feat(datasets): guard unsupported remote streaming decode cases

Raise clear errors when remote streaming (hf://) cannot decode video: RuntimeError when RGB decoding requires torchcodec but it is unavailable, and NotImplementedError for depth video which pyav cannot read over hf://.
refactor(datasets): align streaming video decoding with non-streaming reader
2026-08-08 17:39:44 +00:00 · 2026-08-06 13:21:29 +02:00 · 2026-08-06 13:21:28 +02:00 · 2026-08-06 13:20:47 +02:00 · 2026-08-06 13:20:47 +02:00 · 2026-08-06 13:20:47 +02:00
231 changed files with 6066 additions and 1682 deletions
@@ -53,7 +53,7 @@ permissions:
  contents: read

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

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

    steps:
-      - uses: actions/checkout@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

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

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.8.0"
+  UV_VERSION: "0.11.30"
  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
@@ -52,21 +52,21 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@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 }}
@@ -48,7 +48,7 @@ permissions:

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

 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
@@ -69,7 +69,7 @@ jobs:
      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@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 }}
@@ -37,7 +37,7 @@ permissions:

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

@@ -63,7 +63,7 @@ jobs:
      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
    steps:
-      - uses: actions/checkout@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
@@ -27,7 +27,7 @@ on:

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

@@ -48,12 +48,12 @@ jobs:
    outputs:
      changed: ${{ steps.diff.outputs.changed }}
    steps:
-      - uses: actions/checkout@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'

@@ -21,7 +21,7 @@ on:

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

 jobs:
@@ -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
@@ -19,8 +19,8 @@ on:
  workflow_dispatch:

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

 env:
  CLOSE_ISSUE_MESSAGE: >
@@ -31,7 +31,7 @@ env:
    Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
  WARN_ISSUE_MESSAGE: >
    This issue has been automatically marked as stale because it has not had
-    recent activity (1 year). It will be closed if no further activity occurs.
+    recent activity (1 year). It will be closed if no further activity occurs within 30 days.
    Any change, comment or update to this issue will reset this count.
    Thank you for your contributions.
  WARN_PR_MESSAGE: >
@@ -61,8 +61,8 @@ jobs:
          exempt-pr-labels: never-stale
          days-before-issue-stale: 365
          days-before-issue-close: 30
-          days-before-pr-stale: 365
-          days-before-pr-close: 30
+          days-before-pr-stale: -1
+          days-before-pr-close: -1
          delete-branch: true
          close-issue-message: ${{ env.CLOSE_ISSUE_MESSAGE }}
          close-pr-message: ${{ env.CLOSE_PR_MESSAGE }}
@@ -128,6 +128,23 @@ lerobot-eval \

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

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

 - **[Documentation](https://huggingface.co/docs/lerobot/index):** The complete guide to tutorials & API.
@@ -165,6 +165,8 @@
    title: OpenArm
  - local: rebot_b601
    title: reBot B601-DM
+  - local: third_party_robots
+    title: Third-Party Robots & Teleoperators
  title: "Robots"
 - sections:
  - local: phone_teleop
@@ -175,6 +177,8 @@
 - sections:
  - local: cameras
    title: Cameras
+  - local: third_party_sensors
+    title: Third-Party Cameras & Sensors
  title: "Sensors"
 - sections:
  - local: notebooks
@@ -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 😉
@@ -164,8 +164,8 @@ includes the range reported by the sensor. Requesting an unsupported control als
 Omitted controls leave the sensor's existing automatic or manual setting unchanged. These options
 require `use_rgb=True`.

-On the RealSense D405, the color stream is provided by the Stereo Module, so changing manual
-exposure or gain also affects the depth stream.
+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>
@@ -23,18 +23,18 @@ The broader EVO1 project may include additional training scripts and dataset too
 2. Install EVO1 dependencies:

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

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

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

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

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

 ## Data Requirements

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

 ### Stage 2

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

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

 ### LIBERO Evaluation

-> [!NOTE]
-> Benchmark results for a `lerobot`-hosted LIBERO checkpoint trained with this implementation
-> will be added once training completes.
+#### Reference result

-The official EVO1 LIBERO rollout protocol uses the raw LIBERO camera feature names
+> [!NOTE]
+> The released Stage-2 checkpoint passed clean-download and rollout verification:
+> [`zuoxingdong/evo1_libero`](https://huggingface.co/zuoxingdong/evo1_libero), revision
+> [`515921f4a2c1d3f3ad523721eafa26fdf2af315b`](https://huggingface.co/zuoxingdong/evo1_libero/commit/515921f4a2c1d3f3ad523721eafa26fdf2af315b).
+> The clean-download evaluation used LeRobot revision
+> [`e40b58a8dfa9e7b86918c374791599d070518d11`](https://github.com/huggingface/lerobot/commit/e40b58a8dfa9e7b86918c374791599d070518d11).
+
+The single-run Stage-2 checkpoint at step 70,000 produced:
+
+| Suite          | Successful episodes |  Episodes | Success rate |
+| -------------- | ------------------: | --------: | -----------: |
+| LIBERO Spatial |                 485 |       500 |        97.0% |
+| LIBERO Object  |                 496 |       500 |        99.2% |
+| LIBERO Goal    |                 483 |       500 |        96.6% |
+| LIBERO-10      |                 469 |       500 |        93.8% |
+| **Overall**    |           **1,933** | **2,000** |   **96.65%** |
+
+These results use one trained checkpoint and evaluation seed `1000`; they are not a multi-seed
+mean or confidence estimate.
+
+#### Reference training recipe
+
+The released checkpoint records the complete resolved Stage-2 configuration in
+[`train_config.json`](https://huggingface.co/zuoxingdong/evo1_libero/blob/515921f4a2c1d3f3ad523721eafa26fdf2af315b/train_config.json).
+The measured run used two H100 GPUs with two DDP processes and batch 64 per process, giving global batch 128. Both stages used the same topology. The base VLM came from revision
+`014c0583a0d4bedf29fbe2dbff4f865eb998e171` of `OpenGVLab/InternVL3-1B-hf`.
+The released artifact does not record the exact LeRobot training commit or its original dependency lock,
+so the commands below reproduce the recorded configuration and topology from a current checkout rather
+than reconstructing the software environment bit for bit.
+
+From a LeRobot source checkout, install the locked dependencies and download that exact VLM revision:
+
+```bash
+uv sync --locked --extra training --extra evo1 --extra libero
+VLM_DIR=$(uv run hf download OpenGVLab/InternVL3-1B-hf \
+  --revision=014c0583a0d4bedf29fbe2dbff4f865eb998e171)
+```
+
+Stage 1 freezes the VLM and trains the action head for 5,000 steps:
+
+```bash
+uv run accelerate launch --num_processes=2 -m lerobot.scripts.lerobot_train \
+  --dataset.repo_id=lerobot/libero \
+  --dataset.revision=a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4 \
+  --dataset.video_backend=torchcodec \
+  --dataset.return_uint8=true \
+  --dataset.image_transforms.enable=true \
+  --dataset.use_imagenet_stats=true \
+  --dataset.eval_split=0.0 \
+  --policy.type=evo1 \
+  --policy.training_stage=stage1 \
+  --policy.apply_training_stage_defaults=true \
+  --policy.vlm_model_name="${VLM_DIR}" \
+  --policy.vlm_num_layers=14 \
+  --policy.vlm_dtype=bfloat16 \
+  --policy.device=cuda \
+  --policy.use_amp=true \
+  --policy.use_flash_attn=true \
+  --policy.enable_gradient_checkpointing=true \
+  --policy.gradient_checkpointing_use_reentrant=false \
+  --policy.image_resolution='[448,448]' \
+  --policy.chunk_size=50 \
+  --policy.n_action_steps=50 \
+  --policy.max_state_dim=24 \
+  --policy.max_action_dim=24 \
+  --policy.dropout=0.2 \
+  --policy.optimizer_lr=1e-5 \
+  --policy.optimizer_weight_decay=1e-3 \
+  --policy.optimizer_grad_clip_norm=1.0 \
+  --policy.scheduler_warmup_steps=1000 \
+  --policy.push_to_hub=false \
+  --use_policy_training_preset=true \
+  --batch_size=64 \
+  --steps=5000 \
+  --save_checkpoint=true \
+  --save_checkpoint_to_hub=false \
+  --save_freq=2500 \
+  --log_freq=10 \
+  --env_eval_freq=0 \
+  --num_workers=4 \
+  --prefetch_factor=2 \
+  --persistent_workers=true \
+  --seed=1000 \
+  --wandb.enable=false \
+  --output_dir=./outputs/evo1-libero-stage1-g128-5k
+```
+
+Stage 2 loads the Stage-1 policy but starts a fresh optimizer and scheduler. It trains for 80,000 steps;
+the reported checkpoint is the save at step 70,000:
+
+```bash
+uv run accelerate launch --num_processes=2 -m lerobot.scripts.lerobot_train \
+  --dataset.repo_id=lerobot/libero \
+  --dataset.revision=a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4 \
+  --dataset.video_backend=torchcodec \
+  --dataset.return_uint8=true \
+  --dataset.image_transforms.enable=true \
+  --dataset.use_imagenet_stats=true \
+  --dataset.eval_split=0.0 \
+  --policy.path=./outputs/evo1-libero-stage1-g128-5k/checkpoints/005000/pretrained_model \
+  --policy.training_stage=stage2 \
+  --policy.apply_training_stage_defaults=true \
+  --policy.vlm_model_name="${VLM_DIR}" \
+  --policy.vlm_num_layers=14 \
+  --policy.vlm_dtype=float32 \
+  --policy.device=cuda \
+  --policy.use_amp=true \
+  --policy.use_flash_attn=true \
+  --policy.enable_gradient_checkpointing=true \
+  --policy.gradient_checkpointing_use_reentrant=false \
+  --policy.image_resolution='[448,448]' \
+  --policy.chunk_size=50 \
+  --policy.n_action_steps=50 \
+  --policy.max_state_dim=24 \
+  --policy.max_action_dim=24 \
+  --policy.dropout=0.2 \
+  --policy.optimizer_lr=1e-5 \
+  --policy.optimizer_weight_decay=1e-3 \
+  --policy.optimizer_grad_clip_norm=1.0 \
+  --policy.scheduler_warmup_steps=1000 \
+  --policy.push_to_hub=false \
+  --use_policy_training_preset=true \
+  --batch_size=64 \
+  --steps=80000 \
+  --resume=false \
+  --save_checkpoint=true \
+  --save_checkpoint_to_hub=false \
+  --save_freq=10000 \
+  --log_freq=10 \
+  --env_eval_freq=0 \
+  --num_workers=4 \
+  --prefetch_factor=2 \
+  --persistent_workers=true \
+  --seed=1000 \
+  --wandb.enable=false \
+  --output_dir=./outputs/evo1-libero-stage2-g128-80k
+```
+
+#### Author-format evaluation profile
+
+The author-format EVO1 LIBERO profile uses the raw LIBERO camera feature names
 (`observation.images.agentview_image` and `observation.images.robot0_eye_in_hand_image`), replans every
 14 actions, and binarizes the gripper command before stepping the simulator. The EVO1 policy postprocessor
 can crop the padded 24D action back to the 7D LIBERO action space and apply that gripper binarization. To
-evaluate a LIBERO checkpoint under the same one-episode-per-task setting, keep the raw camera names instead
-of the default `image`/`image2` mapping and set the LIBERO action postprocessing flags:
+evaluate an author-format checkpoint under the same one-episode-per-task setting, keep the raw camera names
+instead of the default `image`/`image2` mapping and set the LIBERO action postprocessing flags:

