refactor(rtc): share Pi05 training helpers

fix(rollout): align relative state feature order
fix(rtc): bootstrap trained inference after cold start
2026-08-01 05:59:44 +00:00 · 2026-07-28 12:33:35 +02:00 · 2026-07-28 12:33:35 +02:00 · 2026-07-28 12:33:34 +02:00 · 2026-07-28 12:33:34 +02:00 · 2026-07-28 12:33:34 +02:00
310 changed files with 12758 additions and 5218 deletions
@@ -53,7 +53,7 @@ permissions:
  contents: read

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.11.30"
+  UV_VERSION: "0.8.0"
  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
@@ -52,21 +52,21 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
      - name: Build and push Docker image CPU
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: ./docker/Dockerfile.user
@@ -87,21 +87,21 @@ jobs:
          sudo apt-get update
          sudo apt-get install git-lfs
          git lfs install
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
        with:
          lfs: true
          persist-credentials: false
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v4 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          cache-binary: false
      - name: Login to Docker Hub
-        uses: docker/login-action@v4.4.0 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
        with:
          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
      - name: Build and push Docker image GPU
-        uses: docker/build-push-action@v7 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
        with:
          context: .
          file: ./docker/Dockerfile.internal
@@ -33,7 +33,7 @@ jobs:
      github.event.workflow_run.event == 'pull_request' &&
      github.event.workflow_run.conclusion == 'success' &&
      github.repository == 'huggingface/lerobot'
-    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@6108e850ae1cf2f71bb0815a600bcd50c39abfa7  # main
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@2430c1ec91d04667414e2fa31ecfc36c153ea391  # main
    with:
      package_name: lerobot
    secrets:
@@ -55,7 +55,7 @@ jobs:
      github.repository == 'huggingface/lerobot'
    permissions:
      contents: read
-    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@6108e850ae1cf2f71bb0815a600bcd50c39abfa7  # main
+    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@e60a538eea9817ab312196d0d233604b01697265  # main
    with:
      commit_sha: ${{ github.sha }}
      package: lerobot
@@ -78,7 +78,7 @@ jobs:
    permissions:
      contents: read
      pull-requests: write
-    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@6108e850ae1cf2f71bb0815a600bcd50c39abfa7  # main
+    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@e60a538eea9817ab312196d0d233604b01697265  # main
    with:
      commit_sha: ${{ github.event.pull_request.head.sha }}
      pr_number: ${{ github.event.number }}
@@ -48,7 +48,7 @@ permissions:

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

      - name: Secret Scanning
-        uses: trufflesecurity/trufflehog@27b0417c16317ca9a472a9a8092acce143b49c55  # v3.95.9
+        uses: trufflesecurity/trufflehog@eafb8c5f6a06175141c27f17bcc17941853d0047  # v3.90.0
        with:
          extra_args: --only-verified
@@ -61,20 +61,16 @@ Full details in [`docs/source/so101.mdx`](./docs/source/so101.mdx) and [`docs/so
 **4.1 Install**

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

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

 ```bash
@@ -102,10 +102,10 @@ lerobot-train \
 ```

 | Category                   | Models                                                                                                                                                                                                                                                                                                                                                                                                                       |
-| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md), [Multitask DiT Policy](./docs/source/policy_multi_task_dit_README.md)                                                                                                                                                                                                      |
 | **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                                                                                                                                                                                                                                  |
-| **VLAs Models**            | [Pi0](./docs/source/pi0.mdx), [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.7](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx), [EO-1](./docs/source/eo1.mdx), [MolmoAct2](./docs/source/molmoact2.mdx), [WALL-OSS](./docs/source/walloss.mdx), [EVO1](./docs/source/evo1.mdx) |
+| **VLAs Models**            | [Pi0](./docs/source/pi0.mdx), [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [Pi052](./docs/source/pi052.mdx), [GR00T N1.7](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx), [EO-1](./docs/source/eo1.mdx), [MolmoAct2](./docs/source/molmoact2.mdx), [WALL-OSS](./docs/source/walloss.mdx), [EVO1](./docs/source/evo1.mdx) |
 | **World Models**           | [VLA-JEPA](./docs/source/vla_jepa.mdx), [LingBot-VA](./docs/source/lingbot_va.mdx), [FastWAM](./docs/source/fastwam.mdx)                                                                                                                                                                                                                                                                                                     |
 | **Reward Models**          | [SARM](./docs/source/sarm.mdx), [TOPReward](./docs/source/topreward.mdx), [Robometer](./docs/source/robometer.mdx)                                                                                                                                                                                                                                                                                                           |

@@ -68,16 +68,17 @@ ENV HOME=/home/user_lerobot \
 # issues with MuJoCo and OpenGL drivers.
 RUN uv venv --python python${PYTHON_VERSION}

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

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

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

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

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

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

 # Set the default command
 CMD ["/bin/bash"]
@@ -63,6 +63,8 @@
    title: π₀-FAST (Pi0Fast)
  - local: pi05
    title: π₀.₅ (Pi05)
+  - local: pi052
+    title: π₀.₅ with language supervision (Pi052)
  - local: molmoact2
    title: MolmoAct2
  - local: vla_jepa
@@ -65,7 +65,7 @@ In summary, you need to specify instructions for:
 Importantly,

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

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

 ## Hardware API redesign

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

 ### What changed?

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

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

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

 ```diff
 action_values = predict_action(
@@ -191,6 +191,162 @@ def make_my_policy_pre_post_processors(

 ---

+## Adding high- and low-level language control
+
+The policy API above is sufficient for training and standard evaluation. To use a language-conditioned policy with interactive `lerobot-rollout`, also register a runtime adapter. The adapter keeps policy-specific prompting and tokenization out of the generic control loop.
+
+The runtime supports two policy shapes:
+
+| Policy shape     | Behavior                                                                | Adapter                                        |
+| ---------------- | ----------------------------------------------------------------------- | ---------------------------------------------- |
+| Low-level / flat | The operator's task or subtask directly conditions action prediction.   | Reuse `DirectTaskPolicyAdapter`.               |
+| High + low level | The policy generates subtasks or memory, then conditions actions on it. | Subclass `BaseLanguageAdapter`, as PI052 does. |
+
+During a rollout, `RuntimeState` stores the high-level task and the active language context:
+
+```text
+task ──> adapter.generate_text("subtask", ...) ──> state.language_context["subtask"]
+                                                       │
+observation ──> processors ──> adapter.select_action() ─┴─> action chunk ──> robot
+```
+
+The generic runtime handles generation frequency, pause/resume, prompt replacement, action queues, and dispatch. The adapter only translates between that runtime contract and your policy.
+
+### Low-level policies
+
+If your policy already consumes the live task through its normal preprocessor and implements `predict_action_chunk`, register the shared direct adapter. PI0.5 and MolmoAct2 use this path:
+
+```python
+# src/lerobot/runtime/registry.py
+_ADAPTERS = {
+    # ...
+    "my_policy": "lerobot.runtime.adapter:DirectTaskPolicyAdapter",
+}
+```
+
+Run it with direct-subtask mode so the operator supplies the instruction used by the action policy:
+
+```bash
+lerobot-rollout \
+    --language \
+    --policy.path=user/my_policy_checkpoint \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --direct_subtask
+```
+
+The rollout context builds the observation batch with the current instruction before `DirectTaskPolicyAdapter` calls `policy.predict_action_chunk(observation)`. No text-generation method is required.
+
+### Hierarchical policies
+
+For a policy that generates language and actions, subclass [`BaseLanguageAdapter`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/runtime/adapter.py) and implement two methods:
+
+- `generate_text(kind, observation, state, user_text=None) -> str` generates a `subtask`, `memory`, or interjection response.
+- `select_action(observation, state)` builds the low-level prompt from the active context and returns an action chunk.
+
+This abbreviated adapter follows [`PI052PolicyAdapter`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi052/inference/pi052_adapter.py):
+
+```python
+# inference/my_policy_adapter.py
+from typing import Any
+
+from lerobot.runtime import RuntimeState
+from lerobot.runtime.adapter import BaseLanguageAdapter
+from lerobot.utils.constants import (
+    OBS_LANGUAGE_ATTENTION_MASK,
+    OBS_LANGUAGE_TOKENS,
+)
+
+
+class MyPolicyAdapter(BaseLanguageAdapter):
+    def select_action(self, observation: dict[str, Any], state: RuntimeState):
+        instruction = state.language_context.get("subtask") or state.task or ""
+        tokens, attention_mask = tokenize_instruction(instruction)
+
+        batch = dict(observation)
+        batch[OBS_LANGUAGE_TOKENS] = tokens
+        batch[OBS_LANGUAGE_ATTENTION_MASK] = attention_mask
+        return self.policy.predict_action_chunk(batch)
+
+    def generate_text(
+        self,
+        kind: str,
+        observation: dict[str, Any] | None,
+        state: RuntimeState,
+        user_text: str | None = None,
+    ) -> str:
+        messages = self.build_messages(kind, state, user_text)
+        batch, tokenizer = tokenize_messages(messages, observation)
+        return self.policy.select_message(
+            batch,
+            tokenizer=tokenizer,
+            min_new_tokens=self.gen.min_new_tokens,
+            temperature=self.gen.temperature,
+            top_p=self.gen.top_p,
+        )
+
+    def build_messages(
+        self, kind: str, state: RuntimeState, user_text: str | None
+    ) -> list[dict[str, str]]:
+        if kind == "subtask":
+            return [{"role": "user", "content": state.task or ""}]
+        if kind == "memory":
+            return [
+                {"role": "user", "content": state.task or ""},
+                {
+                    "role": "user",
+                    "content": f"Completed subtask: {state.extra.get('prior_subtask', '')}",
+                },
+            ]
+        if kind == "interjection":
+            return [
+                {"role": "user", "content": state.task or ""},
+                {"role": "user", "content": user_text or ""},
+            ]
+        raise ValueError(f"Unsupported text kind: {kind}")
+```
+
+`tokenize_instruction` and `tokenize_messages` are policy-specific helpers. They must reproduce the prompt format used during training; PI052, for example, adds the discretized robot state to its low-level subtask prompt and uses the same PaliGemma formatting for `select_message`.
+
+`BaseLanguageAdapter` provides the default hierarchy: regenerate a subtask at action-chunk boundaries, update memory when the subtask changes, and handle user interjections. Override `_regenerate_context` only if your policy uses a different hierarchy.
+
+Register the adapter with a lazy import so importing LeRobot does not load the model or its optional dependencies:
+
+```python
+# src/lerobot/runtime/registry.py
+_ADAPTERS = {
+    # ...
+    "my_policy": "lerobot.policies.my_policy.inference.my_policy_adapter:MyPolicyAdapter",
+}
+```
+
+The key must match the policy's registered type. Once registered, the same checkpoint works through the shared entry point:
+
+```bash
+lerobot-rollout \
+    --language \
+    --policy.path=user/my_hierarchical_checkpoint \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --task="put the cup in the sink"
+```
+
+For RoboCasa-compatible policies, replace the robot arguments with `--sim --sim.task=<task>`. Without `--direct_subtask`, the adapter generates the low-level subtask; with it, the operator bypasses high-level generation and supplies each subtask.
+
+### Keep training and deployment aligned
+
+The adapter is intentionally small, but its prompts are part of the model contract:
+
+- Use the same tokenizer, role formatting, special tokens, image ordering, and state encoding as training.
+- Condition `select_action` on `state.language_context["subtask"]`, falling back to `state.task` for direct or not-yet-generated prompts.
+- Return a full action chunk from `select_action`; the runtime handles control-rate dispatch.
+- Keep optional model dependencies inside lazy imports.
+- Test adapter selection, generated-message routing, action-batch construction, and direct-subtask behavior with a lightweight fake policy.
+
+PI052 is the complete in-tree reference: its [processor](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi052/processor_pi052.py) renders the training recipe, its [policy](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi052/modeling_pi052.py) exposes text and action generation, and its [adapter](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi052/inference/pi052_adapter.py) reconstructs those same prompts at deployment.
+
+---
+
 ## Path A: Out-of-tree plugin

 The fastest way to ship a policy: package it as a standalone Python distribution and install it alongside LeRobot. No PR required, you own the release cycle, and you can publish to PyPI under your own namespace.
@@ -136,10 +136,6 @@ config = RealSenseCameraConfig(
    height=480,
    color_mode=ColorMode.RGB,
    use_depth=True,
-    # Optional fixed color controls. Omit them to leave the current sensor settings unchanged.
-    exposure=120,
-    gain=64,
-    white_balance=4600,
    rotation=Cv2Rotation.NO_ROTATION
 )

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

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

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

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

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

 ### 3. **Easier Experimentation**

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

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

@@ -40,10 +40,10 @@ This tutorial guides you through updating the firmware of Feetech motors using t
 For each motor you want to update:

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

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

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

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

 ### Record

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

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

 ## Cameras

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

 ## Teleoperate with cameras

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

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

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

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

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

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

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

-`lerobot-rollout` is the single CLI for deploying trained policies on real robots. It supports multiple execution strategies and inference backends, from quick evaluation to continuous recording and human-in-the-loop data collection.
+`lerobot-rollout` is the single CLI for deploying trained policies on real robots or in an interactive simulator. It supports multiple execution strategies and inference backends, from quick evaluation to continuous recording, language-driven control, and human-in-the-loop data collection.

 ## Quick Start

@@ -150,12 +150,11 @@ lerobot-rollout \
 Foot pedal input is also supported via `--strategy.input_device=pedal`. Configure pedal codes with `--strategy.pedal.*` flags.

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

 ### Episodic (`--strategy.type=episodic`)
@@ -188,14 +187,59 @@ Teleop is optional — if omitted the robot holds its position during the reset
 | `ESC`       | Stop the recording session       |

 | Flag                                            | Description                                                                |
-| ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| ----------------------------------------------- | -------------------------------------------------------------------------- |
 | `--dataset.num_episodes`                        | Number of episodes to record                                               |
 | `--dataset.episode_time_s`                      | Duration of each recording episode in seconds                              |
 | `--dataset.reset_time_s`                        | Duration of the reset phase between episodes in seconds                    |
 | `--teleop.type`                                 | Optional. Teleoperator to drive the robot during resets                    |
 | `--strategy.reset_to_initial_position`          | Whether to reset the robot to its initial position between episodes        |
 | `--strategy.smooth_leader_to_follower_handover` | Whether to turn on or off the leader -> follower smooth handover behavior. |
-| `--strategy.smooth_handover`                    | Smoothly hand control to the teleop at reset start (default: true). Disable for clutch-style teleops that re-reference at the current robot pose on engage |
+
+---
+
+## Interactive language control
+
+Language-conditioned policies can expose a high-level text head in addition to
+their action head. Add `--language` to open-prompt one of these policies on a
+real robot. Language-only flags such as `--direct_subtask` select this mode
+automatically.
+
+MolmoAct2 has no high-level planner, so use direct-subtask mode and type each
+next low-level instruction yourself:
+
+```bash
+lerobot-rollout \
+    --policy.path=lerobot/MolmoAct2-SO100_101-LeRobot \
+    --policy.device=cuda \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM1 \
+    --robot.cameras='{"cam0":{"type":"opencv","index_or_path":"/dev/video0","width":640,"height":480,"fps":30,"fourcc":"MJPG","backend":200},"cam1":{"type":"opencv","index_or_path":"/dev/video2","width":640,"height":480,"fps":30,"fourcc":"MJPG","backend":200}}' \
+    --direct_subtask \
+    --robot.max_relative_target='{"shoulder_pan":5,"shoulder_lift":5,"elbow_flex":5,"wrist_flex":5,"wrist_roll":5,"gripper":5}'
+```
+
+The robot starts paused. Type a subtask, then use `/resume` and `/pause` to
+control action dispatch. Check the workspace and motion limits before resuming.
+Without `--direct_subtask`, a policy such as PI052 generates its active subtask
+from the high-level `--task` itself.
+
+RoboCasa uses the same runtime and processor path. `--sim` selects it
+automatically, so no robot configuration is needed:
+
+```bash
+MUJOCO_GL=egl lerobot-rollout \
+    --policy.path=lerobot/pi052_robocasa \
+    --sim --sim.task=CloseFridge --sim.split=pretrain \
+    --task="close the fridge" \
+    --disable_memory \
+    --sim.render_size=384 \
+    --sim.views=robot0_agentview_left,robot0_eye_in_hand,robot0_agentview_right \
+    --mode=action --ctrl_hz=20
+```
+
+Open `http://localhost:8010` for the live simulator view. Add
+`--sim.direct_subtask` to bypass the language planner and make each typed prompt
+the action policy's current subtask.

 ---

@@ -231,11 +275,12 @@ lerobot-rollout \
 ```

