feat(rl): add bus-control primitives and smooth move functionality for leader intervention

This reverts commit 38b88c414c.

.github/PULL_REQUEST_TEMPLATE.md

@@ -2,6 +2,11 @@
 
 Short, imperative summary (e.g., "fix(robots): handle None in sensor parser"). See [CONTRIBUTING.md](../CONTRIBUTING.md) for PR conventions.
 
+## Type / Scope
+
+- **Type**: (Bug | Feature | Docs | Performance | Test | CI | Chore)
+- **Scope**: (optional — name of module or package affected)
+
 ## Summary / Motivation
 
 - One-paragraph description of what changes and why.
@@ -14,14 +19,28 @@ Short, imperative summary (e.g., "fix(robots): handle None in sensor parser"). S
 
 ## What changed
 
-- Short, concrete bullets explaining the functional changes (how the behavior or output differs now).
+- Short, concrete bullets of the modifications (files/behaviour).
 - Short note if this introduces breaking changes and migration steps.
 
 ## How was this tested (or how to run locally)
 
-- Tests added: list new tests or test files. `pytest -q tests/ -k <keyword>`
+- Tests added: list new tests or test files.
 - Manual checks / dataset runs performed.
-- Instructions for the reviewer for reproducing with a quick example or CLI (if applicable)
+- Instructions for the reviewer
+
+Example:
+
+- Ran the relevant tests:
+
+  ```bash
+  pytest -q tests/ -k <keyword>
+  ```
+
+- Reproduce with a quick example or CLI (if applicable):
+
+  ```bash
+  lerobot-train --some.option=true
+  ```
 
 ## Checklist (required before merge)
 
@@ -29,7 +48,6 @@ Short, imperative summary (e.g., "fix(robots): handle None in sensor parser"). S
 - [ ] All tests pass locally (`pytest`)
 - [ ] Documentation updated
 - [ ] CI is green
-- [ ] Community Review: I have reviewed another contributor's open PR and linked it here: # (insert PR number/link)
 
 ## Reviewer notes
 

.github/workflows/benchmark_tests.yml

@@ -83,13 +83,10 @@ jobs:
           cache-binary: false
 
       - name: Login to Docker Hub
-        if: ${{ env.DOCKERHUB_USERNAME != '' }}
         uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
         with:
           username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
           password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
 
       # Build the benchmark-specific image. The Dockerfile separates dep-install
       # from source-copy, so code-only changes skip the slow uv-sync layer
@@ -118,7 +115,7 @@ jobs:
             bash -c "
               hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
               lerobot-eval \
-                --policy.path=lerobot/smolvla_libero \
+                --policy.path=pepijn223/smolvla_libero \
                 --env.type=libero \
                 --env.task=libero_spatial \
                 --eval.batch_size=1 \
@@ -147,7 +144,7 @@ jobs:
             --artifacts-dir /tmp/libero-artifacts \
             --env libero \
             --task libero_spatial \
-            --policy lerobot/smolvla_libero
+            --policy pepijn223/smolvla_libero
 
       - name: Upload Libero rollout video
         if: always()
@@ -241,13 +238,10 @@ jobs:
           cache-binary: false
 
       - name: Login to Docker Hub
-        if: ${{ env.DOCKERHUB_USERNAME != '' }}
         uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
         with:
           username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
           password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
 
       - name: Build MetaWorld benchmark image
         uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
@@ -270,7 +264,7 @@ jobs:
             bash -c "
               hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
               lerobot-eval \
-                --policy.path=lerobot/smolvla_metaworld \
+                --policy.path=pepijn223/smolvla_metaworld \
                 --env.type=metaworld \
                 --env.task=metaworld-push-v3 \
                 --eval.batch_size=1 \
@@ -299,7 +293,7 @@ jobs:
             --artifacts-dir /tmp/metaworld-artifacts \
             --env metaworld \
             --task metaworld-push-v3 \
-            --policy lerobot/smolvla_metaworld
+            --policy pepijn223/smolvla_metaworld
 
       - name: Upload MetaWorld rollout video
         if: always()
@@ -316,630 +310,3 @@ jobs:
           name: metaworld-metrics
           path: /tmp/metaworld-artifacts/metrics.json
           if-no-files-found: warn
-
-  # ── ROBOTWIN 2.0 ──────────────────────────────────────────────────────────
-  # Isolated image: full RoboTwin 2.0 stack — SAPIEN, mplib, CuRobo,
-  # pytorch3d, + simulation assets (~4 GB).
-  # Build takes ~20 min on first run; subsequent runs hit the layer cache.
-  # Requires an NVIDIA GPU runner with CUDA 12.1 drivers.
-  robotwin-integration-test:
-    name: RoboTwin 2.0 — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-      ROBOTWIN_POLICY: lerobot/smolvla_robotwin
-      ROBOTWIN_TASKS: beat_block_hammer,click_bell,handover_block,stack_blocks_two,click_alarmclock,open_microwave,adjust_bottle,lift_pot,stamp_seal,turn_switch
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        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@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      # Build the full-install image: SAPIEN, mplib, CuRobo, pytorch3d +
-      # simulation assets (~4 GB). Layer cache lives in the runner's local
-      # Docker daemon — reused across re-runs on the same machine.
-      - name: Build RoboTwin 2.0 benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.robotwin
-          push: false
-          load: true
-          tags: lerobot-benchmark-robotwin:ci
-          cache-from: type=local,src=/tmp/.buildx-cache-robotwin
-          cache-to: type=local,dest=/tmp/.buildx-cache-robotwin,mode=max
-
-      - name: Run RoboTwin 2.0 smoke eval (10 tasks, 1 episode each)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          # Named container (no --rm) so we can docker cp artifacts out.
-          docker run --name robotwin-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e ROBOTWIN_POLICY="${ROBOTWIN_POLICY}" \
-            -e ROBOTWIN_TASKS="${ROBOTWIN_TASKS}" \
-            lerobot-benchmark-robotwin:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              cd /opt/robotwin && lerobot-eval \
-                --policy.path=\"\$ROBOTWIN_POLICY\" \
-                --env.type=robotwin \
-                --env.task=\"\$ROBOTWIN_TASKS\" \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--rename_map={\"observation.images.head_camera\": \"observation.images.camera1\", \"observation.images.left_camera\": \"observation.images.camera2\", \"observation.images.right_camera\": \"observation.images.camera3\"}' \
-                --output_dir=/tmp/eval-artifacts
-              python /lerobot/scripts/ci/extract_task_descriptions.py \
-                --env robotwin \
-                --task \"\$ROBOTWIN_TASKS\" \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy RoboTwin artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/robotwin-artifacts
-          docker cp robotwin-eval:/tmp/eval-artifacts/. /tmp/robotwin-artifacts/ 2>/dev/null || true
-          docker rm -f robotwin-eval || true
-
-      - name: Parse RoboTwin eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/robotwin-artifacts \
-            --env robotwin \
-            --task "${ROBOTWIN_TASKS}" \
-            --policy "${ROBOTWIN_POLICY}"
-
-      - name: Upload RoboTwin rollout video
-        if: always()
-        uses: actions/upload-artifact@v4
-        with:
-          name: robotwin-rollout-video
-          path: /tmp/robotwin-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload RoboTwin eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4
-        with:
-          name: robotwin-metrics
-          path: /tmp/robotwin-artifacts/metrics.json
-          if-no-files-found: warn
-
-  # ── ROBOCASA365 ──────────────────────────────────────────────────────────
-  # Isolated image: robocasa + robosuite installed manually as editable
-  # clones (no `lerobot[robocasa]` extra — robocasa's setup.py pins
-  # `lerobot==0.3.3`, which would shadow this repo's lerobot).
-  robocasa-integration-test:
-    name: RoboCasa365 — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        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@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      - name: Build RoboCasa365 benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.robocasa
-          push: false
-          load: true
-          tags: lerobot-benchmark-robocasa:ci
-
-      - name: Run RoboCasa365 smoke eval (10 atomic tasks, 1 episode each)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          docker run --name robocasa-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
-            -e MUJOCO_GL=egl \
-            lerobot-benchmark-robocasa:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              lerobot-eval \
-                --policy.path=lerobot/smolvla_robocasa \
-                --env.type=robocasa \
-                --env.task=CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove,CloseToasterOvenDoor,SlideDishwasherRack,TurnOnSinkFaucet,NavigateKitchen,TurnOnElectricKettle \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--rename_map={\"observation.images.robot0_agentview_left\": \"observation.images.camera1\", \"observation.images.robot0_eye_in_hand\": \"observation.images.camera2\", \"observation.images.robot0_agentview_right\": \"observation.images.camera3\"}' \
-                --output_dir=/tmp/eval-artifacts
-              python scripts/ci/extract_task_descriptions.py \
-                --env robocasa \
-                --task CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove,CloseToasterOvenDoor,SlideDishwasherRack,TurnOnSinkFaucet,NavigateKitchen,TurnOnElectricKettle \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy RoboCasa365 artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/robocasa-artifacts
-          docker cp robocasa-eval:/tmp/eval-artifacts/. /tmp/robocasa-artifacts/ 2>/dev/null || true
-          docker rm -f robocasa-eval || true
-
-      - name: Parse RoboCasa365 eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/robocasa-artifacts \
-            --env robocasa \
-            --task atomic_smoke_10 \
-            --policy lerobot/smolvla_robocasa
-
-      - name: Upload RoboCasa365 rollout video
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robocasa-rollout-video
-          path: /tmp/robocasa-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload RoboCasa365 eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robocasa-metrics
-          path: /tmp/robocasa-artifacts/metrics.json
-          if-no-files-found: warn
-
-  # ── ROBOCEREBRA ───────────────────────────────────────────────────────────
-  # Reuses the LIBERO simulator (libero_10 suite) with RoboCerebra camera
-  # defaults (image/wrist_image). The image is layered on
-  # huggingface/lerobot-gpu, which already ships [libero] as part of [all].
-  robocerebra-integration-test:
-    name: RoboCerebra — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        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@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      - name: Build RoboCerebra benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.robocerebra
-          push: false
-          load: true
-          tags: lerobot-benchmark-robocerebra:ci
-          cache-from: type=local,src=/tmp/.buildx-cache-robocerebra
-          cache-to: type=local,dest=/tmp/.buildx-cache-robocerebra,mode=max
-
-      - name: Run RoboCerebra smoke eval (1 episode)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          docker run --name robocerebra-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
-            -e LIBERO_DATA_FOLDER=/tmp/libero_data \
-            lerobot-benchmark-robocerebra:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              lerobot-eval \
-                --policy.path=lerobot/smolvla_robocerebra \
-                --env.type=libero \
-                --env.task=libero_10 \
-                --env.fps=20 \
-                --env.obs_type=pixels_agent_pos \
-                --env.observation_height=256 \
-                --env.observation_width=256 \
-                '--env.camera_name_mapping={\"agentview_image\": \"image\", \"robot0_eye_in_hand_image\": \"wrist_image\"}' \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--rename_map={\"observation.images.image\": \"observation.images.camera1\", \"observation.images.wrist_image\": \"observation.images.camera2\"}' \
-                --policy.empty_cameras=1 \
-                --output_dir=/tmp/eval-artifacts
-              python scripts/ci/extract_task_descriptions.py \
-                --env libero --task libero_10 \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy RoboCerebra artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/robocerebra-artifacts
-          docker cp robocerebra-eval:/tmp/eval-artifacts/. /tmp/robocerebra-artifacts/ 2>/dev/null || true
-          docker rm -f robocerebra-eval || true
-
-      - name: Parse RoboCerebra eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/robocerebra-artifacts \
-            --env robocerebra \
-            --task libero_10 \
-            --policy lerobot/smolvla_robocerebra
-
-      - name: Upload RoboCerebra rollout video
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robocerebra-rollout-video
-          path: /tmp/robocerebra-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload RoboCerebra eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robocerebra-metrics
-          path: /tmp/robocerebra-artifacts/metrics.json
-          if-no-files-found: warn
-
-  # ── ROBOMME ───────────────────────────────────────────────────────────────
-  # Isolated image: mani-skill/SAPIEN/Vulkan chain with gymnasium and numpy
-  # overrides (robomme can't be a pyproject extra due to numpy<2 pin).
-  robomme-integration-test:
-    name: RoboMME — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-      ROBOMME_POLICY: lerobot/smolvla_robomme
-      ROBOMME_TASKS: PickXtimes,BinFill,StopCube,MoveCube,InsertPeg,SwingXtimes,VideoUnmask,ButtonUnmask,PickHighlight,PatternLock
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        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@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      - name: Build RoboMME benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.robomme
-          push: false
-          load: true
-          tags: lerobot-benchmark-robomme:ci
-
-      - name: Run RoboMME smoke eval (10 tasks, 1 episode each)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          docker run --name robomme-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
-            -e ROBOMME_POLICY="${ROBOMME_POLICY}" \
-            -e ROBOMME_TASKS="${ROBOMME_TASKS}" \
-            lerobot-benchmark-robomme:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              lerobot-eval \
-                --policy.path=\"\$ROBOMME_POLICY\" \
-                --env.type=robomme \
-                --env.task=\"\$ROBOMME_TASKS\" \
-                --env.dataset_split=test \
-                --env.task_ids=[0] \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--rename_map={\"observation.images.image\": \"observation.images.camera1\", \"observation.images.wrist_image\": \"observation.images.camera2\"}' \
-                --policy.empty_cameras=3 \
-                --output_dir=/tmp/eval-artifacts
-              python scripts/ci/extract_task_descriptions.py \
-                --env robomme --task \"\$ROBOMME_TASKS\" \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy RoboMME artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/robomme-artifacts
-          docker cp robomme-eval:/tmp/eval-artifacts/. /tmp/robomme-artifacts/ 2>/dev/null || true
-          docker rm -f robomme-eval || true
-
-      - name: Parse RoboMME eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/robomme-artifacts \
-            --env robomme \
-            --task "${ROBOMME_TASKS}" \
-            --policy "${ROBOMME_POLICY}"
-
-      - name: Upload RoboMME rollout video
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robomme-rollout-video
-          path: /tmp/robomme-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload RoboMME eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: robomme-metrics
-          path: /tmp/robomme-artifacts/metrics.json
-          if-no-files-found: warn
-
-  # ── LIBERO-plus ───────────────────────────────────────────────────────────
-  # Isolated image: LIBERO-plus fork cloned into /home/user_lerobot on top of
-  # huggingface/lerobot-gpu (see docker/Dockerfile.benchmark.libero_plus).
-  libero-plus-integration-test:
-    name: LIBERO-plus — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-      LIBERO_PLUS_SUITE: libero_spatial
-      LIBERO_PLUS_POLICY: lerobot/smolvla_libero_plus
-      LIBERO_PLUS_TASK_IDS: "[0,100,260,500,1000,1500,2000,2400]"
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        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@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      - name: Build LIBERO-plus benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.libero_plus
-          push: false
-          load: true
-          tags: lerobot-benchmark-libero-plus:ci
-          cache-from: type=local,src=/tmp/.buildx-cache-libero-plus
-          cache-to: type=local,dest=/tmp/.buildx-cache-libero-plus,mode=max
-
-      - name: Run LIBERO-plus smoke eval (1 episode)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          docker run --name libero-plus-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
-            -e LIBERO_PLUS_SUITE="${LIBERO_PLUS_SUITE}" \
-            -e LIBERO_PLUS_POLICY="${LIBERO_PLUS_POLICY}" \
-            -e LIBERO_PLUS_TASK_IDS="${LIBERO_PLUS_TASK_IDS}" \
-            lerobot-benchmark-libero-plus:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              lerobot-eval \
-                --policy.path=\"\$LIBERO_PLUS_POLICY\" \
-                --env.type=libero_plus \
-                --env.task=\"\$LIBERO_PLUS_SUITE\" \
-                --env.task_ids=\"\$LIBERO_PLUS_TASK_IDS\" \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--env.camera_name_mapping={\"agentview_image\": \"camera1\", \"robot0_eye_in_hand_image\": \"camera2\"}' \
-                --policy.empty_cameras=1 \
-                --output_dir=/tmp/eval-artifacts
-              python scripts/ci/extract_task_descriptions.py \
-                --env libero_plus --task \"\$LIBERO_PLUS_SUITE\" \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy LIBERO-plus artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/libero-plus-artifacts
-          docker cp libero-plus-eval:/tmp/eval-artifacts/. /tmp/libero-plus-artifacts/ 2>/dev/null || true
-          docker rm -f libero-plus-eval || true
-
-      - name: Parse LIBERO-plus eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/libero-plus-artifacts \
-            --env libero_plus \
-            --task "${LIBERO_PLUS_SUITE}" \
-            --policy "${LIBERO_PLUS_POLICY}"
-
-      - name: Upload LIBERO-plus rollout video
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: libero-plus-rollout-video
-          path: /tmp/libero-plus-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload LIBERO-plus eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: libero-plus-metrics
-          path: /tmp/libero-plus-artifacts/metrics.json
-          if-no-files-found: warn
-
-  # ── VLABENCH ─────────────────────────────────────────────────────────────
-  # Isolated image: lerobot[vlabench] only (VLABench, mujoco==3.2.2, dm-control chain)
-  vlabench-integration-test:
-    name: VLABench — build image + 1-episode eval
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          persist-credentials: false
-          lfs: true
-
-      - name: Set up Docker Buildx
-        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@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-        env:
-          DOCKERHUB_USERNAME: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-
-      - name: Build VLABench benchmark image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: docker/Dockerfile.benchmark.vlabench
-          push: false
-          load: true
-          tags: lerobot-benchmark-vlabench:ci
-          build-args: |
-            VLABENCH_ASSETS_REPO=lerobot/vlabench-assets
-
-      - name: Run VLABench smoke eval (10 tasks, 1 episode each)
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          docker run --name vlabench-eval --gpus all \
-            --shm-size=4g \
-            -e HF_HOME=/tmp/hf \
-            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
-            -e MUJOCO_GL=egl \
-            lerobot-benchmark-vlabench:ci \
-            bash -c "
-              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
-              lerobot-eval \
-                --policy.path=lerobot/smolvla_vlabench \
-                --env.type=vlabench \
-                --env.task=select_fruit,select_toy,select_book,select_painting,select_drink,select_ingredient,select_billiards,select_poker,add_condiment,insert_flower \
-                --eval.batch_size=1 \
-                --eval.n_episodes=1 \
-                --eval.use_async_envs=false \
-                --policy.device=cuda \
-                '--rename_map={\"observation.images.image\": \"observation.images.camera1\", \"observation.images.second_image\": \"observation.images.camera2\", \"observation.images.wrist_image\": \"observation.images.camera3\"}' \
-                --output_dir=/tmp/eval-artifacts
-              python scripts/ci/extract_task_descriptions.py \
-                --env vlabench \
-                --task select_fruit,select_toy,select_book,select_painting,select_drink,select_ingredient,select_billiards,select_poker,add_condiment,insert_flower \
-                --output /tmp/eval-artifacts/task_descriptions.json
-            "
-
-      - name: Copy VLABench artifacts from container
-        if: always()
-        run: |
-          mkdir -p /tmp/vlabench-artifacts
-          docker cp vlabench-eval:/tmp/eval-artifacts/. /tmp/vlabench-artifacts/ 2>/dev/null || true
-          docker rm -f vlabench-eval || true
-
-      - name: Parse VLABench eval metrics
-        if: always()
-        run: |
-          python3 scripts/ci/parse_eval_metrics.py \
-            --artifacts-dir /tmp/vlabench-artifacts \
-            --env vlabench \
-            --task select_fruit,select_toy,select_book,select_painting,select_drink,select_ingredient,select_billiards,select_poker,add_condiment,insert_flower \
-            --policy lerobot/smolvla_vlabench
-
-      - name: Upload VLABench rollout video
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: vlabench-rollout-video
-          path: /tmp/vlabench-artifacts/videos/
-          if-no-files-found: warn
-
-      - name: Upload VLABench eval metrics
-        if: always()
-        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
-        with:
-          name: vlabench-metrics
-          path: /tmp/vlabench-artifacts/metrics.json
-          if-no-files-found: warn

