Merge remote-tracking branch 'origin/main' into codex/episode-video-streaming-byte-cache

# Conflicts: # src/lerobot/scripts/lerobot_train.py
Remove streaming benchmark prototypes
2026-08-08 17:39:44 +00:00 · 2026-07-28 11:18:45 +02:00 · 2026-07-28 11:14:10 +02:00 · 2026-07-27 20:49:19 +02:00 · 2026-07-27 16:07:48 +02:00 · 2026-07-27 09:41:57 +02:00
387 changed files with 8765 additions and 17421 deletions
@@ -53,7 +53,7 @@ permissions:
  contents: read

 env:
-  UV_VERSION: "0.11.30"
+  UV_VERSION: "0.8.0"
  PYTHON_VERSION: "3.12"

 # Cancel in-flight runs for the same branch/PR.
@@ -72,19 +72,19 @@ jobs:
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}

    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false
          lfs: true

      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false

      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
@@ -95,7 +95,7 @@ jobs:
      # from source-copy, so code-only changes skip the slow uv-sync layer
      # when the runner has a warm Docker daemon cache.
      - name: Build Libero benchmark image
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: docker/Dockerfile.benchmark.libero
@@ -151,7 +151,7 @@ jobs:

      - name: Upload Libero rollout video
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: libero-rollout-video
          path: /tmp/libero-artifacts/videos/
@@ -159,7 +159,7 @@ jobs:

      - name: Upload Libero eval metrics
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: libero-metrics
          path: /tmp/libero-artifacts/metrics.json
@@ -214,7 +214,7 @@ jobs:

      - name: Upload Libero train-smoke eval video
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: libero-train-smoke-video
          path: /tmp/libero-train-smoke-artifacts/eval/
@@ -230,19 +230,19 @@ jobs:
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}

    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false
          lfs: true

      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false

      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
@@ -250,7 +250,7 @@ jobs:
          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}

      - name: Build MetaWorld benchmark image
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: docker/Dockerfile.benchmark.metaworld
@@ -303,7 +303,7 @@ jobs:

      - name: Upload MetaWorld rollout video
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: metaworld-rollout-video
          path: /tmp/metaworld-artifacts/videos/
@@ -311,7 +311,7 @@ jobs:

      - name: Upload MetaWorld eval metrics
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: metaworld-metrics
          path: /tmp/metaworld-artifacts/metrics.json
@@ -332,19 +332,19 @@ jobs:
      ROBOTWIN_TASKS: beat_block_hammer,click_bell,handover_block,stack_blocks_two,click_alarmclock,open_microwave,adjust_bottle,lift_pot,stamp_seal,turn_switch

    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false
          lfs: true

      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false

      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
@@ -355,7 +355,7 @@ jobs:
      # simulation assets (~4 GB). Layer cache lives in the runner's local
      # Docker daemon — reused across re-runs on the same machine.
      - name: Build RoboTwin 2.0 benchmark image
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: docker/Dockerfile.benchmark.robotwin
@@ -413,7 +413,7 @@ jobs:

      - name: Upload RoboTwin rollout video
        if: always()
-        uses: actions/upload-artifact@v7
+        uses: actions/upload-artifact@v4
        with:
          name: robotwin-rollout-video
          path: /tmp/robotwin-artifacts/videos/
@@ -421,7 +421,7 @@ jobs:

      - name: Upload RoboTwin eval metrics
        if: always()
-        uses: actions/upload-artifact@v7
+        uses: actions/upload-artifact@v4
        with:
          name: robotwin-metrics
          path: /tmp/robotwin-artifacts/metrics.json
@@ -439,19 +439,19 @@ jobs:
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}

    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false
          lfs: true

      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false

      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
@@ -459,7 +459,7 @@ jobs:
          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}

      - name: Build RoboCasa365 benchmark image
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: docker/Dockerfile.benchmark.robocasa
@@ -514,7 +514,7 @@ jobs:

      - name: Upload RoboCasa365 rollout video
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: robocasa-rollout-video
          path: /tmp/robocasa-artifacts/videos/
@@ -522,7 +522,7 @@ jobs:

      - name: Upload RoboCasa365 eval metrics
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: robocasa-metrics
          path: /tmp/robocasa-artifacts/metrics.json
@@ -540,19 +540,19 @@ jobs:
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}

    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false
          lfs: true

      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false

      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
@@ -560,7 +560,7 @@ jobs:
          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}

      - name: Build RoboCerebra benchmark image
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: docker/Dockerfile.benchmark.robocerebra
@@ -621,7 +621,7 @@ jobs:

      - name: Upload RoboCerebra rollout video
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: robocerebra-rollout-video
          path: /tmp/robocerebra-artifacts/videos/
@@ -629,7 +629,7 @@ jobs:

      - name: Upload RoboCerebra eval metrics
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: robocerebra-metrics
          path: /tmp/robocerebra-artifacts/metrics.json
@@ -648,19 +648,19 @@ jobs:
      ROBOMME_TASKS: PickXtimes,BinFill,StopCube,MoveCube,InsertPeg,SwingXtimes,VideoUnmask,ButtonUnmask,PickHighlight,PatternLock

    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
        with:
          persist-credentials: false
          lfs: true

      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false

      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
@@ -668,7 +668,7 @@ jobs:
          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}

      - name: Build RoboMME benchmark image
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: docker/Dockerfile.benchmark.robomme
@@ -726,7 +726,7 @@ jobs:

      - name: Upload RoboMME rollout video
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: robomme-rollout-video
          path: /tmp/robomme-artifacts/videos/
@@ -734,7 +734,7 @@ jobs:

      - name: Upload RoboMME eval metrics
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: robomme-metrics
          path: /tmp/robomme-artifacts/metrics.json
@@ -754,19 +754,19 @@ jobs:
      LIBERO_PLUS_TASK_IDS: "[0,100,260,500,1000,1500,2000,2400]"

    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false
          lfs: true

      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false

      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
@@ -774,7 +774,7 @@ jobs:
          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}

      - name: Build LIBERO-plus benchmark image
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: docker/Dockerfile.benchmark.libero_plus
@@ -834,7 +834,7 @@ jobs:

      - name: Upload LIBERO-plus rollout video
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: libero-plus-rollout-video
          path: /tmp/libero-plus-artifacts/videos/
@@ -842,7 +842,7 @@ jobs:

      - name: Upload LIBERO-plus eval metrics
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: libero-plus-metrics
          path: /tmp/libero-plus-artifacts/metrics.json
@@ -858,19 +858,19 @@ jobs:
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}

    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false
          lfs: true

      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false

      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
@@ -878,7 +878,7 @@ jobs:
          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}

      - name: Build VLABench benchmark image
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: docker/Dockerfile.benchmark.vlabench
@@ -936,7 +936,7 @@ jobs:

      - name: Upload VLABench rollout video
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: vlabench-rollout-video
          path: /tmp/vlabench-artifacts/videos/
@@ -944,7 +944,7 @@ jobs:

      - name: Upload VLABench eval metrics
        if: always()
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: vlabench-metrics
          path: /tmp/vlabench-artifacts/metrics.json
@@ -47,7 +47,7 @@ jobs:
    timeout-minutes: 30
    steps:
      - name: Checkout code
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false

@@ -27,7 +27,7 @@ on:

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.11.30"
+  UV_VERSION: "0.8.0"
  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
@@ -52,21 +52,21 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
      - name: Build and push Docker image CPU
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: ./docker/Dockerfile.user
@@ -87,21 +87,21 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
      - name: Build and push Docker image GPU
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: ./docker/Dockerfile.internal
@@ -33,7 +33,7 @@ jobs:
      github.event.workflow_run.event == 'pull_request' &&
      github.event.workflow_run.conclusion == 'success' &&
      github.repository == 'huggingface/lerobot'
-    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@6108e850ae1cf2f71bb0815a600bcd50c39abfa7  # main
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
    with:
      package_name: lerobot
    secrets:
@@ -24,24 +24,19 @@ on:
        required: false
        type: string

-  # Triggers on pushes to main that touch the docs or the sources the API reference is generated from.
-  # `src/**` is included because the API reference is built from docstrings via `[[autodoc]]`: without it,
-  # published API pages would go stale as soon as a docstring changed.
+  # Triggers the workflow on push events to main for the docs folder
  push:
    branches:
      - main
    paths:
      - "docs/**"
-      - "src/**"

-  # Same for pull requests, so a docstring change gets a preview build and a broken `[[autodoc]]` path
-  # fails the PR rather than main.
+  # Triggers the workflow on pull request events targeting main for the docs folder
  pull_request:
    branches:
      - main
    paths:
      - "docs/**"
-      - "src/**"

  release:
    types: [published]
@@ -60,25 +55,16 @@ jobs:
      github.repository == 'huggingface/lerobot'
    permissions:
      contents: read
-    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@6108e850ae1cf2f71bb0815a600bcd50c39abfa7  # main
+    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@e60a538eea9817ab312196d0d233604b01697265  # main
    with:
      commit_sha: ${{ github.sha }}
      package: lerobot
-      # doc-builder ships a mock-deps registry entry for lerobot, so the reusable workflow takes its
-      # "light install" path: `pip install ./lerobot --no-deps` plus a handful of real dependencies.
-      # That is not enough to import lerobot — draccus runs `register_subclass` at import time and
-      # `processor/converters.py` calls `functools.singledispatch.register(torch.Tensor)`, neither of
-      # which works against a mock. Install the package for real before the build.
-      pre_command: uv pip install "./lerobot[dataset]"
-      # `--version main` is load-bearing: without `--not_python_module`, doc-builder falls back to
-      # `lerobot.__version__` and only maps that to the default branch when it contains "dev". Our main
-      # branch carries a release version (0.6.2), so omitting this would publish the main docs to
-      # /lerobot/v0.6.2/ instead of /lerobot/main/ and disable notebook building.
      additional_args: >-
+        --not_python_module
        ${{
          (github.event_name == 'release' && format('--version {0}', github.event.release.tag_name)) ||
          (inputs.version != '' && format('--version {0}', inputs.version)) ||
-          '--version main'
+          ''
        }}
    secrets:
      token: ${{ secrets.HUGGINGFACE_PUSH }}
@@ -92,11 +78,9 @@ jobs:
    permissions:
      contents: read
      pull-requests: write
-    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@6108e850ae1cf2f71bb0815a600bcd50c39abfa7  # main
+    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@e60a538eea9817ab312196d0d233604b01697265  # main
    with:
      commit_sha: ${{ github.event.pull_request.head.sha }}
      pr_number: ${{ github.event.number }}
      package: lerobot
-      # See the comment on build_main_docs. The PR workflow passes its own `--version pr_<n>`, so no
-      # additional_args are needed here.
-      pre_command: uv pip install "./lerobot[dataset]"
+      additional_args: --not_python_module
@@ -48,7 +48,7 @@ permissions:

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.11.30"
+  UV_VERSION: "0.8.0"
  PYTHON_VERSION: "3.12"

 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
@@ -69,7 +69,7 @@ jobs:
      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false
          lfs: true
@@ -87,7 +87,7 @@ jobs:
          libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev

      - name: Setup uv and Python
-        uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990  # v8.3.2
+        uses: astral-sh/setup-uv@d0cc045d04ccac9d8b7881df0226f9e82c39688e  # v6
        with:
          enable-cache: true
          version: ${{ env.UV_VERSION }}
@@ -37,7 +37,7 @@ permissions:

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.11.30"
+  UV_VERSION: "0.8.0"
  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu

@@ -63,7 +63,7 @@ jobs:
      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          lfs: true
          persist-credentials: false
@@ -80,7 +80,7 @@ jobs:
          speech-dispatcher libgeos-dev portaudio19-dev

      - name: Setup uv and Python
-        uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990  # v8.3.2
+        uses: astral-sh/setup-uv@d0cc045d04ccac9d8b7881df0226f9e82c39688e  # v6
        with:
          enable-cache: true
          version: ${{ env.UV_VERSION }}
@@ -137,21 +137,21 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          lfs: true
          persist-credentials: false
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c  # v4.2.0
+        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f  # v3
        with:
          cache-binary: false
      - name: Login to Docker Hub
-        uses: docker/login-action@af1e73f918a031802d376d3c8bbc3fe56130a9b0  # v4.4.0
+        uses: docker/login-action@c94ce9fb468520275223c153574b00df6fe4bcc9  # v3
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
      - name: Build and push Docker image
-        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a  # v7.3.0
+        uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8  # v6
        with:
          context: .
          file: ./docker/Dockerfile.internal
@@ -29,7 +29,7 @@ jobs:
    runs-on: ubuntu-latest
    if: github.repository == 'huggingface/lerobot'
    steps:
-      - uses: actions/github-script@v9
+      - uses: actions/github-script@v8
        with:
          script: |
            // Setup Input Text
@@ -27,7 +27,7 @@ on:

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.11.30"
+  UV_VERSION: "0.8.0"
  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:latest-deps

@@ -48,12 +48,12 @@ jobs:
    outputs:
      changed: ${{ steps.diff.outputs.changed }}
    steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
        with:
          persist-credentials: false

      - name: Setup uv and Python
-        uses: astral-sh/setup-uv@v8.3.2 # zizmor: ignore[unpinned-uses]
+        uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
        with:
          version: ${{ env.UV_VERSION }}
          python-version: ${{ env.PYTHON_VERSION }}
@@ -74,7 +74,7 @@ jobs:

      - name: Upload updated lockfile
        if: steps.diff.outputs.changed == 'true'
-        uses: actions/upload-artifact@v7 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: uv-lock
          path: uv.lock
@@ -93,13 +93,13 @@ jobs:
      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false

      - name: Download updated lockfile
-        uses: actions/download-artifact@v8 # zizmor: ignore[unpinned-uses]
+        uses: actions/download-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: uv-lock

@@ -115,7 +115,7 @@ jobs:
          speech-dispatcher libgeos-dev portaudio19-dev

      - name: Setup uv and Python
-        uses: astral-sh/setup-uv@v8.3.2 # zizmor: ignore[unpinned-uses]
+        uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
        with:
          enable-cache: true
          version: ${{ env.UV_VERSION }}
@@ -153,27 +153,27 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false

      - name: Download updated lockfile
-        uses: actions/download-artifact@v8 # zizmor: ignore[unpinned-uses]
+        uses: actions/download-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: uv-lock

      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
      - name: Build and push Docker image
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: ./docker/Dockerfile.internal
@@ -247,12 +247,12 @@ jobs:
    env:
      GH_TOKEN: ${{ secrets.UPDATE_LOCK_TOKEN }}
    steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
        with:
          persist-credentials: false

      - name: Download updated lockfile
-        uses: actions/download-artifact@v8 # zizmor: ignore[unpinned-uses]
+        uses: actions/download-artifact@v4 # zizmor: ignore[unpinned-uses]
        with:
          name: uv-lock

@@ -33,7 +33,7 @@ jobs:
    runs-on: ubuntu-latest
    if: github.repository == 'huggingface/lerobot' && !github.event.pull_request.draft
    steps:
-      - uses: actions/labeler@v7
+      - uses: actions/labeler@v6
        with:
          repo-token: ${{ secrets.GITHUB_TOKEN }}
          sync-labels: true # Removes labels if files are removed from the PR
@@ -43,12 +43,12 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout code
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false

      - name: Set up Python
-        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v6
+        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6
        with:
          python-version: '3.12'

@@ -56,41 +56,3 @@ jobs:
        uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd  # v3.0.1
        with:
          extra_args: --all-files --show-diff-on-failure --color=always
-
-  # This job runs the examples in our docstrings and validates the doctest allowlist.
-  # See docs/source/writing_docstrings.mdx for the standard these enforce.
-  doc-checks:
-    name: Run Documentation Checks (Doctests)
-    runs-on: ubuntu-latest
-    env:
-      # Examples that need a physical robot, a serial port or a Hub download are skipped by content.
-      # Everything else has to actually run. See src/lerobot/utils/doctest_utils.py.
-      SKIP_HARDWARE_DOCTEST: "1"
-      SKIP_CUDA_DOCTEST: "1"
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
-        with:
-          persist-credentials: false
-
-      - name: Setup uv and Python
-        uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990  # v8.3.2
-        with:
-          enable-cache: true
-          version: "0.11.30"
-          python-version: "3.12"
-
-      - name: Install dependencies
-        run: uv sync --locked --extra test --extra dataset
-
-      - name: Check the doctest list is sorted and its paths exist
-        run: make check-doctest-list
-
-      - name: Check documented arguments match their signatures
-        run: make check-docstrings
-
-      - name: Check docstring coverage has not regressed
-        run: uv run --with interrogate interrogate --config=pyproject.toml
-
-      - name: Run doctests
-        run: make doctest
@@ -21,7 +21,7 @@ on:

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.11.30"
+  UV_VERSION: "0.8.0"
  PYTHON_VERSION: "3.12"

 jobs:
@@ -38,12 +38,12 @@ jobs:

    steps:
      - name: Checkout code
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false

      - name: Set up Python
-        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v6
+        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6
        with:
          python-version: '3.12'

@@ -104,7 +104,7 @@ jobs:
      - name: Publish to TestPyPI for pre-releases
        # True for tags like 'v0.2.0-rc1'
        if: startsWith(github.ref, 'refs/tags/v') && contains(github.ref, '-')
-        uses: pypa/gh-action-pypi-publish@ba38be9e461d3875417946c167d0b5f3d385a247  # v1.14.1
+        uses: pypa/gh-action-pypi-publish@ed0c53931b1dc9bd32cbe73a98c7f6766f8a527e  # v1.13.0
        with:
          repository-url: https://test.pypi.org/legacy/
          verbose: true
@@ -112,7 +112,7 @@ jobs:

      - name: Publish to PyPI
        if: startsWith(github.ref, 'refs/tags/v') && !contains(github.ref, '-')
-        uses: pypa/gh-action-pypi-publish@ba38be9e461d3875417946c167d0b5f3d385a247  # v1.14.1
+        uses: pypa/gh-action-pypi-publish@ed0c53931b1dc9bd32cbe73a98c7f6766f8a527e  # v1.13.0
        with:
          verbose: true
          print-hash: true
@@ -127,7 +127,7 @@ jobs:
    env:
      MUJOCO_GL: egl
    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          lfs: true
          persist-credentials: false
@@ -137,7 +137,7 @@ jobs:
          git curl libglib2.0-0 libegl1-mesa-dev ffmpeg libusb-1.0-0-dev \
          speech-dispatcher libgeos-dev portaudio19-dev
      - name: Setup uv and Python
-        uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990  # v8.3.2
+        uses: astral-sh/setup-uv@d0cc045d04ccac9d8b7881df0226f9e82c39688e  # v6
        with:
          enable-cache: true # zizmor: ignore[cache-poisoning]
          version: ${{ env.UV_VERSION }}
@@ -43,12 +43,12 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout code
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          fetch-depth: 0
          persist-credentials: false

      - name: Secret Scanning
-        uses: trufflesecurity/trufflehog@27b0417c16317ca9a472a9a8092acce143b49c55  # v3.95.9
+        uses: trufflesecurity/trufflehog@eafb8c5f6a06175141c27f17bcc17941853d0047  # v3.90.0
        with:
          extra_args: --only-verified
@@ -19,8 +19,8 @@ on:
  workflow_dispatch:

  # Runs at 02:00
-  schedule:
-    - cron: "0 2 * * *"
+  # schedule:
+  #   - cron: "0 2 * * *"

 env:
  CLOSE_ISSUE_MESSAGE: >
@@ -31,7 +31,7 @@ env:
    Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
  WARN_ISSUE_MESSAGE: >
    This issue has been automatically marked as stale because it has not had
-    recent activity (1 year). It will be closed if no further activity occurs within 30 days.
+    recent activity (1 year). It will be closed if no further activity occurs.
    Any change, comment or update to this issue will reset this count.
    Thank you for your contributions.
  WARN_PR_MESSAGE: >
@@ -61,8 +61,8 @@ jobs:
          exempt-pr-labels: never-stale
          days-before-issue-stale: 365
          days-before-issue-close: 30
-          days-before-pr-stale: -1
-          days-before-pr-close: -1
+          days-before-pr-stale: 365
+          days-before-pr-close: 30
          delete-branch: true
          close-issue-message: ${{ env.CLOSE_ISSUE_MESSAGE }}
          close-pr-message: ${{ env.CLOSE_PR_MESSAGE }}
@@ -67,11 +67,7 @@ repos:
        args: [--prose-wrap=preserve]
        # Jinja2 model-card templates use a .md extension but contain {% ... %} /
        # {{ ... }} tags that prettier's Markdown formatter mangles (e.g. table loops).
-        #
-        # docs/source/api/ holds the generated API reference. Its `[[autodoc]]` blocks restrict output
-        # to an indented `- member` list, which prettier reads as a lazy paragraph continuation and
-        # joins onto one line — silently turning a member list into part of the directive.
-        exclude: ^(src/lerobot/templates/.*\.md|docs/source/api/.*\.mdx)$
+        exclude: ^src/lerobot/templates/.*\.md$

  ##### Security #####
  - repo: https://github.com/gitleaks/gitleaks
@@ -108,13 +104,8 @@ repos:
  #       args: ["--docstring-style", "google", "-v", "2"]
  #       exclude: ^tests/.*$

-  # interrogate runs in CI (quality.yml, doc-checks job) rather than here. Its 1.7.0 release still imports
-  # the deprecated `py` package, which resolves against whatever `py` happens to be importable in
-  # pre-commit's isolated env — on a machine with miniconda on the path that is a stray `py.py` and the
-  # hook dies before it reads any config. The gate is the same either way; the CI step is just reliable.
  # - repo: https://github.com/econchick/interrogate
  #   rev: 1.7.0
  #   hooks:
  #     - id: interrogate
-  #       args: ["--config=pyproject.toml"]
-  #       pass_filenames: false
+  #       args: ["-vv", "--config=pyproject.toml"]
@@ -61,20 +61,16 @@ Full details in [`docs/source/so101.mdx`](./docs/source/so101.mdx) and [`docs/so
 **4.1 Install**

 ```bash
-# uv (recommended — see AGENTS.md and CLAUDE.md)
-uv sync --locked --extra feetech          # SO-100/SO-101 motor stack
-# uv sync --locked --extra all            # everything
-# uv sync --locked --extra smolvla        # add SmolVLA deps
-
-# pip (alternative, e.g. when not working from source)
-# pip install 'lerobot[feetech]'
-# pip install 'lerobot[all]'
-# pip install 'lerobot[smolvla]'
-
+pip install 'lerobot[feetech]'              # SO-100/SO-101 motor stack
+# pip install 'lerobot[all]'                # everything
+# pip install 'lerobot[aloha,pusht]'        # specific features
+# pip install 'lerobot[smolvla]'            # add SmolVLA deps
 git lfs install && git lfs pull
-hf auth login                             # required to push datasets/policies
+hf auth login                               # required to push datasets/policies
 ```

+Contributors can alternatively use `uv sync --locked --extra feetech` (see `AGENTS.md`).
+
 **4.2 Find USB ports** — run once per arm, unplug when prompted.

 ```bash
@@ -50,10 +50,6 @@ To run checks manually on all files:
 pre-commit run --all-files
 ```

-### Docstrings
-
-The API reference is generated from the docstrings in `src/lerobot/`. If you add or change anything public, follow the [docstring standard](https://huggingface.co/docs/lerobot/writing_docstrings) — the format is parsed by the renderer and checked in CI.
-
 ### Running Tests

 We use `pytest`. First, ensure you have test artifacts by installing **git-lfs**:
@@ -184,29 +184,3 @@ test-smolvla-ete-eval:
 # backend, so it does not require a real model checkpoint or GPU.
 annotation-e2e:
 	uv run python -m tests.annotations.run_e2e_smoke
-
-# Docstring & doctest checks. See docs/source/writing_docstrings.mdx for the standard these enforce.
-
-# Run the examples in the docstrings listed in utils/documentation_tests.txt. Hardware and GPU examples are
-# skipped by content (see src/lerobot/utils/doctest_utils.py); CI sets both flags.
-doctest:
-	@files=$$(grep -v '^\s*#' utils/documentation_tests.txt | grep -v '^\s*$$'); \
-	if [ -z "$$files" ]; then \
-		echo "utils/documentation_tests.txt lists no files; nothing to run."; \
-	else \
-		SKIP_HARDWARE_DOCTEST=1 uv run pytest --doctest-modules --no-header -q $$files; \
-	fi
-
-check-doctest-list:
-	uv run python utils/check_doctest_list.py
-
-fix-doctest-list:
-	uv run python utils/check_doctest_list.py --fix_and_overwrite
-
-check-docstrings:
-	uv run python utils/check_docstrings.py
-	uv run python utils/check_config_docstrings.py
-
-fix-docstrings:
-	uv run python utils/check_docstrings.py --fix_and_overwrite
-	uv run python utils/check_doctest_list.py --fix_and_overwrite
@@ -128,23 +128,6 @@ lerobot-eval \

 Learn how to implement your own simulation environment or benchmark and distribute it from the HF Hub by following the [EnvHub Documentation](https://huggingface.co/docs/lerobot/envhub).

-### Third-Party Hardware
-
-Beyond the natively supported hardware, the community maintains a growing ecosystem of plugins for other robots, teleoperators, cameras, and sensors - UFACTORY xArm, Universal Robots UR5e, Franka, AgileX Piper, Trossen WidowX, ARX5, I2RT YAM, GELLO, SpaceMouse, Meta Quest, ROS 2 bridges, tactile and depth cameras, and more.
-
-Plugins are auto-discovered by package name: LeRobot imports any installed package prefixed with `lerobot_robot_`, `lerobot_teleoperator_`, or `lerobot_camera_`. Install one and use the `type` it registers straight from the CLI:
-
-```bash
-pip install lerobot_robot_<name> lerobot_teleoperator_<name>
-
-lerobot-record \
-  --robot.type=<robot_name> \
-  --teleop.type=<teleoperator_name> \
-  --dataset.repo_id=${HF_USER}/my-dataset
-```
-
-Browse the full list in the [Third-Party Robots & Teleoperators](https://huggingface.co/docs/lerobot/main/third_party_robots) and [Third-Party Cameras & Sensors](https://huggingface.co/docs/lerobot/main/third_party_sensors) documentation.
-
 ## Resources

 - **[Documentation](https://huggingface.co/docs/lerobot/index):** The complete guide to tutorials & API.
@@ -1,60 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""Root conftest: makes doctest collection use LeRobot's parser.
-
-This only affects `--doctest-modules` runs (see `make doctest`). The test suite itself is configured by
-`tests/conftest.py`.
-"""
-
-import doctest
-
-import _pytest.doctest
-
-from lerobot.utils.doctest_utils import LeRobotDoctestModule, LeRobotDocTestParser
-
-# Lets an example opt out of output comparison with `# doctest: +IGNORE_RESULT`, for calls whose output is
-# a progress bar or otherwise not reproducible.
-IGNORE_RESULT = doctest.register_optionflag("IGNORE_RESULT")
-
-OutputChecker = doctest.OutputChecker
-
-
-class CustomOutputChecker(OutputChecker):
-    """An output checker that honours the `IGNORE_RESULT` flag."""
-
-    def check_output(self, want, got, optionflags):
-        """Return `True` when `IGNORE_RESULT` is set, otherwise defer to stdlib.
-
-        Args:
-            want (`str`):
-                The expected output.
-            got (`str`):
-                The actual output.
-            optionflags (`int`):
-                Bitmask of active doctest option flags.
-
-        Returns:
-            `bool`: Whether the output is considered a match.
-        """
-        if IGNORE_RESULT & optionflags:
-            return True
-        return OutputChecker.check_output(self, want, got, optionflags)
-
-
-# Reassigning these module attributes is how doctest behaviour is customised; mypy sees it as assigning to
-# a type, which is exactly what is intended here.
-doctest.OutputChecker = CustomOutputChecker  # type: ignore[misc]
-_pytest.doctest.DoctestModule = LeRobotDoctestModule
-doctest.DocTestParser = LeRobotDocTestParser  # type: ignore[misc]
@@ -68,16 +68,17 @@ ENV HOME=/home/user_lerobot \
 # issues with MuJoCo and OpenGL drivers.
 RUN uv venv --python python${PYTHON_VERSION}

-# Install third-party dependencies separately for layer caching
+# Install Python dependencies for caching
 COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
-RUN uv sync --locked --extra all --no-install-project --no-cache
+COPY --chown=user_lerobot:user_lerobot src/ src/
+
+RUN uv sync --locked --extra all --no-cache

 RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas

-# Copy the application source code and install the local project
+# Copy the rest of the application source code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .
-RUN uv sync --locked --extra all --no-cache

 # Set the default command
 CMD ["/bin/bash"]
@@ -60,14 +60,15 @@ ENV HOME=/home/user_lerobot \
 # run other Python projects in the same container without dependency conflicts.
 RUN uv venv

-# Install third-party dependencies separately for layer caching
+# Install Python dependencies for caching
 COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
-RUN uv sync --locked --extra all --no-install-project --no-cache
+COPY --chown=user_lerobot:user_lerobot src/ src/

-# Copy the application code and install the local project
+RUN uv sync --locked --extra all --no-cache
+
+# Copy the rest of the application code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .
-RUN uv sync --locked --extra all --no-cache