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

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

 - [EVO1 repository](https://github.com/MINT-SJTU/Evo-1)
@@ -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`.
@@ -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="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/robots_control_video.webp" width="640px" alt="Reachy 2 Demo">
+</div>
+
+## How It Works
+
+**Teleoperate → Record → Train → Deploy**
+
+1. **Teleoperate** - control the robot yourself (with a leader arm, keyboard, or phone) so it can learn from your movements.
+2. **Record** - each demonstration is saved as a dataset: synchronized camera video plus the actions you took.
+3. **Train** - a policy (the neural network that will control the robot) learns to imitate your demonstrations.
+4. **Deploy** - run the trained policy on the robot and watch it complete the task on its own.
+
+## Get Started
+
+New here? [Install LeRobot](./installation), then pick your path:
+
+<div class="grid grid-cols-1 md:grid-cols-3 gap-4 my-6">
+  <div class="border dark:border-gray-700 rounded-lg p-4 shadow">
+    <div class="text-lg font-semibold mb-2">🔧 I have a robot</div>
+    <p class="text-gray-700 dark:text-gray-300 text-sm">
+      LeRobot supports a wide range of arms and mobile robots. Popular picks:
+    </p>
+    <ul class="text-gray-700 dark:text-gray-300 text-sm list-disc pl-5 mb-2">
+      <li>
+        <a href="./so101">SO-101</a> - our flagship, low-cost arm
+      </li>
+      <li>
+        <a href="./lekiwi">LeKiwi</a> - a mobile base with an arm on top
+      </li>
+      <li>
+        <a href="./koch">Koch v1.1</a> - a long-time community favorite
+      </li>
+      <li>
+        or find yours under <strong>Robots</strong> in the sidebar
+      </li>
+    </ul>
+    <p class="text-gray-700 dark:text-gray-300 text-sm">
+      Once it's assembled and calibrated, record a dataset and train your first
+      policy with the <a href="./il_robots">imitation learning tutorial</a> - or
+      skip the CLI entirely with <a href="./lelab">LeLab</a>, a browser GUI for
+      the same workflow.
+    </p>
+  </div>
+  <div class="border dark:border-gray-700 rounded-lg p-4 shadow">
+    <div class="text-lg font-semibold mb-2">💻 No hardware yet</div>
+    <p class="text-gray-700 dark:text-gray-300 text-sm">
+      You can still train and evaluate policies without owning a robot:
+    </p>
+    <ul class="text-gray-700 dark:text-gray-300 text-sm list-disc pl-5 mb-2">
+      <li>
+        train on an existing
+        <a href="https://huggingface.co/datasets?other=LeRobot">
+          LeRobot dataset
+        </a>
+        from the Hub
+      </li>
+      <li>
+        evaluate in <a href="./envhub">simulation</a>, against benchmarks like
+        LIBERO or Meta-World
+      </li>
+      <li>
+        try the free <a href="./notebooks">Colab notebooks</a> - nothing to
+        install
+      </li>
+    </ul>
+  </div>
+  <div class="border dark:border-gray-700 rounded-lg p-4 shadow">
+    <div class="text-lg font-semibold mb-2">🤝 I want to contribute</div>
+    <p class="text-gray-700 dark:text-gray-300 text-sm">
+      Start with the <a href="./contributing">Contributing guide</a>, then
+      <a href="./bring_your_own_policies">add a new policy</a> or
+      <a href="./integrate_hardware">bring your own hardware</a>.
+    </p>
+  </div>
+</div>
+
+## Explore the Docs
+
+<div class="grid grid-cols-1 md:grid-cols-3 gap-4 my-6">
+  <a
+    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
+    href="./cheat-sheet"
+  >
+    <div class="font-semibold mb-1">📋 Cheat Sheet</div>
+    <p class="text-gray-700 dark:text-gray-300 text-sm">
+      Every LeRobot CLI command, copy-paste ready.
+    </p>
+  </a>
+  <a
+    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
+    href="./hardware_guide"
+  >
+    <div class="font-semibold mb-1">🖥️ Compute & Hardware Guide</div>
+    <p class="text-gray-700 dark:text-gray-300 text-sm">
+      Which policy fits your GPU, and how long training takes.
+    </p>
+  </a>
+  <a
+    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
+    href="./lerobot-dataset-v3"
+  >
+    <div class="font-semibold mb-1">🗂️ LeRobotDataset</div>
+    <p class="text-gray-700 dark:text-gray-300 text-sm">
+      Load, stream, and visualize robot datasets from the Hub.
+    </p>
+  </a>
+  <a
+    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
+    href="./lelab"
+  >
+    <div class="font-semibold mb-1">🖼 LeLab</div>
+    <p class="text-gray-700 dark:text-gray-300 text-sm">
+      A browser GUI for calibrating, recording, and training - no CLI required.
+    </p>
+  </a>
+  <a
+    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
+    href="./act"
+  >
+    <div class="font-semibold mb-1">🧠 Policies</div>
+    <p class="text-gray-700 dark:text-gray-300 text-sm">
+      Start with ACT, our recommended first policy - or browse SmolVLA, π₀, and
+      more in the sidebar.
+    </p>
+  </a>
+  <a
+    class="!no-underline border dark:border-gray-700 rounded-lg p-4 shadow hover:shadow-lg"
+    href="./envhub"
+  >
+    <div class="font-semibold mb-1">🎮 Simulation & Benchmarks</div>
+    <p class="text-gray-700 dark:text-gray-300 text-sm">
+      Train and evaluate in simulated environments before touching real
+      hardware.
+    </p>
+  </a>
+</div>
+
+## Common Problems
+
+Running into issues? A few of the most frequent ones:
+
+- **Blurry or unusable camera footage** - lighting matters more than resolution. See the [Cameras](./cameras) guide.
+- **Build or install errors** (`cmake`, `ffmpeg`, CUDA) - see the Troubleshooting section of the [Installation guide](./installation#troubleshooting).
+- **Not sure which policy fits your GPU** - check the [Compute & Hardware Guide](./hardware_guide).
+- **Still stuck?** Ask on [Discord](https://discord.gg/s3KuuzsPFb) - the community (and the LeRobot team) is there to help.
@@ -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                                             |
-| ------------------------------------ | ------------------------------------------------------- |
-| `--strategy.num_episodes`            | Number of correction episodes to record (default: 10)   |
-| `--strategy.record_autonomous`       | Record autonomous frames too (default: false)           |
-| `--strategy.upload_every_n_episodes` | Push to Hub every N episodes (default: 5)               |
-| `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard) |
-| `--teleop.type`                      | **Required.** Teleoperator type                         |
+| Flag                                 | Description                                                                                                                                                    |
+| ------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `--strategy.num_episodes`            | Number of correction episodes to record (default: 10)                                                                                                          |
+| `--strategy.record_autonomous`       | Record autonomous frames too (default: false)                                                                                                                  |
+| `--strategy.upload_every_n_episodes` | Push to Hub every N episodes (default: 5)                                                                                                                      |
+| `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard)                                                                                                        |
+| `--strategy.smooth_handover`         | Smoothly hand control over at pause / correction start (default: true). Disable for clutch-style teleops that re-reference at the current robot pose on engage |
+| `--teleop.type`                      | **Required.** Teleoperator type                                                                                                                                |

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

 ---

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

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

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

 ## Layer 3 — training format

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

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

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

-If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.
-If an event-scoped branch is selected on a frame without the required event row, rendering returns `None`, allowing a loader to retry another sample.
+If both language columns are missing, `None`, or empty, `RenderMessagesStep` uses
+the task string as low-level supervision when available and otherwise leaves the
+sample unchanged. For an annotated sample, if no recipe branch applies and no
+task fallback exists, rendering returns `None`, allowing a loader to retry another sample.
@@ -142,6 +142,22 @@ repo_id = "yaak-ai/L2D-v3"
 dataset = StreamingLeRobotDataset(repo_id)  # streams directly from the Hub
 ```

+Datasets stored in an [HF Storage Bucket](https://huggingface.co/docs/hub/storage-buckets) (`hf://buckets/`) can be streamed the same way by passing `repo_type="bucket"`:
+
+```python
+dataset = StreamingLeRobotDataset("my-org/my-bucket", repo_type="bucket")
+```
+
+Both options are available in `lerobot-train` through `--dataset.streaming=true`, and `--dataset.repo_type=bucket` to stream from a bucket instead of a Hub dataset repo:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=my-org/my-bucket \
+  --dataset.repo_type=bucket \
+  --dataset.streaming=true \
+  ...
+```
+
 <div style="display:flex; justify-content:center; gap:12px; flex-wrap:wrap;">
  <figure style="margin:0; text-align:center;">
    <img
@@ -92,6 +92,20 @@ LIBERO supports two control modes — `relative` (default) and `absolute`. Diffe
 --env.control_mode=relative   # or "absolute"
 ```