.github/workflows/documentation-upload-pr.yml

@@ -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@9ad2de8582b56c017cb530c1165116d40433f1c6  # main
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
     with:
       package_name: lerobot
     secrets:

.github/workflows/latest_deps_tests.yml

@@ -217,24 +217,6 @@ jobs:
       - name: Run end-to-end tests
         run: make test-end-to-end
 
-  slack-notification:
-    name: Slack Notification
-    needs: [cpu-tests, gpu-tests, upgrade-lock]
-    if: always() && needs.upgrade-lock.outputs.changed == 'true'
-    runs-on: ubuntu-latest
-    permissions:
-      contents: read
-    env:
-      CI_SLACK_CHANNEL: ${{ secrets.CI_SLACK_CHANNEL }}
-    steps:
-      - name: Post to a Slack channel
-        uses: huggingface/hf-workflows/.github/actions/post-slack@a88e7fa2eaee28de5a4d6142381b1fb792349b67  # main
-        with:
-          slack_channel: ${{ env.CI_SLACK_CHANNEL }}
-          title: "Results of the latest dependency tests (CPU + GPU)"
-          status: ${{ (needs.cpu-tests.result == 'success' && needs.gpu-tests.result == 'success') && 'success' || 'failure' }}
-          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
-
   # This job creates or updates a PR with the upgraded lockfile
   open-pr:
     name: Open PR

AGENTS.md

@@ -1,7 +1,5 @@
 This file provides guidance to AI agents when working with code in this repository.
 
-> **User-facing help → [`AGENT_GUIDE.md`](./AGENT_GUIDE.md)** (SO-101 setup, recording, picking a policy, training duration, eval — with copy-pasteable commands).
-
 ## Project Overview
 
 LeRobot is a PyTorch-based library for real-world robotics, providing datasets, pretrained policies, and tools for training, evaluation, data collection, and robot control. It integrates with Hugging Face Hub for model/dataset sharing.

AGENT_GUIDE.md

@@ -1,410 +0,0 @@
-# AGENT_GUIDE.md — LeRobot Helper for AI Agents & Users
-
-This file is a practical, copy-paste-friendly companion for any AI agent (Cursor, Claude, ChatGPT, Codex, etc.) helping a user work with LeRobot. It complements [`AGENTS.md`](./AGENTS.md) (dev/contributor context) with **user-facing guidance**: how to start, what to train, how long, how to record, and how to calibrate an SO-101.
-
----
-
-## 1. Start here — ask the user first (MANDATORY)
-
-Before suggesting any command, an agent MUST ask the user at least these questions and wait for answers:
-
-1. **What's your goal?** (e.g. "teach my SO-101 to fold a cloth", "train a policy on an existing HF dataset", "contribute a PR", "understand the codebase")
-2. **What hardware do you have?**
-   - Robot: none / SO-100 / SO-101 / Koch / LeKiwi / Reachy / other
-   - Teleop: leader arm / phone / keyboard / gamepad / none
-   - Cameras: how many, resolution, fixed or moving?
-3. **What machine will you train on?**
-   - GPU model + VRAM (e.g. "laptop 3060 6 GB", "RTX 4090 24 GB", "A100 80 GB", "CPU only")
-   - OS: macOS / Linux / Windows
-4. **Skill level & time budget?** First time, some ML, experienced? Hours, days, a weekend?
-5. **Do you already have a dataset?** Yes (HF repo id?) / no / want to record one
-6. **How can I help right now?** (pick one concrete next step)
-
-Only after you have answers, propose a concrete path. If something is ambiguous, ask again rather than guessing. Bias toward **the simplest thing that works** for the user's hardware and goal.
-
----
-
-## 2. LeRobot in 60 seconds
-
-LeRobot = **datasets + policies + envs + robot control**, unified by a small set of strong abstractions.
-
-- **`LeRobotDataset`** — episode-aware dataset (video or images + actions + state), loadable from the Hub or disk.
-- **Policies** (`ACT`, `Diffusion`, `SmolVLA`, `π0`, `π0.5`, `Wall-X`, `X-VLA`, `VQ-BeT`, `TD-MPC`, …) — all inherit `PreTrainedPolicy` and can be pushed/pulled from the Hub.
-- **Processors** — small composable transforms between dataset → policy → robot.
-- **Envs** (sim) and **Robots** (real) — same action/observation contract so code swaps cleanly.
-- **CLI** — `lerobot-record`, `lerobot-train`, `lerobot-eval`, `lerobot-teleoperate`, `lerobot-calibrate`, `lerobot-find-port`, `lerobot-setup-motors`, `lerobot-replay`.
-
-See [`AGENTS.md`](./AGENTS.md) for repo architecture.
-
----
-
-## 3. Quickstart paths (pick one)
-
-### Path A — "I have an SO-101 and want my first trained policy"
-
-Go to §4 (SO-101 end-to-end), then §5 (data tips), then §6 (pick a policy — likely **ACT**), then §7 (how long), then §8 (eval).
-
-### Path B — "No hardware, I want to train on an existing dataset"
-
-Skip §4. Pick a policy in §6, pick a duration in §7, then run `lerobot-train` per §4.9 with a Hub `--dataset.repo_id` and an `--env.type` for eval. Finish with §8.
-
-### Path C — "I just want to understand the codebase"
-
-Read §2 above, then `AGENTS.md` "Architecture", then open `src/lerobot/policies/act/` and `src/lerobot/datasets/lerobot_dataset.py` as canonical examples.
-
----
-
-## 4. SO-101 end-to-end cheat-sheet
-
-Full details in [`docs/source/so101.mdx`](./docs/source/so101.mdx) and [`docs/source/il_robots.mdx`](./docs/source/il_robots.mdx). Minimum commands in order. Confirm arms are assembled + powered before issuing.
-
-**4.1 Install**
-
-```bash
-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
-lerobot-find-port
-```
-
-macOS: `/dev/tty.usbmodem...`; Linux: `/dev/ttyACM0` (may need `sudo chmod 666 /dev/ttyACM0`).
-
-**4.3 Setup motor IDs & baudrate** (one-time, per arm)
-
-```bash
-lerobot-setup-motors --robot.type=so101_follower --robot.port=<FOLLOWER_PORT>
-lerobot-setup-motors --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>
-```
-
-**4.4 Calibrate** — center all joints, press Enter, sweep each joint through its full range. The `id` is the calibration key — reuse it everywhere.
-
-```bash
-lerobot-calibrate --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower
-lerobot-calibrate --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>   --teleop.id=my_leader
-```
-
-**4.5 Teleoperate** (sanity check, no recording)
-
-```bash
-lerobot-teleoperate \
-  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
-  --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>  --teleop.id=my_leader \
-  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-  --display_data=true
-```
-
-> **Feetech timeout / comms error on SO-100 / SO-101?** Before touching software, check the **red motor LEDs** on the daisy chain.
->
-> - **All steady red, gripper → base chain** → wiring OK.
-> - **One or more motors dark / chain stops mid-way** → wiring issue: reseat the 3-pin cables, check the controller-board power supply, and make sure each motor is fully clicked in.
-> - **LEDs blinking** → the motor is in an **error state**: usually overload (forcing a joint past its limit) **or wrong power supply voltage**. SO-100 / SO-101 ship in two variants — a **5 V / 7.4 V** build and a **12 V** build — they are NOT interchangeable. Using a 12 V PSU on a 5 V / 7.4 V arm (or vice-versa) will trip this error; confirm your motor variant before powering up.
->
-> Most "timeout" errors are physical, not code.
-
-**4.6 Record a dataset** — keys: **→** next, **←** redo, **ESC** finish & upload.
-
-```bash
-HF_USER=$(NO_COLOR=1 hf auth whoami | awk -F': *' 'NR==1 {print $2}')
-
-lerobot-record \
-  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
-  --teleop.type=so101_leader  --teleop.port=<LEADER_PORT>  --teleop.id=my_leader \
-  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-  --dataset.repo_id=${HF_USER}/my_task \
-  --dataset.single_task="<describe the task in one sentence>" \
-  --dataset.num_episodes=50 \
-  --dataset.episode_time_s=30 \
-  --dataset.reset_time_s=10 \
-  --display_data=true
-```
-
-**4.7 Visualize** — **always** do this before training. Look for missing frames, camera blur, unreachable targets, inconsistent object positions.
-After upload: https://huggingface.co/spaces/lerobot/visualize_dataset → paste `${HF_USER}/my_task`. Works for **any LeRobot-formatted Hub dataset** — use it to scout other datasets, inspect episode quality, or debug your own data before retraining.
-
-**4.8 Replay an episode** (sanity check)
-
-```bash
-lerobot-replay --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
-  --dataset.repo_id=${HF_USER}/my_task --dataset.episode=0
-```
-
-**4.9 Train** (default: ACT — fastest, lowest memory). Apple silicon: `--policy.device=mps`. See §6/§7 for policy and duration.
-
-```bash
-lerobot-train \
-  --dataset.repo_id=${HF_USER}/my_task \
-  --policy.type=act \
-  --policy.device=cuda \
-  --output_dir=outputs/train/act_my_task \
-  --job_name=act_my_task \
-  --batch_size=8 \
-  --wandb.enable=true \
-  --policy.repo_id=${HF_USER}/act_my_task
-```
-
-**4.10 Evaluate on the real robot** — compare success rate to a teleoperated baseline.
-
-```bash
-lerobot-record \
-  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
-  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-  --dataset.repo_id=${HF_USER}/eval_my_task \
-  --dataset.single_task="<same task description as training>" \
-  --dataset.num_episodes=10 \
-  --policy.path=${HF_USER}/act_my_task
-```
-
----
-
-## 5. Data collection tips (beginner → reliable policy)
-
-Good data beats clever models. Adopt these defaults and deviate only with evidence.
-
-### 5.1 Setup & ergonomics
-
-- **Fix the rig and cameras** before touching the software. If the rig vibrates or the operator gets frustrated, fix that first — more bad data won't help.
-- **Lighting matters more than resolution.** Diffuse, consistent light. Avoid moving shadows.
-- **"Can you do the task from the camera view alone?"** If no, your cameras are wrong. Fix before recording.
-- Enable **action interpolation** for rollouts when available for smoother trajectories.
-
-### 5.2 Practice before you record
-
-- Do 5–10 demos without recording. Build a deliberate, repeatable strategy.
-- Hesitant or inconsistent demos teach the model hesitation.
-
-### 5.3 Quality over speed
-
-Deliberate, high-quality execution beats fast sloppy runs. Optimize for speed only **after** strategy is dialed in — never trade quality for it.
-
-### 5.4 Consistency within and across episodes
-
-Same grasp, approach vector, and timing. Coherent strategies are much easier to learn than wildly varying movements.
-
-### 5.5 Start small, then extend (the golden rule)
-
-- **First 50 episodes = constrained version** of the task: one object, fixed position, fixed camera setup, one operator.
-- Train a quick ACT model. See what fails.
-- **Then add diversity** along one axis at a time: more positions → more lighting → more objects → more operators.
-- Don't try to collect the "perfect dataset" on day one. Iterate.
-
-### 5.6 Policy choice for beginners
-
-- **Laptop / first time / want results fast → ACT.** Works surprisingly well, trains fast even on a laptop GPU.
-- **Bigger GPU / language-conditioned / multi-task → SmolVLA.** Unfreezing the vision encoder (see §7) is a big win here.
-- Defer π0 / π0.5 / Wall-X / X-VLA until you have a proven ACT baseline and a 20+ GB GPU.
-
-### 5.7 Recommended defaults for your first task
-
-| Setting          | Value                                                                                                                                                 |
-| ---------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------- |
-| Episodes         | **50** to start, scale to 100–300 after first training                                                                                                |
-| Episode length   | 20–45 s (shorter is fine for grasp/place)                                                                                                             |
-| Reset time       | 10 s                                                                                                                                                  |
-| FPS              | 30                                                                                                                                                    |
-| Cameras          | **2 cameras recommended**: 1 fixed front + 1 wrist. Multi-view often outperforms single-view. A single fixed camera also works to keep things simple. |
-| Task description | Short, specific, action-phrased sentence                                                                                                              |
-
-### 5.8 Troubleshooting signal
-
-- Policy fails at one specific stage → record 10–20 more episodes **targeting that stage**.
-- Policy flaps / oscillates → likely inconsistent demos, or need more training; re-record worst episodes (use **←** to redo).
-- Policy ignores the object → camera framing or lighting issue, not a model issue.
-
-See also: [What makes a good dataset](https://huggingface.co/blog/lerobot-datasets#what-makes-a-good-dataset).
-
----
-
-## 6. Which policy should I train?
-
-Match the policy to the user's **GPU memory** and **time budget**. Numbers below come from an internal profiling run (one training update per policy). They are **indicative only** — see caveats.
-
-### 6.1 Profiling snapshot (indicative)
-
-All policies typically train for **5–10 epochs** (see §7).
-
-| Policy      | Batch | Update (ms) | Peak GPU mem (GB) | Best for                                                                                         |
-| ----------- | ----: | ----------: | ----------------: | ------------------------------------------------------------------------------------------------ |
-| `act`       |     4 |    **83.9** |          **0.94** | First-time users, laptops, single-task. Fast and reliable.                                       |
-| `diffusion` |     4 |       168.6 |              4.94 | Multi-modal action distributions; needs mid-range GPU.                                           |
-| `smolvla`   |     1 |       357.8 |              3.93 | Language-conditioned, multi-task, small VLA. **Unfreeze vision encoder for big gains** (see §7). |
-| `xvla`      |     1 |       731.6 |             15.52 | Large VLA, multi-task.                                                                           |
-| `wall_x`    |     1 |       716.5 |             15.95 | Large VLA with world-model objective.                                                            |
-| `pi0`       |     1 |       940.3 |             15.50 | Strong large VLA baseline (Physical Intelligence).                                               |
-| `pi05`      |     1 |      1055.8 |             16.35 | Newer π policy; similar footprint to `pi0`.                                                      |
-
-**Critical caveats:**
-
-- **Optimizer:** measured with **SGD**. LeRobot's default is **AdamW**, which keeps extra optimizer state → **peak memory will be noticeably higher** with the default, especially for `pi0`, `pi05`, `wall_x`, `xvla`.
-- **Batch size:** the large policies were profiled at batch 1. In practice use a **larger batch** for stable training (see §7.4). Memory scales roughly linearly with batch.
-
-### 6.2 Decision rules
-
-- **< 8 GB VRAM (laptop, 3060, M-series Mac):** → `act`. Maybe `diffusion` if you have ~6–8 GB free.
-- **12–16 GB VRAM (4070/4080, A4000):** → `smolvla` with defaults, or `act`/`diffusion` with larger batch. `pi0`/`pi05`/`wall_x`/`xvla` feasible only with small batch + gradient accumulation.
-- **24+ GB VRAM (3090/4090/A5000):** → any policy. Prefer `smolvla` (unfrozen) for multi-task; `act` for single-task grasp-and-place (still often the best ROI). Could experiment with `pi0` or `pi05` or `xvla`
-- **80 GB (A100/H100):** → any, with healthy batch. `pi05`, `xvla`, `wall_x` become comfortable.
-- **CPU only:** → don't train here. Use Google Colab (see [`docs/source/notebooks.mdx`](./docs/source/notebooks.mdx)) or a rented GPU.
-
----
-
-## 7. How long should I train?
-
-Robotics imitation learning usually converges in a **few epochs over the dataset**, not hundreds of thousands of raw steps. Think **epochs first**, then translate to steps.
-
-### 7.1 Rule of thumb
-
-- **Typical total: 5–10 epochs.** Start at 5, eval, then decide if more helps.
-- Very small datasets (< 30 episodes) may want slightly more epochs — but first, **collect more data**.
-- VLAs with a pretrained vision backbone typically need **fewer** epochs than training from scratch.
-
-### 7.2 Steps ↔ epochs conversion
-
-```
-total_frames     = sum of frames over all episodes      # e.g. 50 eps × 30 fps × 30 s ≈ 45,000
-steps_per_epoch  = ceil(total_frames / batch_size)
-total_steps      = epochs × steps_per_epoch
-```
-
-Examples for `--batch_size=8`:
-
-| Dataset size            |  Frames | Steps / epoch | 5 epochs | 10 epochs |
-| ----------------------- | ------: | ------------: | -------: | --------: |
-| 50 eps × 30 s @ 30 fps  |  45,000 |        ~5,625 |      28k |       56k |
-| 100 eps × 30 s @ 30 fps |  90,000 |       ~11,250 |      56k |      113k |
-| 300 eps × 30 s @ 30 fps | 270,000 |       ~33,750 |     169k |      338k |
-
-Pass the resulting total with `--steps=<N>`; eval at intermediate checkpoints (`outputs/train/.../checkpoints/`).
-
-### 7.3 Per-policy starting points (single-task, ~50 episodes)
-
-| Policy         | Batch | Steps (first run) | Notes                                                             |
-| -------------- | ----: | ----------------: | ----------------------------------------------------------------- |
-| `act`          |  8–16 |           30k–80k | Usually converges under 50k for single-task.                      |
-| `diffusion`    |  8–16 |          80k–150k | Benefits from longer training than ACT.                           |
-| `smolvla`      |   4–8 |           30k–80k | Pretrained VLM → converges fast.                                  |
-| `pi0` / `pi05` |   1–4 |           30k–80k | Memory-bound; use gradient accumulation for effective batch ≥ 16! |
-
-### 7.4 Batch size guidance
-
-- **Bigger batch is preferable** for stable gradients on teleop data.
-- If GPU memory is the bottleneck, use **gradient accumulation** to raise _effective_ batch without raising peak memory.
-- Scale **learning rate** gently with batch; most LeRobot defaults work fine for a 2–4× batch change.
-
-### 7.5 Scale LR schedule & checkpoints with `--steps`
-
-LeRobot's default schedulers (e.g. SmolVLA's cosine decay) use `scheduler_decay_steps=30_000`, which is sized for long training runs. When you shorten training (e.g. 5k–10k steps on a small dataset), **scale the scheduler down to match** — otherwise the LR stays near the peak and never decays. Same for checkpoint frequency.
-
-```bash
-lerobot-train ... \
-  --steps=5000 \
-  --policy.scheduler_decay_steps=5000 \
-  --save_freq=5000
-```
-
-Rule of thumb: set `scheduler_decay_steps ≈ steps`, and `save_freq` to whatever granularity you want for eval (e.g. every 1k–5k steps). Match `scheduler_warmup_steps` proportionally if your run is very short.
-
-### 7.6 SmolVLA: unfreeze the vision encoder for real gains
-
-SmolVLA ships with `freeze_vision_encoder=True`. Unfreezing usually **improves performance substantially** on specialized tasks, at the cost of more VRAM and slower steps. Enable with:
-
-```bash
-lerobot-train ... --policy.type=smolvla \
-  --policy.freeze_vision_encoder=false \
-  --policy.train_expert_only=false
-```
-
-### 7.7 Signals to stop / keep going
-
-- Train loss plateaus → stop, save a Hub checkpoint.
-- Train loss still dropping and you're under 10 epochs → keep going.
-
----
-
-## 8. Evaluation & benchmarks
-
-Two flavors of evaluation:
-
-### 8.1 Real-robot eval (SO-101, etc.)
-
-Reuse `lerobot-record` with `--policy.path` to run the trained policy on-robot and save the run as an eval dataset. Convention: prefix the dataset with `eval_`.
-
-```bash
-lerobot-record \
-  --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.id=my_follower \
-  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-  --dataset.repo_id=${HF_USER}/eval_my_task \
-  --dataset.single_task="<same task description used during training>" \
-  --dataset.num_episodes=10 \
-  --policy.path=${HF_USER}/act_my_task
-```
-
-Report success rate across episodes. Compare to a teleoperated baseline and to an earlier checkpoint to catch regressions.
-
-### 8.2 Sim-benchmark eval
-
-For policies trained on sim datasets (PushT, Aloha, LIBERO, MetaWorld, RoboCasa, …) use `lerobot-eval` against the matching `env.type`:
-
-```bash
-lerobot-eval \
-  --policy.path=${HF_USER}/diffusion_pusht \
-  --env.type=pusht \
-  --eval.n_episodes=50 \
-  --eval.batch_size=10 \
-  --policy.device=cuda
-```
-
-- Use `--policy.path=outputs/train/.../checkpoints/<step>/pretrained_model` for local checkpoints.
-- `--eval.n_episodes` should be ≥ 50 for a stable success-rate estimate.
-- Available envs live in `src/lerobot/envs/`. See [`docs/source/libero.mdx`](./docs/source/libero.mdx), [`metaworld.mdx`](./docs/source/metaworld.mdx), [`robocasa.mdx`](./docs/source/robocasa.mdx), [`vlabench.mdx`](./docs/source/vlabench.mdx) for specific benchmarks.
-- To add a new benchmark, see [`docs/source/adding_benchmarks.mdx`](./docs/source/adding_benchmarks.mdx) and [`envhub.mdx`](./docs/source/envhub.mdx).
-
-### 8.2b Dockerfiles for benchmark eval
-
-Benchmark envs have native dependencies that are painful to install locally. The repo ships **pre-baked Dockerfiles** for each supported benchmark — use these to run `lerobot-eval` in a reproducible environment:
-
-| Benchmark   | Dockerfile                                                                             |
-| ----------- | -------------------------------------------------------------------------------------- |
-| LIBERO      | [`docker/Dockerfile.benchmark.libero`](./docker/Dockerfile.benchmark.libero)           |
-| LIBERO+     | [`docker/Dockerfile.benchmark.libero_plus`](./docker/Dockerfile.benchmark.libero_plus) |
-| MetaWorld   | [`docker/Dockerfile.benchmark.metaworld`](./docker/Dockerfile.benchmark.metaworld)     |
-| RoboCasa    | [`docker/Dockerfile.benchmark.robocasa`](./docker/Dockerfile.benchmark.robocasa)       |
-| RoboCerebra | [`docker/Dockerfile.benchmark.robocerebra`](./docker/Dockerfile.benchmark.robocerebra) |
-| RoboMME     | [`docker/Dockerfile.benchmark.robomme`](./docker/Dockerfile.benchmark.robomme)         |
-| RoboTwin    | [`docker/Dockerfile.benchmark.robotwin`](./docker/Dockerfile.benchmark.robotwin)       |
-| VLABench    | [`docker/Dockerfile.benchmark.vlabench`](./docker/Dockerfile.benchmark.vlabench)       |
-
-Build and run (adapt to your benchmark):
-
-```bash
-docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-bench-robomme .
-docker run --gpus all --rm -it \
-  -v $HOME/.cache/huggingface:/root/.cache/huggingface \
-  lerobot-bench-robomme \
-  lerobot-eval --policy.path=<your_policy> --env.type=<env> --eval.n_episodes=50
-```
-
-See [`docker/README.md`](./docker/README.md) for base-image details.
-
-### 8.3 Target success rates
-
-Single-task grasp-and-place with 50 clean episodes: ACT should reach **> 70% success** on the training configuration. Less → data problem (see §5), not model problem. Expect a drop when generalizing to new positions — scale episodes or diversity to recover.
-
----
-
-## 9. Further reading & resources
-
-- **Getting started:** [`installation.mdx`](./docs/source/installation.mdx) · [`il_robots.mdx`](./docs/source/il_robots.mdx) · [What makes a good dataset](https://huggingface.co/blog/lerobot-datasets)
-- **Per-policy docs:** browse [`docs/source/*.mdx`](./docs/source/) (policies, hardware, benchmarks, advanced training).
-- **Community:** [Discord](https://discord.com/invite/s3KuuzsPFb) · [Hub `LeRobot` tag](https://huggingface.co/datasets?other=LeRobot) · [Dataset visualizer](https://huggingface.co/spaces/lerobot/visualize_dataset)
-
-> Keep this file current. If you learn a rule that would prevent a class of user mistakes, add it here and in [`AGENTS.md`](./AGENTS.md).