 | Flag                                        | Description                                                                     |
-| ------------------------------------------- | -------------------------------------------------------------- |
+| ------------------------------------------- | ------------------------------------------------------------------------------- |
 | `--inference.rtc.execution_horizon`         | Steps to blend with previous chunk (default: varies by policy)                  |
+| `--inference.rtc.mode`                      | `guided` (default) or trained-prefix `trained` for compatible Pi052 checkpoints |
 | `--inference.rtc.max_guidance_weight`       | Consistency enforcement strength (default: varies by policy)                    |
 | `--inference.rtc.prefix_attention_schedule` | Blend schedule: `LINEAR`, `EXP`, `ONES`, `ZEROS`                                |
-| `--inference.queue_threshold`               | Max queue size before backpressure (default: 30)               |
+| `--inference.queue_threshold`               | Backpressure threshold; trained RTC requires at least its maximum delay         |

 See the [Real-Time Chunking](./rtc) guide for details on tuning RTC parameters.

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

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

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

@@ -141,6 +141,17 @@ 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.
+
+A message recipe with a supervised assistant turn on the `low_level` stream trains
+the π0.5 paper's joint sequence instead of a blend: the target span gets text CE
+while also conditioning the action losses in the same forward.
+`recipes/subtask_joint.yaml` is the provided example; pair it with
+`--policy.joint_subtask_conditioning=true` at inference.
+
 ## Graceful absence

 If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.
@@ -51,7 +51,7 @@ In addition to these instructions, you need to install the Feetech SDK & ZeroMQ
 pip install -e ".[lekiwi]"
 ```

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

 # Step-by-Step Assembly Instructions
@@ -92,20 +92,6 @@ LIBERO supports two control modes — `relative` (default) and `absolute`. Diffe
 --env.control_mode=relative   # or "absolute"
 ```

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

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

 ### Recommended evaluation episodes

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

 ## Training

 ### Dataset

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

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

 For reference, the original dataset published by Physical Intelligence:

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

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

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

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

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

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

 **Observations:**
@@ -36,12 +36,6 @@ This diverse training mixture creates a "curriculum" that enables generalization
   pip install -e ".[pi]"
   ```

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

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

 ## Training

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

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

