feat(openarm): add end-effector kinematics support

Expose optional URDF-based forward/inverse kinematics for the OpenArm follower so it can be recorded/commanded in end-effector (Cartesian) space, mirroring the SO-100 kinematics pattern. - config: add optional urdf_path / target_frame_name fields - robot: add arm_motor_names property and make_kinematics() helper that builds a RobotKinematics solver (lazy import; placo only needed when used) No behavior change when urdf_path is unset (joint-space only). Co-authored-by: Cursor <cursoragent@cursor.com>
fix(rollout): apply torch compile mode correctly (#4268 )
2026-07-31 13:39:40 +00:00 · 2026-07-31 15:24:25 +02:00 · 2026-07-31 14:59:17 +02:00 · 2026-07-31 14:53:12 +02:00 · 2026-07-31 14:49:30 +02:00 · 2026-07-31 14:48:43 +02:00
232 changed files with 4507 additions and 1253 deletions
@@ -72,19 +72,19 @@ jobs:
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          persist-credentials: false
          lfs: true
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          persist-credentials: false
          lfs: true
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          persist-credentials: false
          lfs: true
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@v4
+        uses: actions/upload-artifact@v7
        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@v4
+        uses: actions/upload-artifact@v7
        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@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          persist-credentials: false
          lfs: true
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          persist-credentials: false
          lfs: true
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
        with:
          persist-credentials: false
          lfs: true
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          persist-credentials: false
          lfs: true
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          persist-credentials: false
          lfs: true
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
        if: ${{ env.DOCKERHUB_USERNAME != '' }}
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          persist-credentials: false
@@ -52,21 +52,21 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
        with:
          lfs: true
          persist-credentials: false
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@v6
+      - uses: actions/checkout@v7
        with:
          lfs: true
          persist-credentials: false
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@6108e850ae1cf2f71bb0815a600bcd50c39abfa7  # main
    with:
      package_name: lerobot
    secrets:
@@ -55,7 +55,7 @@ jobs:
      github.repository == 'huggingface/lerobot'
    permissions:
      contents: read
-    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@e60a538eea9817ab312196d0d233604b01697265  # main
+    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@6108e850ae1cf2f71bb0815a600bcd50c39abfa7  # main
    with:
      commit_sha: ${{ github.sha }}
      package: lerobot
@@ -78,7 +78,7 @@ jobs:
    permissions:
      contents: read
      pull-requests: write
-    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@e60a538eea9817ab312196d0d233604b01697265  # main
+    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@6108e850ae1cf2f71bb0815a600bcd50c39abfa7  # main
    with:
      commit_sha: ${{ github.event.pull_request.head.sha }}
      pr_number: ${{ github.event.number }}
@@ -69,7 +69,7 @@ jobs:
      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        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@d0cc045d04ccac9d8b7881df0226f9e82c39688e  # v6
+        uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990  # v8.3.2
        with:
          enable-cache: true
          version: ${{ env.UV_VERSION }}
@@ -63,7 +63,7 @@ jobs:
      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        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@d0cc045d04ccac9d8b7881df0226f9e82c39688e  # v6
+        uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990  # v8.3.2
        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@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          lfs: true
          persist-credentials: false
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f  # v3
+        uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c  # v4.2.0
        with:
          cache-binary: false
      - name: Login to Docker Hub
-        uses: docker/login-action@c94ce9fb468520275223c153574b00df6fe4bcc9  # v3
+        uses: docker/login-action@af1e73f918a031802d376d3c8bbc3fe56130a9b0  # v4.4.0
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
      - name: Build and push Docker image
-        uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8  # v6
+        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a  # v7.3.0
        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@v8
+      - uses: actions/github-script@v9
        with:
          script: |
            // Setup Input Text
@@ -48,12 +48,12 @@ jobs:
    outputs:
      changed: ${{ steps.diff.outputs.changed }}
    steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
        with:
          persist-credentials: false
      - name: Setup uv and Python
-        uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
+        uses: astral-sh/setup-uv@v8.3.2 # 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@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/upload-artifact@v7 # 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@v6
+      - uses: actions/checkout@v7
        with:
          lfs: true
          persist-credentials: false
      - name: Download updated lockfile
-        uses: actions/download-artifact@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/download-artifact@v8 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: astral-sh/setup-uv@v8.3.2 # 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@v6
+      - uses: actions/checkout@v7
        with:
          lfs: true
          persist-credentials: false
      - name: Download updated lockfile
-        uses: actions/download-artifact@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/download-artifact@v8 # zizmor: ignore[unpinned-uses]
        with:
          name: uv-lock
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v4.4.0 # 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@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v7 # 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@v6
+      - uses: actions/checkout@v7
        with:
          persist-credentials: false
      - name: Download updated lockfile
-        uses: actions/download-artifact@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/download-artifact@v8 # 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@v6
+      - uses: actions/labeler@v7
        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@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          persist-credentials: false
      - name: Set up Python
-        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6
+        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v6
        with:
          python-version: '3.12'
@@ -38,12 +38,12 @@ jobs:
    steps:
      - name: Checkout code
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          persist-credentials: false
      - name: Set up Python
-        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6
+        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # 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@ed0c53931b1dc9bd32cbe73a98c7f6766f8a527e  # v1.13.0
+        uses: pypa/gh-action-pypi-publish@ba38be9e461d3875417946c167d0b5f3d385a247  # v1.14.1
        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@ed0c53931b1dc9bd32cbe73a98c7f6766f8a527e  # v1.13.0
+        uses: pypa/gh-action-pypi-publish@ba38be9e461d3875417946c167d0b5f3d385a247  # v1.14.1
        with:
          verbose: true
          print-hash: true
@@ -127,7 +127,7 @@ jobs:
    env:
      MUJOCO_GL: egl
    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        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@d0cc045d04ccac9d8b7881df0226f9e82c39688e  # v6
+        uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990  # v8.3.2
        with:
          enable-cache: true # zizmor: ignore[cache-poisoning]
          version: ${{ env.UV_VERSION }}
@@ -43,12 +43,12 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout code
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
        with:
          fetch-depth: 0
          persist-credentials: false
      - name: Secret Scanning
-        uses: trufflesecurity/trufflehog@eafb8c5f6a06175141c27f17bcc17941853d0047  # v3.90.0
+        uses: trufflesecurity/trufflehog@27b0417c16317ca9a472a9a8092acce143b49c55  # v3.95.9
        with:
          extra_args: --only-verified
@@ -61,15 +61,19 @@ Full details in [`docs/source/so101.mdx`](./docs/source/so101.mdx) and [`docs/so
 **4.1 Install**
 ```bash
-pip install 'lerobot[feetech]'              # SO-100/SO-101 motor stack
+# uv (recommended — see AGENTS.md and CLAUDE.md)
-# pip install 'lerobot[all]'                # everything
+uv sync --locked --extra feetech          # SO-100/SO-101 motor stack
-# pip install 'lerobot[aloha,pusht]'        # specific features
+# uv sync --locked --extra all            # everything
-# pip install 'lerobot[smolvla]'            # add SmolVLA deps
+# uv sync --locked --extra smolvla        # add SmolVLA deps
 git lfs install && git lfs pull
 hf auth login                               # required to push datasets/policies
 ```
-Contributors can alternatively use `uv sync --locked --extra feetech` (see `AGENTS.md`).
+# pip (alternative, e.g. when not working from source)
 # pip install 'lerobot[feetech]'
 # pip install 'lerobot[all]'
 # pip install 'lerobot[smolvla]'
 git lfs install && git lfs pull
 hf auth login                             # required to push datasets/policies
 ```
 **4.2 Find USB ports** — run once per arm, unplug when prompted.
@@ -68,17 +68,16 @@ ENV HOME=/home/user_lerobot \
 # issues with MuJoCo and OpenGL drivers.
 RUN uv venv --python python${PYTHON_VERSION}
-# Install Python dependencies for caching
+# Install third-party dependencies separately for layer caching
 COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
-COPY --chown=user_lerobot:user_lerobot src/ src/
+RUN uv sync --locked --extra all --no-install-project --no-cache
 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 rest of the application source code
+# Copy the application source code and install the local project
 # 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"]
@@ -60,15 +60,14 @@ ENV HOME=/home/user_lerobot \
 # run other Python projects in the same container without dependency conflicts.
 RUN uv venv
-# Install Python dependencies for caching
+# Install third-party dependencies separately for layer caching
 COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
-COPY --chown=user_lerobot:user_lerobot src/ src/
+RUN uv sync --locked --extra all --no-install-project --no-cache
-RUN uv sync --locked --extra all --no-cache
+# Copy the application code and install the local project
 # 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"]
@@ -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](NOTE:addlinktoLOC))
+- `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))
 - `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 a 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 an 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 to first try and and replay an episode of the dataset your policy was trained on using the section above.
+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 on your inference script (shown here in the `record.py` script):
+Then, add these same transformations to your inference script (shown here in the `record.py` script):
 ```diff
 action_values = predict_action(
@@ -136,6 +136,10 @@ 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
 )
@@ -154,6 +158,15 @@ 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>
@@ -88,20 +88,6 @@ 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(
@@ -116,6 +102,7 @@ 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**
@@ -145,7 +132,7 @@ 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**
@@ -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 on Upgrade tab**:
+2. **Click the Upgrade tab**:
-3. **Click on Online button**:
+3. **Click the Online button**:
-   - If an potential firmware update is found, it will be displayed in the box
+   - If a potential firmware update is found, it will be displayed in the box
-4. **Click on Upgrade button**:
+4. **Click the Upgrade button**:
   - The update progress will be displayed
 ## Step 6: Verify Update
@@ -59,6 +59,7 @@ 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/settings).
+This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data).
 ```bash
 lerobot-record \
@@ -98,7 +98,7 @@ The teleoperate command will automatically:
 ## Cameras
-To add cameras to your setup, follow this [Guide](./cameras#setup-cameras).
+To add cameras to your setup, follow this [Guide](./cameras).
 ## Teleoperate with cameras
@@ -1,23 +1,177 @@
 # LeRobot
 <div class="flex justify-center">
  <a target="_blank" href="https://huggingface.co/lerobot">
    <img
-      alt="HuggingFace Expert Acceleration Program"
+      alt="LeRobot, Hugging Face Robotics Library"
      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 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.
+🤗 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 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.
+🤗 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 already provides a set of pretrained models, datasets with human collected demonstrations, and simulated environments so that everyone can get started.
+The goal: lower the barrier to entry for robotics, so that everyone can contribute to, and benefit from, shared datasets and pretrained models.
-🤗 LeRobot hosts pretrained models and datasets on the LeRobot HuggingFace page.
+<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>
-Join the LeRobot community on [Discord](https://discord.gg/s3KuuzsPFb)
+<div align="center">
  <img src="../../media/readme/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.
@@ -149,13 +149,14 @@ lerobot-rollout \
 Foot pedal input is also supported via `--strategy.input_device=pedal`. Configure pedal codes with `--strategy.pedal.*` flags.
-| Flag                                 | Description                                             |
+| Flag                                 | Description                                                                                                                                                    |
-| ------------------------------------ | ------------------------------------------------------- |
+| ------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `--strategy.num_episodes`            | Number of correction episodes to record (default: 10)   |
+| `--strategy.num_episodes`            | Number of correction episodes to record (default: 10)                                                                                                          |
-| `--strategy.record_autonomous`       | Record autonomous frames too (default: false)           |
+| `--strategy.record_autonomous`       | Record autonomous frames too (default: false)                                                                                                                  |
-| `--strategy.upload_every_n_episodes` | Push to Hub every N episodes (default: 5)               |
+| `--strategy.upload_every_n_episodes` | Push to Hub every N episodes (default: 5)                                                                                                                      |
-| `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard) |
+| `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard)                                                                                                        |
-| `--teleop.type`                      | **Required.** Teleoperator type                         |
+| `--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                                                                                                                                |
 ### Episodic (`--strategy.type=episodic`)
@@ -186,14 +187,15 @@ 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                                                                |
+| Flag                                            | Description                                                                                                                                                |
-| ----------------------------------------------- | -------------------------------------------------------------------------- |
+| ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `--dataset.num_episodes`                        | Number of episodes to record                                               |
+| `--dataset.num_episodes`                        | Number of episodes to record                                                                                                                               |
-| `--dataset.episode_time_s`                      | Duration of each recording episode in seconds                              |
+| `--dataset.episode_time_s`                      | Duration of each recording episode in seconds                                                                                                              |
-| `--dataset.reset_time_s`                        | Duration of the reset phase between episodes 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                    |
+| `--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.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_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 |
 ---
@@ -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/so101_follower/so101_follower.py)
+For a good example of how it can be used, you can have a look at our own [SO101 follower implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/so_follower/so_follower.py)
 Use these if compatible. Otherwise, you'll need to find or write a Python interface (not covered in this tutorial):
@@ -51,7 +51,7 @@ In addition to these instructions, you need to install the Feetech SDK & ZeroMQ
 pip install -e ".[lekiwi]"
 ```
-Great :hugs:! You are now done installing LeRobot, and we can begin assembling the SO100/SO101 arms and the mobile base :robot:.