CONTRIBUTING.md

@@ -78,9 +78,6 @@ Use the templates for required fields and examples.
 - **Issues:** Follow the [ticket template](https://github.com/huggingface/lerobot/blob/main/.github/ISSUE_TEMPLATE/bug-report.yml).
 - **Pull requests:** Rebase on `upstream/main`, use a descriptive branch (don't work on `main`), run `pre-commit` and tests locally, and follow the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md).
 
-> [!IMPORTANT]
-> Community Review Policy: To help scale our efforts and foster a collaborative environment, we ask contributors to review at least one other person's open PR before their own receives attention. This shared responsibility multiplies our review capacity and helps everyone's code get merged faster!
-
-Once you have submitted your PR and completed a peer review, a member of the LeRobot team will review your contribution.
+One member of the LeRobot team will then review your contribution.
 
 Thank you for contributing to LeRobot!

docker/Dockerfile.benchmark.libero_plus

@@ -1,84 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for LIBERO-plus integration tests.
-# Extends the nightly GPU image (which has lerobot[all]) with the LIBERO-plus
-# fork source + its 6.4 GB perturbation assets.
-#
-# Build:  docker build -f docker/Dockerfile.benchmark.libero_plus -t lerobot-benchmark-libero-plus .
-# Run:    docker run --gpus all --rm lerobot-benchmark-libero-plus lerobot-eval ...
-
-FROM huggingface/lerobot-gpu:latest
-ENV MUJOCO_GL=egl
-
-# unzip for the 6.4 GB assets.zip; the rest are LIBERO-plus build-time extras
-# (wand / ImageMagick / fontconfig) not in the nightly base.
-USER root
-RUN apt-get update \
-    && apt-get install -y --no-install-recommends \
-         unzip libexpat1 libfontconfig1-dev libmagickwand-dev \
-    && apt-get clean && rm -rf /var/lib/apt/lists/*
-USER user_lerobot
-
-# robosuite==1.4.1 is mandatory (the fork uses `single_arm_env` removed in
-# v1.5+). The rest are LIBERO-plus runtime deps pulled from its setup.py.
-# We install these explicitly instead of via the [libero_plus] extra because
-# the extra's `libero @ git+...` dep installs as a namespace package and then
-# clone and PYTHONPATH-override it below.
-RUN uv pip install --no-cache \
-        "robosuite==1.4.1" \
-        "bddl==1.0.1" \
-        "easydict==1.13" \
-        "mujoco==3.7.0" \
-        "matplotlib==3.10.8" \
-        "Wand==0.6.13" \
-        "scikit-image==0.25.2" \
-        "gym==0.26.2"
-
-# Clone LIBERO-plus and make it importable as `libero`. The nightly base has
-# hf-libero (10 tasks) preinstalled via lerobot[libero]; uninstall it so
-# Python resolves `import libero` to the 2402-task LIBERO-plus module instead.
-# Pinned to the current upstream main SHA so benchmark builds stay reproducible.
-ARG LIBERO_PLUS_SHA=4976dc3
-ENV LIBERO_PLUS_ROOT=/home/user_lerobot/libero-plus/libero/libero
-RUN git clone https://github.com/sylvestf/LIBERO-plus.git /home/user_lerobot/libero-plus \
-    && git -C /home/user_lerobot/libero-plus checkout ${LIBERO_PLUS_SHA} \
-    && cd /home/user_lerobot/libero-plus && uv pip install --no-cache --no-deps -e "." \
-    && (uv pip uninstall hf-libero 2>/dev/null || true)
-ENV PYTHONPATH="/home/user_lerobot/libero-plus:${PYTHONPATH}"
-
-# Perturbation textures/scenes: bddl_base_domain.py resolves XMLs via
-# DIR_PATH/../assets (package-relative, ignoring ~/.libero/config.yaml). All
-# 2402 tasks reference files that ship only in Sylvest/LIBERO-plus's
-# assets.zip (6.4 GB) under a deep author-internal prefix — extract and
-# flatten it under ${LIBERO_PLUS_ROOT}/assets.
-RUN python -c "\
-from huggingface_hub import hf_hub_download; \
-hf_hub_download(repo_id='Sylvest/LIBERO-plus', repo_type='dataset', \
-                filename='assets.zip', local_dir='/tmp/libero-plus-dl')" \
-    && unzip -q /tmp/libero-plus-dl/assets.zip -d /tmp/libero-plus-dl/extract \
-    && ASSETS_DIR=$(find /tmp/libero-plus-dl/extract -type d -name assets | head -1) \
-    && mv "${ASSETS_DIR}" ${LIBERO_PLUS_ROOT}/assets \
-    && rm -rf /tmp/libero-plus-dl
-
-# Point ~/.libero/config.yaml at the clone so LIBERO-plus's imports are
-# non-interactive (it calls input() when the config is missing).
-RUN mkdir -p /home/user_lerobot/.libero \
-    && printf "assets: ${LIBERO_PLUS_ROOT}/assets\nbddl_files: ${LIBERO_PLUS_ROOT}/bddl_files\ndatasets: ${LIBERO_PLUS_ROOT}/../datasets\ninit_states: ${LIBERO_PLUS_ROOT}/init_files\n" \
-       > /home/user_lerobot/.libero/config.yaml
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-CMD ["/bin/bash"]

docker/Dockerfile.benchmark.robocasa

@@ -1,71 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for RoboCasa365 integration tests.
-# Extends the nightly GPU image (which already has all extras installed)
-# with the PR's source code and RoboCasa-specific asset setup.
-#
-# Build:  docker build -f docker/Dockerfile.benchmark.robocasa -t lerobot-benchmark-robocasa .
-# Run:    docker run --gpus all --rm lerobot-benchmark-robocasa lerobot-eval ...
-
-FROM huggingface/lerobot-gpu:latest
-
-# Install robocasa + robosuite as editable clones. pip-installing from git
-# omits data files like robocasa/models/assets/box_links/box_links_assets.json
-# (not declared in package_data), which download_kitchen_assets needs at import.
-#
-# `--no-deps` on robocasa is deliberate: its setup.py pins `lerobot==0.3.3`
-# in install_requires, which would shadow the editable lerobot baked into
-# this image. We install robocasa's actual runtime deps explicitly instead.
-# Pinned SHAs for reproducible benchmark runs. Bump when you need an
-# upstream fix; don't rely on `main`/`master` drift.
-ARG ROBOCASA_SHA=56e355ccc64389dfc1b8a61a33b9127b975ba681
-ARG ROBOSUITE_SHA=aaa8b9b214ce8e77e82926d677b4d61d55e577ab
-RUN git clone https://github.com/robocasa/robocasa.git ~/robocasa && \
-    git -C ~/robocasa checkout ${ROBOCASA_SHA} && \
-    git clone https://github.com/ARISE-Initiative/robosuite.git ~/robosuite && \
-    git -C ~/robosuite checkout ${ROBOSUITE_SHA} && \
-    uv pip install --no-cache -e ~/robocasa --no-deps && \
-    uv pip install --no-cache -e ~/robosuite && \
-    uv pip install --no-cache \
-      "numpy==2.2.5" "numba==0.61.2" "scipy==1.15.3" "mujoco==3.3.1" \
-      "pygame==2.6.1" "Pillow==12.2.0" "opencv-python==4.13.0.92" \
-      "pyyaml==6.0.3" "pynput==1.8.1" "tqdm==4.67.3" "termcolor==3.3.0" \
-      "imageio==2.37.3" "h5py==3.16.0" "lxml==6.0.4" "hidapi==0.14.0.post4" \
-      "tianshou==0.4.10" "gymnasium==1.2.3"
-
-# Set up robocasa macros and download kitchen assets. We need:
-#   - tex              : base environment textures
-#   - tex_generative   : AI-generated textures; kitchen fixture XMLs embed
-#                        refs to generative_textures/wall/tex*.png
-#                        unconditionally, so MjModel.from_xml_string fails
-#                        at reset time without them (even if the env is
-#                        constructed with generative_textures=None).
-#   - fixtures_lw      : lightwheel kitchen fixtures (fridge, counters...)
-#   - objs_lw          : lightwheel object meshes (stools, misc props)
-# We skip the objaverse/aigen object packs (~30GB combined) by pairing
-# this with --env.obj_registries=["lightwheel"] on the lerobot side.
-# The download script prompts interactively, so pipe 'y' to auto-accept.
-RUN python -m robocasa.scripts.setup_macros && \
-    yes y | python -m robocasa.scripts.download_kitchen_assets \
-      --type tex tex_generative fixtures_lw objs_lw
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-# Re-install lerobot editably so the new source (with RoboCasaEnv registration)
-# replaces the stale package baked into the nightly image.
-RUN uv pip install --no-cache --no-deps -e .
-
-CMD ["/bin/bash"]

docker/Dockerfile.benchmark.robocerebra

@@ -1,43 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for RoboCerebra integration tests.
-# RoboCerebra reuses LIBERO's simulator (libero_10 suite) with a different
-# rename_map, so this image is identical to the LIBERO benchmark image —
-# extends the nightly GPU base with LIBERO assets + the PR's source code.
-#
-# Build:  docker build -f docker/Dockerfile.benchmark.robocerebra -t lerobot-benchmark-robocerebra .
-# Run:    docker run --gpus all --rm lerobot-benchmark-robocerebra lerobot-eval ...
-
-FROM huggingface/lerobot-gpu:latest
-
-# Pre-download lerobot/libero-assets from HF Hub so nothing is fetched at
-# runtime (which times out on CI). Point the libero config at the cached path.
-# libero/libero/__init__.py calls input() when ~/.libero/config.yaml is missing,
-# so we write the config before any libero import can happen.
-RUN LIBERO_DIR=$(python -c \
-      "import importlib.util, os; s=importlib.util.find_spec('libero'); \
-       print(os.path.join(os.path.dirname(s.origin), 'libero'))") && \
-    mkdir -p /home/user_lerobot/.libero && \
-    python -c "\
-from huggingface_hub import snapshot_download; \
-snapshot_download(repo_id='lerobot/libero-assets', repo_type='dataset', \
-                  local_dir='/home/user_lerobot/.libero/assets')" && \
-    printf "assets: /home/user_lerobot/.libero/assets\nbddl_files: ${LIBERO_DIR}/bddl_files\ndatasets: ${LIBERO_DIR}/../datasets\ninit_states: ${LIBERO_DIR}/init_files\n" \
-    > /home/user_lerobot/.libero/config.yaml
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-CMD ["/bin/bash"]