 ```bash
 lerobot-train \
-    --dataset.repo_id=lerobot/libero \
+    --dataset.repo_id=your_dataset \
    --policy.type=pi05 \
-    --policy.pretrained_path=lerobot/pi05_libero_base \
-    --policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}' \
-    --policy.n_action_steps=10 \
-    --policy.empty_cameras=1 \
-    --policy.freeze_vision_encoder=true \
-    --policy.train_expert_only=true \
+    --output_dir=./outputs/pi05_training \
+    --job_name=pi05_training \
+    --policy.repo_id=your_repo_id \
+    --policy.pretrained_path=lerobot/pi05_base \
+    --policy.compile_model=true \
    --policy.gradient_checkpointing=true \
+    --wandb.enable=true \
    --policy.dtype=bfloat16 \
+    --policy.freeze_vision_encoder=false \
+    --policy.train_expert_only=false \
+    --steps=3000 \
    --policy.device=cuda \
-    --policy.push_to_hub=false \
-    --output_dir=./outputs/pi05_libero_expert \
-    --job_name=pi05_libero_expert \
-    --batch_size=64 \
-    --num_workers=8 \
-    --steps=30000 \
-    --save_freq=5000 \
-    --seed=1000
+    --batch_size=32
 ```

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

 ### Training Parameters Explained

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

 **💡 Tip**: Setting `train_expert_only=true` freezes the VLM and trains only the action expert and projections, allowing finetuning with reduced memory usage.

+If your dataset is not converted with `quantiles`, you can convert it with the following command:
+
+```bash
+python src/lerobot/scripts/augment_dataset_quantile_stats.py \
+    --repo-id=your_dataset \
+```
+
+Or train pi05 with this normalization mapping: `--policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}'`
+
 ## Relative Actions

 By default, π₀.₅ predicts absolute actions. You can enable **relative actions** so the model predicts offsets relative to the current robot state. This can improve training stability for certain setups.
@@ -0,0 +1,334 @@
+# π₀.₅ with language supervision (Pi052)
+
+Pi052 extends [Pi05](./pi05) with a trainable PaliGemma language head and a
+runtime that alternates language generation with action generation. A single
+checkpoint can predict a low-level subtask, optionally update memory or answer
+visual questions, and condition its flow-matching action expert on that text.
+
+Use Pi05 when you only need task-conditioned actions. Use Pi052 when the policy
+must generate or consume intermediate language during a rollout.
+
+## How Pi052 differs from Pi05
+
+| Capability          | Pi05                                                   | Pi052                                                                             |
+| ------------------- | ------------------------------------------------------ | --------------------------------------------------------------------------------- |
+| Action model        | PaliGemma vision-language prefix + Gemma action expert | Same base architecture                                                            |
+| Language head       | Not trained for runtime generation                     | Re-enabled and trained with text cross-entropy                                    |
+| Action conditioning | Episode task                                           | Active low-level subtask plus normalized robot state                              |
+| Training targets    | Flow-matching actions                                  | Flow actions, recipe-selected text, and optional FAST action tokens               |
+| Dataset requirement | Standard images, state, actions, and task              | The same fields plus language annotations for every language capability you train |
+| Rollout             | Direct task-to-action policy                           | Hierarchical task → subtask → action loop, with optional memory and VQA           |
+
+Pi052 can initialize from a Pi05 checkpoint. The policy architecture remains
+compatible, while Pi052 builds its own processors so recipe labels and FAST
+labels are not silently replaced by the Pi05 processor stack.
+
+## Install
+
+Install LeRobot with the PI dependencies:
+
+```bash
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+python -m venv .venv
+source .venv/bin/activate
+pip install -e ".[pi]"
+```
+
+The `pi` extra includes the PaliGemma/FAST dependencies. Install
+`liger-kernel` for the supported fused training kernels; optional FlashRT
+backends also require the Hugging Face `kernels` package and a supported CUDA
+GPU.
+
+## Prepare language-annotated data
+
+Pi052 does not infer supervised subtasks from a normal LeRobot dataset during
+training. The dataset must contain the language targets used by the selected
+recipe in the optional `language_persistent` and `language_events` columns.
+
+At minimum, annotate a continuous `subtask` timeline so each training frame has
+an active low-level instruction. Add `memory`, VQA, interjections, and speech
+annotations only if the recipe trains those capabilities.
+
+The provided recipes are:
+
+| Recipe                                | Required annotations                                                    | Trains                                                             |
+| ------------------------------------- | ----------------------------------------------------------------------- | ------------------------------------------------------------------ |
+| `recipes/subtask.yaml`                | `subtask`                                                               | Subtask prediction and subtask-conditioned actions                 |
+| `recipes/subtask_joint.yaml`          | `subtask`                                                               | Paper-style joint sequence: subtask text and actions in one sample |
+| `recipes/subtask_mem.yaml`            | `subtask`, `memory`                                                     | Subtasks, actions, and memory updates                              |
+| `recipes/subtask_mem_vqa_speech.yaml` | `subtask`, `memory`, `vqa`; interjection/speech rows for those branches | Subtasks, actions, memory, VQA, and spoken replies                 |
+
+The blend recipes factorize training into separate high-level (task → subtask)
+and low-level (subtask → actions) samples, matching how inference decomposes
+π(a|o, subtask)·π(subtask|o, task). `recipes/subtask_joint.yaml` instead uses
+the π0.5 paper's single-sequence layout — the supervised subtask span is
+attended causally and conditions the FAST and flow losses in the same forward.
+Checkpoints trained with the joint recipe must set
+`--policy.joint_subtask_conditioning=true` at inference so the flow prefix
+rebuilds the same layout (task turn with state, then the generated subtask as a
+causal assistant turn); leave it `false` for the blend recipes.
+
+Use `lerobot-annotate` to generate these columns. The repository includes a
+Hugging Face Jobs launcher that you can edit for your source and destination
+datasets. For a local annotation run, first install
+`pip install -e ".[annotations]"`:
+
+```bash
+HF_TOKEN=hf_... uv run python examples/annotations/run_hf_job.py
+```
+
+Before a long training run, inspect several episodes and verify that subtasks
+are temporally correct and cover the full demonstration. See
+[Annotation Pipeline](./annotation_pipeline) for generation and validation, and
+[Language Columns and Recipes](./language_and_recipes) for the schema and
+recipe resolver.
+
+<Tip>
+  If a dataset has no language columns, recipe rendering becomes a no-op and
+  Pi052 falls back to the plain Pi05 prompt path. This is useful for
+  compatibility but does not train the language planner.
+</Tip>
+
+## Train Pi052
+
+This example initializes Pi052 from the native Pi052 initialization checkpoint
+and trains the default subtask-and-memory recipe:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=${HF_USER}/my_language_annotated_dataset \
+    --policy.type=pi052 \
+    --policy.pretrained_path=lerobot/pi052_base \
+    --policy.recipe_path=recipes/subtask_mem.yaml \
+    --policy.dtype=bfloat16 \
+    --policy.device=cuda \
+    --policy.freeze_vision_encoder=false \
+    --policy.gradient_checkpointing=true \
+    --batch_size=8 \
+    --steps=30000 \
+    --output_dir=outputs/pi052 \
+    --job_name=pi052 \
+    --wandb.enable=true
+```
+
+For subtask-only data, change the recipe to `recipes/subtask.yaml` and disable
+memory during rollout. Start with a small run and confirm that W&B examples show
+the expected prompt, text target, and action endpoints before scaling up.
+
+### Main training controls
+
+| Option                              |                    Default | Purpose                                                             |
+| ----------------------------------- | -------------------------: | ------------------------------------------------------------------- |
+| `policy.recipe_path`                | `recipes/subtask_mem.yaml` | Selects the language/action objective mixture                       |
+| `policy.text_loss_weight`           |                      `1.0` | Language-head cross-entropy weight; `0` disables text training      |
+| `policy.flow_loss_weight`           |                     `10.0` | Continuous action flow-loss weight                                  |
+| `policy.enable_fast_action_loss`    |                     `true` | Adds discrete FAST action-token supervision                         |
+| `policy.fast_action_loss_weight`    |                      `1.0` | FAST cross-entropy weight                                           |
+| `policy.knowledge_insulation`       |                     `true` | Blocks action-loss gradients through the VLM K/V path               |
+| `policy.flow_num_repeats`           |                        `5` | Reuses one VLM prefix for independent denoising targets             |
+| `policy.rtc_training_max_delay`     |                        `0` | Maximum clean-prefix delay; `0` disables training-time RTC          |
+| `policy.lm_head_lr_scale`           |                      `1.0` | Scales language-head learning rate; `1.0` uses the base rate        |
+| `policy.fast_skip_tokens`           |                     `1152` | FAST id offset; skips `<seg>`+`<loc>` so VQA and FAST never collide |
+| `policy.joint_subtask_conditioning` |                    `false` | Rebuilds the joint-sequence prefix at inference (see recipes)       |
+
+`fast_skip_tokens=1152` places FAST codes below PaliGemma's `<loc>` range.
+openpi's pi0-FAST convention is `128` (FAST occupies the `<loc>` ids); use that
+value only to stay weight-compatible with checkpoints trained that way, and
+avoid combining it with the VQA recipe, whose `<loc>` targets would share
+embedding rows with FAST codes.
+
+The loss weights are starting points, not dataset-independent constants. Track
+flow loss and text/FAST losses separately, and inspect generated subtasks rather
+than selecting a checkpoint from total loss alone.
+
+### Training-time RTC
+
+Pi052 optionally supports training-time action conditioning from
+[Training-Time Action Conditioning for Efficient Real-Time Chunking](https://arxiv.org/abs/2512.05964).
+It simulates inference latency by sampling a clean action prefix for every flow
+draw, passing a per-action flow timestep to the action expert, and computing the
+flow loss only on the remaining postfix. The default value of `0` leaves the
+standard Pi052 objective unchanged.
+
+```bash
+lerobot-train \
+    --dataset.repo_id=${HF_USER}/my_language_annotated_dataset \
+    --policy.type=pi052 \
+    --policy.pretrained_path=lerobot/pi05_base \
+    --policy.recipe_path=recipes/subtask_mem.yaml \
+    --policy.rtc_training_max_delay=10 \
+    --policy.dtype=bfloat16 \
+    --policy.device=cuda \
+    --batch_size=8 \
+    --steps=30000 \
+    --output_dir=outputs/pi052_rtc \
+    --job_name=pi052_rtc
+```
+
+`rtc_training_max_delay` is measured in controller steps and must be smaller
+than `chunk_size`. Choose it to cover the largest inference latency expected at
+deployment: at 50 Hz, for example, 10 steps correspond to 200 ms. A delay of
+zero is included in the uniform sampling distribution, so the checkpoint also
+continues to receive ordinary flow-matching examples. Set rollout's
+`inference.rtc.execution_horizon` and `inference.queue_threshold` to at least
+this maximum so inference starts early enough and the previous chunk retains
+every action needed for the committed prefix.
+
+Run the resulting checkpoint with the asynchronous `lerobot-rollout` backend
+and select the trained-prefix path explicitly:
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --policy.path=outputs/pi052_rtc/checkpoints/last/pretrained_model \
+    --inference.type=rtc \
+    --inference.rtc.mode=trained \
+    --inference.rtc.execution_horizon=10 \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --task="pick up the cube" \
+    --fps=50 \
+    --device=cuda
+```
+
+The rollout engine measures latency continuously, carries the still-unexecuted
+actions from the previous chunk into the next prediction, and discards the
+prefix that elapsed during inference. If the measured delay exceeds the
+checkpoint's `rtc_training_max_delay`, rollout stops with an explicit error
+instead of silently extrapolating beyond the training distribution. Use
+`--inference.rtc.mode=guided` for the original Jacobian-guided RTC path; it does
+not require a training-time RTC checkpoint but adds backward-pass work during
+denoising.
+
+### Dataset-specific FAST tokenizer
+
+The universal FAST tokenizer works out of the box. For a large or
+embodiment-specific dataset, Pi052 can fit and cache a tokenizer on normalized
+actions before training:
+
+```bash
+lerobot-train \
+    ... \
+    --policy.auto_fit_fast_tokenizer=true \
+    --policy.fast_tokenizer_fit_samples=4096
+```
+
+The fit runs once per dataset/tokenizer configuration. Keep
+`auto_fit_fast_tokenizer=false` when you do not want the extra preprocessing
+pass.
+
+## Training performance
+
+Pi052 uses optimized training paths by default:
+
+- batches repeated flow targets and suffix projections instead of replaying
+  small operations in Python;
+- caches constant action masks and computes RoPE positions once per forward;
+- selects the text/FAST cross-entropy implementation from target shape and
+  sparsity;
+- skips the mathematically dead VLM/vision backward on knowledge-insulated,
+  flow-only batches;
+- uses native non-reentrant SigLIP layer checkpointing when gradient
+  checkpointing is enabled; and
+- retains the Liger RoPE/GeGLU kernels while avoiding the slower LayerNorm
+  patch at SigLIP shapes.
+
+Optional training backends are disabled by default:
+
+| Option                                 | When to try it                                                                                    |
+| -------------------------------------- | ------------------------------------------------------------------------------------------------- |
+| `policy.use_flashrt_adarms=true`       | Fused adaptive RMSNorm and gated residuals on supported CUDA GPUs                                 |
+| `policy.use_compiled_text_ce=true`     | Compiled materialized-logit CE buckets                                                            |
+| `policy.use_compiled_vision=true`      | Compiled vision only when the vision pass has no gradients                                        |
+| `policy.use_flex_attention=true`       | Profiled CUDA setups with knowledge insulation and `flow_num_repeats > 1`; otherwise SDPA is used |
+| `policy.use_manual_attention=true`     | Explicitly profiled shapes where materialized attention is faster                                 |
+| `policy.manual_attention_scope=action` | Restricts manual attention to action queries                                                      |
+
+Do not enable every backend blindly. Flex and manual attention are mutually
+exclusive, and attention/AdaRMS alternatives require knowledge insulation.
+The benchmark-best configuration used compiled text CE and FlashRT AdaRMS,
+with Flex/manual attention and compiled vision disabled.
+
+### Reported training benchmarks
+
+These benchmarks measure complete optimizer steps with three real camera
+inputs, BF16 transformer/action execution, FP32 vision, fused AdamW, and no
+video decoding or network I/O. Results vary with GPU, batch shape, annotation
+mixture, and checkpointing:
+
+| Workload                   |     RTX PRO 6000 Blackwell |                 A100 80 GB |
+| -------------------------- | -------------------------: | -------------------------: |
+| Full flow + text, batch 1  | 4.75× vs checkpointing off | 3.33× vs checkpointing off |
+| Full flow + text, batch 8  | 2.16× vs checkpointing off | 1.66× vs checkpointing off |
+| Full flow + text, batch 64 |  1.24× vs checkpointing on |  1.15× vs checkpointing on |
+| Flow-only, batch 1         | 3.70× vs checkpointing off | 3.58× vs checkpointing off |
+| Flow-only, batch 64        |  3.76× vs checkpointing on |  3.61× vs checkpointing on |
+
+On those 80 GB GPUs, full training was fastest without gradient checkpointing
+through batch 8, then required checkpointing at batch 16 and above. Treat that
+as a tuning rule to test on your hardware, not a universal threshold. Flow-only
+means both text and FAST supervision are disabled; it is useful for action-only
+ablation or post-training but does not learn the language runtime.
+
+## Inference performance
+
+Pi052 has two inference loops, and both avoid repeatedly encoding the expensive
+multimodal prefix:
+
+1. **Action denoising** encodes the image/language prefix once, reuses its KV
+   cache across flow steps, precomputes the timestep schedule on-device, and
+   crops temporary suffix K/V instead of cloning the prefix cache.
+2. **Language decoding** uses autoregressive KV caching, so each new token only
+   processes the sampled token against cached image/language keys instead of
+   rerunning the full prefix.
+
+The runtime also runs language and actions at different rates. Increase
+`--subtask_chunks_per_gen` when a subtask remains valid across several action
+chunks, lower `--high_level_hz`, or use `--direct_subtask` to bypass language
+generation entirely. These settings reduce compute but also slow replanning.
+
+`--fp8` enables the optional FlashRT inference MLP swap on supported CUDA GPUs.
+It calibrates on the first observation and falls back to BF16 when unavailable;
+because FP8 can change outputs slightly, validate task success before using it
+for production rollouts.
+
+## Run a checkpoint
+
+RoboCasa:
+
+```bash
+MUJOCO_GL=egl lerobot-rollout \
+    --policy.path=lerobot/pi052_robocasa \
+    --sim --sim.task=CloseFridge --sim.split=pretrain \
+    --task="close the fridge" \
+    --disable_memory \
+    --sim.render_size=384 \
+    --sim.views=robot0_agentview_left,robot0_eye_in_hand,robot0_agentview_right \
+    --mode=action --ctrl_hz=20
+```
+
+Open `http://localhost:8010` for the live view. Without
+`--sim.direct_subtask`, Pi052 generates the low-level subtask; with it, each
+prompt becomes the action policy's subtask directly.
+
+The same runtime supports real robots. See [Interactive language
+control](./inference#interactive-language-control) for the real-arm command,
+safety behavior, and runtime controls.
+
+## Troubleshooting
+
+- **No text loss or generated subtasks:** confirm the selected recipe can bind
+  the annotations on sampled frames and that `policy.text_loss_weight > 0`.
+- **Subtasks look plausible but actions fail:** verify subtask boundaries,
+  normalized state/action statistics, and that low-level recipe samples are
+  present.
+- **Text collapses to repeated or location tokens:** inspect text-target
+  coverage, language-head learning rate, and the balance between flow, FAST,
+  and text losses.
+- **Out of memory:** reduce batch size first, then enable gradient
+  checkpointing. Do not enable compiled or alternative attention backends
+  without profiling their memory on your camera count.
+- **Slow rollout:** separate action latency from language latency, then tune
+  `--subtask_chunks_per_gen`, `--high_level_hz`, and the number of flow
+  inference steps.
@@ -110,15 +110,21 @@ lerobot-train \
 ### Key Training Parameters

 | Parameter                               | Description                                        | Default                         |
-| -------------------------------------- | -------------------------------------------------- | ------------------------------- |
+| --------------------------------------- | -------------------------------------------------- | ------------------------------- |
 | `--policy.gradient_checkpointing=true`  | Reduces memory usage significantly during training | `false`                         |
 | `--policy.dtype=bfloat16`               | Use mixed precision training for efficiency        | `float32`                       |
 | `--policy.chunk_size`                   | Number of action steps to predict (action horizon) | `50`                            |
-| `--policy.n_action_steps`              | Number of action steps to execute                  | `50`                            |
+| `--policy.n_action_steps`               | Number of decoded action steps to execute          | `50`                            |
 | `--policy.max_action_tokens`            | Maximum number of FAST tokens per action chunk     | `256`                           |
 | `--policy.action_tokenizer_name`        | FAST tokenizer to use                              | `lerobot/fast-action-tokenizer` |
+| `--policy.auto_fit_fast_tokenizer=true` | Fit and cache a tokenizer for the training dataset | `false`                         |
 | `--policy.compile_model=true`           | Enable torch.compile for faster training           | `false`                         |

+Set `--policy.auto_fit_fast_tokenizer=true` to sample action chunks from the
+training dataset and cache a fitted tokenizer under
+`~/.cache/lerobot/fast_tokenizers`. This also works when fine-tuning with
+`--policy.path`; leave it disabled to retain the checkpoint's tokenizer.
+
 ## Inference

 ### KV-Caching for Fast Inference
@@ -151,6 +157,9 @@ actions = policy.predict_action_chunk(batch)

 The model takes images, text instructions, and robot state as input, and outputs discrete FAST tokens that are decoded back to continuous actions.

+PI0-FAST always decodes a complete `chunk_size` action chunk. `n_action_steps` controls only
+how many actions from that chunk are executed before the policy predicts again.
+
 ## Configuration Options

 | Parameter            | Description                                     | Default    |
@@ -174,7 +183,7 @@ The model takes images, text instructions, and robot state as input, and outputs

 ## Reproducing π₀Fast results

-We reproduce the results of π₀Fast on the LIBERO benchmark using the LeRobot implementation. We take the LeRobot PiFast base model [lerobot/pi0fast-base](https://huggingface.co/lerobot/pi0fast-base) and finetune for an additional 40k steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
+We reproduce the results of π₀Fast on the LIBERO benchmark using the LeRobot implementation. We take the LeRobot PiFast base model [lerobot/pi0fast-base](https://huggingface.co/lerobot/pi0fast-base) and finetune for an additional 40kk steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).

 The finetuned model can be found here:

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

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

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

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

 ## Dataset feature contracts

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

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

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

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

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

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

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

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

 ## Running Evaluation

@@ -1,6 +1,6 @@
 # Real-Time Chunking (RTC)

-Real-Time Chunking (RTC) is an inference-time method that allows large, flow-matching based robotic policies, such as [Pi0](./pi0), [Pi0.5](./pi05), and [SmolVLA](./smolvla), to produce smooth, continuous, and reactive motion despite having high inference latency.
+Real-Time Chunking (RTC) allows large, flow-matching based robotic policies, such as [Pi0](./pi0), [Pi0.5](./pi05), and [SmolVLA](./smolvla), to produce smooth, continuous, and reactive motion despite having high inference latency. LeRobot provides the original inference-time guided mode and, for compatible Pi052 checkpoints, training-time action conditioning with cheap hard-prefix inference.

 These policies generate chunks of future actions (e.g., 50 steps at a time) instead of single actions.
 Because the models are large, producing each chunk takes longer than the time it takes the robot to execute it.
@@ -57,7 +57,7 @@ policy_cfg.rtc_config = RTCConfig(
 policy = PI0Policy.from_pretrained("lerobot/pi0_base", policy_cfg=policy_cfg, device="cuda")

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

 # Initialize the action queue
 action_queue = ActionQueue(policy_cfg.rtc_config)
@@ -92,6 +92,15 @@ for step in range(num_steps):

 `RTCConfig` has the following parameters to tune:

+**`mode`** selects the action-prefix conditioning method:
+
+- `guided` (default) applies the original Jacobian guidance during denoising and works with ordinary flow-matching checkpoints.
+- `trained` hard-inpaints the previous chunk's prefix with per-action flow timesteps. It currently requires a Pi052 checkpoint trained with `policy.rtc_training_max_delay > 0` and avoids the guidance backward pass.
+
+For trained mode, both `execution_horizon` and the rollout backend's
+`inference.queue_threshold` must be at least the checkpoint's
+`rtc_training_max_delay`; rollout validates this before connecting the robot.
+
 **`execution_horizon`**: How many timesteps from the previous chunk to maintain consistency with. Higher values mean smoother transitions but potentially less reactivity.

 Typical values: 8-12 steps
@@ -100,7 +109,7 @@ Typical values: 8-12 steps
 RTCConfig(execution_horizon=10)
 ```

