get robot observation
+            robot_obs = robot.get_observation()

-        # Dataset EE -> robot joints
-        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+            # Dataset EE -> robot joints
+            joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))

-        # Send action to robot
-        _ = robot.send_action(joint_action)
+            # Send action to robot
+            _ = robot.send_action(joint_action)

-        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
-
-    # Clean up
-    robot.disconnect()
+            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+    finally:
+        # Clean up
+        robot.disconnect()


 if __name__ == "__main__":
@@ -16,22 +16,22 @@
 import time

 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
    robot_action_observation_to_transition,
    transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
    EEBoundsAndSafety,
    EEReferenceAndDelta,
    GripperVelocityToJoint,
    InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, 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

@@ -22,7 +22,8 @@ from pathlib import Path
 import numpy as np
 import tensorflow_datasets as tfds

-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.utils.utils import get_elapsed_time_in_days_hours_minutes_seconds

 DROID_SHARDS = 2048
@@ -26,7 +26,7 @@ from huggingface_hub import HfApi
 from huggingface_hub.constants import REPOCARD_NAME
 from port_droid import DROID_SHARDS

-from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
 from lerobot.datasets.utils import create_lerobot_dataset_card
 from lerobot.utils.utils import init_logging

@@ -155,7 +155,7 @@ class UploadDataset(PipelineStep):
        from datasets.utils.tqdm import disable_progress_bars
        from huggingface_hub import CommitOperationAdd, preupload_lfs_files

-        from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+        from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
        from lerobot.utils.utils import init_logging

        init_logging()
@@ -27,8 +27,8 @@ measuring consistency and ground truth alignment.
 Usage:
    # Basic usage with smolvla policy
    uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
        --rtc.execution_horizon=8 \
        --device=mps \
        --rtc.max_guidance_weight=10.0 \
@@ -58,16 +58,16 @@ Usage:
        --device=cuda

    uv run python examples/rtc/eval_dataset.py \
-        --policy.path=lipsop/reuben_pi0 \
-        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
+        --policy.path=<USER>/reuben_pi0 \
+        --dataset.repo_id=<USER>/so101_cube_in_cup \
        --rtc.execution_horizon=8 \
        --device=cuda

    # With torch.compile for faster inference (PyTorch 2.0+)
    # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
    uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
        --rtc.execution_horizon=8 \
        --device=mps \
        --use_torch_compile=true \
@@ -75,8 +75,8 @@ Usage:

    # With torch.compile on CUDA (CUDA graphs disabled by default)
    uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
        --rtc.execution_horizon=8 \
        --device=cuda \
        --use_torch_compile=true \
@@ -84,8 +84,8 @@ Usage:

    # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
    uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
        --use_torch_compile=true \
        --torch_compile_backend=inductor \
        --torch_compile_mode=max-autotune \
@@ -113,8 +113,9 @@ 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, LeRobotDatasetMetadata
+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.policies.rtc.debug_visualizer import RTCDebugVisualizer
@@ -28,7 +28,7 @@ For simulation environments, see eval_with_simulation.py
 Usage:
    # Run RTC with Real robot with RTC
    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
        --policy.device=mps \
        --rtc.enabled=true \
        --rtc.execution_horizon=20 \
@@ -41,7 +41,7 @@ Usage:

    # Run RTC with Real robot without RTC
    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
        --policy.device=mps \
        --rtc.enabled=false \
        --robot.type=so100_follower \
@@ -53,7 +53,7 @@ Usage:

    # Run RTC with Real robot with pi0.5 policy
    uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=helper2424/pi05_check_rtc \
+        --policy.path=<USER>/pi05_check_rtc \
        --policy.device=mps \
        --rtc.enabled=true \
        --rtc.execution_horizon=20 \
@@ -78,10 +78,11 @@ 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.utils import build_dataset_frame, hw_to_dataset_features
+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.policies.rtc.action_queue import ActionQueue
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
@@ -94,9 +95,10 @@ from lerobot.rl.process import ProcessSignalHandler
 from lerobot.robots import (  # noqa: F401
    Robot,
    RobotConfig,
+    bi_so_follower,
    koch_follower,
-    so100_follower,
-    so101_follower,
+    so_follower,
+    unitree_g1,
 )
 from lerobot.robots.utils import make_robot_from_config
 from lerobot.utils.constants import OBS_IMAGES
@@ -16,15 +16,13 @@

 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.datasets.utils import combine_feature_dicts
 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.processor import (
-    RobotAction,
-    RobotObservation,
    RobotProcessorPipeline,
    make_default_teleop_action_processor,
 )
@@ -34,13 +32,13 @@ from lerobot.processor.converters import (
    transition_to_observation,
    transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
    ForwardKinematicsJointsToEE,
    InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 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.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
@@ -143,38 +141,24 @@ def main():
    listener, events = init_keyboard_listener()
    init_rerun(session_name="so100_so100_evaluate")

-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")

-    print("Starting evaluate loop...")
-    episode_idx = 0
-    for episode_idx in range(NUM_EPISODES):
-        log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting evaluate loop...")
+        episode_idx = 0
+        for episode_idx in range(NUM_EPISODES):
+            log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")

-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=make_default_teleop_action_processor(),
-            robot_action_processor=robot_ee_to_joints_processor,
-            robot_observation_processor=robot_joints_to_ee_pose_processor,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
-            log_say("Reset the environment")
+            # Main record loop
            record_loop(
                robot=robot,
                events=events,
                fps=FPS,
+                policy=policy,
+                preprocessor=preprocessor,  # Pass the pre and post policy processors
+                postprocessor=postprocessor,
+                dataset=dataset,
                control_time_s=EPISODE_TIME_SEC,
                single_task=TASK_DESCRIPTION,
                display_data=True,
@@ -183,24 +167,41 @@ def main():
                robot_observation_processor=robot_joints_to_ee_pose_processor,
            )

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

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

-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        listener.stop()

-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()


 if __name__ == "__main__":
@@ -16,27 +16,26 @@


 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.datasets.utils import combine_feature_dicts
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
    observation_to_transition,
    robot_action_observation_to_transition,
    transition_to_observation,
    transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
    EEBoundsAndSafety,
    ForwardKinematicsJointsToEE,
    InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.scripts.lerobot_record import record_loop
-from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
-from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
+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.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
@@ -148,38 +147,23 @@ def main():
    listener, events = init_keyboard_listener()
    init_rerun(session_name="recording_phone")

-    if not leader.is_connected or not follower.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+    try:
+        if not leader.is_connected or not follower.is_connected:
+            raise ValueError("Robot or teleop is not connected!")

-    print("Starting record loop...")
-    episode_idx = 0
-    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting record loop...")
+        episode_idx = 0
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")

-        # Main record loop
-        record_loop(
-            robot=follower,
-            events=events,
-            fps=FPS,
-            teleop=leader,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=leader_joints_to_ee,
-            robot_action_processor=ee_to_follower_joints,
-            robot_observation_processor=follower_joints_to_ee,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-            log_say("Reset the environment")
+            # Main record loop
            record_loop(
                robot=follower,
                events=events,
                fps=FPS,
                teleop=leader,
-                control_time_s=RESET_TIME_SEC,
+                dataset=dataset,
+                control_time_s=EPISODE_TIME_SEC,
                single_task=TASK_DESCRIPTION,
                display_data=True,
                teleop_action_processor=leader_joints_to_ee,
@@ -187,25 +171,44 @@ def main():
                robot_observation_processor=follower_joints_to_ee,
            )

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

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

-    # Clean up
-    log_say("Stop recording")
-    leader.disconnect()
-    follower.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1

-    dataset.finalize()
-    dataset.push_to_hub()
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        leader.disconnect()
+        follower.disconnect()
+        listener.stop()
+
+        dataset.finalize()
+        dataset.push_to_hub()


 if __name__ == "__main__":
@@ -19,16 +19,16 @@ import time

 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
    robot_action_observation_to_transition,
    transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
    InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
@@ -75,32 +75,35 @@ def main():
    # Connect to the robot
    robot.connect()

-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")

-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+        print("Starting replay loop...")
+        log_say(f"Replaying episode {EPISODE_IDX}")
+        for idx in range(len(episode_frames)):
+            t0 = time.perf_counter()

-        # Get recorded action from dataset
-        ee_action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+            # Get recorded action from dataset
+            ee_action = {
+                name: float(actions[idx][ACTION][i])
+                for i, name in enumerate(dataset.features[ACTION]["names"])
+            }

-        # Get robot observation
-        robot_obs = robot.get_observation()
+            # Get robot observation
+            robot_obs = robot.get_observation()

-        # Dataset EE -> robot joints
-        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+            # Dataset EE -> robot joints
+            joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))

-        # Send action to robot
-        _ = robot.send_action(joint_action)
+            # Send action to robot
+            _ = robot.send_action(joint_action)

-        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
+            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))

-    # Clean up
-    robot.disconnect()
+    finally:
+        # Clean up
+        robot.disconnect()


 if __name__ == "__main__":
@@ -17,21 +17,20 @@
 import time

 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
    robot_action_observation_to_transition,
    robot_action_to_transition,
    transition_to_robot_action,
 )
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.robot_kinematic_processor import (
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.robots.so_follower.robot_kinematic_processor import (
    EEBoundsAndSafety,
    ForwardKinematicsJointsToEE,
    InverseKinematicsEEToJoints,
 )
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
-from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
+from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
+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

@@ -19,8 +19,9 @@ from pathlib import Path
 import torch

 from lerobot.configs.types import FeatureType
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.datasets.utils import dataset_to_policy_features
+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
@@ -20,9 +20,9 @@ from pathlib import Path
 import torch

 from lerobot.configs.types import FeatureType
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+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.datasets.utils import dataset_to_policy_features
 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
@@ -5,8 +5,9 @@ from pathlib import Path
 import torch

 from lerobot.configs.types import FeatureType
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.datasets.utils import dataset_to_policy_features
+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
@@ -1,12 +1,11 @@
 import torch

 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+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.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig

 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20
@@ -4,7 +4,7 @@ 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.robots.so100_follower import SO100FollowerConfig
+from lerobot.robots.so_follower import SO100FollowerConfig


 def main():
@@ -30,6 +30,7 @@ def main():
        robot=robot_cfg,
        server_address=server_address,
        policy_device="mps",
+        client_device="cpu",
        policy_type="act",
        pretrained_name_or_path="<user>/robot_learning_tutorial_act",
        chunk_size_threshold=0.5,  # g
@@ -5,8 +5,9 @@ from pathlib import Path
 import torch

 from lerobot.configs.types import FeatureType
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.datasets.utils import dataset_to_policy_features
+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
@@ -1,12 +1,11 @@
 import torch

 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+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.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig

 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20
@@ -1,12 +1,11 @@
 import torch

 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.utils import hw_to_dataset_features
+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.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig

 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20
@@ -6,16 +6,16 @@ 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.utils import hw_to_dataset_features
 from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
 from lerobot.policies.sac.configuration_sac 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.so100_follower import SO100FollowerConfig
-from lerobot.teleoperators.so100_leader import SO100LeaderConfig
+from lerobot.robots.so_follower import SO100FollowerConfig
+from lerobot.teleoperators.so_leader import SO100LeaderConfig
 from lerobot.teleoperators.utils import TeleopEvents