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

 **Observations:**
@@ -114,38 +128,58 @@ 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 \
-  --env.type=libero \
-  --env.task=libero_10 \
-  --output_dir=./outputs/ \
+  --policy.push_to_hub=false \
+  --dataset.repo_id=lerobot/libero \
+  --dataset.video_backend=torchcodec \
+  --output_dir=./outputs/libero_smolvla \
  --steps=100000 \
-  --batch_size=4 \
-  --eval.batch_size=1 \
-  --eval.n_episodes=1 \
-  --env_eval_freq=1000
+  --batch_size=64
 ```

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

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

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

 **Observations:**
@@ -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 \
-    --job_name=pi05_training \
-    --policy.repo_id=your_repo_id \
-    --policy.pretrained_path=lerobot/pi05_base \
-    --policy.compile_model=true \
-    --policy.gradient_checkpointing=true \
-    --wandb.enable=true \
-    --policy.dtype=bfloat16 \
+    --policy.pretrained_path=lerobot/pi05_libero_base \
+    --policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}' \
+    --policy.n_action_steps=10 \
+    --policy.empty_cameras=1 \
    --policy.freeze_vision_encoder=false \
    --policy.train_expert_only=false \
-    --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.
@@ -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)
-docker build -f docker/Dockerfile.eval-base -t lerobot-eval-base .
-
-# Build RoboMME eval image (applies gymnasium + numpy pin overrides)
-docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-robomme .
+# Build the RoboMME evaluation image from the repo root
+docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-benchmark-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

@@ -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.

@@ -0,0 +1,339 @@
+# Third-Party Robots & Teleoperators
+
+The LeRobot ecosystem extends far beyond its officially supported hardware. Thanks to LeRobot's plugin architecture, the community has built integrations for a wide range of robot arms and teleoperation devices — from industrial manipulators to affordable hobbyist platforms, VR headsets, haptic devices, and full arm-plus-teleoperator kits. This page showcases community-maintained integrations you can use for teleoperation, data collection, and policy deployment.
+
+> [!IMPORTANT]
+> These projects are developed and maintained by third parties. Please refer to each repository for installation instructions, hardware requirements, and support.
+
+Drop-in plugins are auto-discovered by package name: LeRobot imports any installed package prefixed with `lerobot_robot_` or `lerobot_teleoperator_`. Once installed, reference the `type` the plugin registers (see its README — it may differ from the package name) directly from any LeRobot command:
+
+```bash
+pip install lerobot_robot_<name> lerobot_teleoperator_<name>
+
+lerobot-record \
+    --robot.type=<robot_name> \
+    --teleop.type=<teleoperator_name> \
+    --dataset.repo_id=${HF_USER}/my-dataset \
+    --dataset.num_episodes=5
+```
+
+> [!TIP]
+> ⚠️ marks projects that are forks/extensions of LeRobot. They may require custom setup rather than working with an unmodified install. All other entries are drop-in plugins.
+
+## Industrial & Collaborative Arms
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/SpesRobotics/lerobot-robot-xarm">lerobot-robot-xarm</a></td>
+      <td>Plugin for the xArm collaborative arm series from <a href="https://www.ufactory.cc/">UFACTORY</a>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/lebai-robotics/lerobot_lebai">lerobot_lebai</a></td>
+      <td>Plugin for the six-axis collaborative arms from <a href="https://lebai.ltd/en/">Lebai</a>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/wengmister/LeFranX">LeFranX</a> ⚠️</td>
+      <td>LeRobot extension for the <a href="https://franka.de/">Franka</a> research arm, paired with the <a href="https://www.robotera.com/">RobotEra XHand</a> hand for VR teleoperation.</td>
+    </tr>
+  </tbody>
+</table>
+
+#### Universal Robots UR5e
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/yechen056/UR5e-LeRobot">UR5e-LeRobot</a> ⚠️</td>
+      <td>LeRobot extension for the <a href="https://www.universal-robots.com/">Universal Robots UR5e</a>, with single-arm and bimanual support.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/scy-v/lerobot_ur5e_auto">lerobot_ur5e_auto</a> ⚠️</td>
+      <td>LeRobot extension for a mobile <a href="https://www.universal-robots.com/">Universal Robots UR5e</a>, adding automated recording at scale with minimal supervision.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/F-Fer/lerobot_ur5e_gello">lerobot_robot_ur5e</a></td>
+      <td>Plugin for the <a href="https://www.universal-robots.com/">Universal Robots UR5e</a> with a <a href="https://robotiq.com/">Robotiq</a> gripper, over RTDE control.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Research & Learning Arms
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/TrossenRobotics/lerobot_trossen">lerobot_trossen</a></td>
+      <td>Plugin for the WidowX and ALOHA-style arms from <a href="https://www.trossenrobotics.com/">Trossen Robotics</a>.</td>
+    </tr>
+  </tbody>
+</table>
+
+#### AgileX Piper
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/AgRoboticsResearch/lerobot_robot_piper">lerobot_robot_piper (AgRobotics Research)</a></td>
+      <td>Plugin for the <a href="https://global.agilex.ai/">AgileX Piper</a> arm.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/WeGo-Robotics/lerobot_robot_piper">lerobot_robot_piper (WeGo Robotics)</a></td>
+      <td>Plugin for the <a href="https://global.agilex.ai/">AgileX Piper</a> arm, with multi-arm teleoperation, safety limits, and GUI tools.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Affordable & Hobbyist Arms
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/servodevelop/fashionstar-lerobot-robot-cello">fashionstar-lerobot-robot-cello</a></td>
+      <td>Plugin for the StarAI Cello 6+1 degrees of freedom robot arm from <a href="https://fashionstar.com.hk/">FashionStar</a>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/servodevelop/fashionstar-lerobot-robot-viola">fashionstar-lerobot-robot-viola</a></td>
+      <td>Plugin for the compact StarAI Viola 6+1 degrees of freedom robot arm from <a href="https://fashionstar.com.hk/">FashionStar</a>.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Service, Mobile & Utility Robots
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/ugo-plus/lerobot-robot-ugo-pro">lerobot-robot-ugo-pro</a></td>
+      <td>Plugin for the ugo Pro dual-arm service robot from <a href="https://ugo.plus/products/ugo-pro/">ugo</a>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/zuoxingdong/lerobot_robot_lekiwi_pincopen">lerobot_robot_lekiwi_pincopen</a></td>
+      <td>Plugin for a LeKiwi mobile manipulator with a <a href="https://github.com/pollen-robotics/PincOpen">PincOpen</a> gripper.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/KillingJacky/lerobot-robot-dummy">lerobot-robot-dummy</a></td>
+      <td>Plugin simulating a robot for recording without hardware. Useful for debugging !</td>
+    </tr>
+  </tbody>
+</table>
+
+## Teleoperators
+
+### VR & Motion Controllers
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/SpesRobotics/lerobot-teleoperator-teleop">lerobot-teleoperator-teleop</a></td>
+      <td>Plugin turning a phone or VR headset into a teleoperator via <a href="https://immersiveweb.dev">WebXR</a>, wrapping the open-source <a href="https://github.com/SpesRobotics/teleop"><code>teleop</code></a> library.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/Jas000n/lerobot-teleoperator-spacemouse">lerobot-teleoperator-spacemouse</a></td>
+      <td>Plugin for the <a href="https://3dconnexion.com/">3Dconnexion SpaceMouse</a>, with inverse kinematics for SO-ARMS robots.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/Dream-Machines-Robotics/vr-teleop-kit">vr-teleop-kit</a></td>
+      <td>Plugin teleoperating arms from a <a href="https://www.meta.com/quest/">Meta Quest</a> (WebXR), relying on URDF descriptions for inverse kinematics.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/xensedyl/lerobot-teleoperator-pico4">lerobot-teleoperator-pico4</a></td>
+      <td>Plugin for the <a href="https://www.picoxr.com/">PICO 4</a> VR headset, with a companion controller-free <a href="https://github.com/xensedyl/lerobot-teleoperator-pico4-hand">hand-tracking variant</a>.</td>
+    </tr>
+  </tbody>
+</table>
+
+### Leader Arms
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/F-Fer/lerobot_ur5e_gello">lerobot_teleoperator_gello</a></td>
+      <td>Plugin for the 7 degrees of freedom <a href="https://wuphilipp.github.io/gello_site/">GELLO</a> teleoperator.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/uynitsuj/lerobot_teleoperator_yamactiveleader">lerobot_teleoperator_yamactiveleader</a></td>
+      <td>Plugin for the active YAM teleoperator from <a href="https://i2rt.com/">I2RT</a>, a bilateral force-feedback arm.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/charlie8612/lerobot_teleoperator_omy">lerobot_teleoperator_omy</a></td>
+      <td>Plugin for the OMY-L100 6 degrees of freedom teleoperator from <a href="https://www.robotis.com/">ROBOTIS</a>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://pypi.org/project/lerobot-teleoperator-pipermate/">lerobot-teleoperator-pipermate</a></td>
+      <td>Plugin for the PiperMate teleoperator (<a href="https://fashionstar.com.hk/">FashionStar</a> UART servos), driving the <a href="https://global.agilex.ai/">AgileX Piper</a> arm.</td>
+    </tr>
+  </tbody>
+</table>
+
+### Haptic Devices
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/chohh7391/lerobot_teleoperator_inverse3">lerobot_teleoperator_inverse3</a></td>
+      <td>Plugin for the <a href="https://www.haply.co/">Haply Inverse3</a> haptic device, adding force-feedback teleoperation.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/hzhz112/lerobot_teleoperator_omega7">lerobot_teleoperator_omega7</a></td>
+      <td>Plugin for the <a href="https://www.forcedimension.com/">Force Dimension omega.7</a> haptic device, adding force-feedback teleoperation.</td>
+    </tr>
+  </tbody>
+</table>
+
+### Networked & Remote
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://pypi.org/project/lerobot-teleoperator-livekit/">lerobot-teleoperator-livekit</a></td>
+      <td>Plugin receiving teleoperation commands over a <a href="https://livekit.io/">LiveKit</a> Portal (WebRTC) for remote control.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Full Kits (Robot + Teleoperator)
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/villekuosmanen/lerobot-arx5">lerobot-arx5</a></td>
+      <td>Plugin for the <a href="https://www.arx-x.com/">ARX5</a> arm: <a href="https://pypi.org/project/lerobot-robot-arx5/"><code>lerobot-arx5</code></a> robot arm with its <a href="https://pypi.org/project/lerobot-teleoperator-arx5/"><code>lerobot-teleoperator-arx5</code></a> teleoperator arm.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/robertorobotics/Nextis-AIRA-3D">Nextis-AIRA-3D</a></td>
+      <td>Plugin for the 7 degrees of freedom arm from <a href="https://www.nextis.tech">Nextis</a>: robot arm <code>aira_follower</code> and teleoperator arm <code>aira_leader</code>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/pravsels/lerobot_yam">lerobot_yam</a></td>
+      <td>Plugin suite for the YAM arm from <a href="https://i2rt.com/">I2RT</a>: robot arm <code>yam_follower</code> and teleoperator arm <code>yam_leader</code>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/robot-learning-co/trlc-dk1">trlc-dk1</a></td>
+      <td>Plugin for the development kit from <a href="https://www.robot-learning.co/">The Robot Learning Company</a>: single and bimanual arms follower/teleoperator types.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/hexfellow/hex_lerobot_drivers">hex_lerobot_drivers</a></td>
+      <td>Plugin suite for <a href="https://hexfellow.com/">HEXFELLOW</a> devices: robots, teleoperators, and cameras (see <a href="./third_party_sensors">Cameras &amp; Sensors</a>).</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/Hiwonder-official/lerobot-robot-nexarm-follower">lerobot-robot-nexarm-follower</a></td>
+      <td>Plugin for the NexArm from <a href="https://www.hiwonder.com/">Hiwonder</a>: the <a href="https://github.com/Hiwonder-official/lerobot-robot-nexarm-follower">robot arm</a> and its matching <a href="https://github.com/Hiwonder-official/lerobot-teleoperator-nexarm-leader">teleoperator arm</a>.</td>
+    </tr>
+  </tbody>
+</table>
+
+## ROS 2 Bridges
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/ngres/leros2">leros2</a></td>
+      <td>Plugin bridging ROS 2 topics and actions to LeRobot robots and teleoperators.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/ROBOTIS-GIT/lerobot_robot_ros2_zenoh">lerobot_robot_ros2_zenoh</a></td>
+      <td>Plugin bridging ROS 2 robots to LeRobot over <a href="https://zenoh.io">Zenoh</a> pub/sub transport.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Contributing
+
+Built your own LeRobot hardware integration? The plugin system makes it straightforward to add new robots and teleoperators — check out the [Bring Your Own Hardware](./integrate_hardware) guide to get started, and share your project with the community!
@@ -0,0 +1,99 @@
+# Third-Party Cameras & Sensors
+
+The LeRobot ecosystem extends far beyond its natively supported cameras (OpenCV, Intel RealSense, ZMQ, Reachy 2). Thanks to LeRobot's plugin architecture, the community has built drop-in camera and sensor integrations — from depth cameras to vision-based tactile sensors. This page showcases community-maintained camera and sensor integrations you can use for teleoperation, data collection, and policy deployment.
+
+> [!IMPORTANT]
+> These projects are developed and maintained by third parties. Please refer to each repository for installation instructions, hardware requirements, and support.
+
+Drop-in plugins are auto-discovered by package name: LeRobot imports any installed package prefixed with `lerobot_camera_`. Once installed, reference the camera `type` the plugin registers (see its README — it may differ from the package name) directly from any LeRobot command:
+
+```bash
+pip install lerobot_camera_<name>
+
+lerobot-record \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.cameras="{ front: {type: <name>, width: 640, height: 480, fps: 30} }" \
+    --dataset.repo_id=${HF_USER}/my-dataset \
+    --dataset.num_episodes=5
+```
+
+## Tactile Sensors
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/xensedyl/lerobot-camera-xense">lerobot-camera-xense</a></td>
+      <td>Plugin for <a href="https://www.xenserobotics.com/">Xense</a> vision-based tactile sensors, exposing rectified/difference images, depth, and 2D markers.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Depth Cameras
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/hexfellow/hex_lerobot_drivers/tree/main/lerobot_camera_berxel">lerobot_camera_berxel</a></td>
+      <td>Plugin for the <a href="https://www.berxel.com/">Berxel</a> depth camera, part of the broader <a href="https://hexfellow.com/">HEXFELLOW</a> driver suite.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Networked Cameras
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/F-Fer/lerobot_ur5e_gello">lerobot_camera_zmq</a></td>
+      <td>Plugin streaming <a href="https://www.stereolabs.com/">Stereolabs ZED</a> and USB camera frames from a Raspberry Pi over the network.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Virtual Cameras
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/hexfellow/hex_lerobot_drivers/tree/main/lerobot_camera_dummy">lerobot_camera_dummy</a></td>
+      <td>Plugin simulating a camera for recording without hardware. Useful for debugging !</td>
+    </tr>
+  </tbody>
+</table>
+
+## Contributing
+
+Built your own LeRobot camera or sensor integration? Package it as an installable `lerobot_camera_<name>` plugin and it will be auto-discovered by the LeRobot CLI — see the [Bring Your Own Hardware](./integrate_hardware) guide and the [Cameras](./cameras) reference to get started, then share your project with the community!
@@ -290,47 +290,6 @@ lerobot-rollout \

 ---

-## Part 5: SONIC Whole-Body Control (Latent-Token Policies)
-
-LeRobot now ports the **decoder** half of NVIDIA's [SONIC](https://github.com/NVIDIA/sonic) whole-body deploy stack as the `SonicWholeBodyController`. Instead of commanding joints, the policy outputs a 64-D **latent motion token** (`motion_token.{i}.pos`) each tick; the SONIC decoder maps that token plus recent proprioception history onto a residual action that is scaled and added to the standing pose, producing 50 Hz joint-position targets for all 29 DoF.
-
-The controller loads its ONNX decoder and all deploy constants (PD gains, `default_angles`, `action_scale`, and the neutral token) from [`lerobot/sonic_decoder`](https://huggingface.co/lerobot/sonic_decoder). The repo also ships a distilled low-latency decoder, selectable with the controller's `policy_type="low_latency"` argument (defaults to `"default"`).
-
-### Test in Simulation
-
-Any token-output policy trained on SONIC motion tokens (e.g. [`nepyope/sonic_walk`](https://huggingface.co/nepyope/sonic_walk)) can drive it:
-
-```bash
-lerobot-rollout \
-  --strategy.type=base \
-  --policy.path=nepyope/sonic_walk \
-  --policy.device=cuda \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=true \
-  --robot.controller=SonicWholeBodyController \
-  --robot.cameras='{"ego_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30, "warmup_s": 15}}'
-```
-
-### Run on the Physical Robot
-
-Start the robot server (see Part 2), then point the rollout at the robot instead of the sim:
-
-```bash
-lerobot-rollout \
-  --strategy.type=base \
-  --policy.path=nepyope/sonic_walk \
-  --policy.device=cuda \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=false \
-  --robot.robot_ip=<ROBOT_IP> \
-  --robot.controller=SonicWholeBodyController \
-  --robot.cameras='{"ego_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30, "warmup_s": 5}}'
-```
-
-On connect, the robot eases every joint to the decoder's `default_angles` before the controller takes over, so the first policy commands don't snap from the connect-time pose.
-
---
-
 ## Additional Resources

 - [Unitree SDK Documentation](https://github.com/unitreerobotics/unitree_sdk2_python)
@@ -341,4 +300,4 @@ On connect, the robot eases every joint to the decoder's `default_angles` before

 ---

-_Last updated: July 2026_
+_Last updated: March 2026_
@@ -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)
-7. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
-8. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
+6. **Modify Tasks** - Change the natural-language task descriptions associated with episodes
+7. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
+8. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
+9. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.

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

 ---

@@ -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

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"

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

    # Config & Hub
-    "draccus==0.10.0", # TODO: Relax version constraint
-    "huggingface-hub>=1.0.0,<2.0.0",
+    "draccus>=0.11.6,<0.12.0",
+    "huggingface-hub>=1.6.0,<2.0.0",
    "requests>=2.32.0,<3.0.0",

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

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

 # Optional dependencies
@@ -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]",
@@ -261,7 +261,7 @@ annotations = [
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools>=1.73.1,<2.0.0", "mypy>=1.19.1", "ruff>=0.14.1", "lerobot[notebook]"]
 notebook = ["jupyter>=1.0.0,<2.0.0", "ipykernel>=6.0.0,<7.0.0"]
-test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
+test = ["pytest>=8.1.0,<10.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]

 # Simulation
@@ -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
@@ -109,6 +109,11 @@ class RealSenseCamera(Camera):
        ```
    """

+    # Maximum number of warmup attempts made by connect(). A failed attempt is first
+    # retried with a plain pipeline stop/start, which is usually enough to recover the
+    # stream; a USB hardware reset is performed before the final attempt as a last resort.
+    _MAX_CONNECT_ATTEMPTS = 3
+
    def __init__(self, config: RealSenseCameraConfig):
        """
        Initializes the RealSenseCamera instance.
@@ -173,6 +178,76 @@ class RealSenseCamera(Camera):
        """Checks if the camera pipeline is started and streams are active."""
        return self.rs_pipeline is not None and self.rs_profile is not None