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

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

@@ -124,6 +133,10 @@ python examples/rtc/eval_dataset.py \
    --device=cuda
 ```

+Add `--rtc.mode=trained` when evaluating a compatible training-time RTC Pi052
+checkpoint. Unsupported policies reject trained mode instead of falling back to
+guided RTC.
+
 The script generates a visualization of the denoising process, comparing standard generation (left) with RTC (right). In the RTC plots, you can see how the first few steps (blue/purple lines) are guided to match the red ground truth trajectory (previous chunk's tail), ensuring a smooth transition between chunks.

 <p align="center">
@@ -141,6 +154,7 @@ lerobot-rollout \
    --strategy.type=base \
    --policy.path=${HF_USERNAME}/policy_repo_id \
    --inference.type=rtc \
+    --inference.rtc.mode=guided \
    --inference.rtc.execution_horizon=10 \
    --inference.rtc.max_guidance_weight=10.0 \
    --robot.type=so100_follower \
@@ -151,6 +165,24 @@ lerobot-rollout \
    --device=cuda
 ```

+For a training-time RTC Pi052 checkpoint, change the mode to `trained`. The
+checkpoint records its maximum supported delay, and rollout validates measured
+latency against it:
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --policy.path=${HF_USERNAME}/pi052_training_rtc \
+    --inference.type=rtc \
+    --inference.rtc.mode=trained \
+    --inference.rtc.execution_horizon=10 \
+    --robot.type=so100_follower \
+    --robot.port=/dev/tty.usbmodem58FA0834591 \
+    --task="Move green small object into the purple platform" \
+    --duration=120 \
+    --device=cuda
+```
+
 ## How It Differs from the Async Inference in LeRobot

 Both RTC and [async inference](./async) improve real-time robot control, but they solve different problems.
@@ -189,3 +221,5 @@ See `examples/rtc/eval_dataset.py` for a complete example of offline RTC visuali
 - [Smooth-As-Butter Robot Policies](https://alexander-soare.github.io/robotics/2025/08/05/smooth-as-butter-robot-policies.html) - Excellent technical explanation with real robot results
 - [Physical Intelligence - Real-Time Chunking](https://www.physicalintelligence.company/research/real_time_chunking) - Original paper and research
 - [Kinetix RTC Implementation](https://github.com/Physical-Intelligence/real-time-chunking-kinetix) - Reference implementation from Physical Intelligence
+- [Training-Time Action Conditioning](https://arxiv.org/abs/2512.05964) - Efficient RTC with clean-prefix conditioning during training
+- [RLDX-1](https://github.com/RLWRLD/RLDX-1) - PyTorch reference used for the training-time RTC integration
@@ -93,7 +93,7 @@ lerobot-train --help

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

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

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

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

@@ -11,10 +11,9 @@ LeRobot provides several utilities for manipulating datasets:
 3. **Merge Datasets** - Combine multiple datasets into one. The datasets must have identical features, and episodes are concatenated in the order specified in `repo_ids`
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
-6. **Modify Tasks** - Change the natural-language task descriptions associated with episodes
-7. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
-8. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
-9. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
+6. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
+7. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
+8. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.

 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
@@ -51,11 +50,11 @@ lerobot-edit-dataset \
 Divide a dataset into multiple subsets.

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

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

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

 Convert an image-based dataset to video format, creating a new LeRobotDataset where images are stored as videos. This is useful for reducing storage requirements and improving data loading performance. The new dataset will have the exact same structure as the original, but with images encoded as MP4 videos in the proper LeRobot format.
@@ -50,11 +50,11 @@ lerobot-record \
 All flags below are prefixed with `--dataset.rgb_encoder.` on the CLI.

 | Parameter       | Type             | Default       | Description                                                                                                                                                                            |
-| --------------- | ---------------- | ------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| --------------- | ---------------- | ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `vcodec`        | `str`            | `"libsvtav1"` | Video codec name. `"auto"` picks the first available hardware encoder from a fixed preference list, falling back to `libsvtav1`.                                                       |
 | `pix_fmt`       | `str`            | `"yuv420p"`   | Output pixel format. Must be supported by the chosen codec in your FFmpeg build.                                                                                                       |
 | `g`             | `int`            | `2`           | GOP size — a keyframe every `g` frames. Emitted as FFmpeg option `g`.                                                                                                                  |
-| `crf`           | `int` or `float` | `30`          | Abstract quality value, mapped per codec (see the [mapping](https://github.com/huggingface/lerobot/blob/main/src/lerobot/configs/video.py#L197)). Lower → higher quality / larger output where the mapping is monotone. |
+| `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.                                                                                                                             |
@@ -44,7 +44,6 @@ from typing import Protocol

 import numpy as np

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

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

 import numpy as np

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

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

 import numpy as np

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

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

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

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

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

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

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

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

 import time

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

@@ -306,6 +306,7 @@ class RTCEvaluator:
        # Configure RTC
        rtc_config = RTCConfig(
            enabled=rtc_enabled,
+            mode=self.cfg.rtc.mode,
            execution_horizon=self.cfg.rtc.execution_horizon,
            max_guidance_weight=self.cfg.rtc.max_guidance_weight,
            prefix_attention_schedule=self.cfg.rtc.prefix_attention_schedule,
@@ -417,7 +418,7 @@ class RTCEvaluator:
    def run_evaluation(self):
        """Run evaluation on two random dataset samples using three separate policies.

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

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

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

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

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

 import time

-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
    RobotProcessorPipeline,
@@ -31,6 +30,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
    InverseKinematicsEEToJoints,
 )
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data

@@ -67,8 +67,8 @@ dependencies = [
    "einops>=0.8.0,<0.9.0",

    # Config & Hub
-    "draccus>=0.11.6,<0.12.0",
-    "huggingface-hub>=1.6.0,<2.0.0",
+    "draccus==0.10.0", # TODO: Relax version constraint
+    "huggingface-hub>=1.0.0,<2.0.0",
    "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,<82.0.0",  # torch 2.11 requires setuptools<82; a higher cap makes the resolver downgrade torch
+    "setuptools>=71.0.0,<81.0.0",
 ]

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

 # ── Feature-scoped extras ──────────────────────────────────
 dataset = [
-    "datasets>=4.8.0,<5.0.0",
+    "datasets>=4.7.0,<5.0.0",
    "pandas>=2.0.0,<3.0.0", # NOTE: Transitive dependency of datasets
    "pyarrow>=21.0.0,<30.0.0", # NOTE: Transitive dependency of datasets
    "lerobot[av-dep]",
@@ -150,6 +150,7 @@ pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 # There is no cmeel-urdfdom 5.x; <5 selects the 4.x ABI the placo/pin wheels are built against.
 placo-dep = ["placo>=0.9.6,<0.9.16", "cmeel-urdfdom>=4,<5", "cmeel-tinyxml2<11"]
 transformers-dep = ["transformers>=5.4.0,<5.6.0"]
+sentencepiece-dep = ["sentencepiece>=0.2.0,<0.3.0"] # FAST action tokenizer backend (pi052, pi0_fast)
 grpcio-dep = ["grpcio>=1.73.1,<2.0.0", "protobuf>=6.31.1,<8.0.0"]
 accelerate-dep = ["accelerate>=1.14.0,<2.0.0"]
 can-dep = ["python-can>=4.2.0,<5.0.0"]
@@ -212,7 +213,7 @@ wallx = [
    "torchdiffeq>=0.2.4,<0.3.0",
    "lerobot[qwen-vl-utils-dep]",
 ]
-pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
+pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]", "lerobot[sentencepiece-dep]"]
 molmoact2 = ["lerobot[transformers-dep]", "lerobot[peft-dep]", "lerobot[scipy-dep]"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "lerobot[accelerate-dep]"]
 multi_task_dit = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]"]
@@ -261,7 +262,7 @@ annotations = [
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools>=1.73.1,<2.0.0", "mypy>=1.19.1", "ruff>=0.14.1", "lerobot[notebook]"]
 notebook = ["jupyter>=1.0.0,<2.0.0", "ipykernel>=6.0.0,<7.0.0"]
-test = ["pytest>=8.1.0,<10.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
+test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]

 # Simulation
@@ -494,19 +495,6 @@ ignore_errors = true
 module = "lerobot.envs.*"
 ignore_errors = false

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

 # [[tool.mypy.overrides]]
 # module = "lerobot.utils.*"
@@ -38,7 +38,6 @@ import draccus
 import grpc
 import torch

-from lerobot.lerobot_types import PolicyAction
 from lerobot.policies import get_policy_class, make_pre_post_processors
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.transport import (
@@ -46,6 +45,7 @@ from lerobot.transport import (
    services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import receive_bytes_in_chunks
+from lerobot.types import PolicyAction

 from .configs import PolicyServerConfig
 from .constants import SUPPORTED_POLICIES
@@ -120,22 +120,14 @@ class OpenCVCamera(Camera):
        self.rotation: int | None = get_cv2_rotation(config.rotation)
        self.backend: int = config.backend

-        self.capture_width: int | None = None
-        self.capture_height: int | None = None
-        self._reset_connection_settings()
-
-    def __str__(self) -> str:
-        return f"{self.__class__.__name__}({self.index_or_path})"
-
-    def _reset_connection_settings(self) -> None:
-        """Restore settings that may have been auto-detected during a failed connection."""
-        self.fps = self.config.fps
-        self.width = self.config.width
-        self.height = self.config.height
+        if self.height and self.width:
            self.capture_width, self.capture_height = self.width, self.height
            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
                self.capture_width, self.capture_height = self.height, self.width

+    def __str__(self) -> str:
+        return f"{self.__class__.__name__}({self.index_or_path})"
+
    @property
    def is_connected(self) -> bool:
        """Checks if the camera is currently connected and opened."""
@@ -172,7 +164,6 @@ class OpenCVCamera(Camera):
                f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
            )

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

@@ -184,13 +175,6 @@ class OpenCVCamera(Camera):
            with self.frame_lock:
                if self.latest_frame is None:
                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
-        except BaseException:
-            try:
-                self._cleanup_resources()
-            except Exception:
-                logger.exception(f"Failed to fully clean up {self} after connect() failed.")
-            self._reset_connection_settings()
-            raise

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

@@ -328,7 +312,6 @@ class OpenCVCamera(Camera):

        for target in targets_to_scan:
            camera = cv2.VideoCapture(target)
-            try:
            if camera.isOpened():
                default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
                default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
@@ -338,9 +321,7 @@ class OpenCVCamera(Camera):
                # Get FOURCC code and convert to string
                default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
                default_fourcc_code_int = int(default_fourcc_code)
-                    default_fourcc = "".join(
-                        [chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)]
-                    )
+                default_fourcc = "".join([chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])

                camera_info = {
                    "name": f"OpenCV Camera @ {target}",
@@ -357,7 +338,6 @@ class OpenCVCamera(Camera):
                }

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

        return found_cameras_info
@@ -516,26 +496,6 @@ class OpenCVCamera(Camera):
            self.latest_timestamp = None
            self.new_frame_event.clear()

-    def _cleanup_resources(self) -> None:
-        """Stop background reads and release the capture, including after partial setup."""
-        read_thread = self.thread
-        videocapture = self.videocapture
-
-        try:
-            self._stop_read_thread()
-        finally:
-            self.videocapture = None
-            try:
-                if videocapture is not None:
-                    videocapture.release()
-            finally:
-                # Releasing the device may unblock a hardware read that outlived
-                # the first bounded join in _stop_read_thread().
-                if read_thread is not None and read_thread.is_alive():
-                    read_thread.join(timeout=2.0)
-                    if read_thread.is_alive():  # pragma: no cover
-                        logger.warning(f"{self} read thread remained alive after releasing the capture.")
-
    @check_if_not_connected
    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
        """
@@ -626,6 +586,16 @@ class OpenCVCamera(Camera):
        if not self.is_connected and self.thread is None:
            raise DeviceNotConnectedError(f"{self} not connected.")