docker/Dockerfile.benchmark.robomme

@@ -1,56 +0,0 @@
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for RoboMME integration tests.
-# Extends the nightly GPU image (which has lerobot[all]) with Vulkan system
-# libs for ManiSkill/SAPIEN and the robomme extra. robomme isn't in [all]
-# because mani-skill hard-pins gymnasium==0.29.1 and numpy<2.0.0 which
-# conflict with lerobot's defaults; both are safe at runtime:
-#   - gymnasium 0.29.x has the same 5-tuple step() API as 1.x (since 0.26)
-#   - numpy 1.26.4 is API-compatible with lerobot's actual usage.
-#
-# Build:  docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-benchmark-robomme .
-# Run:    docker run --gpus all --rm lerobot-benchmark-robomme lerobot-eval ...
-
-FROM huggingface/lerobot-gpu:latest
-
-# NVIDIA Container Toolkit: expose Vulkan driver capability for headless rendering.
-ENV NVIDIA_DRIVER_CAPABILITIES=all \
-    VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json
-
-# ManiSkill/SAPIEN's renderer needs Vulkan, which isn't in the base image.
-USER root
-RUN apt-get update \
-    && apt-get install -y --no-install-recommends \
-         libvulkan1 libvulkan-dev mesa-vulkan-drivers \
-    && mkdir -p /usr/share/vulkan/icd.d \
-    && echo '{"file_format_version":"1.0.0","ICD":{"library_path":"libGLX_nvidia.so.0","api_version":"1.3.0"}}' \
-       > /usr/share/vulkan/icd.d/nvidia_icd.json \
-    && apt-get clean && rm -rf /var/lib/apt/lists/*
-USER user_lerobot
-
-# Install smolvla + av-dep via the PR's pyproject, then layer robomme on top
-# with gymnasium/numpy overrides. robomme isn't a pyproject extra because its
-# mani-skill pin conflicts with lerobot's base numpy>=2 (see pyproject.toml).
-COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
-RUN printf 'gymnasium==0.29.1\nnumpy==1.26.4\n' > /tmp/robomme_override.txt \
-    && uv pip install --no-cache --override /tmp/robomme_override.txt \
-         -e ".[smolvla,av-dep]" \
-         "robomme @ git+https://github.com/RoboMME/robomme_benchmark.git@main" \
-    && python -c "import robomme; print('robomme import OK')"
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-CMD ["/bin/bash"]

docker/Dockerfile.benchmark.robotwin

@@ -1,138 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for RoboTwin 2.0 integration tests.
-# Extends the nightly GPU image with the RoboTwin simulator stack:
-#   sapien/mplib/pytorch3d + NVlabs CuRobo + embodiments.zip + objects.zip
-# (~3.96 GB of assets; background_texture.zip ~11 GB skipped for smoke eval).
-#
-# Build: docker build -f docker/Dockerfile.benchmark.robotwin -t lerobot-benchmark-robotwin .
-# Run:   docker run --gpus all --rm lerobot-benchmark-robotwin \
-#            lerobot-eval --env.type=robotwin --env.task=beat_block_hammer ...
-
-FROM huggingface/lerobot-gpu:latest
-
-ENV NVIDIA_DRIVER_CAPABILITIES=all \
-    VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json \
-    ROBOTWIN_ROOT=/opt/robotwin
-
-# The nightly base is CUDA -base (no compiler, no Vulkan loader). CuRobo's
-# `pip install -e .` runs nvcc, and SAPIEN renders via Vulkan — add both.
-USER root
-# Pinned upstream SHA for reproducible benchmark runs. Bump when we need
-# an upstream fix; don't rely on `main` drift.
-ARG ROBOTWIN_SHA=0aeea2d669c0f8516f4d5785f0aa33ba812c14b4
-RUN apt-get update \
-    && apt-get install -y --no-install-recommends \
-         cuda-nvcc-12-4 cuda-cudart-dev-12-4 \
-         libvulkan1 vulkan-tools \
-    && mkdir -p /usr/share/vulkan/icd.d \
-    && echo '{"file_format_version":"1.0.0","ICD":{"library_path":"libGLX_nvidia.so.0","api_version":"1.3.0"}}' \
-       > /usr/share/vulkan/icd.d/nvidia_icd.json \
-    && git clone https://github.com/RoboTwin-Platform/RoboTwin.git ${ROBOTWIN_ROOT} \
-    && git -C ${ROBOTWIN_ROOT} checkout ${ROBOTWIN_SHA} \
-    && chown -R user_lerobot:user_lerobot ${ROBOTWIN_ROOT} \
-    && apt-get clean && rm -rf /var/lib/apt/lists/*
-USER user_lerobot
-
-# RoboTwin runtime deps (av is already in the base via [av-dep]).
-RUN uv pip install --no-cache \
-        "sapien==3.0.0b1" "mplib==0.2.1" "transforms3d==0.4.2" "trimesh==4.4.3" \
-        "open3d==0.19.0" "imageio==2.34.2" termcolor zarr pydantic h5py
-
-# pytorch3d has no universal wheel; must be built from source (~10 min, cached).
-RUN uv pip install --no-cache --no-build-isolation \
-        "git+https://github.com/facebookresearch/pytorch3d.git@stable"
-
-# CuRobo — NVlabs motion generator; TORCH_CUDA_ARCH_LIST must be set or the
-# build aborts on an empty arch list. RoboTwin's own installer pins v0.7.8,
-# which still exposes the v1 API (`curobo.types.math`) that RoboTwin imports.
-ARG CUROBO_REF=v0.7.8
-RUN cd ${ROBOTWIN_ROOT}/envs \
-    && git clone --branch ${CUROBO_REF} --depth 1 https://github.com/NVlabs/curobo.git \
-    && cd curobo \
-    && TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;8.9;9.0" \
-       uv pip install -e . --no-build-isolation --no-cache
-
-# Upstream patches (mirror RoboTwin's script/_install.sh).
-# These patches target the exact versions pinned above; re-check when upgrading.
-# mplib==0.2.1: drop a broken `or collide` clause in planner.py.
-#   Safe to remove once mplib > 0.2.1 ships with the fix upstream.
-# sapien==3.0.0b1: fix URDF loader encoding + .srdf extension check.
-#   Safe to remove once sapien > 3.0.0b1 ships with the fix upstream.
-RUN python - <<'EOF'
-import pathlib, re, site
-for d in site.getsitepackages():
-    p = pathlib.Path(d) / "mplib" / "planner.py"
-    if p.exists():
-        p.write_text(re.sub(r"\bor collide\b", "", p.read_text(), count=1))
-        print(f"mplib patch applied: {p}")
-    p = pathlib.Path(d) / "sapien" / "wrapper" / "urdf_loader.py"
-    if p.exists():
-        src = p.read_text().replace(
-            "with open(srdf_path) as f:", 'with open(srdf_path, encoding="utf-8") as f:'
-        ).replace('"srdf"', '".srdf"')
-        p.write_text(src)
-        print(f"sapien patch applied: {p}")
-EOF
-
-# Simulation assets from TianxingChen/RoboTwin2.0: embodiments (~220 MB) +
-# objects (~3.74 GB). background_texture (~11 GB) is intentionally skipped.
-# The dataset is public — no auth token needed.
-RUN python - <<'EOF'
-import os, pathlib, zipfile
-from huggingface_hub import hf_hub_download
-
-assets_dir = pathlib.Path(os.environ["ROBOTWIN_ROOT"]) / "assets"
-assets_dir.mkdir(parents=True, exist_ok=True)
-for fname in ("embodiments.zip", "objects.zip"):
-    local = hf_hub_download(
-        repo_id="TianxingChen/RoboTwin2.0",
-        repo_type="dataset",
-        filename=fname,
-        local_dir=str(assets_dir),
-    )
-    with zipfile.ZipFile(local, "r") as z:
-        z.extractall(str(assets_dir))
-    pathlib.Path(local).unlink()
-EOF
-
-WORKDIR ${ROBOTWIN_ROOT}
-RUN python script/update_embodiment_config_path.py
-
-ENV PYTHONPATH="${ROBOTWIN_ROOT}"
-
-# Fail the image build early if the CuRobo package layout regresses. Importing
-# RoboTwin's planner here is too eager because CuRobo constructs CUDA-backed
-# defaults at import time, while Docker builds don't have access to an NVIDIA
-# driver.
-RUN python - <<'EOF'
-from pathlib import Path
-
-from curobo.types.math import Pose
-
-planner_src = (Path("/opt/robotwin/envs/robot/planner.py")).read_text()
-assert "from curobo.types.math import Pose as CuroboPose" in planner_src
-
-print("CuRobo import OK:", Pose.__name__)
-print("RoboTwin planner import references curobo.types.math")
-EOF
-
-# Return to the lerobot source directory (set by base image) before overlaying.
-WORKDIR /lerobot
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-CMD ["/bin/bash"]

docker/Dockerfile.benchmark.vlabench

@@ -1,99 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Benchmark image for VLABench integration tests.
-# Extends the nightly GPU image with the PR's source code and VLABench setup.
-#
-# Build:  docker build -f docker/Dockerfile.benchmark.vlabench -t lerobot-benchmark-vlabench .
-# Run:    docker run --gpus all --rm lerobot-benchmark-vlabench lerobot-eval ...
-
-FROM huggingface/lerobot-gpu:latest
-
-# Install VLABench from GitHub (not on PyPI) and pin MuJoCo/dm-control.
-# Shallow-clone without submodule recursion (nested SSH-only submodules fail in CI).
-# Editable install (-e) because VLABench/utils/ has no __init__.py, so
-# find_packages() omits it from wheels; editable mode uses the source tree directly.
-# rrt-algorithms has the same packaging issue (rrt/ dir missing __init__.py).
-# Patch: constant.py calls os.listdir on ~100 asset/obj/meshes/* dirs at import
-# time. Guard the call so missing dirs return [] instead of crashing (in case
-# the asset download is partial).
-#
-# Pinned upstream SHAs for reproducible benchmark runs. Bump when you need
-# an upstream fix; don't rely on `main`/`develop` drift.
-ARG VLABENCH_SHA=cf588fe60c0c7282174fe979f5913170cfe69017
-ARG RRT_ALGORITHMS_SHA=e51d95ee489a225220d6ae2a764c4111f6ba7d85
-RUN git clone https://github.com/OpenMOSS/VLABench.git ~/VLABench && \
-    git -C ~/VLABench checkout ${VLABENCH_SHA} && \
-    git clone https://github.com/motion-planning/rrt-algorithms.git ~/rrt-algorithms && \
-    git -C ~/rrt-algorithms checkout ${RRT_ALGORITHMS_SHA} && \
-    python3 -c "\
-import pathlib; \
-p = pathlib.Path.home() / 'VLABench/VLABench/configs/constant.py'; \
-t = p.read_text(); \
-p.write_text(t.replace( \
-    'subdirs = os.listdir(xml_dir)', \
-    'if not os.path.isdir(xml_dir): return []\n    subdirs = os.listdir(xml_dir)'))" && \
-    uv pip install --no-cache -e ~/VLABench -e ~/rrt-algorithms \
-      mujoco==3.2.2 dm-control==1.0.22 \
-      open3d colorlog scikit-learn openai gdown
-
-# Download VLABench mesh assets. Task configs reference object meshes
-# (obj/meshes/fruit/, containers/basket/, tablewares/plates/, etc.); without
-# them the task builder picks from an empty mesh list and crashes with
-# IndexError at task-build time (random.choice([]) in config_manager.py).
-#
-# Preferred source: an HF Hub mirror. Set VLABENCH_ASSETS_REPO at build time
-# (e.g. --build-arg VLABENCH_ASSETS_REPO=lerobot/vlabench-assets) and we'll
-# snapshot_download the repo into VLABench's assets dir. This is the reliable
-# path for CI — Google Drive frequently returns HTTP 429 ("Too many users have
-# viewed or downloaded this file recently") on shared academic files.
-#
-# After download we *validate* that at least one XML exists under each
-# task-critical subtree and fail the build loudly if not. Silent-empty asset
-# dirs are the #1 cause of VLABench runtime crashes in CI, so we surface them
-# here rather than after a 10-minute eval build.
-#
-# Fallback: VLABench's own gdown-based script. Best-effort only.
-ARG VLABENCH_ASSETS_REPO=""
-RUN ASSETS_DIR="$HOME/VLABench/VLABench/assets" && \
-    if [ -n "${VLABENCH_ASSETS_REPO}" ]; then \
-        echo "Downloading VLABench assets from HF Hub: ${VLABENCH_ASSETS_REPO}" && \
-        uv pip install --no-cache "huggingface_hub[hf_xet]>=0.26" && \
-        python -c "from huggingface_hub import snapshot_download; \
-p = snapshot_download(repo_id='${VLABENCH_ASSETS_REPO}', repo_type='dataset', \
-    local_dir='${ASSETS_DIR}', allow_patterns=['obj/**', 'scenes/**']); \
-print('snapshot_download returned:', p)"; \
-    else \
-        echo "No VLABENCH_ASSETS_REPO set — falling back to gdown" && \
-        python ~/VLABench/scripts/download_assets.py --choice all; \
-    fi && \
-    python -c "\
-from pathlib import Path; \
-import sys; \
-root = Path('${ASSETS_DIR}'); \
-checks = ['obj/meshes/tablewares/plates', 'obj/meshes/containers/basket', 'obj/meshes/fruit', 'obj/meshes/containers/tray']; \
-failed = []; \
-print(f'Validating VLABench assets under {root}'); \
-[print(f'  {c}: {len(list((root/c).rglob(\"*.xml\")))} XMLs') for c in checks]; \
-[failed.append(c) for c in checks if not any((root/c).rglob('*.xml'))]; \
-sys.exit(f'Empty asset dirs (no *.xml): {failed}') if failed else print('All asset dirs populated.')"
-
-# Overlay the PR's source code on top of the nightly image.
-COPY --chown=user_lerobot:user_lerobot . .
-
-# Re-install lerobot editably so the new source (with VLABenchEnv registration
-# and updated obs handling) replaces the stale package baked into the nightly image.
-RUN uv pip install --no-cache --no-deps -e .
-
-CMD ["/bin/bash"]

docs/source/_toctree.yml

@@ -77,22 +77,10 @@
     title: Adding a New Benchmark
   - local: libero
     title: LIBERO
-  - local: libero_plus
-    title: LIBERO-plus
   - local: metaworld
     title: Meta-World
-  - local: robotwin
-    title: RoboTwin 2.0
-  - local: robocasa
-    title: RoboCasa365
-  - local: robocerebra
-    title: RoboCerebra
-  - local: robomme
-    title: RoboMME
   - local: envhub_isaaclab_arena
     title: NVIDIA IsaacLab Arena Environments
-  - local: vlabench
-    title: VLABench
   title: "Benchmarks"
 - sections:
   - local: introduction_processors

docs/source/hilserl.mdx

@@ -820,10 +820,10 @@ The LeRobot system uses a distributed actor-learner architecture for training. T
 
 Create a training configuration file (example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/train_config.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/configs/train.py`.
 
-1. Configure the policy settings (`type="sac"`, `device`, etc.)
+1. Configure the policy settings (`type="gaussian_actor"`, `device`, etc.)
 2. Set `dataset` to your cropped dataset
 3. Configure environment settings with crop parameters
-4. Check the other parameters related to SAC in [configuration_sac.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/sac/configuration_sac.py#L79).
+4. Check the other parameters related to the Gaussian Actor in [configuration_gaussian_actor.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/gaussian_actor/configuration_gaussian_actor.py#L79).
 5. Verify that the `policy` config is correct with the right `input_features` and `output_features` for your task.
 
 **Starting the Learner**
@@ -926,7 +926,7 @@ The ideal behaviour is that your intervention rate should drop gradually during
 
 Some configuration values have a disproportionate impact on training stability and speed:
 
-- **`temperature_init`** (`policy.temperature_init`) – initial entropy temperature in SAC. Higher values encourage more exploration; lower values make the policy more deterministic early on. A good starting point is `1e-2`. We observed that setting it too high can make human interventions ineffective and slow down learning.
+- **`temperature_init`** (`algorithm.temperature_init`) – initial entropy temperature in SAC. Higher values encourage more exploration; lower values make the policy more deterministic early on. A good starting point is `1e-2`. We observed that setting it too high can make human interventions ineffective and slow down learning.
 - **`policy_parameters_push_frequency`** (`policy.actor_learner_config.policy_parameters_push_frequency`) – interval in _seconds_ between two weight pushes from the learner to the actor. The default is `4 s`. Decrease to **1-2 s** to provide fresher weights (at the cost of more network traffic); increase only if your connection is slow, as this will reduce sample efficiency.
 - **`storage_device`** (`policy.storage_device`) – device on which the learner keeps the policy parameters. If you have spare GPU memory, set this to `"cuda"` (instead of the default `"cpu"`). Keeping the weights on-GPU removes CPU→GPU transfer overhead and can significantly increase the number of learner updates per second.
 

docs/source/il_robots.mdx

@@ -32,12 +32,6 @@ Once you’ve gathered enough trajectories, you’ll train a neural network to i
 
 If you run into any issues at any point, jump into our [Discord community](https://discord.com/invite/s3KuuzsPFb) for support.
 