 # Set the default command
 CMD ["/bin/bash"]
@@ -37,6 +37,8 @@
 - sections:
  - local: lerobot-dataset-v3
    title: Using LeRobotDataset
+  - local: training_dataset_streaming
+    title: Training Dataset Streaming
  - local: porting_datasets_v3
    title: Porting Large Datasets
  - local: using_dataset_tools
@@ -165,8 +167,6 @@
    title: OpenArm
  - local: rebot_b601
    title: reBot B601-DM
-  - local: third_party_robots
-    title: Third-Party Robots & Teleoperators
  title: "Robots"
 - sections:
  - local: phone_teleop
@@ -177,8 +177,6 @@
 - sections:
  - local: cameras
    title: Cameras
-  - local: third_party_sensors
-    title: Third-Party Cameras & Sensors
  title: "Sensors"
 - sections:
  - local: notebooks
@@ -191,28 +189,6 @@
 - sections:
  - local: contributing
    title: Contribute to LeRobot
-  - local: writing_docstrings
-    title: Writing docstrings
  - local: backwardcomp
    title: Backward compatibility
  title: "About"
- sections:
-  - local: api/robots
-    title: Robots
-  - local: api/teleoperators
-    title: Teleoperators
-  - local: api/cameras
-    title: Cameras
-  - local: api/motors
-    title: Motors
-  - local: api/datasets
-    title: Datasets
-  - local: api/policies
-    title: Policies
-  - local: api/processor
-    title: Processors
-  - local: api/envs
-    title: Environments
-  - local: api/configs
-    title: Configuration
-  title: "API Reference"
@@ -1,73 +0,0 @@
-# Cameras
-
-Cameras supply the image observations a policy sees. Every backend — OpenCV, Intel RealSense, Reachy 2 —
-implements the [`Camera`] interface, so swapping hardware does not change the code that reads frames.
-
-See the [Cameras guide](../cameras) for choosing and configuring a camera, and
-[Third-Party Cameras & Sensors](../third_party_sensors) for devices outside the core set.
-
-## Camera
-
-[[autodoc]] lerobot.cameras.Camera
-    - connect
-    - disconnect
-    - read
-    - async_read
-    - find_cameras
-
-## CameraConfig
-
-[[autodoc]] lerobot.cameras.CameraConfig
-
-## make_cameras_from_configs
-
-[[autodoc]] lerobot.cameras.make_cameras_from_configs
-
-## ColorMode
-
-[[autodoc]] lerobot.cameras.ColorMode
-
-## Cv2Rotation
-
-[[autodoc]] lerobot.cameras.Cv2Rotation
-
-## Cv2Backends
-
-[[autodoc]] lerobot.cameras.Cv2Backends
-
-## OpenCVCamera
-
-The default backend for USB webcams and video files, built on OpenCV.
-
-[[autodoc]] lerobot.cameras.opencv.OpenCVCamera
-    - all
-
-[[autodoc]] lerobot.cameras.opencv.OpenCVCameraConfig
-
-## RealSenseCamera
-
-Intel RealSense cameras, with optional depth sensing.
-
-[[autodoc]] lerobot.cameras.realsense.RealSenseCamera
-    - all
-
-[[autodoc]] lerobot.cameras.realsense.RealSenseCameraConfig
-
-## Reachy2Camera
-
-Cameras exposed by a Reachy 2 robot's own camera manager.
-
-[[autodoc]] lerobot.cameras.reachy2_camera.Reachy2Camera
-    - all
-
-[[autodoc]] lerobot.cameras.reachy2_camera.Reachy2CameraConfig
-
-## ZMQCamera
-
-Reads frames published over a ZeroMQ socket by `ImageServer`, for cameras attached to a different machine
-than the one running the policy.
-
-[[autodoc]] lerobot.cameras.zmq.ZMQCamera
-    - all
-
-[[autodoc]] lerobot.cameras.zmq.ZMQCameraConfig
@@ -1,27 +0,0 @@
-# Configuration
-
-LeRobot configuration is plain dataclasses parsed by [draccus](https://github.com/dlwh/draccus), so every
-field is settable from the CLI. [`TrainPipelineConfig`] is the top-level object for `lerobot-train`.
-
-Polymorphic configs (policies, robots, environments) use `draccus.ChoiceRegistry`: a subclass registers
-itself with `@register_subclass("name")` and is then selectable by that name on the command line.
-
-## TrainPipelineConfig
-
-[[autodoc]] lerobot.configs.train.TrainPipelineConfig
-
-## PreTrainedConfig
-
-[[autodoc]] lerobot.configs.PreTrainedConfig
-
-## DatasetConfig
-
-[[autodoc]] lerobot.configs.DatasetConfig
-
-## EvalConfig
-
-[[autodoc]] lerobot.configs.EvalConfig
-
-## WandBConfig
-
-[[autodoc]] lerobot.configs.WandBConfig
@@ -1,23 +0,0 @@
-# Datasets
-
-[`LeRobotDataset`] is the format every LeRobot script reads and writes. It is episode-aware, decodes video
-observations on the fly, and round-trips to the Hugging Face Hub.
-
-See [Using LeRobotDataset](../lerobot-dataset-v3) for the format and the common operations,
-[Porting Large Datasets](../porting_datasets_v3) for migration, and [Tools](../tools) for the CLI.
-
-## LeRobotDataset
-
-[[autodoc]] lerobot.datasets.LeRobotDataset
-
-## LeRobotDatasetMetadata
-
-[[autodoc]] lerobot.datasets.LeRobotDatasetMetadata
-
-## MultiLeRobotDataset
-
-[[autodoc]] lerobot.datasets.MultiLeRobotDataset
-
-## StreamingLeRobotDataset
-
-[[autodoc]] lerobot.datasets.StreamingLeRobotDataset
@@ -1,19 +0,0 @@
-# Environments
-
-Simulation environments are configured through [`EnvConfig`] and built by [`make_env`]. Each subclass
-declares its `gym_kwargs` and how to construct the vectorised environments.
-
-See [Environments from the Hub](../envhub) for using published environments and
-[Adding a New Benchmark](../adding_benchmarks) for contributing one.
-
-## EnvConfig
-
-[[autodoc]] lerobot.envs.EnvConfig
-
-## make_env
-
-[[autodoc]] lerobot.envs.make_env
-
-## make_env_config
-
-[[autodoc]] lerobot.envs.make_env_config
@@ -1,23 +0,0 @@
-# Motors
-
-`MotorsBus` is the low-level interface to a chain of servos on a serial bus. Robots use it to read positions
-and write goal positions; you rarely touch it directly unless you are adding hardware.
-
-See [Bring Your Own Hardware](../integrate_hardware) for adding a new bus, and
-[Updating Feetech Firmware](../feetech) and [Damiao Motors and CAN Bus](../damiao) for device-specific notes.
-
-## MotorsBus
-
-[[autodoc]] lerobot.motors.motors_bus.MotorsBus
-
-## Motor
-
-[[autodoc]] lerobot.motors.Motor
-
-## MotorCalibration
-
-[[autodoc]] lerobot.motors.MotorCalibration
-
-## MotorNormMode
-
-[[autodoc]] lerobot.motors.MotorNormMode
@@ -1,20 +0,0 @@
-# Policies
-
-Every policy inherits [`PreTrainedPolicy`], which combines a `torch.nn.Module` with the Hub mixin, so any
-policy can be pushed to and loaded from the Hugging Face Hub with the same two calls.
-
-Each policy has its own guide with training recipes and results — [ACT](../act), [SmolVLA](../smolvla),
-[π₀](../pi0), [π₀.₅](../pi05) and the rest are listed under Policies. To add one, see
-[Adding a Policy](../bring_your_own_policies).
-
-## PreTrainedPolicy
-
-[[autodoc]] lerobot.policies.pretrained.PreTrainedPolicy
-
-## PreTrainedConfig
-
-[[autodoc]] lerobot.configs.PreTrainedConfig
-
-## make_policy
-
-[[autodoc]] lerobot.policies.factory.make_policy
@@ -1,20 +0,0 @@
-# Processors
-
-Processors are the data transformation layer between a robot, a dataset and a policy. A pipeline is a chain
-of [`ProcessorStep`]s; each step declares how it transforms both the data and the feature contract.
-
-See [Introduction to Robot Processors](../introduction_processors) for the concepts,
-[Implement your own processor](../implement_your_own_processor) to write a step, and
-[Debug your processor pipeline](../debug_processor_pipeline) when a pipeline misbehaves.
-
-## ProcessorStep
-
-[[autodoc]] lerobot.processor.pipeline.ProcessorStep
-
-## DataProcessorPipeline
-
-[[autodoc]] lerobot.processor.pipeline.DataProcessorPipeline
-
-## PolicyProcessorPipeline
-
-[[autodoc]] lerobot.processor.pipeline.PolicyProcessorPipeline
@@ -1,147 +0,0 @@
-# Robots
-
-Every robot in LeRobot implements the [`Robot`] interface: connect, read an observation, send an action,
-disconnect. Writing a policy or a recording script against that interface means it works with any supported
-arm without change.
-
-This page is the generated reference. For wiring, calibration and first-run instructions, start with the
-hardware guides — [SO-101](../so101), [LeKiwi](../lekiwi), [Hope Jr](../hope_jr), [Reachy 2](../reachy2),
-[OpenArm](../openarm) — or [Imitation Learning for Robots](../il_robots) for the end-to-end workflow. To add
-a robot of your own, see [Bring Your Own Hardware](../integrate_hardware).
-
-## Robot
-
-The abstract base class. Subclasses implement every method below; the contract described here is what a
-policy or recording loop can rely on.
-
-[[autodoc]] lerobot.robots.Robot
-    - connect
-    - disconnect
-    - configure
-    - calibrate
-    - get_observation
-    - send_action
-    - observation_features
-    - action_features
-    - is_connected
-    - is_calibrated
-
-## RobotConfig
-
-[[autodoc]] lerobot.robots.RobotConfig
-
-## make_robot_from_config
-
-[[autodoc]] lerobot.robots.make_robot_from_config
-
-## SO-100 and SO-101 followers
-
-`SO100Follower` and `SO101Follower` are aliases of the same `SOFollower` class; the two arms differ in their
-configuration, not their control code. `SO100FollowerConfig` and `SO101FollowerConfig` are likewise aliases
-of `SOFollowerRobotConfig`.
-
-[[autodoc]] lerobot.robots.so_follower.SOFollower
-    - all
-
-[[autodoc]] lerobot.robots.so_follower.SOFollowerRobotConfig
-
-## BiSOFollower
-
-Two SO followers driven as one bimanual robot.
-
-[[autodoc]] lerobot.robots.bi_so_follower.BiSOFollower
-    - all
-
-[[autodoc]] lerobot.robots.bi_so_follower.BiSOFollowerConfig
-
-## KochFollower
-
-[[autodoc]] lerobot.robots.koch_follower.KochFollower
-    - all
-
-[[autodoc]] lerobot.robots.koch_follower.KochFollowerConfig
-
-## LeKiwi
-
-`LeKiwi` runs on the robot itself. `LeKiwiClient` is the host-side proxy that talks to it over the network
-and presents the same [`Robot`] interface.
-
-[[autodoc]] lerobot.robots.lekiwi.LeKiwi
-    - all
-
-[[autodoc]] lerobot.robots.lekiwi.LeKiwiConfig
-
-[[autodoc]] lerobot.robots.lekiwi.LeKiwiClient
-    - all
-
-[[autodoc]] lerobot.robots.lekiwi.LeKiwiClientConfig
-
-## OpenArmFollower
-
-[[autodoc]] lerobot.robots.openarm_follower.OpenArmFollower
-    - all
-
-[[autodoc]] lerobot.robots.openarm_follower.OpenArmFollowerConfig
-
-## BiOpenArmFollower
-
-[[autodoc]] lerobot.robots.bi_openarm_follower.BiOpenArmFollower
-    - all
-
-[[autodoc]] lerobot.robots.bi_openarm_follower.BiOpenArmFollowerConfig
-
-## OmxFollower
-
-[[autodoc]] lerobot.robots.omx_follower.OmxFollower
-    - all
-
-[[autodoc]] lerobot.robots.omx_follower.OmxFollowerConfig
-
-## Reachy2Robot
-
-[[autodoc]] lerobot.robots.reachy2.Reachy2Robot
-    - all
-
-[[autodoc]] lerobot.robots.reachy2.Reachy2RobotConfig
-
-## UnitreeG1
-
-[[autodoc]] lerobot.robots.unitree_g1.UnitreeG1
-    - all
-
-[[autodoc]] lerobot.robots.unitree_g1.UnitreeG1Config
-
-## Hope Jr
-
-The Hope Jr humanoid is exposed as two independent robots, an arm and a hand.
-
-[[autodoc]] lerobot.robots.hope_jr.HopeJrArm
-    - all
-
-[[autodoc]] lerobot.robots.hope_jr.HopeJrArmConfig
-
-[[autodoc]] lerobot.robots.hope_jr.HopeJrHand
-    - all
-
-[[autodoc]] lerobot.robots.hope_jr.HopeJrHandConfig
-
-## RebotB601Follower
-
-[[autodoc]] lerobot.robots.rebot_b601_follower.RebotB601Follower
-    - all
-
-[[autodoc]] lerobot.robots.rebot_b601_follower.RebotB601FollowerRobotConfig
-
-## BiRebotB601Follower
-
-[[autodoc]] lerobot.robots.bi_rebot_b601_follower.BiRebotB601Follower
-    - all
-
-[[autodoc]] lerobot.robots.bi_rebot_b601_follower.BiRebotB601FollowerConfig
-
-## EarthRoverMiniPlus
-
-[[autodoc]] lerobot.robots.earthrover_mini_plus.EarthRoverMiniPlus
-    - all
-
-[[autodoc]] lerobot.robots.earthrover_mini_plus.EarthRoverMiniPlusConfig
@@ -1,30 +0,0 @@
-# Teleoperators
-
-A teleoperator produces actions for a robot to follow — a leader arm, a gamepad, a keyboard, a phone. All of
-them implement the [`Teleoperator`] interface, so a recording script written against it works with any input
-device.
-
-See [Phone teleoperation](../phone_teleop) and [Isaac Teleop](../isaac_teleop) for setup guides, and
-[Imitation Learning for Robots](../il_robots) for the recording workflow.
-
-## Teleoperator
-
-[[autodoc]] lerobot.teleoperators.Teleoperator
-    - connect
-    - disconnect
-    - configure
-    - calibrate
-    - get_action
-    - send_feedback
-    - action_features
-    - feedback_features
-    - is_connected
-    - is_calibrated
-
-## TeleoperatorConfig
-
-[[autodoc]] lerobot.teleoperators.TeleoperatorConfig
-
-## make_teleoperator_from_config
-
-[[autodoc]] lerobot.teleoperators.make_teleoperator_from_config
@@ -65,7 +65,7 @@ In summary, you need to specify instructions for:
 Importantly,

 - `actions_per_chunk` and `chunk_size_threshold` are key parameters to tune for your setup.
- `aggregate_fn_name` is the function to aggregate actions on overlapping portions. You can either add a new one to a registry of functions, or add your own in `robot_client.py` (see [here](https://github.com/huggingface/lerobot/blob/main/src/lerobot/async_inference/robot_client.py#L224))
+- `aggregate_fn_name` is the function to aggregate actions on overlapping portions. You can either add a new one to a registry of functions, or add your own in `robot_client.py` (see [here](NOTE:addlinktoLOC))
 - `debug_visualize_queue_size` is a useful tool to tune the `CLIENT` parameters.

 ## Done! You should see your robot moving around by now 😉
@@ -58,7 +58,7 @@ final_action = postprocessor(action)

 ## Hardware API redesign

-PR [#777](https://github.com/huggingface/lerobot/pull/777) improves the LeRobot calibration but is **not backward-compatible**. Below is an overview of what changed and how you can continue to work with datasets created before this pull request.
+PR [#777](https://github.com/huggingface/lerobot/pull/777) improves the LeRobot calibration but is **not backward-compatible**. Below is a overview of what changed and how you can continue to work with datasets created before this pull request.

 ### What changed?

@@ -129,8 +129,8 @@ python examples/backward_compatibility/replay.py \

 Policies output actions in the same format as the datasets (`torch.Tensors`). Therefore, the same transformations should be applied.

-To find these transformations, we recommend first replaying an episode of the dataset your policy was trained on using the section above.
-Then, add these same transformations to your inference script (shown here in the `record.py` script):
+To find these transformations, we recommend to first try and and replay an episode of the dataset your policy was trained on using the section above.
+Then, add these same transformations on your inference script (shown here in the `record.py` script):

 ```diff
 action_values = predict_action(
@@ -161,16 +161,6 @@ The methods called by the train/eval loops:

 Batches are flat dictionaries keyed by the constants in [`lerobot.utils.constants`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/utils/constants.py): `OBS_STATE` (`observation.state.<motor>`), `OBS_IMAGES` (`observation.images.<camera>`), `OBS_LANGUAGE`, `ACTION`, etc. Reuse the constants — don't invent new prefixes.

-If your model is large enough to warrant [sharded multi-GPU training](./multi_gpu_training#sharded-training-fsdp), also declare its FSDP wrap units — the repeated block classes sharding operates on:
-
-```python
-class MyPolicy(PreTrainedPolicy):
-    ...
-    _fsdp_wrap_modules = ["MyTransformerBlock"]
-```
-
-With this one declaration, `--parallelism.dp_shard=N` works out of the box for your policy (users can still override it with `--accelerator.fsdp.wrap_modules`). Without any wrap source, sharded runs fail at startup by design.
-
 ### Processor functions

 LeRobot uses `PolicyProcessorPipeline`s to normalize inputs and de-normalize outputs around your policy. For a concrete reference, see [`processor_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/processor_act.py) or [`processor_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/processor_diffusion.py).
@@ -310,7 +300,7 @@ The file names are load-bearing: the factory does lazy imports by name, and the
 Two places need to know about your policy. All by name.

 1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. This import is what registers your policy: `@PreTrainedConfig.register_subclass("my_policy")` runs, and from then on the factory resolves everything by convention. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast).
-2. **`templates/lerobot_modelcard_template.md` and the root `README.md`** — the template is what the end-of-training publisher renders into the model card of every checkpoint trained with your policy: add a one-line description of your policy in the `model_name` branches, map it in `policy_docs` so cards link to your MDX guide, and optionally add an architecture image to `diagrams`. Then add your policy to the models table in the root `README.md`, under the right category, linking to your doc page.
+2. **`templates/lerobot_modelcard_template.md` and the root `README.md`** — the template is what `push_model_to_hub` renders into the model card of every checkpoint trained with your policy: add a one-line description of your policy in the `model_name` branches, map it in `policy_docs` so cards link to your MDX guide, and optionally add an architecture image to `diagrams`. Then add your policy to the models table in the root `README.md`, under the right category, linking to your doc page.

 Mirror an existing policy that's structurally similar to yours; the diff is small.

@@ -354,7 +344,7 @@ A new policy is much easier to review — and far more useful — when it ships

 **Pick at least one in-tree benchmark.** LeRobot ships sim benchmarks with per-benchmark Docker images (LIBERO, LIBERO-plus, Meta-World, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, VLABench and more). Pick the one that matches your policy's modality — VLAs usually go to LIBERO or VLABench; image-only BC to LIBERO or Meta-World. The full list lives under [Benchmarks](./libero) in the docs sidebar.

-**Push the checkpoint & processors** to the Hub under `lerobot/<policy>_<benchmark>` (or your namespace if you don't have write access; a maintainer can mirror it). The easiest way is training with `--policy.repo_id=<namespace>/<repo>` and `--policy.push_to_hub=true`: `lerobot-train` publishes the model, both processors, and a model card at the end of the run. To publish an existing checkpoint after the fact, upload its `pretrained_model/` directory (e.g. `huggingface-cli upload`), or use `lerobot-convert-dcp --push_to_hub=...` for sharded-format checkpoints.
+**Push the checkpoint & processors** to the Hub under `lerobot/<policy>_<benchmark>` (or your namespace if you don't have write access; a maintainer can mirror it). Use `PreTrainedPolicy.push_model_to_hub` so the repo gets `config.json`, `model.safetensors`, and a model card.

 **Report results in your policy's MDX**, with the exact `lerobot-eval` command and hardware so anyone can re-run:

@@ -136,10 +136,6 @@ config = RealSenseCameraConfig(
    height=480,
    color_mode=ColorMode.RGB,
    use_depth=True,
-    # Optional fixed color controls. Omit them to leave the current sensor settings unchanged.
-    exposure=120,
-    gain=64,
-    white_balance=4600,
    rotation=Cv2Rotation.NO_ROTATION
 )

@@ -158,15 +154,6 @@ finally:
 ```
 <!-- prettier-ignore-end -->

-Manual color controls disable the corresponding automatic exposure or white-balance mode. Their
-supported ranges vary by camera model; an invalid value raises an error at connection time that
-includes the range reported by the sensor. Requesting an unsupported control also raises an error.
-Omitted controls leave the sensor's existing automatic or manual setting unchanged. These options
-require `use_rgb=True`.
-
-Manual color controls require a dedicated RGB module. Cameras without one, such as the RealSense
-D405, do not support them and raise an error at connection time.
-
 </hfoption>
 </hfoptions>

@@ -88,6 +88,20 @@ policy_preprocessor = NormalizerProcessorStep(stats=dataset_stats)

 The same policy can work with different environment processors, and the same environment processor can work with different policies:

+````python
+# Use SmolVLA policy with LIBERO environment
+# Use SmolVLA policy with LIBERO environment
+libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
+    env_cfg=libero_cfg,
+    policy_cfg=smolvla_cfg,
+)
+smolvla_preprocessor, smolvla_postprocessor = make_pre_post_processors(smolvla_cfg)
+# Or use ACT policy with the same LIBERO environment
+libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
+    env_cfg=libero_cfg,
+    policy_cfg=act_cfg,
+)
+act_preprocessor, act_postprocessor = make_pre_post_processors(act_cfg)
 ```python
 # Use SmolVLA policy with LIBERO environment
 libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
@@ -102,7 +116,6 @@ libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
    policy_cfg=act_cfg,
 )
 act_preprocessor, act_postprocessor = make_pre_post_processors(act_cfg)
-```

 ### 3. **Easier Experimentation**

@@ -132,7 +145,7 @@ class LiberoVelocityProcessorStep(ObservationProcessorStep):
        state = torch.cat([eef_pos, eef_axisangle, eef_vel,
                          gripper_pos, gripper_vel], dim=-1)  # 14D
        return state
-```
+````

 ### 4. **Cleaner Environment Code**

@@ -23,18 +23,18 @@ The broader EVO1 project may include additional training scripts and dataset too
 2. Install EVO1 dependencies:

   ```bash
-   pip install -e ".[training,evo1]"
+   pip install -e ".[evo1]"
   ```

-   For LIBERO training and evaluation, install the LIBERO extra as well:
+   For LIBERO evaluation, install the LIBERO extra as well:

   ```bash
-   pip install -e ".[training,evo1,libero]"
+   pip install -e ".[evo1,libero]"
   ```

 3. Install a `flash-attn` wheel only if it is compatible with your Python, PyTorch, CUDA, and GPU stack. EVO1 falls back to standard attention when `flash_attn` is not available.

-EVO1 uses the native Hugging Face `transformers` InternVL implementation, so `policy.vlm_model_name` must point to a natively converted checkpoint such as `OpenGVLab/InternVL3-1B-hf` (note the `-hf` suffix). The first run downloads the configured VLM checkpoint and later runs reuse it from the Hugging Face cache.
+EVO1 uses the native Hugging Face `transformers` InternVL implementation, so `policy.vlm_model_name` must point to a natively converted checkpoint such as `OpenGVLab/InternVL3-1B-hf` (note the `-hf` suffix). The first run may download the configured VLM checkpoint unless `policy.vlm_model_name` points to a local model directory.

 ## Data Requirements

@@ -92,7 +92,7 @@ lerobot-train \

 ### Stage 2

-Stage 2 loads the Stage 1 policy, but starts a fresh optimizer and scheduler:
+Stage 2 finetunes the VLM branches and action head. A common workflow starts from a Stage 1 checkpoint:

 ```bash
 lerobot-train \
@@ -152,154 +152,16 @@ lerobot-rollout \

 ### LIBERO Evaluation

-#### Reference result
-
 > [!NOTE]
-> The released Stage-2 checkpoint passed clean-download and rollout verification:
-> [`zuoxingdong/evo1_libero`](https://huggingface.co/zuoxingdong/evo1_libero), revision
-> [`515921f4a2c1d3f3ad523721eafa26fdf2af315b`](https://huggingface.co/zuoxingdong/evo1_libero/commit/515921f4a2c1d3f3ad523721eafa26fdf2af315b).
-> The clean-download evaluation used LeRobot revision
-> [`e40b58a8dfa9e7b86918c374791599d070518d11`](https://github.com/huggingface/lerobot/commit/e40b58a8dfa9e7b86918c374791599d070518d11).
+> Benchmark results for a `lerobot`-hosted LIBERO checkpoint trained with this implementation
+> will be added once training completes.

-The single-run Stage-2 checkpoint at step 70,000 produced:
-
-| Suite          | Successful episodes |  Episodes | Success rate |
-| -------------- | ------------------: | --------: | -----------: |
-| LIBERO Spatial |                 485 |       500 |        97.0% |
-| LIBERO Object  |                 496 |       500 |        99.2% |
-| LIBERO Goal    |                 483 |       500 |        96.6% |
-| LIBERO-10      |                 469 |       500 |        93.8% |
-| **Overall**    |           **1,933** | **2,000** |   **96.65%** |
-
-These results use one trained checkpoint and evaluation seed `1000`; they are not a multi-seed
-mean or confidence estimate.
-
-#### Reference training recipe
-
-The released checkpoint records the complete resolved Stage-2 configuration in
-[`train_config.json`](https://huggingface.co/zuoxingdong/evo1_libero/blob/515921f4a2c1d3f3ad523721eafa26fdf2af315b/train_config.json).
-The measured run used two H100 GPUs with two DDP processes and batch 64 per process, giving global batch 128. Both stages used the same topology. The base VLM came from revision
-`014c0583a0d4bedf29fbe2dbff4f865eb998e171` of `OpenGVLab/InternVL3-1B-hf`.
-The released artifact does not record the exact LeRobot training commit or its original dependency lock,
-so the commands below reproduce the recorded configuration and topology from a current checkout rather
-than reconstructing the software environment bit for bit.
-
-From a LeRobot source checkout, install the locked dependencies and download that exact VLM revision:
-
-```bash
-uv sync --locked --extra training --extra evo1 --extra libero
-VLM_DIR=$(uv run hf download OpenGVLab/InternVL3-1B-hf \
-  --revision=014c0583a0d4bedf29fbe2dbff4f865eb998e171)
-```
-
-Stage 1 freezes the VLM and trains the action head for 5,000 steps:
-
-```bash
-uv run accelerate launch --num_processes=2 -m lerobot.scripts.lerobot_train \
-  --dataset.repo_id=lerobot/libero \
-  --dataset.revision=a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4 \
-  --dataset.video_backend=torchcodec \
-  --dataset.return_uint8=true \
-  --dataset.image_transforms.enable=true \
-  --dataset.use_imagenet_stats=true \
-  --dataset.eval_split=0.0 \
-  --policy.type=evo1 \
-  --policy.training_stage=stage1 \
-  --policy.apply_training_stage_defaults=true \
-  --policy.vlm_model_name="${VLM_DIR}" \
-  --policy.vlm_num_layers=14 \
-  --policy.vlm_dtype=bfloat16 \
-  --policy.device=cuda \
-  --policy.use_amp=true \
-  --policy.use_flash_attn=true \
-  --policy.enable_gradient_checkpointing=true \
-  --policy.gradient_checkpointing_use_reentrant=false \
-  --policy.image_resolution='[448,448]' \
-  --policy.chunk_size=50 \
-  --policy.n_action_steps=50 \
-  --policy.max_state_dim=24 \
-  --policy.max_action_dim=24 \
-  --policy.dropout=0.2 \
-  --policy.optimizer_lr=1e-5 \
-  --policy.optimizer_weight_decay=1e-3 \
-  --policy.optimizer_grad_clip_norm=1.0 \
-  --policy.scheduler_warmup_steps=1000 \
-  --policy.push_to_hub=false \
-  --use_policy_training_preset=true \
-  --batch_size=64 \
-  --steps=5000 \
-  --save_checkpoint=true \
-  --save_checkpoint_to_hub=false \
-  --save_freq=2500 \
-  --log_freq=10 \
-  --env_eval_freq=0 \
-  --num_workers=4 \
-  --prefetch_factor=2 \
-  --persistent_workers=true \
-  --seed=1000 \
-  --wandb.enable=false \
-  --output_dir=./outputs/evo1-libero-stage1-g128-5k
-```
-
-Stage 2 loads the Stage-1 policy but starts a fresh optimizer and scheduler. It trains for 80,000 steps;
-the reported checkpoint is the save at step 70,000:
-
-```bash
-uv run accelerate launch --num_processes=2 -m lerobot.scripts.lerobot_train \
-  --dataset.repo_id=lerobot/libero \
-  --dataset.revision=a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4 \
-  --dataset.video_backend=torchcodec \
-  --dataset.return_uint8=true \
-  --dataset.image_transforms.enable=true \
-  --dataset.use_imagenet_stats=true \
-  --dataset.eval_split=0.0 \
-  --policy.path=./outputs/evo1-libero-stage1-g128-5k/checkpoints/005000/pretrained_model \
-  --policy.training_stage=stage2 \
-  --policy.apply_training_stage_defaults=true \
-  --policy.vlm_model_name="${VLM_DIR}" \
-  --policy.vlm_num_layers=14 \
-  --policy.vlm_dtype=float32 \
-  --policy.device=cuda \
-  --policy.use_amp=true \
-  --policy.use_flash_attn=true \
-  --policy.enable_gradient_checkpointing=true \
-  --policy.gradient_checkpointing_use_reentrant=false \
-  --policy.image_resolution='[448,448]' \
-  --policy.chunk_size=50 \
-  --policy.n_action_steps=50 \
-  --policy.max_state_dim=24 \
-  --policy.max_action_dim=24 \
-  --policy.dropout=0.2 \
-  --policy.optimizer_lr=1e-5 \
-  --policy.optimizer_weight_decay=1e-3 \
-  --policy.optimizer_grad_clip_norm=1.0 \
-  --policy.scheduler_warmup_steps=1000 \
-  --policy.push_to_hub=false \
-  --use_policy_training_preset=true \
-  --batch_size=64 \
-  --steps=80000 \
-  --resume=false \
-  --save_checkpoint=true \
-  --save_checkpoint_to_hub=false \
-  --save_freq=10000 \
-  --log_freq=10 \
-  --env_eval_freq=0 \
-  --num_workers=4 \
-  --prefetch_factor=2 \
-  --persistent_workers=true \
-  --seed=1000 \
-  --wandb.enable=false \
-  --output_dir=./outputs/evo1-libero-stage2-g128-80k
-```
-
-#### Author-format evaluation profile
-
-The author-format EVO1 LIBERO profile uses the raw LIBERO camera feature names
+The official EVO1 LIBERO rollout protocol uses the raw LIBERO camera feature names
 (`observation.images.agentview_image` and `observation.images.robot0_eye_in_hand_image`), replans every
 14 actions, and binarizes the gripper command before stepping the simulator. The EVO1 policy postprocessor
 can crop the padded 24D action back to the 7D LIBERO action space and apply that gripper binarization. To
-evaluate an author-format checkpoint under the same one-episode-per-task setting, keep the raw camera names
-instead of the default `image`/`image2` mapping and set the LIBERO action postprocessing flags:
+evaluate a LIBERO checkpoint under the same one-episode-per-task setting, keep the raw camera names instead
+of the default `image`/`image2` mapping and set the LIBERO action postprocessing flags:

 ```bash
 lerobot-eval \
@@ -319,75 +181,6 @@ lerobot-eval \
  --eval.n_episodes=1
 ```