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

        logger.info(f"{self} disconnected.")
@@ -121,9 +121,6 @@ class RealSenseCamera(Camera):

        self.config = config

-        self.width: int | None = config.width
-        self.height: int | None = config.height
-
        if config.serial_number_or_name.isdigit():
            self.serial_number = config.serial_number_or_name
        else:
@@ -134,9 +131,6 @@ class RealSenseCamera(Camera):
        self.use_rgb = config.use_rgb
        self.use_depth = config.use_depth
        self.warmup_s = config.warmup_s
-        self.exposure: int | None = config.exposure
-        self.gain: int | None = config.gain
-        self.white_balance: int | None = config.white_balance

        self.rs_pipeline: rs.pipeline | None = None
        self.rs_profile: rs.pipeline_profile | None = None
@@ -151,23 +145,14 @@ class RealSenseCamera(Camera):

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

-        self.capture_width: int | None = None
-        self.capture_height: int | None = None
-        self._reset_connection_settings()
-
-    def __str__(self) -> str:
-        return f"{self.__class__.__name__}({self.serial_number})"
-
-    def _reset_connection_settings(self) -> None:
-        """Restore settings that may have been auto-detected during a failed connection."""
-        self.fps = self.config.fps
-        self.width = self.config.width
-        self.height = self.config.height
-        self.warmup_s = self.config.warmup_s
+        if self.height and self.width:
            self.capture_width, self.capture_height = self.width, self.height
            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
                self.capture_width, self.capture_height = self.height, self.width

+    def __str__(self) -> str:
+        return f"{self.__class__.__name__}({self.serial_number})"
+
    @property
    def is_connected(self) -> bool:
        """Checks if the camera pipeline is started and streams are active."""
@@ -187,8 +172,7 @@ class RealSenseCamera(Camera):

        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.
        """
@@ -206,9 +190,7 @@ class RealSenseCamera(Camera):
                f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
            ) from e

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

        # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
@@ -224,13 +206,6 @@ class RealSenseCamera(Camera):
                self.use_depth and self.latest_depth_frame is None
            ):
                raise ConnectionError(f"{self} failed to capture frames during warmup.")
-        except BaseException:
-            try:
-                self._cleanup_resources()
-            except Exception:
-                logger.exception(f"Failed to fully clean up {self} after connect() failed.")
-            self._reset_connection_settings()
-            raise

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

@@ -364,114 +339,6 @@ class RealSenseCamera(Camera):
        self.new_frame_event.clear()
        return self._async_read(timeout_ms=10000, read_depth=read_depth)

-    def _get_color_sensor(self) -> "rs.sensor":
-        """Returns the dedicated "RGB Camera" sensor that controls the color stream.
-
-        Manual color controls are only applied to a dedicated RGB module. Cameras
-        without one (e.g. the D405, whose color stream comes from the shared
-        "Stereo Module") are unsupported, so we never fall back to another sensor
-        to avoid altering the depth stream.
-        """
-        if self.rs_profile is None:
-            raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
-
-        device = self.rs_profile.get_device()
-        sensors = {s.get_info(rs.camera_info.name): s for s in device.query_sensors()}
-
-        if "RGB Camera" in sensors:
-            return sensors["RGB Camera"]
-
-        available = list(sensors.keys())
-        raise RuntimeError(
-            f"{self}: manual color controls require a dedicated 'RGB Camera' module, which this camera does not have. ",
-            f"Available sensors: {available}.",
-        )
-
-    def _set_sensor_option(self, sensor: "rs.sensor", option: "rs.option", value: float, label: str) -> None:
-        """Sets a sensor option, re-raising range errors with actionable diagnostics."""
-        try:
-            sensor.set_option(option, value)
-        except Exception as e:
-            range_info = ""
-            try:
-                option_range = sensor.get_option_range(option)
-                range_info = (
-                    f" (supported range: min={option_range.min}, max={option_range.max}, "
-                    f"step={option_range.step}, default={option_range.default})"
-                )
-            except Exception:
-                range_info = " (option range unavailable)"
-            raise ValueError(
-                f"{self}: failed to set {label} to {value}{range_info}. Original error: {e}"
-            ) from e
-
-    def _configure_sensor_options(self) -> None:
-        """Applies manual sensor options (exposure, gain, white balance) to the color sensor.
-
-        When exposure or gain is set, auto-exposure is disabled first. When white_balance
-        is set, auto white balance is disabled first. An omitted option is left unchanged,
-        and configuration is skipped entirely if all options are omitted.
-
-        Raises:
-            ValueError: If the sensor does not support a requested option or a requested
-                value is invalid. Invalid-value errors include the option name, requested
-                value, and supported range when available.
-        """
-        if self.exposure is None and self.gain is None and self.white_balance is None:
-            return
-
-        color_sensor = self._get_color_sensor()
-
-        requested_options = (
-            (rs.option.exposure, self.exposure, "exposure"),
-            (rs.option.gain, self.gain, "gain"),
-            (rs.option.white_balance, self.white_balance, "white balance"),
-        )
-        unsupported_options = [
-            label
-            for option, value, label in requested_options
-            if value is not None and not color_sensor.supports(option)
-        ]
-        if unsupported_options:
-            raise ValueError(
-                f"{self}: color sensor does not support requested manual options: {unsupported_options}."
-            )
-
-        manual_exposure_requested = self.exposure is not None or self.gain is not None
-        if manual_exposure_requested:
-            if color_sensor.supports(rs.option.enable_auto_exposure):
-                self._set_sensor_option(color_sensor, rs.option.enable_auto_exposure, 0, "auto-exposure")
-                logger.info(f"{self} auto-exposure disabled.")
-            else:
-                logger.warning(
-                    f"{self} sensor does not support disabling auto-exposure; "
-                    "applying manual exposure/gain directly."
-                )
-
-        if self.exposure is not None:
-            self._set_sensor_option(color_sensor, rs.option.exposure, self.exposure, "exposure")
-            logger.info(f"{self} exposure set to {self.exposure}.")
-
-        if self.gain is not None:
-            self._set_sensor_option(color_sensor, rs.option.gain, self.gain, "gain")
-            logger.info(f"{self} gain set to {self.gain}.")
-
-        if self.white_balance is not None:
-            if color_sensor.supports(rs.option.enable_auto_white_balance):
-                self._set_sensor_option(
-                    color_sensor, rs.option.enable_auto_white_balance, 0, "auto white balance"
-                )
-                logger.info(f"{self} auto white balance disabled.")
-            else:
-                logger.warning(
-                    f"{self} sensor does not support disabling auto white balance; "
-                    "applying manual white balance directly."
-                )
-            self._set_sensor_option(
-                color_sensor, rs.option.white_balance, self.white_balance, "white balance"
-            )
-            logger.info(f"{self} white balance set to {self.white_balance}.")
-
    @check_if_not_connected
    def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
        """
@@ -674,27 +541,6 @@ class RealSenseCamera(Camera):
            self.latest_timestamp = None
            self.new_frame_event.clear()

-    def _cleanup_resources(self) -> None:
-        """Stop background reads and stop the pipeline, including after partial setup."""
-        read_thread = self.thread
-        rs_pipeline = self.rs_pipeline
-
-        try:
-            self._stop_read_thread()
-        finally:
-            self.rs_pipeline = None
-            self.rs_profile = None
-            try:
-                if rs_pipeline is not None:
-                    rs_pipeline.stop()
-            finally:
-                # Stopping the pipeline may unblock a hardware read that outlived
-                # the first bounded join in _stop_read_thread().
-                if read_thread is not None and read_thread.is_alive():
-                    read_thread.join(timeout=2.0)
-                    if read_thread.is_alive():  # pragma: no cover
-                        logger.warning(f"{self} read thread remained alive after stopping the pipeline.")
-
    def _async_read(self, timeout_ms: float, read_depth: bool = False) -> NDArray[Any]:
        """Shared helper for :meth:`async_read`/:meth:`async_read_depth`: return the latest buffered frame."""
        if self.thread is None or not self.thread.is_alive():
@@ -838,5 +684,18 @@ class RealSenseCamera(Camera):
                f"Attempted to disconnect {self}, but it appears already disconnected."
            )

-        self._cleanup_resources()
+        if self.thread is not None:
+            self._stop_read_thread()
+
+        if self.rs_pipeline is not None:
+            self.rs_pipeline.stop()
+            self.rs_pipeline = None
+            self.rs_profile = None
+
+        with self.frame_lock:
+            self.latest_color_frame = None
+            self.latest_depth_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
        logger.info(f"{self} disconnected.")
@@ -46,17 +46,6 @@ class RealSenseCameraConfig(CameraConfig):
        use_depth: Whether to enable depth stream. Defaults to False.
        rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
        warmup_s: Time reading frames before returning from connect (in seconds)
-        exposure: Manual exposure value for the color sensor. When set, auto-exposure is
-            disabled and this fixed value is used. Valid ranges are camera-model specific
-            and reported if the value is rejected. Defaults to None (leave unchanged).
-        gain: Manual gain value for the color sensor. When set, auto-exposure is disabled
-            and this fixed gain is used, which also freezes exposure at its current value
-            when no exposure is configured. Valid ranges are camera-model specific and
-            reported if the value is rejected. Defaults to None (leave unchanged).
-        white_balance: Manual white balance value for the color sensor. When set, auto
-            white balance is disabled and this fixed value is used. Valid ranges are
-            camera-model specific and reported if the value is rejected. Defaults to None
-            (leave unchanged).

    Note:
        - Either name or serial_number must be specified.
@@ -72,9 +61,6 @@ class RealSenseCameraConfig(CameraConfig):
    use_depth: bool = False
    rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
    warmup_s: int = 1
-    exposure: int | None = None
-    gain: int | None = None
-    white_balance: int | None = None

    def __post_init__(self) -> None:
        self.color_mode = ColorMode(self.color_mode)
@@ -83,18 +69,6 @@ class RealSenseCameraConfig(CameraConfig):
        if not self.use_rgb and not self.use_depth:
            raise ValueError("At least one of `use_rgb` or `use_depth` must be enabled.")

-        manual_color_options = {
-            "exposure": self.exposure,
-            "gain": self.gain,
-            "white_balance": self.white_balance,
-        }
-        configured_color_options = [name for name, value in manual_color_options.items() if value is not None]
-        if configured_color_options and not self.use_rgb:
-            raise ValueError(
-                "Manual color sensor options require `use_rgb=True`. "
-                f"Configured options: {configured_color_options}."
-            )
-
        values = (self.fps, self.width, self.height)
        if any(v is not None for v in values) and any(v is None for v in values):
            raise ValueError(
@@ -35,9 +35,9 @@ else:

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


 def predict_action(
@@ -52,17 +52,6 @@ def get_step_checkpoint_dir(output_dir: Path, total_steps: int, step: int) -> Pa
    return output_dir / CHECKPOINTS_DIR / step_identifier


-def should_save_checkpoint(step: int, save_freq: int, total_steps: int) -> bool:
-    """Whether a checkpoint should be saved at ``step``.
-
-    A checkpoint is saved every ``save_freq`` steps and always after the final step. A
-    non-positive ``save_freq`` disables periodic saving (only the final checkpoint is
-    written), mirroring how ``log_freq``/``eval_freq`` treat non-positive values and
-    avoiding a ``ZeroDivisionError`` from ``step % 0``.
-    """
-    return (save_freq > 0 and step % save_freq == 0) or step == total_steps
-
-
 def save_training_step(
    step: int, save_dir: Path, num_processes: int | None = None, batch_size: int | None = None
 ) -> None:
@@ -71,19 +71,13 @@ class DatasetRecordConfig:
    # Number of threads per encoder instance. None = auto (codec default).
    # Lower values reduce CPU usage, maps to 'lp' (via svtav1-params) for libsvtav1 and 'threads' for h264/hevc..
    encoder_threads: int | None = None
-    # Skip appending the date-time tag to repo_id, keeping the user-provided name as-is
-    # (e.g. self-managed versioned names intended for a later `lerobot-edit-dataset merge`).
-    no_stamp: bool = False

    def stamp_repo_id(self) -> None:
        """Append a date-time tag to ``repo_id`` so each recording session gets a unique name.

        Must be called explicitly at dataset *creation* time — not on resume,
        where the existing ``repo_id`` (already stamped) must be preserved.
-        No-op when ``no_stamp`` is set, preserving a user-managed ``repo_id``.
        """
-        if self.no_stamp:
-            return
        if self.repo_id:
            timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
            self.repo_id = f"{self.repo_id}_{timestamp}"
@@ -33,6 +33,8 @@ class DatasetConfig:
    # 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
@@ -62,6 +64,10 @@ 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 and any(ep < 0 for ep in self.exclude_episodes):
+            raise ValueError(
+                f"exclude_episodes must be non-negative, got: {[ep for ep in self.exclude_episodes if ep < 0]}"
+            )


@dataclass
@@ -27,13 +27,6 @@ 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

@@ -79,18 +72,11 @@ 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}."
-    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])
+    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)

    return attr_level_args

@@ -98,12 +84,10 @@ 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:]
-    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]
+    prefix = f"--{arg_name}="
+    for arg in args:
+        if arg.startswith(prefix):
+            return arg[len(prefix) :]
    return None


@@ -111,7 +95,7 @@ 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 accepts arguments in the formats '--key=value' and '--key value' and returns them as a dictionary.
+    It processes arguments in the format '--key=value' and returns them as a dictionary.