 LOG_EVERY = 10
@@ -1,12 +1,11 @@
 import torch

 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.utils import hw_to_dataset_features
+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.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig

 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20
@@ -1,347 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""
-Example: GR00T Locomotion with Pre-loaded Policies
-
-This example demonstrates the NEW pattern for loading GR00T policies externally
-and passing them to the robot class.
-"""
-
-import argparse
-import logging
-import threading
-import time
-from collections import deque
-
-import numpy as np
-import onnxruntime as ort
-from huggingface_hub import hf_hub_download
-
-from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
-from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
-
-logger = logging.getLogger(__name__)
-
-GROOT_DEFAULT_ANGLES = np.zeros(29, dtype=np.float32)
-GROOT_DEFAULT_ANGLES[[0, 6]] = -0.1  # hip pitch
-GROOT_DEFAULT_ANGLES[[3, 9]] = 0.3  # knee
-GROOT_DEFAULT_ANGLES[[4, 10]] = -0.2  # ankle pitch
-
-MISSING_JOINTS = []
-G1_MODEL = "g1_23"  # or "g1_29"
-if G1_MODEL == "g1_23":
-    MISSING_JOINTS = [12, 14, 20, 21, 27, 28]  # waist yaw/pitch, wrist pitch/yaw
-
-LOCOMOTION_ACTION_SCALE = 0.25
-
-LOCOMOTION_CONTROL_DT = 0.02
-
-ANG_VEL_SCALE: float = 0.25
-DOF_POS_SCALE: float = 1.0
-DOF_VEL_SCALE: float = 0.05
-CMD_SCALE: list = [2.0, 2.0, 0.25]
-
-
-DEFAULT_GROOT_REPO_ID = "nepyope/GR00T-WholeBodyControl_g1"
-
-
-def load_groot_policies(
-    repo_id: str = DEFAULT_GROOT_REPO_ID,
-) -> tuple[ort.InferenceSession, ort.InferenceSession]:
-    """Load GR00T dual-policy system (Balance + Walk) from Hugging Face Hub.
-
-    Args:
-        repo_id: Hugging Face Hub repository ID containing the ONNX policies.
-    """
-    logger.info(f"Loading GR00T dual-policy system from Hugging Face Hub ({repo_id})...")
-
-    # Download ONNX policies from Hugging Face Hub
-    balance_path = hf_hub_download(
-        repo_id=repo_id,
-        filename="GR00T-WholeBodyControl-Balance.onnx",
-    )
-    walk_path = hf_hub_download(
-        repo_id=repo_id,
-        filename="GR00T-WholeBodyControl-Walk.onnx",
-    )
-
-    # Load ONNX policies
-    policy_balance = ort.InferenceSession(balance_path)
-    policy_walk = ort.InferenceSession(walk_path)
-
-    logger.info("GR00T policies loaded successfully")
-
-    return policy_balance, policy_walk
-
-
-class GrootLocomotionController:
-    """
-    Handles GR00T-style locomotion control for the Unitree G1 robot.
-
-    This controller manages:
-    - Dual-policy system (Balance + Walk)
-    - 29-joint observation processing
-    - 15D action output (legs + waist)
-    - Policy inference and motor command generation
-    """
-
-    def __init__(self, policy_balance, policy_walk, robot, config):
-        self.policy_balance = policy_balance
-        self.policy_walk = policy_walk
-        self.robot = robot
-        self.config = config
-
-        self.locomotion_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)  # vx, vy, theta_dot
-
-        # GR00T-specific state
-        self.groot_qj_all = np.zeros(29, dtype=np.float32)
-        self.groot_dqj_all = np.zeros(29, dtype=np.float32)
-        self.groot_action = np.zeros(15, dtype=np.float32)
-        self.groot_obs_single = np.zeros(86, dtype=np.float32)
-        self.groot_obs_history = deque(maxlen=6)
-        self.groot_obs_stacked = np.zeros(516, dtype=np.float32)
-        self.groot_height_cmd = 0.74  # Default base height
-        self.groot_orientation_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)
-
-        # input to gr00t is 6 frames (6*86D=516)
-        for _ in range(6):
-            self.groot_obs_history.append(np.zeros(86, dtype=np.float32))
-
-        # Thread management
-        self.locomotion_running = False
-        self.locomotion_thread = None
-
-        logger.info("GrootLocomotionController initialized")
-
-    def groot_locomotion_run(self):
-        # get current observation
-        robot_state = self.robot.get_observation()
-
-        if robot_state is None:
-            return
-
-        # get command from remote controller
-        if robot_state.wireless_remote is not None:
-            self.robot.remote_controller.set(robot_state.wireless_remote)
-            if self.robot.remote_controller.button[0]:  # R1 - raise waist
-                self.groot_height_cmd += 0.001
-                self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
-            if self.robot.remote_controller.button[4]:  # R2 - lower waist
-                self.groot_height_cmd -= 0.001
-                self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
-        else:
-            self.robot.remote_controller.lx = 0.0
-            self.robot.remote_controller.ly = 0.0
-            self.robot.remote_controller.rx = 0.0
-            self.robot.remote_controller.ry = 0.0
-
-        self.locomotion_cmd[0] = self.robot.remote_controller.ly  # forward/backward
-        self.locomotion_cmd[1] = self.robot.remote_controller.lx * -1  # left/right
-        self.locomotion_cmd[2] = self.robot.remote_controller.rx * -1  # rotation rate
-
-        for i in range(29):
-            self.groot_qj_all[i] = robot_state.motor_state[i].q
-            self.groot_dqj_all[i] = robot_state.motor_state[i].dq
-
-        # adapt observation for g1_23dof
-        for idx in MISSING_JOINTS:
-            self.groot_qj_all[idx] = 0.0
-            self.groot_dqj_all[idx] = 0.0
-
-        # Scale joint positions and velocities
-        qj_obs = self.groot_qj_all.copy()
-        dqj_obs = self.groot_dqj_all.copy()
-
-        # express imu data in gravity frame of reference
-        quat = robot_state.imu_state.quaternion
-        ang_vel = np.array(robot_state.imu_state.gyroscope, dtype=np.float32)
-        gravity_orientation = self.robot.get_gravity_orientation(quat)
-
-        # scale joint positions and velocities before policy inference
-        qj_obs = (qj_obs - GROOT_DEFAULT_ANGLES) * DOF_POS_SCALE
-        dqj_obs = dqj_obs * DOF_VEL_SCALE
-        ang_vel_scaled = ang_vel * ANG_VEL_SCALE
-
-        # build single frame observation
-        self.groot_obs_single[:3] = self.locomotion_cmd * np.array(CMD_SCALE)
-        self.groot_obs_single[3] = self.groot_height_cmd
-        self.groot_obs_single[4:7] = self.groot_orientation_cmd
-        self.groot_obs_single[7:10] = ang_vel_scaled
-        self.groot_obs_single[10:13] = gravity_orientation
-        self.groot_obs_single[13:42] = qj_obs
-        self.groot_obs_single[42:71] = dqj_obs
-        self.groot_obs_single[71:86] = self.groot_action  # 15D previous actions
-
-        # Add to history and stack observations (6 frames × 86D = 516D)
-        self.groot_obs_history.append(self.groot_obs_single.copy())
-
-        # Stack all 6 frames into 516D vector
-        for i, obs_frame in enumerate(self.groot_obs_history):
-            start_idx = i * 86
-            end_idx = start_idx + 86
-            self.groot_obs_stacked[start_idx:end_idx] = obs_frame
-
-        # Run policy inference (ONNX) with 516D stacked observation
-
-        cmd_magnitude = np.linalg.norm(self.locomotion_cmd)
-
-        selected_policy = (
-            self.policy_balance if cmd_magnitude < 0.05 else self.policy_walk
-        )  # balance/standing policy for small commands, walking policy for movement commands
-
-        # run policy inference
-        ort_inputs = {selected_policy.get_inputs()[0].name: np.expand_dims(self.groot_obs_stacked, axis=0)}
-        ort_outs = selected_policy.run(None, ort_inputs)
-        self.groot_action = ort_outs[0].squeeze()
-
-        # transform action back to target joint positions
-        target_dof_pos_15 = GROOT_DEFAULT_ANGLES[:15] + self.groot_action * LOCOMOTION_ACTION_SCALE
-
-        # command motors
-        for i in range(15):
-            motor_idx = i
-            self.robot.msg.motor_cmd[motor_idx].q = target_dof_pos_15[i]
-            self.robot.msg.motor_cmd[motor_idx].qd = 0
-            self.robot.msg.motor_cmd[motor_idx].kp = self.robot.kp[motor_idx]
-            self.robot.msg.motor_cmd[motor_idx].kd = self.robot.kd[motor_idx]
-            self.robot.msg.motor_cmd[motor_idx].tau = 0
-
-        # adapt action for g1_23dof
-        for joint_idx in MISSING_JOINTS:
-            self.robot.msg.motor_cmd[joint_idx].q = 0.0
-            self.robot.msg.motor_cmd[joint_idx].qd = 0
-            self.robot.msg.motor_cmd[joint_idx].kp = self.robot.kp[joint_idx]
-            self.robot.msg.motor_cmd[joint_idx].kd = self.robot.kd[joint_idx]
-            self.robot.msg.motor_cmd[joint_idx].tau = 0
-
-        # send action to robot
-        self.robot.send_action(self.robot.msg)
-
-    def _locomotion_thread_loop(self):
-        """Background thread that runs the locomotion policy at specified rate."""
-        logger.info("Locomotion thread started")
-        while self.locomotion_running:
-            start_time = time.time()
-            try:
-                self.groot_locomotion_run()
-            except Exception as e:
-                logger.error(f"Error in locomotion loop: {e}")
-
-            # Sleep to maintain control rate
-            elapsed = time.time() - start_time
-            sleep_time = max(0, LOCOMOTION_CONTROL_DT - elapsed)
-            time.sleep(sleep_time)
-        logger.info("Locomotion thread stopped")
-
-    def start_locomotion_thread(self):
-        if self.locomotion_running:
-            logger.warning("Locomotion thread already running")
-            return
-
-        logger.info("Starting locomotion control thread...")
-        self.locomotion_running = True
-        self.locomotion_thread = threading.Thread(target=self._locomotion_thread_loop, daemon=True)
-        self.locomotion_thread.start()
-
-        logger.info("Locomotion control thread started!")
-
-    def stop_locomotion_thread(self):
-        if not self.locomotion_running:
-            return
-
-        logger.info("Stopping locomotion control thread...")
-        self.locomotion_running = False
-        if self.locomotion_thread:
-            self.locomotion_thread.join(timeout=2.0)
-        logger.info("Locomotion control thread stopped")
-
-    def reset_robot(self):
-        """Move robot legs to default standing position over 2 seconds (arms are not moved)."""
-        total_time = 3.0
-        num_step = int(total_time / self.robot.control_dt)
-
-        # Only control legs, not arms (first 12 joints)
-        default_pos = GROOT_DEFAULT_ANGLES  # First 12 values are leg angles
-        dof_size = len(default_pos)
-
-        # Get current lowstate
-        robot_state = self.robot.get_observation()
-
-        # Record the current leg positions
-        init_dof_pos = np.zeros(dof_size, dtype=np.float32)
-        for i in range(dof_size):
-            init_dof_pos[i] = robot_state.motor_state[i].q
-
-        # Move legs to default pos
-        for i in range(num_step):
-            alpha = i / num_step
-            for motor_idx in range(dof_size):
-                target_pos = default_pos[motor_idx]
-                self.robot.msg.motor_cmd[motor_idx].q = (
-                    init_dof_pos[motor_idx] * (1 - alpha) + target_pos * alpha
-                )
-                self.robot.msg.motor_cmd[motor_idx].qd = 0
-                self.robot.msg.motor_cmd[motor_idx].kp = self.robot.kp[motor_idx]
-                self.robot.msg.motor_cmd[motor_idx].kd = self.robot.kd[motor_idx]
-                self.robot.msg.motor_cmd[motor_idx].tau = 0
-            self.robot.msg.crc = self.robot.crc.Crc(self.robot.msg)
-            self.robot.lowcmd_publisher.Write(self.robot.msg)
-            time.sleep(self.robot.control_dt)
-        logger.info("Reached default position (legs only)")
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="GR00T Locomotion Controller for Unitree G1")
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        default=DEFAULT_GROOT_REPO_ID,
-        help=f"Hugging Face Hub repo ID for GR00T policies (default: {DEFAULT_GROOT_REPO_ID})",
-    )
-    args = parser.parse_args()
-
-    # load policies
-    policy_balance, policy_walk = load_groot_policies(repo_id=args.repo_id)
-
-    # initialize robot
-    config = UnitreeG1Config()
-    robot = UnitreeG1(config)
-
-    # initialize gr00t locomotion controller
-    groot_controller = GrootLocomotionController(
-        policy_balance=policy_balance,
-        policy_walk=policy_walk,
-        robot=robot,
-        config=config,
-    )
-
-    # reset legs and start locomotion thread
-    try:
-        groot_controller.reset_robot()
-        groot_controller.start_locomotion_thread()
-
-        # log status
-        logger.info("Robot initialized with GR00T locomotion policies")
-        logger.info("Locomotion controller running in background thread")
-        logger.info("Press Ctrl+C to stop")
-
-        # keep robot alive
-        while True:
-            time.sleep(1.0)
-    except KeyboardInterrupt:
-        print("\nStopping locomotion...")
-        groot_controller.stop_locomotion_thread()
-        print("Done!")
@@ -25,11 +25,11 @@ discord = "https://discord.gg/s3KuuzsPFb"