-#### Native `lerobot/libero` v3 profile
-
-Revision `a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4` stores camera features as `image` and
-`image2`. This example evaluates all ten LIBERO Object tasks, launching each task in a fresh process:
-
-```bash
-export MUJOCO_GL=egl
-export PYOPENGL_PLATFORM=egl
-
-suite=libero_object
-horizon=280
-for task_id in {0..9}; do
-  lerobot-eval \
-    --policy.path=zuoxingdong/evo1_libero \
-    --policy.pretrained_revision=515921f4a2c1d3f3ad523721eafa26fdf2af315b \
-    --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
-    --policy.device=cuda \
-    --policy.use_amp=true \
-    --policy.vlm_dtype=bfloat16 \
-    --policy.use_flash_attn=false \
-    --policy.enable_gradient_checkpointing=false \
-    --policy.vlm_num_layers=14 \
-    --policy.image_resolution='[448,448]' \
-    --policy.max_text_length=1024 \
-    --policy.chunk_size=50 \
-    --policy.n_action_steps=14 \
-    --policy.max_state_dim=24 \
-    --policy.max_action_dim=24 \
-    --policy.num_inference_timesteps=32 \
-    --policy.postprocess_action_dim=7 \
-    --policy.binarize_gripper=true \
-    --policy.gripper_threshold=0.0 \
-    --policy.gripper_below_threshold_value=-1.0 \
-    --policy.gripper_above_threshold_value=1.0 \
-    --env.type=libero \
-    --env.task="${suite}" \
-    --env.task_ids="[${task_id}]" \
-    --env.camera_name=agentview_image,robot0_eye_in_hand_image \
-    --env.camera_name_mapping="{agentview_image: image, robot0_eye_in_hand_image: image2}" \
-    --env.control_mode=relative \
-    --env.obs_type=pixels_agent_pos \
-    --env.observation_width=448 \
-    --env.observation_height=448 \
-    --env.init_states=true \
-    --env.episode_length="${horizon}" \
-    --env.render_mode=rgb_array \
-    --env.max_parallel_tasks=1 \
-    --eval.n_episodes=50 \
-    --eval.batch_size=1 \
-    --eval.use_async_envs=false \
-    --eval.recording=false \
-    --seed=1000 \
-    --output_dir="./outputs/evo1-libero-stage2-70k-eval/${suite}/task-${task_id}" \
-    --job_name="evo1-libero-stage2-70k-${suite}-task-${task_id}"
-done
-```
-
-Run all ten task IDs for each suite with these horizons:
-
-| `env.task`       | `env.episode_length` |
-| ---------------- | -------------------: |
-| `libero_spatial` |                `280` |
-| `libero_object`  |                `280` |
-| `libero_goal`    |                `300` |
-| `libero_10`      |                `520` |
-
-Set `suite` and `horizon` for each row. This gives 500 episodes per suite and 2,000 episodes overall, while
-the loop's fresh process per task matches the measured RNG-reset topology.
-
 ## References

 - [EVO1 repository](https://github.com/MINT-SJTU/Evo-1)
@@ -40,10 +40,10 @@ This tutorial guides you through updating the firmware of Feetech motors using t
 For each motor you want to update:

 1. **Select the motor** from the list by clicking on it
-2. **Click the Upgrade tab**:
-3. **Click the Online button**:
-   - If a potential firmware update is found, it will be displayed in the box
-4. **Click the Upgrade button**:
+2. **Click on Upgrade tab**:
+3. **Click on Online button**:
+   - If an potential firmware update is found, it will be displayed in the box
+4. **Click on Upgrade button**:
   - The update progress will be displayed

 ## Step 6: Verify Update
@@ -59,7 +59,6 @@ The `lerobot-rollout --strategy.type=dagger` mode requires **teleoperators with

 - `bi_openarm_mini` - Bimanual OpenArm Mini
 - `so_leader` - SO100 / SO101 leader arm
- `bi_so_leader` - Bimanual SO100 / SO101 leader arms

 > [!IMPORTANT]
 > The provided commands default to `bi_openarm_follower` + `bi_openarm_mini`.
@@ -211,7 +211,7 @@ Record, Replay and Train with Hope-JR is still experimental.

 ### Record

-This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data).
+This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data/settings).

 ```bash
 lerobot-record \
@@ -98,7 +98,7 @@ The teleoperate command will automatically:

 ## Cameras

-To add cameras to your setup, follow this [Guide](./cameras).
+To add cameras to your setup, follow this [Guide](./cameras#setup-cameras).

 ## Teleoperate with cameras

@@ -1,177 +1,23 @@
-# LeRobot
-
 <div class="flex justify-center">
  <a target="_blank" href="https://huggingface.co/lerobot">
    <img
-      alt="LeRobot, Hugging Face Robotics Library"
+      alt="HuggingFace Expert Acceleration Program"
      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-logo-thumbnail.png"
+      style="width: 100%"
    ></img>
  </a>
 </div>

+# LeRobot
+
 **State-of-the-art machine learning for real-world robotics**

-🤗 LeRobot provides a hardware-agnostic, Python-native interface for controlling real robots - from affordable arms like the SO-ARM101 to full humanoids. Plus the tools to record, store, and share the datasets they generate. Every dataset uses the standardized **LeRobotDataset** format (synchronized video + action/state data) and can be streamed directly from the [Hugging Face Hub](https://huggingface.co/lerobot).
+🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier for entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.

-🤗 On top of that data, LeRobot implements state-of-the-art policies - from lightweight imitation-learning models like ACT to large vision-language-action models like π₀ and SmolVLA - all trainable, shareable, and deployable with the same handful of CLI commands.
+🤗 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real-world with a focus on imitation learning and reinforcement learning.

-The goal: lower the barrier to entry for robotics, so that everyone can contribute to, and benefit from, shared datasets and pretrained models.
+🤗 LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulated environments so that everyone can get started.

-<div align="center" style="display: flex; justify-content: center; gap: 8px; flex-wrap: wrap; margin: 20px 0;">
-  <a href="https://discord.gg/s3KuuzsPFb" target="_blank">
-    <img alt="Discord" src="https://img.shields.io/badge/Discord-Join_the_Community-5865F2?style=flat&logo=discord&logoColor=white">
-  </a>
-  <a href="https://x.com/LeRobotHF" target="_blank">
-    <img alt="X (Twitter)" src="https://img.shields.io/badge/X-Follow_%40LeRobotHF-black?style=flat&logo=x&logoColor=white">
-  </a>
-  <a href="https://huggingface.co/lerobot" target="_blank">
-    <img alt="Hugging Face Hub" src="https://img.shields.io/badge/HF_Hub-Models_%26_Datasets-FFD21E?style=flat">
-  </a>
-</div>
+🤗 LeRobot hosts pretrained models and datasets on the LeRobot HuggingFace page.

-<div align="center">
-  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/robots_control_video.webp" width="640px" alt="Reachy 2 Demo">
-</div>
-
-## How It Works
-
-**Teleoperate → Record → Train → Deploy**
-
-1. **Teleoperate** - control the robot yourself (with a leader arm, keyboard, or phone) so it can learn from your movements.
-2. **Record** - each demonstration is saved as a dataset: synchronized camera video plus the actions you took.
-3. **Train** - a policy (the neural network that will control the robot) learns to imitate your demonstrations.
-4. **Deploy** - run the trained policy on the robot and watch it complete the task on its own.
-
-## Get Started
-
-New here? [Install LeRobot](./installation), then pick your path:
-
-<div class="grid grid-cols-1 md:grid-cols-3 gap-4 my-6">
-  <div class="border dark:border-gray-700 rounded-lg p-4 shadow">
-    <div class="text-lg font-semibold mb-2">🔧 I have a robot</div>
-    <p class="text-gray-700 dark:text-gray-300 text-sm">
-      LeRobot supports a wide range of arms and mobile robots. Popular picks:
-    </p>
-    <ul class="text-gray-700 dark:text-gray-300 text-sm list-disc pl-5 mb-2">
-      <li>
-        <a href="./so101">SO-101</a> - our flagship, low-cost arm
-      </li>
-      <li>
-        <a href="./lekiwi">LeKiwi</a> - a mobile base with an arm on top
-      </li>
-      <li>
-        <a href="./koch">Koch v1.1</a> - a long-time community favorite
-      </li>
-      <li>
-        or find yours under <strong>Robots</strong> in the sidebar
-      </li>
-    </ul>
-    <p class="text-gray-700 dark:text-gray-300 text-sm">
-      Once it's assembled and calibrated, record a dataset and train your first
-      policy with the <a href="./il_robots">imitation learning tutorial</a> - or
-      skip the CLI entirely with <a href="./lelab">LeLab</a>, a browser GUI for
-      the same workflow.
-    </p>
-  </div>
-  <div class="border dark:border-gray-700 rounded-lg p-4 shadow">
-    <div class="text-lg font-semibold mb-2">💻 No hardware yet</div>
-    <p class="text-gray-700 dark:text-gray-300 text-sm">
-      You can still train and evaluate policies without owning a robot:
-    </p>
-    <ul class="text-gray-700 dark:text-gray-300 text-sm list-disc pl-5 mb-2">
-      <li>
-        train on an existing
-        <a href="https://huggingface.co/datasets?other=LeRobot">
-          LeRobot dataset
-        </a>
-        from the Hub
-      </li>
-      <li>
-        evaluate in <a href="./envhub">simulation</a>, against benchmarks like
-        LIBERO or Meta-World
-      </li>
-      <li>
-        try the free <a href="./notebooks">Colab notebooks</a> - nothing to
-        install
-      </li>
-    </ul>
-  </div>
-  <div class="border dark:border-gray-700 rounded-lg p-4 shadow">
-    <div class="text-lg font-semibold mb-2">🤝 I want to contribute</div>
-    <p class="text-gray-700 dark:text-gray-300 text-sm">
-      Start with the <a href="./contributing">Contributing guide</a>, then
-      <a href="./bring_your_own_policies">add a new policy</a> or
-      <a href="./integrate_hardware">bring your own hardware</a>.
-    </p>
-  </div>
-</div>
-
-## Explore the Docs
-
-<div class="grid grid-cols-1 md:grid-cols-3 gap-4 my-6">
-  <a
-    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
-    href="./cheat-sheet"
-  >
-    <div class="font-semibold mb-1">📋 Cheat Sheet</div>
-    <p class="text-gray-700 dark:text-gray-300 text-sm">
-      Every LeRobot CLI command, copy-paste ready.
-    </p>
-  </a>
-  <a
-    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
-    href="./hardware_guide"
-  >
-    <div class="font-semibold mb-1">🖥️ Compute & Hardware Guide</div>
-    <p class="text-gray-700 dark:text-gray-300 text-sm">
-      Which policy fits your GPU, and how long training takes.
-    </p>
-  </a>
-  <a
-    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
-    href="./lerobot-dataset-v3"
-  >
-    <div class="font-semibold mb-1">🗂️ LeRobotDataset</div>
-    <p class="text-gray-700 dark:text-gray-300 text-sm">
-      Load, stream, and visualize robot datasets from the Hub.
-    </p>
-  </a>
-  <a
-    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
-    href="./lelab"
-  >
-    <div class="font-semibold mb-1">🖼 LeLab</div>
-    <p class="text-gray-700 dark:text-gray-300 text-sm">
-      A browser GUI for calibrating, recording, and training - no CLI required.
-    </p>
-  </a>
-  <a
-    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
-    href="./act"
-  >
-    <div class="font-semibold mb-1">🧠 Policies</div>
-    <p class="text-gray-700 dark:text-gray-300 text-sm">
-      Start with ACT, our recommended first policy - or browse SmolVLA, π₀, and
-      more in the sidebar.
-    </p>
-  </a>
-  <a
-    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
-    href="./envhub"
-  >
-    <div class="font-semibold mb-1">🎮 Simulation & Benchmarks</div>
-    <p class="text-gray-700 dark:text-gray-300 text-sm">
-      Train and evaluate in simulated environments before touching real
-      hardware.
-    </p>
-  </a>
-</div>
-
-## Common Problems
-
-Running into issues? A few of the most frequent ones:
-
- **Blurry or unusable camera footage** - lighting matters more than resolution. See the [Cameras](./cameras) guide.
- **Build or install errors** (`cmake`, `ffmpeg`, CUDA) - see the Troubleshooting section of the [Installation guide](./installation#troubleshooting).
- **Not sure which policy fits your GPU** - check the [Compute & Hardware Guide](./hardware_guide).
- **Still stuck?** Ask on [Discord](https://discord.gg/s3KuuzsPFb) - the community (and the LeRobot team) is there to help.
+Join the LeRobot community on [Discord](https://discord.gg/s3KuuzsPFb)
@@ -149,14 +149,13 @@ lerobot-rollout \

 Foot pedal input is also supported via `--strategy.input_device=pedal`. Configure pedal codes with `--strategy.pedal.*` flags.

-| Flag                                 | Description                                                                                                                                                    |
-| ------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `--strategy.num_episodes`            | Number of correction episodes to record (default: 10)                                                                                                          |
-| `--strategy.record_autonomous`       | Record autonomous frames too (default: false)                                                                                                                  |
-| `--strategy.upload_every_n_episodes` | Push to Hub every N episodes (default: 5)                                                                                                                      |
-| `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard)                                                                                                        |
-| `--strategy.smooth_handover`         | Smoothly hand control over at pause / correction start (default: true). Disable for clutch-style teleops that re-reference at the current robot pose on engage |
-| `--teleop.type`                      | **Required.** Teleoperator type                                                                                                                                |
+| Flag                                 | Description                                             |
+| ------------------------------------ | ------------------------------------------------------- |
+| `--strategy.num_episodes`            | Number of correction episodes to record (default: 10)   |
+| `--strategy.record_autonomous`       | Record autonomous frames too (default: false)           |
+| `--strategy.upload_every_n_episodes` | Push to Hub every N episodes (default: 5)               |
+| `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard) |
+| `--teleop.type`                      | **Required.** Teleoperator type                         |

 ### Episodic (`--strategy.type=episodic`)

@@ -187,15 +186,14 @@ Teleop is optional — if omitted the robot holds its position during the reset
 | `←` (left)  | Discard episode and re-record it |
 | `ESC`       | Stop the recording session       |

-| Flag                                            | Description                                                                                                                                                |
-| ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `--dataset.num_episodes`                        | Number of episodes to record                                                                                                                               |
-| `--dataset.episode_time_s`                      | Duration of each recording episode in seconds                                                                                                              |
-| `--dataset.reset_time_s`                        | Duration of the reset phase between episodes in seconds                                                                                                    |
-| `--teleop.type`                                 | Optional. Teleoperator to drive the robot during resets                                                                                                    |
-| `--strategy.reset_to_initial_position`          | Whether to reset the robot to its initial position between episodes                                                                                        |
-| `--strategy.smooth_leader_to_follower_handover` | Whether to turn on or off the leader -> follower smooth handover behavior.                                                                                 |
-| `--strategy.smooth_handover`                    | Smoothly hand control to the teleop at reset start (default: true). Disable for clutch-style teleops that re-reference at the current robot pose on engage |
+| Flag                                            | Description                                                                |
+| ----------------------------------------------- | -------------------------------------------------------------------------- |
+| `--dataset.num_episodes`                        | Number of episodes to record                                               |
+| `--dataset.episode_time_s`                      | Duration of each recording episode in seconds                              |
+| `--dataset.reset_time_s`                        | Duration of the reset phase between episodes in seconds                    |
+| `--teleop.type`                                 | Optional. Teleoperator to drive the robot during resets                    |
+| `--strategy.reset_to_initial_position`          | Whether to reset the robot to its initial position between episodes        |
+| `--strategy.smooth_leader_to_follower_handover` | Whether to turn on or off the leader -> follower smooth handover behavior. |

 ---

@@ -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/so_follower/so_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):

@@ -108,7 +108,6 @@ own binding plus a matching image block, e.g.

 ```yaml
 ask_vqa_top:
-  route: vqa
  bindings:
    vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
    vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
@@ -128,9 +127,7 @@ ask_vqa_top:
      }
 ```

-Add one such sub-recipe per camera the dataset records. The explicit
-`route: vqa` marker makes a matching sparse VQA annotation take precedence
-over normal weighted blend selection; component names are purely descriptive.
+Add one such sub-recipe per camera the dataset records.

 ## Layer 3 — training format

@@ -144,20 +141,7 @@ sample["target_message_indices"]

 The renderer does not apply a tokenizer chat template. Policy processors decide how to serialize the messages for their backbone, which keeps the same dataset usable across SmolVLA, Pi0.5, and any future VLM that expects OpenAI-style chat messages.

-## Blends
-
-Blend recipes select one weighted sub-recipe deterministically from the sample index.
-`recipes/subtask_mem.yaml` trains the compact core blend — high-level subtask prediction, low-level execution, and memory. `recipes/subtask_mem_vqa_speech.yaml` is the fuller variant that also adds VQA and spoken interjection responses.
-
-`recipes/subtask_joint.yaml` demonstrates joint sequence training rather than a
-weighted blend. For the same sample, its assistant subtask is supervised with
-text cross-entropy on the `low_level` stream while action prediction remains
-active, matching the joint setup from the π0.5 paper. Enable
-`--policy.joint_subtask_conditioning=true` to use that subtask conditioning at inference.
-
 ## Graceful absence

-If both language columns are missing, `None`, or empty, `RenderMessagesStep` uses
-the task string as low-level supervision when available and otherwise leaves the
-sample unchanged. For an annotated sample, if no recipe branch applies and no
-task fallback exists, rendering returns `None`, allowing a loader to retry another sample.
+If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.
+If an event-scoped branch is selected on a frame without the required event row, rendering returns `None`, allowing a loader to retry another sample.
@@ -51,7 +51,7 @@ In addition to these instructions, you need to install the Feetech SDK & ZeroMQ
 pip install -e ".[lekiwi]"
 ```

-Great 🤗! You are now done installing LeRobot, and we can begin assembling the SO100/SO101 arms and the mobile base 🤖.
+Great :hugs:! You are now done installing LeRobot, and we can begin assembling the SO100/SO101 arms and the mobile base :robot:.
 Every time you now want to use LeRobot, you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.

 # Step-by-Step Assembly Instructions
@@ -131,31 +131,54 @@ for batch in data_loader:
    # model.forward(batch)
 ```

-## Stream a dataset (no downloads)
+## Stream a dataset during training

-Use `StreamingLeRobotDataset` to iterate directly from the Hub without local copies. This allows to stream large datasets without the need to downloading them onto disk or loading them onto memory, and is a key feature of the new dataset format.
+Enable training-time streaming with the public `--dataset.streaming=true` flag:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=yaak-ai/L2D-v3 \
+  --dataset.streaming=true \
+  --policy.type=act \
+  --output_dir=outputs/train/act_streaming
+```
+
+This is separate from `--dataset.streaming_encoding=true`, which controls video encoding while
+recording. Training-time streaming leaves the map-style `LeRobotDataset` path unchanged.
+
+`StreamingLeRobotDataset` assigns complete episodes disjointly across distributed ranks and
+DataLoader workers, reads only the selected episode rows from Parquet, and keeps a bounded pool of
+episode video bytes. Every selected frame is visited exactly once per streaming epoch.
+
+On first use, LeRobot looks for a revision-matched MP4 index sidecar. If none is published with the
+dataset, it builds one in the revision-keyed local LeRobot cache under a process lock and installs it
+atomically. Training never uploads this sidecar. Dataset maintainers can build and publish one
+explicitly with `scripts/build_mp4_sidecar.py --push`.
+
+The default memory cap is 8 GiB per DataLoader worker. It can be adjusted along with episode mixing
+and prefetch:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=yaak-ai/L2D-v3 \
+  --dataset.streaming=true \
+  --dataset.streaming_byte_budget_gb=4 \
+  --dataset.streaming_episode_pool_size=16 \
+  --dataset.streaming_prefetch_episodes=4 \
+  --policy.type=act \
+  --output_dir=outputs/train/act_streaming
+```
+
+Use `--dataset.streaming_data_root=hf://buckets/OWNER/BUCKET/PREFIX` when metadata lives in a
+dataset repository but Parquet and MP4 data are mirrored in an HF Bucket.
+
+The Python API uses the same implementation:

 ```python
 from lerobot.datasets import StreamingLeRobotDataset

 repo_id = "yaak-ai/L2D-v3"
-dataset = StreamingLeRobotDataset(repo_id)  # streams directly from the Hub
-```
-
-Datasets stored in an [HF Storage Bucket](https://huggingface.co/docs/hub/storage-buckets) (`hf://buckets/`) can be streamed the same way by passing `repo_type="bucket"`:
-
-```python
-dataset = StreamingLeRobotDataset("my-org/my-bucket", repo_type="bucket")
-```
-
-Both options are available in `lerobot-train` through `--dataset.streaming=true`, and `--dataset.repo_type=bucket` to stream from a bucket instead of a Hub dataset repo:
-
-```bash
-lerobot-train \
-  --dataset.repo_id=my-org/my-bucket \
-  --dataset.repo_type=bucket \
-  --dataset.streaming=true \
-  ...
+dataset = StreamingLeRobotDataset(repo_id)
 ```

 <div style="display:flex; justify-content:center; gap:12px; flex-wrap:wrap;">
@@ -92,20 +92,6 @@ LIBERO supports two control modes — `relative` (default) and `absolute`. Diffe
 --env.control_mode=relative   # or "absolute"
 ```

-### Reset performance
-
-By default, LeRobot preserves LIBERO's hard-reset behavior. With fixed initial
-states enabled, you can opt into soft resets to skip rebuilding the simulator
-model and renderer on every episode:
-
-```bash
--env.init_states=true --env.hard_reset=false
-```
-
-Soft resets are faster but are not bit-identical to hard resets after the
-environment's settling steps, so camera observations and policy results may
-differ slightly. Use hard resets when reproducing benchmark results.
-
 ### Policy inputs and outputs

 **Observations:**
@@ -128,58 +114,38 @@ differ slightly. Use hard resets when reproducing benchmark results.

 ### Recommended evaluation episodes

-For reproducible benchmarking, use **10 episodes per task** across all four standard suites (Spatial, Object, Goal, Long). This gives 400 total episodes and matches the protocol used for published results. Success rates may vary by a few percent across evaluation seeds, so we recommend averaging over 3 seeds.
-
-<Tip>
-  To compare two policies on the same episodes, use the same `--seed`, keep
-  `--env.init_states=true`, and run each task in a single batch
-  (`--eval.batch_size` equal to episodes per task).
-</Tip>
+For reproducible benchmarking, use **10 episodes per task** across all four standard suites (Spatial, Object, Goal, Long). This gives 400 total episodes and matches the protocol used for published results.

 ## Training

 ### Dataset

-Two preprocessed LIBERO datasets are fully compatible with LeRobot. They contain the same demonstrations with the same schema and differ in how camera frames are stored:
+We provide a preprocessed LIBERO dataset fully compatible with LeRobot:

-|                           | [lerobot/libero](https://huggingface.co/datasets/lerobot/libero) | [HuggingFaceVLA/libero](https://huggingface.co/datasets/HuggingFaceVLA/libero) |
-| ------------------------- | ---------------------------------------------------------------- | ------------------------------------------------------------------------------ |
-| episodes / frames / tasks | 1,693 / 273,465 / 40                                             | 1,693 / 273,465 / 40                                                           |
-| cameras                   | 2× 256×256×3                                                     | 2× 256×256×3                                                                   |
-| state / action dims       | 8 / 7                                                            | 8 / 7                                                                          |
-| dataset format            | v3.0                                                             | v3.0                                                                           |
-| camera encoding           | MP4 video                                                        | PNG in parquet                                                                 |
-| download size             | **1.9 GB**                                                       | 69.9 GB                                                                        |
-| extra dependency          | video backend (`torchcodec` or `pyav`)                           | none                                                                           |
-
-**We recommend [lerobot/libero](https://huggingface.co/datasets/lerobot/libero)**: **37× smaller download** with **equivalent loading speed** (~330 samples/s per worker). Video re-encoding is slightly lossy; use the image-based variant if you cannot install a video decoding backend.
+- [HuggingFaceVLA/libero](https://huggingface.co/datasets/HuggingFaceVLA/libero)

 For reference, the original dataset published by Physical Intelligence:

 - [physical-intelligence/libero](https://huggingface.co/datasets/physical-intelligence/libero)

-<Tip>
-Pin `--dataset.revision=<commit-sha>` when reporting results — Hub datasets can be re-uploaded, and success rates are only comparable against the same data revision.
-</Tip>
-
 ### Example training command

-Train SmolVLA on the recommended dataset:
-
 ```bash
 lerobot-train \
  --policy.type=smolvla \
+  --policy.repo_id=${HF_USER}/libero-test \
  --policy.load_vlm_weights=true \
-  --policy.push_to_hub=false \
-  --dataset.repo_id=lerobot/libero \
-  --dataset.video_backend=torchcodec \
-  --output_dir=./outputs/libero_smolvla \
+  --dataset.repo_id=HuggingFaceVLA/libero \
+  --env.type=libero \
+  --env.task=libero_10 \
+  --output_dir=./outputs/ \
  --steps=100000 \
-  --batch_size=64
+  --batch_size=4 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=1 \
+  --env_eval_freq=1000
 ```

-To share the result on the Hub, replace `--policy.push_to_hub=false` with `--policy.repo_id=${HF_USER}/libero-smolvla`. Evaluate saved checkpoints with `lerobot-eval` as shown in the [Evaluation](#evaluation) section.
-
 ## Reproducing published results

 We reproduce the results of Pi0.5 on the LIBERO benchmark. We take the Physical Intelligence LIBERO base model (`pi05_libero`) and finetune for an additional 6k steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
@@ -134,20 +134,6 @@ LIBERO-plus supports two control modes — `relative` (default) and `absolute`.
 --env.control_mode=relative   # or "absolute"
 ```

-### Reset performance
-
-By default, LeRobot preserves LIBERO's hard-reset behavior. With fixed initial
-states enabled, you can opt into soft resets to skip rebuilding the simulator
-model and renderer on every episode:
-
-```bash
--env.init_states=true --env.hard_reset=false
-```
-
-Soft resets are faster but are not bit-identical to hard resets after the
-environment's settling steps, so camera observations and policy results may
-differ slightly. Use hard resets when reproducing benchmark results.
-
 ### Policy inputs and outputs

 **Observations:**
@@ -1,29 +1,28 @@
 # Multi-GPU Training

-LeRobot trains on multiple GPUs through [Hugging Face Accelerate](https://huggingface.co/docs/accelerate). Three data-parallel layouts are supported:
-
-| Layout   | What it does                                                  | Config                                                  |
-| -------- | ------------------------------------------------------------- | ------------------------------------------------------- |
-| **DDP**  | Replicates the full model on every GPU                        | default on any multi-GPU launch                         |
-| **FSDP** | Shards parameters, gradients, and optimizer state across GPUs | `--parallelism.dp_shard=N`                              |
-| **HSDP** | Shards within groups of GPUs, replicates across groups        | `--parallelism.dp_replicate=R --parallelism.dp_shard=S` |
+This guide shows you how to train policies on multiple GPUs using [Hugging Face Accelerate](https://huggingface.co/docs/accelerate).

 ## Installation

-`accelerate` is included in the `training` extra:
+`accelerate` is included in the `training` extra. Install it with:

 ```bash
 pip install 'lerobot[training]'
 ```

-## Launching
+## Training with Multiple GPUs

-Distributed training can be launched through both `torchrun` and `accelerate launch`. Accelerate is used as a plain launcher: it does not manage the training configuration, and every distributed training setting lives in LeRobot's own config system.
+You can launch training in two ways:

-With `torchrun`:
+### Option 1: Without config (specify parameters directly)
+
+You can specify all parameters directly in the command without running `accelerate config`:

 ```bash
-torchrun --nproc-per-node=2 $(which lerobot-train) \
+accelerate launch \
+  --multi_gpu \
+  --num_processes=2 \
+  $(which lerobot-train) \
  --dataset.repo_id=${HF_USER}/my_dataset \
  --policy.type=act \
  --policy.repo_id=${HF_USER}/my_trained_policy \
@@ -32,145 +31,150 @@ torchrun --nproc-per-node=2 $(which lerobot-train) \
  --wandb.enable=true
 ```

-With `accelerate launch` (as a plain launcher):
+**Key accelerate parameters:**
+
+- `--multi_gpu`: Enable multi-GPU training
+- `--num_processes=2`: Number of GPUs to use
+- `--mixed_precision=fp16`: Use fp16 mixed precision (or `bf16` if supported)
+
+### Option 2: Using accelerate config
+
+If you prefer to save your configuration, you can optionally configure accelerate for your hardware setup by running:

 ```bash
+accelerate config
+```
+
+This interactive setup will ask you questions about your training environment (number of GPUs, mixed precision settings, etc.) and saves the configuration for future use. For a simple multi-GPU setup on a single machine, you can use these recommended settings:
+
+- Compute environment: This machine
+- Number of machines: 1
+- Number of processes: (number of GPUs you want to use)
+- GPU ids to use: (leave empty to use all)
+- Mixed precision: fp16 or bf16 (recommended for faster training)
+
+Then launch training with:
+
+```bash
+accelerate launch $(which lerobot-train) \
+  --dataset.repo_id=${HF_USER}/my_dataset \
+  --policy.type=act \
+  --policy.repo_id=${HF_USER}/my_trained_policy \
+  --output_dir=outputs/train/act_multi_gpu \
+  --job_name=act_multi_gpu \
+  --wandb.enable=true
+```
+
+## How It Works
+
+When you launch training with accelerate:
+
+1. **Automatic detection**: LeRobot automatically detects if it's running under accelerate
+2. **Data distribution**: Your batch is automatically split across GPUs
+3. **Gradient synchronization**: Gradients are synchronized across GPUs during backpropagation
+4. **Single process logging**: Only the main process logs to wandb and saves checkpoints
+
+## Learning Rate and Training Steps Scaling
+
+**Important:** LeRobot does **NOT** automatically scale learning rates or training steps based on the number of GPUs. This gives you full control over your training hyperparameters.
+
+### Why No Automatic Scaling?
+
+Many distributed training frameworks automatically scale the learning rate by the number of GPUs (e.g., `lr = base_lr × num_gpus`).
+However, LeRobot keeps the learning rate exactly as you specify it.
+
+### When and How to Scale
+
+If you want to scale your hyperparameters when using multiple GPUs, you should do it manually:
+
+**Learning Rate Scaling:**
+
+```bash
+# Example: 2 GPUs with linear LR scaling
+# Base LR: 1e-4, with 2 GPUs -> 2e-4
 accelerate launch --num_processes=2 $(which lerobot-train) \
-  --dataset.repo_id=${HF_USER}/my_dataset \
-  --policy.type=act \
-  --policy.repo_id=${HF_USER}/my_trained_policy \
-  --output_dir=outputs/train/act_multi_gpu \
-  --job_name=act_multi_gpu \
-  --wandb.enable=true
+  --optimizer.lr=2e-4 \
+  --dataset.repo_id=lerobot/pusht \
+  --policy.type=act
 ```

-With no `--parallelism.*` flags, a multi-process launch runs plain DDP. Multi-node runs use the standard `torchrun --nnodes/--node-rank/--rdzv-endpoint` flags (or `accelerate launch --num_machines/--machine_rank/--main_process_ip`).
+**Training Steps Scaling:**

-> [!WARNING]
-> Accelerate's YAML config files (`accelerate launch --config_file some.yaml`, `accelerate config`) are not supported. They configure the engine through environment variables, bypassing LeRobot's configuration system, so `train_config.json` would no longer describe the settings a run actually used. `lerobot-train` therefore refuses to start when [accelerate environment variables](https://huggingface.co/docs/accelerate/usage_guides/fsdp) are set. Put the settings in `--parallelism.*` / `--accelerator.*` flags instead, or set `LEROBOT_ALLOW_ACCELERATE_ENV=1` to acknowledge the override and proceed anyway.
-
-## Batch semantics, learning rate, and steps
-
-Each of the `dp_replicate × dp_shard` data-parallel workers loads its own `--batch_size` micro-batch every step, so one training step consumes `batch_size × dp_world_size` samples, and `× gradient_accumulation_steps` of those go into each optimizer update:
-
-```
-effective_batch_size = batch_size × dp_world_size × gradient_accumulation_steps
-```
-
-The training banner prints this factorization at startup. `--steps` counts loop steps (micro-batches per worker), not optimizer updates.
-
-Gradient accumulation is a first-class flag:
+Since the effective batch size `bs` increases with multiple GPUs (batch_size × num_gpus), you may want to reduce the number of training steps proportionally:

 ```bash
-torchrun --nproc-per-node=2 $(which lerobot-train) \
-  --batch_size=8 --accelerator.gradient_accumulation.steps=4 ...
+# Example: 2 GPUs with effective batch size 2x larger
+# Original: batch_size=8, steps=100000
+# With 2 GPUs: batch_size=8 (16 in total), steps=50000
+accelerate launch --num_processes=2 $(which lerobot-train) \
+  --batch_size=8 \
+  --steps=50000 \
+  --dataset.repo_id=lerobot/pusht \
+  --policy.type=act
 ```

-**LeRobot does not auto-scale the learning rate or the number of steps** when the effective batch size grows. If you scale out and want equivalent training, please adjust manually, e.g. with 2 GPUs: double `--optimizer.lr` (linear scaling), or halve `--steps`.
+## Training Large Models with FSDP

-## Sharded training (FSDP)
+DDP replicates the full model on every GPU, so a model that doesn't fit on one GPU won't fit under
+DDP either. For large models, use **FSDP** (Fully Sharded Data Parallel), which shards parameters,
+gradients, and optimizer state across GPUs. See the [accelerate FSDP guide](https://huggingface.co/docs/accelerate/usage_guides/fsdp) for background.

-If a model is too large to train with DDP, shard it with FSDP2:
+An example on how to launch LeRobot training with FSDP across 4 GPUs (1 machine):

 ```bash
-torchrun --nproc-per-node=4 $(which lerobot-train) \
+accelerate launch --config_file fsdp.yaml --num_processes=4 $(which lerobot-train) \
  --dataset.repo_id=${HF_USER}/my_dataset \
  --policy.type=<your_policy> \
-  --parallelism.dp_shard=4 \
-  --accelerator.mixed_precision=bf16 \
  --output_dir=outputs/train/my_policy_fsdp
 ```

-`--parallelism.dp_shard=-1` shards over however many processes the launcher started.
+A minimal `fsdp.yaml` (FSDP1; shards params/grads/optimizer — ZeRO-3-equivalent):

-### Wrap units
-
-FSDP shards the model in units (typically the repeated transformer block) and gathers one unit at a time during forward/backward. Policies declare their wrap units via `_fsdp_wrap_modules` on the policy class. For example, ACT declares `["ACTEncoderLayer", "ACTDecoderLayer"]` and FastWAM declares `["MoTLayer"]`. For a policy without a `_fsdp_wrap_modules` declaration, pass one of the flags below. You can specify the module class name explicitly, or use a size-based policy instead:
-
-```bash
--accelerator.fsdp.wrap_modules='["MyTransformerBlock"]'   # explicit class names
--accelerator.fsdp.min_num_params=1000000                  # or: wrap every submodule above 1M params
+```yaml
+compute_environment: LOCAL_MACHINE
+distributed_type: FSDP
+mixed_precision: bf16
+num_machines: 1
+num_processes: 4
+fsdp_config:
+  fsdp_version: 1
+  fsdp_sharding_strategy: FULL_SHARD # params + grads + optimizer (ZeRO-3)
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: <YourTransformerBlock> # repeated block class to shard
+  fsdp_use_orig_params: true # required: optimizer is built pre-prepare
+  fsdp_state_dict_type: FULL_STATE_DICT
 ```

-If a policy doesn't declare `_fsdp_wrap_modules` and no `--accelerator.fsdp.wrap_modules` or `--accelerator.fsdp.min_num_params` is passed, the run fails at startup rather than silently wrapping only the root module (which would forfeit all sharding memory savings).
+Set `fsdp_transformer_layer_cls_to_wrap` to your model's repeated transformer-block class so each
+block is sharded as its own unit. `fsdp_use_orig_params: true` is required because LeRobot builds the
+optimizer before `accelerator.prepare()`.

-Other sharding settings:
+### FSDP checkpoints

- `--accelerator.fsdp.reshard_after_forward`: whether to keep each unit's parameters resident after forward.
- `--accelerator.fsdp.cpu_offload`: keeps parameters, gradients and optimizer states on CPU.
- `--accelerator.fsdp.ignored_modules`: a regex of module paths to keep unsharded.
+LeRobot gathers the full state dict across all ranks and the main process writes it as a single
+`model.safetensors`, loadable as usual with `Policy.from_pretrained(...)`. Two things to look out for:

-### HSDP
-
-Hybrid Sharded Data Parallel: parameters, gradients and optimizer states are sharded across `dp_shard` ranks, and that sharding is replicated `dp_replicate` times. Parameter all-gathers and gradient reduce-scatters stay inside a shard group; only the all-reduce that synchronizes the replicas crosses between groups. The two degrees must multiply to the world size:
-
-```bash
-# 16 GPUs = 2 nodes × 8: shard within each node, replicate across nodes
-torchrun --nnodes=2 --nproc-per-node=8 ... $(which lerobot-train) \
-  --parallelism.dp_replicate=2 --parallelism.dp_shard=8 ...
-```
-
-## Checkpoints
-
-Every checkpoint contains a `pretrained_model/` directory and a `training_state/` directory:
-
-```text
-005000/  # the training step at that checkpoint
-├── pretrained_model/
-│   ├── config.json  # policy config
-│   ├── train_config.json  # the full training config
-│   ├── model.safetensors  # full weights (checkpoint_format ∈ {safetensors, safetensors_dcp}, or any non-sharded run)
-│   ├── pytorch_model_fsdp_0/  # DCP weight shards (checkpoint_format ∈ {dcp, safetensors_dcp})
-│   ├── policy_preprocessor.json  # preprocessor config (when the run has a preprocessor)
-│   ├── policy_preprocessor_step_*.safetensors  # state of the stateful preprocessor steps
-│   ├── policy_postprocessor.json  # postprocessor config (when the run has a postprocessor)
-│   └── policy_postprocessor_step_*.safetensors  # state of the stateful postprocessor steps
-└── training_state/
-    ├── training_step.json  # step counter, topology, and batch semantics
-    ├── rng_state.safetensors  # rng states
-    ├── scheduler_state.json  # scheduler state (when the run has a scheduler)
-    ├── optimizer_state.safetensors  # full optimizer state (non-sharded runs)
-    ├── optimizer_param_groups.json  # optimizer param groups (non-sharded runs)
-    └── optimizer_0/  # DCP optimizer shards (sharded runs)
-```
-
-During single-GPU or DDP training, the pipeline serializes each state dict into a single file: `model.safetensors` for the model and `optimizer_state.safetensors` for the optimizer.
-
-During sharded training, the optimizer state is saved as DCP shards under `training_state/optimizer_0/`, and the layout of the model under `pretrained_model/` can be configured through `--checkpoint_format`:
-
-| `--checkpoint_format`     | Weights artifact                             | Use when                                                              |
-| ------------------------- | -------------------------------------------- | --------------------------------------------------------------------- |
-| `safetensors` _(default)_ | single `model.safetensors` only              | you want every checkpoint immediately loadable with `from_pretrained` |
-| `dcp`                     | `pytorch_model_fsdp_0/` shard directory only | gathering the full weights makes saves and resumes too slow           |
-| `safetensors_dcp`         | both                                         | you want fast resume _and_ immediately loadable checkpoints           |
-
-Two things to know about gathered (`safetensors`) checkpoints from sharded runs:
-
- **They store fp32 weights.** Under mixed precision training, FSDP keeps an fp32 master copy, and the checkpoint saves the master copy to make sure training resumes consistently.
- The gather is collective (all ranks participate) but only the main process writes.
-
-### Converting DCP checkpoints
-
-`lerobot-convert-dcp` merges a DCP shard directory into a regular `model.safetensors`, offline and without GPUs:
-
-```bash
-lerobot-convert-dcp --checkpoint_dir=outputs/train/run/checkpoints/005000
-lerobot-convert-dcp --checkpoint_dir=... --delete_dcp=true --push_to_hub=${HF_USER}/my_policy
-```
-
-`--push_to_hub` publishes the converted directory as a model repo.
-
-### Resuming
-
-Resume with `--resume=true --config_path=.../checkpoints/last/pretrained_model/train_config.json`. Resuming from a DCP checkpoint supports resharding the model and optimizer state to the _current_ topology, which means you can resume with a different `dp_replicate/dp_shard` split. The data sampler can always resume at the right epoch and offset, but is only _sample-exact_ when the world size and batch size match the original run (a warning is logged otherwise).
-
-> [!NOTE]
-> FSDP checkpoints written by LeRobot 0.6.x and earlier used a different on-disk layout (a gathered full optimizer state) and **cannot be resumed**.
+- **Checkpoints store fp32 weights.** Under mixed precision (`bf16`/`fp16`) FSDP keeps an fp32 master
+  copy, and the checkpoint saves it (~2× the bf16 size on disk) so training can resume consistently
+  with the fp32 optimizer state; `from_pretrained` casts back to the policy dtype on load. FSDP-specific
+  caveat: an fp32 checkpoint is materialized in full precision on the target device _before_ casting,
+  so loading it for inference on a tight GPU can OOM even when the bf16 model would fit — load on CPU
+  first, or cast `model.safetensors` to the deployment dtype offline.
+- The sharded optimizer state is gathered into a full (world-size-independent) state dict and saved
+  alongside the model in the same `optimizer_state.safetensors` / `optimizer_param_groups.json`
+  format as single-GPU training, so **resume-from-checkpoint is supported** with `--resume=true`.
+  Resume reshards both the model and the optimizer state to the _current_ FSDP topology, so you can
+  resume an FSDP checkpoint on a different number of GPUs. Note that the data sampler is only
+  sample-exact when the world size and batch size match the original run (a warning is logged
+  otherwise); the optimizer/model state itself is unaffected.

 ## Notes

- Checkpoint saves and end-of-training publishes are collective (every rank enters them). Gathered weights, sidecar files and Hub uploads are written by the main process alone.
- Metrics are reduced across ranks before logging: losses are averaged, and `samples/s` reports cluster-wide throughput.
- Learning-rate scheduling is stepped once per training step regardless of the number of processes (`step_scheduler_with_optimizer=False` is baked in).
+- The `--policy.use_amp` flag in `lerobot-train` is only used when **not** running with accelerate. When using accelerate, mixed precision is controlled by accelerate's configuration.
+- Training logs, checkpoints, and hub uploads are only done by the main process to avoid conflicts. Non-main processes have console logging disabled to prevent duplicate output.
+- The effective batch size is `batch_size × num_gpus`. If you use 4 GPUs with `--batch_size=8`, your effective batch size is 32.
+- Learning rate scheduling is handled correctly across multiple processes—LeRobot sets `step_scheduler_with_optimizer=False` to prevent accelerate from adjusting scheduler steps based on the number of processes.
+- When saving or pushing models, LeRobot automatically unwraps the model from accelerate's distributed wrapper to ensure compatibility.
+- WandB integration automatically initializes only on the main process, preventing multiple runs from being created.

-For background on the underlying machinery, see the [Accelerate FSDP guide](https://huggingface.co/docs/accelerate/usage_guides/fsdp). To go deeper on large-scale training, check out the [Ultrascale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook).
+For more advanced configurations and troubleshooting, see the [Accelerate documentation](https://huggingface.co/docs/accelerate). If you want to learn more about how to train on a large number of GPUs, checkout this awesome guide: [Ultrascale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook).
@@ -36,12 +36,6 @@ This diverse training mixture creates a "curriculum" that enables generalization
   pip install -e ".[pi]"
   ```

-   If you installed LeRobot from PyPI:
-
-   ```bash
-   pip install 'lerobot[pi]'
-   ```
-
 ## Usage

 To use π₀.₅ in your LeRobot configuration, specify the policy type as:
@@ -52,106 +46,27 @@ policy.type=pi05

 ## Training

-### Quickstart on LIBERO
-
-Finetune the LIBERO base model on [lerobot/libero](https://huggingface.co/datasets/lerobot/libero), a ~1.9 GB video-encoded copy of the demonstrations behind the [results below](#libero-benchmark-results).
-
-It carries the keys π₀.₅ reads, which are also the ones the LIBERO environment produces at evaluation time:
-
-| Feature                     | Shape in the dataset | How π₀.₅ consumes it                                    |
-| --------------------------- | -------------------- | ------------------------------------------------------- |
-| `observation.images.image`  | 256×256×3, agentview | resized to 224×224                                      |
-| `observation.images.image2` | 256×256×3, wrist     | resized to 224×224                                      |
-| `observation.state`         | 8                    | discretized into 256 bins and written into the prompt   |
-| `action`                    | 7                    | padded to 32 internally; the loss uses the first 7 dims |
-
-**No `--rename_map` is needed here** — the keys already match; see [Rename Map and Empty Cameras](./rename_map) if yours differ.
-
-<Tip>
-  π₀.₅ uses the gated
-  [google/paligemma-3b-pt-224](https://huggingface.co/google/paligemma-3b-pt-224)
-  tokenizer — accept its license on the Hub and log in with `hf auth login`
-  before training.
-</Tip>
-
-Sized for a single 80 GB GPU:
-
-```bash
-lerobot-train \
-    --dataset.repo_id=lerobot/libero \
-    --policy.type=pi05 \
-    --policy.pretrained_path=lerobot/pi05_libero_base \
-    --policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}' \
-    --policy.n_action_steps=10 \
-    --policy.empty_cameras=1 \
-    --policy.freeze_vision_encoder=false \
-    --policy.train_expert_only=false \
-    --policy.gradient_checkpointing=true \
-    --policy.dtype=bfloat16 \
-    --policy.device=cuda \
-    --policy.push_to_hub=false \
-    --output_dir=./outputs/pi05_libero \
-    --job_name=pi05_libero \
-    --batch_size=64 \
-    --num_workers=8 \
-    --steps=30000 \
-    --save_freq=5000 \
-    --seed=1000
-```
-
-**Mean/std normalization, not π₀.₅'s [quantile default](#quantile-statistics)** — matching [pi05_libero_finetuned_v044](https://huggingface.co/lerobot/pi05_libero_finetuned_v044), the checkpoint the results below were measured on.
-
-**`--policy.n_action_steps=10` and `--policy.empty_cameras=1` are explicit** because `--policy.pretrained_path` loads weights only — `lerobot/pi05_libero_base` stores both, and they would otherwise fall back to `50` and `0` (see [Loading a checkpoint](#loading-a-checkpoint)).
-
-Then evaluate a checkpoint with `lerobot-eval` and compare against the reference success rates — see [LIBERO](./libero).
-
-### Quantile statistics
-
-π₀.₅ normalizes `STATE` and `ACTION` with quantiles, so your dataset's `meta/stats.json` needs `q01` and `q99`. Older datasets carry only `min`/`max`/`mean`/`std` and fail on the first batch:
-
-```
-ValueError: QUANTILES normalization mode requires q01 and q99 stats
-```
-
-Recompute them:
-
-```bash
-lerobot-edit-dataset \
-    --repo_id your_dataset \
-    --new_repo_id your_dataset \
-    --operation.type recompute_stats \
-    --operation.overwrite true
-```
-
-**The result lands in `$HF_LEROBOT_HOME/your_dataset`**, not the cache `--dataset.repo_id` reads — so train with `--dataset.root=$HF_LEROBOT_HOME/your_dataset`, or add `--push_to_hub true` above.
-
-Or keep the dataset as-is and pass `--policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}'`.
-
 ### Training Command Example

-The same finetune with the VLM frozen: less memory, at some cost in success rate. Swap `--dataset.repo_id` for your own dataset.
+Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:

 ```bash
 lerobot-train \
-    --dataset.repo_id=lerobot/libero \
+    --dataset.repo_id=your_dataset \
    --policy.type=pi05 \
-    --policy.pretrained_path=lerobot/pi05_libero_base \
-    --policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}' \
-    --policy.n_action_steps=10 \
-    --policy.empty_cameras=1 \
-    --policy.freeze_vision_encoder=true \
-    --policy.train_expert_only=true \
+    --output_dir=./outputs/pi05_training \
+    --job_name=pi05_training \
+    --policy.repo_id=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 \
-    --policy.push_to_hub=false \
-    --output_dir=./outputs/pi05_libero_expert \
-    --job_name=pi05_libero_expert \
-    --batch_size=64 \
-    --num_workers=8 \
-    --steps=30000 \
-    --save_freq=5000 \
-    --seed=1000
+    --batch_size=32
 ```

 ### Key Training Parameters
@@ -159,24 +74,10 @@ lerobot-train \
 - **`--policy.compile_model=true`**: Enables model compilation for faster training
 - **`--policy.gradient_checkpointing=true`**: Reduces memory usage significantly during training
 - **`--policy.dtype=bfloat16`**: Use mixed precision training for efficiency
- **`--batch_size=64`**: Batch size for training, adapt this based on your GPU memory
+- **`--batch_size=32`**: Batch size for training, adapt this based on your GPU memory
 - **`--policy.pretrained_path=lerobot/pi05_base`**: The base π₀.₅ model you want to finetune, options are:
  - [lerobot/pi05_base](https://huggingface.co/lerobot/pi05_base)
-  - [lerobot/pi05_libero_base](https://huggingface.co/lerobot/pi05_libero_base) (specifically trained on the Libero dataset)
-
-### Loading a checkpoint
-
-The two forms are not interchangeable:
-
-|                                        | `--policy.path`                                | `--policy.pretrained_path`           |
-| -------------------------------------- | ---------------------------------------------- | ------------------------------------ |
-| Loads                                  | weights **and** the checkpoint's `config.json` | weights only                         |
-| Feature names                          | from the checkpoint                            | from your dataset                    |
-| Stored settings, e.g. `n_action_steps` | inherited                                      | reset to the defaults                |
-| `--policy.type`                        | must be omitted                                | required                             |
-| `--rename_map`                         | needed when your camera keys differ            | never — the keys come from your data |
-
-Passing a `--rename_map` alongside `--policy.pretrained_path` renames the batch away from those names, and the first batch fails with `All image features are missing from the batch`.
+  - [lerobot/pi05_libero](https://huggingface.co/lerobot/pi05_libero) (specifically trained on the Libero dataset)

 ### Training Parameters Explained

@@ -187,6 +88,15 @@ Passing a `--rename_map` alongside `--policy.pretrained_path` renames the batch

 **💡 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
+python src/lerobot/scripts/augment_dataset_quantile_stats.py \
+    --repo-id=your_dataset \
+```
+
+Or train pi05 with this normalization mapping: `--policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}'`
+
 ## Relative Actions

 By default, π₀.₅ predicts absolute actions. You can enable **relative actions** so the model predicts offsets relative to the current robot state. This can improve training stability for certain setups.
@@ -174,7 +174,7 @@ The model takes images, text instructions, and robot state as input, and outputs

 ## 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 40k steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
+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:

@@ -22,7 +22,7 @@ With processors, you choose the learning features you want to use for your polic
 ## Three pipelines

 We often compose three pipelines. Depending on your setup, some can be empty if action and observation spaces already match.
-Each of these pipelines handles different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.
+Each of these pipelines handle different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.

 1. Pipeline 1: Teleop action space → dataset action space (phone pose → EE targets)
 2. Pipeline 2: Dataset action space → robot command space (EE targets → joints)
@@ -74,15 +74,15 @@ In the phone to SO-100 follower examples we use the following adapters:
 - `robot_action_to_transition`: transforms the teleop action dict to a pipeline transition.
 - `transition_to_robot_action`: transforms the pipeline transition to a robot action dict.
 - `observation_to_transition`: transforms the robot observation dict to a pipeline transition.
- `transition_to_observation`: transforms the pipeline transition to an observation dict.
+- `transition_to_observation`: transforms the pipeline transition to a observation dict.

-Check out [src/lerobot/processor/converters.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/converters.py) for more details.
+Checkout [src/lerobot/processor/converters.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/converters.py) for more details.

 ## Dataset feature contracts

 Dataset features are determined by the keys saved in the dataset. Each step can declare what features it modifies in a contract called `transform_features(...)`. Once you build a processor, the processor can then aggregate all of these features with `aggregate_pipeline_dataset_features()` and merge multiple feature dicts with `combine_feature_dicts(...)`.

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

 ```python
    def transform_features(
@@ -82,8 +82,6 @@ By default the env samples objects only from the `lightwheel` registry (what `--

 All eval snippets below mirror the CI command (see `.github/workflows/benchmark_tests.yml`). The `--rename_map` argument maps RoboCasa's native camera keys (`robot0_agentview_left` / `robot0_eye_in_hand` / `robot0_agentview_right`) onto the three-camera (`camera1` / `camera2` / `camera3`) input layout the released `smolvla_robocasa` policy was trained on.

-By default, each task uses the rollout horizon registered by RoboCasa. Set `--env.episode_length=<steps>` to apply the same explicit horizon to every selected task.
-
 ### Single-task evaluation (recommended for quick iteration)

 ```bash
@@ -35,11 +35,14 @@ pip install --override <(printf 'gymnasium==0.29.1\nnumpy==1.26.4\n') \
 ### Docker (recommended)

 ```bash
-# Build the RoboMME evaluation image from the repo root
-docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-benchmark-robomme .
+# Build base image first (from repo root)
+docker build -f docker/Dockerfile.eval-base -t lerobot-eval-base .
+
+# Build RoboMME eval image (applies gymnasium + numpy pin overrides)
+docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-robomme .
 ```

-The benchmark Dockerfile extends the published `huggingface/lerobot-gpu:latest` image, then overrides `gymnasium==0.29.1` and `numpy==1.26.4`. Both versions are runtime-safe for lerobot's actual API usage.
+The `docker/Dockerfile.benchmark.robomme` image overrides `gymnasium==0.29.1` and `numpy==1.26.4` after lerobot's install. Both versions are runtime-safe for lerobot's actual API usage.

 ## Running Evaluation

@@ -57,7 +57,7 @@ policy_cfg.rtc_config = RTCConfig(
 policy = PI0Policy.from_pretrained("lerobot/pi0_base", policy_cfg=policy_cfg, device="cuda")

 # Now use predict_action_chunk with RTC parameters
-inference_delay = 4  # How many steps of inference latency, this value should be calculated based on the inference latency of the policy
+inference_delay = 4  # How many steps of inference latency, this values should be calculated based on the inference latency of the policy

 # Initialize the action queue
 action_queue = ActionQueue(policy_cfg.rtc_config)
@@ -100,7 +100,7 @@ Typical values: 8-12 steps
 RTCConfig(execution_horizon=10)
 ```

-**`max_guidance_weight`**: How strongly to enforce consistency with the previous chunk. This is a hyperparameter that can be tuned to balance the smoothness of the transitions and the reactivity of the policy. For 10 steps flow matching (SmolVLA, Pi0, Pi0.5), a value of 10.0 is an optimal value.
+**`max_guidance_weight`**: How strongly to enforce consistency with the previous chunk. This is a hyperparameter that can be tuned to balance the smoothness of the transitions and the reactivity of the policy. For 10 steps flow matching (SmolVLA, Pi0, Pi0.5), a value of 10.0 is a optimal value.