+Great 🤗! You are now done installing LeRobot, and we can begin assembling the SO100/SO101 arms and the mobile base 🤖.
 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
@@ -114,38 +114,58 @@ LIBERO supports two control modes — `relative` (default) and `absolute`. Diffe
 ### 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.
+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>
 ## Training
 ### Dataset
-We provide a preprocessed LIBERO dataset fully compatible with LeRobot:
+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:
- [HuggingFaceVLA/libero](https://huggingface.co/datasets/HuggingFaceVLA/libero)
+|                           | [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.
 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 \
-  --dataset.repo_id=HuggingFaceVLA/libero \
+  --policy.push_to_hub=false \
-  --env.type=libero \
+  --dataset.repo_id=lerobot/libero \
-  --env.task=libero_10 \
+  --dataset.video_backend=torchcodec \
-  --output_dir=./outputs/ \
+  --output_dir=./outputs/libero_smolvla \
  --steps=100000 \
-  --batch_size=4 \
+  --batch_size=64
  --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).
@@ -36,6 +36,12 @@ 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:
@@ -46,27 +52,106 @@ policy.type=pi05
 ## Training
-### Training Command Example
+### Quickstart on LIBERO
-Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:
+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=your_dataset \
+    --dataset.repo_id=lerobot/libero \
    --policy.type=pi05 \
-    --output_dir=./outputs/pi05_training \
+    --policy.pretrained_path=lerobot/pi05_libero_base \
-    --job_name=pi05_training \
+    --policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}' \
-    --policy.repo_id=your_repo_id \
+    --policy.n_action_steps=10 \
-    --policy.pretrained_path=lerobot/pi05_base \
+    --policy.empty_cameras=1 \
    --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.gradient_checkpointing=true \
    --policy.dtype=bfloat16 \
    --policy.device=cuda \
-    --batch_size=32
+    --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.
 ```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=true \
    --policy.train_expert_only=true \
    --policy.gradient_checkpointing=true \
    --policy.dtype=bfloat16 \
    --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
 ```
 ### Key Training Parameters
@@ -74,10 +159,24 @@ 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=32`**: Batch size for training, adapt this based on your GPU memory
+- **`--batch_size=64`**: 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](https://huggingface.co/lerobot/pi05_libero) (specifically trained on the Libero dataset)
+  - [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`.
 ### Training Parameters Explained
@@ -88,15 +187,6 @@ lerobot-train \
 **💡 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 40kk 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 40k 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 handle different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.
+Each of these pipelines handles 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 a observation dict.
+- `transition_to_observation`: transforms the pipeline transition to an observation dict.
-Checkout [src/lerobot/processor/converters.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/converters.py) for more details.
+Check out [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 and example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
+Below is an example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
 ```python
    def transform_features(
@@ -82,6 +82,8 @@ 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
@@ -35,14 +35,11 @@ pip install --override <(printf 'gymnasium==0.29.1\nnumpy==1.26.4\n') \
 ### Docker (recommended)
 ```bash
-# Build base image first (from repo root)
+# Build the RoboMME evaluation image from the repo root
-docker build -f docker/Dockerfile.eval-base -t lerobot-eval-base .
+docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-benchmark-robomme .
 # Build RoboMME eval image (applies gymnasium + numpy pin overrides)
 docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-robomme .
 ```
-The `docker/Dockerfile.benchmark.robomme` image overrides `gymnasium==0.29.1` and `numpy==1.26.4` after lerobot's install. Both versions are runtime-safe for lerobot's actual API usage.
+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.
 ## 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 values should be calculated based on the inference latency of the policy
+inference_delay = 4  # How many steps of inference latency, this value 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 a 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 an 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).
+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).
 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 motor 5 using 1 M2x6mm screw.
+- Attach the handle to the leader holder 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.
@@ -11,9 +11,10 @@ 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. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
+6. **Modify Tasks** - Change the natural-language task descriptions associated with episodes
-7. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
+7. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
-8. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
+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.
 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`.
@@ -50,11 +51,11 @@ lerobot-edit-dataset \
 Divide a dataset into multiple subsets.
 ```bash
-# Split by fractions (e.g. 80% train, 20% test, 20% val)
+# Split by fractions (e.g. 60% train, 20% val, 20% test)
 lerobot-edit-dataset \
    --repo_id lerobot/pusht \
    --operation.type split \
-    --operation.splits '{"train": 0.8, "test": 0.2, "val": 0.2}'
+    --operation.splits '{"train": 0.6, "val": 0.2, "test": 0.2}'
 # Split by specific episode indices
 lerobot-edit-dataset \
@@ -89,6 +90,53 @@ 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                                                                                                                                                                            |
+| 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`.                                                       |
+| `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.                                                                                                       |
+| `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`.                                                                                                                  |
+| `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. |
+| `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`.                                                               |
+| `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.                                   |
+| `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.                                                                                                                             |
+| `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.                                                     |
+| `extra_options` | `dict`           | `{}`          | Extra FFmpeg or codec specific options merged after the structured fields above. Cannot override keys already set by those fields.                                                                                      |
 ---
@@ -44,6 +44,7 @@ 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,
@@ -56,7 +57,6 @@ 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.types import RobotAction
+from lerobot.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.types import RobotAction
+from lerobot.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
@@ -21,6 +21,7 @@ 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
@@ -38,7 +39,6 @@ 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
@@ -16,6 +16,7 @@
 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,
@@ -36,7 +37,6 @@ 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
@@ -17,6 +17,7 @@
 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,
@@ -27,7 +28,6 @@ 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
@@ -27,6 +27,7 @@ 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,
@@ -43,7 +44,6 @@ 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
@@ -15,6 +15,7 @@
 import time
 from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -31,7 +32,6 @@ 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 deinitalized after each step to free memory. Large models
+        Note: Policies are deinitialized 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.
@@ -21,6 +21,7 @@ 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
@@ -38,7 +39,6 @@ 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
@@ -17,6 +17,7 @@
 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,
@@ -33,7 +34,6 @@ 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
@@ -18,6 +18,7 @@
 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,
@@ -28,7 +29,6 @@ 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
@@ -25,6 +25,7 @@ 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,
@@ -41,7 +42,6 @@ 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
@@ -16,6 +16,7 @@
 import time
 from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -30,7 +31,6 @@ 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
@@ -67,8 +67,8 @@ dependencies = [
    "einops>=0.8.0,<0.9.0",
    # Config & Hub
-    "draccus==0.10.0", # TODO: Relax version constraint
+    "draccus>=0.11.6,<0.12.0",
-    "huggingface-hub>=1.0.0,<2.0.0",
+    "huggingface-hub>=1.6.0,<2.0.0",
    "requests>=2.32.0,<3.0.0",
    # Environments
@@ -95,7 +95,7 @@ dependencies = [
 # ── Feature-scoped extras ──────────────────────────────────
 dataset = [
-    "datasets>=4.7.0,<5.0.0",
+    "datasets>=4.8.0,<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]",
@@ -494,6 +494,19 @@ 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.*"
@@ -38,6 +38,7 @@ 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 (
@@ -45,7 +46,6 @@ 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
@@ -120,14 +120,22 @@ class OpenCVCamera(Camera):
        self.rotation: int | None = get_cv2_rotation(config.rotation)
        self.backend: int = config.backend
-        if self.height and self.width:
+        self.capture_width: int | None = None
-            self.capture_width, self.capture_height = self.width, self.height
+        self.capture_height: int | None = None
-            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+        self._reset_connection_settings()
                self.capture_width, self.capture_height = self.height, self.width
    def __str__(self) -> str:
        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."""
@@ -164,17 +172,25 @@ class OpenCVCamera(Camera):
                f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
            )
-        self._configure_capture_settings()
+        try:
-        self._start_read_thread()
+            self._configure_capture_settings()
            self._start_read_thread()
-        if warmup and self.warmup_s > 0:
+            if warmup and self.warmup_s > 0:
-            start_time = time.time()
+                start_time = time.time()
-            while time.time() - start_time < self.warmup_s:
+                while time.time() - start_time < self.warmup_s:
-                self.async_read(timeout_ms=self.warmup_s * 1000)
+                    self.async_read(timeout_ms=self.warmup_s * 1000)
-                time.sleep(0.1)
+                    time.sleep(0.1)
-            with self.frame_lock:
+                with self.frame_lock:
-                if self.latest_frame is None:
+                    if self.latest_frame is None:
-                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
+                        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
        logger.info(f"{self} connected.")
@@ -312,32 +328,36 @@ class OpenCVCamera(Camera):
        for target in targets_to_scan:
            camera = cv2.VideoCapture(target)
-            if camera.isOpened():
+            try:
-                default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
+                if camera.isOpened():
-                default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                    default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
-                default_fps = camera.get(cv2.CAP_PROP_FPS)
+                    default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
-                default_format = camera.get(cv2.CAP_PROP_FORMAT)
+                    default_fps = camera.get(cv2.CAP_PROP_FPS)
                    default_format = camera.get(cv2.CAP_PROP_FORMAT)
-                # Get FOURCC code and convert to string
+                    # Get FOURCC code and convert to string
-                default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
+                    default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
-                default_fourcc_code_int = int(default_fourcc_code)
+                    default_fourcc_code_int = int(default_fourcc_code)
-                default_fourcc = "".join([chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])
+                    default_fourcc = "".join(
                        [chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)]
                    )
-                camera_info = {
+                    camera_info = {
-                    "name": f"OpenCV Camera @ {target}",
+                        "name": f"OpenCV Camera @ {target}",
-                    "type": "OpenCV",
+                        "type": "OpenCV",
-                    "id": target,
+                        "id": target,
-                    "backend_api": camera.getBackendName(),
+                        "backend_api": camera.getBackendName(),
-                    "default_stream_profile": {
+                        "default_stream_profile": {
-                        "format": default_format,
+                            "format": default_format,
-                        "fourcc": default_fourcc,
+                            "fourcc": default_fourcc,
-                        "width": default_width,
+                            "width": default_width,
-                        "height": default_height,
+                            "height": default_height,
-                        "fps": default_fps,
+                            "fps": default_fps,
-                    },
+                        },
-                }
+                    }
-                found_cameras_info.append(camera_info)
+                    found_cameras_info.append(camera_info)
            finally:
                camera.release()
        return found_cameras_info
@@ -496,6 +516,26 @@ 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]:
        """
@@ -586,16 +626,6 @@ class OpenCVCamera(Camera):
        if not self.is_connected and self.thread is None:
            raise DeviceNotConnectedError(f"{self} not connected.")
-        if self.thread is not None:
+        self._cleanup_resources()
            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.")
@@ -121,6 +121,9 @@ 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:
@@ -131,6 +134,9 @@ 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
@@ -145,14 +151,23 @@ class RealSenseCamera(Camera):
        self.rotation: int | None = get_cv2_rotation(config.rotation)
-        if self.height and self.width:
+        self.capture_width: int | None = None
-            self.capture_width, self.capture_height = self.width, self.height
+        self.capture_height: int | None = None
-            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+        self._reset_connection_settings()
                self.capture_width, self.capture_height = self.height, self.width
    def __str__(self) -> str:
        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."""
@@ -172,7 +187,8 @@ class RealSenseCamera(Camera):
        Raises:
            DeviceAlreadyConnectedError: If the camera is already connected.
-            ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
+            ValueError: If the configuration is invalid, a requested sensor option is unsupported,
                or a requested sensor value is invalid.
            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.