    Args:
        plugin_arg_suffix (str): The suffix to identify plugin-related arguments.
@@ -128,17 +112,12 @@ def parse_plugin_args(plugin_arg_suffix: str, args: Sequence[str]) -> dict[str,
        {'env.discover_packages_path': 'my_package'}
    """
    plugin_args = {}
-    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]
+    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
    return plugin_args

@@ -206,82 +185,10 @@ 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 []
-    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
+    return [arg for arg in args if not arg.startswith(f"--{field_to_filter}=")]


 def filter_path_args(fields_to_filter: str | list[str], args: Sequence[str] | None = None) -> list[str]:
@@ -313,23 +220,7 @@ 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}",
                )
-            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
+            filtered_args = [arg for arg in filtered_args if not arg.startswith(f"--{field}.")]

    return filtered_args

@@ -408,9 +299,6 @@ 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__()
@@ -418,7 +306,6 @@ 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)
-                try:
                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)
@@ -430,9 +317,6 @@ def wrap(config_path: Path | None = None) -> Callable[[F], F]:
                        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

@@ -163,10 +163,8 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
        return None

    def _save_pretrained(self, save_directory: Path) -> None:
-        # 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)
+        with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
+            draccus.dump(self, f, indent=4)

    @classmethod
    def from_pretrained(
@@ -78,7 +78,7 @@ 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:
@@ -147,7 +147,7 @@ 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."""
+        """Validate bindings and require text or low-level action supervision."""
        assert self.messages is not None
        known_bindings = set(DEFAULT_BINDINGS) | set(self.bindings or {}) | {"task"}

@@ -156,8 +156,14 @@ 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."""
@@ -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 +1,70 @@
+# 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
+    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
+    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}
@@ -103,10 +103,8 @@ class RewardModelConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
        pass

    def _save_pretrained(self, save_directory: Path) -> None:
-        # 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)
+        with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
+            draccus.dump(self, f, indent=4)

    @classmethod
    def from_pretrained(
@@ -119,7 +119,6 @@ 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
@@ -195,11 +194,7 @@ class TrainPipelineConfig(HubMixin):
            )

        if Path(config_path).resolve().exists():
-            # `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
+            policy_dir = Path(config_path).parent
            self.checkpoint_path = policy_dir.parent
        elif self.job.is_remote:
            return
@@ -19,7 +19,6 @@ import copy
 import logging
 import shutil
 from pathlib import Path
-from typing import Any, NotRequired, TypedDict

 import datasets
 import pandas as pd
@@ -50,32 +49,8 @@ from .utils import (
 )
 from .video_utils import concatenate_video_files, get_video_duration_in_s

-logger = logging.getLogger(__name__)

-type FeatureDict = dict[str, dict[str, Any]]
-type ChunkFile = tuple[int, int]
-
-
-class IndexState(TypedDict):
-    chunk: int
-    file: int
-    src_to_dst: NotRequired[dict[ChunkFile, ChunkFile]]
-
-
-class VideoIndex(TypedDict):
-    chunk: int
-    file: int
-    latest_duration: float
-    episode_duration: float
-    src_to_offset: NotRequired[dict[ChunkFile, float]]
-    src_to_dst: NotRequired[dict[ChunkFile, ChunkFile]]
-    dst_file_durations: NotRequired[dict[ChunkFile, float]]
-
-
-type VideoIndexState = dict[str, VideoIndex]
-
-
-def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> FeatureDict:
+def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> dict[str, dict]:
    """Create a merged video feature info dictionary for aggregation. The video encoder info is merged field-by-field: each key is kept only when every source agrees; otherwise that key is set to ``null`` (or ``{}`` for ``video.extra_options``) and a warning is logged.

    Args:
@@ -84,14 +59,14 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
    Returns:
        dict: A dictionary of merged video feature info.
    """
-    merged_info: FeatureDict = copy.deepcopy(all_metadata[0].features)
+    merged_info = copy.deepcopy(all_metadata[0].features)
    video_keys = [k for k in merged_info if merged_info[k].get("dtype") == "video"]

    for vk in video_keys:
        video_infos = [m.features.get(vk, {}).get("info") or {} for m in all_metadata]
        base_video_info = video_infos[0]

-        merged_encoder_info: dict[str, Any] = {}
+        merged_encoder_info: dict = {}
        fallback_keys: list[str] = []
        for info_key in VIDEO_ENCODER_INFO_KEYS:
            values = [info.get(info_key, None) for info in video_infos]
@@ -105,7 +80,7 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
                merged_encoder_info[info_key] = {} if info_key == "video.extra_options" else None

        if fallback_keys:
-            logger.warning(
+            logging.warning(
                f"Merging heterogeneous or incomplete video encoder metadata for feature {vk}. "
                f"Setting these keys to null: {fallback_keys}.",
            )
@@ -117,7 +92,7 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
    return merged_info


-def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]) -> tuple[int, str | None, FeatureDict]:
+def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
    """Validates that all dataset metadata have consistent properties.

    Ensures all datasets have the same fps, robot_type, and features to guarantee
@@ -154,9 +129,7 @@ def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]) -> tuple[i
    return fps, robot_type, features


-def update_data_df(
-    df: pd.DataFrame, src_meta: LeRobotDatasetMetadata, dst_meta: LeRobotDatasetMetadata
-) -> pd.DataFrame:
+def update_data_df(df, src_meta, dst_meta):
    """Updates a data DataFrame with new indices and task mappings for aggregation.

    Adjusts episode indices, frame indices, and task indices to account for
@@ -181,12 +154,12 @@ def update_data_df(


 def update_meta_data(
-    df: pd.DataFrame,
-    dst_meta: LeRobotDatasetMetadata,
-    meta_idx: IndexState,
-    data_idx: IndexState,
-    videos_idx: VideoIndexState,
-) -> pd.DataFrame:
+    df,
+    dst_meta,
+    meta_idx,
+    data_idx,
+    videos_idx,
+):
    """Updates metadata DataFrame with new chunk, file, and timestamp indices.

    Adjusts all indices and timestamps to account for previously aggregated
@@ -316,7 +289,7 @@ def aggregate_datasets(
    chunk_size: int | None = None,
    concatenate_videos: bool = True,
    concatenate_data: bool = True,
-) -> None:
+):
    """Aggregates multiple LeRobot datasets into a single unified dataset.

    This is the main function that orchestrates the aggregation process by:
@@ -336,7 +309,7 @@ def aggregate_datasets(
        concatenate_videos: When False, keep one mp4 per source file instead of packing into shards.
        concatenate_data: When False, keep one parquet per source file instead of packing into shards.
    """
-    logger.info("Start aggregate_datasets")
+    logging.info("Start aggregate_datasets")

    if data_files_size_in_mb is None:
        data_files_size_in_mb = DEFAULT_DATA_FILE_SIZE_IN_MB
@@ -368,15 +341,15 @@ def aggregate_datasets(
        video_files_size_in_mb=video_files_size_in_mb,
    )

-    logger.info("Find all tasks")
+    logging.info("Find all tasks")
    unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
    dst_meta.tasks = pd.DataFrame(
        {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
    )

-    meta_idx: IndexState = {"chunk": 0, "file": 0}
-    data_idx: IndexState = {"chunk": 0, "file": 0}
-    videos_idx: VideoIndexState = {
+    meta_idx = {"chunk": 0, "file": 0}
+    data_idx = {"chunk": 0, "file": 0}
+    videos_idx = {
        key: {"chunk": 0, "file": 0, "latest_duration": 0, "episode_duration": 0} for key in video_keys
    }

@@ -400,17 +373,12 @@ def aggregate_datasets(
        dst_meta.info.total_frames += src_meta.total_frames

    finalize_aggregation(dst_meta, all_metadata)
-    logger.info("Aggregation complete.")
+    logging.info("Aggregation complete.")


 def aggregate_videos(
-    src_meta: LeRobotDatasetMetadata,
-    dst_meta: LeRobotDatasetMetadata,
-    videos_idx: VideoIndexState,
-    video_files_size_in_mb: float,
-    chunk_size: int,
-    concatenate_videos: bool = True,
-) -> VideoIndexState:
+    src_meta, dst_meta, videos_idx, video_files_size_in_mb, chunk_size, concatenate_videos=True
+):
    """Aggregates video chunks from a source dataset into the destination dataset.

    Handles video file concatenation and rotation based on file size limits.
@@ -438,8 +406,7 @@ def aggregate_videos(
            videos_idx[key]["dst_file_durations"] = {}

    for key, video_idx in videos_idx.items():
-        unique_chunk_file_pairs: list[ChunkFile] = sorted(
-            {
+        unique_chunk_file_pairs = {
            (chunk, file)
            for chunk, file in zip(
                src_meta.episodes[f"videos/{key}/chunk_index"],
@@ -447,7 +414,7 @@ def aggregate_videos(
                strict=False,
            )
        }
-        )
+        unique_chunk_file_pairs = sorted(unique_chunk_file_pairs)

        chunk_idx = video_idx["chunk"]
        file_idx = video_idx["file"]
@@ -522,14 +489,7 @@ def aggregate_videos(
    return videos_idx


-def aggregate_data(
-    src_meta: LeRobotDatasetMetadata,
-    dst_meta: LeRobotDatasetMetadata,
-    data_idx: IndexState,
-    data_files_size_in_mb: float,
-    chunk_size: int,
-    concatenate_data: bool = True,
-) -> IndexState:
+def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_size, concatenate_data=True):
    """Aggregates data chunks from a source dataset into the destination dataset.

    Reads source data files, updates indices to match the aggregated dataset,
@@ -550,16 +510,14 @@ def aggregate_data(
    Returns:
        dict: Updated data_idx with current chunk and file indices.
    """
-    unique_chunk_file_ids: list[ChunkFile] = sorted(
-        {
+    unique_chunk_file_ids = {
        (c, f)
        for c, f in zip(
-                src_meta.episodes["data/chunk_index"],
-                src_meta.episodes["data/file_index"],
-                strict=False,
+            src_meta.episodes["data/chunk_index"], src_meta.episodes["data/file_index"], strict=False
        )
    }
-    )
+
+    unique_chunk_file_ids = sorted(unique_chunk_file_ids)
    contains_images = len(dst_meta.image_keys) > 0

    # retrieve features schema for proper image typing in parquet
@@ -567,7 +525,7 @@ def aggregate_data(

    # Track source to destination file mapping for metadata update
    # This is critical for handling datasets that are already results of a merge
-    src_to_dst: dict[ChunkFile, ChunkFile] = {}
+    src_to_dst: dict[tuple[int, int], tuple[int, int]] = {}

    for src_chunk_idx, src_file_idx in unique_chunk_file_ids:
        src_path = src_meta.root / DEFAULT_DATA_PATH.format(
@@ -606,13 +564,7 @@ def aggregate_data(
    return data_idx


-def aggregate_metadata(
-    src_meta: LeRobotDatasetMetadata,
-    dst_meta: LeRobotDatasetMetadata,
-    meta_idx: IndexState,
-    data_idx: IndexState,
-    videos_idx: VideoIndexState,
-) -> IndexState:
+def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
    """Aggregates metadata from a source dataset into the destination dataset.

    Reads source metadata files, updates all indices and timestamps,
@@ -628,8 +580,7 @@ def aggregate_metadata(
    Returns:
        dict: Updated meta_idx with current chunk and file indices.
    """
-    chunk_file_ids: list[ChunkFile] = sorted(
-        {
+    chunk_file_ids = {
        (c, f)
        for c, f in zip(
            src_meta.episodes["meta/episodes/chunk_index"],
@@ -637,7 +588,8 @@ def aggregate_metadata(
            strict=False,
        )
    }
-    )
+
+    chunk_file_ids = sorted(chunk_file_ids)
    for chunk_idx, file_idx in chunk_file_ids:
        src_path = src_meta.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
        df = pd.read_parquet(src_path)
@@ -670,16 +622,16 @@ def aggregate_metadata(
 def append_or_create_parquet_file(
    df: pd.DataFrame,
    src_path: Path,
-    idx: IndexState,
+    idx: dict[str, int],
    max_mb: float,
    chunk_size: int,
    default_path: str,
    contains_images: bool = False,
-    aggr_root: Path | None = None,
+    aggr_root: Path = None,
    hf_features: datasets.Features | None = None,
    concatenate: bool = True,
    one_row_group_per_episode: bool = False,
-) -> tuple[IndexState, ChunkFile]:
+) -> tuple[dict[str, int], tuple[int, int]]:
    """Appends data to an existing parquet file or creates a new one based on size constraints.

    Manages file rotation when size limits are exceeded to prevent individual files
@@ -702,13 +654,7 @@ def append_or_create_parquet_file(
    Returns:
        tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
               and (dst_chunk, dst_file) is the actual destination file the data was written to.
-
-    Raises:
-        ValueError: If aggr_root is not provided.
    """
-    if aggr_root is None:
-        raise ValueError("aggr_root must be provided.")
-
    dst_chunk, dst_file = idx["chunk"], idx["file"]
    dst_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)

@@ -752,9 +698,7 @@ def append_or_create_parquet_file(
    return idx, (dst_chunk, dst_file)


-def finalize_aggregation(
-    aggr_meta: LeRobotDatasetMetadata, all_metadata: list[LeRobotDatasetMetadata]
-) -> None:
+def finalize_aggregation(aggr_meta, all_metadata):
    """Finalizes the dataset aggregation by writing summary files and statistics.