 **`prefix_attention_schedule`**: How to weight consistency across the overlap region.

@@ -93,7 +93,7 @@ lerobot-train --help

 ## Evaluate the finetuned model and run it in real-time

-Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./il_robots#record-a-dataset).
+Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./il_robots).
 Once you are logged in, you can run inference in your setup by doing:

 ```bash
@@ -338,7 +338,7 @@ It is advisable to install one 3-pin cable in the motor after placing them befor
 <hfoption id="Leader">

 - Mount the leader holder onto the wrist and secure it with 4 M3x6mm screws.
- Attach the handle to the leader holder using 1 M2x6mm screw.
+- Attach the handle to motor 5 using 1 M2x6mm screw.
 - Insert the gripper motor, secure it with 2 M2x6mm screws on each side, attach a motor horn using a M3x6mm horn screw.
 - Attach the follower trigger with 4 M3x6mm screws.

@@ -1,339 +0,0 @@
-# Third-Party Robots & Teleoperators
-
-The LeRobot ecosystem extends far beyond its officially supported hardware. Thanks to LeRobot's plugin architecture, the community has built integrations for a wide range of robot arms and teleoperation devices — from industrial manipulators to affordable hobbyist platforms, VR headsets, haptic devices, and full arm-plus-teleoperator kits. This page showcases community-maintained integrations you can use for teleoperation, data collection, and policy deployment.
-
-> [!IMPORTANT]
-> These projects are developed and maintained by third parties. Please refer to each repository for installation instructions, hardware requirements, and support.
-
-Drop-in plugins are auto-discovered by package name: LeRobot imports any installed package prefixed with `lerobot_robot_` or `lerobot_teleoperator_`. Once installed, reference the `type` the plugin registers (see its README — it may differ from the package name) directly from any LeRobot command:
-
-```bash
-pip install lerobot_robot_<name> lerobot_teleoperator_<name>
-
-lerobot-record \
-    --robot.type=<robot_name> \
-    --teleop.type=<teleoperator_name> \
-    --dataset.repo_id=${HF_USER}/my-dataset \
-    --dataset.num_episodes=5
-```
-
-> [!TIP]
-> ⚠️ marks projects that are forks/extensions of LeRobot. They may require custom setup rather than working with an unmodified install. All other entries are drop-in plugins.
-
-## Industrial & Collaborative Arms
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/SpesRobotics/lerobot-robot-xarm">lerobot-robot-xarm</a></td>
-      <td>Plugin for the xArm collaborative arm series from <a href="https://www.ufactory.cc/">UFACTORY</a>.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/lebai-robotics/lerobot_lebai">lerobot_lebai</a></td>
-      <td>Plugin for the six-axis collaborative arms from <a href="https://lebai.ltd/en/">Lebai</a>.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/wengmister/LeFranX">LeFranX</a> ⚠️</td>
-      <td>LeRobot extension for the <a href="https://franka.de/">Franka</a> research arm, paired with the <a href="https://www.robotera.com/">RobotEra XHand</a> hand for VR teleoperation.</td>
-    </tr>
-  </tbody>
-</table>
-
-#### Universal Robots UR5e
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/yechen056/UR5e-LeRobot">UR5e-LeRobot</a> ⚠️</td>
-      <td>LeRobot extension for the <a href="https://www.universal-robots.com/">Universal Robots UR5e</a>, with single-arm and bimanual support.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/scy-v/lerobot_ur5e_auto">lerobot_ur5e_auto</a> ⚠️</td>
-      <td>LeRobot extension for a mobile <a href="https://www.universal-robots.com/">Universal Robots UR5e</a>, adding automated recording at scale with minimal supervision.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/F-Fer/lerobot_ur5e_gello">lerobot_robot_ur5e</a></td>
-      <td>Plugin for the <a href="https://www.universal-robots.com/">Universal Robots UR5e</a> with a <a href="https://robotiq.com/">Robotiq</a> gripper, over RTDE control.</td>
-    </tr>
-  </tbody>
-</table>
-
-## Research & Learning Arms
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/TrossenRobotics/lerobot_trossen">lerobot_trossen</a></td>
-      <td>Plugin for the WidowX and ALOHA-style arms from <a href="https://www.trossenrobotics.com/">Trossen Robotics</a>.</td>
-    </tr>
-  </tbody>
-</table>
-
-#### AgileX Piper
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/AgRoboticsResearch/lerobot_robot_piper">lerobot_robot_piper (AgRobotics Research)</a></td>
-      <td>Plugin for the <a href="https://global.agilex.ai/">AgileX Piper</a> arm.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/WeGo-Robotics/lerobot_robot_piper">lerobot_robot_piper (WeGo Robotics)</a></td>
-      <td>Plugin for the <a href="https://global.agilex.ai/">AgileX Piper</a> arm, with multi-arm teleoperation, safety limits, and GUI tools.</td>
-    </tr>
-  </tbody>
-</table>
-
-## Affordable & Hobbyist Arms
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/servodevelop/fashionstar-lerobot-robot-cello">fashionstar-lerobot-robot-cello</a></td>
-      <td>Plugin for the StarAI Cello 6+1 degrees of freedom robot arm from <a href="https://fashionstar.com.hk/">FashionStar</a>.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/servodevelop/fashionstar-lerobot-robot-viola">fashionstar-lerobot-robot-viola</a></td>
-      <td>Plugin for the compact StarAI Viola 6+1 degrees of freedom robot arm from <a href="https://fashionstar.com.hk/">FashionStar</a>.</td>
-    </tr>
-  </tbody>
-</table>
-
-## Service, Mobile & Utility Robots
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/ugo-plus/lerobot-robot-ugo-pro">lerobot-robot-ugo-pro</a></td>
-      <td>Plugin for the ugo Pro dual-arm service robot from <a href="https://ugo.plus/products/ugo-pro/">ugo</a>.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/zuoxingdong/lerobot_robot_lekiwi_pincopen">lerobot_robot_lekiwi_pincopen</a></td>
-      <td>Plugin for a LeKiwi mobile manipulator with a <a href="https://github.com/pollen-robotics/PincOpen">PincOpen</a> gripper.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/KillingJacky/lerobot-robot-dummy">lerobot-robot-dummy</a></td>
-      <td>Plugin simulating a robot for recording without hardware. Useful for debugging !</td>
-    </tr>
-  </tbody>
-</table>
-
-## Teleoperators
-
-### VR & Motion Controllers
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/SpesRobotics/lerobot-teleoperator-teleop">lerobot-teleoperator-teleop</a></td>
-      <td>Plugin turning a phone or VR headset into a teleoperator via <a href="https://immersiveweb.dev">WebXR</a>, wrapping the open-source <a href="https://github.com/SpesRobotics/teleop"><code>teleop</code></a> library.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/Jas000n/lerobot-teleoperator-spacemouse">lerobot-teleoperator-spacemouse</a></td>
-      <td>Plugin for the <a href="https://3dconnexion.com/">3Dconnexion SpaceMouse</a>, with inverse kinematics for SO-ARMS robots.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/Dream-Machines-Robotics/vr-teleop-kit">vr-teleop-kit</a></td>
-      <td>Plugin teleoperating arms from a <a href="https://www.meta.com/quest/">Meta Quest</a> (WebXR), relying on URDF descriptions for inverse kinematics.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/xensedyl/lerobot-teleoperator-pico4">lerobot-teleoperator-pico4</a></td>
-      <td>Plugin for the <a href="https://www.picoxr.com/">PICO 4</a> VR headset, with a companion controller-free <a href="https://github.com/xensedyl/lerobot-teleoperator-pico4-hand">hand-tracking variant</a>.</td>
-    </tr>
-  </tbody>
-</table>
-
-### Leader Arms
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/F-Fer/lerobot_ur5e_gello">lerobot_teleoperator_gello</a></td>
-      <td>Plugin for the 7 degrees of freedom <a href="https://wuphilipp.github.io/gello_site/">GELLO</a> teleoperator.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/uynitsuj/lerobot_teleoperator_yamactiveleader">lerobot_teleoperator_yamactiveleader</a></td>
-      <td>Plugin for the active YAM teleoperator from <a href="https://i2rt.com/">I2RT</a>, a bilateral force-feedback arm.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/charlie8612/lerobot_teleoperator_omy">lerobot_teleoperator_omy</a></td>
-      <td>Plugin for the OMY-L100 6 degrees of freedom teleoperator from <a href="https://www.robotis.com/">ROBOTIS</a>.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://pypi.org/project/lerobot-teleoperator-pipermate/">lerobot-teleoperator-pipermate</a></td>
-      <td>Plugin for the PiperMate teleoperator (<a href="https://fashionstar.com.hk/">FashionStar</a> UART servos), driving the <a href="https://global.agilex.ai/">AgileX Piper</a> arm.</td>
-    </tr>
-  </tbody>
-</table>
-
-### Haptic Devices
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/chohh7391/lerobot_teleoperator_inverse3">lerobot_teleoperator_inverse3</a></td>
-      <td>Plugin for the <a href="https://www.haply.co/">Haply Inverse3</a> haptic device, adding force-feedback teleoperation.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/hzhz112/lerobot_teleoperator_omega7">lerobot_teleoperator_omega7</a></td>
-      <td>Plugin for the <a href="https://www.forcedimension.com/">Force Dimension omega.7</a> haptic device, adding force-feedback teleoperation.</td>
-    </tr>
-  </tbody>
-</table>
-
-### Networked & Remote
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://pypi.org/project/lerobot-teleoperator-livekit/">lerobot-teleoperator-livekit</a></td>
-      <td>Plugin receiving teleoperation commands over a <a href="https://livekit.io/">LiveKit</a> Portal (WebRTC) for remote control.</td>
-    </tr>
-  </tbody>
-</table>
-
-## Full Kits (Robot + Teleoperator)
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/villekuosmanen/lerobot-arx5">lerobot-arx5</a></td>
-      <td>Plugin for the <a href="https://www.arx-x.com/">ARX5</a> arm: <a href="https://pypi.org/project/lerobot-robot-arx5/"><code>lerobot-arx5</code></a> robot arm with its <a href="https://pypi.org/project/lerobot-teleoperator-arx5/"><code>lerobot-teleoperator-arx5</code></a> teleoperator arm.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/robertorobotics/Nextis-AIRA-3D">Nextis-AIRA-3D</a></td>
-      <td>Plugin for the 7 degrees of freedom arm from <a href="https://www.nextis.tech">Nextis</a>: robot arm <code>aira_follower</code> and teleoperator arm <code>aira_leader</code>.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/pravsels/lerobot_yam">lerobot_yam</a></td>
-      <td>Plugin suite for the YAM arm from <a href="https://i2rt.com/">I2RT</a>: robot arm <code>yam_follower</code> and teleoperator arm <code>yam_leader</code>.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/robot-learning-co/trlc-dk1">trlc-dk1</a></td>
-      <td>Plugin for the development kit from <a href="https://www.robot-learning.co/">The Robot Learning Company</a>: single and bimanual arms follower/teleoperator types.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/hexfellow/hex_lerobot_drivers">hex_lerobot_drivers</a></td>
-      <td>Plugin suite for <a href="https://hexfellow.com/">HEXFELLOW</a> devices: robots, teleoperators, and cameras (see <a href="./third_party_sensors">Cameras &amp; Sensors</a>).</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/Hiwonder-official/lerobot-robot-nexarm-follower">lerobot-robot-nexarm-follower</a></td>
-      <td>Plugin for the NexArm from <a href="https://www.hiwonder.com/">Hiwonder</a>: the <a href="https://github.com/Hiwonder-official/lerobot-robot-nexarm-follower">robot arm</a> and its matching <a href="https://github.com/Hiwonder-official/lerobot-teleoperator-nexarm-leader">teleoperator arm</a>.</td>
-    </tr>
-  </tbody>
-</table>
-
-## ROS 2 Bridges
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/ngres/leros2">leros2</a></td>
-      <td>Plugin bridging ROS 2 topics and actions to LeRobot robots and teleoperators.</td>
-    </tr>
-    <tr style="border:0">
-      <td><a href="https://github.com/ROBOTIS-GIT/lerobot_robot_ros2_zenoh">lerobot_robot_ros2_zenoh</a></td>
-      <td>Plugin bridging ROS 2 robots to LeRobot over <a href="https://zenoh.io">Zenoh</a> pub/sub transport.</td>
-    </tr>
-  </tbody>
-</table>
-
-## Contributing
-
-Built your own LeRobot hardware integration? The plugin system makes it straightforward to add new robots and teleoperators — check out the [Bring Your Own Hardware](./integrate_hardware) guide to get started, and share your project with the community!
@@ -1,99 +0,0 @@
-# Third-Party Cameras & Sensors
-
-The LeRobot ecosystem extends far beyond its natively supported cameras (OpenCV, Intel RealSense, ZMQ, Reachy 2). Thanks to LeRobot's plugin architecture, the community has built drop-in camera and sensor integrations — from depth cameras to vision-based tactile sensors. This page showcases community-maintained camera and sensor integrations you can use for teleoperation, data collection, and policy deployment.
-
-> [!IMPORTANT]
-> These projects are developed and maintained by third parties. Please refer to each repository for installation instructions, hardware requirements, and support.
-
-Drop-in plugins are auto-discovered by package name: LeRobot imports any installed package prefixed with `lerobot_camera_`. Once installed, reference the camera `type` the plugin registers (see its README — it may differ from the package name) directly from any LeRobot command:
-
-```bash
-pip install lerobot_camera_<name>
-
-lerobot-record \
-    --robot.type=so101_follower \
-    --robot.port=/dev/ttyACM0 \
-    --robot.cameras="{ front: {type: <name>, width: 640, height: 480, fps: 30} }" \
-    --dataset.repo_id=${HF_USER}/my-dataset \
-    --dataset.num_episodes=5
-```
-
-## Tactile Sensors
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/xensedyl/lerobot-camera-xense">lerobot-camera-xense</a></td>
-      <td>Plugin for <a href="https://www.xenserobotics.com/">Xense</a> vision-based tactile sensors, exposing rectified/difference images, depth, and 2D markers.</td>
-    </tr>
-  </tbody>
-</table>
-
-## Depth Cameras
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/hexfellow/hex_lerobot_drivers/tree/main/lerobot_camera_berxel">lerobot_camera_berxel</a></td>
-      <td>Plugin for the <a href="https://www.berxel.com/">Berxel</a> depth camera, part of the broader <a href="https://hexfellow.com/">HEXFELLOW</a> driver suite.</td>
-    </tr>
-  </tbody>
-</table>
-
-## Networked Cameras
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/F-Fer/lerobot_ur5e_gello">lerobot_camera_zmq</a></td>
-      <td>Plugin streaming <a href="https://www.stereolabs.com/">Stereolabs ZED</a> and USB camera frames from a Raspberry Pi over the network.</td>
-    </tr>
-  </tbody>
-</table>
-
-## Virtual Cameras
-
-<!-- prettier-ignore -->
-<table width="100%" style="display:table; width:100%; table-layout:fixed;">
-  <colgroup>
-    <col width="30%" />
-    <col width="70%" />
-  </colgroup>
-  <thead style="border:0">
-    <tr style="border:0"><th>Project</th><th>Description</th></tr>
-  </thead>
-  <tbody>
-    <tr style="border:0">
-      <td><a href="https://github.com/hexfellow/hex_lerobot_drivers/tree/main/lerobot_camera_dummy">lerobot_camera_dummy</a></td>
-      <td>Plugin simulating a camera for recording without hardware. Useful for debugging !</td>
-    </tr>
-  </tbody>
-</table>
-
-## Contributing
-
-Built your own LeRobot camera or sensor integration? Package it as an installable `lerobot_camera_<name>` plugin and it will be auto-discovered by the LeRobot CLI — see the [Bring Your Own Hardware](./integrate_hardware) guide and the [Cameras](./cameras) reference to get started, then share your project with the community!
@@ -40,15 +40,3 @@ lerobot-eval \
 ```

 However, in most cases, presence of an accelerator is detected automatically and `policy.device` parameter can be omitted from CLI commands.
-
-## Mixed precision
-
-Training precision is owned by `--accelerator.mixed_precision`, which accepts `no` (default) and `bf16`:
-
-```bash
-lerobot-train \
-    --policy.type=act \
-    --accelerator.mixed_precision=bf16 ...
-```
-
-`bf16` requires an accelerator that supports it.
@@ -0,0 +1,121 @@
+# Training Dataset Streaming
+
+Training-time dataset streaming lets `lerobot-train` consume a LeRobotDataset v3 without first
+downloading its complete Parquet and video payload. Enable it through the existing public switch:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=OWNER/DATASET \
+  --dataset.streaming=true \
+  --policy.type=act \
+  --output_dir=outputs/train/act_streaming
+```
+
+This feature is independent of `--dataset.streaming_encoding=true`. `streaming_encoding` controls
+how videos are written while recording; `dataset.streaming` controls how an existing dataset is read
+during training. Recording and rollout encoders are not used by this training path.
+
+## How an epoch is read
+
+Each distributed rank owns a deterministic, frame-balanced set of complete episodes. One logical
+exact-coverage pool per rank mixes those episodes while visiting every selected frame once per
+rank-local coverage epoch. Parquet columns and compressed MP4 byte ranges are prefetched from the
+same byte-aware admission frontier. Temporal history and future windows are resolved inside the
+complete episode, including the same boundary padding masks as map-style loading.
+
+When `--num_workers` is nonzero, training uses one dedicated DataLoader process per rank. Its
+bounded result queue holds decoded batches while the policy trains. The configured worker count is
+instead used as internal Parquet and byte-range fetch concurrency, so increasing it does not create
+independent samplers or multiply the cache budget.
+
+The default map-style `LeRobotDataset` behavior is unchanged when `--dataset.streaming=false`.
+
+## MP4 sidecars and the first run
+
+Video streaming uses a small MP4 index sidecar. Dataset initialization first checks the
+revision-keyed local cache, then looks for a valid published sidecar. If neither is available,
+LeRobot builds the sidecar locally while holding a process lock and installs it atomically. A
+failed or interrupted build does not replace the previous valid file.
+
+Training is read-only: it never uploads a sidecar or modifies the dataset repository. On a cluster
+with node-local caches, the first job may build once per node. A shared LeRobot cache avoids that
+duplication.
+
+Dataset maintainers can build a sidecar ahead of time:
+
+```bash
+uv run python scripts/build_mp4_sidecar.py \
+  --repo-id=OWNER/DATASET \
+  --revision=COMMIT_SHA \
+  --data-root=hf://datasets/OWNER/DATASET@COMMIT_SHA \
+  --output=/tmp/dataset-mp4-sidecar.npz
+```
+
+Publication is always explicit. Add `--push` only after validating the complete-dataset sidecar.
+Subset sidecars cannot be published.
+
+## Configuration
+
+The production defaults are:
+
+| Option                              |        Default | Meaning                                             |
+| ----------------------------------- | -------------: | --------------------------------------------------- |
+| `streaming_episode_pool_size`       |             32 | Maximum complete episodes mixed by each rank        |
+| `streaming_prefetch_episodes`       |              8 | Episodes fetched ahead of the active pool           |
+| `streaming_byte_budget_gb`          |              8 | Maximum synthesized MP4 bytes per rank              |
+| `streaming_decode_threads`          |              2 | Parallel sample assembly and video decode workers   |
+| `streaming_decoded_queue_size`      |              8 | Decoded samples buffered ahead, in planner order    |
+| `streaming_max_open_decoders`       | pool × cameras | Independent open-decoder LRU cap per rank           |
+| `streaming_native_http_connections` |          unset | Native HTTP connection cap per rank                 |
+| `streaming_native_http_subranges`   |              1 | Concurrent subranges per native HTTP range read     |
+| `streaming_data_root`               |          unset | Optional local, Hub, bucket, or fsspec payload root |
+
+The active episode set is capped by both episode count and the exact synthesized mini-MP4 sizes
+computed from the sidecar. An episode larger than the complete rank budget fails before training
+fetches its payload. The cache, decoder LRU, and decoded-batch queue remain independently bounded.
+Start with a smaller pool or budget on memory-constrained hosts:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=OWNER/DATASET \
+  --dataset.streaming=true \
+  --dataset.streaming_episode_pool_size=16 \
+  --dataset.streaming_prefetch_episodes=4 \
+  --dataset.streaming_byte_budget_gb=4 \
+  --dataset.streaming_decode_threads=2 \
+  --dataset.streaming_decoded_queue_size=8 \
+  --num_workers=4 \
+  --policy.type=act \
+  --output_dir=outputs/train/act_streaming
+```
+
+If metadata remains in a dataset repository while payload files are mirrored elsewhere, set
+`--dataset.streaming_data_root`. Supported values include local paths, revision-qualified
+`hf://datasets/...` roots, `hf://buckets/...` roots, and fsspec URLs.
+
+## Resume and shuffle migration
+
+The earlier streaming reader used a bounded row shuffle buffer. The episode reader instead has
+deterministic exact-coverage ordering derived from the seed and epoch. Checkpoint resume restores the
+per-rank sample offset using the checkpoint batch size. Changing distributed world size or
+batch size changes ownership or batch boundaries. For sample-exact comparisons, resume with the
+same world size and batch size. Keep the same internal fetch concurrency when comparing performance.
+
+Streaming checkpoints created by the earlier multi-worker sampler record a different ownership
+topology and are rejected for sample-exact resume. Start a new run from the saved policy weights
+rather than claiming that its dataset stream resumes exactly.
+
+## Troubleshooting
+
+- **The first batch takes a long time:** check logs for a sidecar build. Reuse a shared cache or
+  publish a validated complete sidecar explicitly.
+- **A sidecar lock times out:** another process may still be indexing the same revision. Confirm it
+  is healthy before removing a stale lock.
+- **A rank owns no data:** reduce the number of ranks so every rank owns at least one selected
+  episode.
+- **The byte budget is exceeded:** lower the episode pool, increase the per-rank byte budget, or
+  use an explicitly prepared payload layout with smaller episode ranges.
+- **Remote reads cannot authenticate:** verify the normal Hugging Face token or fsspec credentials
+  are available in every worker environment. Credentials are never embedded in the sidecar.
+- **Refill stalls are high:** compare p95/p99 batch wait, reduce network contention, raise prefetch
+  gradually, and verify that the dataset sidecar matches the exact revision.
@@ -11,10 +11,9 @@ LeRobot provides several utilities for manipulating datasets:
 3. **Merge Datasets** - Combine multiple datasets into one. The datasets must have identical features, and episodes are concatenated in the order specified in `repo_ids`
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
-6. **Modify Tasks** - Change the natural-language task descriptions associated with episodes
-7. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
-8. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
-9. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
+6. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
+7. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
+8. **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`.
@@ -51,11 +50,11 @@ lerobot-edit-dataset \
 Divide a dataset into multiple subsets.

 ```bash
-# Split by fractions (e.g. 60% train, 20% val, 20% test)
+# Split by fractions (e.g. 80% train, 20% test, 20% val)
 lerobot-edit-dataset \
    --repo_id lerobot/pusht \
    --operation.type split \
-    --operation.splits '{"train": 0.6, "val": 0.2, "test": 0.2}'
+    --operation.splits '{"train": 0.8, "test": 0.2, "val": 0.2}'

 # Split by specific episode indices
 lerobot-edit-dataset \
@@ -90,53 +89,6 @@ lerobot-edit-dataset \
    --operation.feature_names "['observation.images.top']"
 ```

-#### Modify Tasks
-
-Change the natural-language task descriptions attached to episodes. This is useful for fixing typos, standardizing wording, or re-labeling episodes.
-
-> [!WARNING]
-> `modify_tasks` modifies the dataset **in-place** (updating `meta/tasks.parquet`, the `task_index` column in the data files, the `tasks` column in the episode metadata, and `total_tasks` in `meta/info.json`). The `--new_repo_id` and `--new_root` parameters are ignored for this operation.
-
-```bash
-# Set a single task for all episodes
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht \
-    --operation.type modify_tasks \
-    --operation.new_task "Pick up the cube and place it"
-
-# Set different tasks for specific episodes
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht \
-    --operation.type modify_tasks \
-    --operation.episode_tasks '{"0": "Task A", "1": "Task B", "2": "Task A"}'
-
-# Replace existing task strings wherever they appear
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht \
-    --operation.type modify_tasks \
-    --operation.task_replacements '{"Pick up the red cube": "Lift the red cube"}'
-
-# Combine modes in a single run
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht \
-    --operation.type modify_tasks \
-    --operation.new_task "Default task" \
-    --operation.task_replacements '{"Pick up the red cube": "Lift the red cube"}' \
-    --operation.episode_tasks '{"5": "Special task for episode 5"}'
-```
-
-**Parameters:**
-
- `new_task`: A single task string used as the default for episodes not otherwise covered.
- `episode_tasks`: Mapping from episode index to task string.
- `task_replacements`: Mapping from existing task strings to their replacements, applied to episodes whose current task matches a key. Every key must be an existing task in the dataset.
-
-The modes can be combined in a single run. Per episode, the task is resolved with the following precedence:
-
-`episode_tasks` > `task_replacements` > `new_task` > original task
-
-At least one of `new_task`, `episode_tasks`, or `task_replacements` must be specified. An episode that ends up with no task raises an error.
-
 #### Convert to Video

 Convert an image-based dataset to video format, creating a new LeRobotDataset where images are stored as videos. This is useful for reducing storage requirements and improving data loading performance. The new dataset will have the exact same structure as the original, but with images encoded as MP4 videos in the proper LeRobot format.
@@ -49,16 +49,16 @@ lerobot-record \

 All flags below are prefixed with `--dataset.rgb_encoder.` on the CLI.

-| Parameter       | Type             | Default       | Description                                                                                                                                                                                                             |
-| --------------- | ---------------- | ------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `vcodec`        | `str`            | `"libsvtav1"` | Video codec name. `"auto"` picks the first available hardware encoder from a fixed preference list, falling back to `libsvtav1`.                                                                                        |
-| `pix_fmt`       | `str`            | `"yuv420p"`   | Output pixel format. Must be supported by the chosen codec in your FFmpeg build.                                                                                                                                        |
-| `g`             | `int`            | `2`           | GOP size — a keyframe every `g` frames. Emitted as FFmpeg option `g`.                                                                                                                                                   |
-| `crf`           | `int` or `float` | `30`          | Abstract quality value, mapped per codec (see the [mapping](https://github.com/huggingface/lerobot/blob/main/src/lerobot/configs/video.py#L197)). Lower → higher quality / larger output where the mapping is monotone. |
-| `preset`        | `int` or `str`   | `12` \*       | Encoder speed preset; meaning depends on the codec. <br/>\* When unset and `vcodec=libsvtav1`, LeRobot defaults to `12`.                                                                                                |
-| `fast_decode`   | `int`            | `0`           | `libsvtav1`: `0–2`, passed via `svtav1-params`. <br/>`h264` / `hevc` (software): if `>0`, sets `tune=fastdecode`. <br/>Other codecs: usually unused.                                                                    |
-| `video_backend` | `str`            | `"pyav"`      | Only `"pyav"` is currently implemented for video encoding.                                                                                                                                                              |
-| `extra_options` | `dict`           | `{}`          | Extra FFmpeg or codec specific options merged after the structured fields above. Cannot override keys already set by those fields.                                                                                      |
+| Parameter       | Type             | Default       | Description                                                                                                                                                                            |
+| --------------- | ---------------- | ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `vcodec`        | `str`            | `"libsvtav1"` | Video codec name. `"auto"` picks the first available hardware encoder from a fixed preference list, falling back to `libsvtav1`.                                                       |
+| `pix_fmt`       | `str`            | `"yuv420p"`   | Output pixel format. Must be supported by the chosen codec in your FFmpeg build.                                                                                                       |
+| `g`             | `int`            | `2`           | GOP size — a keyframe every `g` frames. Emitted as FFmpeg option `g`.                                                                                                                  |
+| `crf`           | `int` or `float` | `30`          | Abstract quality value, mapped per codec (see the [mapping](#mapping-videoencoderconfig--ffmpeg-options) below). Lower → higher quality / larger output where the mapping is monotone. |
+| `preset`        | `int` or `str`   | `12` \*       | Encoder speed preset; meaning depends on the codec. <br/>\* When unset and `vcodec=libsvtav1`, LeRobot defaults to `12`.                                                               |
+| `fast_decode`   | `int`            | `0`           | `libsvtav1`: `0–2`, passed via `svtav1-params`. <br/>`h264` / `hevc` (software): if `>0`, sets `tune=fastdecode`. <br/>Other codecs: usually unused.                                   |
+| `video_backend` | `str`            | `"pyav"`      | Only `"pyav"` is currently implemented for video encoding.                                                                                                                             |
+| `extra_options` | `dict`           | `{}`          | Extra FFmpeg or codec specific options merged after the structured fields above. Cannot override keys already set by those fields.                                                     |

 ---

@@ -1,287 +0,0 @@
-# Writing docstrings
-
-LeRobot's API reference is generated directly from the docstrings in `src/lerobot/`. A docstring is not a
-comment — it is the published documentation for that object, and the format below is what the renderer and
-the CI checks parse.
-
-This page is the contract. If you are adding or editing anything public in `src/lerobot/`, follow it.
-
-> [!IMPORTANT]
-> **An undocumented public method is an invisible one.** `[[autodoc]]` silently skips members that have no
-> docstring — no warning, no error, it simply does not appear on the rendered page. Coverage and
-> API-reference completeness are the same problem.
-
-## The format in one example
-
-Google section headers, Hugging Face type formatting. Both, not one or the other.
-
-````python
-def send_action(self, action: RobotAction, rate_hz: float = 30.0) -> RobotAction:
-    """Command the robot to move to a target joint configuration.
-
-    Values are clipped by the configured maximum relative target before reaching the motors, so the
-    returned action may differ from the requested one.
-
-    Args:
-        action (`dict[str, float]`):
-            Target values keyed by motor name, e.g. `{"shoulder_pan.pos": 0.0}`. Keys must match the
-            robot's action features.
-        rate_hz (`float`, *optional*, defaults to `30.0`):
-            Control loop frequency.
-
-    Returns:
-        `dict[str, float]`: The action actually written to the motors after safety clipping.
-
-    Raises:
-        DeviceNotConnectedError: If the robot has not been connected.
-
-    Example:
-        ```python
-        >>> from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
-        >>> robot = SO101Follower(SO101FollowerConfig(port="/dev/ttyACM0"))  # doctest: +SKIP
-        >>> robot.connect()  # doctest: +SKIP
-        >>> robot.send_action({"shoulder_pan.pos": 0.0})  # doctest: +SKIP
-        ```
-    """
-````
-
-Cross-references are omitted from the examples on this page — see [Cross-references](#cross-references) for
-their syntax and why they cannot be shown inside a code block.
-
-## Rules
-
-### Sections
-
-`Args:` · `Returns:` · `Raises:` · `Yields:` · `Example:` · `Note:`
-
-In that order. No other section headers. A one-line summary comes first, then an optional free-form
-description, then the sections.
-
-### The `Args:` line is machine-parsed
-
-```
-name (`type`, *optional*, defaults to `X`):
-    Description, indented on its own line.
-```
-
-The `*optional*, defaults to` clause is **checked against the real signature default** by
-`make check-docstrings`. It is not decorative — if you write a default that has drifted from the code, CI
-fails. Omit the clause entirely for required parameters:
-
-```python
-    Args:
-        port (`str`):
-            Serial port the arm is connected to, e.g. `/dev/ttyACM0`.
-        max_relative_target (`float | dict[str, float]`, *optional*):
-            Caps the magnitude of the relative positional target vector. `None` disables clipping.
-        use_degrees (`bool`, *optional*, defaults to `True`):
-            Keep `True` for backward compatibility with existing policies and datasets.
-```
-
-Types go in backticks. Use `*optional*` with no `defaults to` when the default is `None` or is otherwise not
-worth restating.
-
-### `Returns:` is type-first
-
-One indented line, type first, then a colon, then the description:
-
-```python
-    Returns:
-        `dict[str, float]`: The action actually written to the motors after safety clipping.
-```
-
-`Yields:` takes the same shape.
-
-### `**Attributes**:`, never `Attributes:`
-
-doc-builder parses a bare `Attributes:` as a **synonym for `Parameters:`**, so your attributes get rendered
-as constructor arguments. This is silent and wrong. Whenever the attributes differ from the constructor
-parameters, use the bold form with a `--` separator:
-
-```python
-class Robot(abc.ABC):
-    """The base abstract class for all LeRobot-compatible robots.
-
-    **Attributes**:
-        - **config_class** (`type[RobotConfig]`) -- The expected configuration class for this robot.
-        - **name** (`str`) -- The unique robot name used to identify this robot type.
-    """
-```
-
-Note `--`, not `:`.
-
-### Cross-references
-
-Use doc-builder's bracket syntax: a square-bracketed backtick-quoted path. **Sphinx roles (`:pymeth:`,
-`:pyattr:`) are not supported** and render as literal text on the page.
-
-| Want                         | Write                               |
-| ---------------------------- | ----------------------------------- |
-| Class in the main package    | &#91;`Robot`&#93;                   |
-| Method, show the full path   | &#91;`Robot.connect`&#93;           |
-| Method, show the bare name   | &#91;`~Robot.connect`&#93;          |
-| Nested path                  | &#91;`~robots.Robot.connect`&#93;   |
-| Object in another HF library | &#91;`~accelerate.Accelerator`&#93; |
-
-The `~` strips the path from the **link text only**; the link still resolves to the full path.
-
-> [!NOTE]
-> doc-builder resolves this syntax everywhere in a page — including inside fenced code blocks. That is why
-> the docstring examples on this page use plain prose instead of cross-references: a code block containing
-> one would render the resolved link rather than the syntax you need to type. In your own docstrings, use
-> cross-references freely; this restriction only affects documentation _about_ the syntax.
-
-### Callouts
-
-Use GitHub-style blockquotes:
-
-```markdown
-> [!TIP]
-> Call this once at startup — it takes about two seconds.
-
-> [!WARNING]
-> Torque is disabled on disconnect. The arm will drop if it is holding a load.
-```
-
-The `<Tip>` component is legacy per doc-builder; don't add new ones.
-
-### Examples must be fenced
-
-An example lives inside a fenced ` ```python ` block containing `>>> `. The fence is what makes it render
-as a code block, and it is what the doctest preprocessor's regex looks for:
-
-````python
-    Example:
-        ```python
-        >>> from lerobot.robots.so_follower import SO101FollowerConfig
-        >>> cfg = SO101FollowerConfig(port="/dev/ttyACM0")
-        >>> cfg.use_degrees
-        True
-        ```
-````
-
-> [!WARNING]
-> An unfenced `>>>` is still collected — doctest finds prompts anywhere in a docstring. What you lose is the
-> rendering, so it shows up as a wall of prose on the page. Every example needs the fence.
-
-Every example either executes in CI or carries `# doctest: +SKIP`. Anything that touches hardware, a GPU, or
-downloads from the Hub gets `+SKIP`:
-
-````python
-    Example:
-        ```python
-        >>> robot.connect()  # doctest: +SKIP
-        >>> policy = ACTPolicy.from_pretrained("lerobot/act_aloha_sim_transfer_cube_human")  # doctest: +SKIP
-        ```
-````
-
-Add files containing runnable examples to `utils/documentation_tests.txt`.
-
-Put examples on the three to five genuine entry points of a module. Examples on trivial accessors are noise.
-
-## Three patterns you will hit constantly
-
-### Config dataclasses
-
-Configuration fields are historically documented with `#` comments above each field. **doc-builder cannot
-see inline comments** — such a class renders with every field listed and not a single description. Move them
-into an `Args:` block on the class docstring:
-
-```python
-@dataclass
-class SOFollowerConfig:
-    """Configuration for SO-family follower arms.
-
-    Args:
-        port (`str`):
-            Serial port the arm is connected to, e.g. `/dev/ttyACM0`.
-        max_relative_target (`float | dict[str, float]`, *optional*):
-            Caps the magnitude of the relative positional target vector. A scalar applies to all motors;
-            a dict maps motor name to a per-motor cap. `None` disables clipping.
-        use_degrees (`bool`, *optional*, defaults to `True`):
-            Keep `True` for backward compatibility with existing policies and datasets.
-    """
-
-    port: str
-    max_relative_target: float | dict[str, float] | None = None
-    use_degrees: bool = True
-```
-
-> [!IMPORTANT]
-> **doc-builder does not inherit docstrings from base classes.** LeRobot's registered config classes are
-> often thin multiple-inheritance shims:
->
-> ```python
-> @RobotConfig.register_subclass("so101_follower")
-> @dataclass
-> class SOFollowerRobotConfig(RobotConfig, SOFollowerConfig):
->     pass
-> ```
->
-> That class renders **every** field — including the ones it inherits — with no descriptions at all, no
-> matter how well the bases are documented. The `Args:` block must live on the concrete class that
-> `[[autodoc]]` names, and it must cover inherited fields too.
-
-### Base class, then concrete subclass
-
-The abstract base carries the canonical contract. Subclasses document only what deviates — port semantics,
-calibration quirks, motor layout, supported feature keys. Do not copy the base contract into every subclass.
-
-`Robot`, `Teleoperator`, `Camera`, `MotorsBus`, `ProcessorStep`, and `PreTrainedPolicy` all follow this
-shape.
-
-### Module-level aliases
-
-Several public names are aliases rather than distinct classes:
-
-```python
-SO100FollowerConfig = SOFollowerRobotConfig
-SO101FollowerConfig = SOFollowerRobotConfig
-```
-
-`[[autodoc]]` resolves the alias and renders the **canonical** class name, so a `## SO101FollowerConfig`
-heading will show `class lerobot.robots.so_follower.SOFollowerRobotConfig` in the body. Document the
-canonical class once, and mention the aliases in the page's prose rather than giving each alias its own
-autodoc block.
-
-## What not to document
-
- **Private members.** Anything starting with `_` is not part of the public API.
- **The type annotation restated as prose.** `port (`str`): A string.` adds nothing. Say what it is for.
- **Vendored upstream code.** `src/lerobot/policies/molmoact2/molmoact2_hf_model/` is vendored from
-  `transformers` and already carries upstream-style docstrings. Leave it alone — restyling it only creates
-  conflicts on the next sync. It is excluded from the API reference and from the docstring checks.
-
-## How this is enforced
-
-| Check                     | What it catches                                                                                            |
-| ------------------------- | ---------------------------------------------------------------------------------------------------------- |
-| `make check-docstrings`   | An `Args:` entry that doesn't match the signature; a documented default that has drifted from the real one |
-| `make doctest`            | Examples that no longer run                                                                                |
-| `make check-doctest-list` | Stale or unsorted entries in `utils/documentation_tests.txt`                                               |
-| `ruff` (`D` rules)        | Google-convention style violations                                                                         |
-| `interrogate`             | Docstring coverage falling below the current threshold                                                     |
-| doc-builder               | A `[[autodoc]]` path that points at something that doesn't exist — this breaks the docs build              |
-
-Run them together before opening a PR:
-
-```bash
-make check-docstrings && make doctest && pre-commit run --all-files
-```
-
-Then render the page and actually look at it:
-
-```bash
-doc-builder build lerobot docs/source/ --build_dir /tmp/doc-build
-```
-
-## Checklist
-
- [ ] Every public member you touched has a docstring.
- [ ] Every `Args:` entry matches the signature, including the `*optional*, defaults to` clause.
- [ ] `Returns:` is type-first on one indented line.
- [ ] No bare `Attributes:` — use `**Attributes**:` with `--` separators.
- [ ] No Sphinx roles — cross-references use &#91;`~module.Class.method`&#93;.
- [ ] Examples are inside a fenced ` ```python ` block, and either run in CI or carry `# doctest: +SKIP`.
- [ ] Config dataclass fields are in an `Args:` block on the concrete class, not `#` comments.
- [ ] The rendered page has been eyeballed.
@@ -44,7 +44,6 @@ from typing import Protocol