-<Tip>
-
-Want to quickly get the right commands for your setup? The [quickstart notebook](https://github.com/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb) [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb) lets you configure your robot once and generates all the commands below ready to paste.
-
-</Tip>
-
 ## Set up and Calibrate
 
 If you haven't yet set up and calibrated your robot and teleop device, please do so by following the robot-specific tutorial.

docs/source/libero_plus.mdx

@@ -1,188 +0,0 @@
-# LIBERO-plus
-
-LIBERO-plus is a **robustness benchmark** for Vision-Language-Action (VLA) models built on top of [LIBERO](./libero). It systematically stress-tests policies by applying **seven independent perturbation dimensions** to the original LIBERO task set, exposing failure modes that standard benchmarks miss.
-
-- Paper: [In-depth Robustness Analysis of Vision-Language-Action Models](https://arxiv.org/abs/2510.13626)
-- GitHub: [sylvestf/LIBERO-plus](https://github.com/sylvestf/LIBERO-plus)
-- Dataset: [lerobot/libero_plus](https://huggingface.co/datasets/lerobot/libero_plus)
-
-![An overview of the LIBERO-plus benchmark perturbation dimensions](https://github.com/sylvestf/LIBERO-plus/raw/main/static/images/libero-plus.jpg)
-
-## Perturbation dimensions
-
-LIBERO-plus creates ~10 000 task variants by perturbing each original LIBERO task along these axes:
-
-| Dimension             | What changes                                          |
-| --------------------- | ----------------------------------------------------- |
-| Objects layout        | Target position, presence of confounding objects      |
-| Camera viewpoints     | Camera position, orientation, field-of-view           |
-| Robot initial states  | Manipulator start pose                                |
-| Language instructions | LLM-rewritten task description (paraphrase / synonym) |
-| Light conditions      | Intensity, direction, color, shadow                   |
-| Background textures   | Scene surface and object appearance                   |
-| Sensor noise          | Photometric distortions and image degradation         |
-
-## Available task suites
-
-LIBERO-plus covers the same five suites as LIBERO:
-
-| Suite          | CLI name         | Tasks | Max steps | Description                                        |
-| -------------- | ---------------- | ----- | --------- | -------------------------------------------------- |
-| LIBERO-Spatial | `libero_spatial` | 10    | 280       | Tasks requiring reasoning about spatial relations  |
-| LIBERO-Object  | `libero_object`  | 10    | 280       | Tasks centered on manipulating different objects   |
-| LIBERO-Goal    | `libero_goal`    | 10    | 300       | Goal-conditioned tasks with changing targets       |
-| LIBERO-90      | `libero_90`      | 90    | 400       | Short-horizon tasks from the LIBERO-100 collection |
-| LIBERO-Long    | `libero_10`      | 10    | 520       | Long-horizon tasks from the LIBERO-100 collection  |
-
-<Tip warning={true}>
-  Installing LIBERO-plus **replaces** vanilla LIBERO — it uninstalls `hf-libero`
-  so that `import libero` resolves to the LIBERO-plus fork. You cannot have both
-  installed at the same time. To switch back to vanilla LIBERO, uninstall the
-  fork and reinstall with `pip install -e ".[libero]"`.
-</Tip>
-
-## Installation
-
-### System dependencies (Linux only)
-
-```bash
-sudo apt install libexpat1 libfontconfig1-dev libmagickwand-dev
-```
-
-### Python package
-
-```bash
-pip install -e ".[libero]" "robosuite==1.4.1" bddl easydict mujoco wand scikit-image gym
-git clone https://github.com/sylvestf/LIBERO-plus.git
-cd LIBERO-plus && pip install --no-deps -e .
-pip uninstall -y hf-libero  # so `import libero` resolves to the fork
-```
-
-LIBERO-plus is installed from its GitHub fork rather than a pyproject extra — the fork ships as a namespace package that pip can't handle, so it must be cloned and added to `PYTHONPATH`. See `docker/Dockerfile.benchmark.libero_plus` for the canonical install. MuJoCo is required, so only Linux is supported.
-
-<Tip>
-Set the MuJoCo rendering backend before running evaluation:
-
-```bash
-export MUJOCO_GL=egl   # headless / HPC / cloud
-```
-
-</Tip>
-
-### Download LIBERO-plus assets
-
-LIBERO-plus ships its extended asset pack separately. Download `assets.zip` from the [Hugging Face dataset](https://huggingface.co/datasets/Sylvest/LIBERO-plus/tree/main) and extract it into the LIBERO-plus package directory:
-
-```bash
-# After installing the package, find where it was installed:
-python -c "import libero; print(libero.__file__)"
-# Then extract assets.zip into <package_root>/libero/assets/
-```
-
-## Evaluation
-
-### Default evaluation (recommended)
-
-Evaluate across the four standard suites (10 episodes per task):
-
-```bash
-lerobot-eval \
-  --policy.path="your-policy-id" \
-  --env.type=libero_plus \
-  --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --env.max_parallel_tasks=1
-```
-
-### Single-suite evaluation
-
-Evaluate on one LIBERO-plus suite:
-
-```bash
-lerobot-eval \
-  --policy.path="your-policy-id" \
-  --env.type=libero_plus \
-  --env.task=libero_spatial \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10
-```
-
-- `--env.task` picks the suite (`libero_spatial`, `libero_object`, etc.).
-- `--env.task_ids` restricts to specific task indices (`[0]`, `[1,2,3]`, etc.). Omit to run all tasks in the suite.
-- `--eval.batch_size` controls how many environments run in parallel.
-- `--eval.n_episodes` sets how many episodes to run per task.
-
-### Multi-suite evaluation
-
-Benchmark a policy across multiple suites at once by passing a comma-separated list:
-
-```bash
-lerobot-eval \
-  --policy.path="your-policy-id" \
-  --env.type=libero_plus \
-  --env.task=libero_spatial,libero_object \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10
-```
-
-### Control mode
-
-LIBERO-plus supports two control modes — `relative` (default) and `absolute`. Different VLA checkpoints are trained with different action parameterizations, so make sure the mode matches your policy:
-
-```bash
---env.control_mode=relative   # or "absolute"
-```
-
-### Policy inputs and outputs
-
-**Observations:**
-
-- `observation.state` — 8-dim proprioceptive features (eef position, axis-angle orientation, gripper qpos)
-- `observation.images.image` — main camera view (`agentview_image`), HWC uint8
-- `observation.images.image2` — wrist camera view (`robot0_eye_in_hand_image`), HWC uint8
-
-**Actions:**
-
-- Continuous control in `Box(-1, 1, shape=(7,))` — 6D end-effector delta + 1D gripper
-
-### 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.
-
-## Training
-
-### Dataset
-
-A LeRobot-format training dataset for LIBERO-plus is available at:
-
-- [lerobot/libero_plus](https://huggingface.co/datasets/lerobot/libero_plus)
-
-### Example training command
-
-```bash
-lerobot-train \
-    --policy.type=smolvla \
-    --policy.repo_id=${HF_USER}/smolvla_libero_plus \
-    --policy.load_vlm_weights=true \
-    --dataset.repo_id=lerobot/libero_plus \
-    --env.type=libero_plus \
-    --env.task=libero_spatial \
-    --output_dir=./outputs/ \
-    --steps=100000 \
-    --batch_size=4 \
-    --eval.batch_size=1 \
-    --eval.n_episodes=1 \
-    --eval_freq=1000
-```
-
-## Relationship to LIBERO
-
-LIBERO-plus is a drop-in extension of LIBERO:
-
-- Same Python gym interface (`LiberoEnv`, `LiberoProcessorStep`)
-- Same camera names and observation/action format
-- Same task suite names
-- Installs under the same `libero` Python package name (different GitHub repo)
-
-To use the original LIBERO benchmark, see [LIBERO](./libero) and use `--env.type=libero`.

docs/source/robocasa.mdx

@@ -1,188 +0,0 @@
-# RoboCasa365
-
-[RoboCasa365](https://robocasa.ai) is a large-scale simulation framework for training and benchmarking **generalist robots** in everyday kitchen tasks. It ships 365 diverse manipulation tasks across 2,500 kitchen environments, 3,200+ object assets and 600+ hours of human demonstration data, on a PandaOmron 12-DOF mobile manipulator (Franka arm on a holonomic base).
-
-- Paper: [RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots](https://arxiv.org/abs/2406.02523)
-- GitHub: [robocasa/robocasa](https://github.com/robocasa/robocasa)
-- Project website: [robocasa.ai](https://robocasa.ai)
-- Pretrained policy: [`lerobot/smolvla_robocasa`](https://huggingface.co/lerobot/smolvla_robocasa)
-- Single-task dataset (CloseFridge): [`pepijn223/robocasa_CloseFridge`](https://huggingface.co/datasets/pepijn223/robocasa_CloseFridge)
-
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/robocasa-banner.webp"
-  alt="RoboCasa365 benchmark overview"
-  width="85%"
-/>
-
-## Available tasks
-
-RoboCasa365 organizes its 365 tasks into two families and three upstream benchmark groups that LeRobot exposes as first-class `--env.task` shortcuts:
-
-| Family    | Tasks | Description                                                                     |
-| --------- | ----- | ------------------------------------------------------------------------------- |
-| Atomic    | ~65   | Single-skill tasks: pick-and-place, door/drawer manipulation, appliance control |
-| Composite | ~300  | Multi-step tasks across 60+ categories: cooking, cleaning, organizing, etc.     |
-
-**Atomic task examples:** `CloseFridge`, `OpenDrawer`, `OpenCabinet`, `TurnOnMicrowave`, `TurnOffStove`, `NavigateKitchen`, `PickPlaceCounterToStove`.
-
-**Composite task categories:** baking, boiling, brewing, chopping, clearing table, defrosting food, loading dishwasher, making tea, microwaving food, washing dishes, and more.
-
-`--env.task` accepts three forms:
-
-- a single task name (`CloseFridge`)
-- a comma-separated list (`CloseFridge,OpenBlenderLid,PickPlaceCoffee`)
-- a benchmark-group shortcut — `atomic_seen`, `composite_seen`, `composite_unseen`, `pretrain50`, `pretrain100`, `pretrain200`, `pretrain300` — which auto-expands to the upstream task list and auto-sets the dataset `split` (`target` or `pretrain`).
-
-## Installation
-
-RoboCasa and its dependency `robosuite` are not published on PyPI, and RoboCasa's own `setup.py` hardcodes `lerobot==0.3.3`, which conflicts with this repo's `lerobot`. LeRobot therefore does **not** expose a `robocasa` extra — install the two packages manually as editable clones (using `--no-deps` on `robocasa` to skip its shadowed `lerobot` pin):
-
-```bash
-# After following the standard LeRobot installation instructions.
-
-git clone https://github.com/robocasa/robocasa.git ~/robocasa
-git clone https://github.com/ARISE-Initiative/robosuite.git ~/robosuite
-pip install -e ~/robocasa --no-deps
-pip install -e ~/robosuite
-
-# Robocasa's runtime deps (the ones its setup.py would have pulled, minus
-# the bad lerobot pin).
-pip install numpy numba scipy mujoco pygame Pillow opencv-python \
-            pyyaml pynput tqdm termcolor imageio h5py lxml hidapi \
-            tianshou gymnasium
-
-python -m robocasa.scripts.setup_macros
-# Lightweight assets (lightwheel object meshes + textures). Enough for
-# the default env out of the box.
-python -m robocasa.scripts.download_kitchen_assets \
-  --type tex tex_generative fixtures_lw objs_lw
-# Optional: full objaverse/aigen registries (~30GB) for richer object
-# variety. Enable at eval time via --env.obj_registries (see below).
-# python -m robocasa.scripts.download_kitchen_assets --type objs_objaverse
-```
-
-<Tip>
-RoboCasa requires MuJoCo. Set the rendering backend before training or evaluation:
-
-```bash
-export MUJOCO_GL=egl  # for headless servers (HPC, cloud)
-```
-
-</Tip>
-
-### Object registries
-
-By default the env samples objects only from the `lightwheel` registry (what `--type objs_lw` ships), which avoids a `Probabilities contain NaN` crash when the objaverse / aigen packs aren't on disk. If you've downloaded the full asset set, enable the full registry at runtime:
-
-```bash
---env.obj_registries='[objaverse,lightwheel]'
-```
-
-## Evaluation
-
-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.
-
-### Single-task evaluation (recommended for quick iteration)
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_robocasa \
-  --env.type=robocasa \
-  --env.task=CloseFridge \
-  --eval.batch_size=1 \
-  --eval.n_episodes=20 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
-```
-
-### Multi-task evaluation
-
-Pass a comma-separated list of tasks:
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_robocasa \
-  --env.type=robocasa \
-  --env.task=CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove \
-  --eval.batch_size=1 \
-  --eval.n_episodes=20 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
-```
-
-### Benchmark-group evaluation
-
-Run an entire upstream group (e.g. all 18 `atomic_seen` tasks with `split=target`):
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_robocasa \
-  --env.type=robocasa \
-  --env.task=atomic_seen \
-  --eval.batch_size=1 \
-  --eval.n_episodes=20 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.robot0_agentview_left": "observation.images.camera1", "observation.images.robot0_eye_in_hand": "observation.images.camera2", "observation.images.robot0_agentview_right": "observation.images.camera3"}'
-```
-
-### Recommended evaluation episodes
-
-**20 episodes per task** for reproducible benchmarking. Matches the protocol used in published results.
-
-## Policy inputs and outputs
-
-**Observations** (raw RoboCasa camera names are preserved verbatim):
-
-- `observation.state` — 16-dim proprioceptive state (base position, base quaternion, relative end-effector position, relative end-effector quaternion, gripper qpos)
-- `observation.images.robot0_agentview_left` — left agent view, 256×256 HWC uint8
-- `observation.images.robot0_eye_in_hand` — wrist camera view, 256×256 HWC uint8
-- `observation.images.robot0_agentview_right` — right agent view, 256×256 HWC uint8
-
-**Actions:**
-
-- Continuous control in `Box(-1, 1, shape=(12,))` — base motion (4D) + control mode (1D) + end-effector position (3D) + end-effector rotation (3D) + gripper (1D).
-
-## Training
-
-### Single-task example
-
-A ready-to-use single-task dataset is on the Hub:
-[`pepijn223/robocasa_CloseFridge`](https://huggingface.co/datasets/pepijn223/robocasa_CloseFridge).
-
-Fine-tune a SmolVLA base on `CloseFridge`:
-
-```bash
-lerobot-train \
-  --policy.type=smolvla \
-  --policy.repo_id=${HF_USER}/smolvla_robocasa_CloseFridge \
-  --policy.load_vlm_weights=true \
-  --policy.push_to_hub=true \
-  --dataset.repo_id=pepijn223/robocasa_CloseFridge \
-  --env.type=robocasa \
-  --env.task=CloseFridge \
-  --output_dir=./outputs/smolvla_robocasa_CloseFridge \
-  --steps=100000 \
-  --batch_size=4 \
-  --eval_freq=5000 \
-  --eval.batch_size=1 \
-  --eval.n_episodes=5 \
-  --save_freq=10000
-```
-
-Evaluate the resulting checkpoint:
-
-```bash
-lerobot-eval \
-  --policy.path=${HF_USER}/smolvla_robocasa_CloseFridge \
-  --env.type=robocasa \
-  --env.task=CloseFridge \
-  --eval.batch_size=1 \
-  --eval.n_episodes=20
-```
-
-## Reproducing published results
-
-The released checkpoint [`lerobot/smolvla_robocasa`](https://huggingface.co/lerobot/smolvla_robocasa) is evaluated with the commands in the [Evaluation](#evaluation) section. CI runs a 10-atomic-task smoke eval (one episode each) on every PR touching the benchmark, picking fixture-centric tasks that don't require the objaverse asset pack.