 [project]
 name = "lerobot"
-version = "0.4.3"
+version = "0.5.1"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
-readme = "README.md"
+dynamic = ["readme"]
 license = { text = "Apache-2.0" }
-requires-python = ">=3.10"
+requires-python = ">=3.12"
 authors = [
    { name = "Rémi Cadène", email = "re.cadene@gmail.com" },
    { name = "Simon Alibert", email = "alibert.sim@gmail.com" },
@@ -50,7 +50,8 @@ classifiers = [
    "Intended Audience :: Education",
    "Intended Audience :: Science/Research",
    "License :: OSI Approved :: Apache Software License",
-    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
    "Topic :: Software Development :: Build Tools",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
@@ -59,28 +60,30 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [

    # Hugging Face dependencies
-    "datasets>=4.0.0,<4.2.0",
+    "datasets>=4.0.0,<5.0.0",
    "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub[hf-transfer,cli]>=0.34.2,<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",
+
+    "torch>=2.2.1,<2.11.0",
+    "torchcodec>=0.2.1,<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')",
+    "torchvision>=0.21.0,<0.26.0",
+
    "einops>=0.8.0,<0.9.0",
-    "opencv-python-headless>=4.9.0,<4.13.0",
+    "opencv-python-headless>=4.9.0,<4.14.0",
    "av>=15.0.0,<16.0.0",
    "jsonlines>=4.0.0,<5.0.0",
-    "packaging>=24.2,<26.0",
-    "pynput>=1.7.7,<1.9.0",
+    "pynput>=1.7.8,<1.9.0",
    "pyserial>=3.5,<4.0",
-    "wandb>=0.20.0,<0.22.0", # TODO: Bumb dependency (compatible with protobuf)

-    "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
-    "torchcodec>=0.2.1,<0.6.0; 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')", # TODO: Bumb dependency
-    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
-
-    "draccus==0.10.0", # TODO: Remove ==
+    "wandb>=0.24.0,<0.25.0",
+    "draccus==0.10.0", # TODO: Relax version constraint
    "gymnasium>=1.1.1,<2.0.0",
    "rerun-sdk>=0.24.0,<0.27.0",

@@ -95,43 +98,56 @@ dependencies = [

 # Common
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-placo-dep = ["placo>=0.9.6,<0.10.0"]
-transformers-dep = ["transformers>=4.57.1,<5.0.0"]
-grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"] # TODO: Bumb dependency (compatible with wandb)
+placo-dep = ["placo>=0.9.6,<0.9.17"]
+transformers-dep = ["transformers>=5.3.0,<6.0.0"]
+grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
+can-dep = ["python-can>=4.2.0,<5.0.0"]
+peft-dep = ["peft>=0.18.0,<1.0.0"]
+scipy-dep = ["scipy>=1.14.0,<2.0.0"]
+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.

 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
+damiao = ["lerobot[can-dep]"]
+robstride = ["lerobot[can-dep]"]

 # Robots
+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"]
 unitree_g1 = [
+    # "unitree-sdk2==1.0.1",
    "pyzmq>=26.2.1,<28.0.0",
-    "onnxruntime>=1.16.0"
+    "onnxruntime>=1.16.0,<2.0.0",
+    "onnx>=1.16.0,<2.0.0",
+    "meshcat>=0.3.0,<0.4.0",
+    "lerobot[matplotlib-dep]",
+    "lerobot[pygame-dep]",
 ]
-reachy2 = ["reachy2_sdk>=1.0.14,<1.1.0"]
+reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
    "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
-    "pyrealsense2-macosx>=2.54,<2.55.0 ; sys_platform == 'darwin'",
+    "pyrealsense2-macosx>=2.54,<2.57.0 ; sys_platform == 'darwin'",
 ]
-phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0", "lerobot[scipy-dep]"]

 # Policies
 wallx = [
-    "transformers==4.49.0",
-    "peft==0.17.1",
-    "scipy==1.15.3",
-    "torchdiffeq==0.2.5",
-    "qwen_vl_utils==0.0.11"
+    "lerobot[transformers-dep]",
+    "lerobot[peft]",
+    "lerobot[scipy-dep]",
+    "torchdiffeq>=0.2.4,<0.3.0",
+    "lerobot[qwen-vl-utils-dep]",
 ]
-pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi"]
+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"]
 groot = [
    "lerobot[transformers-dep]",
-    "peft>=0.13.0,<1.0.0",
+    "lerobot[peft]",
    "dm-tree>=0.1.8,<1.0.0",
    "timm>=1.0.0,<1.1.0",
    "safetensors>=0.4.3,<1.0.0",
@@ -140,13 +156,13 @@ groot = [
    "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", "matplotlib>=3.10.3,<4.0.0", "qwen-vl-utils>=0.0.14"]
+sarm = ["lerobot[transformers-dep]", "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]"]

 # Features
-async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
-peft = ["lerobot[transformers-dep]", "peft>=0.18.0"]
+async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
+peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]

 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1"]
@@ -154,13 +170,19 @@ test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]

 # Simulation
-aloha = ["gym-aloha>=0.1.2,<0.2.0"]
+# 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"]
-metaworld = ["metaworld==3.0.0"]
+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]"]

 # All
 all = [
+    # 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[gamepad]",
    "lerobot[hopejr]",
@@ -168,7 +190,7 @@ all = [
    "lerobot[reachy2]",
    "lerobot[kinematics]",
    "lerobot[intelrealsense]",
-    # "lerobot[wallx]",
+    "lerobot[wallx]",
    "lerobot[pi]",
    "lerobot[smolvla]",
    # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
@@ -181,10 +203,11 @@ all = [
    "lerobot[aloha]",
    "lerobot[pusht]",
    "lerobot[phone]",
-    "lerobot[libero]",
+    "lerobot[libero]; sys_platform == 'linux'",
    "lerobot[metaworld]",
    "lerobot[sarm]",
    "lerobot[peft]",
+    # "lerobot[unitree_g1]", TODO: Unitree requires specific installation instructions for unitree_sdk2
 ]

 [project.scripts]
@@ -197,18 +220,23 @@ lerobot-setup-motors="lerobot.scripts.lerobot_setup_motors:main"
 lerobot-teleoperate="lerobot.scripts.lerobot_teleoperate:main"
 lerobot-eval="lerobot.scripts.lerobot_eval:main"
 lerobot-train="lerobot.scripts.lerobot_train:main"
+lerobot-train-tokenizer="lerobot.scripts.lerobot_train_tokenizer:main"
 lerobot-dataset-viz="lerobot.scripts.lerobot_dataset_viz:main"
 lerobot-info="lerobot.scripts.lerobot_info:main"
 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"

 # ---------------- Tool Configurations ----------------
+[tool.setuptools.package-data]
+lerobot = ["envs/*.json"]
+
 [tool.setuptools.packages.find]
 where = ["src"]

 [tool.ruff]
-target-version = "py310"
+target-version = "py312"
 line-length = 110
 exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]

@@ -277,6 +305,7 @@ default.extend-ignore-identifiers-re = [
    "thw",
    "inpt",
    "ROBOTIS",
+    "OT_VALUE"
 ]

 # TODO: Uncomment when ready to use
@@ -299,7 +328,7 @@ default.extend-ignore-identifiers-re = [
 # Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations

 [tool.mypy]
-python_version = "3.10"
+python_version = "3.12"
 ignore_missing_imports = true
 follow_imports = "skip"
 # warn_return_any = true
@@ -351,9 +380,9 @@ ignore_errors = false
 module = "lerobot.cameras.*"
 ignore_errors = false

-# [[tool.mypy.overrides]]
-# module = "lerobot.motors.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.motors.*"
+ignore_errors = false