    Writes the tasks file, info file with total counts and splits, and
@@ -764,16 +708,16 @@ def finalize_aggregation(
        aggr_meta: Aggregated dataset metadata.
        all_metadata: List of all source dataset metadata objects.
    """
-    logger.info("write tasks")
+    logging.info("write tasks")
    write_tasks(aggr_meta.tasks, aggr_meta.root)

-    logger.info("write info")
+    logging.info("write info")
    aggr_meta.info.total_tasks = len(aggr_meta.tasks)
    aggr_meta.info.total_episodes = sum(m.total_episodes for m in all_metadata)
    aggr_meta.info.total_frames = sum(m.total_frames for m in all_metadata)
    aggr_meta.info.splits = {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"}
    write_info(aggr_meta.info, aggr_meta.root)

-    logger.info("write stats")
+    logging.info("write stats")
    aggr_meta.stats = aggregate_stats([m.stats for m in all_metadata])
    write_stats(aggr_meta.stats, aggr_meta.root)
@@ -18,15 +18,13 @@ 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, sync_bucket
-from huggingface_hub.utils import WeakFileLock
+from huggingface_hub import snapshot_download

 from lerobot.configs import DEPTH_METER_UNIT, VideoEncoderConfig
 from lerobot.utils.constants import DEFAULT_FEATURES, HF_LEROBOT_HOME, HF_LEROBOT_HUB_CACHE
@@ -76,7 +74,6 @@ 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.
@@ -99,38 +96,21 @@ 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
-        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.root = self._requested_root if self._requested_root is not None else 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

-        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")
-
-        with metadata_lock:
        try:
            if force_cache_sync or (
                self._requested_root is None and has_legacy_hub_download_metadata(self.root)
@@ -138,7 +118,7 @@ class LeRobotDatasetMetadata:
                raise FileNotFoundError
            self._load_metadata()
        except (FileNotFoundError, NotADirectoryError):
-                if self.repo_type != "bucket" and is_valid_version(self.revision):
+            if is_valid_version(self.revision):
                if token is None:
                    self.revision = get_safe_version(self.repo_id, self.revision)
                else:
@@ -174,12 +154,6 @@ 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)

@@ -214,8 +188,8 @@ class LeRobotDatasetMetadata:
    def _load_metadata(self):
        self.info = load_info(self.root)
        check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
-        self.tasks = load_tasks(self.root) if self.total_tasks > 0 else None
-        self.episodes = load_episodes(self.root) if self.total_episodes > 0 else None
+        self.tasks = load_tasks(self.root)
+        self.episodes = load_episodes(self.root)
        self.stats = load_stats(self.root)

    def ensure_readable(self) -> None:
@@ -258,16 +232,6 @@ 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(
@@ -298,8 +262,6 @@ 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
@@ -163,10 +163,40 @@ 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)
+        # Annotated datasets may have language columns absent from metadata.
+        # Extend the schema before the strict Parquet cast.
+        features = self._extend_features_with_language_columns(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 _extend_features_with_language_columns(self, features: datasets.Features) -> datasets.Features:
+        """Register language columns found in Parquet but missing from metadata."""
+        # Leave empty datasets to fail through the normal loading path.
+        try:
+            sample = next((self.root / "data").glob("*/*.parquet"))
+        except StopIteration:
+            return features
+
+        from pyarrow import parquet as _pq  # noqa: PLC0415
+
+        schema_names = set(_pq.read_schema(sample).names)
+        from .language import (  # noqa: PLC0415
+            LANGUAGE_EVENTS,
+            LANGUAGE_PERSISTENT,
+            language_events_column_feature,
+            language_persistent_column_feature,
+        )
+
+        extra: dict[str, object] = {}
+        if LANGUAGE_PERSISTENT in schema_names and LANGUAGE_PERSISTENT not in features:
+            extra[LANGUAGE_PERSISTENT] = language_persistent_column_feature()
+        if LANGUAGE_EVENTS in schema_names and LANGUAGE_EVENTS not in features:
+            extra[LANGUAGE_EVENTS] = language_events_column_feature()
+        if not extra:
+            return features
+        return datasets.Features({**features, **extra})
+
    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:
@@ -1045,12 +1045,10 @@ def _copy_data_with_feature_changes(
                    df[feature_name] = feature_values
                else:
                    feature_slice = values[frame_idx:end_idx]
-                    if feature_slice.ndim == 1:
-                        df[feature_name] = feature_slice
-                    elif feature_slice.ndim == 2 and feature_slice.shape[1] == 1:
+                    if len(feature_slice.shape) > 1 and feature_slice.shape[1] == 1:
                        df[feature_name] = feature_slice.flatten()
                    else:
-                        df[feature_name] = list(feature_slice)
+                        df[feature_name] = feature_slice
            frame_idx = end_idx

        # Write using the same chunk/file structure as source
@@ -1437,18 +1435,15 @@ 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`)

-    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.
+    You can combine both: `new_task` sets the default, and `episode_tasks` overrides
+    specific episodes.

    The dataset is modified in-place, updating only the task-related files:
    - meta/tasks.parquet
@@ -1458,14 +1453,11 @@ def modify_tasks(

    Args:
        dataset: The source LeRobotDataset to modify.
-        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.
+        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.

-    At least one of `new_task`, `episode_tasks`, or `task_replacements` must be provided.

    Examples:
        Set a single task for all episodes:
@@ -1483,17 +1475,11 @@ def modify_tasks(
                new_task="Default task",
                episode_tasks={5: "Special task for episode 5"}
            )
-
-        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 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:
+    if episode_tasks is not None:
        valid_indices = set(range(dataset.meta.total_episodes))
        invalid = set(episode_tasks.keys()) - valid_indices
        if invalid:
@@ -1503,29 +1489,19 @@ 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 task_replacements and original_task in task_replacements:
-            episode_to_task[ep_idx] = task_replacements[original_task]
-        elif new_task:
+        elif new_task is not None:
            episode_to_task[ep_idx] = new_task
-        elif original_task:
-            # Keep original task if not overridden and no default provided
-            episode_to_task[ep_idx] = original_task
        else:
-            raise ValueError(f"Episode {ep_idx} has no task; provide new_task or episode_tasks")
+            # 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]

    # Collect all unique tasks and create new task mapping
    unique_tasks = sorted(set(episode_to_task.values()))
@@ -386,9 +386,6 @@ 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()
@@ -66,6 +66,17 @@ def resolve_delta_timestamps(
    return delta_timestamps


+def _resolve_episodes(
+    episodes: list[int] | None, exclude_episodes: list[int] | None, total_episodes: int
+) -> list[int] | None:
+    """Apply an episode exclusion list on top of an optional allowlist."""
+    if not exclude_episodes:
+        return episodes
+    base = episodes if episodes is not None else list(range(total_episodes))
+    excluded = set(exclude_episodes)
+    return [episode for episode in base if episode not in excluded]
+
+
 def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDataset:
    """Handles the logic of setting up delta timestamps and image transforms before creating a dataset.

@@ -87,11 +98,14 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
            cfg.dataset.repo_id, root=cfg.dataset.root, revision=cfg.dataset.revision
        )
        delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, ds_meta)
+        episodes = _resolve_episodes(
+            cfg.dataset.episodes, cfg.dataset.exclude_episodes, ds_meta.total_episodes
+        )
        if not cfg.dataset.streaming:
            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,
@@ -104,7 +118,7 @@ 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,
@@ -162,14 +162,28 @@ def render_sample(
    task: str | None = None,
    dataset_ctx: Any | None = None,
 ) -> RenderedMessages | None:
-    """Render the chat-style messages for a single dataset sample.
+    """Resolve one sample's bindings and render its message recipe.

-    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.
+    Returns ``None`` when no text or low-level action supervision applies.
    """
    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 +197,55 @@ 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.
+    """
+    assert recipe.blend is not None
+    renderable: list[tuple[float, RenderedMessages]] = []
+    for name, component in recipe.blend.items():
+        if not name.startswith("ask_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:
+            renderable.append((float(component.weight or 0.0), rendered))
+
+    if not renderable:
+        return None
+    if len(renderable) == 1:
+        return renderable[0][1]
+
+    # Choose among matching cameras by relative weight, or uniformly when all weights are zero.
+    total = sum(w for w, _ in renderable) or float(len(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 w, rendered in renderable:
+        cumulative += w or (total / len(renderable))
+        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:
@@ -346,7 +409,9 @@ 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 +468,12 @@ 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(
@@ -17,8 +17,8 @@ from collections.abc import Sequence
 from typing import Any

 from lerobot.configs import PipelineFeatureType
-from lerobot.lerobot_types import RobotAction, RobotObservation
 from lerobot.processor import DataProcessorPipeline
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE, OBS_STR
 from lerobot.utils.feature_utils import hw_to_dataset_features

@@ -16,7 +16,6 @@
 from collections import deque
 from collections.abc import Callable, Generator, Iterable, Iterator
 from pathlib import Path
-from typing import Literal

 import datasets
 import numpy as np
@@ -59,10 +58,6 @@ 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
@@ -183,7 +178,7 @@ 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:
        """
@@ -262,16 +257,15 @@ 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. 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 ``$HF_LEROBOT_HOME/hub``.
+            root (Path | None, optional): Local directory to use for local datasets. When omitted, Hub
+                metadata is resolved through a revision-safe snapshot cache under
+                ``$HF_LEROBOT_HOME/hub``.
            episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                their episode_index in this list.
            image_transforms (Callable | None, optional): Transform to apply to image data.
@@ -286,8 +280,6 @@ 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``
@@ -295,14 +287,10 @@ 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.repo_type = repo_type
-        self._requested_root = Path(root) if root is not None else None
+        self._requested_root = Path(root) if root 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 and self.repo_type == "dataset"
+        self.streaming_from_local = root is not None

        self.image_transforms = image_transforms
        self.episodes = episodes
@@ -329,7 +317,6 @@ 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
@@ -358,18 +345,7 @@ 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 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 = {}
+        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",
@@ -446,7 +422,10 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
                    else:
                        frames_buffer.append(frame)
                    break  # random shard sampled, switch shard
-            except _ShardExhaustedError:
+            except (
+                RuntimeError,
+                StopIteration,
+            ):  # NOTE: StopIteration inside a generator throws a RuntimeError since python 3.7
                del idx_to_backtrack_dataset[shard_key]  # Remove exhausted shard, onto another shard

        # Once shards are all exhausted, shuffle the buffer and yield the remaining frames
@@ -524,11 +503,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):

    def make_frame(self, dataset_iterator: Backtrackable) -> Generator:
        """Makes a frame starting from a dataset iterator"""
-        try:
        item = next(dataset_iterator)
-        except StopIteration as e:
-            # Translate exhaustion here, before PEP 479 turns it into an indistinguishable RuntimeError.
-            raise _ShardExhaustedError from e
        item = item_to_torch(item)

        updates = []  # list of "updates" to apply to the item retrieved from hf_dataset (w/o camera features)
@@ -26,6 +26,7 @@ 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

@@ -50,17 +51,6 @@ 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): ...

@@ -378,7 +368,7 @@ def get_safe_version(
        str: The safe version string (e.g., "v1.2.3") to use as a revision.