 import numpy as np

-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -57,6 +56,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
    EEBoundsAndSafety,
    InverseKinematicsEEToJoints,
 )
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import HF_LEROBOT_CALIBRATION, HF_LEROBOT_HOME, TELEOPERATORS
 from lerobot.utils.robot_utils import precise_sleep

@@ -38,7 +38,7 @@ from typing import TYPE_CHECKING

 import numpy as np

-from lerobot.lerobot_types import RobotAction
+from lerobot.types import RobotAction

 from .base import _GRIPPER_MOTOR_SCALE, IsaacTeleopTeleoperator, _isaacteleop_available
 from .config_isaac_teleop import SO101LeaderArmConfig
@@ -32,7 +32,7 @@ from typing import TYPE_CHECKING, Any

 import numpy as np

-from lerobot.lerobot_types import RobotAction
+from lerobot.types import RobotAction

 from .base import IsaacTeleopTeleoperator, _isaacteleop_available
 from .config_isaac_teleop import XRControllerConfig
@@ -26,8 +26,8 @@ from __future__ import annotations
 from dataclasses import dataclass

 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
-from lerobot.lerobot_types import RobotAction
 from lerobot.processor import ProcessorStepRegistry, RobotActionProcessorStep
+from lerobot.types import RobotAction
 from lerobot.utils.rotation import Rotation

 from .base import _GRIPPER_MOTOR_SCALE
@@ -21,7 +21,6 @@ from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.common.control_utils import predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
@@ -39,6 +38,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
    ForwardKinematicsJointsToEE,
    InverseKinematicsEEToJoints,
 )
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
 from lerobot.utils.keyboard_input import init_keyboard_listener
@@ -16,7 +16,6 @@

 from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -37,6 +36,7 @@ from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.phone import Phone, PhoneConfig
 from lerobot.teleoperators.phone.config_phone import PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.utils import log_say
@@ -17,7 +17,6 @@
 import time

 from lerobot.datasets import LeRobotDataset
-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -28,6 +27,7 @@ from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.robots.so_follower.robot_kinematic_processor import (
    InverseKinematicsEEToJoints,
 )
+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
@@ -27,7 +27,6 @@ Highlight, or DAgger via ``lerobot-rollout --strategy.type=...``.

 from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.configs import PreTrainedConfig
-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -44,6 +43,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
 from lerobot.rollout.inference import SyncInferenceConfig
 from lerobot.rollout.strategies import BaseStrategy
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.process import ProcessSignalHandler
 from lerobot.utils.utils import init_logging

@@ -15,7 +15,6 @@

 import time

-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -32,6 +31,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.teleoperators.phone import Phone, PhoneConfig
 from lerobot.teleoperators.phone.config_phone import PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data

@@ -417,7 +417,7 @@ class RTCEvaluator:
    def run_evaluation(self):
        """Run evaluation on two random dataset samples using three separate policies.

-        Note: Policies are deinitialized after each step to free memory. Large models
+        Note: Policies are deinitalized after each step to free memory. Large models
        (e.g., VLA models with billions of parameters) cannot fit three instances in
        memory simultaneously. By deleting and garbage collecting after each step,
        we ensure only one policy is loaded at a time.
@@ -21,7 +21,6 @@ from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.common.control_utils import predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
@@ -39,6 +38,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
    ForwardKinematicsJointsToEE,
    InverseKinematicsEEToJoints,
 )
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
 from lerobot.utils.keyboard_input import init_keyboard_listener
@@ -17,7 +17,6 @@

 from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -34,6 +33,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 )
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.feature_utils import combine_feature_dicts
 from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.utils import log_say
@@ -18,7 +18,6 @@
 import time

 from lerobot.datasets import LeRobotDataset
-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -29,6 +28,7 @@ from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.robots.so_follower.robot_kinematic_processor import (
    InverseKinematicsEEToJoints,
 )
+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
@@ -25,7 +25,6 @@ forward/inverse kinematics.

 from lerobot.cameras.opencv import OpenCVCameraConfig
 from lerobot.configs import PreTrainedConfig
-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -42,6 +41,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.rollout import BaseStrategyConfig, RolloutConfig, build_rollout_context
 from lerobot.rollout.inference import SyncInferenceConfig
 from lerobot.rollout.strategies import BaseStrategy
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.process import ProcessSignalHandler
 from lerobot.utils.utils import init_logging

@@ -16,7 +16,6 @@

 import time

-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -31,6 +30,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
    InverseKinematicsEEToJoints,
 )
 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

@@ -12,8 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-"""This script demonstrates how to train a Diffusion Policy on the PushT environment,
-using a dataset processed in streaming mode."""
+"""Train directly from episode-scoped Parquet and MP4 streams.
+
+For normal training, prefer ``lerobot-train --dataset.streaming=true`` so distributed sharding,
+checkpoint resume, and device placement are configured by the training pipeline. This lower-level
+example shows the underlying Python API.
+"""

 from pathlib import Path

@@ -64,18 +68,28 @@ def main():
        ACTION: [t / dataset_metadata.fps for t in range(cfg.n_action_steps)],
    }

-    # Instantiating the training dataset in streaming mode allows to not consume up memory as the data is fetched
-    # iteratively rather than being load into memory all at once. Retrieved frames are shuffled across epochs
-    dataset = StreamingLeRobotDataset(dataset_id, delta_timestamps=delta_timestamps, tolerance_s=1e-3)
+    # The first run resolves or locally builds a revision-safe MP4 index sidecar. It is never
+    # uploaded implicitly. Episode rows and video byte ranges are then prefetched together.
+    dataset = StreamingLeRobotDataset(
+        dataset_id,
+        delta_timestamps=delta_timestamps,
+        tolerance_s=1e-3,
+        episode_pool_size=16,
+        prefetch_episodes=4,
+        byte_budget_gb=4,
+        repeat=True,
+    )

    optimizer = torch.optim.Adam(policy.parameters(), lr=1e-4)
    dataloader = torch.utils.data.DataLoader(
        dataset,
-        num_workers=4,
+        # One worker owns the rank-level pool. Internal fetch concurrency is configured through
+        # StreamingLeRobotDataset.max_num_shards (the lerobot-train CLI derives it from num_workers).
+        num_workers=1,
        batch_size=16,
        pin_memory=device.type != "cpu",
-        drop_last=True,
-        prefetch_factor=2,  # loads batches with multiprocessing while policy trains
+        prefetch_factor=2,  # bounded decoded-batch queue while the policy trains
+        persistent_workers=True,
    )

    # Run training loop.
@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"

 [project]
 name = "lerobot"
-version = "0.6.2"
+version = "0.6.1"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
@@ -67,8 +67,11 @@ dependencies = [
    "einops>=0.8.0,<0.9.0",

    # Config & Hub
-    "draccus>=0.11.6,<0.12.0",
-    "huggingface-hub>=1.6.0,<2.0.0",
+    "draccus==0.10.0", # TODO: Relax version constraint
+    "huggingface-hub>=1.0.0,<2.0.0",
+    "filelock>=3.12.0,<4.0.0",
+    "fsspec>=2023.5.0,<2027.0.0",
+    "httpx>=0.27.0,<1.0.0",
    "requests>=2.32.0,<3.0.0",

    # Environments
@@ -87,7 +90,7 @@ dependencies = [

    # Build tools (required by opencv-python-headless on some platforms)
    "cmake>=3.29.0.1,<4.2.0",
-    "setuptools>=71.0.0,<82.0.0",  # torch 2.11 requires setuptools<82; a higher cap makes the resolver downgrade torch
+    "setuptools>=71.0.0,<81.0.0",
 ]

 # Optional dependencies
@@ -95,7 +98,7 @@ dependencies = [

 # ── Feature-scoped extras ──────────────────────────────────
 dataset = [
-    "datasets>=4.8.0,<5.0.0",
+    "datasets>=4.7.0,<5.0.0",
    "pandas>=2.0.0,<3.0.0", # NOTE: Transitive dependency of datasets
    "pyarrow>=21.0.0,<30.0.0", # NOTE: Transitive dependency of datasets
    "lerobot[av-dep]",
@@ -261,7 +264,7 @@ annotations = [
 # 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,<2.0.0", "mypy>=1.19.1", "ruff>=0.14.1", "lerobot[notebook]"]
 notebook = ["jupyter>=1.0.0,<2.0.0", "ipykernel>=6.0.0,<7.0.0"]
-test = ["pytest>=8.1.0,<10.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
+test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]

 # Simulation
@@ -346,7 +349,6 @@ lerobot-record="lerobot.scripts.lerobot_record:main"
 lerobot-replay="lerobot.scripts.lerobot_replay:main"
 lerobot-setup-motors="lerobot.scripts.lerobot_setup_motors:main"
 lerobot-teleoperate="lerobot.scripts.lerobot_teleoperate:main"
-lerobot-convert-dcp="lerobot.scripts.lerobot_convert_dcp:main"
 lerobot-eval="lerobot.scripts.lerobot_eval:main"
 lerobot-train="lerobot.scripts.lerobot_train:main"
 lerobot-train-tokenizer="lerobot.scripts.lerobot_train_tokenizer:main"
@@ -401,63 +403,19 @@ exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]
 # N: pep8-naming
 # TODO: Uncomment rules when ready to use
 select = [
-    "E", "W", "F", "I", "B", "C4", "T20", "N", "UP", "SIM", "D" #, "A", "S", "RUF"
+    "E", "W", "F", "I", "B", "C4", "T20", "N", "UP", "SIM" #, "A", "S", "D", "RUF"
 ]
 ignore = [
    "E501", # Line too long
    "T201", # Print statement found
    "T203", # Pprint statement found
    "B008", # Perform function call in argument defaults
-    # D100/D104: module- and package-level docstrings. The API reference is generated from class and
-    # function docstrings; a banner at the top of every file and every __init__.py would not appear on any
-    # rendered page. Coverage of the things that do get rendered is enforced by interrogate instead.
-    "D100",
-    "D104",
 ]

 [tool.ruff.lint.per-file-ignores]
-"__init__.py" = ["F401", "F403", "E402", "D104"]
+"__init__.py" = ["F401", "F403", "E402"]
 # E402: conditional-import guards (TYPE_CHECKING / is_package_available) must precede the imports they protect
 "src/lerobot/scripts/convert_dataset_v21_to_v30.py" = ["E402"]
-
-# D (pydocstyle) is enabled globally, but only holds for code that has been converted to the docstring
-# standard in docs/source/writing_docstrings.mdx. Every module below is still on the old style; each entry
-# is deleted as that module is converted, and this block can be removed once it is empty.
-#
-# Not part of the API reference and not planned for conversion: tests, examples, benchmarks, templates,
-# CI helper scripts and the packaging shim.
-"tests/**" = ["D"]
-"examples/**" = ["D"]
-"benchmarks/**" = ["D"]
-"scripts/**" = ["D"]
-"setup.py" = ["D"]
-"src/lerobot/templates/**" = ["D"]
-# Vendored from transformers; keeps its upstream docstring style so syncs stay clean.
-"src/lerobot/policies/molmoact2/molmoact2_hf_model/**" = ["D"]
-# Awaiting conversion, one PR per module.
-"src/lerobot/annotations/**" = ["D"]
-"src/lerobot/async_inference/**" = ["D"]
-"src/lerobot/common/**" = ["D"]
-"src/lerobot/configs/**" = ["D"]
-"src/lerobot/data_processing/**" = ["D"]
-"src/lerobot/datasets/**" = ["D"]
-"src/lerobot/distributed/**" = ["D"]
-"src/lerobot/envs/**" = ["D"]
-"src/lerobot/jobs/**" = ["D"]
-"src/lerobot/model/**" = ["D"]
-"src/lerobot/motors/**" = ["D"]
-"src/lerobot/optim/**" = ["D"]
-"src/lerobot/policies/**" = ["D"]
-"src/lerobot/processor/**" = ["D"]
-"src/lerobot/rewards/**" = ["D"]
-"src/lerobot/rl/**" = ["D"]
-"src/lerobot/rollout/**" = ["D"]
-"src/lerobot/scripts/**" = ["D"]
-"src/lerobot/teleoperators/**" = ["D"]
-"src/lerobot/transforms/**" = ["D"]
-"src/lerobot/transport/**" = ["D"]
-"src/lerobot/utils/**" = ["D"]
-"src/lerobot/lerobot_types.py" = ["D"]
 [tool.ruff.lint.isort]
 combine-as-imports = true
 known-first-party = ["lerobot"]
@@ -481,6 +439,7 @@ exclude_dirs = [
 skips = ["B101", "B311", "B404", "B603", "B615"]

 [tool.typos]
+default.extend-words = { trak = "trak" }
 default.extend-ignore-re = [
    "(?Rm)^.*(#|//)\\s*spellchecker:disable-line$",                      # spellchecker:disable-line
    "(?s)(#|//)\\s*spellchecker:off.*?\\n\\s*(#|//)\\s*spellchecker:on", # spellchecker:<on|off>
@@ -501,34 +460,25 @@ default.extend-ignore-identifiers-re = [
    "seperated_timestep",
 ]

-# Docstring coverage gate. `fail-under` is a RATCHET, not a target: it is set just below the currently
-# measured coverage so it passes today, and is raised in the same PR that documents a module. Never set it
-# to a value that fails on main. The destination is 100; see docs/source/writing_docstrings.mdx.
-[tool.interrogate]
-ignore-init-module = true
-ignore-init-method = true
-ignore-nested-functions = false
-ignore-magic = false
-ignore-semiprivate = false
-ignore-private = false
-ignore-property-decorators = false
-ignore-module = false
-ignore-setters = false
-fail-under = 55
-output-format = "term-missing"
-color = true
-paths = ["src/lerobot"]
-exclude = ["src/lerobot/policies/molmoact2/molmoact2_hf_model"]
+# TODO: Uncomment when ready to use
+# [tool.interrogate]
+# ignore-init-module = true
+# ignore-init-method = true
+# ignore-nested-functions = false
+# ignore-magic = false
+# ignore-semiprivate = false
+# ignore-private = false
+# ignore-property-decorators = false
+# ignore-module = false
+# ignore-setters = false
+# fail-under = 80
+# output-format = "term-missing"
+# color = true
+# paths = ["src/lerobot"]

 # TODO: Enable mypy gradually module by module across multiple PRs
 # Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations

-[tool.pytest.ini_options]
-markers = [
-    "multigpu: distributed tests needing 2-4 GPUs (CI: docker_publish.yml lane)",
-    "multigpu_heavy: 8-GPU sweeps and soak tests; never run in CI",
-]
-
 [tool.mypy]
 python_version = "3.12"
 ignore_missing_imports = true
@@ -548,19 +498,6 @@ ignore_errors = true
 module = "lerobot.envs.*"
 ignore_errors = false

-[[tool.mypy.overrides]]
-module = "lerobot.annotations.*"
-ignore_errors = false
-disallow_untyped_defs = true
-disallow_incomplete_defs = true
-check_untyped_defs = true
-
-[[tool.mypy.overrides]]
-module = "lerobot.transforms.*"
-ignore_errors = false
-disallow_untyped_defs = true
-disallow_incomplete_defs = true
-check_untyped_defs = true

 # [[tool.mypy.overrides]]
 # module = "lerobot.utils.*"
@@ -575,15 +512,6 @@ disallow_untyped_defs = true
 disallow_incomplete_defs = true
 check_untyped_defs = true

-[[tool.mypy.overrides]]
-module = "lerobot.distributed.*"
-ignore_errors = false
-
-# extra strictness for the distributed engine
-disallow_untyped_defs = true
-disallow_incomplete_defs = true
-check_untyped_defs = true
-
 [[tool.mypy.overrides]]
 module = "lerobot.optim.*"
 ignore_errors = false
@@ -0,0 +1,95 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+from __future__ import annotations
+
+import argparse
+import time
+from pathlib import Path
+
+import fsspec
+
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.streaming_sidecar import (
+    build_mp4_sidecar,
+    make_sidecar_spec,
+    published_sidecar_url,
+    range_backend_for_root,
+)
+from lerobot.streaming.sidecar import SidecarSpec
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Build a reusable MP4 byte-index sidecar for streaming.")
+    parser.add_argument("--repo-id", required=True)
+    parser.add_argument("--revision", default=None)
+    parser.add_argument("--data-root", required=True)
+    parser.add_argument("--output", required=True)
+    parser.add_argument("--episodes", type=int, default=None)
+    parser.add_argument("--workers", type=int, default=8)
+    parser.add_argument("--range-backend", choices=("fsspec", "native-http"), default=None)
+    parser.add_argument("--max-probe-mb", type=int, default=64)
+    parser.add_argument("--push", action="store_true", help="Explicitly publish the sidecar to data_root.")
+    return parser.parse_args()
+
+
+def push_sidecar(local_path: str, spec: SidecarSpec) -> list[str]:
+    if not spec.data_root.startswith("hf://"):
+        raise ValueError("--push currently supports only hf:// data roots")
+
+    fs = fsspec.filesystem("hf")
+    remote = published_sidecar_url(spec)
+    fs.put(str(Path(local_path)), remote)
+    return [remote]
+
+
+def main() -> None:
+    args = parse_args()
+
+    meta = LeRobotDatasetMetadata(args.repo_id, revision=args.revision)
+    meta.ensure_readable()
+    total = (
+        int(meta.total_episodes) if args.episodes is None else min(args.episodes, int(meta.total_episodes))
+    )
+    spec = make_sidecar_spec(meta, args.data_root)
+    if total != int(meta.total_episodes):
+        selected_paths = {
+            str(meta.get_video_file_path(ep_idx, key)) for ep_idx in range(total) for key in meta.video_keys
+        }
+        spec = SidecarSpec(
+            repo_id=spec.repo_id,
+            revision=spec.revision,
+            data_root=spec.data_root,
+            source_files=tuple(item for item in spec.source_files if item[0] in selected_paths),
+        )
+
+    start = time.perf_counter()
+    build_mp4_sidecar(
+        args.output,
+        spec,
+        range_backend=args.range_backend or range_backend_for_root(args.data_root),
+        workers=args.workers,
+        max_probe_bytes=args.max_probe_mb * 1024 * 1024,
+    )
+    elapsed = time.perf_counter() - start
+    print(f"wrote {args.output}")
+    print(f"episodes={total} files={len(spec.source_files)} elapsed_s={elapsed:.2f}")
+    if args.push:
+        if total != int(meta.total_episodes):
+            raise ValueError("Only a complete dataset sidecar can be published")
+        pushed = push_sidecar(args.output, spec)
+        for remote in pushed:
+            print(f"pushed {remote}")
+    else:
+        print("push_skipped: pass --push for explicit publication")
+
+
+if __name__ == "__main__":
+    main()
@@ -14,7 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-"""LeRobot -- PyTorch library for real-world robotics.
+"""
+LeRobot -- PyTorch library for real-world robotics.

 Provides datasets, pretrained policies, and tools for training, evaluation,
 data collection, and robot control. Integrates with Hugging Face Hub for
@@ -13,7 +13,7 @@
 # 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.
-"""To enable `lerobot.__version__`."""
+"""To enable `lerobot.__version__`"""

 from importlib.metadata import PackageNotFoundError, version

@@ -38,7 +38,6 @@ import draccus
 import grpc
 import torch

-from lerobot.lerobot_types import PolicyAction
 from lerobot.policies import get_policy_class, make_pre_post_processors
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.transport import (
@@ -46,6 +45,7 @@ from lerobot.transport import (
    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
@@ -33,10 +33,10 @@ class Camera(abc.ABC):
    - Connection/disconnection
    - Frame capture (sync/async/latest)

-    **Attributes**:
-        - **fps** (`int | None`) -- Configured frames per second.
-        - **width** (`int | None`) -- Frame width in pixels.
-        - **height** (`int | None`) -- Frame height in pixels.
+    Attributes:
+        fps (int | None): Configured frames per second
+        width (int | None): Frame width in pixels
+        height (int | None): Frame height in pixels
    """

    def __init__(self, config: CameraConfig):
@@ -50,23 +50,23 @@ class Camera(abc.ABC):
        self.height: int | None = config.height

    def __enter__(self):
-        """Context manager entry.
-
+        """
+        Context manager entry.
        Automatically connects to the camera.
        """
        self.connect()
        return self

    def __exit__(self, exc_type, exc_value, traceback) -> None:
-        """Context manager exit.
-
+        """
+        Context manager exit.
        Automatically disconnects, ensuring resources are released even on error.
        """
        self.disconnect()

    def __del__(self) -> None:
-        """Destructor safety net.
-
+        """
+        Destructor safety net.
        Attempts to disconnect if the object is garbage collected without cleanup.
        """
        try:
@@ -90,7 +90,6 @@ class Camera(abc.ABC):
    @abc.abstractmethod
    def find_cameras() -> list[dict[str, Any]]:
        """Detects available cameras connected to the system.
-
        Returns:
            List[Dict[str, Any]]: A list of dictionaries,
            where each dictionary contains information about a detected camera.
@@ -22,36 +22,15 @@ import draccus  # type: ignore  # TODO: add type stubs for draccus


 class ColorMode(str, Enum):
-    """Color channel order for frames returned by a camera.
-
-    **Attributes**:
-        - **RGB** -- Red-green-blue channel order.
-        - **BGR** -- Blue-green-red channel order, OpenCV's native order.
-    """
-
    RGB = "rgb"
    BGR = "bgr"

    @classmethod
    def _missing_(cls, value: object) -> None:
-        """Reject a value that is not a valid `ColorMode`.
-
-        Raises:
-            ValueError: Always, naming the invalid value and the valid choices.
-        """
        raise ValueError(f"`color_mode` is expected to be in {list(cls)}, but {value} is provided.")


 class Cv2Rotation(int, Enum):
-    """Clockwise rotation to apply to a frame after capture, in degrees.
-
-    **Attributes**:
-        - **NO_ROTATION** -- No rotation.
-        - **ROTATE_90** -- Rotate 90° clockwise.
-        - **ROTATE_180** -- Rotate 180°.
-        - **ROTATE_270** -- Rotate 270° clockwise (90° counter-clockwise).
-    """
-
    NO_ROTATION = 0
    ROTATE_90 = 90
    ROTATE_180 = 180
@@ -59,31 +38,11 @@ class Cv2Rotation(int, Enum):

    @classmethod
    def _missing_(cls, value: object) -> None:
-        """Reject a value that is not a valid `Cv2Rotation`.
-
-        Raises:
-            ValueError: Always, naming the invalid value and the valid choices.
-        """
        raise ValueError(f"`rotation` is expected to be in {list(cls)}, but {value} is provided.")


 # Subset from https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html
 class Cv2Backends(int, Enum):
-    """OpenCV capture backend to request when opening a device.
-
-    See the [OpenCV `VideoCaptureAPIs` reference](https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html)
-    for the full list this is a subset of.
-
-    **Attributes**:
-        - **ANY** -- Let OpenCV auto-detect the backend.
-        - **V4L2** -- Video4Linux2, the usual choice on Linux.
-        - **DSHOW** -- DirectShow, a Windows backend.
-        - **PVAPI** -- PvAPI, for Prosilica GigE cameras.
-        - **ANDROID** -- Android's native camera API.
-        - **AVFOUNDATION** -- AVFoundation, the usual choice on macOS.
-        - **MSMF** -- Microsoft Media Foundation, a Windows backend.
-    """
-
    ANY = 0
    V4L2 = 200
    DSHOW = 700
@@ -94,40 +53,15 @@ class Cv2Backends(int, Enum):

    @classmethod
    def _missing_(cls, value: object) -> None:
-        """Reject a value that is not a valid `Cv2Backends`.
-
-        Raises:
-            ValueError: Always, naming the invalid value and the valid choices.
-        """
        raise ValueError(f"`backend` is expected to be in {list(cls)}, but {value} is provided.")


@dataclass(kw_only=True)
 class CameraConfig(draccus.ChoiceRegistry, abc.ABC):  # type: ignore  # TODO: add type stubs for draccus
-    """Base configuration shared by every camera backend.
-
-    Concrete backends subclass this and register themselves with
-    `@CameraConfig.register_subclass("name")`, which is what makes `--camera.type=name` work on the
-    command line. Subclasses inherit the three fields below and must document them alongside their own.
-
-    Args:
-        fps (`int`, *optional*):
-            Requested frames per second for the color stream. `None` leaves it at the backend's default.
-        width (`int`, *optional*):
-            Requested frame width in pixels. `None` leaves it at the backend's default.
-        height (`int`, *optional*):
-            Requested frame height in pixels. `None` leaves it at the backend's default.
-    """
-
    fps: int | None = None
    width: int | None = None
    height: int | None = None

    @property
    def type(self) -> str:
-        """Return the registered name this config was registered under.
-
-        Returns:
-            `str`: The name passed to `@CameraConfig.register_subclass`, e.g. `"opencv"`.
-        """
        return str(self.get_choice_name(self.__class__))
@@ -12,7 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-"""Provides the OpenCVCamera class for capturing frames from cameras using OpenCV."""
+"""
+Provides the OpenCVCamera class for capturing frames from cameras using OpenCV.
+"""

 import logging
 import math
@@ -48,7 +50,8 @@ logger = logging.getLogger(__name__)


 class OpenCVCamera(Camera):
-    """Manages camera interactions using OpenCV for efficient frame recording.
+    """
+    Manages camera interactions using OpenCV for efficient frame recording.