docs/source/robocerebra.mdx

@@ -1,99 +0,0 @@
-# RoboCerebra
-
-[RoboCerebra](https://robocerebra-project.github.io/) is a long-horizon manipulation benchmark that evaluates **high-level reasoning, planning, and memory** in VLAs. Episodes chain multiple sub-goals with language-grounded intermediate instructions, built on top of LIBERO's simulator stack (MuJoCo + robosuite, Franka Panda 7-DOF).
-
-- Paper: [RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation](https://arxiv.org/abs/2506.06677)
-- Project website: [robocerebra-project.github.io](https://robocerebra-project.github.io/)
-- Dataset: [`lerobot/robocerebra_unified`](https://huggingface.co/datasets/lerobot/robocerebra_unified) — LeRobot v3.0, 6,660 episodes / 571,116 frames at 20 fps, 1,728 language-grounded sub-tasks.
-- Pretrained policy: [`lerobot/smolvla_robocerebra`](https://huggingface.co/lerobot/smolvla_robocerebra)
-
-## Available tasks
-
-RoboCerebra reuses LIBERO's simulator, so evaluation runs against the LIBERO `libero_10` long-horizon suite:
-
-| Suite     | CLI name    | Tasks | Description                                                   |
-| --------- | ----------- | ----- | ------------------------------------------------------------- |
-| LIBERO-10 | `libero_10` | 10    | Long-horizon kitchen/living room tasks chaining 3–6 sub-goals |
-
-Each RoboCerebra episode in the dataset is segmented into multiple sub-tasks with natural-language instructions, which the unified dataset exposes as independent supervision signals.
-
-## Installation
-
-RoboCerebra piggybacks on LIBERO, so the `libero` extra is all you need:
-
-```bash
-pip install -e ".[libero]"
-```
-
-<Tip>
-RoboCerebra requires Linux (MuJoCo / robosuite). Set the rendering backend before training or evaluation:
-
-```bash
-export MUJOCO_GL=egl  # for headless servers (HPC, cloud)
-```
-
-</Tip>
-
-## Evaluation
-
-RoboCerebra eval runs against LIBERO's `libero_10` suite with RoboCerebra's camera naming (`image` + `wrist_image`) and an extra empty-camera slot so a three-view-trained policy receives the expected input layout:
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_robocerebra \
-  --env.type=libero \
-  --env.task=libero_10 \
-  --env.fps=20 \
-  --env.obs_type=pixels_agent_pos \
-  --env.observation_height=256 \
-  --env.observation_width=256 \
-  '--env.camera_name_mapping={"agentview_image": "image", "robot0_eye_in_hand_image": "wrist_image"}' \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.wrist_image": "observation.images.camera2"}' \
-  --policy.empty_cameras=1
-```
-
-### Recommended evaluation episodes
-
-**10 episodes per task** across the `libero_10` suite (100 total) for reproducible benchmarking. Matches the protocol used in the RoboCerebra paper.
-
-## Policy inputs and outputs
-
-**Observations:**
-
-- `observation.state` — 8-dim proprioceptive state (7 joint positions + gripper)
-- `observation.images.image` — third-person view, 256×256 HWC uint8
-- `observation.images.wrist_image` — wrist-mounted camera view, 256×256 HWC uint8
-
-**Actions:**
-
-- Continuous control in `Box(-1, 1, shape=(7,))` — end-effector delta (6D) + gripper (1D)
-
-## Training
-
-The unified dataset at [`lerobot/robocerebra_unified`](https://huggingface.co/datasets/lerobot/robocerebra_unified) exposes two RGB streams and language-grounded sub-task annotations:
-
-| Feature                          | Shape         | Description          |
-| -------------------------------- | ------------- | -------------------- |
-| `observation.images.image`       | (256, 256, 3) | Third-person view    |
-| `observation.images.wrist_image` | (256, 256, 3) | Wrist-mounted camera |
-| `observation.state`              | (8,)          | Joint pos + gripper  |
-| `action`                         | (7,)          | EEF delta + gripper  |
-
-Fine-tune a SmolVLA base on it:
-
-```bash
-lerobot-train \
-  --policy.path=lerobot/smolvla_base \
-  --dataset.repo_id=lerobot/robocerebra_unified \
-  --env.type=libero \
-  --env.task=libero_10 \
-  --output_dir=outputs/smolvla_robocerebra
-```
-
-## Reproducing published results
-
-The released checkpoint [`lerobot/smolvla_robocerebra`](https://huggingface.co/lerobot/smolvla_robocerebra) was trained on `lerobot/robocerebra_unified` and evaluated with the command in the [Evaluation](#evaluation) section. CI runs the same command with `--eval.n_episodes=1` as a smoke test on every PR touching the benchmark.

docs/source/robomme.mdx

@@ -1,130 +0,0 @@
-# RoboMME
-
-[RoboMME](https://robomme.github.io) is a memory-augmented manipulation benchmark built on ManiSkill (SAPIEN). It evaluates a robot's ability to retain and use information across an episode — counting, object permanence, reference, and imitation.
-
-- **16 tasks** across 4 memory-skill suites
-- **1,600 training demos** (100 per task, 50 val, 50 test)
-- **Dataset**: [`lerobot/robomme`](https://huggingface.co/datasets/lerobot/robomme) — LeRobot v3.0, 768K frames at 10 fps
-- **Simulator**: ManiSkill / SAPIEN, Panda arm, Linux only
-
-![RoboMME benchmark tasks overview](https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04639/gradient.png)
-
-## Tasks
-
-| Suite                             | Tasks                                                         |
-| --------------------------------- | ------------------------------------------------------------- |
-| **Counting** (temporal memory)    | BinFill, PickXtimes, SwingXtimes, StopCube                    |
-| **Permanence** (spatial memory)   | VideoUnmask, VideoUnmaskSwap, ButtonUnmask, ButtonUnmaskSwap  |
-| **Reference** (object memory)     | PickHighlight, VideoRepick, VideoPlaceButton, VideoPlaceOrder |
-| **Imitation** (procedural memory) | MoveCube, InsertPeg, PatternLock, RouteStick                  |
-
-## Installation
-
-> RoboMME requires **Linux** (ManiSkill/SAPIEN uses Vulkan rendering). Docker is recommended to isolate dependency conflicts.
-
-### Native (Linux)
-
-```bash
-pip install --override <(printf 'gymnasium==0.29.1\nnumpy==1.26.4\n') \
-  -e '.[smolvla,av-dep]' \
-  'robomme @ git+https://github.com/RoboMME/robomme_benchmark.git@main'
-```
-
-> **Dependency note**: `mani-skill` (pulled by `robomme`) pins `gymnasium==0.29.1` and `numpy<2.0.0`, which conflict with lerobot's base `numpy>=2.0.0`. That's why `robomme` is not a pyproject extra — use the override install above, or the Docker approach below to avoid conflicts entirely.
-
-### Docker (recommended)
-
-```bash
-# Build base image first (from repo root)
-docker build -f docker/Dockerfile.eval-base -t lerobot-eval-base .
-
-# Build RoboMME eval image (applies gymnasium + numpy pin overrides)
-docker build -f docker/Dockerfile.benchmark.robomme -t lerobot-robomme .
-```
-
-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
-
-### Default (single task, single episode)
-
-```bash
-lerobot-eval \
-    --policy.path=<your_policy_repo> \
-    --env.type=robomme \
-    --env.task=PickXtimes \
-    --env.dataset_split=test \
-    --env.task_ids=[0] \
-    --eval.batch_size=1 \
-    --eval.n_episodes=1
-```
-
-### Multi-task evaluation
-
-Evaluate multiple tasks in one run by comma-separating task names. Use `task_ids` to control which episodes are evaluated per task. Recommended: 50 episodes per task for the test split.
-
-```bash
-lerobot-eval \
-    --policy.path=<your_policy_repo> \
-    --env.type=robomme \
-    --env.task=PickXtimes,BinFill,StopCube,MoveCube,InsertPeg \
-    --env.dataset_split=test \
-    --env.task_ids=[0,1,2,3,4,5,6,7,8,9] \
-    --eval.batch_size=1 \
-    --eval.n_episodes=50
-```
-
-### Key CLI options for `env.type=robomme`
-
-| Option               | Default       | Description                                        |
-| -------------------- | ------------- | -------------------------------------------------- |
-| `env.task`           | `PickXtimes`  | Any of the 16 task names above (comma-separated)   |
-| `env.dataset_split`  | `test`        | `train`, `val`, or `test`                          |
-| `env.action_space`   | `joint_angle` | `joint_angle` (8-D) or `ee_pose` (7-D)             |
-| `env.episode_length` | `300`         | Max steps per episode                              |
-| `env.task_ids`       | `null`        | List of episode indices to evaluate (null = `[0]`) |
-
-## Dataset
-
-The dataset [`lerobot/robomme`](https://huggingface.co/datasets/lerobot/robomme) is in **LeRobot v3.0 format** and can be loaded directly:
-
-```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-dataset = LeRobotDataset("lerobot/robomme")
-```
-
-### Dataset features
-
-| Feature            | Shape         | Description                     |
-| ------------------ | ------------- | ------------------------------- |
-| `image`            | (256, 256, 3) | Front camera RGB                |
-| `wrist_image`      | (256, 256, 3) | Wrist camera RGB                |
-| `actions`          | (8,)          | Joint angles + gripper          |
-| `state`            | (8,)          | Joint positions + gripper state |
-| `simple_subgoal`   | str           | High-level language annotation  |
-| `grounded_subgoal` | str           | Grounded language annotation    |
-| `episode_index`    | int           | Episode ID                      |
-| `frame_index`      | int           | Frame within episode            |
-
-### Feature key alignment (training)
-
-The env wrapper exposes `pixels/image` and `pixels/wrist_image` as observation keys. The `features_map` in `RoboMMEEnv` maps these to `observation.images.image` and `observation.images.wrist_image` for the policy. State is exposed as `agent_pos` and maps to `observation.state`.
-
-The dataset's `image` and `wrist_image` columns already align with the policy input keys, so no renaming is needed when fine-tuning.
-
-## Action Spaces
-
-| Type          | Dim | Description                                               |
-| ------------- | --- | --------------------------------------------------------- |
-| `joint_angle` | 8   | 7 joint angles + 1 gripper (−1 closed, +1 open, absolute) |
-| `ee_pose`     | 7   | xyz + roll/pitch/yaw + gripper                            |
-
-Set via `--env.action_space=joint_angle` (default) or `--env.action_space=ee_pose`.
-
-## Platform Notes
-
-- **Linux only**: ManiSkill requires SAPIEN/Vulkan. macOS and Windows are not supported.
-- **GPU recommended**: Rendering is CPU-capable but slow; CUDA + Vulkan gives full speed.
-- **gymnasium / numpy conflict**: See installation note above. Docker image handles this automatically.
-- **ManiSkill fork**: `robomme` depends on a specific ManiSkill fork (`YinpeiDai/ManiSkill`), pulled in automatically via the `robomme` package.

docs/source/robotwin.mdx

@@ -1,223 +0,0 @@
-# RoboTwin 2.0
-
-RoboTwin 2.0 is a **large-scale dual-arm manipulation benchmark** built on the SAPIEN physics engine. It provides a standardized evaluation protocol for bimanual robotic policies across 50 tasks (as of upstream `main`) with strong domain randomization (clutter, lighting, background, tabletop height, and language instructions).
-
-- Paper: [RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation](https://arxiv.org/abs/2506.18088)
-- GitHub: [RoboTwin-Platform/RoboTwin](https://github.com/RoboTwin-Platform/RoboTwin)
-- Leaderboard: [robotwin-platform.github.io/leaderboard](https://robotwin-platform.github.io/leaderboard)
-- Dataset: [lerobot/robotwin_unified](https://huggingface.co/datasets/lerobot/robotwin_unified)
-
-![RoboTwin 2.0 benchmark overview](https://www.aitntnews.com/pictures/2025/7/8/9a7f79cb-5ba9-11f0-8581-fa163e47d677.png)
-
-## Overview
-
-| Property      | Value                                                    |
-| ------------- | -------------------------------------------------------- |
-| Tasks         | 50 dual-arm manipulation tasks                           |
-| Robot         | Aloha-AgileX bimanual (14 DOF, 7 per arm)                |
-| Action space  | 14-dim joint-space, continuous in `[-1, 1]`              |
-| Cameras       | `head_camera`, `left_camera`, `right_camera`             |
-| Simulator     | SAPIEN (not MuJoCo)                                      |
-| Eval protocol | 100 episodes/task, 50 demo_clean demonstrations          |
-| Eval settings | **Easy** (`demo_clean`) and **Hard** (`demo_randomized`) |
-
-## Available tasks
-
-RoboTwin 2.0 ships 50 dual-arm manipulation tasks in its upstream `envs/` directory. The canonical list is the `ROBOTWIN_TASKS` tuple in `src/lerobot/envs/robotwin.py`, mirrored verbatim from the upstream repo. Example tasks:
-
-| Task                     | CLI name                 | Category          |
-| ------------------------ | ------------------------ | ----------------- |
-| Beat block with hammer   | `beat_block_hammer`      | Tool use          |
-| Click bell / alarm clock | `click_bell`             | Precision press   |
-| Stack blocks (2 / 3)     | `stack_blocks_two/three` | Stacking          |
-| Stack bowls (2 / 3)      | `stack_bowls_two/three`  | Stacking          |
-| Handover block / mic     | `handover_block`         | Bimanual coord.   |
-| Lift pot                 | `lift_pot`               | Bimanual lift     |
-| Shake bottle             | `shake_bottle`           | Continuous motion |
-| Turn switch              | `turn_switch`            | Articulated obj   |
-| Stamp seal               | `stamp_seal`             | Precision place   |
-| Scan object              | `scan_object`            | Mobile manip.     |
-
-Pass a comma-separated list to `--env.task` to run multiple tasks in a single eval sweep.
-
-<Tip warning={true}>
-  `open_laptop` is currently broken upstream (its `check_success()` uses
-  `self.arm_tag`, which is only set inside the scripted-expert `play_once()`
-  path and therefore unavailable during normal policy eval). Avoid it until the
-  upstream bug is fixed, or patch the task to default `self.arm_tag = "left"` in
-  `load_actors()`.
-</Tip>
-
-## Dataset
-
-The RoboTwin 2.0 dataset is available in **LeRobot v3.0 format** on the Hugging Face Hub:
-
-```
-lerobot/robotwin_unified
-```
-
-It contains over 100,000 pre-collected trajectories across all 50 tasks (79.6 GB, Apache 2.0 license). No format conversion is needed — it is already in the correct LeRobot v3.0 schema with video observations and action labels.
-
-You can load it directly with the HF Datasets library:
-
-```python
-from datasets import load_dataset
-
-ds = load_dataset("lerobot/robotwin_unified", split="train")
-```
-
-## Installation
-
-RoboTwin 2.0 requires **Linux** with an NVIDIA GPU (CUDA 12.1 recommended). Installation takes approximately 20 minutes.
-
-### 1. Create a conda environment
-
-```bash
-conda create -n robotwin python=3.10 -y
-conda activate robotwin
-```
-
-### 2. Install LeRobot
-
-```bash
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-pip install -e "."
-```
-
-### 3. Install RoboTwin 2.0
-
-```bash
-git clone https://github.com/RoboTwin-Platform/RoboTwin.git
-cd RoboTwin
-bash script/_install.sh
-bash script/_download_assets.sh
-```
-
-The install script handles all Python dependencies including SAPIEN, CuRobo, mplib, and pytorch3d.
-
-<Tip warning={true}>
-If the automated install fails, install manually:
-
-```bash
-pip install -r requirements.txt
-pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"
-cd envs && git clone https://github.com/NVlabs/curobo.git && cd curobo
-pip install -e . --no-build-isolation
-```
-
-Then apply the required mplib fix: in `mplib/planner.py` line 807, remove `or collide` from the conditional.
-
-</Tip>
-
-### 4. Add RoboTwin to PYTHONPATH
-
-The RoboTwin task modules must be importable by LeRobot. From within the `RoboTwin/` directory:
-
-```bash
-export PYTHONPATH="${PYTHONPATH}:$(pwd)"
-```
-
-Add this to your shell profile to make it permanent.
-
-## Evaluation
-
-### Standard evaluation (recommended)
-
-Evaluate a policy on a single task with the official protocol (100 episodes):
-
-```bash
-lerobot-eval \
-  --policy.path="your-hf-policy-id" \
-  --env.type=robotwin \
-  --env.task=beat_block_hammer \
-  --eval.batch_size=1 \
-  --eval.n_episodes=100
-```
-
-### Single-task quick check
-
-```bash
-lerobot-eval \
-  --policy.path="your-hf-policy-id" \
-  --env.type=robotwin \
-  --env.task=beat_block_hammer \
-  --eval.batch_size=1 \
-  --eval.n_episodes=5
-```
-
-### Multi-task sweep
-
-Evaluate on several tasks in one run:
-
-```bash
-lerobot-eval \
-  --policy.path="your-hf-policy-id" \
-  --env.type=robotwin \
-  --env.task=beat_block_hammer,click_bell,handover_block,stack_blocks_two \
-  --eval.batch_size=1 \
-  --eval.n_episodes=100
-```
-
-### Full benchmark (all 50 tasks)
-
-```bash
-lerobot-eval \
-  --policy.path="your-hf-policy-id" \
-  --env.type=robotwin \
-  --env.task=adjust_bottle,beat_block_hammer,blocks_ranking_rgb,blocks_ranking_size,click_alarmclock,click_bell,dump_bin_bigbin,grab_roller,handover_block,handover_mic,hanging_mug,lift_pot,move_can_pot,move_pillbottle_pad,move_playingcard_away,move_stapler_pad,open_microwave,pick_diverse_bottles,pick_dual_bottles,place_a2b_left,place_a2b_right,place_bread_basket,place_bread_skillet,place_burger_fries,place_can_basket,place_cans_plasticbox,place_container_plate,place_dual_shoes,place_empty_cup,place_fan,place_mouse_pad,place_object_basket,place_object_scale,place_object_stand,place_phone_stand,place_shoe,press_stapler,put_bottles_dustbin,put_object_cabinet,rotate_qrcode,scan_object,shake_bottle,shake_bottle_horizontally,stack_blocks_three,stack_blocks_two,stack_bowls_three,stack_bowls_two,stamp_seal,turn_switch \
-  --eval.batch_size=1 \
-  --eval.n_episodes=100
-```
-
-<Tip>
-  `open_laptop` is intentionally omitted above because of the upstream
-  `self.arm_tag` bug (see the **Available tasks** section). Re-add it once the
-  upstream fix lands.
-</Tip>
-
-## Camera configuration
-
-By default, all three cameras are included:
-
-| Camera key     | Description                    |
-| -------------- | ------------------------------ |
-| `head_camera`  | Torso-mounted overhead view    |
-| `left_camera`  | Left arm wrist-mounted camera  |
-| `right_camera` | Right arm wrist-mounted camera |
-
-To use a subset of cameras, override `--env.camera_names`:
-
-```bash
-lerobot-eval \
-  --policy.path="your-hf-policy-id" \
-  --env.type=robotwin \
-  --env.task=beat_block_hammer \
-  --env.camera_names="head_camera,left_camera" \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10
-```
-
-## Environment config reference
-
-Key parameters for `RoboTwinEnvConfig`:
-
-| Parameter            | Default                                  | Description                        |
-| -------------------- | ---------------------------------------- | ---------------------------------- |
-| `task`               | `"beat_block_hammer"`                    | Comma-separated task name(s)       |
-| `fps`                | `25`                                     | Simulation FPS                     |
-| `episode_length`     | `300`                                    | Max steps per episode              |
-| `obs_type`           | `"pixels_agent_pos"`                     | `"pixels"` or `"pixels_agent_pos"` |
-| `camera_names`       | `"head_camera,left_camera,right_camera"` | Comma-separated active cameras     |
-| `observation_height` | `240`                                    | Camera pixel height                |
-| `observation_width`  | `320`                                    | Camera pixel width                 |
-
-## Leaderboard submission
-
-Results can be submitted to the [RoboTwin 2.0 leaderboard](https://robotwin-platform.github.io/leaderboard). The official protocol requires:
-
-- Training on 50 `demo_clean` demonstrations per task
-- Evaluating 100 episodes per task
-- Reporting success rate separately for **Easy** (`demo_clean`) and **Hard** (`demo_randomized`) settings
-
-For submission instructions, refer to the [RoboTwin 2.0 documentation](https://robotwin-platform.github.io/doc/).