        """
@@ -190,22 +206,31 @@ class RealSenseCamera(Camera):
                f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
            ) from e
-        self._configure_capture_settings()
+        try:
-        self._start_read_thread()
+            self._configure_capture_settings()
            self._configure_sensor_options()
            self._start_read_thread()
-        # 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.
+            # 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)
+            self.warmup_s = max(self.warmup_s, 1)
-        warmup_read = self.async_read if self.use_rgb else self.async_read_depth
+            warmup_read = self.async_read if self.use_rgb else self.async_read_depth
-        start_time = time.time()
+            start_time = time.time()
-        while time.time() - start_time < self.warmup_s:
+            while time.time() - start_time < self.warmup_s:
-            warmup_read(timeout_ms=self.warmup_s * 1000)
+                warmup_read(timeout_ms=self.warmup_s * 1000)
-            time.sleep(0.1)
+                time.sleep(0.1)
-        with self.frame_lock:
+            with self.frame_lock:
-            if (self.use_rgb and self.latest_color_frame is None) or (
+                if (self.use_rgb and self.latest_color_frame is None) or (
-                self.use_depth and self.latest_depth_frame is None
+                    self.use_depth and self.latest_depth_frame is None
-            ):
+                ):
-                raise ConnectionError(f"{self} failed to capture frames during warmup.")
+                    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
        logger.info(f"{self} connected.")
@@ -339,6 +364,114 @@ 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]:
        """
@@ -541,6 +674,27 @@ 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():
@@ -684,18 +838,5 @@ class RealSenseCamera(Camera):
                f"Attempted to disconnect {self}, but it appears already disconnected."
            )
-        if self.thread is not None:
+        self._cleanup_resources()
            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.")
@@ -46,6 +46,17 @@ class RealSenseCameraConfig(CameraConfig):
        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)
        exposure: 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. Defaults to None (leave unchanged).
        gain: 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. Defaults to None (leave unchanged).
        white_balance: 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. Defaults to None
            (leave unchanged).
    Note:
        - Either name or serial_number must be specified.
@@ -61,6 +72,9 @@ 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:
        self.color_mode = ColorMode(self.color_mode)
@@ -69,6 +83,18 @@ class RealSenseCameraConfig(CameraConfig):
        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(
@@ -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(
@@ -52,6 +52,17 @@ def get_step_checkpoint_dir(output_dir: Path, total_steps: int, step: int) -> Pa
    return output_dir / CHECKPOINTS_DIR / step_identifier
 def should_save_checkpoint(step: int, save_freq: int, total_steps: int) -> bool:
    """Whether a checkpoint should be saved at ``step``.
    A checkpoint is saved every ``save_freq`` steps and always after the final step. A
    non-positive ``save_freq`` disables periodic saving (only the final checkpoint is
    written), mirroring how ``log_freq``/``eval_freq`` treat non-positive values and
    avoiding a ``ZeroDivisionError`` from ``step % 0``.
    """
    return (save_freq > 0 and step % save_freq == 0) or step == total_steps
 def save_training_step(
    step: int, save_dir: Path, num_processes: int | None = None, batch_size: int | None = None
 ) -> None:
@@ -71,13 +71,19 @@ class DatasetRecordConfig:
    # Number of threads per encoder instance. None = auto (codec default).
    # Lower values reduce CPU usage, maps to 'lp' (via svtav1-params) for libsvtav1 and 'threads' for h264/hevc..
    encoder_threads: int | None = None
    # Skip appending the date-time tag to repo_id, keeping the user-provided name as-is
    # (e.g. self-managed versioned names intended for a later `lerobot-edit-dataset merge`).
    no_stamp: bool = False
    def stamp_repo_id(self) -> None:
        """Append a date-time tag to ``repo_id`` so each recording session gets a unique name.
        Must be called explicitly at dataset *creation* time — not on resume,
        where the existing ``repo_id`` (already stamped) must be preserved.
        No-op when ``no_stamp`` is set, preserving a user-managed ``repo_id``.
        """
        if self.no_stamp:
            return
        if self.repo_id:
            timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
            self.repo_id = f"{self.repo_id}_{timestamp}"
@@ -27,6 +27,13 @@ from typing import Any, TypeVar, cast
 import draccus
 import yaml  # type: ignore[import-untyped]
 from draccus.help_formatter import SimpleHelpFormatter
 from draccus.utils import DecodingError
 from draccus.wrappers import DataclassWrapper
 from draccus.wrappers.choice_wrapper import ChoiceWrapper, UnionWrapper
 from draccus.wrappers.field_wrapper import FieldWrapper
 from draccus.wrappers.suppressing_argparse import SuppressingArgumentParser
 from draccus.wrappers.wrapper import AggregateWrapper, Wrapper
 from lerobot.utils.utils import has_method
@@ -72,11 +79,18 @@ def get_cli_overrides(field_name: str, args: Sequence[str] | None = None) -> lis
        args = sys.argv[1:]
    attr_level_args = []
    detect_string = f"--{field_name}."
-    exclude_strings = (f"--{field_name}.{draccus.CHOICE_TYPE_KEY}=", f"--{field_name}.{PATH_KEY}=")
+    excluded_names = (draccus.CHOICE_TYPE_KEY, PATH_KEY)
-    for arg in args:
+    for index, arg in enumerate(args):
-        if arg.startswith(detect_string) and not arg.startswith(exclude_strings):
+        if not arg.startswith(detect_string):
-            denested_arg = f"--{arg.removeprefix(detect_string)}"
+            continue
-            attr_level_args.append(denested_arg)
+
        denested_arg = arg.removeprefix(detect_string)
        if denested_arg.split("=", maxsplit=1)[0] in excluded_names:
            continue
        attr_level_args.append(f"--{denested_arg}")
        if "=" not in arg and index + 1 < len(args) and not args[index + 1].startswith("--"):
            attr_level_args.append(args[index + 1])
    return attr_level_args
@@ -84,10 +98,12 @@ def get_cli_overrides(field_name: str, args: Sequence[str] | None = None) -> lis
 def parse_arg(arg_name: str, args: Sequence[str] | None = None) -> str | None:
    if args is None:
        args = sys.argv[1:]
-    prefix = f"--{arg_name}="
+    option = f"--{arg_name}"
-    for arg in args:
+    for index, arg in enumerate(args):
-        if arg.startswith(prefix):
+        if arg.startswith(f"{option}="):
-            return arg[len(prefix) :]
+            return arg.removeprefix(f"{option}=")
        if arg == option and index + 1 < len(args) and not args[index + 1].startswith("--"):
            return args[index + 1]
    return None
@@ -95,7 +111,7 @@ def parse_plugin_args(plugin_arg_suffix: str, args: Sequence[str]) -> dict[str,
    """Parse plugin-related arguments from command-line arguments.
    This function extracts arguments from command-line arguments that match a specified suffix pattern.
-    It processes arguments in the format '--key=value' and returns them as a dictionary.
+    It accepts arguments in the formats '--key=value' and '--key value' and returns them as a dictionary.
    Args:
        plugin_arg_suffix (str): The suffix to identify plugin-related arguments.
@@ -112,13 +128,18 @@ def parse_plugin_args(plugin_arg_suffix: str, args: Sequence[str]) -> dict[str,
        {'env.discover_packages_path': 'my_package'}
    """
    plugin_args = {}
-    for arg in args:
+    for index, arg in enumerate(args):
-        if "=" in arg and plugin_arg_suffix in arg:
+        if not arg.startswith("--"):
-            key, value = arg.split("=", 1)
+            continue
-            # Remove leading '--' if present
+
-            if key.startswith("--"):
+        key, separator, value = arg[2:].partition("=")
-                key = key[2:]
+        if plugin_arg_suffix not in key:
-            plugin_args[key] = value
+            continue
        if not separator:
            if index + 1 >= len(args) or args[index + 1].startswith("--"):
                continue
            value = args[index + 1]
        plugin_args[key] = value
    return plugin_args
@@ -185,10 +206,82 @@ def get_type_arg(field_name: str, args: Sequence[str] | None = None) -> str | No
    return parse_arg(f"{field_name}.{draccus.CHOICE_TYPE_KEY}", args)
 def _register_scoped_actions(
    wrapper: Wrapper, parser: SuppressingArgumentParser, cli_args: Sequence[str]
 ) -> None:
    """Like draccus's own Wrapper.register_actions, but for a ChoiceType field only recurses into
    the already-selected subclass (per CLI `.type` args), instead of every registered choice.
    This mirrors draccus 0.11.x's internal wrapper traversal because its public parser eagerly registers
    every choice before parsing the command line. Keep this in sync when updating draccus.
    """
    if isinstance(wrapper, ChoiceWrapper):
        group = parser.add_argument_group(title=wrapper.title, description=wrapper.description)
        children = wrapper._children
        arg_name = f"{wrapper.dest}.{draccus.CHOICE_TYPE_KEY}" if wrapper.dest else draccus.CHOICE_TYPE_KEY
        group.add_argument(
            f"--{arg_name}",
            choices=list(children.keys()),
            help=f"Which type of {wrapper.title} to use",
            required=wrapper.required,
        )
        selected = get_type_arg(wrapper.dest, cli_args) if wrapper.dest else None
        if selected in children:
            _register_scoped_actions(children[selected], parser, cli_args)
    elif isinstance(wrapper, DataclassWrapper):
        group = parser.add_argument_group(title=wrapper.title, description=wrapper.description)
        for child in wrapper._children:
            if isinstance(child, AggregateWrapper):
                parser.add_argument(
                    f"--{child.name}", type=str, required=False, help=f"Config file for {child.name}"
                )
                _register_scoped_actions(child, parser, cli_args)
            elif isinstance(child, FieldWrapper):
                child.add_action(group)
    elif isinstance(wrapper, UnionWrapper):
        group = parser.add_argument_group(title=wrapper.title, description=wrapper.description)
        has_field_wrapper = False
        for child in wrapper._children:
            if isinstance(child, (DataclassWrapper, ChoiceWrapper)):
                _register_scoped_actions(child, parser, cli_args)
            elif isinstance(child, FieldWrapper):
                has_field_wrapper = True
        if has_field_wrapper:
            group.add_argument(f"--{wrapper.dest}", required=False)
    else:
        wrapper.register_actions(parser)
 def print_scoped_help(config_class: type, cli_args: Sequence[str]) -> None:
    """Prints --help output scoped to the choices already resolved on the CLI (e.g. --env.type=pusht),
    instead of draccus's default of expanding every registered subclass of every ChoiceType field."""
    parser = SuppressingArgumentParser(formatter_class=SimpleHelpFormatter)
    parser.add_argument(
        f"--{draccus.utils.CONFIG_ARG}", type=str, help="Path for a config file to parse with draccus"
    )
    _register_scoped_actions(DataclassWrapper(config_class), parser, cli_args)
    parser.print_help()
 def filter_arg(field_to_filter: str, args: Sequence[str] | None = None) -> list[str]:
    if args is None:
        return []
-    return [arg for arg in args if not arg.startswith(f"--{field_to_filter}=")]
+    option = f"--{field_to_filter}"
    filtered_args = []
    index = 0
    while index < len(args):
        arg = args[index]
        if arg == option:
            index += 1
            if index < len(args) and not args[index].startswith("--"):
                index += 1
            continue
        if arg.startswith(f"{option}="):
            index += 1
            continue
        filtered_args.append(arg)
        index += 1
    return filtered_args
 def filter_path_args(fields_to_filter: str | list[str], args: Sequence[str] | None = None) -> list[str]:
@@ -220,7 +313,23 @@ def filter_path_args(fields_to_filter: str | list[str], args: Sequence[str] | No
                    argument=None,
                    message=f"Cannot specify both --{field}.{PATH_KEY} and --{field}.{draccus.CHOICE_TYPE_KEY}",
                )
-            filtered_args = [arg for arg in filtered_args if not arg.startswith(f"--{field}.")]