 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"
@@ -383,40 +412,3 @@ ignore_errors = false
 # [[tool.mypy.overrides]]
 # module = "lerobot.scripts.*"
 # ignore_errors = false
-
-[tool.uv]
-# wallx requires transformers==4.49.0 which conflicts with other extras that need >=4.53.0
-conflicts = [
-    [
-        { extra = "wallx" },
-        { extra = "transformers-dep" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "pi" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "smolvla" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "groot" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "xvla" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "hilserl" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "libero" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "all" },
-    ],
-]
@@ -1,76 +1,73 @@
 #
-# This file is autogenerated by pip-compile with Python 3.10
+# This file is autogenerated by pip-compile with Python 3.12
 # by the following command:
 #
 #    pip-compile --output-file=requirements-macos.txt requirements.in
 #
 -e .[all]
    # via -[all]
-absl-py==2.3.1
+absl-py==2.4.0
    # via
    #   dm-control
    #   dm-env
    #   dm-tree
    #   labmaze
    #   mujoco
-    #   tensorboard
-accelerate==1.11.0
+accelerate==1.13.0
    # via
    #   lerobot
    #   peft
 aiohappyeyeballs==2.6.1
    # via aiohttp
-aiohttp==3.13.1
+aiohttp==3.13.3
    # via fsspec
 aiosignal==1.4.0
    # via aiohttp
+annotated-doc==0.0.4
+    # via
+    #   fastapi
+    #   typer
 annotated-types==0.7.0
    # via pydantic
-antlr4-python3-runtime==4.9.3
-    # via
-    #   hydra-core
-    #   omegaconf
-anyio==4.11.0
+anyio==4.12.1
    # via
+    #   httpx
    #   starlette
    #   watchfiles
-asttokens==3.0.0
+asttokens==3.0.1
    # via stack-data
-async-timeout==5.0.1
-    # via aiohttp
 attrs==25.4.0
    # via
    #   aiohttp
    #   dm-tree
    #   jsonlines
-    #   jsonschema
-    #   referencing
    #   rerun-sdk
 av==15.1.0
-    # via lerobot
-bddl==1.0.1
-    # via libero
-certifi==2025.10.5
    # via
+    #   lerobot
+    #   qwen-vl-utils
+certifi==2026.2.25
+    # via
+    #   httpcore
+    #   httpx
    #   requests
    #   sentry-sdk
 cffi==2.0.0
    # via pymunk
-cfgv==3.4.0
+cfgv==3.5.0
    # via pre-commit
-charset-normalizer==3.4.4
+charset-normalizer==3.4.5
    # via requests
-click==8.3.0
+click==8.3.1
    # via
+    #   typer
    #   uvicorn
    #   wandb
-cloudpickle==3.1.1
-    # via
-    #   gymnasium
-    #   libero
-cmake==4.1.0
+cloudpickle==3.1.2
+    # via gymnasium
+cmake==4.1.3
    # via lerobot
-cmeel==0.57.3
+cmeel==0.59.0
    # via
    #   cmeel-assimp
    #   cmeel-boost
@@ -108,15 +105,17 @@ cmeel-zlib==1.3.1
    # via cmeel-assimp
 coal-library==3.0.1
    # via pin
-contourpy==1.3.2
-    # via matplotlib
-coverage[toml]==7.11.0
+contourpy==1.3.3
+    # via
+    #   lerobot
+    #   matplotlib
+coverage[toml]==7.13.4
    # via pytest-cov
 cycler==0.12.1
    # via matplotlib
-datasets==4.1.1
+datasets==4.6.1
    # via lerobot
-debugpy==1.8.17
+debugpy==1.8.20
    # via lerobot
 decorator==5.2.1
    # via ipython
@@ -130,7 +129,7 @@ dill==0.4.0
    #   multiprocess
 distlib==0.4.0
    # via virtualenv
-dm-control==1.0.34
+dm-control==1.0.37
    # via gym-aloha
 dm-env==1.6
    # via dm-control
@@ -138,69 +137,55 @@ dm-tree==0.1.9
    # via
    #   dm-control
    #   dm-env
-    #   lerobot
 docopt==0.6.2
    # via num2words
 draccus==0.10.0
    # via lerobot
 dynamixel-sdk==3.8.4
    # via lerobot
-easydict==1.13
-    # via libero
-egl-probe @ git+https://github.com/huggingface/egl_probe.git
-    # via
-    #   libero
-    #   robomimic
 eigenpy==3.10.3
    # via coal-library
-einops==0.8.1
-    # via
-    #   lerobot
-    #   libero
+einops==0.8.2
+    # via lerobot
 eiquadprog==1.2.9
    # via placo
-etils[epath,epy]==1.13.0
+etils[epath,epy]==1.14.0
    # via mujoco
-exceptiongroup==1.3.0
-    # via
-    #   anyio
-    #   ipython
-    #   pytest
 executing==2.2.1
    # via stack-data
+faker==34.0.2
+    # via lerobot
 farama-notifications==0.0.4
    # via gymnasium
-fastapi==0.119.1
-    # via teleop
-fastjsonschema==2.21.2
-    # via nbformat
+fastapi==0.135.1
+    # via
+    #   lerobot
+    #   teleop
 feetech-servo-sdk==1.0.0
    # via lerobot
-filelock==3.20.0
+filelock==3.25.0
    # via
    #   datasets
    #   diffusers
    #   huggingface-hub
+    #   python-discovery
    #   torch
-    #   transformers
    #   virtualenv
-fonttools==4.60.1
+fonttools==4.61.1
    # via matplotlib
 frozenlist==1.8.0
    # via
    #   aiohttp
    #   aiosignal
-fsspec[http]==2025.9.0
+fsspec[http]==2026.2.0
    # via
    #   datasets
    #   etils
    #   huggingface-hub
    #   torch
-future==1.0.0
-    # via libero
 gitdb==4.0.12
    # via gitpython
-gitpython==3.1.45
+gitpython==3.1.46
    # via wandb
 glfw==2.10.0
    # via
@@ -212,7 +197,6 @@ grpcio==1.73.1
    #   lerobot
    #   reachy2-sdk
    #   reachy2-sdk-api
-    #   tensorboard
 grpcio-tools==1.73.1
    # via
    #   lerobot
@@ -223,71 +207,67 @@ gym-hil==0.1.13
    # via lerobot
 gym-pusht==0.1.6
    # via lerobot
-gymnasium==1.2.1
+gymnasium==1.2.3
    # via
    #   gym-aloha
    #   gym-hil
    #   gym-pusht
    #   lerobot
-    #   libero
    #   metaworld
 h11==0.16.0
-    # via uvicorn
-h5py==3.15.1
-    # via robomimic
+    # via
+    #   httpcore
+    #   uvicorn
 hebi-py==2.11.0
    # via lerobot
-hf-transfer==0.1.9
-    # via huggingface-hub
-hf-xet==1.1.10
+hf-xet==1.3.2
    # via huggingface-hub
 hidapi==0.14.0.post4
    # via
    #   gym-hil
    #   lerobot
+httpcore==1.0.9
+    # via httpx
 httptools==0.7.1
    # via uvicorn
-huggingface-hub[cli,hf-transfer]==0.35.3
+httpx==0.28.1
+    # via
+    #   datasets
+    #   huggingface-hub
+huggingface-hub==1.6.0
    # via
    #   accelerate
    #   datasets
    #   diffusers
    #   lerobot
    #   peft
-    #   timm
    #   tokenizers
    #   transformers
-hydra-core==1.3.2
-    # via libero
-identify==2.6.15
+identify==2.6.17
    # via pre-commit
 idna==3.11
    # via
    #   anyio
+    #   httpx
    #   requests
    #   yarl
-imageio[ffmpeg]==2.37.0
+imageio[ffmpeg]==2.37.2
    # via
    #   gym-aloha
    #   gym-hil
    #   lerobot
    #   metaworld
-    #   robomimic
    #   scikit-image
 imageio-ffmpeg==0.6.0
-    # via
-    #   imageio
-    #   robomimic
-importlib-metadata==8.7.0
+    # via imageio
+importlib-metadata==8.7.1
    # via diffusers
-importlib-resources==6.5.2
-    # via etils
 iniconfig==2.3.0
    # via pytest
-inquirerpy==0.3.4
-    # via huggingface-hub
-ipython==8.37.0
+ipython==9.11.0
    # via meshcat
+ipython-pygments-lexers==1.1.1
+    # via ipython
 ischedule==1.2.7
    # via placo
 jedi==0.19.2
@@ -296,44 +276,24 @@ jinja2==3.1.6
    # via torch
 jsonlines==4.0.0
    # via lerobot
-jsonschema==4.25.1
-    # via nbformat
-jsonschema-specifications==2025.9.1
-    # via jsonschema
-jupyter-core==5.9.1
-    # via nbformat
-jupytext==1.18.1
-    # via bddl
 kiwisolver==1.4.9
    # via matplotlib
 labmaze==1.0.6
    # via dm-control
-lazy-loader==0.4
+lazy-loader==0.5
    # via scikit-image
-libero @ git+https://github.com/huggingface/lerobot-libero.git@main
-    # via lerobot
-llvmlite==0.45.1
-    # via numba
+librt==0.8.1
+    # via mypy
 lxml==6.0.2
    # via dm-control
-markdown==3.9
-    # via tensorboard
 markdown-it-py==4.0.0
-    # via
-    #   jupytext
-    #   mdit-py-plugins
+    # via rich
 markupsafe==3.0.3
-    # via
-    #   jinja2
-    #   werkzeug
-matplotlib==3.10.7
-    # via
-    #   lerobot
-    #   libero
+    # via jinja2
+matplotlib==3.10.8
+    # via lerobot
 matplotlib-inline==0.2.1
    # via ipython
-mdit-py-plugins==0.5.0
-    # via jupytext
 mdurl==0.1.2
    # via markdown-it-py
 mergedeep==1.3.4
@@ -346,41 +306,35 @@ mock-serial==0.0.1
    # via lerobot
 mpmath==1.3.0
    # via sympy
-mujoco==3.3.7
+mujoco==3.5.0
    # via
    #   dm-control
    #   gym-aloha
    #   gym-hil
-    #   libero
    #   metaworld
-    #   robosuite
-multidict==6.7.0
+multidict==6.7.1
    # via
    #   aiohttp
    #   yarl
-multiprocess==0.70.16
+multiprocess==0.70.18
    # via datasets
+mypy==1.19.1
+    # via lerobot
 mypy-extensions==1.1.0
-    # via typing-inspect
-nbformat==5.10.4
-    # via jupytext
-networkx==3.4.2
    # via
-    #   bddl
+    #   mypy
+    #   typing-inspect
+networkx==3.6.1
+    # via
    #   scikit-image
    #   torch
-ninja==1.13.0
-    # via lerobot
-nodeenv==1.9.1
+nodeenv==1.10.0
    # via pre-commit
 num2words==0.5.14
    # via lerobot
-numba==0.62.1
-    # via robosuite
 numpy==2.2.6
    # via
    #   accelerate
-    #   bddl
    #   cmeel-boost
    #   contourpy
    #   datasets
@@ -389,16 +343,14 @@ numpy==2.2.6
    #   dm-env
    #   dm-tree
    #   gymnasium
-    #   h5py
    #   hebi-py
    #   imageio
    #   labmaze
-    #   libero
+    #   lerobot
    #   matplotlib
    #   meshcat
    #   metaworld
    #   mujoco
-    #   numba
    #   opencv-python
    #   opencv-python-headless
    #   pandas
@@ -406,26 +358,18 @@ numpy==2.2.6
    #   pyquaternion
    #   reachy2-sdk
    #   rerun-sdk
-    #   robomimic
-    #   robosuite
    #   scikit-image
    #   scipy
    #   shapely
    #   teleop
-    #   tensorboard
-    #   tensorboardx
    #   tifffile
    #   torchvision
    #   transformers
    #   transforms3d
-omegaconf==2.3.0
-    # via hydra-core
-opencv-python==4.12.0.88
+opencv-python==4.13.0.92
    # via
    #   gym-pusht
-    #   libero
    #   reachy2-sdk
-    #   robosuite
 opencv-python-headless==4.12.0.88
    # via lerobot
 orderly-set==5.5.0
@@ -435,97 +379,87 @@ packaging==25.0
    #   accelerate
    #   datasets
    #   huggingface-hub
-    #   hydra-core
-    #   jupytext
    #   lazy-loader
    #   lerobot
    #   matplotlib
    #   peft
    #   pytest
+    #   qwen-vl-utils
    #   reachy2-sdk
    #   scikit-image
-    #   tensorboard
-    #   tensorboardx
    #   transformers
    #   wandb
 pandas==2.3.3
    # via
    #   datasets
    #   lerobot
-parso==0.8.5
+parso==0.8.6
    # via jedi
-peft==0.17.1
+pathspec==1.0.4
+    # via mypy
+peft==0.18.1
    # via lerobot
 pexpect==4.9.0
    # via ipython
-pfzy==0.3.4
-    # via inquirerpy
-pillow==12.0.0
+pillow==12.1.1
    # via
    #   diffusers
    #   imageio
-    #   lerobot
    #   matplotlib
    #   meshcat
+    #   qwen-vl-utils
    #   rerun-sdk
-    #   robosuite
    #   scikit-image
-    #   tensorboard
    #   torchvision
 pin==3.4.0
    # via placo
-placo==0.9.14
+placo==0.9.16
    # via lerobot
-platformdirs==4.5.0
+platformdirs==4.9.4
    # via
-    #   jupyter-core
+    #   python-discovery
    #   virtualenv
    #   wandb
 pluggy==1.6.0
    # via
    #   pytest
    #   pytest-cov
-pre-commit==4.3.0
+pre-commit==4.5.1
    # via lerobot
 prompt-toolkit==3.0.52
-    # via
-    #   inquirerpy
-    #   ipython
+    # via ipython
 propcache==0.4.1
    # via
    #   aiohttp
    #   yarl
-protobuf==6.31.0
+protobuf==6.31.1
    # via
    #   dm-control
    #   grpcio-tools
    #   lerobot
    #   reachy2-sdk
    #   reachy2-sdk-api
-    #   tensorboard
-    #   tensorboardx
    #   wandb
-psutil==7.1.1
+psutil==7.2.2
    # via
    #   accelerate
    #   imageio
    #   peft
-    #   robomimic
 ptyprocess==0.7.0
    # via pexpect
 pure-eval==0.2.3
    # via stack-data
-pyarrow==21.0.0
+pyarrow==23.0.1
    # via
    #   datasets
    #   rerun-sdk
-pycparser==2.23
+pycparser==3.0
    # via cffi
-pydantic==2.12.3
+pydantic==2.12.5
    # via
    #   fastapi
    #   wandb
-pydantic-core==2.41.4
+pydantic-core==2.41.5
    # via pydantic
 pygame==2.6.1
    # via
@@ -535,33 +469,35 @@ pygame==2.6.1
 pygments==2.19.2
    # via
    #   ipython
+    #   ipython-pygments-lexers
    #   pytest
+    #   rich
 pymunk==6.11.1
    # via
    #   gym-pusht
    #   lerobot
-pyngrok==7.4.1
+pyngrok==7.5.1
    # via meshcat
 pynput==1.8.1
    # via
    #   gym-hil
    #   lerobot
-pyobjc-core==12.0
+pyobjc-core==12.1
    # via