    Raises:
-        RuntimeError: If the repo has no version tags.
+        RevisionNotFoundError: If the repo has no version tags.
        BackwardCompatibilityError: If only older major versions are available.
        ForwardCompatibilityError: If only newer major versions are available.
    """
@@ -388,7 +378,17 @@ 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 RuntimeError(MISSING_VERSION_TAG_MESSAGE.format(repo_id=repo_id))
+        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")
+            ```
+            """
+        )

    if target_version in hub_versions:
        return f"v{target_version}"
@@ -328,7 +328,6 @@ 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
@@ -357,8 +356,6 @@ 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(
@@ -419,7 +416,6 @@ 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
@@ -511,7 +507,7 @@ class MetaworldEnv(EnvConfig):
 class RoboCasaEnv(EnvConfig):
    task: str = "CloseFridge"
    fps: int = 20
-    episode_length: int | None = None
+    episode_length: int = 1000
    obs_type: str = "pixels_agent_pos"
    render_mode: str = "rgb_array"
    camera_name: str = "robot0_agentview_left,robot0_eye_in_hand,robot0_agentview_right"
@@ -564,7 +560,13 @@ class RoboCasaEnv(EnvConfig):
            kwargs["split"] = self.split
        return kwargs

-    def create_envs(self, n_envs: int, use_async_envs: bool = False):
+    def create_envs(
+        self,
+        n_envs: int,
+        use_async_envs: bool = False,
+        terminate_on_success: bool = True,
+        horizon: int | None = None,
+    ):
        from .robocasa import create_robocasa_envs

        if self.task is None:
@@ -578,6 +580,8 @@ class RoboCasaEnv(EnvConfig):
            env_cls=env_cls,
            episode_length=self.episode_length,
            obj_registries=tuple(self.obj_registries),
+            terminate_on_success=terminate_on_success,
+            horizon=horizon,
        )


@@ -30,7 +30,7 @@ from gymnasium import spaces
 from libero.libero import benchmark, get_libero_path
 from libero.libero.envs import OffScreenRenderEnv

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

 from .utils import _LazyAsyncVectorEnv, parse_camera_names

@@ -128,13 +128,10 @@ 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
@@ -161,7 +158,6 @@ 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
@@ -269,9 +265,6 @@ 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
@@ -384,20 +377,14 @@ class LiberoEnv(gym.Env):
            }
        )
        observation = self._format_raw_obs(raw_obs)
-        # 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.
+        if terminated:
+            self.reset()
        truncated = False
        return observation, reward, terminated, truncated, info

    def close(self):
        if self._env is not None:
-            try:
            self._env.close()
-            finally:
-                # LIBERO deletes its inner env on close, so this wrapper must
-                # be recreated before the next reset.
-                self._env = None


 def _make_env_fns(
@@ -484,7 +471,6 @@ 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:
@@ -521,10 +507,6 @@ 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.lerobot_types import RobotObservation
+from 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.lerobot_types import RobotObservation
+from lerobot.types import RobotObservation

 from .utils import _LazyAsyncVectorEnv, parse_camera_names

@@ -33,8 +33,8 @@ logger = logging.getLogger(__name__)

 # Dimensions for the flat action/state vectors used by the LeRobot wrapper.
 # These correspond to the PandaOmron robot in RoboCasa365.
-OBS_STATE_DIM = 16  # base_pos(3) + base_quat(4) + ee_pos_rel(3) + ee_quat_rel(4) + gripper_qpos(2)
-ACTION_DIM = 12  # base_motion(4) + control_mode(1) + ee_pos(3) + ee_rot(3) + gripper(1)
+OBS_STATE_DIM = 16  # ee_pos_rel(3) + ee_quat_rel(4) + base_pos(3) + base_quat(4) + gripper_qpos(2)
+ACTION_DIM = 12  # ee_pos(3) + ee_rot(3) + gripper(1) + base_motion(4) + control_mode(1)
 ACTION_LOW = -1.0
 ACTION_HIGH = 1.0

@@ -98,30 +98,18 @@ 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.

-    Layout: base_motion(4) + control_mode(1) + ee_pos(3) + ee_rot(3) + gripper(1)
+    Layout (openpi / robocasa.utils.env_utils.convert_action order):
+        ee_pos(3) + ee_rot(3) + gripper(1) + base_motion(4) + control_mode(1)
    """
    return {
-        "action.base_motion": flat_action[0:4],
-        "action.control_mode": flat_action[4:5],
-        "action.end_effector_position": flat_action[5:8],
-        "action.end_effector_rotation": flat_action[8:11],
-        "action.gripper_close": flat_action[11:12],
+        "action.end_effector_position": flat_action[0:3],
+        "action.end_effector_rotation": flat_action[3:6],
+        "action.gripper_close": flat_action[6:7],
+        "action.base_motion": flat_action[7:11],
+        "action.control_mode": flat_action[11:12],
    }


@@ -149,9 +137,16 @@ class RoboCasaEnv(gym.Env):
        episode_length: int | None = None,
        obj_registries: Sequence[str] = DEFAULT_OBJ_REGISTRIES,
        episode_index: int = 0,
+        terminate_on_success: bool = True,
+        horizon: int | None = None,
    ):
        super().__init__()
        self.task = task
+        # When False, a task-success does NOT end/reset the episode — used by the
+        # interactive sim so one kitchen persists across sequential prompts.
+        self.terminate_on_success = terminate_on_success
+        # Underlying robosuite horizon (steps before truncation). None -> default.
+        self.horizon = horizon
        self.obs_type = obs_type
        self.render_mode = render_mode
        self.observation_width = observation_width
@@ -167,7 +162,7 @@ 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 _get_task_horizon(task)
+        self._max_episode_steps = episode_length if episode_length is not None else 1000

        # Deferred — created on first reset() inside the worker subprocess
        # to avoid inheriting stale GPU/EGL contexts across fork().
@@ -223,12 +218,16 @@ class RoboCasaEnv(gym.Env):
        # (only None/"all"/"pretrain"/"target" are valid). Always pass a
        # valid value so we don't hit that default. Extra kwargs are
        # forwarded to the underlying kitchen env via create_env/robosuite.make.
+        extra_kwargs: dict[str, Any] = {}
+        if self.horizon is not None:
+            extra_kwargs["horizon"] = int(self.horizon)
        self._env = RoboCasaGymEnv(
            env_name=self.task,
            camera_widths=self.observation_width,
            camera_heights=self.observation_height,
            split=self.split if self.split is not None else "all",
            obj_registries=self.obj_registries,
+            **extra_kwargs,
        )

        ep_meta = self._env.env.get_ep_meta()
@@ -243,12 +242,14 @@ class RoboCasaEnv(gym.Env):
            return {"pixels": images}

        # `state.*` keys come from PandaOmronKeyConverter inside the wrapper.
+        # openpi state order: ee first, then base, then gripper (matches the
+        # openpi robocasa pipeline / examples/robocasa/main.py state layout).
        agent_pos = np.concatenate(
            [
-                raw_obs.get("state.base_position", np.zeros(3)),
-                raw_obs.get("state.base_rotation", np.zeros(4)),
                raw_obs.get("state.end_effector_position_relative", np.zeros(3)),
                raw_obs.get("state.end_effector_rotation_relative", np.zeros(4)),
+                raw_obs.get("state.base_position", np.zeros(3)),
+                raw_obs.get("state.base_rotation", np.zeros(4)),
                raw_obs.get("state.gripper_qpos", np.zeros(2)),
            ],
            axis=-1,
@@ -293,7 +294,7 @@ class RoboCasaEnv(gym.Env):
        raw_obs, reward, done, truncated, info = self._env.step(action_dict)

        is_success = bool(info.get("success", False))
-        terminated = done or is_success
+        terminated = done or (is_success and self.terminate_on_success)
        info.update({"task": self.task, "done": done, "is_success": is_success})

        observation = self._format_raw_obs(raw_obs)
@@ -326,6 +327,8 @@ def _make_env_fns(
    split: str | None,
    episode_length: int | None,
    obj_registries: Sequence[str],
+    terminate_on_success: bool = True,
+    horizon: int | None = None,
 ) -> list[Callable[[], RoboCasaEnv]]:
    """Build n_envs factory callables for a single task.

@@ -348,6 +351,8 @@ def _make_env_fns(
            episode_length=episode_length,
            obj_registries=obj_registries,
            episode_index=episode_index,
+            terminate_on_success=terminate_on_success,
+            horizon=horizon,
        )

    return [partial(_make_env, i) for i in range(n_envs)]
@@ -361,6 +366,8 @@ def create_robocasa_envs(
    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
    episode_length: int | None = None,
    obj_registries: Sequence[str] = DEFAULT_OBJ_REGISTRIES,
+    terminate_on_success: bool = True,
+    horizon: int | None = None,
 ) -> dict[str, dict[int, Any]]:
    """Create vectorized RoboCasa365 environments with a consistent return shape.

@@ -422,6 +429,8 @@ def create_robocasa_envs(
            split=split,
            episode_length=episode_length,
            obj_registries=obj_registries,
+            terminate_on_success=terminate_on_success,
+            horizon=horizon,
        )

        if is_async:
@@ -28,7 +28,7 @@ import numpy as np
 import torch
 from gymnasium import spaces

-from lerobot.lerobot_types import RobotObservation
+from lerobot.types import RobotObservation
 from lerobot.utils.import_utils import _scipy_available

 from .utils import _LazyAsyncVectorEnv
@@ -384,9 +384,7 @@ class RoboTwinEnv(gym.Env):

        self._env: Any | None = None  # deferred — created on first reset() inside worker
        self._step_count: int = 0
-        self._black_frame: np.ndarray = np.zeros(
-            (self.observation_height, self.observation_width, 3), dtype=np.uint8
-        )
+        self._black_frame = np.zeros((self.observation_height, self.observation_width, 3), dtype=np.uint8)

        image_spaces = {
            cam: spaces.Box(
@@ -212,12 +212,7 @@ class _LazyAsyncVectorEnv:

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

    @property
    def unwrapped(self):
@@ -37,7 +37,7 @@ import numpy as np
 from gymnasium import spaces
 from scipy.spatial.transform import Rotation

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

 from .utils import _LazyAsyncVectorEnv

@@ -373,7 +373,7 @@ class VLABenchEnv(gym.Env):

        if action.shape[0] != 7:
            # Unknown layout — fall back to zero-pad so the sim doesn't crash.
-            padded: np.ndarray = np.zeros(ctrl_dim, dtype=np.float64)
+            padded = np.zeros(ctrl_dim, dtype=np.float64)
            padded[: min(action.shape[0], ctrl_dim)] = action[:ctrl_dim]
            return padded

@@ -122,9 +122,6 @@ MODEL_ENCODING_TABLE = {
    "xm430-w350": X_SERIES_ENCODINGS_TABLE,
    "xm540-w270": X_SERIES_ENCODINGS_TABLE,
    "xc430-w150": X_SERIES_ENCODINGS_TABLE,
-    "xh540-w150": X_SERIES_ENCODINGS_TABLE,
-    "xc330-t288": X_SERIES_ENCODINGS_TABLE,
-    "xc330-t181": X_SERIES_ENCODINGS_TABLE,
 }

 # {model: model_resolution}
@@ -137,9 +134,6 @@ MODEL_RESOLUTION = {
    "xm430-w350": 4096,
    "xm540-w270": 4096,
    "xc430-w150": 4096,
-    "xh540-w150": 4096,
-    "xc330-t288": 4096,
-    "xc330-t181": 4096,
 }

 # {model: model_number}
@@ -151,9 +145,6 @@ MODEL_NUMBER_TABLE = {
    "xm430-w350": 1020,
    "xm540-w270": 1120,
    "xc430-w150": 1070,
-    "xh540-w150": 1110,
-    "xc330-t288": 1220,
-    "xc330-t181": 1210,
 }

 # {model: available_operating_modes}
@@ -165,9 +156,6 @@ MODEL_OPERATING_MODES = {
    "xm430-w350": [0, 1, 3, 4, 5, 16],
    "xm540-w270": [0, 1, 3, 4, 5, 16],
    "xc430-w150": [1, 3, 4, 16],
-    "xh540-w150": [0, 1, 3, 4, 5, 16],
-    "xc330-t288": [0, 1, 3, 4, 5, 16],
-    "xc330-t181": [0, 1, 3, 4, 5, 16],
 }

 MODEL_CONTROL_TABLE = {
@@ -178,9 +166,6 @@ MODEL_CONTROL_TABLE = {
    "xm430-w350": X_SERIES_CONTROL_TABLE,
    "xm540-w270": X_SERIES_CONTROL_TABLE,
    "xc430-w150": X_SERIES_CONTROL_TABLE,
-    "xh540-w150": X_SERIES_CONTROL_TABLE,
-    "xc330-t288": X_SERIES_CONTROL_TABLE,
-    "xc330-t181": X_SERIES_CONTROL_TABLE,
 }

 MODEL_BAUDRATE_TABLE = {
@@ -191,9 +176,6 @@ MODEL_BAUDRATE_TABLE = {
    "xm430-w350": X_SERIES_BAUDRATE_TABLE,
    "xm540-w270": X_SERIES_BAUDRATE_TABLE,
    "xc430-w150": X_SERIES_BAUDRATE_TABLE,
-    "xh540-w150": X_SERIES_BAUDRATE_TABLE,
-    "xc330-t288": X_SERIES_BAUDRATE_TABLE,
-    "xc330-t181": X_SERIES_BAUDRATE_TABLE,
 }

 AVAILABLE_BAUDRATES = [
@@ -104,6 +104,8 @@ class AdamWConfig(OptimizerConfig):
    eps: float = 1e-8
    weight_decay: float = 1e-2
    grad_clip_norm: float = 10.0
+    foreach: bool | None = None
+    fused: bool | None = None

    def build(self, params: OptimizerParams) -> torch.optim.Optimizer:
        kwargs = asdict(self)
@@ -28,6 +28,7 @@ from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig as M
 from .pi0.configuration_pi0 import PI0Config as PI0Config
 from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
 from .pi05.configuration_pi05 import PI05Config as PI05Config
+from .pi052.configuration_pi052 import PI052Config as PI052Config
 from .pretrained import PreTrainedPolicy as PreTrainedPolicy
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
@@ -56,6 +57,7 @@ __all__ = [
    "PI0Config",
    "PI0FastConfig",
    "PI05Config",
+    "PI052Config",
    "SmolVLAConfig",
    "TDMPCConfig",
    "VLAJEPAConfig",
@@ -41,21 +41,20 @@ else:
 def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedding` (exact copy)
    time: torch.Tensor, dimension: int, min_period: float, max_period: float, device="cpu"
 ) -> Tensor:
-    """Computes sine-cosine positional embedding vectors for scalar positions."""
+    """Compute sine-cosine embeddings for scalar or per-action positions."""
    if dimension % 2 != 0:
        raise ValueError(f"dimension ({dimension}) must be divisible by 2")

-    if time.ndim != 1:
-        raise ValueError("The time tensor is expected to be of shape `(batch_size, )`.")
+    if time.ndim not in (1, 2):
+        raise ValueError("The time tensor must have shape (batch_size,) or (batch_size, action_horizon).")

    dtype = get_safe_dtype(torch.float64, device.type)
    fraction = torch.linspace(0.0, 1.0, dimension // 2, dtype=dtype, device=device)
    period = min_period * (max_period / min_period) ** fraction

-    # Compute the outer product
    scaling_factor = 1.0 / period * 2 * math.pi
-    sin_input = scaling_factor[None, :] * time[:, None]
-    return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+    sin_input = time[..., None] * scaling_factor
+    return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=-1)


 def make_att_2d_masks(pad_masks: Tensor, att_masks: Tensor) -> Tensor:  # see openpi (exact copy)
@@ -22,7 +22,6 @@ 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,
@@ -32,6 +31,7 @@ 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

--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Pepijn	12740f6be0	refactor(rtc): share Pi05 training helpers	2026-07-28 12:33:35 +02:00
Pepijn	d9c05b76aa	fix(rollout): align relative state feature order	2026-07-28 12:33:35 +02:00
Pepijn	afca390c30	fix(rtc): bootstrap trained inference after cold start	2026-07-28 12:33:34 +02:00
pepijn	777d3b126a	feat(pi05): add training-time RTC Extend trained RTC conditioning and postfix-only flow loss to plain PI05 so single-task checkpoints can use the same real-time chunking path as PI052. Co-authored-by: Cursor <cursoragent@cursor.com>	2026-07-28 12:33:34 +02:00
Pepijn	f71c99c7c9	fix(rtc): validate trained prefix capacity	2026-07-28 12:33:34 +02:00
Pepijn	f168fa4223	fix(rtc): align trained mode with RLDX	2026-07-28 12:33:34 +02:00
Pepijn	b94a93847f	feat(pi052): add training-time RTC	2026-07-28 12:33:34 +02:00
Pepijn	f2c8867df1	refactor(pi052): load native base checkpoint	2026-07-28 12:32:43 +02:00
Pepijn	6ac10f2a13	refactor(pi052): remove redundant policy code	2026-07-28 11:56:15 +02:00
Pepijn	04397777b6	docs(pi052): shorten configuration descriptions	2026-07-28 11:20:04 +02:00
Pepijn	ac197d9ad0	test(pi0_fast): cover shared tokenizer contract	2026-07-28 11:20:02 +02:00
pepijn	76171662fb	fix(pi0_fast): apply CI formatting Keep the parity tests compatible with the repository's current Ruff hooks and remove the requested migration warning from the docs. Co-authored-by: Cursor <cursoragent@cursor.com>	2026-07-28 11:18:02 +02:00
pepijn	7a05b31f83	fix(pi0_fast): align FAST semantics with OpenPI Use OpenPI-compatible normalization, token boundaries, balanced loss, and strict full-chunk decoding so training and inference share one sequence contract. Co-authored-by: Cursor <cursoragent@cursor.com>	2026-07-28 11:17:59 +02:00
Pepijn	a6f533a6dd	feat(pi052): add language-supervised policy	2026-07-28 10:08:20 +02:00
Pepijn	f2b90e3ad6	feat(runtime): add interactive language rollouts	2026-07-28 10:05:27 +02:00
Pepijn	3f093d8927	feat(data): add recipe-driven language supervision	2026-07-28 10:04:28 +02:00