docs/source/vlabench.mdx

@@ -1,176 +0,0 @@
-# VLABench
-
-[VLABench](https://github.com/OpenMOSS/VLABench) is a large-scale benchmark for **language-conditioned robotic manipulation with long-horizon reasoning**. The upstream suite covers 100 task categories across 2,000+ objects and evaluates six dimensions of robot intelligence: mesh & texture understanding, spatial reasoning, world-knowledge transfer, semantic instruction comprehension, physical-law understanding, and long-horizon planning. Built on MuJoCo / dm_control with a Franka Panda 7-DOF arm. LeRobot exposes **43 of these tasks** through `--env.task` (21 primitives + 22 composites, see [Available tasks](#available-tasks) below).
-
-- Paper: [VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning](https://arxiv.org/abs/2412.18194)
-- GitHub: [OpenMOSS/VLABench](https://github.com/OpenMOSS/VLABench)
-- Project website: [vlabench.github.io](https://vlabench.github.io)
-- Pretrained policy: [`lerobot/smolvla_vlabench`](https://huggingface.co/lerobot/smolvla_vlabench)
-
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/vlabench.png"
-  alt="VLABench benchmark overview"
-  width="85%"
-/>
-
-## Available tasks
-
-VLABench ships two task suites covering **43 task categories** in LeRobot's `--env.task` surface:
-
-| Suite     | CLI name    | Tasks | Description                                                      |
-| --------- | ----------- | ----- | ---------------------------------------------------------------- |
-| Primitive | `primitive` | 21    | Single / few-skill combinations (select, insert, physics QA)     |
-| Composite | `composite` | 22    | Multi-step reasoning and long-horizon planning (cook, rearrange) |
-
-**Primitive tasks:** `select_fruit`, `select_toy`, `select_chemistry_tube`, `add_condiment`, `select_book`, `select_painting`, `select_drink`, `insert_flower`, `select_billiards`, `select_ingredient`, `select_mahjong`, `select_poker`, and physical-reasoning tasks (`density_qa`, `friction_qa`, `magnetism_qa`, `reflection_qa`, `simple_cuestick_usage`, `simple_seesaw_usage`, `sound_speed_qa`, `thermal_expansion_qa`, `weight_qa`).
-
-**Composite tasks:** `cluster_billiards`, `cluster_book`, `cluster_drink`, `cluster_toy`, `cook_dishes`, `cool_drink`, `find_unseen_object`, `get_coffee`, `hammer_nail`, `heat_food`, `make_juice`, `play_mahjong`, `play_math_game`, `play_poker`, `play_snooker`, `rearrange_book`, `rearrange_chemistry_tube`, `set_dining_table`, `set_study_table`, `store_food`, `take_chemistry_experiment`, `use_seesaw_complex`.
-
-`--env.task` accepts three forms:
-
-- a single task name (`select_fruit`)
-- a comma-separated list (`select_fruit,heat_food`)
-- a suite shortcut (`primitive`, `composite`, or `primitive,composite`)
-
-## Installation
-
-VLABench is **not on PyPI** — its only distribution is the [OpenMOSS/VLABench](https://github.com/OpenMOSS/VLABench) GitHub repo — so LeRobot does not expose a `vlabench` extra. Install it manually as an editable clone, alongside the MuJoCo / dm_control pins VLABench needs, then fetch the mesh assets:
-
-```bash
-# After following the standard LeRobot installation instructions.
-
-git clone https://github.com/OpenMOSS/VLABench.git ~/VLABench
-git clone https://github.com/motion-planning/rrt-algorithms.git ~/rrt-algorithms
-pip install -e ~/VLABench -e ~/rrt-algorithms
-pip install "mujoco==3.2.2" "dm-control==1.0.22" \
-            open3d colorlog scikit-learn openai gdown
-
-python ~/VLABench/scripts/download_assets.py
-```
-
-<Tip>
-VLABench requires Linux (`sys_platform == 'linux'`) and Python 3.10+. Set the MuJoCo rendering backend before running:
-
-```bash
-export MUJOCO_GL=egl  # for headless servers (HPC, cloud)
-```
-
-</Tip>
-
-## Evaluation
-
-All eval snippets below mirror the command CI runs (see `.github/workflows/benchmark_tests.yml`). The `--rename_map` argument maps VLABench's `image` / `second_image` / `wrist_image` camera keys onto the three-camera (`camera1` / `camera2` / `camera3`) input layout the released `smolvla_vlabench` policy was trained on.
-
-### Single-task evaluation (recommended for quick iteration)
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_vlabench \
-  --env.type=vlabench \
-  --env.task=select_fruit \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
-```
-
-### Multi-task evaluation
-
-Pass a comma-separated list of tasks:
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_vlabench \
-  --env.type=vlabench \
-  --env.task=select_fruit,select_toy,add_condiment,heat_food \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
-```
-
-### Suite-wide evaluation
-
-Run an entire suite (all 21 primitives or all 22 composites):
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_vlabench \
-  --env.type=vlabench \
-  --env.task=primitive \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  --env.max_parallel_tasks=1 \
-  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
-```
-
-Or both suites:
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/smolvla_vlabench \
-  --env.type=vlabench \
-  --env.task=primitive,composite \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --eval.use_async_envs=false \
-  --policy.device=cuda \
-  --env.max_parallel_tasks=1 \
-  '--rename_map={"observation.images.image": "observation.images.camera1", "observation.images.second_image": "observation.images.camera2", "observation.images.wrist_image": "observation.images.camera3"}'
-```
-
-### Recommended evaluation episodes
-
-**10 episodes per task** for reproducible benchmarking (210 total for the full primitive suite, 220 for composite). Matches the protocol in the VLABench paper.
-
-## Policy inputs and outputs
-
-**Observations:**
-
-- `observation.state` — 7-dim end-effector state (position xyz + Euler xyz + gripper)
-- `observation.images.image` — front camera, 480×480 HWC uint8
-- `observation.images.second_image` — second camera, 480×480 HWC uint8
-- `observation.images.wrist_image` — wrist camera, 480×480 HWC uint8
-
-**Actions:**
-
-- Continuous control in `Box(-1, 1, shape=(7,))` — 3D position + 3D Euler orientation + 1D gripper.
-
-## Training
-
-### Datasets
-
-Pre-collected VLABench datasets in LeRobot format on the Hub:
-
-- [`VLABench/vlabench_primitive_ft_lerobot_video`](https://huggingface.co/datasets/VLABench/vlabench_primitive_ft_lerobot_video) — 5,000 episodes, 128 tasks, 480×480 images.
-- [`VLABench/vlabench_composite_ft_lerobot_video`](https://huggingface.co/datasets/VLABench/vlabench_composite_ft_lerobot_video) — 5,977 episodes, 167 tasks, 224×224 images.
-
-### Example training command
-
-Fine-tune a SmolVLA base on the primitive suite:
-
-```bash
-lerobot-train \
-  --policy.type=smolvla \
-  --policy.repo_id=${HF_USER}/smolvla_vlabench_primitive \
-  --policy.load_vlm_weights=true \
-  --policy.push_to_hub=true \
-  --dataset.repo_id=VLABench/vlabench_primitive_ft_lerobot_video \
-  --env.type=vlabench \
-  --env.task=select_fruit \
-  --output_dir=./outputs/smolvla_vlabench_primitive \
-  --steps=100000 \
-  --batch_size=4 \
-  --eval_freq=5000 \
-  --eval.batch_size=1 \
-  --eval.n_episodes=1 \
-  --save_freq=10000
-```
-
-## Reproducing published results
-
-The released checkpoint [`lerobot/smolvla_vlabench`](https://huggingface.co/lerobot/smolvla_vlabench) was trained on the primitive-suite dataset above and is evaluated with the [Single-task](#single-task-evaluation-recommended-for-quick-iteration) / [Suite-wide](#suite-wide-evaluation) commands. CI runs a 10-primitive-task smoke eval (one episode each) on every PR touching the benchmark.