    This class provides a high-level interface to connect to, configure, and read
    frames from cameras compatible with OpenCV's VideoCapture. It supports both
@@ -90,7 +93,8 @@ class OpenCVCamera(Camera):
    """

    def __init__(self, config: OpenCVCameraConfig):
-        """Initializes the OpenCVCamera instance.
+        """
+        Initializes the OpenCVCamera instance.

        Args:
            config: The configuration settings for the camera.
@@ -116,23 +120,14 @@ class OpenCVCamera(Camera):
        self.rotation: int | None = get_cv2_rotation(config.rotation)
        self.backend: int = config.backend

-        self.capture_width: int | None = None
-        self.capture_height: int | None = None
-        self._reset_connection_settings()
+        if self.height and self.width:
+            self.capture_width, self.capture_height = self.width, self.height
+            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+                self.capture_width, self.capture_height = self.height, self.width

    def __str__(self) -> str:
-        """Return a short representation naming the class and its `index_or_path`."""
        return f"{self.__class__.__name__}({self.index_or_path})"

-    def _reset_connection_settings(self) -> None:
-        """Restore settings that may have been auto-detected during a failed connection."""
-        self.fps = self.config.fps
-        self.width = self.config.width
-        self.height = self.config.height
-        self.capture_width, self.capture_height = self.width, self.height
-        if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-            self.capture_width, self.capture_height = self.height, self.width
-
    @property
    def is_connected(self) -> bool:
        """Checks if the camera is currently connected and opened."""
@@ -140,7 +135,8 @@ class OpenCVCamera(Camera):

    @check_if_already_connected
    def connect(self, warmup: bool = True) -> None:
-        """Connects to the OpenCV camera specified in the configuration.
+        """
+        Connects to the OpenCV camera specified in the configuration.

        Initializes the OpenCV VideoCapture object, sets desired camera properties
        (FPS, width, height), starts the background reading thread and performs initial checks.
@@ -154,6 +150,7 @@ class OpenCVCamera(Camera):
            ConnectionError: If the specified camera index/path is not found or fails to open.
            RuntimeError: If the camera opens but fails to apply requested settings.
        """
+
        # Use 1 thread for OpenCV operations to avoid potential conflicts or
        # blocking in multi-threaded applications, especially during data collection.
        cv2.setNumThreads(1)
@@ -167,31 +164,24 @@ class OpenCVCamera(Camera):
                f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
            )

-        try:
-            self._configure_capture_settings()
-            self._start_read_thread()
+        self._configure_capture_settings()
+        self._start_read_thread()

-            if warmup and self.warmup_s > 0:
-                start_time = time.time()
-                while time.time() - start_time < self.warmup_s:
-                    self.async_read(timeout_ms=self.warmup_s * 1000)
-                    time.sleep(0.1)
-                with self.frame_lock:
-                    if self.latest_frame is None:
-                        raise ConnectionError(f"{self} failed to capture frames during warmup.")
-        except BaseException:
-            try:
-                self._cleanup_resources()
-            except Exception:
-                logger.exception(f"Failed to fully clean up {self} after connect() failed.")
-            self._reset_connection_settings()
-            raise
+        if warmup and self.warmup_s > 0:
+            start_time = time.time()
+            while time.time() - start_time < self.warmup_s:
+                self.async_read(timeout_ms=self.warmup_s * 1000)
+                time.sleep(0.1)
+            with self.frame_lock:
+                if self.latest_frame is None:
+                    raise ConnectionError(f"{self} failed to capture frames during warmup.")

        logger.info(f"{self} connected.")

    @check_if_not_connected
    def _configure_capture_settings(self) -> None:
-        """Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
+        """
+        Applies the specified FOURCC, FPS, width, and height settings to the connected camera.

        This method attempts to set the camera properties via OpenCV. It checks if
        the camera successfully applied the settings and raises an error if not.
@@ -208,6 +198,7 @@ class OpenCVCamera(Camera):
                          to the requested value.
            DeviceNotConnectedError: If the camera is not connected.
        """
+
        if self.videocapture is None:
            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")

@@ -239,6 +230,7 @@ class OpenCVCamera(Camera):

    def _validate_fps(self) -> None:
        """Validates and sets the camera's frames per second (FPS)."""
+
        if self.videocapture is None:
            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")

@@ -253,6 +245,7 @@ class OpenCVCamera(Camera):

    def _validate_fourcc(self) -> None:
        """Validates and sets the camera's FOURCC code."""
+
        fourcc_code = cv2.VideoWriter_fourcc(*self.config.fourcc)

        if self.videocapture is None:
@@ -273,6 +266,7 @@ class OpenCVCamera(Camera):

    def _validate_width_and_height(self) -> None:
        """Validates and sets the camera's frame capture width and height."""
+
        if self.videocapture is None:
            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")

@@ -296,7 +290,8 @@ class OpenCVCamera(Camera):

    @staticmethod
    def find_cameras() -> list[dict[str, Any]]:
-        """Detects available OpenCV cameras connected to the system.
+        """
+        Detects available OpenCV cameras connected to the system.

        On Linux, it scans '/dev/video*' paths. On other systems (like macOS, Windows),
        it checks indices from 0 up to `MAX_OPENCV_INDEX`.
@@ -317,36 +312,32 @@ class OpenCVCamera(Camera):

        for target in targets_to_scan:
            camera = cv2.VideoCapture(target)
-            try:
-                if camera.isOpened():
-                    default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
-                    default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
-                    default_fps = camera.get(cv2.CAP_PROP_FPS)
-                    default_format = camera.get(cv2.CAP_PROP_FORMAT)
+            if camera.isOpened():
+                default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
+                default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                default_fps = camera.get(cv2.CAP_PROP_FPS)
+                default_format = camera.get(cv2.CAP_PROP_FORMAT)

-                    # Get FOURCC code and convert to string
-                    default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
-                    default_fourcc_code_int = int(default_fourcc_code)
-                    default_fourcc = "".join(
-                        [chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)]
-                    )
+                # Get FOURCC code and convert to string
+                default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
+                default_fourcc_code_int = int(default_fourcc_code)
+                default_fourcc = "".join([chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])

-                    camera_info = {
-                        "name": f"OpenCV Camera @ {target}",
-                        "type": "OpenCV",
-                        "id": target,
-                        "backend_api": camera.getBackendName(),
-                        "default_stream_profile": {
-                            "format": default_format,
-                            "fourcc": default_fourcc,
-                            "width": default_width,
-                            "height": default_height,
-                            "fps": default_fps,
-                        },
-                    }
+                camera_info = {
+                    "name": f"OpenCV Camera @ {target}",
+                    "type": "OpenCV",
+                    "id": target,
+                    "backend_api": camera.getBackendName(),
+                    "default_stream_profile": {
+                        "format": default_format,
+                        "fourcc": default_fourcc,
+                        "width": default_width,
+                        "height": default_height,
+                        "fps": default_fps,
+                    },
+                }

-                    found_cameras_info.append(camera_info)
-            finally:
+                found_cameras_info.append(camera_info)
                camera.release()

        return found_cameras_info
@@ -364,7 +355,8 @@ class OpenCVCamera(Camera):

    @check_if_not_connected
    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
-        """Reads a single frame synchronously from the camera.
+        """
+        Reads a single frame synchronously from the camera.

        This is a blocking call. It waits for the next available frame from the
        camera hardware via OpenCV.
@@ -380,6 +372,7 @@ class OpenCVCamera(Camera):
                          received frame dimensions don't match expectations before rotation.
            ValueError: If an invalid `color_mode` is requested.
        """
+
        start_time = time.perf_counter()

        if color_mode is not None:
@@ -399,7 +392,8 @@ class OpenCVCamera(Camera):
        return frame

    def _postprocess_image(self, image: NDArray[Any]) -> NDArray[Any]:
-        """Applies color conversion, dimension validation, and rotation to a raw frame.
+        """
+        Applies color conversion, dimension validation, and rotation to a raw frame.

        Args:
            image (np.ndarray): The raw image frame (expected BGR format from OpenCV).
@@ -412,6 +406,7 @@ class OpenCVCamera(Camera):
            RuntimeError: If the raw frame dimensions do not match the configured
                          `width` and `height`.
        """
+
        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
            raise ValueError(
                f"Invalid color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
@@ -437,7 +432,8 @@ class OpenCVCamera(Camera):
        return processed_image

    def _read_loop(self) -> None:
-        """Internal loop run by the background thread for asynchronous reading.
+        """
+        Internal loop run by the background thread for asynchronous reading.

        On each iteration:
        1. Reads a color frame (blocking call)
@@ -500,29 +496,10 @@ class OpenCVCamera(Camera):
            self.latest_timestamp = None
            self.new_frame_event.clear()

-    def _cleanup_resources(self) -> None:
-        """Stop background reads and release the capture, including after partial setup."""
-        read_thread = self.thread
-        videocapture = self.videocapture
-
-        try:
-            self._stop_read_thread()
-        finally:
-            self.videocapture = None
-            try:
-                if videocapture is not None:
-                    videocapture.release()
-            finally:
-                # Releasing the device may unblock a hardware read that outlived
-                # the first bounded join in _stop_read_thread().
-                if read_thread is not None and read_thread.is_alive():
-                    read_thread.join(timeout=2.0)
-                    if read_thread.is_alive():  # pragma: no cover
-                        logger.warning(f"{self} read thread remained alive after releasing the capture.")
-
    @check_if_not_connected
    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
-        """Reads the latest available frame asynchronously.
+        """
+        Reads the latest available frame asynchronously.

        This method retrieves the most recent frame captured by the background
        read thread. It does not block waiting for the camera hardware directly,
@@ -542,6 +519,7 @@ class OpenCVCamera(Camera):
            TimeoutError: If no frame becomes available within the specified timeout.
            RuntimeError: If an unexpected error occurs.
        """
+
        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")

@@ -576,6 +554,7 @@ class OpenCVCamera(Camera):
            DeviceNotConnectedError: If the camera is not connected.
            RuntimeError: If the camera is connected but has not captured any frames yet.
        """
+
        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")

@@ -595,7 +574,8 @@ class OpenCVCamera(Camera):
        return frame

    def disconnect(self) -> None:
-        """Disconnects from the camera and cleans up resources.
+        """
+        Disconnects from the camera and cleans up resources.

        Stops the background read thread (if running) and releases the OpenCV
        VideoCapture object.
@@ -606,6 +586,16 @@ class OpenCVCamera(Camera):
        if not self.is_connected and self.thread is None:
            raise DeviceNotConnectedError(f"{self} not connected.")

-        self._cleanup_resources()
+        if self.thread is not None:
+            self._stop_read_thread()
+
+        if self.videocapture is not None:
+            self.videocapture.release()
+            self.videocapture = None
+
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()

        logger.info(f"{self} disconnected.")
@@ -40,28 +40,17 @@ class OpenCVCameraConfig(CameraConfig):
    OpenCVCameraConfig(0, 30, 1280, 720, fourcc="YUYV")     # With YUYV format
    ```

-    Args:
-        index_or_path (`int | Path`):
-            Either an integer representing the camera device index, or a Path object pointing to a video
-            file.
-        color_mode (`ColorMode`, *optional*, defaults to `ColorMode.RGB`):
-            Color mode for image output.
-        rotation (`Cv2Rotation`, *optional*, defaults to `Cv2Rotation.NO_ROTATION`):
-            Image rotation setting (0°, 90°, 180°, or 270°).
-        warmup_s (`int`, *optional*, defaults to 1):
-            Time reading frames before returning from connect (in seconds).
-        fourcc (`str`, *optional*):
-            FOURCC code for video format (e.g., `"MJPG"`, `"YUYV"`, `"I420"`). `None` auto-detects.
-        backend (`Cv2Backends`, *optional*, defaults to `Cv2Backends.ANY`):
-            OpenCV backend identifier. See [`~cameras.Cv2Backends`] for the supported values.
-        fps (`int`, *optional*):
-            Requested frames per second for the color stream. `None` leaves it at the backend's default.
-        width (`int`, *optional*):
-            Requested frame width in pixels for the color stream. `None` leaves it at the backend's
-            default.
-        height (`int`, *optional*):
-            Requested frame height in pixels for the color stream. `None` leaves it at the backend's
-            default.
+    Attributes:
+        index_or_path: Either an integer representing the camera device index,
+                      or a Path object pointing to a video file.
+        fps: Requested frames per second for the color stream.
+        width: Requested frame width in pixels for the color stream.
+        height: Requested frame height in pixels for the color stream.
+        color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
+        rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
+        warmup_s: Time reading frames before returning from connect (in seconds)
+        fourcc: FOURCC code for video format (e.g., "MJPG", "YUYV", "I420"). Defaults to None (auto-detect).
+        backend: OpenCV backend identifier (https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html). Defaults to ANY.

    Note:
        - Only 3-channel color output (RGB/BGR) is currently supported.
@@ -77,11 +66,6 @@ class OpenCVCameraConfig(CameraConfig):
    backend: Cv2Backends = Cv2Backends.ANY

    def __post_init__(self) -> None:
-        """Normalize `color_mode`, `rotation`, and `backend`, and validate `fourcc`.
-
-        Raises:
-            ValueError: If `fourcc` is set and is not a 4-character string.
-        """
        self.color_mode = ColorMode(self.color_mode)
        self.rotation = Cv2Rotation(self.rotation)
        self.backend = Cv2Backends(self.backend)
@@ -43,24 +43,16 @@ class Reachy2CameraConfig(CameraConfig):
    )  # Left teleop camera, 640x480 @ 30FPS
    ```

-    Args:
-        name (`str`):
-            Name of the camera device. Either `"teleop"` or `"depth"`.
-        image_type (`str`):
-            Type of image stream. For the `"teleop"` camera, either `"left"` or `"right"`. For the
-            `"depth"` camera, either `"rgb"` or `"depth"` (depth is not supported yet).
-        color_mode (`ColorMode`, *optional*, defaults to `ColorMode.RGB`):
-            Color mode for image output.
-        ip_address (`str`, *optional*, defaults to `"localhost"`):
-            IP address of the robot.
-        port (`int`, *optional*, defaults to 50065):
-            Port number for the camera server.
-        fps (`int`, *optional*):
-            Requested frames per second for the color stream. Not configurable for Reachy 2 cameras.
-        width (`int`, *optional*):
-            Requested frame width in pixels for the color stream.
-        height (`int`, *optional*):
-            Requested frame height in pixels for the color stream.
+    Attributes:
+        name: Name of the camera device. Can be "teleop" or "depth".
+        image_type: Type of image stream. For "teleop" camera, can be "left" or "right".
+                    For "depth" camera, can be "rgb" or "depth". (depth is not supported yet)
+        fps: Requested frames per second for the color stream. Not configurable for Reachy 2 cameras.
+        width: Requested frame width in pixels for the color stream.
+        height: Requested frame height in pixels for the color stream.
+        color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
+        ip_address: IP address of the robot. Defaults to "localhost".
+        port: Port number for the camera server. Defaults to 50065.

    Note:
        - Only 3-channel color output (RGB/BGR) is currently supported.
@@ -73,11 +65,6 @@ class Reachy2CameraConfig(CameraConfig):
    port: int = 50065

    def __post_init__(self) -> None:
-        """Normalize `color_mode` and validate `name`/`image_type`.
-
-        Raises:
-            ValueError: If `name` is not `"teleop"`/`"depth"`, or `image_type` is not valid for `name`.
-        """
        if self.name not in ["teleop", "depth"]:
            raise ValueError(f"`name` is expected to be 'teleop' or 'depth', but {self.name} is provided.")
        if (self.name == "teleop" and self.image_type not in ["left", "right"]) or (
@@ -12,7 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-"""Provides the Reachy2Camera class for capturing frames from Reachy 2 cameras using Reachy 2's CameraManager."""
+"""
+Provides the Reachy2Camera class for capturing frames from Reachy 2 cameras using Reachy 2's CameraManager.
+"""

 from __future__ import annotations

@@ -40,8 +42,6 @@ else:
    CameraManager = None

    class CameraView:
-        """Fallback stand-in for `reachy2_sdk`'s `CameraView` when the SDK is not installed."""
-
        LEFT = 0
        RIGHT = 1

@@ -55,7 +55,8 @@ logger = logging.getLogger(__name__)


 class Reachy2Camera(Camera):
-    """Manages Reachy 2 camera using Reachy 2 CameraManager.
+    """
+    Manages Reachy 2 camera using Reachy 2 CameraManager.

    This class provides a high-level interface to connect to, configure, and read
    frames from Reachy 2 cameras. It supports both synchronous and asynchronous
@@ -69,7 +70,8 @@ class Reachy2Camera(Camera):
    """

    def __init__(self, config: Reachy2CameraConfig):
-        """Initializes the Reachy2Camera instance.
+        """
+        Initializes the Reachy2Camera instance.

        Args:
            config: The configuration settings for the camera.
@@ -86,7 +88,6 @@ class Reachy2Camera(Camera):
        self.cam_manager: CameraManager | None = None

    def __str__(self) -> str:
-        """Return a short representation naming the class, camera name, and image type."""
        return f"{self.__class__.__name__}({self.config.name}, {self.config.image_type})"

    @property
@@ -104,7 +105,8 @@ class Reachy2Camera(Camera):
            raise ValueError(f"Invalid camera name '{self.config.name}'. Expected 'teleop' or 'depth'.")

    def connect(self, warmup: bool = True) -> None:
-        """Connects to the Reachy2 CameraManager as specified in the configuration.
+        """
+        Connects to the Reachy2 CameraManager as specified in the configuration.

        Raises:
            DeviceNotConnectedError: If the camera is not connected.
@@ -118,12 +120,15 @@ class Reachy2Camera(Camera):

    @staticmethod
    def find_cameras() -> list[dict[str, Any]]:
-        """Detection not implemented for Reachy2 cameras."""
+        """
+        Detection not implemented for Reachy2 cameras.
+        """
        raise NotImplementedError("Camera detection is not implemented for Reachy2 cameras.")

    @check_if_not_connected
    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
-        """Reads a single frame synchronously from the camera.
+        """
+        Reads a single frame synchronously from the camera.

        This method retrieves the most recent frame available in Reachy 2's low-level software.

@@ -182,7 +187,8 @@ class Reachy2Camera(Camera):

    @check_if_not_connected
    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
-        """Same as [`~cameras.reachy2_camera.Reachy2Camera.read`]; Reachy 2 has no separate async path.
+        """
+        Same as read()

        Returns:
            np.ndarray: The latest captured frame as a NumPy array in the format
@@ -193,6 +199,7 @@ class Reachy2Camera(Camera):
            TimeoutError: If no frame becomes available within the specified timeout.
            RuntimeError: If an unexpected error occurs.
        """
+
        return self.read()

    @check_if_not_connected
@@ -213,6 +220,7 @@ class Reachy2Camera(Camera):
            DeviceNotConnectedError: If the camera is not connected.
            RuntimeError: If the camera is connected but has not captured any frames yet.
        """
+
        if self.latest_frame is None or self.latest_timestamp is None:
            raise RuntimeError(f"{self} has not captured any frames yet.")

@@ -226,11 +234,13 @@ class Reachy2Camera(Camera):

    @check_if_not_connected
    def disconnect(self) -> None:
-        """Stops the background read thread (if running).
+        """
+        Stops the background read thread (if running).

        Raises:
            DeviceNotConnectedError: If the camera is already disconnected.
        """
+
        if self.cam_manager is not None:
            self.cam_manager.disconnect()

@@ -12,7 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-"""Provides the RealSenseCamera class for capturing frames from Intel RealSense cameras."""
+"""
+Provides the RealSenseCamera class for capturing frames from Intel RealSense cameras.
+"""

 import logging
 import sys
@@ -44,7 +46,8 @@ pkg_name = "pyrealsense2-macosx" if sys.platform == "darwin" else "pyrealsense2"


 class RealSenseCamera(Camera):
-    """Manages interactions with Intel RealSense cameras for frame and depth recording.
+    """
+    Manages interactions with Intel RealSense cameras for frame and depth recording.

    This class provides an interface similar to `OpenCVCamera` but tailored for
    RealSense devices, leveraging the `pyrealsense2` library. It uses the camera's
@@ -106,13 +109,9 @@ class RealSenseCamera(Camera):
        ```
    """

-    # Maximum number of warmup attempts made by connect(). A failed attempt is first
-    # retried with a plain pipeline stop/start, which is usually enough to recover the
-    # stream; a USB hardware reset is performed before the final attempt as a last resort.
-    _MAX_CONNECT_ATTEMPTS = 3
-
    def __init__(self, config: RealSenseCameraConfig):
-        """Initializes the RealSenseCamera instance.
+        """
+        Initializes the RealSenseCamera instance.

        Args:
            config: The configuration settings for the camera.
@@ -122,9 +121,6 @@ class RealSenseCamera(Camera):

        self.config = config

-        self.width: int | None = config.width
-        self.height: int | None = config.height
-
        if config.serial_number_or_name.isdigit():
            self.serial_number = config.serial_number_or_name
        else:
@@ -135,9 +131,6 @@ class RealSenseCamera(Camera):
        self.use_rgb = config.use_rgb
        self.use_depth = config.use_depth
        self.warmup_s = config.warmup_s
-        self.exposure: int | None = config.exposure
-        self.gain: int | None = config.gain
-        self.white_balance: int | None = config.white_balance

        self.rs_pipeline: rs.pipeline | None = None
        self.rs_profile: rs.pipeline_profile | None = None
@@ -152,77 +145,54 @@ class RealSenseCamera(Camera):

        self.rotation: int | None = get_cv2_rotation(config.rotation)

-        self.capture_width: int | None = None
-        self.capture_height: int | None = None
-        self._reset_connection_settings()
+        if self.height and self.width:
+            self.capture_width, self.capture_height = self.width, self.height
+            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+                self.capture_width, self.capture_height = self.height, self.width

    def __str__(self) -> str:
-        """Return a short representation naming the class and its `serial_number`."""
        return f"{self.__class__.__name__}({self.serial_number})"

-    def _reset_connection_settings(self) -> None:
-        """Restore settings that may have been auto-detected during a failed connection."""
-        self.fps = self.config.fps
-        self.width = self.config.width
-        self.height = self.config.height
-        self.warmup_s = self.config.warmup_s
-        self.capture_width, self.capture_height = self.width, self.height
-        if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-            self.capture_width, self.capture_height = self.height, self.width
-
    @property
    def is_connected(self) -> bool:
        """Checks if the camera pipeline is started and streams are active."""
        return self.rs_pipeline is not None and self.rs_profile is not None

-    def _hardware_reset(self, wait_s: float = 5.0) -> None:
-        """Issue a USB hardware reset to recover an unresponsive device (common on D405)."""
-        context = rs.context()
-        for device in context.query_devices():
-            if device.get_info(rs.camera_info.serial_number) == self.serial_number:
-                logger.info(f"{self} performing hardware reset.")
-                device.hardware_reset()
-                time.sleep(wait_s)
-                return
-        logger.warning(f"{self} device not found for hardware reset, skipping.")
+    @check_if_already_connected
+    def connect(self, warmup: bool = True) -> None:
+        """
+        Connects to the RealSense camera specified in the configuration.

-    def _open_pipeline(self) -> None:
-        """Initializes the RealSense pipeline, starts it, and starts the background read thread.
+        Initializes the RealSense pipeline, configures the required streams (color
+        and optionally depth), starts the pipeline, and validates the actual stream settings.
+
+        Args:
+            warmup (bool): If True, waits at connect() time until at least one valid frame
+                           has been captured by the background thread. Defaults to True.

        Raises:
-            ValueError: If the configuration is invalid, a requested sensor option is unsupported,
-                or a requested sensor value is invalid.
+            DeviceAlreadyConnectedError: If the camera is already connected.
+            ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
            ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
            RuntimeError: If the pipeline starts but fails to apply requested settings.
        """
-        rs_pipeline = rs.pipeline()
+
+        self.rs_pipeline = rs.pipeline()
        rs_config = rs.config()
        self._configure_rs_pipeline_config(rs_config)

        try:
-            rs_profile = rs_pipeline.start(rs_config)
+            self.rs_profile = self.rs_pipeline.start(rs_config)
        except RuntimeError as e:
+            self.rs_profile = None
+            self.rs_pipeline = None
            raise ConnectionError(
                f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
            ) from e

-        self.rs_pipeline = rs_pipeline
-        self.rs_profile = rs_profile
+        self._configure_capture_settings()
+        self._start_read_thread()

-        try:
-            self._configure_capture_settings()
-            self._configure_sensor_options()
-            self._start_read_thread()
-        except BaseException:
-            self._release_after_failed_setup()
-            raise
-
-    def _run_warmup(self) -> None:
-        """Blocks until at least one valid frame has been captured by the background thread.
-
-        Raises:
-            ConnectionError: If no frame arrives before ``warmup_s`` elapses.
-        """
        # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
        self.warmup_s = max(self.warmup_s, 1)

@@ -237,71 +207,12 @@ class RealSenseCamera(Camera):
            ):
                raise ConnectionError(f"{self} failed to capture frames during warmup.")

-    def _release_after_failed_setup(self) -> None:
-        """Releases the device handle and restores auto-detected settings after a failed attempt."""
-        try:
-            self._cleanup_resources()
-        except Exception:
-            logger.exception(f"Failed to fully clean up {self} after connect() failed.")
-        self._reset_connection_settings()
-
-    @check_if_already_connected
-    def connect(self, warmup: bool = True) -> None:
-        """Connects to the RealSense camera specified in the configuration.
-
-        Initializes the RealSense pipeline, configures the required streams (color
-        and optionally depth), starts the pipeline, and validates the actual stream settings.
-
-        If the pipeline starts but no frames arrive during warmup, retries up to
-        ``_MAX_CONNECT_ATTEMPTS`` times, performing a USB hardware reset before the
-        final attempt.
-
-        Args:
-            warmup (bool): If True, waits at connect() time until at least one valid frame
-                           has been captured by the background thread. Defaults to True.
-
-        Raises:
-            DeviceAlreadyConnectedError: If the camera is already connected.
-            ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
-            ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
-            RuntimeError: If the pipeline starts but fails to apply requested settings.
-        """
-        if not warmup:
-            self._open_pipeline()
-            logger.info(f"{self} connected.")
-            return
-
-        last_error: Exception | None = None
-
-        for attempt in range(1, self._MAX_CONNECT_ATTEMPTS + 1):
-            if attempt == self._MAX_CONNECT_ATTEMPTS:
-                self._hardware_reset()
-
-            self._open_pipeline()
-
-            connected = False
-            try:
-                self._run_warmup()
-                connected = True
-            except (TimeoutError, ConnectionError) as e:
-                last_error = e
-            finally:
-                if not connected:
-                    self._release_after_failed_setup()
-
-            if connected:
-                logger.info(f"{self} connected.")
-                return
-
-            logger.warning(f"{self} warmup failed (attempt {attempt}/{self._MAX_CONNECT_ATTEMPTS}).")
-
-        raise ConnectionError(
-            f"{self} failed to capture frames after {self._MAX_CONNECT_ATTEMPTS} attempts."
-        ) from last_error
+        logger.info(f"{self} connected.")

    @staticmethod
    def find_cameras() -> list[dict[str, Any]]:
-        """Detects available Intel RealSense cameras connected to the system.
+        """
+        Detects available Intel RealSense cameras connected to the system.

        Returns:
            List[Dict[str, Any]]: A list of dictionaries,
@@ -400,6 +311,7 @@ class RealSenseCamera(Camera):
        Raises:
            DeviceNotConnectedError: If device is not connected.
        """
+
        if self.rs_profile is None:
            raise RuntimeError(f"{self}: rs_profile must be initialized before use.")

@@ -427,117 +339,10 @@ class RealSenseCamera(Camera):
        self.new_frame_event.clear()
        return self._async_read(timeout_ms=10000, read_depth=read_depth)

-    def _get_color_sensor(self) -> "rs.sensor":
-        """Returns the dedicated "RGB Camera" sensor that controls the color stream.
-
-        Manual color controls are only applied to a dedicated RGB module. Cameras
-        without one (e.g. the D405, whose color stream comes from the shared
-        "Stereo Module") are unsupported, so we never fall back to another sensor
-        to avoid altering the depth stream.
-        """
-        if self.rs_profile is None:
-            raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
-
-        device = self.rs_profile.get_device()
-        sensors = {s.get_info(rs.camera_info.name): s for s in device.query_sensors()}
-
-        if "RGB Camera" in sensors:
-            return sensors["RGB Camera"]
-
-        available = list(sensors.keys())
-        raise RuntimeError(
-            f"{self}: manual color controls require a dedicated 'RGB Camera' module, which this camera does not have. ",
-            f"Available sensors: {available}.",
-        )
-
-    def _set_sensor_option(self, sensor: "rs.sensor", option: "rs.option", value: float, label: str) -> None:
-        """Sets a sensor option, re-raising range errors with actionable diagnostics."""
-        try:
-            sensor.set_option(option, value)
-        except Exception as e:
-            range_info = ""
-            try:
-                option_range = sensor.get_option_range(option)
-                range_info = (
-                    f" (supported range: min={option_range.min}, max={option_range.max}, "
-                    f"step={option_range.step}, default={option_range.default})"
-                )
-            except Exception:
-                range_info = " (option range unavailable)"
-            raise ValueError(
-                f"{self}: failed to set {label} to {value}{range_info}. Original error: {e}"
-            ) from e
-
-    def _configure_sensor_options(self) -> None:
-        """Applies manual sensor options (exposure, gain, white balance) to the color sensor.
-
-        When exposure or gain is set, auto-exposure is disabled first. When white_balance
-        is set, auto white balance is disabled first. An omitted option is left unchanged,
-        and configuration is skipped entirely if all options are omitted.
-
-        Raises:
-            ValueError: If the sensor does not support a requested option or a requested
-                value is invalid. Invalid-value errors include the option name, requested
-                value, and supported range when available.
-        """
-        if self.exposure is None and self.gain is None and self.white_balance is None:
-            return
-
-        color_sensor = self._get_color_sensor()
-
-        requested_options = (
-            (rs.option.exposure, self.exposure, "exposure"),
-            (rs.option.gain, self.gain, "gain"),
-            (rs.option.white_balance, self.white_balance, "white balance"),
-        )
-        unsupported_options = [
-            label
-            for option, value, label in requested_options
-            if value is not None and not color_sensor.supports(option)
-        ]
-        if unsupported_options:
-            raise ValueError(
-                f"{self}: color sensor does not support requested manual options: {unsupported_options}."
-            )
-
-        manual_exposure_requested = self.exposure is not None or self.gain is not None
-        if manual_exposure_requested:
-            if color_sensor.supports(rs.option.enable_auto_exposure):
-                self._set_sensor_option(color_sensor, rs.option.enable_auto_exposure, 0, "auto-exposure")
-                logger.info(f"{self} auto-exposure disabled.")
-            else:
-                logger.warning(
-                    f"{self} sensor does not support disabling auto-exposure; "
-                    "applying manual exposure/gain directly."
-                )
-
-        if self.exposure is not None:
-            self._set_sensor_option(color_sensor, rs.option.exposure, self.exposure, "exposure")
-            logger.info(f"{self} exposure set to {self.exposure}.")
-
-        if self.gain is not None:
-            self._set_sensor_option(color_sensor, rs.option.gain, self.gain, "gain")
-            logger.info(f"{self} gain set to {self.gain}.")
-
-        if self.white_balance is not None:
-            if color_sensor.supports(rs.option.enable_auto_white_balance):
-                self._set_sensor_option(
-                    color_sensor, rs.option.enable_auto_white_balance, 0, "auto white balance"
-                )
-                logger.info(f"{self} auto white balance disabled.")
-            else:
-                logger.warning(
-                    f"{self} sensor does not support disabling auto white balance; "
-                    "applying manual white balance directly."
-                )
-            self._set_sensor_option(
-                color_sensor, rs.option.white_balance, self.white_balance, "white balance"
-            )
-            logger.info(f"{self} white balance set to {self.white_balance}.")
-
    @check_if_not_connected
    def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
-        """Reads a single frame (depth) synchronously from the camera.
+        """
+        Reads a single frame (depth) synchronously from the camera.