+    def _hardware_reset(self, wait_s: float = 5.0) -> None:
+        """Issue a USB hardware reset to recover an unresponsive device (common on D405)."""
+        context = rs.context()
+        for device in context.query_devices():
+            if device.get_info(rs.camera_info.serial_number) == self.serial_number:
+                logger.info(f"{self} performing hardware reset.")
+                device.hardware_reset()
+                time.sleep(wait_s)
+                return
+        logger.warning(f"{self} device not found for hardware reset, skipping.")
+
+    def _open_pipeline(self) -> None:
+        """Initializes the RealSense pipeline, starts it, and starts the background read thread.
+
+        Raises:
+            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.
+        """
+        rs_pipeline = rs.pipeline()
+        rs_config = rs.config()
+        self._configure_rs_pipeline_config(rs_config)
+
+        try:
+            rs_profile = rs_pipeline.start(rs_config)
+        except RuntimeError as e:
+            raise ConnectionError(
+                f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
+            ) from e
+
+        self.rs_pipeline = rs_pipeline
+        self.rs_profile = rs_profile
+
+        try:
+            self._configure_capture_settings()
+            self._configure_sensor_options()
+            self._start_read_thread()
+        except BaseException:
+            self._release_after_failed_setup()
+            raise
+
+    def _run_warmup(self) -> None:
+        """Blocks until at least one valid frame has been captured by the background thread.
+
+        Raises:
+            ConnectionError: If no frame arrives before ``warmup_s`` elapses.
+        """
+        # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
+        self.warmup_s = max(self.warmup_s, 1)
+
+        warmup_read = self.async_read if self.use_rgb else self.async_read_depth
+        start_time = time.time()
+        while time.time() - start_time < self.warmup_s:
+            warmup_read(timeout_ms=self.warmup_s * 1000)
+            time.sleep(0.1)
+        with self.frame_lock:
+            if (self.use_rgb and self.latest_color_frame is None) or (
+                self.use_depth and self.latest_depth_frame is None
+            ):
+                raise ConnectionError(f"{self} failed to capture frames during warmup.")
+
+    def _release_after_failed_setup(self) -> None:
+        """Releases the device handle and restores auto-detected settings after a failed attempt."""
+        try:
+            self._cleanup_resources()
+        except Exception:
+            logger.exception(f"Failed to fully clean up {self} after connect() failed.")
+        self._reset_connection_settings()
+
    @check_if_already_connected
    def connect(self, warmup: bool = True) -> None:
        """
@@ -181,58 +256,53 @@ class RealSenseCamera(Camera):
        Initializes the RealSense pipeline, configures the required streams (color
        and optionally depth), starts the pipeline, and validates the actual stream settings.

+        If the pipeline starts but no frames arrive during warmup, retries up to
+        ``_MAX_CONNECT_ATTEMPTS`` times, performing a USB hardware reset before the
+        final attempt.
+
        Args:
            warmup (bool): If True, waits at connect() time until at least one valid frame
                           has been captured by the background thread. Defaults to True.

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

-        self.rs_pipeline = rs.pipeline()
-        rs_config = rs.config()
-        self._configure_rs_pipeline_config(rs_config)
+        if not warmup:
+            self._open_pipeline()
+            logger.info(f"{self} connected.")
+            return

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

-        try:
-            self._configure_capture_settings()
-            self._configure_sensor_options()
-            self._start_read_thread()
+        for attempt in range(1, self._MAX_CONNECT_ATTEMPTS + 1):
+            if attempt == self._MAX_CONNECT_ATTEMPTS:
+                self._hardware_reset()

-            # 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._open_pipeline()

-            warmup_read = self.async_read if self.use_rgb else self.async_read_depth
-            start_time = time.time()
-            while time.time() - start_time < self.warmup_s:
-                warmup_read(timeout_ms=self.warmup_s * 1000)
-                time.sleep(0.1)
-            with self.frame_lock:
-                if (self.use_rgb and self.latest_color_frame is None) or (
-                    self.use_depth and self.latest_depth_frame is None
-                ):
-                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
-        except BaseException:
+            connected = False
            try:
-                self._cleanup_resources()
-            except Exception:
-                logger.exception(f"Failed to fully clean up {self} after connect() failed.")
-            self._reset_connection_settings()
-            raise
+                self._run_warmup()
+                connected = True
+            except (TimeoutError, ConnectionError) as e:
+                last_error = e
+            finally:
+                if not connected:
+                    self._release_after_failed_setup()

-        logger.info(f"{self} connected.")
+            if connected:
+                logger.info(f"{self} connected.")
+                return
+
+            logger.warning(f"{self} warmup failed (attempt {attempt}/{self._MAX_CONNECT_ATTEMPTS}).")
+
+        raise ConnectionError(
+            f"{self} failed to capture frames after {self._MAX_CONNECT_ATTEMPTS} attempts."
+        ) from last_error

    @staticmethod
    def find_cameras() -> list[dict[str, Any]]:
@@ -365,11 +435,12 @@ class RealSenseCamera(Camera):
        return self._async_read(timeout_ms=10000, read_depth=read_depth)

    def _get_color_sensor(self) -> "rs.sensor":
-        """Returns the sensor that controls the color stream.
+        """Returns the dedicated "RGB Camera" sensor that controls the color stream.

-        Most RealSense cameras expose "RGB Camera" for color. The D405 has no
-        separate RGB module — its color stream comes from "Stereo Module".
-        We try RGB Camera first, then fall back to Stereo Module.
+        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.")
@@ -377,12 +448,14 @@ class RealSenseCamera(Camera):
        device = self.rs_profile.get_device()
        sensors = {s.get_info(rs.camera_info.name): s for s in device.query_sensors()}

-        for name in ("RGB Camera", "Stereo Module"):
-            if name in sensors:
-                return sensors[name]
+        if "RGB Camera" in sensors:
+            return sensors["RGB Camera"]

        available = list(sensors.keys())
-        raise RuntimeError(f"{self}: no color sensor found. Available sensors: {available}")
+        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."""
@@ -626,6 +699,9 @@ class RealSenseCamera(Camera):
                capture_time = time.perf_counter()

                with self.frame_lock:
+                    # Under the lock, so a late frame cannot resurrect the buffer _stop_read_thread() cleared.
+                    if stop_event.is_set():
+                        break
                    if self.use_rgb:
                        self.latest_color_frame = processed_color_frame
                    if self.use_depth:
@@ -836,4 +912,5 @@ class RealSenseCamera(Camera):
            )

        self._cleanup_resources()
+
        logger.info(f"{self} disconnected.")
@@ -102,6 +102,7 @@ class ImageServer:
                fps=self.fps,
                width=shape[1],
                height=shape[0],
+                fourcc=cfg.get("fourcc", "MJPG"),
                color_mode=ColorMode.RGB,
            )
            camera = OpenCVCamera(cam_config)
@@ -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:
@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

+import logging
 from dataclasses import dataclass, field

 from lerobot.transforms import ImageTransformsConfig
@@ -21,6 +22,8 @@ from lerobot.utils.import_utils import get_safe_default_video_backend

 from .video import DEFAULT_DEPTH_UNIT, DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT

+logger = logging.getLogger(__name__)
+

@dataclass
 class DatasetConfig:
@@ -29,10 +32,15 @@ class DatasetConfig:
    # "dataset_index" into the returned item. The index mapping is made according to the order in which the
    # datasets are provided.
    repo_id: str
+    # Hub repository type: "dataset" (default) or "bucket" for an HF Storage Bucket streamed over
+    # hf://buckets/. Buckets are streaming-only, so "bucket" requires streaming=true.
+    repo_type: str = "dataset"
    # Root directory for a concrete local dataset tree (e.g. 'dataset/path'). If None, local datasets are
    # looked up under $HF_LEROBOT_HOME/repo_id and Hub downloads use a revision-safe cache under $HF_LEROBOT_HOME/hub.
    root: str | None = None
    episodes: list[int] | None = None
+    # Episode indices to drop (e.g. corrupt or heterogeneous ones). Applied on top of `episodes`.
+    exclude_episodes: list[int] | None = None
    image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)
    revision: str | None = None
    use_imagenet_stats: bool = True
@@ -48,6 +56,16 @@ class DatasetConfig:
    eval_split: float = 0.0

    def __post_init__(self) -> None:
+        if self.repo_type not in ("dataset", "bucket"):
+            raise ValueError(f"repo_type must be 'dataset' or 'bucket', got {self.repo_type!r}")
+        if self.repo_type == "bucket" and not self.streaming:
+            raise ValueError(
+                "repo_type='bucket' is streaming-only: set streaming=true to train from an HF Storage Bucket."
+            )
+        if self.repo_type == "bucket" and self.eval_split != 0.0:
+            raise ValueError(
+                "eval_split requires map-style datasets and is not supported with repo_type='bucket'."
+            )
        if self.depth_output_unit not in (DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT):
            raise ValueError(
                f"depth_output_unit must be '{DEPTH_METER_UNIT}' or '{DEPTH_MILLIMETER_UNIT}', got {self.depth_output_unit!r}"
@@ -62,6 +80,14 @@ class DatasetConfig:
            if len(self.episodes) != len(set(self.episodes)):
                duplicates = sorted({ep for ep in self.episodes if self.episodes.count(ep) > 1})
                raise ValueError(f"Episode indices contain duplicates: {duplicates}")
+        if self.exclude_episodes is not None:
+            negative_episodes = [episode for episode in self.exclude_episodes if episode < 0]
+            if negative_episodes:
+                logger.warning(
+                    "Ignoring negative exclude_episodes entries: %s",
+                    negative_episodes,
+                )
+                self.exclude_episodes = [episode for episode in self.exclude_episodes if episode >= 0]


@dataclass
@@ -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}=")
-    for arg in args:
-        if arg.startswith(detect_string) and not arg.startswith(exclude_strings):
-            denested_arg = f"--{arg.removeprefix(detect_string)}"
-            attr_level_args.append(denested_arg)
+    excluded_names = (draccus.CHOICE_TYPE_KEY, PATH_KEY)
+    for index, arg in enumerate(args):
+        if not arg.startswith(detect_string):
+            continue
+
+        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}="
-    for arg in args:
-        if arg.startswith(prefix):
-            return arg[len(prefix) :]
+    option = f"--{arg_name}"
+    for index, arg in enumerate(args):
+        if arg.startswith(f"{option}="):
+            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:
-        if "=" in arg and plugin_arg_suffix in arg:
-            key, value = arg.split("=", 1)
-            # Remove leading '--' if present
-            if key.startswith("--"):
-                key = key[2:]
-            plugin_args[key] = value
+    for index, arg in enumerate(args):
+        if not arg.startswith("--"):
+            continue
+
+        key, separator, value = arg[2:].partition("=")
+        if plugin_arg_suffix not in key:
+            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:
-                    cli_args = filter_arg("config_path", cli_args)
-                    cfg = argtype.from_pretrained(config_path_cli, cli_args=cli_args)
-                else:
-                    if config_path_cli:
+                try:
+                    if has_method(argtype, "from_pretrained") and config_path_cli:
                        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,
-                    )
+                        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(
+                            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"):
-            draccus.dump(self, f, indent=4)
+        # Encode against the base class so draccus includes the choice "type" key,
+        # 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(
@@ -23,6 +23,7 @@ from typing import Any, Literal, get_args

 MessageRole = Literal["user", "assistant", "system", "tool"]
 MessageStream = Literal["high_level", "low_level"]
+RecipeRoute = Literal["vqa"]

 DEFAULT_BINDINGS = {
    "subtask": "active_at(t, style=subtask)",
@@ -40,6 +41,7 @@ discovery (here) and rendered-message substitution (in ``language_render``)."""

 _VALID_ROLES = frozenset(get_args(MessageRole))
 _VALID_STREAMS = frozenset(get_args(MessageStream))
+_VALID_ROUTES = frozenset(get_args(RecipeRoute))


@dataclass
@@ -78,7 +80,7 @@ class MessageTurn:
            raise ValueError(f"Unsupported message stream: {self.stream!r}")
        if self.content is None and self.tool_calls_from is None:
            raise ValueError("MessageTurn.content is required unless tool_calls_from is set.")
-        if self.content is not None and not isinstance(self.content, (str, list)):
+        if self.content is not None and not isinstance(self.content, str | list):
            raise TypeError("MessageTurn.content must be a string, a list of HF-style blocks, or None.")
        if isinstance(self.content, list):
            for block in self.content:
@@ -99,13 +101,16 @@ class TrainingRecipe:

    A recipe is either a *message recipe* (``messages`` plus optional
    ``bindings``) or a *blend recipe* (``blend`` mapping names to weighted
-    sub-recipes). ``weight`` is only meaningful inside a blend.
+    sub-recipes). ``weight`` and ``route`` are only meaningful inside a blend;
+    ``route: vqa`` gives sparse VQA annotations priority over normal weighted
+    selection.
    """

    messages: list[MessageTurn] | None = None
    bindings: dict[str, str] | None = None
    blend: dict[str, TrainingRecipe] | None = None
    weight: float | None = None
+    route: RecipeRoute | None = None

    def __post_init__(self) -> None:
        """Validate that exactly one of ``messages`` or ``blend`` is set."""
@@ -113,6 +118,10 @@ class TrainingRecipe:
            raise ValueError("TrainingRecipe must set only one of messages or blend.")
        if self.messages is None and self.blend is None:
            raise ValueError("TrainingRecipe must set one of messages or blend.")
+        if self.route is not None and self.route not in _VALID_ROUTES:
+            raise ValueError(f"Unsupported recipe route: {self.route!r}")
+        if self.blend is not None and self.route is not None:
+            raise ValueError("TrainingRecipe.route may only be set on a message recipe inside a blend.")

        if self.messages is not None:
            self._validate_message_recipe()
@@ -147,8 +156,9 @@ class TrainingRecipe:
        return cls.from_dict(data)

    def _validate_message_recipe(self) -> None:
-        """Ensure every templated binding is known and at least one turn is a target."""
-        assert self.messages is not None
+        """Validate bindings and require text or low-level action supervision."""
+        if self.messages is None:
+            raise ValueError("Cannot validate a message recipe without messages.")
        known_bindings = set(DEFAULT_BINDINGS) | set(self.bindings or {}) | {"task"}

        for turn in self.messages:
@@ -156,12 +166,19 @@ class TrainingRecipe:
            if missing:
                raise ValueError(f"MessageTurn references unknown binding(s): {sorted(missing)}")

-        if not any(turn.target for turn in self.messages):
-            raise ValueError("Message recipes must contain at least one target turn.")
+        has_target = any(turn.target for turn in self.messages)
+        has_low_level = any(turn.stream == "low_level" for turn in self.messages)
+        if not (has_target or has_low_level):
+            raise ValueError(
+                "Message recipes must contain at least one supervised turn — "
+                "either ``target: true`` (text CE) or ``stream: low_level`` "
+                "(flow/action loss)."
+            )

    def _validate_blend_recipe(self) -> None:
        """Ensure each blend component is a non-empty, weighted message recipe."""
-        assert self.blend is not None
+        if self.blend is None:
+            raise ValueError("Cannot validate a blend recipe without blend components.")
        if not self.blend:
            raise ValueError("Blend recipes must contain at least one component.")

@@ -0,0 +1,16 @@
+# Predicts subtasks from tasks and trains subtask-conditioned action flow without memory or plans.
+# Requires `subtask` annotations; samples with missing `if_present` bindings do not render.
+
+blend:
+
+  high_level_subtask:
+    weight: 0.30
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+
+  low_level_execution:
+    weight: 0.70
+    messages:
+      # The low-level stream trains action flow on the generated or annotated subtask.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
@@ -0,0 +1,13 @@
+# Paper-style joint sequence (pi0.5 §IV-B): one sample supervises the subtask
+# text with CE and, because the assistant turn is part of the prefix, conditions
+# the FAST and flow action losses on the same annotated subtask in one forward.
+# The supervised span is attended causally; the action losses see task + subtask.
+#
+# Pair with `--policy.joint_subtask_conditioning=true` at inference so the flow
+# prefix reproduces this layout (task turn with state + causal generated subtask).
+# Samples without a `subtask` annotation fall back to a plain task-prompt
+# low-level sample via `if_present`.
+
+messages:
+  - {role: user, content: "${task}", stream: low_level}
+  - {role: assistant, content: "${subtask}", stream: low_level, target: true, if_present: subtask}
@@ -0,0 +1,30 @@
+# Trains subtask prediction, subtask-conditioned action flow, and memory updates without plans.
+# Requires `subtask` and `memory`; missing `if_present` bindings skip the affected sub-recipe.
+
+blend:
+
+  high_level_subtask:
+    weight: 0.25
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+
+  low_level_execution:
+    weight: 0.60
+    messages:
+      # The low-level stream trains action flow on the generated or annotated subtask.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  memory_update:
+    # `active_at` densifies sparse boundaries while preserving the prior-memory/subtask mapping.
+    # Inference controls update timing through `subtask_change` events.
+    weight: 0.15
+    bindings:
+      prior_memory: "nth_prev(style=memory, offset=1)"
+      current_memory: "active_at(t, style=memory)"
+      completed_subtask: "nth_prev(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "Previous memory: ${prior_memory}", stream: high_level, if_present: prior_memory}
+      - {role: user, content: "Completed subtask: ${completed_subtask}", stream: high_level, if_present: completed_subtask}
+      - {role: assistant, content: "${current_memory}", stream: high_level, target: true, if_present: current_memory}
@@ -0,0 +1,72 @@
+# Adds memory, spoken interjection responses, and camera-grounded VQA to subtask/action training.
+# Missing optional annotations skip only their sub-recipe; `say` tool calls tokenize as `<say>...</say>`.
+
+blend:
+
+  high_level_subtask:
+    weight: 0.25
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+
+  low_level_execution:
+    weight: 0.40
+    messages:
+      # The low-level stream trains action flow on the generated or annotated subtask.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  memory_update:
+    # `active_at` densifies sparse boundaries while preserving the prior-memory/subtask mapping.
+    # Inference controls update timing through `subtask_change` events.
+    weight: 0.10
+    bindings:
+      prior_memory: "nth_prev(style=memory, offset=1)"
+      current_memory: "active_at(t, style=memory)"