+            option_prefix = f"--{field}."
            retained_args = []
            index = 0
            while index < len(filtered_args):
                arg = filtered_args[index]
                if arg.startswith(option_prefix):
                    index += 1
                    if (
                        "=" not in arg
                        and index < len(filtered_args)
                        and not filtered_args[index].startswith("--")
                    ):
                        index += 1
                    continue
                retained_args.append(arg)
                index += 1
            filtered_args = retained_args
    return filtered_args
@@ -299,6 +408,9 @@ def wrap(config_path: Path | None = None) -> Callable[[F], F]:
                        # add the relevant CLI arg to the error message
                        raise PluginLoadError(f"{e}\nFailed plugin CLI Arg: {plugin_cli_arg}") from e
                    cli_args = filter_arg(plugin_cli_arg, cli_args)
                if "--help" in cli_args or "-h" in cli_args:
                    print_scoped_help(argtype, cli_args)
                    sys.exit(0)
                config_path_cli = parse_arg("config_path", cli_args)
                if has_method(argtype, "__get_path_fields__"):
                    path_fields = argtype.__get_path_fields__()
@@ -306,17 +418,21 @@ def wrap(config_path: Path | None = None) -> Callable[[F], F]:
                    # Also extract path fields from the YAML/JSON config file
                    if config_path_cli:
                        config_path_cli = extract_path_fields_from_config(config_path_cli, path_fields)
-                if has_method(argtype, "from_pretrained") and config_path_cli:
+                try:
-                    cli_args = filter_arg("config_path", cli_args)
+                    if has_method(argtype, "from_pretrained") and config_path_cli:
                    cfg = argtype.from_pretrained(config_path_cli, cli_args=cli_args)
                else:
                    if config_path_cli:
                        cli_args = filter_arg("config_path", cli_args)
-                    cfg = draccus.parse(
+                        cfg = argtype.from_pretrained(config_path_cli, cli_args=cli_args)
-                        config_class=argtype,
+                    else:
-                        config_path=config_path_cli or config_path,
+                        if config_path_cli:
-                        args=cli_args,
+                            cli_args = filter_arg("config_path", cli_args)
-                    )
+                        cfg = draccus.parse(
                            config_class=argtype,
                            config_path=config_path_cli or config_path,
                            args=cli_args,
                        )
                except DecodingError as e:
                    print(f"error: {e}", file=sys.stderr)
                    sys.exit(1)
            response = fn(cfg, *args, **kwargs)
            return response
@@ -163,8 +163,10 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
        return None
    def _save_pretrained(self, save_directory: Path) -> None:
-        with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
+        # Encode against the base class so draccus includes the choice "type" key,
-            draccus.dump(self, f, indent=4)
+        # which `from_pretrained` needs to resolve the concrete subclass.
        with open(save_directory / CONFIG_NAME, "w") as f:
            json.dump(draccus.encode(self, PreTrainedConfig), f, indent=4)
    @classmethod
    def from_pretrained(
@@ -103,8 +103,10 @@ class RewardModelConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
        pass
    def _save_pretrained(self, save_directory: Path) -> None:
-        with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
+        # Encode against the base class so draccus includes the choice "type" key,
-            draccus.dump(self, f, indent=4)
+        # which `from_pretrained` needs to resolve the concrete subclass.
        with open(save_directory / CONFIG_NAME, "w") as f:
            json.dump(draccus.encode(self, RewardModelConfig), f, indent=4)
    @classmethod
    def from_pretrained(
@@ -119,6 +119,7 @@ class TrainPipelineConfig(HubMixin):
    tolerance_s: float = 1e-4
    save_checkpoint: bool = True
    # Checkpoint is saved every `save_freq` training iterations and after the last training step.
    # A non-positive value disables periodic saving, keeping only the final checkpoint.
    save_freq: int = 20_000
    use_policy_training_preset: bool = True
    optimizer: OptimizerConfig | None = None
@@ -194,7 +195,11 @@ class TrainPipelineConfig(HubMixin):
            )
        if Path(config_path).resolve().exists():
-            policy_dir = Path(config_path).parent
+            # `config_path` may point at the checkpoint's train_config.json or at its
            # pretrained_model/ directory (both documented above) — resolve either to
            # the pretrained_model/ directory.
            config_path_obj = Path(config_path)
            policy_dir = config_path_obj.parent if config_path_obj.is_file() else config_path_obj
            self.checkpoint_path = policy_dir.parent
        elif self.job.is_remote:
            return
@@ -19,6 +19,7 @@ import copy
 import logging
 import shutil
 from pathlib import Path
 from typing import Any, NotRequired, TypedDict
 import datasets
 import pandas as pd
@@ -49,8 +50,32 @@ from .utils import (
 )
 from .video_utils import concatenate_video_files, get_video_duration_in_s
 logger = logging.getLogger(__name__)
-def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> dict[str, dict]:
+type FeatureDict = dict[str, dict[str, Any]]
 type ChunkFile = tuple[int, int]
 class IndexState(TypedDict):
    chunk: int
    file: int
    src_to_dst: NotRequired[dict[ChunkFile, ChunkFile]]
 class VideoIndex(TypedDict):
    chunk: int
    file: int
    latest_duration: float
    episode_duration: float
    src_to_offset: NotRequired[dict[ChunkFile, float]]
    src_to_dst: NotRequired[dict[ChunkFile, ChunkFile]]
    dst_file_durations: NotRequired[dict[ChunkFile, float]]
 type VideoIndexState = dict[str, VideoIndex]
 def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> FeatureDict:
    """Create a merged video feature info dictionary for aggregation. The video encoder info is merged field-by-field: each key is kept only when every source agrees; otherwise that key is set to ``null`` (or ``{}`` for ``video.extra_options``) and a warning is logged.
    Args:
@@ -59,14 +84,14 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
    Returns:
        dict: A dictionary of merged video feature info.
    """
-    merged_info = copy.deepcopy(all_metadata[0].features)
+    merged_info: FeatureDict = copy.deepcopy(all_metadata[0].features)
    video_keys = [k for k in merged_info if merged_info[k].get("dtype") == "video"]
    for vk in video_keys:
        video_infos = [m.features.get(vk, {}).get("info") or {} for m in all_metadata]
        base_video_info = video_infos[0]
-        merged_encoder_info: dict = {}
+        merged_encoder_info: dict[str, Any] = {}
        fallback_keys: list[str] = []
        for info_key in VIDEO_ENCODER_INFO_KEYS:
            values = [info.get(info_key, None) for info in video_infos]
@@ -80,7 +105,7 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
                merged_encoder_info[info_key] = {} if info_key == "video.extra_options" else None
        if fallback_keys:
-            logging.warning(
+            logger.warning(
                f"Merging heterogeneous or incomplete video encoder metadata for feature {vk}. "
                f"Setting these keys to null: {fallback_keys}.",
            )
@@ -92,7 +117,7 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
    return merged_info
-def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
+def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]) -> tuple[int, str | None, FeatureDict]:
    """Validates that all dataset metadata have consistent properties.
    Ensures all datasets have the same fps, robot_type, and features to guarantee
@@ -129,7 +154,9 @@ def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
    return fps, robot_type, features
-def update_data_df(df, src_meta, dst_meta):
+def update_data_df(
    df: pd.DataFrame, src_meta: LeRobotDatasetMetadata, dst_meta: LeRobotDatasetMetadata
 ) -> pd.DataFrame:
    """Updates a data DataFrame with new indices and task mappings for aggregation.
    Adjusts episode indices, frame indices, and task indices to account for
@@ -154,12 +181,12 @@ def update_data_df(df, src_meta, dst_meta):
 def update_meta_data(
-    df,
+    df: pd.DataFrame,
-    dst_meta,
+    dst_meta: LeRobotDatasetMetadata,
-    meta_idx,
+    meta_idx: IndexState,
-    data_idx,
+    data_idx: IndexState,
-    videos_idx,
+    videos_idx: VideoIndexState,
-):
+) -> pd.DataFrame:
    """Updates metadata DataFrame with new chunk, file, and timestamp indices.
    Adjusts all indices and timestamps to account for previously aggregated
@@ -289,7 +316,7 @@ def aggregate_datasets(
    chunk_size: int | None = None,
    concatenate_videos: bool = True,
    concatenate_data: bool = True,
-):
+) -> None:
    """Aggregates multiple LeRobot datasets into a single unified dataset.
    This is the main function that orchestrates the aggregation process by:
@@ -309,7 +336,7 @@ def aggregate_datasets(
        concatenate_videos: When False, keep one mp4 per source file instead of packing into shards.
        concatenate_data: When False, keep one parquet per source file instead of packing into shards.
    """
-    logging.info("Start aggregate_datasets")
+    logger.info("Start aggregate_datasets")
    if data_files_size_in_mb is None:
        data_files_size_in_mb = DEFAULT_DATA_FILE_SIZE_IN_MB
@@ -341,15 +368,15 @@ def aggregate_datasets(
        video_files_size_in_mb=video_files_size_in_mb,
    )
-    logging.info("Find all tasks")
+    logger.info("Find all tasks")
    unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
    dst_meta.tasks = pd.DataFrame(
        {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
    )
-    meta_idx = {"chunk": 0, "file": 0}
+    meta_idx: IndexState = {"chunk": 0, "file": 0}
-    data_idx = {"chunk": 0, "file": 0}
+    data_idx: IndexState = {"chunk": 0, "file": 0}
-    videos_idx = {
+    videos_idx: VideoIndexState = {
        key: {"chunk": 0, "file": 0, "latest_duration": 0, "episode_duration": 0} for key in video_keys
    }
@@ -373,12 +400,17 @@ def aggregate_datasets(
        dst_meta.info.total_frames += src_meta.total_frames
    finalize_aggregation(dst_meta, all_metadata)
-    logging.info("Aggregation complete.")
+    logger.info("Aggregation complete.")
 def aggregate_videos(
-    src_meta, dst_meta, videos_idx, video_files_size_in_mb, chunk_size, concatenate_videos=True
+    src_meta: LeRobotDatasetMetadata,
-):
+    dst_meta: LeRobotDatasetMetadata,
    videos_idx: VideoIndexState,
    video_files_size_in_mb: float,
    chunk_size: int,
    concatenate_videos: bool = True,
 ) -> VideoIndexState:
    """Aggregates video chunks from a source dataset into the destination dataset.
    Handles video file concatenation and rotation based on file size limits.
@@ -406,15 +438,16 @@ def aggregate_videos(
            videos_idx[key]["dst_file_durations"] = {}
    for key, video_idx in videos_idx.items():
-        unique_chunk_file_pairs = {
+        unique_chunk_file_pairs: list[ChunkFile] = sorted(
-            (chunk, file)
+            {
-            for chunk, file in zip(
+                (chunk, file)
-                src_meta.episodes[f"videos/{key}/chunk_index"],
+                for chunk, file in zip(
-                src_meta.episodes[f"videos/{key}/file_index"],
+                    src_meta.episodes[f"videos/{key}/chunk_index"],
-                strict=False,
+                    src_meta.episodes[f"videos/{key}/file_index"],
-            )
+                    strict=False,
-        }
+                )
-        unique_chunk_file_pairs = sorted(unique_chunk_file_pairs)
+            }
        )
        chunk_idx = video_idx["chunk"]
        file_idx = video_idx["file"]
@@ -489,7 +522,14 @@ def aggregate_videos(
    return videos_idx
-def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_size, concatenate_data=True):
+def aggregate_data(
    src_meta: LeRobotDatasetMetadata,
    dst_meta: LeRobotDatasetMetadata,
    data_idx: IndexState,
    data_files_size_in_mb: float,
    chunk_size: int,
    concatenate_data: bool = True,
 ) -> IndexState:
    """Aggregates data chunks from a source dataset into the destination dataset.
    Reads source data files, updates indices to match the aggregated dataset,
@@ -510,14 +550,16 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
    Returns:
        dict: Updated data_idx with current chunk and file indices.
    """
-    unique_chunk_file_ids = {
+    unique_chunk_file_ids: list[ChunkFile] = sorted(
-        (c, f)
+        {
-        for c, f in zip(
+            (c, f)
-            src_meta.episodes["data/chunk_index"], src_meta.episodes["data/file_index"], strict=False
+            for c, f in zip(
-        )
+                src_meta.episodes["data/chunk_index"],
-    }
+                src_meta.episodes["data/file_index"],
-
+                strict=False,
-    unique_chunk_file_ids = sorted(unique_chunk_file_ids)
+            )
        }
    )
    contains_images = len(dst_meta.image_keys) > 0
    # retrieve features schema for proper image typing in parquet
@@ -525,7 +567,7 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
    # Track source to destination file mapping for metadata update
    # This is critical for handling datasets that are already results of a merge
-    src_to_dst: dict[tuple[int, int], tuple[int, int]] = {}
+    src_to_dst: dict[ChunkFile, ChunkFile] = {}
    for src_chunk_idx, src_file_idx in unique_chunk_file_ids:
        src_path = src_meta.root / DEFAULT_DATA_PATH.format(
@@ -564,7 +606,13 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
    return data_idx
-def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
+def aggregate_metadata(
    src_meta: LeRobotDatasetMetadata,
    dst_meta: LeRobotDatasetMetadata,
    meta_idx: IndexState,
    data_idx: IndexState,
    videos_idx: VideoIndexState,
 ) -> IndexState:
    """Aggregates metadata from a source dataset into the destination dataset.