    #   pyobjc-framework-applicationservices
    #   pyobjc-framework-cocoa
    #   pyobjc-framework-coretext
    #   pyobjc-framework-quartz
-pyobjc-framework-applicationservices==12.0
+pyobjc-framework-applicationservices==12.1
    # via pynput
-pyobjc-framework-cocoa==12.0
+pyobjc-framework-cocoa==12.1
    # via
    #   pyobjc-framework-applicationservices
    #   pyobjc-framework-coretext
    #   pyobjc-framework-quartz
-pyobjc-framework-coretext==12.0
+pyobjc-framework-coretext==12.1
    # via pyobjc-framework-applicationservices
-pyobjc-framework-quartz==12.0
+pyobjc-framework-quartz==12.1
    # via
    #   pynput
    #   pyobjc-framework-applicationservices
@@ -570,13 +506,13 @@ pyopengl==3.1.10
    # via
    #   dm-control
    #   mujoco
-pyparsing==3.2.5
+pyparsing==3.3.2
    # via
    #   dm-control
    #   matplotlib
 pyquaternion==0.9.9
    # via reachy2-sdk
-pyrealsense2-macosx==2.54.2
+pyrealsense2-macosx==2.56.5
    # via lerobot
 pyserial==3.5
    # via
@@ -585,7 +521,6 @@ pyserial==3.5
    #   lerobot
 pytest==8.4.2
    # via
-    #   bddl
    #   lerobot
    #   pytest-cov
    #   pytest-timeout
@@ -596,11 +531,14 @@ pytest-timeout==2.4.0
    # via lerobot
 python-dateutil==2.9.0.post0
    # via
+    #   faker
    #   matplotlib
    #   pandas
-python-dotenv==1.1.1
+python-discovery==1.1.1
+    # via virtualenv
+python-dotenv==1.2.2
    # via uvicorn
-pytz==2025.2
+pytz==2026.1.post1
    # via pandas
 pyyaml==6.0.3
    # via
@@ -609,13 +547,10 @@ pyyaml==6.0.3
    #   draccus
    #   hebi-py
    #   huggingface-hub
-    #   jupytext
-    #   omegaconf
    #   peft
    #   pre-commit
    #   pyngrok
    #   pyyaml-include
-    #   timm
    #   transformers
    #   uvicorn
    #   wandb
@@ -625,15 +560,13 @@ pyzmq==27.1.0
    # via
    #   lerobot
    #   meshcat
-reachy2-sdk==1.0.14
+qwen-vl-utils==0.0.14
+    # via lerobot
+reachy2-sdk==1.0.15
    # via lerobot
 reachy2-sdk-api==1.0.21
    # via reachy2-sdk
-referencing==0.37.0
-    # via
-    #   jsonschema
-    #   jsonschema-specifications
-regex==2025.10.23
+regex==2026.2.28
    # via
    #   diffusers
    #   transformers
@@ -642,184 +575,150 @@ requests==2.32.5
    #   datasets
    #   diffusers
    #   dm-control
-    #   huggingface-hub
+    #   qwen-vl-utils
    #   teleop
-    #   transformers
    #   wandb
-rerun-sdk==0.26.1
+rerun-sdk==0.26.2
    # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
    # via placo
-robomimic==0.2.0
-    # via libero
-robosuite==1.4.0
-    # via libero
-rpds-py==0.28.0
-    # via
-    #   jsonschema
-    #   referencing
-safetensors==0.6.2
+rich==14.3.3
+    # via typer
+safetensors==0.7.0
    # via
    #   accelerate
    #   diffusers
    #   lerobot
    #   peft
-    #   timm
    #   transformers
 scikit-image==0.25.2
    # via
    #   gym-pusht
    #   lerobot
-scipy==1.15.3
+scipy==1.17.1
    # via
    #   dm-control
+    #   lerobot
    #   metaworld
-    #   robosuite
    #   scikit-image
-sentry-sdk==2.42.1
+    #   torchdiffeq
+sentry-sdk==2.54.0
    # via wandb
 shapely==2.1.2
    # via gym-pusht
+shellingham==1.5.4
+    # via typer
 six==1.17.0
    # via
    #   pynput
    #   python-dateutil
-smmap==5.0.2
+smmap==5.0.3
    # via gitdb
-sniffio==1.3.1
-    # via anyio
 stack-data==0.6.3
    # via ipython
-starlette==0.48.0
+starlette==0.52.1
    # via fastapi
 sympy==1.14.0
    # via torch
-teleop==0.1.2
+teleop==0.1.4
    # via lerobot
-tensorboard==2.20.0
-    # via robomimic
-tensorboard-data-server==0.7.2
-    # via tensorboard
-tensorboardx==2.6.4
-    # via robomimic
-termcolor==3.1.0
-    # via
-    #   lerobot
-    #   robomimic
-thop==0.1.1.post2209072238
-    # via libero
-tifffile==2025.5.10
+termcolor==3.3.0
+    # via lerobot
+tifffile==2026.3.3
    # via scikit-image
-timm==1.0.20
-    # via lerobot
-tokenizers==0.22.1
+tokenizers==0.22.2
    # via transformers
 toml==0.10.2
    # via draccus
-tomli==2.3.0
-    # via
-    #   cmeel
-    #   coverage
-    #   jupytext
-    #   pytest
-torch==2.7.1
+torch==2.10.0
    # via
    #   accelerate
    #   lerobot
    #   peft
-    #   robomimic
-    #   thop
-    #   timm
+    #   torchdiffeq
    #   torchvision
-torchcodec==0.5
+torchcodec==0.10.0
    # via lerobot
-torchvision==0.22.1
-    # via
-    #   lerobot
-    #   robomimic
-    #   timm
-tornado==6.5.2
+torchdiffeq==0.2.5
+    # via lerobot
+torchvision==0.25.0
+    # via lerobot
+tornado==6.5.4
    # via meshcat
-tqdm==4.67.1
+tqdm==4.67.3
    # via
    #   datasets
    #   dm-control
    #   huggingface-hub
    #   peft
-    #   robomimic
    #   transformers
 traitlets==5.14.3
    # via
    #   ipython
-    #   jupyter-core
    #   matplotlib-inline
-    #   nbformat
-transformers==4.57.1
+transformers==5.3.0
    # via
    #   lerobot
-    #   libero
    #   peft
 transforms3d==0.4.2
    # via teleop
+typer==0.24.1
+    # via
+    #   huggingface-hub
+    #   transformers
 typing-extensions==4.15.0
    # via
    #   aiosignal
    #   anyio
    #   etils
-    #   exceptiongroup
+    #   faker
    #   fastapi
    #   gymnasium
    #   huggingface-hub
-    #   ipython
-    #   multidict
+    #   mypy
    #   pydantic
    #   pydantic-core
-    #   referencing
    #   rerun-sdk
    #   starlette
    #   torch
    #   typing-inspect
    #   typing-inspection
-    #   uvicorn
-    #   virtualenv
    #   wandb
 typing-inspect==0.9.0
    # via draccus
 typing-inspection==0.4.2
-    # via pydantic
-tzdata==2025.2
+    # via
+    #   fastapi
+    #   pydantic
+tzdata==2025.3
    # via pandas
 u-msgpack-python==2.8.0
    # via meshcat
-urllib3==2.5.0
+urllib3==2.6.3
    # via
    #   requests
    #   sentry-sdk
-uvicorn[standard]==0.38.0
+uvicorn[standard]==0.41.0
    # via teleop
 uvloop==0.22.1
    # via uvicorn
-virtualenv==20.35.3
+virtualenv==21.1.0
    # via pre-commit
-wandb==0.21.4
-    # via
-    #   lerobot
-    #   libero
+wandb==0.24.2
+    # via lerobot
 watchfiles==1.1.1
    # via uvicorn
-wcwidth==0.2.14
+wcwidth==0.6.0
    # via prompt-toolkit
 websocket-client==1.9.0
    # via teleop
-websockets==15.0.1
+websockets==16.0
    # via uvicorn
-werkzeug==3.1.3
-    # via tensorboard
-wrapt==2.0.0
+wrapt==2.1.2
    # via dm-tree
 xxhash==3.6.0
    # via datasets
-yarl==1.22.0
+yarl==1.23.0
    # via aiohttp
 zipp==3.23.0
    # via
@@ -1,12 +1,12 @@
 #
-# This file is autogenerated by pip-compile with Python 3.10
+# This file is autogenerated by pip-compile with Python 3.12
 # by the following command:
 #
 #    pip-compile --output-file=requirements-ubuntu.txt requirements.in
 #
 -e .[all]
    # via -[all]
-absl-py==2.3.1
+absl-py==2.4.0
    # via
    #   dm-control
    #   dm-env
@@ -14,30 +14,33 @@ absl-py==2.3.1
    #   labmaze
    #   mujoco
    #   tensorboard
-accelerate==1.11.0
+accelerate==1.13.0
    # via
    #   lerobot
    #   peft
 aiohappyeyeballs==2.6.1
    # via aiohttp
-aiohttp==3.13.1
+aiohttp==3.13.3
    # via fsspec
 aiosignal==1.4.0
    # via aiohttp
+annotated-doc==0.0.4
+    # via
+    #   fastapi
+    #   typer
 annotated-types==0.7.0
    # via pydantic
 antlr4-python3-runtime==4.9.3
    # via
    #   hydra-core
    #   omegaconf
-anyio==4.11.0
+anyio==4.12.1
    # via
+    #   httpx
    #   starlette
    #   watchfiles
-asttokens==3.0.0
+asttokens==3.0.1
    # via stack-data
-async-timeout==5.0.1
-    # via aiohttp
 attrs==25.4.0
    # via
    #   aiohttp
@@ -47,30 +50,35 @@ attrs==25.4.0
    #   referencing
    #   rerun-sdk
 av==15.1.0
-    # via lerobot
-bddl==1.0.1
-    # via libero
-certifi==2025.10.5
    # via
+    #   lerobot
+    #   qwen-vl-utils
+bddl==1.0.1
+    # via hf-libero
+certifi==2026.2.25
+    # via
+    #   httpcore
+    #   httpx
    #   requests
    #   sentry-sdk
 cffi==2.0.0
    # via pymunk
-cfgv==3.4.0
+cfgv==3.5.0
    # via pre-commit
-charset-normalizer==3.4.4
+charset-normalizer==3.4.5
    # via requests
-click==8.3.0
+click==8.3.1
    # via
+    #   typer
    #   uvicorn
    #   wandb
-cloudpickle==3.1.1
+cloudpickle==3.1.2
    # via
    #   gymnasium
-    #   libero
-cmake==4.1.0
+    #   hf-libero
+cmake==4.1.3
    # via lerobot
-cmeel==0.57.3
+cmeel==0.59.0
    # via
    #   cmeel-assimp
    #   cmeel-boost
@@ -108,20 +116,24 @@ cmeel-zlib==1.3.1
    # via cmeel-assimp
 coal-library==3.0.1
    # via pin
-contourpy==1.3.2
-    # via matplotlib
-coverage[toml]==7.11.0
+contourpy==1.3.3
+    # via
+    #   lerobot
+    #   matplotlib
+coverage[toml]==7.13.4
    # via pytest-cov
+cuda-bindings==12.9.4
+    # via torch
+cuda-pathfinder==1.4.1
+    # via cuda-bindings
 cycler==0.12.1
    # via matplotlib
-datasets==4.1.1
+datasets==4.6.1
    # via lerobot
-debugpy==1.8.17
+debugpy==1.8.20
    # via lerobot
 decorator==5.2.1
    # via ipython
-decord==0.6.0
-    # via lerobot
 deepdiff==8.6.1
    # via lerobot
 diffusers==0.35.2
@@ -132,7 +144,7 @@ dill==0.4.0
    #   multiprocess
 distlib==0.4.0
    # via virtualenv
-dm-control==1.0.34
+dm-control==1.0.37
    # via gym-aloha
 dm-env==1.6
    # via dm-control
@@ -140,7 +152,6 @@ dm-tree==0.1.9
    # via
    #   dm-control
    #   dm-env
-    #   lerobot
 docopt==0.6.2
    # via num2words
 draccus==0.10.0
@@ -148,66 +159,60 @@ draccus==0.10.0
 dynamixel-sdk==3.8.4
    # via lerobot
 easydict==1.13
-    # via libero
-egl-probe @ git+https://github.com/huggingface/egl_probe.git
-    # via
-    #   libero
-    #   robomimic
+    # via hf-libero
+egl-probe==1.0.2
+    # via robomimic
 eigenpy==3.10.3
    # via coal-library
-einops==0.8.1
+einops==0.8.2
    # via
-    #   flash-attn
+    #   hf-libero
    #   lerobot
-    #   libero
 eiquadprog==1.2.9
    # via placo
-etils[epath,epy]==1.13.0
+etils[epath,epy]==1.14.0
    # via mujoco
-evdev==1.9.2
+evdev==1.9.3
    # via pynput
-exceptiongroup==1.3.0
-    # via
-    #   anyio
-    #   ipython
-    #   pytest
 executing==2.2.1
    # via stack-data
+faker==34.0.2
+    # via lerobot
 farama-notifications==0.0.4
    # via gymnasium
-fastapi==0.119.1
-    # via teleop
+fastapi==0.135.1
+    # via
+    #   lerobot
+    #   teleop
 fastjsonschema==2.21.2
    # via nbformat
 feetech-servo-sdk==1.0.0
    # via lerobot
-filelock==3.20.0
+filelock==3.25.0
    # via
    #   datasets
    #   diffusers
    #   huggingface-hub
+    #   python-discovery
    #   torch
-    #   transformers
    #   virtualenv
-flash-attn==2.8.3
-    # via lerobot
-fonttools==4.60.1
+fonttools==4.61.1
    # via matplotlib
 frozenlist==1.8.0
    # via
    #   aiohttp
    #   aiosignal
-fsspec[http]==2025.9.0
+fsspec[http]==2026.2.0
    # via
    #   datasets
    #   etils
    #   huggingface-hub
    #   torch
 future==1.0.0
-    # via libero
+    # via hf-libero
 gitdb==4.0.12
    # via gitpython
-gitpython==3.1.45
+gitpython==3.1.46
    # via wandb
 glfw==2.10.0
    # via
@@ -230,50 +235,60 @@ gym-hil==0.1.13
    # via lerobot
 gym-pusht==0.1.6
    # via lerobot
-gymnasium==1.2.1
+gymnasium==1.2.3
    # via
    #   gym-aloha
    #   gym-hil
    #   gym-pusht
+    #   hf-libero
    #   lerobot
-    #   libero
    #   metaworld
 h11==0.16.0
-    # via uvicorn
-h5py==3.15.1
+    # via
+    #   httpcore
+    #   uvicorn
+h5py==3.16.0
    # via robomimic
 hebi-py==2.11.0
    # via lerobot
-hf-transfer==0.1.9
-    # via huggingface-hub
-hf-xet==1.1.10
+hf-egl-probe==1.0.2
+    # via hf-libero
+hf-libero==0.1.3
+    # via lerobot
+hf-xet==1.3.2
    # via huggingface-hub
 hidapi==0.14.0.post4
    # via
    #   gym-hil
    #   lerobot
+httpcore==1.0.9
+    # via httpx
 httptools==0.7.1
    # via uvicorn
-huggingface-hub[cli,hf-transfer]==0.35.3
+httpx==0.28.1
+    # via
+    #   datasets
+    #   huggingface-hub
+huggingface-hub==1.6.0
    # via
    #   accelerate
    #   datasets
    #   diffusers
    #   lerobot
    #   peft
-    #   timm
    #   tokenizers
    #   transformers
 hydra-core==1.3.2
-    # via libero
-identify==2.6.15
+    # via hf-libero
+identify==2.6.17
    # via pre-commit
 idna==3.11
    # via
    #   anyio
+    #   httpx
    #   requests