+      completed_subtask: "nth_prev(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "Previous memory: ${prior_memory}", stream: high_level, if_present: prior_memory}
+      - {role: user, content: "Completed subtask: ${completed_subtask}", stream: high_level, if_present: completed_subtask}
+      - {role: assistant, content: "${current_memory}", stream: high_level, target: true, if_present: current_memory}
+
+  user_interjection_response:
+    weight: 0.10
+    bindings:
+      interjection: "emitted_at(t, style=interjection)"
+      speech: "emitted_at(t, role=assistant, tool_name=say)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: user, content: "${interjection}", stream: high_level, if_present: interjection}
+      # The assistant target is a `say` tool call flattened to a `<say>...</say>` marker.
+      - {role: assistant, stream: high_level, target: true, if_present: speech, tool_calls_from: speech}
+
+  # Each camera uses a separate VQA sub-recipe for view-specific binding.
+  ask_vqa_top:
+    weight: 0.075
+    route: vqa
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.front)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.front)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.front}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
+
+  ask_vqa_wrist:
+    weight: 0.075
+    route: vqa
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.wrist)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.wrist)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.wrist}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
@@ -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"):
-            draccus.dump(self, f, indent=4)
+        # Encode against the base class so draccus includes the choice "type" key,
+        # 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
@@ -76,7 +76,7 @@ import torch
 from pydantic import BaseModel, Field
 from transformers import AutoProcessor, Qwen3VLMoeForConditionalGeneration

-from lerobot.datasets import LeRobotDataset
+from lerobot.datasets import LeRobotDataset, resolve_episode_indices


 # Pydantic Models for SARM Subtask Annotation
@@ -1049,7 +1049,10 @@ def main():
    torch_dtype = {"bfloat16": torch.bfloat16, "float16": torch.float16, "float32": torch.float32}[args.dtype]

    # Determine episodes
-    episode_indices = args.episodes or list(range(dataset.meta.total_episodes))
+    resolved_episodes = resolve_episode_indices(args.episodes, dataset.meta.total_episodes)
+    episode_indices = (
+        resolved_episodes if resolved_episodes is not None else list(range(dataset.meta.total_episodes))
+    )

    existing_annotations = load_annotations_from_dataset(dataset.root, prefix="sparse")
    if args.skip_existing:
@@ -52,7 +52,7 @@ from .pipeline_features import aggregate_pipeline_dataset_features, create_initi
 from .pyav_utils import check_video_encoder_parameters_pyav, detect_available_encoders_pyav
 from .sampler import EpisodeAwareSampler, compute_sampler_state
 from .streaming_dataset import StreamingLeRobotDataset
-from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card
+from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card, resolve_episode_indices
 from .video_utils import VideoEncodingManager

 # NOTE: Low-level I/O functions (cast_stats_to_numpy, get_parquet_file_size_in_mb, etc.)
@@ -97,6 +97,7 @@ __all__ = [
    "reencode_dataset",
    "remove_feature",
    "resolve_delta_timestamps",
+    "resolve_episode_indices",
    "safe_stop_image_writer",
    "split_dataset",
    "write_stats",
@@ -22,6 +22,7 @@ from pathlib import Path
 from typing import Any, NotRequired, TypedDict

 import datasets
+import numpy as np
 import pandas as pd
 import tqdm

@@ -303,6 +304,46 @@ def update_meta_data(
    df["dataset_to_index"] = df["dataset_to_index"] + dst_meta.info.total_frames
    df["episode_index"] = df["episode_index"] + dst_meta.info.total_episodes

+    # Per-episode stats still describe the pre-merge values of the bookkeeping columns
+    # reindexed above. index/episode_index shift by a constant; task_index is relabeled,
+    # so recompute it from the episode's (stable) task strings via the unified tasks table.
+    shift_stat_keys = ("min", "max", "mean", "q01", "q10", "q50", "q90", "q99")
+    for name, offset in (
+        ("episode_index", dst_meta.info.total_episodes),
+        ("index", dst_meta.info.total_frames),
+    ):
+        for stat in shift_stat_keys:
+            col = f"stats/{name}/{stat}"
+            if col in df.columns:
+                df[col] = df[col] + offset
+
+    if any(c.startswith("stats/task_index/") for c in df.columns):
+        quantiles = {"q01": 0.01, "q10": 0.10, "q50": 0.50, "q90": 0.90, "q99": 0.99}
+        ids_per_row = [
+            np.array([dst_meta.tasks.loc[t, "task_index"] for t in tasks], dtype=np.float64)
+            for tasks in df["tasks"]
+        ]
+
+        def _task_stat(ids, stat):
+            if stat == "min":
+                return ids.min()
+            if stat == "max":
+                return ids.max()
+            if stat == "std":
+                return ids.std()
+            if stat in quantiles:
+                return np.quantile(ids, quantiles[stat])
+            return ids.mean()
+
+        for stat in ("min", "max", "mean", "std", *quantiles):
+            col = f"stats/task_index/{stat}"
+            if col in df.columns:
+                # np.full_like preserves each cell container and dtype so the parquet schema is unchanged.
+                df[col] = [
+                    np.full_like(orig, _task_stat(ids, stat))
+                    for orig, ids in zip(df[col], ids_per_row, strict=True)
+                ]
+
    return df


@@ -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:
-            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 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)
+        metadata_lock = contextlib.nullcontext()
+        if self.repo_type == "bucket":
+            self.root.parent.mkdir(parents=True, exist_ok=True)
+            metadata_lock = WeakFileLock(self.root.parent / f".{self.root.name}.lock")

-            self._pull_from_repo(allow_patterns="meta/", token=token)
-            self._load_metadata()
+        with metadata_lock:
+            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)

@@ -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
@@ -68,8 +68,9 @@ class DatasetReader:
        Args:
            meta: Dataset metadata instance.
            root: Local dataset root directory.
-            episodes: Optional list of episode indices to select. ``None``
-                means all episodes.
+            episodes: Optional list of episode indices to select, assumed
+                already validated by the caller. ``None`` means
+                all episodes.
            tolerance_s: Timestamp synchronization tolerance in seconds.
            video_backend: Video decoding backend identifier.
            delta_timestamps: Optional dict mapping feature keys to lists of
@@ -151,9 +152,9 @@ class DatasetReader:
    @property
    def num_frames(self) -> int:
        """Number of frames in selected episodes."""
-        if self.episodes is not None and self.hf_dataset is not None:
-            return len(self.hf_dataset)
-        return self._meta.total_frames
+        if self.episodes is None:
+            return self._meta.total_frames
+        return sum(self._meta.episodes[ep]["length"] for ep in self.episodes)

    @property
    def num_episodes(self) -> int:
@@ -163,10 +164,34 @@ class DatasetReader:
    def _load_hf_dataset(self) -> datasets.Dataset:
        """hf_dataset contains all the observations, states, actions, rewards, etc."""
        features = get_hf_features_from_features(self._meta.features)
+        self._validate_language_columns_declared(features)
        hf_dataset = load_nested_dataset(self.root / "data", features=features, episodes=self.episodes)
        hf_dataset.set_transform(hf_transform_to_torch)
        return hf_dataset

+    def _validate_language_columns_declared(self, features: datasets.Features) -> None:
+        """Require language columns stored in Parquet to be declared in metadata."""
+        # Leave empty datasets to fail through the normal loading path.
+        try:
+            sample = next((self.root / "data").glob("*/*.parquet"))
+        except StopIteration:
+            return
+
+        from pyarrow import parquet as _pq  # noqa: PLC0415
+
+        # LeRobot shards are schema-uniform, so one schema represents the dataset.
+        schema_names = set(_pq.read_schema(sample).names)
+        from .language import LANGUAGE_COLUMNS  # noqa: PLC0415
+
+        missing = sorted(set(LANGUAGE_COLUMNS) & schema_names - set(features))
+        if missing:
+            raise ValueError(
+                f"Dataset Parquet files contain language feature(s) missing from metadata: {missing}. "
+                "Metadata must describe the stored data; add the entries returned by "
+                "lerobot.datasets.language.language_feature_info() to meta/info.json['features'] "
+                "or rerun the annotation pipeline's metadata synchronization."
+            )
+
    def _check_cached_episodes_sufficient(self) -> bool:
        """Check if the cached dataset contains all requested episodes and their video files."""
        if self.hf_dataset is None or len(self.hf_dataset) == 0:
@@ -215,7 +240,12 @@ class DatasetReader:
    def _get_query_indices(
        self, abs_idx: int, ep_idx: int
    ) -> tuple[dict[str, list[int]], dict[str, torch.Tensor]]:
-        """Compute query indices for delta timestamps."""
+        """Compute query indices for delta timestamps.
+
+        A delta is padding when ``abs_idx + delta`` falls outside the episode's
+        ``[dataset_from_index, dataset_to_index)`` range, and is clamped back into it
+        otherwise.
+        """
        ep = self._meta.episodes[ep_idx]
        ep_start = ep["dataset_from_index"]
        ep_end = ep["dataset_to_index"]
@@ -225,7 +255,7 @@ class DatasetReader:
        }
        padding = {
            f"{key}_is_pad": torch.BoolTensor(
-                [(abs_idx + delta < ep_start) | (abs_idx + delta >= ep_end) for delta in delta_idx]
+                [(abs_idx + delta < ep_start) or (abs_idx + delta >= ep_end) for delta in delta_idx]
            )
            for key, delta_idx in self.delta_indices.items()
        }
@@ -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
-    specific episodes.
+    Per episode, the task is resolved with precedence:
+    `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
-            is also None, raises an error.
-        episode_tasks: Optional dict mapping episode indices to their task strings.
-            Overrides `new_task` for specific episodes.
+        new_task: Default task applied to any episode not covered by `episode_tasks` or a
+            matching `task_replacements` entry.
+        episode_tasks: Optional dict mapping episode indices to task strings. Takes precedence
+            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"]
-            if not original_tasks:
-                raise ValueError(f"Episode {ep_idx} has no tasks and no default task was provided")
-            episode_to_task[ep_idx] = original_tasks[0]
+            episode_to_task[ep_idx] = original_task
+        else:
+            raise ValueError(f"Episode {ep_idx} has no task; provide new_task or episode_tasks")

    # 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()
@@ -29,6 +29,7 @@ from .dataset_metadata import LeRobotDatasetMetadata
 from .lerobot_dataset import LeRobotDataset
 from .multi_dataset import MultiLeRobotDataset
 from .streaming_dataset import StreamingLeRobotDataset
+from .utils import resolve_episode_indices


 def resolve_delta_timestamps(
@@ -84,14 +85,25 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas

    if isinstance(cfg.dataset.repo_id, str):
        ds_meta = LeRobotDatasetMetadata(
-            cfg.dataset.repo_id, root=cfg.dataset.root, revision=cfg.dataset.revision
+            cfg.dataset.repo_id,
+            root=cfg.dataset.root,
+            revision=cfg.dataset.revision,
+            repo_type=cfg.dataset.repo_type,
        )
        delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, ds_meta)
+        # Resolved here because dataset reader does not apply exclude_episodes.
+        episodes = resolve_episode_indices(
+            cfg.dataset.episodes, ds_meta.total_episodes, cfg.dataset.exclude_episodes
+        )
        if not cfg.dataset.streaming:
+            if cfg.dataset.repo_type == "bucket":
+                raise ValueError(
+                    "repo_type='bucket' is streaming-only: set dataset.streaming=true to train from an HF Storage Bucket."
+                )
            dataset = LeRobotDataset(
                cfg.dataset.repo_id,
                root=cfg.dataset.root,
-                episodes=cfg.dataset.episodes,
+                episodes=episodes,
                delta_timestamps=delta_timestamps,
                image_transforms=image_transforms,
                revision=cfg.dataset.revision,
@@ -104,13 +116,14 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
            dataset = StreamingLeRobotDataset(
                cfg.dataset.repo_id,
                root=cfg.dataset.root,
-                episodes=cfg.dataset.episodes,
+                episodes=episodes,
                delta_timestamps=delta_timestamps,
                image_transforms=image_transforms,
                revision=cfg.dataset.revision,
                max_num_shards=cfg.num_workers,
                tolerance_s=cfg.tolerance_s,
                return_uint8=True,
+                repo_type=cfg.dataset.repo_type,
            )
    else:
        raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")
@@ -162,14 +162,32 @@ def render_sample(
    task: str | None = None,
    dataset_ctx: Any | None = None,
 ) -> RenderedMessages | None:
-    """Render the chat-style messages for a single dataset sample.
+    """Render recipe-defined messages and supervision for one dataset sample.

-    Resolves the recipe's bindings against ``persistent`` and ``events`` rows
-    at frame timestamp ``t``, then expands the recipe's message templates.
-    Returns ``None`` if the resolved sample contains no target message.
+    Resolves bindings against ``persistent`` and ``events`` at frame timestamp
+    ``t``. Blend recipes first route matching sparse VQA annotations, then use
+    deterministic weighted selection for the remaining samples. Returns
+    ``None`` when the selected recipe provides no text or low-level action
+    supervision for this sample.
    """
    persistent_rows = _normalize_rows(persistent or [])
    event_rows = _normalize_rows(events or [])
+
+    # Route sparse VQA frames to a matching view-specific component before weighted selection.
+    # This avoids dropping annotated frames or selecting VQA without annotations.
+    if recipe.blend is not None:
+        vqa_rendered = _render_vqa_if_present(
+            recipe,
+            persistent=persistent_rows,
+            events=event_rows,
+            t=t,
+            sample_idx=sample_idx,
+            task=task,
+            dataset_ctx=dataset_ctx,
+        )
+        if vqa_rendered is not None:
+            return vqa_rendered
+
    selected_recipe = _select_recipe(recipe, sample_idx)
    bindings = _resolve_bindings(
        selected_recipe,
@@ -183,6 +201,58 @@ def render_sample(
    return _render_message_recipe(selected_recipe, bindings)


+def _render_vqa_if_present(
+    recipe: TrainingRecipe,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow],
+    t: float,
+    sample_idx: int,
+    task: str | None,
+    dataset_ctx: Any | None,
+) -> RenderedMessages | None:
+    """Render a matching VQA component, or return ``None`` for normal selection.
+
+    Multiple matching views are selected deterministically by relative weight.
+    """
+    if recipe.blend is None:
+        return None
+    renderable: list[tuple[float, RenderedMessages]] = []
+    for component in recipe.blend.values():
+        if component.route != "vqa":
+            continue
+        bindings = _resolve_bindings(
+            component,
+            persistent=persistent,
+            events=events,
+            t=t,
+            sample_idx=sample_idx,
+            task=task,
+            dataset_ctx=dataset_ctx,
+        )
+        rendered = _render_message_recipe(component, bindings)
+        if rendered is not None:
+            if component.weight is None:
+                raise ValueError("Routed VQA blend components must define a weight.")
+            renderable.append((component.weight, rendered))
+
+    if not renderable:
+        return None
+    if len(renderable) == 1:
+        return renderable[0][1]
+
+    # Choose among matching cameras by their validated positive relative weights.
+    total = sum(weight for weight, _ in renderable)
+    digest = hashlib.blake2b(f"vqa:{sample_idx}".encode(), digest_size=8).digest()
+    draw = int.from_bytes(digest, "big") / 2**64 * total
+    cumulative = 0.0
+    for weight, rendered in renderable:
+        cumulative += weight
+        if draw < cumulative:
+            return rendered
+    return renderable[-1][1]
+
+
 def _select_recipe(recipe: TrainingRecipe, sample_idx: int) -> TrainingRecipe:
    """Pick a deterministic blend component for ``sample_idx`` (or return ``recipe``)."""
    if recipe.blend is None:
@@ -201,7 +271,8 @@ def _select_recipe(recipe: TrainingRecipe, sample_idx: int) -> TrainingRecipe:
        cumulative += component.weight or 0.0
        if draw < cumulative:
            return component
-    assert last_component is not None
+    if last_component is None:
+        raise ValueError("Blend recipes must contain at least one component.")
    return last_component


@@ -321,7 +392,8 @@ def _render_message_recipe(
    bindings: dict[str, LanguageRow | str | None],
 ) -> RenderedMessages | None:
    """Expand ``recipe.messages`` into rendered chat messages using ``bindings``."""
-    assert recipe.messages is not None
+    if recipe.messages is None:
+        raise ValueError("Cannot render a blend recipe as a message recipe.")
    messages: list[dict[str, Any]] = []
    streams: list[str | None] = []
    target_indices: list[int] = []
@@ -346,7 +418,9 @@ def _render_message_recipe(
        if turn.target:
            target_indices.append(message_idx)

-    if not target_indices:
+    # Keep samples with either text targets or low-level action supervision.
+    has_low_level = any(stream == "low_level" for stream in streams)
+    if not target_indices and not has_low_level:
        return None

    rendered = {
@@ -403,14 +477,12 @@ def _validate_rendered(rendered: RenderedMessages) -> None:

    if len(streams) != len(messages):
        raise ValueError("message_streams must be aligned with messages.")
-    if not target_indices:
-        raise ValueError("Rendered samples must contain at least one target message.")
+    # Require text or low-level action supervision.
+    if not target_indices and not any(s == "low_level" for s in streams):
+        raise ValueError("Rendered samples must contain a target message or a low_level-stream message.")
    for idx in target_indices:
        if idx < 0 or idx >= len(messages):
            raise ValueError(f"Target message index {idx} is out of bounds.")
-    # ``stream`` is enforced non-None at MessageTurn construction time
-    # (see ``MessageTurn.__post_init__``), so a missing stream here would
-    # mean the dataclass invariant was bypassed; no need to re-check.


 def _nth_relative(
@@ -34,6 +34,7 @@ from .utils import (
    create_lerobot_dataset_card,
    get_safe_version,
    is_valid_version,
+    resolve_episode_indices,
 )
 from .video_utils import (
    StreamingVideoEncoder,
@@ -237,13 +238,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
        self.revision = self.meta.revision
        self.meta.rescale_depth_stats(self._depth_output_unit)

-        if episodes is not None and any(
-            episode >= self.meta.total_episodes or episode < 0 for episode in episodes
-        ):
-            logger.warning(
-                f"Some episodes in the provided episodes list are out of range for this dataset ({self.meta.total_episodes})."
+        episodes = resolve_episode_indices(episodes, self.meta.total_episodes)
+        if episodes is not None and not episodes:
+            raise ValueError(
+                "No valid episodes: the requested episode selection is empty after resolving "
+                f"against the dataset range [0, {self.meta.total_episodes})."
            )
-
        if episode_filter is not None:
            resolved = self.meta.filter_episodes(episode_filter, candidates=episodes)
            if not resolved:
@@ -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