    Reads source metadata files, updates all indices and timestamps,
@@ -580,16 +628,16 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
    Returns:
        dict: Updated meta_idx with current chunk and file indices.
    """
-    chunk_file_ids = {
+    chunk_file_ids: list[ChunkFile] = sorted(
-        (c, f)
+        {
-        for c, f in zip(
+            (c, f)
-            src_meta.episodes["meta/episodes/chunk_index"],
+            for c, f in zip(
-            src_meta.episodes["meta/episodes/file_index"],
+                src_meta.episodes["meta/episodes/chunk_index"],
-            strict=False,
+                src_meta.episodes["meta/episodes/file_index"],
-        )
+                strict=False,
-    }
+            )
-
+        }
-    chunk_file_ids = sorted(chunk_file_ids)
+    )
    for chunk_idx, file_idx in chunk_file_ids:
        src_path = src_meta.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
        df = pd.read_parquet(src_path)
@@ -622,16 +670,16 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
 def append_or_create_parquet_file(
    df: pd.DataFrame,
    src_path: Path,
-    idx: dict[str, int],
+    idx: IndexState,
    max_mb: float,
    chunk_size: int,
    default_path: str,
    contains_images: bool = False,
-    aggr_root: Path = None,
+    aggr_root: Path | None = None,
    hf_features: datasets.Features | None = None,
    concatenate: bool = True,
    one_row_group_per_episode: bool = False,
-) -> tuple[dict[str, int], tuple[int, int]]:
+) -> tuple[IndexState, ChunkFile]:
    """Appends data to an existing parquet file or creates a new one based on size constraints.
    Manages file rotation when size limits are exceeded to prevent individual files
@@ -654,7 +702,13 @@ def append_or_create_parquet_file(
    Returns:
        tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
               and (dst_chunk, dst_file) is the actual destination file the data was written to.
    Raises:
        ValueError: If aggr_root is not provided.
    """
    if aggr_root is None:
        raise ValueError("aggr_root must be provided.")
    dst_chunk, dst_file = idx["chunk"], idx["file"]
    dst_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
@@ -698,7 +752,9 @@ def append_or_create_parquet_file(
    return idx, (dst_chunk, dst_file)
-def finalize_aggregation(aggr_meta, all_metadata):
+def finalize_aggregation(
    aggr_meta: LeRobotDatasetMetadata, all_metadata: list[LeRobotDatasetMetadata]
 ) -> None:
    """Finalizes the dataset aggregation by writing summary files and statistics.
    Writes the tasks file, info file with total counts and splits, and
@@ -708,16 +764,16 @@ def finalize_aggregation(aggr_meta, all_metadata):
        aggr_meta: Aggregated dataset metadata.
        all_metadata: List of all source dataset metadata objects.
    """
-    logging.info("write tasks")
+    logger.info("write tasks")
    write_tasks(aggr_meta.tasks, aggr_meta.root)
-    logging.info("write info")
+    logger.info("write info")
    aggr_meta.info.total_tasks = len(aggr_meta.tasks)
    aggr_meta.info.total_episodes = sum(m.total_episodes for m in all_metadata)
    aggr_meta.info.total_frames = sum(m.total_frames for m in all_metadata)
    aggr_meta.info.splits = {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"}
    write_info(aggr_meta.info, aggr_meta.root)
-    logging.info("write stats")
+    logger.info("write stats")
    aggr_meta.stats = aggregate_stats([m.stats for m in all_metadata])
    write_stats(aggr_meta.stats, aggr_meta.root)
@@ -18,13 +18,15 @@ import logging
 from collections.abc import Callable, Iterable
 from copy import deepcopy
 from pathlib import Path
 from typing import Literal
 import numpy as np
 import packaging.version
 import pandas as pd
 import pyarrow as pa
 import pyarrow.parquet as pq
-from huggingface_hub import snapshot_download
+from huggingface_hub import snapshot_download, sync_bucket
 from huggingface_hub.utils import WeakFileLock
 from lerobot.configs import DEPTH_METER_UNIT, VideoEncoderConfig
 from lerobot.utils.constants import DEFAULT_FEATURES, HF_LEROBOT_HOME, HF_LEROBOT_HUB_CACHE
@@ -74,6 +76,7 @@ class LeRobotDatasetMetadata:
        force_cache_sync: bool = False,
        metadata_buffer_size: int = 10,
        *,
        repo_type: Literal["dataset", "bucket"] = "dataset",
        token: str | bool | None = None,
    ):
        """Load or download metadata for an existing LeRobot dataset.
@@ -96,36 +99,53 @@ class LeRobotDatasetMetadata:
                even when local files exist.
            metadata_buffer_size: Number of episode metadata records to buffer
                in memory before flushing to parquet.
            repo_type: Repository type: "dataset" (default) or "bucket" for an
                HF Storage Bucket streamed over hf://buckets/.
            token: Authentication token used for Hub requests. Pass a string
                token, ``True`` to require the locally stored token, ``False``
                to disable authentication, or ``None`` to use the Hugging Face
                Hub default.
        """
        if repo_type not in ("dataset", "bucket"):
            raise ValueError(f"repo_type must be 'dataset' or 'bucket', got {repo_type!r}")
        self.repo_id = repo_id
        self.repo_type = repo_type
        self.revision = revision if revision else CODEBASE_VERSION
        self._requested_root = Path(root) if root is not None else None
-        self.root = self._requested_root if self._requested_root is not None else HF_LEROBOT_HOME / repo_id
+        if self._requested_root is not None:
            self.root = self._requested_root
        elif self.repo_type == "bucket":
            self.root = HF_LEROBOT_HUB_CACHE / ("buckets--" + self.repo_id.replace("/", "--"))
        else:
            self.root = HF_LEROBOT_HOME / repo_id
        self._pq_writer = None
        self.latest_episode = None
        self._metadata_buffer: list[dict] = []
        self._metadata_buffer_size = metadata_buffer_size
        self._finalized = False
-        try:
+        metadata_lock = contextlib.nullcontext()
-            if force_cache_sync or (
+        if self.repo_type == "bucket":
-                self._requested_root is None and has_legacy_hub_download_metadata(self.root)
+            self.root.parent.mkdir(parents=True, exist_ok=True)
-            ):
+            metadata_lock = WeakFileLock(self.root.parent / f".{self.root.name}.lock")
                raise FileNotFoundError
            self._load_metadata()
        except (FileNotFoundError, NotADirectoryError):
            if is_valid_version(self.revision):
                if token is None:
                    self.revision = get_safe_version(self.repo_id, self.revision)
                else:
                    self.revision = get_safe_version(self.repo_id, self.revision, token=token)
-            self._pull_from_repo(allow_patterns="meta/", token=token)
+        with metadata_lock:
-            self._load_metadata()
+            try:
                if force_cache_sync or (
                    self._requested_root is None and has_legacy_hub_download_metadata(self.root)
                ):
                    raise FileNotFoundError
                self._load_metadata()
            except (FileNotFoundError, NotADirectoryError):
                if self.repo_type != "bucket" and is_valid_version(self.revision):
                    if token is None:
                        self.revision = get_safe_version(self.repo_id, self.revision)
                    else:
                        self.revision = get_safe_version(self.repo_id, self.revision, token=token)
                self._pull_from_repo(allow_patterns="meta/", token=token)
                self._load_metadata()
    def _flush_metadata_buffer(self) -> None:
        """Write all buffered episode metadata to parquet file."""
@@ -154,6 +174,12 @@ class LeRobotDatasetMetadata:
            self._pq_writer = pq.ParquetWriter(
                path, schema=table.schema, compression="snappy", use_dictionary=True
            )
        else:
            # Column order in `combined_dict` follows the source episode dict's insertion
            # order, which can differ between batches (e.g. episodes originally stored in
            # different parquet shards with different column orders). Realign to the
            # writer's established schema so `write_table` doesn't reject a reordered match.
            table = table.select(self._pq_writer.schema.names)
        self._pq_writer.write_table(table)
@@ -188,8 +214,8 @@ class LeRobotDatasetMetadata:
    def _load_metadata(self):
        self.info = load_info(self.root)
        check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
-        self.tasks = load_tasks(self.root)
+        self.tasks = load_tasks(self.root) if self.total_tasks > 0 else None
-        self.episodes = load_episodes(self.root)
+        self.episodes = load_episodes(self.root) if self.total_episodes > 0 else None
        self.stats = load_stats(self.root)
    def ensure_readable(self) -> None:
@@ -232,6 +258,16 @@ class LeRobotDatasetMetadata:
        *,
        token: str | bool | None = None,
    ) -> None:
        if self.repo_type == "bucket":
            self.root.mkdir(parents=True, exist_ok=True)
            sync_bucket(
                f"hf://buckets/{self.repo_id}/meta",
                str(self.root / "meta"),
                delete=True,
                quiet=True,
                token=token,
            )
            return
        token_kwargs = {} if token is None else {"token": token}
        if self._requested_root is None:
            self.root = Path(
@@ -262,6 +298,8 @@ class LeRobotDatasetMetadata:
    @property
    def url_root(self) -> str:
        """Hugging Face Hub URL root for this dataset."""
        if self.repo_type == "bucket":
            return f"hf://buckets/{self.repo_id}"
        return f"hf://datasets/{self.repo_id}"
    @property
@@ -1045,10 +1045,12 @@ def _copy_data_with_feature_changes(
                    df[feature_name] = feature_values
                else:
                    feature_slice = values[frame_idx:end_idx]
-                    if len(feature_slice.shape) > 1 and feature_slice.shape[1] == 1:
+                    if feature_slice.ndim == 1:
                        df[feature_name] = feature_slice
                    elif feature_slice.ndim == 2 and feature_slice.shape[1] == 1:
                        df[feature_name] = feature_slice.flatten()
                    else:
-                        df[feature_name] = feature_slice
+                        df[feature_name] = list(feature_slice)
            frame_idx = end_idx
        # Write using the same chunk/file structure as source
@@ -1435,15 +1437,18 @@ def modify_tasks(
    dataset: LeRobotDataset,
    new_task: str | None = None,
    episode_tasks: dict[int, str] | None = None,
    task_replacements: dict[str, str] | None = None,
 ) -> LeRobotDataset:
    """Modify tasks in a LeRobotDataset.
    This function allows you to either:
    1. Set a single task for the entire dataset (using `new_task`)
    2. Set specific tasks for specific episodes (using `episode_tasks`)
    3. Replace existing task strings wherever they appear (using `task_replacements`)
-    You can combine both: `new_task` sets the default, and `episode_tasks` overrides
+    Per episode, the task is resolved with precedence:
-    specific episodes.
+    `episode_tasks` > `task_replacements` > `new_task` > original task. An episode that ends
    up with no task (none of the above apply and it had no original task) raises an error.
    The dataset is modified in-place, updating only the task-related files:
    - meta/tasks.parquet
@@ -1453,11 +1458,14 @@ def modify_tasks(
    Args:
        dataset: The source LeRobotDataset to modify.
-        new_task: A single task string to apply to all episodes. If None and episode_tasks
+        new_task: Default task applied to any episode not covered by `episode_tasks` or a
-            is also None, raises an error.
+            matching `task_replacements` entry.
-        episode_tasks: Optional dict mapping episode indices to their task strings.
+        episode_tasks: Optional dict mapping episode indices to task strings. Takes precedence
-            Overrides `new_task` for specific episodes.
+            over both `task_replacements` and `new_task`.
        task_replacements: Optional dict mapping existing task strings to new ones. Applied to
            episodes whose current task matches a key. Every key must be an existing task.
    At least one of `new_task`, `episode_tasks`, or `task_replacements` must be provided.