    #   yarl
-imageio[ffmpeg]==2.37.0
+imageio[ffmpeg]==2.37.2
    # via
    #   gym-aloha
    #   gym-hil
@@ -285,16 +300,14 @@ imageio-ffmpeg==0.6.0
    # via
    #   imageio
    #   robomimic
-importlib-metadata==8.7.0
+importlib-metadata==8.7.1
    # via diffusers
-importlib-resources==6.5.2
-    # via etils
 iniconfig==2.3.0
    # via pytest
-inquirerpy==0.3.4
-    # via huggingface-hub
-ipython==8.37.0
+ipython==9.11.0
    # via meshcat
+ipython-pygments-lexers==1.1.1
+    # via ipython
 ischedule==1.2.7
    # via placo
 jedi==0.19.2
@@ -303,40 +316,41 @@ jinja2==3.1.6
    # via torch
 jsonlines==4.0.0
    # via lerobot
-jsonschema==4.25.1
+jsonschema==4.26.0
    # via nbformat
 jsonschema-specifications==2025.9.1
    # via jsonschema
 jupyter-core==5.9.1
    # via nbformat
-jupytext==1.18.1
+jupytext==1.19.1
    # via bddl
 kiwisolver==1.4.9
    # via matplotlib
 labmaze==1.0.6
    # via dm-control
-lazy-loader==0.4
+lazy-loader==0.5
    # via scikit-image
-libero @ git+https://github.com/huggingface/lerobot-libero.git@main
-    # via lerobot
-llvmlite==0.45.1
+librt==0.8.1
+    # via mypy
+llvmlite==0.46.0
    # via numba
 lxml==6.0.2
    # via dm-control
-markdown==3.9
+markdown==3.10.2
    # via tensorboard
 markdown-it-py==4.0.0
    # via
    #   jupytext
    #   mdit-py-plugins
+    #   rich
 markupsafe==3.0.3
    # via
    #   jinja2
    #   werkzeug
-matplotlib==3.10.7
+matplotlib==3.10.8
    # via
+    #   hf-libero
    #   lerobot
-    #   libero
 matplotlib-inline==0.2.1
    # via ipython
 mdit-py-plugins==0.5.0
@@ -353,36 +367,38 @@ mock-serial==0.0.1
    # via lerobot
 mpmath==1.3.0
    # via sympy
-mujoco==3.3.7
+mujoco==3.5.0
    # via
    #   dm-control
    #   gym-aloha
    #   gym-hil
-    #   libero
+    #   hf-libero
    #   metaworld
    #   robosuite
-multidict==6.7.0
+multidict==6.7.1
    # via
    #   aiohttp
    #   yarl
-multiprocess==0.70.16
+multiprocess==0.70.18
    # via datasets
+mypy==1.19.1
+    # via lerobot
 mypy-extensions==1.1.0
-    # via typing-inspect
+    # via
+    #   mypy
+    #   typing-inspect
 nbformat==5.10.4
    # via jupytext
-networkx==3.4.2
+networkx==3.6.1
    # via
    #   bddl
    #   scikit-image
    #   torch
-ninja==1.13.0
-    # via lerobot
-nodeenv==1.9.1
+nodeenv==1.10.0
    # via pre-commit
 num2words==0.5.14
    # via lerobot
-numba==0.62.1
+numba==0.64.0
    # via robosuite
 numpy==2.2.6
    # via
@@ -391,7 +407,6 @@ numpy==2.2.6
    #   cmeel-boost
    #   contourpy
    #   datasets
-    #   decord
    #   diffusers
    #   dm-control
    #   dm-env
@@ -399,9 +414,10 @@ numpy==2.2.6
    #   gymnasium
    #   h5py
    #   hebi-py
+    #   hf-libero
    #   imageio
    #   labmaze
-    #   libero
+    #   lerobot
    #   matplotlib
    #   meshcat
    #   metaworld
@@ -426,49 +442,51 @@ numpy==2.2.6
    #   torchvision
    #   transformers
    #   transforms3d
-nvidia-cublas-cu12==12.6.4.1
+nvidia-cublas-cu12==12.8.4.1
    # via
    #   nvidia-cudnn-cu12
    #   nvidia-cusolver-cu12
    #   torch
-nvidia-cuda-cupti-cu12==12.6.80
+nvidia-cuda-cupti-cu12==12.8.90
    # via torch
-nvidia-cuda-nvrtc-cu12==12.6.77
+nvidia-cuda-nvrtc-cu12==12.8.93
    # via torch
-nvidia-cuda-runtime-cu12==12.6.77
+nvidia-cuda-runtime-cu12==12.8.90
    # via torch
-nvidia-cudnn-cu12==9.5.1.17
+nvidia-cudnn-cu12==9.10.2.21
    # via torch
-nvidia-cufft-cu12==11.3.0.4
+nvidia-cufft-cu12==11.3.3.83
    # via torch
-nvidia-cufile-cu12==1.11.1.6
+nvidia-cufile-cu12==1.13.1.3
    # via torch
-nvidia-curand-cu12==10.3.7.77
+nvidia-curand-cu12==10.3.9.90
    # via torch
-nvidia-cusolver-cu12==11.7.1.2
+nvidia-cusolver-cu12==11.7.3.90
    # via torch
-nvidia-cusparse-cu12==12.5.4.2
+nvidia-cusparse-cu12==12.5.8.93
    # via
    #   nvidia-cusolver-cu12
    #   torch
-nvidia-cusparselt-cu12==0.6.3
+nvidia-cusparselt-cu12==0.7.1
    # via torch
-nvidia-nccl-cu12==2.26.2
+nvidia-nccl-cu12==2.27.5
    # via torch
-nvidia-nvjitlink-cu12==12.6.85
+nvidia-nvjitlink-cu12==12.8.93
    # via
    #   nvidia-cufft-cu12
    #   nvidia-cusolver-cu12
    #   nvidia-cusparse-cu12
    #   torch
-nvidia-nvtx-cu12==12.6.77
+nvidia-nvshmem-cu12==3.4.5
+    # via torch
+nvidia-nvtx-cu12==12.8.90
    # via torch
 omegaconf==2.3.0
    # via hydra-core
-opencv-python==4.12.0.88
+opencv-python==4.13.0.92
    # via
    #   gym-pusht
-    #   libero
+    #   hf-libero
    #   reachy2-sdk
    #   robosuite
 opencv-python-headless==4.12.0.88
@@ -487,6 +505,7 @@ packaging==25.0
    #   matplotlib
    #   peft
    #   pytest
+    #   qwen-vl-utils
    #   reachy2-sdk
    #   scikit-image
    #   tensorboard
@@ -497,21 +516,21 @@ pandas==2.3.3
    # via
    #   datasets
    #   lerobot
-parso==0.8.5
+parso==0.8.6
    # via jedi
-peft==0.17.1
+pathspec==1.0.4
+    # via mypy
+peft==0.18.1
    # via lerobot
 pexpect==4.9.0
    # via ipython
-pfzy==0.3.4
-    # via inquirerpy
-pillow==12.0.0
+pillow==12.1.1
    # via
    #   diffusers
    #   imageio
-    #   lerobot
    #   matplotlib
    #   meshcat
+    #   qwen-vl-utils
    #   rerun-sdk
    #   robosuite
    #   scikit-image
@@ -519,28 +538,27 @@ pillow==12.0.0
    #   torchvision
 pin==3.4.0
    # via placo
-placo==0.9.14
+placo==0.9.16
    # via lerobot
-platformdirs==4.5.0
+platformdirs==4.9.4
    # via
    #   jupyter-core
+    #   python-discovery
    #   virtualenv
    #   wandb
 pluggy==1.6.0
    # via
    #   pytest
    #   pytest-cov
-pre-commit==4.3.0
+pre-commit==4.5.1
    # via lerobot
 prompt-toolkit==3.0.52
-    # via
-    #   inquirerpy
-    #   ipython
+    # via ipython
 propcache==0.4.1
    # via
    #   aiohttp
    #   yarl
-protobuf==6.31.0
+protobuf==6.31.1
    # via
    #   dm-control
    #   grpcio-tools
@@ -550,7 +568,7 @@ protobuf==6.31.0
    #   tensorboard
    #   tensorboardx
    #   wandb
-psutil==7.1.1
+psutil==7.2.2
    # via
    #   accelerate
    #   imageio
@@ -560,17 +578,17 @@ ptyprocess==0.7.0
    # via pexpect
 pure-eval==0.2.3
    # via stack-data
-pyarrow==21.0.0
+pyarrow==23.0.1
    # via
    #   datasets
    #   rerun-sdk
-pycparser==2.23
+pycparser==3.0
    # via cffi
-pydantic==2.12.3
+pydantic==2.12.5
    # via
    #   fastapi
    #   wandb
-pydantic-core==2.41.4
+pydantic-core==2.41.5
    # via pydantic
 pygame==2.6.1
    # via
@@ -580,12 +598,14 @@ pygame==2.6.1
 pygments==2.19.2
    # via
    #   ipython
+    #   ipython-pygments-lexers
    #   pytest
+    #   rich
 pymunk==6.11.1
    # via
    #   gym-pusht
    #   lerobot
-pyngrok==7.4.1
+pyngrok==7.5.1
    # via meshcat
 pynput==1.8.1
    # via
@@ -595,7 +615,7 @@ pyopengl==3.1.10
    # via
    #   dm-control
    #   mujoco
-pyparsing==3.2.5
+pyparsing==3.3.2
    # via
    #   dm-control
    #   matplotlib
@@ -621,13 +641,16 @@ pytest-timeout==2.4.0
    # via lerobot
 python-dateutil==2.9.0.post0
    # via
+    #   faker
    #   matplotlib
    #   pandas
-python-dotenv==1.1.1
+python-discovery==1.1.1
+    # via virtualenv
+python-dotenv==1.2.2
    # via uvicorn
 python-xlib==0.33
    # via pynput
-pytz==2025.2
+pytz==2026.1.post1
    # via pandas
 pyyaml==6.0.3
    # via
@@ -642,7 +665,6 @@ pyyaml==6.0.3
    #   pre-commit
    #   pyngrok
    #   pyyaml-include
-    #   timm
    #   transformers
    #   uvicorn
    #   wandb
@@ -652,7 +674,9 @@ pyzmq==27.1.0
    # via
    #   lerobot
    #   meshcat
-reachy2-sdk==1.0.14
+qwen-vl-utils==0.0.14
+    # via lerobot
+reachy2-sdk==1.0.15
    # via lerobot
 reachy2-sdk-api==1.0.21
    # via reachy2-sdk
@@ -660,7 +684,7 @@ referencing==0.37.0
    # via
    #   jsonschema
    #   jsonschema-specifications
-regex==2025.10.23
+regex==2026.2.28
    # via
    #   diffusers
    #   transformers
@@ -669,60 +693,62 @@ requests==2.32.5
    #   datasets
    #   diffusers
    #   dm-control
-    #   huggingface-hub
+    #   qwen-vl-utils
    #   teleop
-    #   transformers
    #   wandb
-rerun-sdk==0.26.1
+rerun-sdk==0.26.2
    # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
    # via placo
+rich==14.3.3
+    # via typer
 robomimic==0.2.0
-    # via libero
+    # via hf-libero
 robosuite==1.4.0
-    # via libero
-rpds-py==0.28.0
+    # via hf-libero
+rpds-py==0.30.0
    # via
    #   jsonschema
    #   referencing
-safetensors==0.6.2
+safetensors==0.7.0
    # via
    #   accelerate
    #   diffusers
    #   lerobot
    #   peft
-    #   timm
    #   transformers
 scikit-image==0.25.2
    # via
    #   gym-pusht
    #   lerobot
-scipy==1.15.3
+scipy==1.17.1
    # via
    #   dm-control
+    #   lerobot
    #   metaworld
    #   robosuite
    #   scikit-image
-sentry-sdk==2.42.1
+    #   torchdiffeq
+sentry-sdk==2.54.0
    # via wandb
 shapely==2.1.2
    # via gym-pusht
+shellingham==1.5.4
+    # via typer
 six==1.17.0
    # via
    #   pynput
    #   python-dateutil
    #   python-xlib
-smmap==5.0.2
+smmap==5.0.3
    # via gitdb
-sniffio==1.3.1
-    # via anyio
 stack-data==0.6.3
    # via ipython
-starlette==0.48.0
+starlette==0.52.1
    # via fastapi
 sympy==1.14.0
    # via torch
-teleop==0.1.2
+teleop==0.1.4
    # via lerobot
 tensorboard==2.20.0
    # via robomimic
@@ -730,46 +756,38 @@ tensorboard-data-server==0.7.2
    # via tensorboard
 tensorboardx==2.6.4
    # via robomimic
-termcolor==3.1.0
+termcolor==3.3.0
    # via
    #   lerobot
    #   robomimic
 thop==0.1.1.post2209072238
-    # via libero
-tifffile==2025.5.10
+    # via hf-libero
+tifffile==2026.3.3
    # via scikit-image
-timm==1.0.20
-    # via lerobot
-tokenizers==0.22.1
+tokenizers==0.22.2
    # via transformers
 toml==0.10.2
    # via draccus
-tomli==2.3.0
-    # via
-    #   cmeel
-    #   coverage
-    #   jupytext
-    #   pytest
-torch==2.7.1
+torch==2.10.0
    # via
    #   accelerate
-    #   flash-attn
    #   lerobot
    #   peft
    #   robomimic
    #   thop
-    #   timm
+    #   torchdiffeq
    #   torchvision
-torchcodec==0.5
+torchcodec==0.10.0
    # via lerobot
-torchvision==0.22.1
+torchdiffeq==0.2.5
+    # via lerobot
+torchvision==0.25.0
    # via
    #   lerobot
    #   robomimic
-    #   timm
-tornado==6.5.2
+tornado==6.5.4
    # via meshcat
-tqdm==4.67.1
+tqdm==4.67.3
    # via
    #   datasets
    #   dm-control
@@ -783,26 +801,29 @@ traitlets==5.14.3
    #   jupyter-core
    #   matplotlib-inline
    #   nbformat
-transformers==4.57.1
+transformers==5.3.0
    # via
+    #   hf-libero
    #   lerobot
-    #   libero
    #   peft
 transforms3d==0.4.2
    # via teleop
-triton==3.3.1
+triton==3.6.0
    # via torch
+typer==0.24.1
+    # via
+    #   huggingface-hub
+    #   transformers
 typing-extensions==4.15.0
    # via
    #   aiosignal
    #   anyio
    #   etils
-    #   exceptiongroup
+    #   faker
    #   fastapi
    #   gymnasium
    #   huggingface-hub
-    #   ipython
-    #   multidict
+    #   mypy
    #   pydantic
    #   pydantic-core
    #   referencing
@@ -811,46 +832,46 @@ typing-extensions==4.15.0
    #   torch
    #   typing-inspect
    #   typing-inspection
-    #   uvicorn
-    #   virtualenv
    #   wandb
 typing-inspect==0.9.0
    # via draccus
 typing-inspection==0.4.2
-    # via pydantic
-tzdata==2025.2
+    # via
+    #   fastapi
+    #   pydantic
+tzdata==2025.3
    # via pandas
 u-msgpack-python==2.8.0
    # via meshcat
-urllib3==2.5.0
+urllib3==2.6.3
    # via
    #   requests
    #   sentry-sdk
-uvicorn[standard]==0.38.0
+uvicorn[standard]==0.41.0
    # via teleop
 uvloop==0.22.1
    # via uvicorn
-virtualenv==20.35.3
+virtualenv==21.1.0
    # via pre-commit
-wandb==0.21.4
+wandb==0.24.2
    # via
+    #   hf-libero
    #   lerobot
-    #   libero
 watchfiles==1.1.1
    # via uvicorn
-wcwidth==0.2.14
+wcwidth==0.6.0
    # via prompt-toolkit
 websocket-client==1.9.0
    # via teleop
-websockets==15.0.1
+websockets==16.0
    # via uvicorn
-werkzeug==3.1.3
+werkzeug==3.1.6
    # via tensorboard
-wrapt==2.0.0
+wrapt==2.1.2
    # via dm-tree
 xxhash==3.6.0
    # via datasets
-yarl==1.22.0
+yarl==1.23.0
    # via aiohttp
 zipp==3.23.0
    # via
@@ -1,9 +1,9 @@
 # requirements.in

-# requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.0.1 25A362 arm64).
-# Darwin MacBook-Pro.local 25.0.0 Darwin Kernel Version 25.0.0: Wed Sep 17 21:42:08 PDT 2025; root:xnu-12377.1.9~141/RELEASE_ARM64_T8132 arm64
+# requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.3.1 25D2128 arm64).
+# Darwin MacBook-Pro.local 25.3.0 Darwin Kernel Version 25.3.0: Wed Jan 28 20:54:55 PST 2026; root:xnu-12377.91.3~2/RELEASE_ARM64_T8132 arm64

-# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.3 LTS x86_64).
-# Linux mlerobot-linux 6.14.0-33-generic #33~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Fri Sep 19 17:02:30 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
+# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.4 LTS x86_64).
+# Linux lerobot-linux 6.17.0-14-generic #14~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Thu Jan 15 15:52:10 UTC 2 x86_64 x86_64 x86_64 GNU/Linux

 -e .[all]
@@ -0,0 +1,72 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from setuptools import setup
+
+
+def get_version_from_toml() -> str:
+    """Return the project's version string parsed from `pyproject.toml`.
+
+    The function scans `pyproject.toml` line-by-line looking for a line
+    that starts with ``version`` (for example: ``version = "1.2.3"``)
+    and returns the value without surrounding quotes. If no such line is
+    found a :class:`ValueError` is raised.
+
+    Returns:
+        The version string from `pyproject.toml` (e.g. ``"1.2.3"`` ->
+        ``1.2.3``).
+    """
+
+    version = None
+    with open("pyproject.toml", encoding="utf-8") as f:
+        for line in f:
+            if line.strip().startswith("version"):
+                version = line.split("=")[1].strip().strip('"')
+                break
+    if version is None:
+        raise ValueError("Version not found in pyproject.toml")
+    return version
+
+
+def read_long_description() -> str:
+    """Read and return the project's long description for setup.
+
+    This function reads `README.md` and replaces image links that point
+    to the local `./media/` directory with absolute raw GitHub URLs that
+    reference the release tag corresponding to the version parsed from
+    `pyproject.toml` (for example, ``v1.2.3``). The modified README
+    content is returned as a string suitable for passing to
+    ``setuptools.setup(long_description=...)``.
+
+    Returns:
+        The README content with rewritten media links.
+    """
+
+    with open("README.md", encoding="utf-8") as f:
+        content = f.read()
+
+    version = get_version_from_toml()
+    git_tag = f"v{version}"
+
+    base_raw_url = f"https://raw.githubusercontent.com/huggingface/lerobot/{git_tag}/"
+    content = content.replace('src="./media/', f'src="{base_raw_url}media/')
+
+    return content
+
+
+setup(
+    long_description=read_long_description(),
+    long_description_content_type="text/markdown",
+)
@@ -126,6 +126,12 @@ class RobotClientConfig:

    # Device configuration
    policy_device: str = field(default="cpu", metadata={"help": "Device for policy inference"})
+    client_device: str = field(
+        default="cpu",
+        metadata={
+            "help": "Device to move actions to after receiving from server (e.g., for downstream planners)"
+        },
+    )

    # Control behavior configuration
    chunk_size_threshold: float = field(default=0.5, metadata={"help": "Threshold for chunk size control"})
@@ -161,6 +167,9 @@ class RobotClientConfig:
        if not self.policy_device:
            raise ValueError("policy_device cannot be empty")

+        if not self.client_device:
+            raise ValueError("client_device cannot be empty")
+
        if self.chunk_size_threshold < 0 or self.chunk_size_threshold > 1:
            raise ValueError(f"chunk_size_threshold must be between 0 and 1, got {self.chunk_size_threshold}")

@@ -184,6 +193,7 @@ class RobotClientConfig:
            "policy_type": self.policy_type,
            "pretrained_name_or_path": self.pretrained_name_or_path,
            "policy_device": self.policy_device,
+            "client_device": self.client_device,
            "chunk_size_threshold": self.chunk_size_threshold,
            "fps": self.fps,
            "actions_per_chunk": self.actions_per_chunk,
@@ -23,7 +23,7 @@ DEFAULT_INFERENCE_LATENCY = 1 / DEFAULT_FPS
 DEFAULT_OBS_QUEUE_TIMEOUT = 2

 # All action chunking policies
-SUPPORTED_POLICIES = ["act", "smolvla", "diffusion", "tdmpc", "vqbet", "pi0", "pi05"]
+SUPPORTED_POLICIES = ["act", "smolvla", "diffusion", "tdmpc", "vqbet", "pi0", "pi05", "groot"]

 # TODO: Add all other robots
-SUPPORTED_ROBOTS = ["so100_follower", "so101_follower", "bi_so100_follower", "omx_follower"]
+SUPPORTED_ROBOTS = ["so100_follower", "so101_follower", "bi_so_follower", "omx_follower"]
@@ -18,11 +18,12 @@ import os
 import time
 from dataclasses import dataclass, field
 from pathlib import Path
+from typing import Any

 import torch

 from lerobot.configs.types import PolicyFeature
-from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.datasets.feature_utils import build_dataset_frame, hw_to_dataset_features

 # NOTE: Configs need to be loaded for the client to be able to instantiate the policy config
 from lerobot.policies import (  # noqa: F401
@@ -39,8 +40,8 @@ from lerobot.utils.utils import init_logging

 Action = torch.Tensor

-# observation as received from the robot
-RawObservation = dict[str, torch.Tensor]
+# observation as received from the robot (can be numpy arrays, floats, etc.)
+RawObservation = dict[str, Any]

 # observation as those recorded in LeRobot dataset (keys are different)
 LeRobotObservation = dict[str, torch.Tensor]
@@ -39,15 +39,13 @@ import grpc
 import torch

 from lerobot.policies.factory import get_policy_class, make_pre_post_processors
-from lerobot.processor import (
-    PolicyAction,
-    PolicyProcessorPipeline,
-)
+from lerobot.processor import PolicyProcessorPipeline
 from lerobot.transport import (
    services_pb2,  # type: ignore
    services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import receive_bytes_in_chunks
+from lerobot.types import PolicyAction

 from .configs import PolicyServerConfig
 from .constants import SUPPORTED_POLICIES
@@ -381,6 +379,8 @@ class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
        action_tensor = torch.stack(processed_actions, dim=1).squeeze(0)
        self.logger.debug(f"Postprocessed action shape: {action_tensor.shape}")

+        action_tensor = action_tensor.detach().cpu()
+
        """5. Convert to TimedAction list"""
        action_chunk = self._time_action_chunk(
            observation_t.get_timestamp(), list(action_tensor), observation_t.get_timestep()
@@ -25,6 +25,7 @@ python src/lerobot/async_inference/robot_client.py \
    --policy_type=act \
    --pretrained_name_or_path=user/model \
    --policy_device=mps \
+    --client_device=cpu \
    --actions_per_chunk=50 \
    --chunk_size_threshold=0.5 \
    --aggregate_fn_name=weighted_average \
@@ -51,21 +52,20 @@ from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraCon
 from lerobot.robots import (  # noqa: F401
    Robot,
    RobotConfig,
-    bi_so100_follower,
+    bi_so_follower,
    koch_follower,
    make_robot_from_config,
    omx_follower,
-    so100_follower,
-    so101_follower,
+    so_follower,
 )
 from lerobot.transport import (
    services_pb2,  # type: ignore
    services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
+from lerobot.utils.import_utils import register_third_party_plugins

 from .configs import RobotClientConfig
-from .constants import SUPPORTED_ROBOTS
 from .helpers import (
    Action,
    FPSTracker,
@@ -286,6 +286,21 @@ class RobotClient:
                timed_actions = pickle.loads(actions_chunk.data)  # nosec
                deserialize_time = time.perf_counter() - deserialize_start

+                # Log device type of received actions
+                if len(timed_actions) > 0:
+                    received_device = timed_actions[0].get_action().device.type
+                    self.logger.debug(f"Received actions on device: {received_device}")
+
+                # Move actions to client_device (e.g., for downstream planners that need GPU)
+                client_device = self.config.client_device
+                if client_device != "cpu":
+                    for timed_action in timed_actions:
+                        if timed_action.get_action().device.type != client_device:
+                            timed_action.action = timed_action.get_action().to(client_device)
+                    self.logger.debug(f"Converted actions to device: {client_device}")
+                else:
+                    self.logger.debug(f"Actions kept on device: {client_device}")
+
                self.action_chunk_size = max(self.action_chunk_size, len(timed_actions))

                # Calculate network latency if we have matching observations
@@ -470,8 +485,9 @@ class RobotClient:
 def async_client(cfg: RobotClientConfig):
    logging.info(pformat(asdict(cfg)))

-    if cfg.robot.type not in SUPPORTED_ROBOTS:
-        raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
+    # TODO: Assert if checking robot support is still needed with the plugin system
+    # if cfg.robot.type not in SUPPORTED_ROBOTS:
+    #     raise ValueError(f"Robot {cfg.robot.type} not yet supported!")

    client = RobotClient(cfg)

@@ -497,4 +513,5 @@ def async_client(cfg: RobotClientConfig):


 if __name__ == "__main__":
+    register_third_party_plugins()
    async_client()  # run the client
--- a/Show More
+++ b/Show More