@@ -13,9 +13,12 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import logging
 from collections import deque
 from collections.abc import Callable, Generator, Iterable, Iterator
+from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
+from typing import Literal

 import datasets
 import numpy as np
@@ -24,6 +27,7 @@ from datasets import load_dataset

 from lerobot.configs import DEFAULT_DEPTH_UNIT, DEPTH_METER_UNIT, DepthEncoderConfig
 from lerobot.utils.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
+from lerobot.utils.import_utils import get_safe_default_video_backend

 from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
 from .depth_utils import MM_PER_METRE, dequantize_depth
@@ -31,8 +35,7 @@ from .feature_utils import get_delta_indices
 from .io_utils import item_to_torch
 from .utils import (
    check_version_compatibility,
-    find_float_index,
-    is_float_in_list,
+    resolve_episode_indices,
    safe_shard,
 )
 from .video_utils import (
@@ -41,6 +44,8 @@ from .video_utils import (
    decode_video_frames_torchcodec,
 )

+logger = logging.getLogger(__name__)
+

 class LookBackError(Exception):
    """
@@ -58,6 +63,10 @@ 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 +187,7 @@ 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:
        """
@@ -243,6 +252,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
        repo_id: str,
        root: str | Path | None = None,
        episodes: list[int] | None = None,
+        episode_filter: Callable[[dict], bool] | None = None,
        image_transforms: Callable | None = None,
        delta_timestamps: dict[list[float]] | None = None,
        tolerance_s: float = 1e-4,
@@ -257,17 +267,21 @@ 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
-                metadata is resolved through a revision-safe snapshot cache under
-                ``$HF_LEROBOT_HOME/hub``.
+            root (Path | None, optional): Local directory to use for local datasets. In bucket mode,
+                this is an optional local metadata-cache directory; parquet and video data remain remote.
+                When omitted, Hub metadata is resolved through the cache under ``$HF_LEROBOT_HOME/hub``.
            episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                their episode_index in this list.
+            episode_filter (Callable[[dict], bool] | None, optional): Predicate over per-episode
+                metadata rows used to select episodes (e.g. ``lambda ep: ep["length"] >= 100``).
+                Intersected with ``episodes`` when both are set. Defaults to None.
            image_transforms (Callable | None, optional): Transform to apply to image data.
            tolerance_s (float, optional): Tolerance in seconds for timestamp matching.
            revision (str, optional): Git revision id (branch name, tag, or commit hash).
@@ -280,6 +294,8 @@ 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``
@@ -287,13 +303,16 @@ 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
        self.tolerance_s = tolerance_s
        self.revision = revision if revision else CODEBASE_VERSION
        self.seed = seed
@@ -317,6 +336,7 @@ 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
@@ -325,6 +345,40 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
        # Check version
        check_version_compatibility(self.repo_id, self.meta._version, CODEBASE_VERSION)

+        self.episodes = resolve_episode_indices(episodes, self.meta.total_episodes)
+        if self.episodes is not None and not self.episodes:
+            raise ValueError(
+                "No valid episodes: the requested episode selection is empty after resolving "
+                f"against the dataset range [0, {self.meta.total_episodes})."
+            )
+        if episode_filter is not None:
+            resolved = self.meta.filter_episodes(episode_filter, candidates=self.episodes)
+            if not resolved:
+                raise ValueError(
+                    "The episode filter did not match any episode. Make sure the filter and episodes list are valid and compatible."
+                )
+            logger.info(f"The episode filter matched {len(resolved)} episode(s).")
+            self.episodes = resolved
+
+        # RGB video decodes with torchcodec when available, otherwise pyav (local paths only).
+        # Remote streaming (hf://) needs torchcodec's fsspec-backed reader and can't decode depth
+        # at all (pyav can't read hf:// URLs).
+        is_remote_stream = self.streaming and not self.streaming_from_local
+        self._video_backend = get_safe_default_video_backend()
+        rgb_video_keys = [key for key in self.meta.video_keys if key not in self.meta.depth_keys]
+        if rgb_video_keys and is_remote_stream and self._video_backend != "torchcodec":
+            raise RuntimeError(
+                "Remote StreamingLeRobotDataset requires the 'torchcodec' backend to decode RGB video, "
+                f"but it is not available on this platform (affected keys: {rgb_video_keys}). "
+                "Stream from a local root or use the non-streaming LeRobotDataset instead."
+            )
+        if self.meta.depth_keys and is_remote_stream:
+            raise NotImplementedError(
+                f"Remote streaming of depth video ({self.meta.depth_keys}) is not supported: depth is "
+                "decoded with pyav, which cannot read hf:// URLs. Stream from a local root or use the "
+                "non-streaming LeRobotDataset."
+            )
+
        self._depth_encoder_configs: dict[str, DepthEncoderConfig] = {
            vid_key: DepthEncoderConfig.from_video_info(self.meta.features[vid_key].get("info"))
            for vid_key in self.meta.depth_keys
@@ -345,25 +399,43 @@ 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}
-        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,
-        )
+        token_kwargs = {} if token is None else {"token": token}
+        if self.repo_type == "bucket":
+            self.hf_dataset: datasets.IterableDataset = load_dataset(
+                "parquet",
+                data_files=f"hf://buckets/{self.repo_id}/data/*/*.parquet",
+                split="train",
+                streaming=self.streaming,
+                **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,
+            )
+
+        # Streaming loads all shards, so episode selection is applied as a lazy stream filter.
+        if self.episodes is not None:
+            selected = set(self.episodes)
+            self.hf_dataset = self.hf_dataset.filter(lambda x: x["episode_index"] in selected)

        self.num_shards = min(self.hf_dataset.num_shards, max_num_shards)

    @property
    def num_frames(self):
-        return self.meta.total_frames
+        if self.episodes is None:
+            return self.meta.total_frames
+        return sum(self.meta.episodes[ep]["length"] for ep in self.episodes)

    @property
    def num_episodes(self):
-        return self.meta.total_episodes
+        return len(self.episodes) if self.episodes is not None else self.meta.total_episodes

    @property
    def fps(self):
@@ -422,10 +494,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
                    else:
                        frames_buffer.append(frame)
                    break  # random shard sampled, switch shard
-            except (
-                RuntimeError,
-                StopIteration,
-            ):  # NOTE: StopIteration inside a generator throws a RuntimeError since python 3.7
+            except _ShardExhaustedError:
                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
@@ -457,70 +526,45 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
        lookback, lookahead = self._get_window_steps(self.delta_timestamps)
        return Backtrackable(dataset, history=lookback, lookahead=lookahead)

-    def _make_timestamps_from_indices(
-        self, start_ts: float, indices: dict[str, list[int]] | None = None
-    ) -> dict[str, list[float]]:
-        if indices is not None:
-            return {
-                key: (
-                    start_ts + torch.tensor(indices[key]) / self.fps
-                ).tolist()  # NOTE: why not delta_timestamps directly?
-                for key in self.delta_timestamps
-            }
-        else:
-            return dict.fromkeys(self.meta.video_keys, [start_ts])
+    def _get_query_indices(
+        self, abs_idx: int, ep_idx: int
+    ) -> tuple[dict[str, list[int]], dict[str, torch.BoolTensor]]:
+        """Video-key query indices and padding from integer episode boundaries.

-    def _make_padding_camera_frame(self, camera_key: str):
-        """Variable-shape padding frame for given camera keys, given in (H, W, C)"""
-        return torch.zeros(self.meta.info.features[camera_key]["shape"]).permute(-1, 0, 1)
-
-    def _get_video_frame_padding_mask(
-        self,
-        video_frames: dict[str, torch.Tensor],
-        query_timestamps: dict[str, list[float]],
-        original_timestamps: dict[str, list[float]],
-    ) -> dict[str, torch.BoolTensor]:
-        padding_mask = {}
-
-        for video_key, timestamps in original_timestamps.items():
-            if video_key not in video_frames:
-                continue  # only padding on video keys that are available
-            frames = []
-            mask = []
-            padding_frame = self._make_padding_camera_frame(video_key)
-            for ts in timestamps:
-                if is_float_in_list(ts, query_timestamps[video_key]):
-                    idx = find_float_index(ts, query_timestamps[video_key])
-                    frames.append(video_frames[video_key][idx, :])
-                    mask.append(False)
-                else:
-                    frames.append(padding_frame)
-                    mask.append(True)
-
-            padding_mask[f"{video_key}_is_pad"] = torch.BoolTensor(mask)
-
-        return padding_mask
+        Mirrors ``DatasetReader._get_query_indices`` but only for video keys.
+        """
+        ep = self.meta.episodes[ep_idx]
+        ep_start, ep_end = ep["dataset_from_index"], ep["dataset_to_index"]
+        query_indices = {
+            key: [max(ep_start, min(ep_end - 1, abs_idx + delta)) for delta in delta_idx]
+            for key, delta_idx in self.delta_indices.items()
+            if key in self.meta.video_keys
+        }
+        padding = {
+            f"{key}_is_pad": torch.BoolTensor(
+                [(abs_idx + delta < ep_start) or (abs_idx + delta >= ep_end) for delta in delta_idx]
+            )
+            for key, delta_idx in self.delta_indices.items()
+            if key in self.meta.video_keys
+        }
+        return query_indices, padding

    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)

-        # Get episode index from the item
+        # Get episode index and absolute frame index from the item
        ep_idx = item["episode_index"]
-
-        # "timestamp" restarts from 0 for each episode, whereas we need a global timestep within the single .mp4 file (given by index/fps)
-        current_ts = item["index"] / self.fps
-
-        episode_boundaries_ts = {
-            key: (
-                self.meta.episodes[ep_idx][f"videos/{key}/from_timestamp"],
-                self.meta.episodes[ep_idx][f"videos/{key}/to_timestamp"],
-            )
-            for key in self.meta.video_keys
-        }
+        abs_idx = int(item["index"])
+        ep_start = self.meta.episodes[ep_idx]["dataset_from_index"]
+        current_ts = float(item["timestamp"])

        # Apply delta querying logic if necessary
        if self.delta_indices is not None:
@@ -530,12 +574,19 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):

        # Load video frames, when needed
        if len(self.meta.video_keys) > 0:
-            original_timestamps = self._make_timestamps_from_indices(current_ts, self.delta_indices)
+            query_indices = None
+            if self.delta_indices is not None:
+                query_indices, video_padding = self._get_query_indices(abs_idx, ep_idx)

-            # Some timestamps might not result available considering the episode's boundaries
-            query_timestamps = self._get_query_timestamps(
-                current_ts, self.delta_indices, episode_boundaries_ts
-            )
+            # Episode-local timestamps; `_query_videos` shifts them by the per-key `from_timestamp` at decode.
+            query_timestamps = {
+                key: (
+                    [(idx - ep_start) / self.fps for idx in query_indices[key]]
+                    if query_indices is not None and key in query_indices
+                    else [current_ts]
+                )
+                for key in self.meta.video_keys
+            }
            video_frames = self._query_videos(query_timestamps, ep_idx)

            if self.image_transforms is not None:
@@ -547,10 +598,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):

            if self.delta_indices is not None:
                # We always return the same number of frames. Unavailable frames are padded.
-                padding_mask = self._get_video_frame_padding_mask(
-                    video_frames, query_timestamps, original_timestamps
-                )
-                updates.append(padding_mask)
+                updates.append(video_padding)

        result = item.copy()
        for update in updates:
@@ -569,27 +617,6 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):

        yield result

-    def _get_query_timestamps(
-        self,
-        current_ts: float,
-        query_indices: dict[str, list[int]] | None = None,
-        episode_boundaries_ts: dict[str, tuple[float, float]] | None = None,
-    ) -> dict[str, list[float]]:
-        query_timestamps = {}
-        keys_to_timestamps = self._make_timestamps_from_indices(current_ts, query_indices)
-        for key in self.meta.video_keys:
-            if query_indices is not None and key in query_indices:
-                timestamps = keys_to_timestamps[key]
-                # Clamp out timesteps outside of episode boundaries
-                query_timestamps[key] = torch.clamp(
-                    torch.tensor(timestamps), *episode_boundaries_ts[key]
-                ).tolist()
-
-            else:
-                query_timestamps[key] = [current_ts]
-
-        return query_timestamps
-
    def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict:
        """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
        in the main process (e.g. by using a second Dataloader with num_workers=0). It will result in a
@@ -597,22 +624,28 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
        the main process and a subprocess fails to access it.
        """

-        item = {}
-        for video_key, query_ts in query_timestamps.items():
-            root = self.meta.url_root if self.streaming and not self.streaming_from_local else self.root
-            video_path = f"{root}/{self.meta.get_video_file_path(ep_idx, video_key)}"
-            if video_key in self.meta.depth_keys:
+        ep = self.meta.episodes[ep_idx]
+        root = self.meta.url_root if self.streaming and not self.streaming_from_local else self.root
+
+        def _decode_single(vid_key: str, query_ts: list[float]) -> tuple[str, torch.Tensor]:
+            # Episode-local timestamps restart from 0 each episode; shift by the per-key
+            # `from_timestamp` to reach the episode's segment within its video file, matching
+            # `DatasetReader._decode_single`.
+            from_timestamp = ep[f"videos/{vid_key}/from_timestamp"]
+            shifted_query_ts = [from_timestamp + ts for ts in query_ts]
+            video_path = f"{root}/{self.meta.get_video_file_path(ep_idx, vid_key)}"
+            if vid_key in self.meta.depth_keys:
                # Depth maps are 12-bit quantized and only decodable via pyav; dequantize back
                # to physical units to match the non-streaming reader.
                frames = decode_video_frames(
                    video_path,
-                    query_ts,
+                    shifted_query_ts,
                    self.tolerance_s,
                    backend="pyav",
                    return_uint8=False,
                    is_depth=True,
                )
-                depth_encoder = self._depth_encoder_configs[video_key]
+                depth_encoder = self._depth_encoder_configs[vid_key]
                frames = dequantize_depth(
                    frames,
                    depth_min=depth_encoder.depth_min,
@@ -621,18 +654,36 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
                    use_log=depth_encoder.use_log,
                    output_unit=self._depth_output_unit,
                )
-            else:
+            elif self._video_backend == "torchcodec":
                frames = decode_video_frames_torchcodec(
                    video_path,
-                    query_ts,
+                    shifted_query_ts,
                    self.tolerance_s,
                    decoder_cache=self.video_decoder_cache,
                    return_uint8=self._return_uint8,
                )
+            else:
+                # torchcodec unavailable: only reachable for local streaming
+                frames = decode_video_frames(
+                    video_path,
+                    shifted_query_ts,
+                    self.tolerance_s,
+                    backend="pyav",
+                    return_uint8=self._return_uint8,
+                )

-            item[video_key] = frames.squeeze(0) if len(query_ts) == 1 else frames
+            return vid_key, frames.squeeze(0) if len(shifted_query_ts) == 1 else frames

-        return item
+        items = list(query_timestamps.items())
+
+        # Single camera: no threading overhead
+        if len(items) <= 1:
+            return {vid_key: _decode_single(vid_key, query_ts)[1] for vid_key, query_ts in items}
+
+        # Multi-camera: decode in parallel (video decoding releases the GIL)
+        with ThreadPoolExecutor(max_workers=len(items)) as pool:
+            futures = [pool.submit(_decode_single, k, ts) for k, ts in items]
+            return dict(f.result() for f in futures)

    def _get_delta_frames(self, dataset_iterator: Backtrackable, current_item: dict):
        # TODO(fracapuano): Modularize this function, refactor the code
@@ -18,6 +18,7 @@ import dataclasses
 import importlib.resources
 import json
 import logging
+from collections.abc import Sequence
 from dataclasses import dataclass, field
 from pathlib import Path

@@ -26,7 +27,6 @@ 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 +51,17 @@ 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): ...

@@ -88,6 +99,47 @@ VIDEO_DIR = "videos"

 CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
 IMAGE_FILE_PATTERN = "frame-{frame_index:06d}.png"
+
+
+def resolve_episode_indices(
+    episodes: Sequence[int] | None,
+    total_episodes: int,
+    exclude_episodes: Sequence[int] | None = None,
+) -> list[int] | None:
+    """Resolve an optional episode allowlist and exclusion list against dataset bounds.
+
+    ``None`` is preserved when no filtering is requested so callers can retain
+    their native "all episodes" fast path. Invalid indices are ignored with a
+    warning, and the input order is preserved.
+    """
+    if total_episodes < 0:
+        raise ValueError(f"total_episodes must be non-negative, got {total_episodes}")
+
+    if episodes is None and not exclude_episodes:
+        return None
+
+    candidates = list(range(total_episodes)) if episodes is None else list(episodes)
+    invalid = [episode for episode in candidates if not 0 <= episode < total_episodes]
+    if invalid:
+        logger.warning(
+            "Ignoring episode indices outside the dataset range [0, %d): %s",
+            total_episodes,
+            invalid,
+        )
+    candidates = [episode for episode in candidates if 0 <= episode < total_episodes]
+
+    excluded = set(exclude_episodes or [])
+    invalid_excluded = sorted(episode for episode in excluded if not 0 <= episode < total_episodes)
+    if invalid_excluded:
+        logger.warning(
+            "Ignoring excluded episode indices outside the dataset range [0, %d): %s",
+            total_episodes,
+            invalid_excluded,
+        )
+    excluded = {episode for episode in excluded if 0 <= episode < total_episodes}
+    return [episode for episode in candidates if episode not in excluded]
+
+
 DEPTH_FILE_PATTERN = "frame-{frame_index:06d}.tiff"
 DEFAULT_TASKS_PATH = "meta/tasks.parquet"
 DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
@@ -368,7 +420,7 @@ 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 +430,7 @@ 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(
-            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")
-            ```
-            """
-        )
+        raise RuntimeError(MISSING_VERSION_TAG_MESSAGE.format(repo_id=repo_id))