    Examples:
        Set a single task for all episodes:
@@ -1475,11 +1483,17 @@ def modify_tasks(
                new_task="Default task",
                episode_tasks={5: "Special task for episode 5"}
            )
    """
    if new_task is None and episode_tasks is None:
        raise ValueError("Must specify at least one of new_task or episode_tasks")
-    if episode_tasks is not None:
+        Replace existing task strings in-place:
            dataset = modify_tasks(
                dataset,
                task_replacements={"Pick up the cube": "Lift the cube"}
            )
    """
    if not new_task and not episode_tasks and not task_replacements:
        raise ValueError("Must specify at least one of new_task, episode_tasks, or task_replacements")
    if episode_tasks:
        valid_indices = set(range(dataset.meta.total_episodes))
        invalid = set(episode_tasks.keys()) - valid_indices
        if invalid:
@@ -1489,19 +1503,29 @@ def modify_tasks(
    if dataset.meta.episodes is None:
        dataset.meta.episodes = load_episodes(dataset.root)
    if task_replacements:
        current_tasks = set(dataset.meta.tasks.index)
        invalid_tasks = set(task_replacements) - current_tasks
        if invalid_tasks:
            raise ValueError(f"Task replacements reference unknown tasks: {sorted(invalid_tasks)}")
    # Build the mapping from episode index to task string
    episode_to_task: dict[int, str] = {}
    for ep_idx in range(dataset.meta.total_episodes):
        original_tasks = dataset.meta.episodes[ep_idx]["tasks"]
        original_task = original_tasks[0] if original_tasks else None
        if episode_tasks and ep_idx in episode_tasks:
            episode_to_task[ep_idx] = episode_tasks[ep_idx]
-        elif new_task is not None:
+        elif task_replacements and original_task in task_replacements:
            episode_to_task[ep_idx] = task_replacements[original_task]
        elif new_task:
            episode_to_task[ep_idx] = new_task
-        else:
+        elif original_task:
            # Keep original task if not overridden and no default provided
-            original_tasks = dataset.meta.episodes[ep_idx]["tasks"]
+            episode_to_task[ep_idx] = original_task
-            if not original_tasks:
+        else:
-                raise ValueError(f"Episode {ep_idx} has no tasks and no default task was provided")
+            raise ValueError(f"Episode {ep_idx} has no task; provide new_task or episode_tasks")
            episode_to_task[ep_idx] = original_tasks[0]
    # Collect all unique tasks and create new task mapping
    unique_tasks = sorted(set(episode_to_task.values()))
@@ -386,6 +386,9 @@ class DatasetWriter:
                self._episodes_since_last_encoding = 0
        if episode_data is None:
            # Post-save cleanup deliberately does not go through clear_episode_buffer():
            # staging frames of video cameras must survive here — the (possibly batched)
            # encoder still needs them and deletes them once each video is written.
            if len(self._meta.image_keys) > 0:
                self._delete_camera_frame_dirs(self._meta.image_keys)
            self.episode_buffer = self._create_episode_buffer()
@@ -17,8 +17,8 @@ from collections.abc import Sequence
 from typing import Any
 from lerobot.configs import PipelineFeatureType
 from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.processor import DataProcessorPipeline
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE, OBS_STR
 from lerobot.utils.feature_utils import hw_to_dataset_features
@@ -16,6 +16,7 @@
 from collections import deque
 from collections.abc import Callable, Generator, Iterable, Iterator
 from pathlib import Path
 from typing import Literal
 import datasets
 import numpy as np
@@ -58,6 +59,10 @@ class LookAheadError(Exception):
    pass
 class _ShardExhaustedError(Exception):
    """Raised when a streaming dataset shard has no more items."""
 class Backtrackable[T]:
    """
    Wrap any iterator/iterable so you can step back up to `history` items
@@ -178,7 +183,7 @@ class Backtrackable[T]:
        """
        Check if we can go back `steps` items without raising an IndexError.
        """
-        return steps <= len(self._back_buf) + self._cursor
+        return steps < len(self._back_buf) + self._cursor
    def can_peek_ahead(self, steps: int = 1) -> bool:
        """
@@ -257,15 +262,16 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
        return_uint8: bool = False,
        depth_output_unit: str = DEFAULT_DEPTH_UNIT,
        *,
        repo_type: Literal["dataset", "bucket"] = "dataset",
        token: str | bool | None = None,
    ):
        """Initialize a StreamingLeRobotDataset.
        Args:
            repo_id (str): This is the repo id that will be used to fetch the dataset.
-            root (Path | None, optional): Local directory to use for local datasets. When omitted, Hub
+            root (Path | None, optional): Local directory to use for local datasets. In bucket mode,
-                metadata is resolved through a revision-safe snapshot cache under
+                this is an optional local metadata-cache directory; parquet and video data remain remote.
-                ``$HF_LEROBOT_HOME/hub``.
+                When omitted, Hub metadata is resolved through the cache under ``$HF_LEROBOT_HOME/hub``.
            episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                their episode_index in this list.
            image_transforms (Callable | None, optional): Transform to apply to image data.
@@ -280,6 +286,8 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
            shuffle (bool, optional): Whether to shuffle the dataset across exhaustions. Defaults to True.
            depth_output_unit (str, optional): Physical unit depth maps are dequantized to ("m" or "mm").
                Defaults to "mm".
            repo_type: "dataset" (default) or "bucket" to stream from an HF Storage Bucket
                over ``hf://buckets/``.
            token: Authentication token used while streaming this dataset from
                the Hub. Pass a string token, ``True`` to require the locally
                stored token, ``False`` to disable authentication, or ``None``
@@ -287,10 +295,14 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
                on the dataset instance after initialization.
        """
        super().__init__()
        if repo_type not in ("dataset", "bucket"):
            raise ValueError(f"repo_type must be 'dataset' or 'bucket', got {repo_type!r}")
        self.repo_id = repo_id
-        self._requested_root = Path(root) if root else None
+        self.repo_type = repo_type
        self._requested_root = Path(root) if root is not None else None
        self.root = self._requested_root if self._requested_root is not None else HF_LEROBOT_HOME / repo_id
-        self.streaming_from_local = root is not None
+        self.streaming_from_local = root is not None and self.repo_type == "dataset"
        self.image_transforms = image_transforms
        self.episodes = episodes
@@ -317,6 +329,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
            self._requested_root,
            self.revision,
            force_cache_sync=force_cache_sync,
            repo_type=self.repo_type,
            token=token,
        )
        self.root = self.meta.root
@@ -345,15 +358,26 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
            self.delta_timestamps = delta_timestamps
            self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
-        token_kwargs = {} if token is None or self.streaming_from_local else {"token": token}
+        token_kwargs = {} if token is None else {"token": token}
-        self.hf_dataset: datasets.IterableDataset = load_dataset(
+        if self.repo_type == "bucket":
-            self.repo_id if not self.streaming_from_local else str(self.root),
+            self.hf_dataset: datasets.IterableDataset = load_dataset(
-            split="train",
+                "parquet",
-            streaming=self.streaming,
+                data_files=f"hf://buckets/{self.repo_id}/data/*/*.parquet",
-            data_files="data/*/*.parquet",
+                split="train",
-            revision=self.revision,
+                streaming=self.streaming,
-            **token_kwargs,
+                **token_kwargs,
-        )
+            )
        else:
            if self.streaming_from_local:
                token_kwargs = {}
            self.hf_dataset: datasets.IterableDataset = load_dataset(
                self.repo_id if not self.streaming_from_local else str(self.root),
                split="train",
                streaming=self.streaming,
                data_files="data/*/*.parquet",
                revision=self.revision,
                **token_kwargs,
            )
        self.num_shards = min(self.hf_dataset.num_shards, max_num_shards)
@@ -422,10 +446,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
                    else:
                        frames_buffer.append(frame)
                    break  # random shard sampled, switch shard
-            except (
+            except _ShardExhaustedError:
                RuntimeError,
                StopIteration,
            ):  # NOTE: StopIteration inside a generator throws a RuntimeError since python 3.7
                del idx_to_backtrack_dataset[shard_key]  # Remove exhausted shard, onto another shard
        # Once shards are all exhausted, shuffle the buffer and yield the remaining frames
@@ -503,7 +524,11 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
    def make_frame(self, dataset_iterator: Backtrackable) -> Generator:
        """Makes a frame starting from a dataset iterator"""
-        item = next(dataset_iterator)
+        try:
            item = next(dataset_iterator)
        except StopIteration as e:
            # Translate exhaustion here, before PEP 479 turns it into an indistinguishable RuntimeError.
            raise _ShardExhaustedError from e
        item = item_to_torch(item)
        updates = []  # list of "updates" to apply to the item retrieved from hf_dataset (w/o camera features)
@@ -26,7 +26,6 @@ import numpy as np
 import packaging.version
 import torch
 from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from huggingface_hub.errors import RevisionNotFoundError
 from lerobot.utils.utils import flatten_dict, unflatten_dict
@@ -51,6 +50,17 @@ The dataset you requested ({repo_id}) is only available in {version} format.
 As we cannot ensure forward compatibility with it, please update your current version of lerobot.
 """
 MISSING_VERSION_TAG_MESSAGE = """
 Your dataset must be tagged with a codebase version.
 Assuming _version_ is the codebase_version value in the info.json, you can run this:
 ```python
 from huggingface_hub import HfApi
 hub_api = HfApi()
 hub_api.create_tag("{repo_id}", tag="_version_", repo_type="dataset")
 ```
 """
 class CompatibilityError(Exception): ...
@@ -368,7 +378,7 @@ def get_safe_version(
        str: The safe version string (e.g., "v1.2.3") to use as a revision.
    Raises:
-        RevisionNotFoundError: If the repo has no version tags.
+        RuntimeError: If the repo has no version tags.
        BackwardCompatibilityError: If only older major versions are available.
        ForwardCompatibilityError: If only newer major versions are available.
    """
@@ -378,17 +388,7 @@ def get_safe_version(
    hub_versions = get_repo_versions(repo_id) if token is None else get_repo_versions(repo_id, token=token)
    if not hub_versions:
-        raise RevisionNotFoundError(
+        raise RuntimeError(MISSING_VERSION_TAG_MESSAGE.format(repo_id=repo_id))
            f"""Your dataset must be tagged with a codebase version.
            Assuming _version_ is the codebase_version value in the info.json, you can run this:
            ```python
            from huggingface_hub import HfApi
            hub_api = HfApi()
            hub_api.create_tag("{repo_id}", tag="_version_", repo_type="dataset")
            ```
            """
        )
    if target_version in hub_versions:
        return f"v{target_version}"
@@ -507,7 +507,7 @@ class MetaworldEnv(EnvConfig):
 class RoboCasaEnv(EnvConfig):
    task: str = "CloseFridge"
    fps: int = 20
-    episode_length: int = 1000
+    episode_length: int | None = None
    obs_type: str = "pixels_agent_pos"
    render_mode: str = "rgb_array"
    camera_name: str = "robot0_agentview_left,robot0_eye_in_hand,robot0_agentview_right"
@@ -30,7 +30,7 @@ from gymnasium import spaces
 from libero.libero import benchmark, get_libero_path
 from libero.libero.envs import OffScreenRenderEnv
-from lerobot.types import RobotObservation
+from lerobot.lerobot_types import RobotObservation
 from .utils import _LazyAsyncVectorEnv, parse_camera_names
@@ -384,7 +384,12 @@ class LiberoEnv(gym.Env):
    def close(self):
        if self._env is not None:
-            self._env.close()
+            try:
                self._env.close()
            finally:
                # LIBERO deletes its inner env on close, so this wrapper must
                # be recreated before the next reset.
                self._env = None
 def _make_env_fns(
@@ -25,7 +25,7 @@ import metaworld.policies as policies
 import numpy as np
 from gymnasium import spaces
-from lerobot.types import RobotObservation
+from lerobot.lerobot_types import RobotObservation
 from .utils import _LazyAsyncVectorEnv
@@ -25,7 +25,7 @@ import gymnasium as gym
 import numpy as np
 from gymnasium import spaces
-from lerobot.types import RobotObservation
+from lerobot.lerobot_types import RobotObservation
 from .utils import _LazyAsyncVectorEnv, parse_camera_names
@@ -98,6 +98,19 @@ def _resolve_tasks(task: str) -> tuple[list[str], str | None]:
    return names, None
 def _get_task_horizon(task: str) -> int:
    """Return the rollout horizon registered by RoboCasa for a task."""
    from robocasa.utils.dataset_registry_utils import get_task_horizon
    try:
        return int(get_task_horizon(task))
    except ValueError as exc:
        raise ValueError(
            f"No RoboCasa horizon is registered for task '{task}'. "
            "Set `--env.episode_length=<steps>` explicitly."