        This is a blocking call. It waits for a coherent set of frames (depth)
        from the camera hardware via the RealSense pipeline.
@@ -575,7 +380,8 @@ class RealSenseCamera(Camera):

    @check_if_not_connected
    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 0) -> NDArray[Any]:
-        """Reads a single frame (color) synchronously from the camera.
+        """
+        Reads a single frame (color) synchronously from the camera.

        This is a blocking call. It waits for a coherent set of frames (color)
        from the camera hardware via the RealSense pipeline.
@@ -589,6 +395,7 @@ class RealSenseCamera(Camera):
            RuntimeError: If reading frames from the pipeline fails or frames are invalid.
            ValueError: If an invalid `color_mode` is requested.
        """
+
        start_time = time.perf_counter()

        if color_mode is not None:
@@ -612,12 +419,11 @@ class RealSenseCamera(Camera):
        return frame

    def _postprocess_image(self, image: NDArray[Any], depth_frame: bool = False) -> NDArray[Any]:
-        """Applies color conversion, dimension validation, and rotation to a raw color frame.
+        """
+        Applies color conversion, dimension validation, and rotation to a raw color frame.

        Args:
            image (np.ndarray): The raw image frame (expected RGB format from RealSense).
-            depth_frame (bool): Whether `image` is a single-channel depth frame rather than a 3-channel
-                color frame.

        Returns:
            np.ndarray: The processed image frame according to `self.color_mode` and `self.rotation`.
@@ -627,6 +433,7 @@ class RealSenseCamera(Camera):
            RuntimeError: If the raw frame dimensions do not match the configured
                          `width` and `height`.
        """
+
        if self.color_mode and self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
            raise ValueError(
                f"Invalid requested color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
@@ -655,7 +462,8 @@ class RealSenseCamera(Camera):
        return processed_image

    def _read_loop(self) -> None:
-        """Internal loop run by the background thread for asynchronous reading.
+        """
+        Internal loop run by the background thread for asynchronous reading.

        On each iteration:
        1. Reads a color/depth frame (blocking call with 10s timeout)
@@ -688,9 +496,6 @@ class RealSenseCamera(Camera):
                capture_time = time.perf_counter()

                with self.frame_lock:
-                    # Under the lock, so a late frame cannot resurrect the buffer _stop_read_thread() cleared.
-                    if stop_event.is_set():
-                        break
                    if self.use_rgb:
                        self.latest_color_frame = processed_color_frame
                    if self.use_depth:
@@ -736,27 +541,6 @@ class RealSenseCamera(Camera):
            self.latest_timestamp = None
            self.new_frame_event.clear()

-    def _cleanup_resources(self) -> None:
-        """Stop background reads and stop the pipeline, including after partial setup."""
-        read_thread = self.thread
-        rs_pipeline = self.rs_pipeline
-
-        try:
-            self._stop_read_thread()
-        finally:
-            self.rs_pipeline = None
-            self.rs_profile = None
-            try:
-                if rs_pipeline is not None:
-                    rs_pipeline.stop()
-            finally:
-                # Stopping the pipeline may unblock a hardware read that outlived
-                # the first bounded join in _stop_read_thread().
-                if read_thread is not None and read_thread.is_alive():
-                    read_thread.join(timeout=2.0)
-                    if read_thread.is_alive():  # pragma: no cover
-                        logger.warning(f"{self} read thread remained alive after stopping the pipeline.")
-
    def _async_read(self, timeout_ms: float, read_depth: bool = False) -> NDArray[Any]:
        """Shared helper for :meth:`async_read`/:meth:`async_read_depth`: return the latest buffered frame."""
        if self.thread is None or not self.thread.is_alive():
@@ -779,7 +563,8 @@ class RealSenseCamera(Camera):

    @check_if_not_connected
    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
-        """Reads the latest available frame data (color) asynchronously.
+        """
+        Reads the latest available frame data (color) asynchronously.

        This method retrieves the most recent color frame captured by the background
        read thread. It does not block waiting for the camera hardware directly,
@@ -799,6 +584,7 @@ class RealSenseCamera(Camera):
            TimeoutError: If no frame data becomes available within the specified timeout.
            RuntimeError: If the background thread died unexpectedly or another error occurs.
        """
+
        if not self.use_rgb:
            raise RuntimeError(f"{self}: cannot read color — camera was configured with use_rgb=False.")

@@ -884,18 +670,32 @@ class RealSenseCamera(Camera):
        return self._read_latest(max_age_ms=max_age_ms, read_depth=True)

    def disconnect(self) -> None:
-        """Disconnects from the camera, stops the pipeline, and cleans up resources.
+        """
+        Disconnects from the camera, stops the pipeline, and cleans up resources.

        Stops the background read thread (if running) and stops the RealSense pipeline.

        Raises:
            DeviceNotConnectedError: If the camera is already disconnected (pipeline not running).
        """
+
        if not self.is_connected and self.thread is None:
            raise DeviceNotConnectedError(
                f"Attempted to disconnect {self}, but it appears already disconnected."
            )

-        self._cleanup_resources()
+        if self.thread is not None:
+            self._stop_read_thread()
+
+        if self.rs_pipeline is not None:
+            self.rs_pipeline.stop()
+            self.rs_pipeline = None
+            self.rs_profile = None
+
+        with self.frame_lock:
+            self.latest_color_frame = None
+            self.latest_depth_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()

        logger.info(f"{self} disconnected.")
@@ -36,39 +36,16 @@ class RealSenseCameraConfig(CameraConfig):
    RealSenseCameraConfig("0123456789", 30, 640, 480, rotation=Cv2Rotation.ROTATE_90)  # With 90° rotation
    ```

-    Args:
-        serial_number_or_name (`str`):
-            Unique serial number or human-readable name to identify the camera.
-        color_mode (`ColorMode`, *optional*, defaults to `ColorMode.RGB`):
-            Color mode for image output.
-        use_rgb (`bool`, *optional*, defaults to `True`):
-            Whether to enable the color stream.
-        use_depth (`bool`, *optional*, defaults to `False`):
-            Whether to enable the depth stream.
-        rotation (`Cv2Rotation`, *optional*, defaults to `Cv2Rotation.NO_ROTATION`):
-            Image rotation setting (0°, 90°, 180°, or 270°).
-        warmup_s (`int`, *optional*, defaults to 1):
-            Time reading frames before returning from connect (in seconds).
-        exposure (`int`, *optional*):
-            Manual exposure value for the color sensor. When set, auto-exposure is disabled and this fixed
-            value is used. Valid ranges are camera-model specific and reported if the value is rejected.
-            `None` leaves auto-exposure unchanged.
-        gain (`int`, *optional*):
-            Manual gain value for the color sensor. When set, auto-exposure is disabled and this fixed
-            gain is used, which also freezes exposure at its current value when no exposure is configured.
-            Valid ranges are camera-model specific and reported if the value is rejected. `None` leaves it
-            unchanged.
-        white_balance (`int`, *optional*):
-            Manual white balance value for the color sensor. When set, auto white balance is disabled and
-            this fixed value is used. Valid ranges are camera-model specific and reported if the value is
-            rejected. `None` leaves it unchanged.
-        fps (`int`, *optional*):
-            Requested frames per second for the color stream. The depth stream, when enabled, uses the
-            same FPS. Either all of `fps`, `width`, and `height` must be set, or none of them.
-        width (`int`, *optional*):
-            Requested frame width in pixels for the color stream.
-        height (`int`, *optional*):
-            Requested frame height in pixels for the color stream.
+    Attributes:
+        fps: Requested frames per second for the color stream.
+        width: Requested frame width in pixels for the color stream.
+        height: Requested frame height in pixels for the color stream.
+        serial_number_or_name: Unique serial number or human-readable name to identify the camera.
+        color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
+        use_rgb: Whether to enable the color stream. Defaults to True.
+        use_depth: Whether to enable depth stream. Defaults to False.
+        rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
+        warmup_s: Time reading frames before returning from connect (in seconds)

    Note:
        - Either name or serial_number must be specified.
@@ -84,35 +61,14 @@ class RealSenseCameraConfig(CameraConfig):
    use_depth: bool = False
    rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
    warmup_s: int = 1
-    exposure: int | None = None
-    gain: int | None = None
-    white_balance: int | None = None

    def __post_init__(self) -> None:
-        """Normalize `color_mode`/`rotation` and validate the stream and sensor-option settings.
-
-        Raises:
-            ValueError: If neither `use_rgb` nor `use_depth` is enabled, if a manual color sensor option
-                is set without `use_rgb=True`, or if only some of `fps`/`width`/`height` are set.
-        """
        self.color_mode = ColorMode(self.color_mode)
        self.rotation = Cv2Rotation(self.rotation)

        if not self.use_rgb and not self.use_depth:
            raise ValueError("At least one of `use_rgb` or `use_depth` must be enabled.")

-        manual_color_options = {
-            "exposure": self.exposure,
-            "gain": self.gain,
-            "white_balance": self.white_balance,
-        }
-        configured_color_options = [name for name, value in manual_color_options.items() if value is not None]
-        if configured_color_options and not self.use_rgb:
-            raise ValueError(
-                "Manual color sensor options require `use_rgb=True`. "
-                f"Configured options: {configured_color_options}."
-            )
-
        values = (self.fps, self.width, self.height)
        if any(v is not None for v in values) and any(v is None for v in values):
            raise ValueError(
@@ -23,33 +23,6 @@ from .configs import CameraConfig, Cv2Rotation


 def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[str, Camera]:
-    """Instantiate one [`~cameras.Camera`] per entry in a mapping of configs.
-
-    Dispatches on each config's registered `type` to build the matching backend class. This only
-    constructs the camera objects; call [`~cameras.Camera.connect`] on each before use.
-
-    Args:
-        camera_configs (`dict[str, CameraConfig]`):
-            Camera configs keyed by the name each camera should be identified by, e.g. in a robot's
-            observation features.
-
-    Returns:
-        `dict[str, Camera]`: A camera instance per key, in the same order as `camera_configs`.
-
-    Raises:
-        ValueError: If a config's type is not a known backend and building it via the generic device
-            factory also fails.
-
-    Example:
-        ```python
-        >>> from lerobot.cameras.opencv import OpenCVCameraConfig
-        >>> from lerobot.cameras.utils import make_cameras_from_configs
-        >>> configs = {"top": OpenCVCameraConfig(index_or_path=0, fps=30, width=640, height=480)}
-        >>> cameras = make_cameras_from_configs(configs)
-        >>> list(cameras.keys())
-        ['top']
-        ```
-    """
    cameras: dict[str, Camera] = {}

    for key, cfg in camera_configs.items():
@@ -84,15 +57,6 @@ def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[s


 def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
-    """Map a [`~cameras.Cv2Rotation`] to the OpenCV rotation flag `cv2.rotate` expects.
-
-    Args:
-        rotation (`Cv2Rotation`):
-            The configured rotation.
-
-    Returns:
-        `int | None`: The matching `cv2.ROTATE_*` constant, or `None` for [`~cameras.Cv2Rotation.NO_ROTATION`].
-    """
    import cv2  # type: ignore  # TODO: add type stubs for OpenCV

    if rotation == Cv2Rotation.ROTATE_90:
@@ -14,9 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-"""Captures frames from remote cameras via ZeroMQ.
-
-Uses a JSON protocol of the following form:
+"""
+ZMQCamera - Captures frames from remote cameras via ZeroMQ using JSON protocol in the
+following format:
    {
        "timestamps": {"camera_name": float},
        "images": {"camera_name": "<base64-jpeg>"}
@@ -52,7 +52,8 @@ logger = logging.getLogger(__name__)


 class ZMQCamera(Camera):
-    """Manages camera interactions via ZeroMQ for receiving frames from a remote server.
+    """
+    Manages camera interactions via ZeroMQ for receiving frames from a remote server.

    This class connects to a ZMQ Publisher, subscribes to frame topics, and decodes
    incoming JSON messages containing Base64 encoded images. It supports both
@@ -80,11 +81,6 @@ class ZMQCamera(Camera):
    """

    def __init__(self, config: ZMQCameraConfig):
-        """Initialize the camera with the given configuration.
-
-        Args:
-            config: Camera configuration, including the image server's address and port.
-        """
        require_package("pyzmq", extra="pyzmq-dep", import_name="zmq")
        super().__init__(config)

@@ -109,7 +105,6 @@ class ZMQCamera(Camera):
        self.new_frame_event: Event = Event()

    def __str__(self) -> str:
-        """Return a short representation naming the camera and its server address."""
        return f"ZMQCamera({self.camera_name}@{self.server_address}:{self.port})"

    @property
@@ -125,6 +120,7 @@ class ZMQCamera(Camera):
            warmup (bool): If True, waits for the camera to provide at least one
                           valid frame before returning. Defaults to True.
        """
+
        logger.info(f"Connecting to {self}...")

        try:
@@ -175,11 +171,15 @@ class ZMQCamera(Camera):

    @staticmethod
    def find_cameras() -> list[dict[str, Any]]:
-        """Detection not implemented for ZMQ cameras; they require manual server address/port configuration."""
+        """
+        Detection not implemented for ZMQ cameras. These cameras require manual configuration (server address/port).
+        """
        raise NotImplementedError("Camera detection is not implemented for ZMQ cameras.")

    def _read_from_hardware(self) -> NDArray[Any]:
-        """Reads a single frame directly from the ZMQ socket."""
+        """
+        Reads a single frame directly from the ZMQ socket.
+        """
        if not self.is_connected or self.socket is None:
            raise DeviceNotConnectedError(f"{self} is not connected.")

@@ -215,7 +215,8 @@ class ZMQCamera(Camera):

    @check_if_not_connected
    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
-        """Reads a single frame synchronously from the camera.
+        """
+        Reads a single frame synchronously from the camera.

        This is a blocking call. It waits for the next available frame from the
        camera background thread.
@@ -242,7 +243,9 @@ class ZMQCamera(Camera):
        return frame

    def _read_loop(self) -> None:
-        """Internal loop run by the background thread for asynchronous reading."""
+        """
+        Internal loop run by the background thread for asynchronous reading.
+        """
        stop_event = self.stop_event
        if stop_event is None:
            raise RuntimeError(f"{self}: stop_event is not initialized.")
@@ -303,7 +306,8 @@ class ZMQCamera(Camera):

    @check_if_not_connected
    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
-        """Reads the latest available frame asynchronously.
+        """
+        Reads the latest available frame asynchronously.

        Args:
            timeout_ms (float): Maximum time in milliseconds to wait for a frame
@@ -317,6 +321,7 @@ class ZMQCamera(Camera):
            TimeoutError: If no frame data becomes available within the specified timeout.
            RuntimeError: If the background thread is not running.
        """
+
        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")

@@ -348,6 +353,7 @@ class ZMQCamera(Camera):
            DeviceNotConnectedError: If the camera is not connected.
            RuntimeError: If the camera is connected but has not captured any frames yet.
        """
+
        if self.thread is None or not self.thread.is_alive():
            raise RuntimeError(f"{self} read thread is not running.")

@@ -24,32 +24,6 @@ __all__ = ["ZMQCameraConfig", "ColorMode"]
@CameraConfig.register_subclass("zmq")
@dataclass
 class ZMQCameraConfig(CameraConfig):
-    """Configuration for a camera served over a ZeroMQ socket by `ImageServer`.
-
-    Use this to read frames from a camera attached to a different machine (e.g. a Raspberry Pi on a
-    robot), which streams JPEG-encoded frames to this config's `server_address`/`port` over ZMQ.
-
-    Args:
-        server_address (`str`):
-            Address of the machine running the image server, e.g. `"192.168.1.50"`.
-        port (`int`, *optional*, defaults to 5555):
-            TCP port the image server is publishing on. Must be between 1 and 65535.
-        camera_name (`str`, *optional*, defaults to `"zmq_camera"`):
-            Name used to identify this camera in logs and observation keys.
-        color_mode (`ColorMode`, *optional*, defaults to `ColorMode.RGB`):
-            Color mode for the decoded frames.
-        timeout_ms (`int`, *optional*, defaults to 5000):
-            How long to wait for a frame before raising a timeout, in milliseconds. Must be positive.
-        warmup_s (`int`, *optional*, defaults to 1):
-            Time spent reading frames before returning from connect, in seconds.
-        fps (`int`, *optional*):
-            Requested frames per second. `None` leaves it at the server's own rate.
-        width (`int`, *optional*):
-            Requested frame width in pixels. `None` leaves it at the server's own resolution.
-        height (`int`, *optional*):
-            Requested frame height in pixels. `None` leaves it at the server's own resolution.
-    """
-
    server_address: str
    port: int = 5555
    camera_name: str = "zmq_camera"
@@ -58,12 +32,6 @@ class ZMQCameraConfig(CameraConfig):
    warmup_s: int = 1

    def __post_init__(self) -> None:
-        """Normalize `color_mode` and validate the socket settings.
-
-        Raises:
-            ValueError: If `timeout_ms` is not positive, `server_address` is empty, or `port` is outside
-                the 1-65535 range.
-        """
        self.color_mode = ColorMode(self.color_mode)

        if self.timeout_ms <= 0:
@@ -14,8 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-"""Streams camera images over ZMQ.
-
+"""
+Streams camera images over ZMQ.
 Uses lerobot's OpenCVCamera for capture, encodes images to base64 and sends them over ZMQ.
 """

@@ -47,12 +47,6 @@ class CameraCaptureThread:
    """Background thread that continuously captures and encodes frames from a camera."""

    def __init__(self, camera: OpenCVCamera, name: str):
-        """Initialize the capture thread for one camera.
-
-        Args:
-            camera: The already-connected camera to read frames from.
-            name: Name used to identify this camera in log messages.
-        """
        self.camera = camera
        self.name = name
        self.latest_encoded: str | None = None  # Pre-encoded JPEG as base64
@@ -95,21 +89,7 @@ class CameraCaptureThread:


 class ImageServer:
-    """Reads from one or more local OpenCV cameras and publishes their frames over a ZMQ `PUB` socket.
-
-    Runs on the machine physically connected to the cameras (e.g. a Raspberry Pi on a robot); pair with
-    [`~cameras.zmq.ZMQCamera`] on the machine that consumes the stream.
-    """
-
    def __init__(self, config: dict, port: int = 5555):
-        """Connect every camera listed in `config` and open the publishing socket.
-
-        Args:
-            config: A mapping with an optional `"fps"` (the publish loop rate, not the camera capture
-                rate) and a `"cameras"` mapping of camera name to a dict with `"device_id"`, `"shape"`
-                (`[height, width]`), and `"fourcc"` keys.
-            port: TCP port to publish on.
-        """
        # fps controls the publish loop rate (how often frames are sent over ZMQ), not the camera capture rate
        self.fps = config.get("fps", 30)
        self.cameras: dict[str, OpenCVCamera] = {}
@@ -122,7 +102,6 @@ class ImageServer:
                fps=self.fps,
                width=shape[1],
                height=shape[0],
-                fourcc=cfg.get("fourcc", "MJPG"),
                color_mode=ColorMode.RGB,
            )
            camera = OpenCVCamera(cam_config)
@@ -144,11 +123,6 @@ class ImageServer:
        logger.info(f"ImageServer running on port {port}")

    def run(self):
-        """Start each camera's capture thread and publish frames until interrupted.
-
-        Blocks until `KeyboardInterrupt`, then stops the capture threads, disconnects the cameras, and
-        closes the socket.
-        """
        frame_count = 0
        frame_times = deque(maxlen=60)
        last_published_ts: dict[str, float] = {}
@@ -35,9 +35,9 @@ else:

 if TYPE_CHECKING:
    from lerobot.datasets import LeRobotDataset
-from lerobot.lerobot_types import PolicyAction
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.robots import Robot
+from lerobot.types import PolicyAction


 def predict_action(
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Pepijn	21321d7a27	Merge remote-tracking branch 'origin/main' into codex/episode-video-streaming-byte-cache # Conflicts: # src/lerobot/scripts/lerobot_train.py	2026-07-28 11:18:45 +02:00
Pepijn	fdf6014214	Remove streaming benchmark prototypes	2026-07-28 11:14:10 +02:00
Pepijn	4db03ed535	Derive streaming decoder cap from episode pool	2026-07-27 20:49:19 +02:00
Pepijn	79cfb52a71	Harden production dataset streaming pipeline	2026-07-27 16:07:48 +02:00
Pepijn	c0e9b0bbff	Preserve private Hub auth in streaming readers	2026-07-27 09:41:57 +02:00
Pepijn	b5a13e43ce	Merge origin/main into streaming byte-cache branch	2026-07-27 09:37:33 +02:00
Pepijn	1e7e0b6de5	Fix streaming Parquet imports and video fallback	2026-07-24 16:41:54 +02:00
Pepijn	a3edab661b	chore(streaming): remove obsolete benchmark flags	2026-07-24 13:36:17 +02:00
Pepijn	564ba6395a	Merge remote-tracking branch 'origin/main' into codex/episode-video-streaming-byte-cache	2026-07-24 13:26:42 +02:00
Pepijn	85086fed7a	refactor(streaming): split episode data plane	2026-07-24 13:26:36 +02:00
Pepijn	ee06d1005f	refactor(dataset): make streaming pool rank-owned	2026-07-23 17:12:04 +02:00
Pepijn	1c47809bf6	chore: Merge origin/main into streaming branch	2026-07-23 16:10:46 +02:00
Pepijn	3d70b21aac	feat(dataset): integrate episode streaming into training	2026-07-23 16:10:39 +02:00
Pepijn	b85620657f	chore: Merge origin/main into streaming branch	2026-07-23 14:27:30 +02:00
Pepijn	fbfc861cf2	refactor(streaming): exact coverage is the only pool mode Drop the with-replacement sampled path: delete run_pool_stream_simulation and the --coverage flag; the streaming keep-up sim always uses run_exact_coverage_stream (ExactCoveragePool), so every frame of every episode is decoded exactly once per epoch. --pool-samples-per-episode is kept as a deprecated no-op so existing commands still parse (exact mode evicts an episode only when all its frames are emitted, so a turnover cadence no longer applies). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-03 15:07:11 +02:00
Pepijn	06aa6a0425	feat(streaming): exact-once epoch coverage for the byte-cache episode pool The pool path sampled frames with replacement and never guaranteed a full epoch (episodes rotated on a fixed cadence; frames drawn randomly, none tracked). Add ExactCoveragePool: a deterministic planner that enumerates every frame of every episode exactly once per epoch while keeping at most pool_size episodes resident, so batch mixing stays high (uniform draw over all remaining frames in the pool) but coverage is complete and reproducible. Mechanics (the "evict only when all frames sampled" model): episodes are admitted in a seeded global permutation; each resident episode carries a seeded frame-index shuffle; each draw picks a resident episode with probability proportional to its remaining frames and pops one; an episode is evicted only when its last frame is emitted, then a new one is admitted; the epoch ends when admission is exhausted and every resident episode drains. Order is a pure function of (seed, epoch) -> resumable by deterministic fast-forward. The planner does no I/O and exposes admission_order so callers can prefetch episodes ahead of the sampling frontier. Wired into the benchmark as --coverage {sampled,exact}: run_exact_coverage_stream prefetches stream_prefetch_episodes beyond the frontier so a freshly admitted episode's bytes are resident before it is drawn, then decodes each frame once, paced to target. Tests: 7 planner unit tests (exact-once coverage incl. a 45k-frame epoch, pool-size bound, per-(seed,epoch) determinism with coverage preserved, admission/eviction events, coupon-collector mixing, zero-length episodes) and a mocked-cache structural test of run_exact_coverage_stream asserting every-frame-once-per-camera plus the prefetch-before-decode invariant. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-03 15:01:35 +02:00
Pepijn	be64ded80f	perf(streaming): sub-range parallel fetch + non-blocking pool replacement The 64-vs-128-worker benchmark pair proved a per-host throughput ceiling (~270 MiB/s) on the HF bucket path: doubling connections exactly halved per-connection speed (4.8 -> 2.2 MiB/s) and left the aggregate flat, while per-episode latency doubled (5.7s -> 12s) and keep-up worsened. Steady-state demand (148 MiB/s) is well below the ceiling; the keep-up misses come entirely from consumer stalls (refill_wait 14-19s of ~84s): the sim blocks the training hot path on ensure_ready() for the FIFO-head replacement while episodes take 5.7-12s to arrive. Two fixes: - Non-blocking replacements: EpisodeByteCache.is_ready() (all cameras cached or futures done, no blocking) and the stream sim now swaps a replacement only when it is already resident, deferring otherwise; fetch capacity (~2x demand) repays the debt on later batches. A deferred_swaps metric is reported. - Sub-range parallel fetch (native-http): --range-subranges N splits one camera GET into N concurrent sub-range GETs. Under a per-host ceiling this adds no bandwidth but divides per-episode latency by ~N. Keep workers x subranges near the ~64-connection saturation point (e.g. --workers 16 --range-subranges 4). Verified: sub-range span math + order-preserving concat and is_ready semantics (unit-level, network stubbed). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-03 14:05:46 +02:00
Pepijn	88843ed675	Fix episode pool stream benchmark pacing	2026-06-22 20:40:33 +02:00
Pepijn	f2b5c4a47b	Add distributed episode pool benchmark summaries	2026-06-22 17:08:02 +02:00
Pepijn	9202fcea96	Fix pool sampling camera timestamps	2026-06-22 16:44:37 +02:00
Pepijn	ef47c35178	Benchmark random sampling from episode pool	2026-06-22 16:26:26 +02:00
Pepijn	6d6c82eb8c	Add GOP window range benchmark	2026-06-22 15:10:21 +02:00
Pepijn	9201be92cb	Log failed HTTP range attempts	2026-06-22 12:42:30 +02:00
Pepijn	0064a06205	Instrument HfFileSystem range requests	2026-06-22 12:07:59 +02:00
Pepijn	710171ccac	Clarify native HTTP exception timing	2026-06-22 12:02:12 +02:00
Pepijn	0f8257443c	Retry native HTTP timeout statuses	2026-06-22 11:35:28 +02:00
Pepijn	3a09d0c48a	Track native HTTP failed attempt timing	2026-06-22 11:27:08 +02:00
Pepijn	03fc5e3ea9	Allow dynamic range timing counters	2026-06-19 09:43:57 +02:00
Pepijn	28c3e095bf	Report native HTTP chunk timing	2026-06-19 09:34:22 +02:00
Pepijn	5bfb749a9b	Log episode cache fill progress	2026-06-18 18:43:43 +02:00
Pepijn	51c023a7a1	Tune native HTTP range diagnostics	2026-06-17 21:50:05 +02:00
Pepijn	51ea18cb7a	Allow native HTTP sidecar range diagnostics	2026-06-17 21:36:57 +02:00
Pepijn	04ab43b8d2	Report range read timing breakdown	2026-06-17 21:20:08 +02:00
Pepijn	cdfe192491	Remove random frame benchmark path	2026-06-17 21:14:42 +02:00
Pepijn	3451e53452	Use HfFileSystem for sidecar episode benchmark	2026-06-17 21:01:43 +02:00
Pepijn	30849ce74f	Report memory usage in cache benchmarks	2026-06-17 20:54:12 +02:00
Pepijn	7d6907c444	Add random frame range fetch benchmark	2026-06-17 20:48:46 +02:00
Pepijn	d99e1fe89d	Report episode cache fill stage timings	2026-06-17 20:29:57 +02:00
Pepijn	7fcde61b69	Report full dataset estimate in episode cache benchmark	2026-06-17 20:25:21 +02:00
Pepijn	bdfe8f8ce9	Use full MP4 sidecar for episode cache benchmark	2026-06-17 20:22:04 +02:00
Pepijn	34d0495d03	Retry transient native HTTP range failures	2026-06-17 20:19:54 +02:00
Pepijn	834c282631	Make episode cache benchmark fetch-only by default	2026-06-17 20:16:30 +02:00
Pepijn	f132885cbc	Pin Hub range cache and datasets main sources	2026-06-17 19:46:41 +02:00
Pepijn	d0686be2f5	Add episode video streaming byte cache	2026-06-17 19:31:02 +02:00