    if target_version in hub_versions:
        return f"v{target_version}"
@@ -481,17 +523,6 @@ def create_lerobot_dataset_card(
    )


-def is_float_in_list(target, float_list, threshold=1e-6):
-    return any(abs(target - x) <= threshold for x in float_list)
-
-
-def find_float_index(target, float_list, threshold=1e-6):
-    for i, x in enumerate(float_list):
-        if abs(target - x) <= threshold:
-            return i
-    return -1
-
-
 def safe_shard(dataset: datasets.IterableDataset, index: int, num_shards: int) -> datasets.Dataset:
    """
    Safe shards the dataset.
@@ -328,6 +328,7 @@ class LiberoEnv(EnvConfig):
    render_mode: str = "rgb_array"
    camera_name: str = "agentview_image,robot0_eye_in_hand_image"
    init_states: bool = True
+    hard_reset: bool = True
    camera_name_mapping: dict[str, str] | None = None
    observation_height: int = 360
    observation_width: int = 360
@@ -356,6 +357,8 @@ class LiberoEnv(EnvConfig):
    def __post_init__(self):
        if self.fps <= 0:
            raise ValueError(f"fps must be positive, got {self.fps}")
+        if not self.hard_reset and not self.init_states:
+            raise ValueError("hard_reset=False requires init_states=True")