        ) from exc
 def convert_action(flat_action: np.ndarray) -> dict[str, Any]:
    """Split a flat (12,) action vector into a RoboCasa action dict.
@@ -154,7 +167,7 @@ class RoboCasaEnv(gym.Env):
        self.camera_name = parse_camera_names(camera_name)
-        self._max_episode_steps = episode_length if episode_length is not None else 1000
+        self._max_episode_steps = episode_length if episode_length is not None else _get_task_horizon(task)
        # Deferred — created on first reset() inside the worker subprocess
        # to avoid inheriting stale GPU/EGL contexts across fork().
@@ -28,7 +28,7 @@ import numpy as np
 import torch
 from gymnasium import spaces
-from lerobot.types import RobotObservation
+from lerobot.lerobot_types import RobotObservation
 from lerobot.utils.import_utils import _scipy_available
 from .utils import _LazyAsyncVectorEnv
@@ -384,7 +384,9 @@ class RoboTwinEnv(gym.Env):
        self._env: Any | None = None  # deferred — created on first reset() inside worker
        self._step_count: int = 0
-        self._black_frame = np.zeros((self.observation_height, self.observation_width, 3), dtype=np.uint8)
+        self._black_frame: np.ndarray = np.zeros(
            (self.observation_height, self.observation_width, 3), dtype=np.uint8
        )
        image_spaces = {
            cam: spaces.Box(
@@ -37,7 +37,7 @@ import numpy as np
 from gymnasium import spaces
 from scipy.spatial.transform import Rotation
-from lerobot.types import RobotObservation
+from lerobot.lerobot_types import RobotObservation
 from .utils import _LazyAsyncVectorEnv
@@ -373,7 +373,7 @@ class VLABenchEnv(gym.Env):
        if action.shape[0] != 7:
            # Unknown layout — fall back to zero-pad so the sim doesn't crash.
-            padded = np.zeros(ctrl_dim, dtype=np.float64)
+            padded: np.ndarray = np.zeros(ctrl_dim, dtype=np.float64)
            padded[: min(action.shape[0], ctrl_dim)] = action[:ctrl_dim]
            return padded
@@ -122,6 +122,9 @@ MODEL_ENCODING_TABLE = {
    "xm430-w350": X_SERIES_ENCODINGS_TABLE,
    "xm540-w270": X_SERIES_ENCODINGS_TABLE,
    "xc430-w150": X_SERIES_ENCODINGS_TABLE,
    "xh540-w150": X_SERIES_ENCODINGS_TABLE,
    "xc330-t288": X_SERIES_ENCODINGS_TABLE,
    "xc330-t181": X_SERIES_ENCODINGS_TABLE,
 }
 # {model: model_resolution}
@@ -134,6 +137,9 @@ MODEL_RESOLUTION = {
    "xm430-w350": 4096,
    "xm540-w270": 4096,
    "xc430-w150": 4096,
    "xh540-w150": 4096,
    "xc330-t288": 4096,
    "xc330-t181": 4096,
 }
 # {model: model_number}
@@ -145,6 +151,9 @@ MODEL_NUMBER_TABLE = {
    "xm430-w350": 1020,
    "xm540-w270": 1120,
    "xc430-w150": 1070,
    "xh540-w150": 1110,
    "xc330-t288": 1220,
    "xc330-t181": 1210,
 }
 # {model: available_operating_modes}
@@ -156,6 +165,9 @@ MODEL_OPERATING_MODES = {
    "xm430-w350": [0, 1, 3, 4, 5, 16],
    "xm540-w270": [0, 1, 3, 4, 5, 16],
    "xc430-w150": [1, 3, 4, 16],
    "xh540-w150": [0, 1, 3, 4, 5, 16],
    "xc330-t288": [0, 1, 3, 4, 5, 16],
    "xc330-t181": [0, 1, 3, 4, 5, 16],
 }
 MODEL_CONTROL_TABLE = {
@@ -166,6 +178,9 @@ MODEL_CONTROL_TABLE = {
    "xm430-w350": X_SERIES_CONTROL_TABLE,
    "xm540-w270": X_SERIES_CONTROL_TABLE,
    "xc430-w150": X_SERIES_CONTROL_TABLE,
    "xh540-w150": X_SERIES_CONTROL_TABLE,
    "xc330-t288": X_SERIES_CONTROL_TABLE,
    "xc330-t181": X_SERIES_CONTROL_TABLE,
 }
 MODEL_BAUDRATE_TABLE = {
@@ -176,6 +191,9 @@ MODEL_BAUDRATE_TABLE = {
    "xm430-w350": X_SERIES_BAUDRATE_TABLE,
    "xm540-w270": X_SERIES_BAUDRATE_TABLE,
    "xc430-w150": X_SERIES_BAUDRATE_TABLE,
    "xh540-w150": X_SERIES_BAUDRATE_TABLE,
    "xc330-t288": X_SERIES_BAUDRATE_TABLE,
    "xc330-t181": X_SERIES_BAUDRATE_TABLE,
 }
 AVAILABLE_BAUDRATES = [
@@ -22,6 +22,7 @@ from typing import TYPE_CHECKING, Any
 import torch
 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.lerobot_types import TransitionKey
 from lerobot.processor import (
    ComplementaryDataProcessorStep,
    PolicyAction,
@@ -31,7 +32,6 @@ from lerobot.processor import (
    make_default_policy_processor_steps,
    make_policy_processor_pipelines,
 )
 from lerobot.types import TransitionKey
 from lerobot.utils.constants import OBS_STATE
 from lerobot.utils.import_utils import _transformers_available, require_package
@@ -42,6 +42,9 @@ class Evo1Policy(PreTrainedPolicy):
    config_class = Evo1Config
    name = "evo1"
    def supports_rtc(self) -> bool:
        return True
    def __init__(self, config: Evo1Config, *, vlm_hub_kwargs: dict | None = None, **kwargs):
        super().__init__(config)
        config.validate_features()
@@ -21,6 +21,7 @@ from typing import Any
 import torch
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
    AddBatchDimensionProcessorStep,
    DeviceProcessorStep,
@@ -40,7 +41,6 @@ from lerobot.processor.converters import (
    policy_action_to_transition,
    transition_to_policy_action,
 )
 from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
    ACTION,
    DONE,
@@ -302,6 +302,33 @@ def _pad_evo1_stats(
    return padded_stats
 def _refresh_evo1_normalization_steps(
    config: Evo1Config,
    preprocessor: PolicyProcessorPipeline,
    postprocessor: PolicyProcessorPipeline,
 ) -> None:
    """Re-pad checkpoint-loaded (un)normalizer stats/features to EVO1's fixed widths.
    Loading a checkpoint injects the raw dataset stats (unpadded to max_state_dim/max_action_dim)
    into the (un)normalizer via the generic override path in make_pre_post_processors. Those stats
    and their declared features must be re-padded/reshaped to EVO1's fixed widths, otherwise
    normalization fails against the padded state/action tensors (e.g. state padded to 24 vs. 8-dim
    LIBERO stats). Padding is a no-op when stats are already at the target width.
    """
    normalization_features = _evo1_normalization_features(config)
    action_features = _evo1_action_features(config)
    for step in preprocessor.steps:
        if isinstance(step, NormalizerProcessorStep):
            step.features = normalization_features
            step.stats = _pad_evo1_stats(config, step.stats)
            step.to(device=step.device, dtype=step.dtype)
    for step in postprocessor.steps:
        if isinstance(step, UnnormalizerProcessorStep):
            step.features = action_features
            step.stats = _pad_evo1_stats(config, step.stats)
            step.to(device=step.device, dtype=step.dtype)
 def reconcile_evo1_processors(
    config: Evo1Config,
    preprocessor: PolicyProcessorPipeline,
@@ -309,16 +336,19 @@ def reconcile_evo1_processors(
 ) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
    """Reconcile checkpoint-loaded pipelines with the current EVO1 config.
-    Two things cannot be restored from a serialized pipeline alone: the EVO1 batch converter
+    Three things cannot be restored from a serialized pipeline alone: the EVO1 batch converter
-    (converters are plain functions and are never serialized), and eval-time CLI overrides of the
+    (converters are plain functions and are never serialized), eval-time CLI overrides of the
-    action postprocessing flags (`postprocess_action_dim`, `binarize_gripper`, `gripper_*`). This
+    action postprocessing flags (`postprocess_action_dim`, `binarize_gripper`, `gripper_*`), and the
-    restores the converter and rebuilds the action step from the current config so those overrides
+    (un)normalizer stats/features when the generic override path injects raw, unpadded dataset
-    take effect.
+    stats. This restores the converter, re-pads the normalization stats to EVO1's fixed widths, and
    rebuilds the action step from the current config so those overrides take effect.
    """
    # Pipelines reloaded from a checkpoint come back with the default batch converter, which drops
    # non-observation extras (embodiment_id, state_mask, custom task fields) needed by EVO1.
    preprocessor.to_transition = evo1_batch_to_transition
    _refresh_evo1_normalization_steps(config, preprocessor, postprocessor)
    action_step = Evo1ActionProcessorStep(
        action_dim=_evo1_action_dim(config),
        binarize_gripper=config.binarize_gripper,
@@ -28,6 +28,7 @@ if TYPE_CHECKING:
 from lerobot.configs import FeatureType, PreTrainedConfig
 from lerobot.envs import EnvConfig, env_to_policy_features
 from lerobot.lerobot_types import PolicyAction
 from lerobot.processor import (
    AbsoluteActionsProcessorStep,
    PolicyProcessorPipeline,
@@ -37,19 +38,25 @@ from lerobot.processor import (
    transition_to_batch,
    transition_to_policy_action,
 )
 from lerobot.types import PolicyAction
 from lerobot.utils.constants import (
    ACTION,
    POLICY_POSTPROCESSOR_DEFAULT_NAME,
    POLICY_PREPROCESSOR_DEFAULT_NAME,
 )
 from lerobot.utils.feature_utils import dataset_to_policy_features
 from lerobot.utils.import_utils import _peft_available, require_package
 from .evo1.configuration_evo1 import Evo1Config
 from .groot.configuration_groot import GrootConfig
 from .pretrained import PreTrainedPolicy
 from .utils import validate_visual_features_consistency
 if TYPE_CHECKING or _peft_available:
    from peft import PeftConfig, PeftModel
 else:
    PeftConfig = None
    PeftModel = None
 def _reconnect_relative_absolute_steps(
    preprocessor: PolicyProcessorPipeline, postprocessor: PolicyProcessorPipeline
@@ -334,12 +341,15 @@ def make_policy(
        # Load a pretrained PEFT model on top of the policy. The pretrained path points to the folder/repo
        # of the adapter and the adapter's config contains the path to the base policy. So we need the
        # adapter config first, then load the correct policy and then apply PEFT.
-        from peft import PeftConfig, PeftModel
+        require_package("peft", extra="peft")
        logging.info("Loading policy's PEFT adapter.")
        peft_pretrained_path = str(cfg.pretrained_path)
-        peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
+        peft_config = PeftConfig.from_pretrained(
            peft_pretrained_path,
            revision=cfg.pretrained_revision,
        )
        kwargs["pretrained_name_or_path"] = peft_config.base_model_name_or_path
        if not kwargs["pretrained_name_or_path"]:
@@ -350,9 +360,14 @@ def make_policy(
                "the adapter was trained."
            )
        kwargs["revision"] = peft_config.revision
        policy = policy_cls.from_pretrained(**kwargs)
        policy = PeftModel.from_pretrained(
-            policy, peft_pretrained_path, config=peft_config, is_trainable=True
+            policy,
            peft_pretrained_path,
            config=peft_config,
            revision=cfg.pretrained_revision,
            is_trainable=True,
        )
    else:
@@ -68,6 +68,9 @@ class GrootPolicy(PreTrainedPolicy):
    name = "groot"
    config_class = GrootConfig
    def supports_rtc(self) -> bool:
        return True
    def __init__(self, config: GrootConfig, **kwargs):
        """Initialize Groot policy wrapper."""
        require_package("transformers", extra="groot")
@@ -50,6 +50,7 @@ if TYPE_CHECKING or _datasets_available:
 else:
    LeRobotDataset = None
 from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
    AbsoluteActionsProcessorStep,
    AddBatchDimensionProcessorStep,
@@ -66,7 +67,6 @@ from lerobot.processor import (
    transition_to_batch,
    transition_to_policy_action,
 )
 from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
    ACTION,
    OBS_IMAGE,
@@ -43,11 +43,22 @@ from torch.distributions import Beta
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.utils.constants import ACTION
-from lerobot.utils.import_utils import _scipy_available, _transformers_available, require_package
+from lerobot.utils.import_utils import (
    _peft_available,
    _scipy_available,
    _transformers_available,
    require_package,
 )
 from ..rtc.modeling_rtc import RTCProcessor
 from .configuration_molmoact2 import MolmoAct2Config
 if TYPE_CHECKING or _peft_available:
    from peft import LoraConfig, get_peft_model
 else:
    LoraConfig = None
    get_peft_model = None
 logger = logging.getLogger(__name__)
@@ -509,6 +520,9 @@ class MolmoAct2Policy(PreTrainedPolicy):
    config_class = MolmoAct2Config
    name = "molmoact2"
    def supports_rtc(self) -> bool:
        return self.config.inference_action_mode == "continuous"
    def __init__(
        self,
        config: MolmoAct2Config,
@@ -1731,13 +1745,11 @@ class MolmoAct2Policy(PreTrainedPolicy):
    def _build_inner_lora_config(self):
        require_package("peft", extra="molmoact2")
        from peft import LoraConfig
        return LoraConfig(**self._get_inner_peft_targets())
    def _apply_lora_adapters(self) -> None:
        require_package("peft", extra="molmoact2")
        from peft import get_peft_model
        peft_config = self._build_inner_lora_config()
        self._validate_peft_config(peft_config)
@@ -1801,7 +1801,7 @@ class MolmoAct2VisionBackbone(nn.Module):
        valid = pooled_patches_idx >= 0
        valid_token = torch.any(valid, -1)
-        # Use `pooled_patches_idx` to arange the features for image pooling
+        # Use `pooled_patches_idx` to arrange the features for image pooling
        batch_idx = torch.arange(
            pooled_patches_idx.shape[0],
            dtype=torch.long,
@@ -36,6 +36,7 @@ import torch
 from torch import Tensor
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
    AddBatchDimensionProcessorStep,
    DeviceProcessorStep,
@@ -49,7 +50,6 @@ from lerobot.processor import (
    policy_action_to_transition,
    transition_to_policy_action,
 )
 from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
    ACTION,
    OBS_IMAGES,
@@ -749,6 +749,9 @@ class PI0Policy(PreTrainedPolicy):
    config_class = PI0Config
    name = "pi0"
    def supports_rtc(self) -> bool:
        return True
    def __init__(
        self,
        config: PI0Config,
@@ -714,6 +714,9 @@ class PI05Policy(PreTrainedPolicy):
    config_class = PI05Config
    name = "pi05"
    def supports_rtc(self) -> bool:
        return True
    def __init__(
        self,
        config: PI05Config,
@@ -22,6 +22,7 @@ import numpy as np
 import torch
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
    AbsoluteActionsProcessorStep,
    PolicyAction,
@@ -33,7 +34,6 @@ from lerobot.processor import (
    make_default_policy_processor_steps,
    make_policy_processor_pipelines,
 )
 from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import OBS_STATE
 from .configuration_pi05 import PI05Config
@@ -22,6 +22,7 @@ import numpy as np
 import torch
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.lerobot_types import EnvTransition, TransitionKey
 from lerobot.processor import (
    AbsoluteActionsProcessorStep,
    ActionTokenizerProcessorStep,
@@ -34,7 +35,6 @@ from lerobot.processor import (
    make_default_policy_processor_steps,
    make_policy_processor_pipelines,
 )
 from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import OBS_STATE
 from .configuration_pi0_fast import PI0FastConfig
@@ -34,14 +34,22 @@ from lerobot.configs import PreTrainedConfig
 from lerobot.configs.train import TrainPipelineConfig
 from lerobot.utils.device_utils import resolve_safetensors_device
 from lerobot.utils.hub import HubMixin
 from lerobot.utils.import_utils import _peft_available, require_package
 from .utils import log_model_loading_keys
-T = TypeVar("T", bound="PreTrainedPolicy")
+if TYPE_CHECKING or _peft_available:
    from peft import PEFT_TYPE_TO_CONFIG_MAPPING, PeftType, get_peft_model
 else:
    PEFT_TYPE_TO_CONFIG_MAPPING = None
    PeftType = None
    get_peft_model = None
 if TYPE_CHECKING:
    from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
 T = TypeVar("T", bound="PreTrainedPolicy")
 def _build_card_context(
    cfg: TrainPipelineConfig | None,
@@ -241,6 +249,10 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
        """
        raise NotImplementedError
    def supports_rtc(self) -> bool:
        """Whether this policy implements Real-Time Chunking inference semantics."""
        return False
    # TODO(aliberts, rcadene): split into 'forward' and 'compute_loss'?
    @abc.abstractmethod
    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict | None]:
@@ -384,7 +396,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
            peft_cli_overrides: Optional dict of CLI overrides (method_type, target_modules, r, etc.)
                These are merged with policy defaults to build the final config.
        """
-        from peft import get_peft_model
+        require_package("peft", extra="peft")
        # If user provided a complete config, use it directly (with overrides)
        if peft_config is not None:
@@ -455,7 +467,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
        Returns:
            Preprocessed dict with renamed keys and init_type mapped to method-specific key.
        """
-        from peft import PeftType
+        require_package("peft", extra="peft")
        cli_overrides = cli_overrides.copy()
@@ -480,7 +492,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
    def _build_peft_config(self, cli_overrides: dict):
        """Build a PEFT config from policy defaults and CLI overrides."""
-        from peft import PEFT_TYPE_TO_CONFIG_MAPPING, PeftType
+        require_package("peft", extra="peft")
        # Determine PEFT method type (default to LORA)
        method_type_str = cli_overrides.get("method_type") or "lora"
@@ -507,7 +519,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
    def _apply_peft_cli_overrides(self, peft_config, cli_overrides: dict):
        """Apply CLI overrides to an existing PEFT config."""
-        from peft import PEFT_TYPE_TO_CONFIG_MAPPING, PeftType
+        require_package("peft", extra="peft")
        # Get method type from existing config or CLI override
        method_type_str = cli_overrides.get("method_type")
@@ -145,6 +145,9 @@ class SmolVLAPolicy(PreTrainedPolicy):
    config_class = SmolVLAConfig
    name = "smolvla"
    def supports_rtc(self) -> bool:
        return True
    def __init__(
        self,
        config: SmolVLAConfig,
@@ -168,14 +168,23 @@ class SmolVLMWithExpertModel(nn.Module):
                last_layers.append(self.num_vlm_layers - 2)
            frozen_layers = [
                "lm_head",
-                "text_model.model.norm.weight",
+                "text_model.norm.weight",
            ]
            for layer in last_layers:
-                frozen_layers.append(f"text_model.model.layers.{layer}.")
+                frozen_layers.append(f"text_model.layers.{layer}.")
            unmatched_patterns = set(frozen_layers)
            for name, params in self.vlm.named_parameters():
-                if any(k in name for k in frozen_layers):
+                matched_patterns = [k for k in frozen_layers if k in name]
                if matched_patterns:
                    params.requires_grad = False
                    unmatched_patterns.difference_update(matched_patterns)
            if unmatched_patterns:
                raise RuntimeError(
                    "Some frozen layer patterns matched no VLM parameters, so the corresponding layers "
                    "would silently remain trainable (parameter naming may have changed in transformers): "
                    f"{sorted(unmatched_patterns)}"
                )
        # To avoid unused params issue with distributed training
        for name, params in self.lm_expert.named_parameters():
            if "lm_head" in name:
@@ -22,7 +22,7 @@ import torch
 from torch import nn
 from lerobot.configs import FeatureType, PolicyFeature, PreTrainedConfig
-from lerobot.types import PolicyAction, RobotAction, RobotObservation
+from lerobot.lerobot_types import PolicyAction, RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame
@@ -15,11 +15,13 @@
 from __future__ import annotations
 from dataclasses import dataclass, field
 from typing import Any
 from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
 from lerobot.utils.constants import OBS_STATE
@PreTrainedConfig.register_subclass("vla_jepa")
@@ -122,6 +124,13 @@ class VLAJEPAConfig(PreTrainedConfig):
        if self.robot_state_feature is not None:
            self.state_dim = self.robot_state_feature.shape[0]
    def set_dataset_feature_metadata(self, dataset_features: dict[str, Any]) -> None:
        """Add `observation.state` to `input_features` if missing, so it gets normalized."""
        if OBS_STATE in self.input_features or OBS_STATE not in dataset_features:
            return
        shape = tuple(dataset_features[OBS_STATE]["shape"])
        self.input_features[OBS_STATE] = PolicyFeature(type=FeatureType.STATE, shape=shape)
    def get_optimizer_preset(self) -> AdamWConfig:
        return AdamWConfig(
            lr=self.optimizer_lr,
@@ -16,6 +16,7 @@ from __future__ import annotations
 import logging
 from collections import deque
 from contextlib import nullcontext
 from pathlib import Path
 from typing import TYPE_CHECKING, Any
@@ -26,6 +27,7 @@ from torch import Tensor, nn
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.policies.utils import populate_queues
 from lerobot.utils.constants import ACTION, OBS_STATE
 from lerobot.utils.device_utils import is_amp_available
 from lerobot.utils.import_utils import _transformers_available, require_package
 if TYPE_CHECKING or _transformers_available:
@@ -39,6 +41,21 @@ from .configuration_vla_jepa import VLAJEPAConfig
 from .qwen_interface import Qwen3VLInterface
 from .world_model import ActionConditionedVideoPredictor
 def _get_autocast_context(device_type: str, dtype: torch.dtype = torch.bfloat16):
    """Return an autocast context appropriate for the device.
    MPS does not support ``torch.autocast`` at all.  On CUDA devices
    without bfloat16 support (compute capability < 8.0) we fall back to
    float16.
    """
    if not is_amp_available(device_type):
        return nullcontext()
    if device_type == "cuda" and dtype == torch.bfloat16 and not torch.cuda.is_bf16_supported():
        dtype = torch.float16
    return torch.autocast(device_type=device_type, dtype=dtype)
 # ============================================================================
 # Native VLA-JEPA Model - follows original starVLA VLA_JEPA.py implementation
 # ============================================================================
@@ -183,7 +200,7 @@ class VLAJEPAModel(nn.Module):
            action_idx = action_mask.nonzero(as_tuple=True)
        device_type = next(self.parameters()).device.type
-        with torch.autocast(device_type=device_type, dtype=torch.bfloat16):
+        with _get_autocast_context(device_type, torch.bfloat16):
            last_hidden = self._qwen_last_decoder_hidden(qwen_inputs)  # [B, seq_len, H]
            b, _, h = last_hidden.shape
            embodied_action_tokens = last_hidden[embodied_idx[0], embodied_idx[1], :].view(b, -1, h)
@@ -250,7 +267,7 @@ class VLAJEPAModel(nn.Module):
    ) -> Tensor:
        """Flow-matching action-head loss, repeated over `repeated_diffusion_steps`."""
        device_type = next(self.parameters()).device.type
-        with torch.autocast(device_type=device_type, dtype=torch.float32):
+        with _get_autocast_context(device_type, torch.float32):
            r = self.config.repeated_diffusion_steps
            horizon = self.config.chunk_size
            actions_target = actions[:, -horizon:, :].to(torch.float32).repeat(r, 1, 1)
@@ -399,7 +416,8 @@ class VLAJEPAPolicy(PreTrainedPolicy):
        state = batch.get(OBS_STATE)
        if state is not None:
            if state.ndim > 2:
-                state = state[:, -1, :]
+                # deltas are forward-looking here, so index 0 is the current observation, not -1.
                state = state[:, 0, :]
            inputs["state"] = (state.unsqueeze(1) if state.ndim == 2 else state).float()  # [B, 1, dim]
        return inputs
--- a/Show More
+++ b/Show More