        if self.obs_type == "pixels":
            self.features[LIBERO_KEY_PIXELS_AGENTVIEW] = PolicyFeature(
@@ -416,6 +419,7 @@ class LiberoEnv(EnvConfig):
            "observation_height": self.observation_height,
            "observation_width": self.observation_width,
            "control_freq": self.fps,
+            "hard_reset": self.hard_reset,
        }
        if self.task_ids is not None:
            kwargs["task_ids"] = self.task_ids
@@ -507,7 +511,7 @@ 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

@@ -128,10 +128,13 @@ class LiberoEnv(gym.Env):
        control_freq: int = 20,
        control_mode: str = "relative",
        is_libero_plus: bool = False,
+        hard_reset: bool = True,
    ):
        super().__init__()
        if control_freq <= 0:
            raise ValueError(f"control_freq must be positive, got {control_freq}")
+        if not hard_reset and not init_states:
+            raise ValueError("hard_reset=False requires init_states=True")
        self.task_id = task_id
        self.is_libero_plus = is_libero_plus
        self.obs_type = obs_type
@@ -158,6 +161,7 @@ class LiberoEnv(gym.Env):
        self.camera_name_mapping = camera_name_mapping
        self.num_steps_wait = num_steps_wait
        self.control_freq = control_freq
+        self.hard_reset = hard_reset
        self.episode_index = episode_index
        self.episode_length = episode_length
        # Load once and keep
@@ -265,6 +269,9 @@ class LiberoEnv(gym.Env):
            camera_heights=self.observation_height,
            camera_widths=self.observation_width,
            control_freq=self.control_freq,
+            # Soft resets skip LIBERO's model and renderer rebuild. They are opt-in
+            # because settle steps can make their observations differ from hard resets.
+            hard_reset=self.hard_reset,
        )
        env.reset()
        self._env = env
@@ -377,8 +384,9 @@ class LiberoEnv(gym.Env):
            }
        )
        observation = self._format_raw_obs(raw_obs)
-        if terminated:
-            self.reset()
+        # Return the terminal observation unchanged. The caller owns resetting after
+        # termination; vector envs created below use NEXT_STEP autoreset. Resetting here
+        # would therefore reset twice and skip an initial state.
        truncated = False
        return observation, reward, terminated, truncated, info

@@ -476,6 +484,7 @@ def create_libero_envs(
        print(f"Restricting to task_ids={task_ids_filter}")

    is_async = env_cls is gym.vector.AsyncVectorEnv
+    is_sync = env_cls is gym.vector.SyncVectorEnv

    out: dict[str, dict[int, Any]] = defaultdict(dict)
    for suite_name in suite_names:
@@ -512,6 +521,10 @@ def create_libero_envs(
                    cached_act_space = lazy.action_space
                    cached_metadata = lazy.metadata
                out[suite_name][tid] = lazy
+            elif is_sync:
+                out[suite_name][tid] = gym.vector.SyncVectorEnv(
+                    fns, autoreset_mode=gym.vector.AutoresetMode.NEXT_STEP
+                )
            else:
                out[suite_name][tid] = env_cls(fns)
            print(f"Built vec env | suite={suite_name} | task_id={tid} | n_envs={n_envs}")
@@ -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
@@ -177,6 +177,76 @@ def _sub_env_has_attr(env: gym.vector.VectorEnv, attr: str) -> bool:
        return False


+# Passed in `reset(options=...)` by `rollout()` to mark the start of a new rollout.
+# FreezeAfterEpisodeEnd thaws only on this, so Gymnasium's argument-less autoreset
+# cannot be mistaken for a genuine new episode.
+NEW_ROLLOUT_OPTION = "lerobot_new_rollout"
+
+
+class FreezeAfterEpisodeEnd(gym.Wrapper):
+    """Stop doing simulator work once a sub-env's episode has ended.
+
+    `rollout()` runs `while not np.all(done)` with `done` latched, so a sub-env that
+    terminates early keeps being stepped -- physics and offscreen rendering included --
+    until the slowest sub-env in the batch finishes. The batch runs for
+    `max(episode_lengths)` iterations to complete work that only needs
+    `mean(episode_lengths)`.
+
+    This caches the terminal transition and replays it for any further `step()` or
+    autoreset, so a finished sub-env costs nothing. The rollout already ignores those
+    transitions.
+
+    The freeze survives Gymnasium's autoreset deliberately. Under
+    `AutoresetMode.NEXT_STEP` the vector env resets a terminated sub-env on the
+    following step and runs it through an entire extra episode that the rollout
+    discards, because `done` stays latched. Absorbing that reset is most of the saving.
+
+    Only an explicit reset carrying `NEW_ROLLOUT_OPTION` thaws it, so the signal is
+    explicit rather than inferred: Gymnasium's autoreset calls `reset()` with no
+    arguments, but so would a caller passing `seeds=None`, and confusing the two would
+    strand an env frozen for a whole rollout.
+
+    `AutoresetMode.DISABLED` is not an alternative here — Gymnasium asserts that no
+    terminated env is ever stepped in that mode, so the wrapper is never reached.
+    """
+
+    def __init__(self, env: gym.Env):
+        super().__init__(env)
+        self._frozen: tuple | None = None
+
+    def reset(self, *, seed=None, options=None):
+        if self._frozen is not None and not (options or {}).get(NEW_ROLLOUT_OPTION):
+            # Gymnasium's autoreset for a sub-env the rollout has already finished with.
+            # Replay the terminal observation instead of rebuilding the simulation.
+            obs, _, _, _, info = self._frozen
+            return obs, info
+        self._frozen = None
+        return self.env.reset(seed=seed, options=options)
+
+    def step(self, action):
+        if self._frozen is not None:
+            return self._frozen
+        obs, reward, terminated, truncated, info = self.env.step(action)
+        if terminated or truncated:
+            # Zero the reward on replay so a frozen sub-env cannot inflate a return if a
+            # caller sums rewards over the padded tail.
+            self._frozen = (obs, 0.0, terminated, truncated, info)
+        return obs, reward, terminated, truncated, info
+
+    @property
+    def is_frozen(self) -> bool:
+        return self._frozen is not None
+
+
+def freeze_after_episode_end(env_fn: Callable[[], gym.Env]) -> Callable[[], gym.Env]:
+    """Wrap an env factory so the built env freezes once its episode ends."""
+
+    def _fn() -> gym.Env:
+        return FreezeAfterEpisodeEnd(env_fn())
+
+    return _fn
+
+
 class _LazyAsyncVectorEnv:
    """Defers AsyncVectorEnv creation until first use.

@@ -212,7 +282,12 @@ class _LazyAsyncVectorEnv:

    def _ensure(self) -> None:
        if self._env is None:
-            self._env = gym.vector.AsyncVectorEnv(self._env_fns, context="forkserver", shared_memory=True)
+            self._env = gym.vector.AsyncVectorEnv(
+                [freeze_after_episode_end(fn) for fn in self._env_fns],
+                context="forkserver",
+                shared_memory=True,
+                autoreset_mode=gym.vector.AutoresetMode.NEXT_STEP,
+            )

    @property
    def unwrapped(self):
@@ -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

@@ -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,
@@ -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,
@@ -37,7 +38,6 @@ 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,
@@ -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,
@@ -520,6 +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,
@@ -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
--- a/Show More
+++ b/Show More