docs/source/xvla.mdx

@@ -220,7 +220,7 @@ REAL_DIM = 12
 # Postprocessing: Trim 20D predictions to 12D for deployment
 ```
 
-See the [action_hub.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/action_hub.py) implementation for details.
+See the [action_hub.py](/home/jade_choghari/robot/lerobot/src/lerobot/policies/xvla/action_hub.py) implementation for details.
 
 #### Auto Action Mode (Recommended)
 
@@ -519,9 +519,9 @@ If you use X-VLA in your research, please cite:
 
 - [X-VLA Paper](https://arxiv.org/pdf/2510.10274)
 - [LeRobot Documentation](https://github.com/huggingface/lerobot)
-- [Action Registry Implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/action_hub.py)
-- [Processor Implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/processor_xvla.py)
-- [Model Configuration](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/xvla/configuration_xvla.py)
+- [Action Registry Implementation](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/action_hub.py)
+- [Processor Implementation](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/processor_xvla.py)
+- [Model Configuration](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/configuration_xvla.py)
 
 ## Contributing
 

examples/notebooks/quickstart.ipynb

@@ -1,342 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# 🤗 LeRobot Quickstart\n",
-    "\n",
-    "Calibration → teleoperation → data collection → training → evaluation.\n",
-    "\n",
-    "Install the required dependencies: `pip install -e .[notebook,dataset,training,viz,hardware]`.\n",
-    "\n",
-    "**How to use:**\n",
-    "1. Edit the **Configuration** cell with your settings.\n",
-    "2. Run all cells (`Run All`).\n",
-    "3. Each section prints a ready-to-paste terminal command - copy it and run it.\n",
-    "\n",
-    "Each setup is different, please refer to the [LeRobot documentation](https://huggingface.co/docs/lerobot/il_robots) for more details on each step and available options. <br>\n",
-    "Feel free to make this notebook your own and adapt it to your needs!"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "---\n",
-    "## Utils"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def _cameras_arg(cameras: dict) -> str:\n",
-    "    if not cameras:\n",
-    "        return \"\"\n",
-    "    entries = [f\"{n}: {{{', '.join(f'{k}: {v}' for k, v in cfg.items())}}}\" for n, cfg in cameras.items()]\n",
-    "    return \"{ \" + \", \".join(entries) + \" }\"\n",
-    "\n",
-    "\n",
-    "def print_cmd(*parts: str) -> None:\n",
-    "    \"\"\"Print a shell command with line continuations, skipping empty parts.\"\"\"\n",
-    "    non_empty = [p for p in parts if p]\n",
-    "    print(\" \\\\\\n    \".join(non_empty))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "---\n",
-    "## Configuration\n",
-    "\n",
-    "Edit this cell, then **Run All** to generate all commands below."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Robot (follower) - run `lerobot-find-port` to discover the port\n",
-    "ROBOT_TYPE = \"so101_follower\"\n",
-    "ROBOT_PORT = \"/dev/ttyACM0\"\n",
-    "ROBOT_ID = \"my_follower_arm\"\n",
-    "\n",
-    "# Teleop (leader) - run `lerobot-find-port` to discover the port\n",
-    "TELEOP_TYPE = \"so101_leader\"\n",
-    "TELEOP_PORT = \"/dev/ttyACM1\"\n",
-    "TELEOP_ID = \"my_leader_arm\"\n",
-    "\n",
-    "# Cameras - set to {} to disable\n",
-    "# Run `lerobot-find-cameras opencv` to list available cameras and their indices\n",
-    "CAMERAS = {\n",
-    "    \"top\": {\"type\": \"opencv\", \"index_or_path\": 2, \"width\": 640, \"height\": 480, \"fps\": 30},\n",
-    "    \"wrist\": {\"type\": \"opencv\", \"index_or_path\": 4, \"width\": 640, \"height\": 480, \"fps\": 30},\n",
-    "}\n",
-    "\n",
-    "# Dataset\n",
-    "HF_USER = \"your_hf_username\"  # `huggingface-cli whoami` to find your username\n",
-    "DATASET_NAME = \"my_so101_dataset\"\n",
-    "TASK_DESCRIPTION = \"pick and place the block\"\n",
-    "NUM_EPISODES = 10\n",
-    "\n",
-    "# Training\n",
-    "POLICY_TYPE = \"act\"  # act, diffusion, smolvla, ...\n",
-    "POLICY_DEVICE = \"cuda\"  # cuda / cpu / mps\n",
-    "TRAIN_STEPS = 10_000\n",
-    "SAVE_FREQ = 2_000\n",
-    "OUTPUT_DIR = f\"outputs/train/{DATASET_NAME}\"\n",
-    "\n",
-    "# Inference - Hub repo ID or local checkpoint path\n",
-    "# e.g. set to f\"{OUTPUT_DIR}/checkpoints/last\" to use a local checkpoint\n",
-    "POLICY_PATH = f\"{HF_USER}/{DATASET_NAME}_{POLICY_TYPE}\"\n",
-    "LAST_CHECKPOINT_PATH = f\"{OUTPUT_DIR}/checkpoints/last\"\n",
-    "\n",
-    "# Derived\n",
-    "DATASET_REPO_ID = f\"{HF_USER}/{DATASET_NAME}\"\n",
-    "DATASET_ROOT = f\"data/{DATASET_NAME}\"\n",
-    "POLICY_REPO_ID = f\"{HF_USER}/{DATASET_NAME}_{POLICY_TYPE}\"\n",
-    "EVAL_REPO_ID = f\"{HF_USER}/eval_{DATASET_NAME}\"\n",
-    "CAMERAS_ARG = _cameras_arg(CAMERAS)\n",
-    "CAMERAS_FLAG = f'--robot.cameras=\"{CAMERAS_ARG}\"' if CAMERAS_ARG else \"\"\n",
-    "\n",
-    "print(f\"Robot  : {ROBOT_TYPE} @ {ROBOT_PORT}\")\n",
-    "print(f\"Teleop : {TELEOP_TYPE} @ {TELEOP_PORT}\")\n",
-    "print(f\"Cameras: {list(CAMERAS) or 'none'}\")\n",
-    "print(f\"Dataset: {DATASET_REPO_ID} ({NUM_EPISODES} episodes) saved to {DATASET_ROOT}\")\n",
-    "print(f\"Policy : {POLICY_TYPE} -> {POLICY_REPO_ID}\")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "---\n",
-    "## 1. Calibration\n",
-    "\n",
-    "Run once per arm before first use."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Follower\n",
-    "print_cmd(\n",
-    "    \"lerobot-calibrate\",\n",
-    "    f\"--robot.type={ROBOT_TYPE}\",\n",
-    "    f\"--robot.port={ROBOT_PORT}\",\n",
-    "    f\"--robot.id={ROBOT_ID}\",\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Leader\n",
-    "print_cmd(\n",
-    "    \"lerobot-calibrate\",\n",
-    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
-    "    f\"--teleop.port={TELEOP_PORT}\",\n",
-    "    f\"--teleop.id={TELEOP_ID}\",\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "---\n",
-    "## 2. Teleoperation\n",
-    "\n",
-    "See the [teleoperation docs](https://huggingface.co/docs/lerobot/il_robots#teleoperate) and the [cameras guide](https://huggingface.co/docs/lerobot/cameras) for more options."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "print_cmd(\n",
-    "    \"lerobot-teleoperate\",\n",
-    "    f\"--robot.type={ROBOT_TYPE}\",\n",
-    "    f\"--robot.port={ROBOT_PORT}\",\n",
-    "    f\"--robot.id={ROBOT_ID}\",\n",
-    "    CAMERAS_FLAG,\n",
-    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
-    "    f\"--teleop.port={TELEOP_PORT}\",\n",
-    "    f\"--teleop.id={TELEOP_ID}\",\n",
-    "    \"--display_data=true\",\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "---\n",
-    "## 3. Record Dataset\n",
-    "\n",
-    "See the [recording docs](https://huggingface.co/docs/lerobot/il_robots#record-a-dataset) for tips on gathering good data."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "print_cmd(\n",
-    "    \"lerobot-record\",\n",
-    "    f\"--robot.type={ROBOT_TYPE}\",\n",
-    "    f\"--robot.port={ROBOT_PORT}\",\n",
-    "    f\"--robot.id={ROBOT_ID}\",\n",
-    "    CAMERAS_FLAG,\n",
-    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
-    "    f\"--teleop.port={TELEOP_PORT}\",\n",
-    "    f\"--teleop.id={TELEOP_ID}\",\n",
-    "    f\"--dataset.repo_id={DATASET_REPO_ID}\",\n",
-    "    f\"--dataset.num_episodes={NUM_EPISODES}\",\n",
-    "    f'--dataset.single_task=\"{TASK_DESCRIPTION}\"',\n",
-    "    \"--dataset.streaming_encoding=true\",\n",
-    "    \"--display_data=true\",\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Resume a previously interrupted recording session\n",
-    "print_cmd(\n",
-    "    \"lerobot-record\",\n",
-    "    f\"--robot.type={ROBOT_TYPE}\",\n",
-    "    f\"--robot.port={ROBOT_PORT}\",\n",
-    "    f\"--robot.id={ROBOT_ID}\",\n",
-    "    CAMERAS_FLAG,\n",
-    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
-    "    f\"--teleop.port={TELEOP_PORT}\",\n",
-    "    f\"--teleop.id={TELEOP_ID}\",\n",
-    "    f\"--dataset.repo_id={DATASET_REPO_ID}\",\n",
-    "    f\"--dataset.root={DATASET_ROOT}\",\n",
-    "    f\"--dataset.num_episodes={NUM_EPISODES}\",\n",
-    "    f'--dataset.single_task=\"{TASK_DESCRIPTION}\"',\n",
-    "    \"--dataset.streaming_encoding=true\",\n",
-    "    \"--display_data=true\",\n",
-    "    \"--resume=true\",\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "---\n",
-    "## 4. Train Policy\n",
-    "\n",
-    "See the [training docs](https://huggingface.co/docs/lerobot/il_robots#train-a-policy) for configuration options and tips."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "print_cmd(\n",
-    "    \"lerobot-train\",\n",
-    "    f\"--dataset.repo_id={DATASET_REPO_ID}\",\n",
-    "    f\"--policy.type={POLICY_TYPE}\",\n",
-    "    f\"--policy.device={POLICY_DEVICE}\",\n",
-    "    f\"--policy.repo_id={POLICY_REPO_ID}\",\n",
-    "    f\"--output_dir={OUTPUT_DIR}\",\n",
-    "    f\"--steps={TRAIN_STEPS}\",\n",
-    "    f\"--save_freq={SAVE_FREQ}\",\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Resume a previously interrupted training session\n",
-    "print_cmd(\n",
-    "    \"lerobot-train\",\n",
-    "    f\"--config_path={LAST_CHECKPOINT_PATH}/pretrained_model/train_config.json\",\n",
-    "    \"--resume=true\",\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "---\n",
-    "## 5. Inference\n",
-    "\n",
-    "Uses `POLICY_PATH` from the Configuration cell (defaults to the Hub repo ID). You can also put there the `LAST_CHECKPOINT_PATH`.\n",
-    "\n",
-    "See the [inference docs](https://huggingface.co/docs/lerobot/il_robots#run-inference-and-evaluate-your-policy) for details."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "print_cmd(\n",
-    "    \"lerobot-record\",\n",
-    "    f\"--policy.path={POLICY_PATH}\",\n",
-    "    f\"--robot.type={ROBOT_TYPE}\",\n",
-    "    f\"--robot.port={ROBOT_PORT}\",\n",
-    "    f\"--robot.id={ROBOT_ID}\",\n",
-    "    CAMERAS_FLAG,\n",
-    "    f\"--teleop.type={TELEOP_TYPE}\",\n",
-    "    f\"--teleop.port={TELEOP_PORT}\",\n",
-    "    f\"--teleop.id={TELEOP_ID}\",\n",
-    "    f\"--dataset.repo_id={EVAL_REPO_ID}\",\n",
-    "    f\"--dataset.num_episodes={NUM_EPISODES}\",\n",
-    "    f'--dataset.single_task=\"{TASK_DESCRIPTION}\"',\n",
-    "    \"--dataset.streaming_encoding=true\",\n",
-    ")"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "lerobot (3.12.3)",
-   "language": "python",
-   "name": "python3"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.12.3"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 4
-}

examples/so100_to_so100_EE/teleoperate_with_intervention.py

@@ -0,0 +1,170 @@
+# !/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""SO100 leader / follower teleop with HIL-SERL-style intervention toggle.
+
+This is a position-only standalone demo of the leader-arm intervention pattern
+used by the HIL-SERL training stack (see ``lerobot.processor.LeaderArmInterventionStep``
+and ``lerobot.teleoperators.so_leader.SOLeaderFollower``).
+
+Behaviour:
+    * **Following mode** (default): The follower is idle, the leader is
+      torque-enabled and haptically tracks the follower's pose. The user can
+      grab the leader at any time without fighting the position loop.
+    * **Intervention mode** (toggled by pressing SPACE): The leader's torque is
+      released, the user moves the leader freely and the follower mirrors the
+      leader's end-effector position via ``[delta_x, delta_y, delta_z]`` deltas,
+      identical to how the real HIL-SERL action pipeline records interventions.
+
+Keyboard:
+    * ``SPACE`` -- toggle intervention on/off.
+    * ``q``     -- exit the loop cleanly.
+"""
+
+from __future__ import annotations
+
+import time
+
+import numpy as np
+
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+from lerobot.teleoperators.so_leader import SOLeaderFollower, SOLeaderTeleopConfig
+from lerobot.teleoperators.utils import TeleopEvents
+from lerobot.utils.robot_utils import precise_sleep
+
+FPS = 30
+
+# Per-axis EE-delta normalization (metres). Same convention as
+# `LeaderArmInterventionStep`: the normalised delta is `(p_leader - p_follower) / step`,
+# clipped to [-1, 1]. Keep these small so a single tick is a safe motion.
+EE_STEP_SIZES = {"x": 0.010, "y": 0.010, "z": 0.010}
+
+# Workspace bounds (metres) -- a tight box around the resting pose to keep the
+# follower from running into its joint limits during the demo.
+EE_BOUNDS = {"min": np.array([-0.20, -0.30, 0.02]), "max": np.array([0.30, 0.30, 0.40])}
+
+URDF_PATH = "./SO101/so101_new_calib.urdf"
+TARGET_FRAME = "gripper_frame_link"
+
+
+def _joints_dict_to_array(joints: dict[str, float], motor_names: list[str]) -> np.ndarray:
+    return np.array([joints[f"{m}.pos"] for m in motor_names], dtype=float)
+
+
+def _array_to_joints_dict(arr: np.ndarray, motor_names: list[str]) -> dict[str, float]:
+    return {f"{m}.pos": float(v) for m, v in zip(motor_names, arr, strict=True)}
+
+
+def main() -> None:
+    follower_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411", id="my_follower_arm", use_degrees=True
+    )
+    leader_config = SOLeaderTeleopConfig(
+        port="/dev/tty.usbmodem5A460819811",
+        id="my_leader_arm",
+        use_degrees=True,
+        leader_follower_mode=True,
+        use_gripper=True,
+    )
+
+    follower = SO100Follower(follower_config)
+    leader = SOLeaderFollower(leader_config)
+
+    follower_motor_names = list(follower.bus.motors.keys())
+    leader_motor_names = list(leader.bus.motors.keys())
+
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo:
+    # https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    follower_kinematics = RobotKinematics(
+        urdf_path=URDF_PATH, target_frame_name=TARGET_FRAME, joint_names=follower_motor_names
+    )
+    leader_kinematics = RobotKinematics(
+        urdf_path=URDF_PATH, target_frame_name=TARGET_FRAME, joint_names=leader_motor_names
+    )
+
+    follower.connect()
+    leader.connect()
+
+    print("Starting leader-follower intervention demo...")
+    print("  - Press SPACE to toggle intervention.")
+    print("  - Press 'q' to exit.")
+
+    try:
+        while True:
+            t0 = time.perf_counter()
+
+            # 1. Read both arms.
+            follower_obs = follower.get_observation()
+            follower_joints_dict = {f"{m}.pos": float(follower_obs[f"{m}.pos"]) for m in follower_motor_names}
+            leader_joints_dict = leader.get_action()
+
+            # 2. Haptic follow: push follower joints back to the leader. The
+            # leader's `send_action` gates motor writes on its intervention
+            # state internally (torque on while following, off while intervening).
+            leader.send_action(follower_joints_dict)
+
+            # 3. Pull teleop events (SPACE toggle, 'q' terminate).
+            events = leader.get_teleop_events()
+            if events.get(TeleopEvents.TERMINATE_EPISODE):
+                print("Termination requested -- exiting.")
+                break
+
+            is_intervention = events.get(TeleopEvents.IS_INTERVENTION, False)
+
+            if is_intervention:
+                # 4a. Compute leader/follower EE poses, take the *normalised
+                # position-only delta*, and integrate it onto the follower's
+                # current EE pose to get a target. This mirrors the action
+                # space recorded by `LeaderArmInterventionStep` during HIL-SERL.
+                leader_arr = _joints_dict_to_array(leader_joints_dict, leader_motor_names)
+                follower_arr = _joints_dict_to_array(follower_joints_dict, follower_motor_names)
+
+                p_leader = leader_kinematics.forward_kinematics(leader_arr)[:3, 3]
+                p_follower_mat = follower_kinematics.forward_kinematics(follower_arr)
+                p_follower = p_follower_mat[:3, 3]
+
+                raw_delta = p_leader - p_follower
+                step_vec = np.array([EE_STEP_SIZES["x"], EE_STEP_SIZES["y"], EE_STEP_SIZES["z"]], dtype=float)
+                delta_norm = np.clip(raw_delta / step_vec, -1.0, 1.0)
+                delta_m = delta_norm * step_vec
+
+                target_pose = p_follower_mat.copy()
+                target_pose[:3, 3] = np.clip(p_follower + delta_m, EE_BOUNDS["min"], EE_BOUNDS["max"])
+
+                # IK -> joint-space goal for the follower's arm chain. The
+                # gripper joint is kept separate and driven from the leader's
+                # gripper position directly (no IK).
+                target_joints = follower_kinematics.inverse_kinematics(
+                    current_joint_pos=follower_arr,
+                    desired_ee_pose=target_pose,
+                    orientation_weight=0.0,
+                )
+                follower_action = _array_to_joints_dict(target_joints, follower_motor_names)
+                follower_action["gripper.pos"] = float(leader_joints_dict.get("gripper.pos", 50.0))
+                follower.send_action(follower_action)
+            # 4b. Following mode: leave the follower alone -- the leader just
+            # tracks it haptically. In real HIL-SERL training this is where the
+            # policy would step the follower forward.
+
+            precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+    finally:
+        leader.disconnect()
+        follower.disconnect()
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/rl/hilserl_example.py

@@ -4,13 +4,13 @@ from pathlib import Path
 from queue import Empty, Full
 
 import torch
-import torch.optim as optim
 
 from lerobot.datasets import LeRobotDataset
 from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
-from lerobot.policies import SACConfig
-from lerobot.policies.sac.modeling_sac import SACPolicy
-from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
+from lerobot.policies import GaussianActorConfig
+from lerobot.policies.gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy
+from lerobot.policies.gaussian_actor.reward_model.modeling_classifier import Classifier
+from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
 from lerobot.robots.so_follower import SO100FollowerConfig
@@ -28,7 +28,7 @@ def run_learner(
     transitions_queue: mp.Queue,
     parameters_queue: mp.Queue,
     shutdown_event: mp.Event,
-    policy_learner: SACPolicy,
+    policy_learner: GaussianActorPolicy,
     online_buffer: ReplayBuffer,
     offline_buffer: ReplayBuffer,
     lr: float = 3e-4,
@@ -40,8 +40,9 @@ def run_learner(
     policy_learner.train()
     policy_learner.to(device)
 
-    # Create Adam optimizer from scratch - simple and clean
-    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)
+    algo_config = SACAlgorithmConfig.from_policy_config(policy_learner.config)
+    algorithm = SACAlgorithm(policy=policy_learner, config=algo_config)
+    algorithm.make_optimizers_and_scheduler()
 
     print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
     print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
@@ -83,24 +84,26 @@ def run_learner(
                 else:
                     batch[key] = online_batch[key]
 
-            loss, _ = policy_learner.forward(batch)
+            def batch_iter(b=batch):
+                while True:
+                    yield b
 
-            optimizer.zero_grad()
-            loss.backward()
-            optimizer.step()
+            stats = algorithm.update(batch_iter())
             training_step += 1
 
             if training_step % LOG_EVERY == 0:
+                log_dict = stats.to_log_dict()
                 print(
-                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
+                    f"[LEARNER] Training step {training_step}, "
+                    f"critic_loss: {log_dict.get('critic', 'N/A'):.4f}, "
                     f"Buffers: Online={len(online_buffer)}, Offline={len(offline_buffer)}"
                 )
 
             # Send updated parameters to actor every 10 training steps
             if training_step % SEND_EVERY == 0:
                 try:
-                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
-                    parameters_queue.put_nowait(state_dict)
+                    weights = algorithm.get_weights()
+                    parameters_queue.put_nowait(weights)
                     print("[LEARNER] Sent updated parameters to actor")
                 except Full:
                     # Missing write due to queue not being consumed (should happen rarely)
@@ -113,7 +116,7 @@ def run_actor(
     transitions_queue: mp.Queue,
     parameters_queue: mp.Queue,
     shutdown_event: mp.Event,
-    policy_actor: SACPolicy,
+    policy_actor: GaussianActorPolicy,
     reward_classifier: Classifier,
     env_cfg: HILSerlRobotEnvConfig,
     device: torch.device = "mps",
@@ -144,15 +147,15 @@ def run_actor(
 
             while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
                 try:
-                    new_params = parameters_queue.get_nowait()
-                    policy_actor.load_state_dict(new_params)
+                    new_weights = parameters_queue.get_nowait()
+                    policy_actor.load_state_dict(new_weights)
                     print("[ACTOR] Updated policy parameters from learner")
                 except Empty:  # No new updated parameters available from learner, waiting
                     pass
 
-                # Get action from policy
+                # Get action from policy (returns full action: continuous + discrete)
                 policy_obs = make_policy_obs(obs, device=device)
-                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
+                action_tensor = policy_actor.select_action(policy_obs)
                 action = action_tensor.squeeze(0).cpu().numpy()
 
                 # Step environment
@@ -261,14 +264,14 @@ def main():
     action_features = hw_to_dataset_features(env.robot.action_features, "action")
 
     # Create SAC policy for action selection
-    policy_cfg = SACConfig(
+    policy_cfg = GaussianActorConfig(
         device=device,
         input_features=obs_features,
         output_features=action_features,
     )
 
-    policy_actor = SACPolicy(policy_cfg)
-    policy_learner = SACPolicy(policy_cfg)
+    policy_actor = GaussianActorPolicy(policy_cfg)
+    policy_learner = GaussianActorPolicy(policy_cfg)
 
     demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
     offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)

pyproject.toml

@@ -108,9 +108,9 @@ training = [
     "wandb>=0.24.0,<0.25.0",
 ]
 hardware = [
-    "lerobot[pynput-dep]",
-    "lerobot[pyserial-dep]",
-    "lerobot[deepdiff-dep]",
+    "pynput>=1.7.8,<1.9.0",
+    "pyserial>=3.5,<4.0",
+    "deepdiff>=7.0.1,<9.0.0",
 ]
 viz = [
     "rerun-sdk>=0.24.0,<0.27.0",
@@ -136,14 +136,10 @@ scipy-dep = ["scipy>=1.14.0,<2.0.0"]
 diffusers-dep = ["diffusers>=0.27.2,<0.36.0"]
 qwen-vl-utils-dep = ["qwen-vl-utils>=0.0.11,<0.1.0"]
 matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0", "contourpy>=1.3.0,<2.0.0"] # NOTE: Explicitly listing contourpy helps the resolver converge faster.
-pyserial-dep = ["pyserial>=3.5,<4.0"]
-deepdiff-dep = ["deepdiff>=7.0.1,<9.0.0"]
-pynput-dep = ["pynput>=1.7.8,<1.9.0"]
-pyzmq-dep = ["pyzmq>=26.2.1,<28.0.0"]
 
 # Motors
-feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
-dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
+feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
+dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
 damiao = ["lerobot[can-dep]"]
 robstride = ["lerobot[can-dep]"]
 
@@ -151,11 +147,10 @@ robstride = ["lerobot[can-dep]"]
 openarms = ["lerobot[damiao]"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
-lekiwi = ["lerobot[feetech]", "lerobot[pyzmq-dep]"]
+lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
 unitree_g1 = [
     # "unitree-sdk2==1.0.1",
-    "lerobot[pyzmq-dep]",
-    "lerobot[pyserial-dep]",
+    "pyzmq>=26.2.1,<28.0.0",
     "onnxruntime>=1.16.0,<2.0.0",
     "onnx>=1.16.0,<2.0.0",
     "meshcat>=0.3.0,<0.4.0",
@@ -201,8 +196,7 @@ async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
 peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
 
 # Development
-dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1", "ruff>=0.14.1", "lerobot[notebook]"]
-notebook = ["jupyter>=1.0.0,<2.0.0", "ipykernel>=6.0.0,<7.0.0"]
+dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1", "ruff>=0.14.1"]
 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"]
 
@@ -212,20 +206,6 @@ aloha = ["lerobot[dataset]", "gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
 pusht = ["lerobot[dataset]", "gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
 libero = ["lerobot[dataset]", "lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
 metaworld = ["lerobot[dataset]", "metaworld==3.0.0", "lerobot[scipy-dep]"]
-# NOTE: vlabench is NOT exposed as a `lerobot` extra. Its only distribution
-# is the OpenMOSS/VLABench GitHub repo (package name `VLABench`, no PyPI
-# release), so any `vlabench>=X` pip spec is unresolvable. Install it
-# manually alongside MuJoCo / dm-control — see docs/source/vlabench.mdx
-# for the recipe.
-# NOTE: robomme is NOT a pyproject extra — mani-skill hard-pins numpy<2
-# which conflicts with lerobot's numpy>=2 base pin, so the two trees can't
-# resolve into a single env. Install it only in the RoboMME Docker image
-# via `uv pip install --override` (see docker/Dockerfile.benchmark.robomme).
-# NOTE: robocasa is NOT exposed as a `lerobot` extra. Its setup.py pins
-# `lerobot==0.3.3` in install_requires, which cyclically shadows our own
-# workspace `lerobot` and makes the graph unsolvable under any resolver
-# (uv, pip). Install it manually alongside robosuite — see
-# docs/source/robocasa.mdx for the recipe.
 
 # All
 all = [

scripts/ci/extract_task_descriptions.py

@@ -31,23 +31,9 @@ from __future__ import annotations
 
 import argparse
 import json
-import re
 import sys
 from pathlib import Path
 
-# LIBERO-plus derives task.language by space-joining the perturbation-variant
-# filename (grab_language_from_filename in libero/libero/benchmark/__init__.py),
-# so non-_language_ variants inherit a trailing metadata blob like
-# "view 0 0 100 0 0 initstate 0 noise 45" or "add 16". Strip those tokens so
-# the description matches the base instruction used in the training dataset.
-_LIBERO_PERTURBATION_TAIL_RE = re.compile(
-    r"(?:\s(?:view|initstate|noise|add|tb|table|light|level)(?:\s\d+)+)+$"
-)
-
-
-def _strip_libero_perturbation_tail(instruction: str) -> str:
-    return _LIBERO_PERTURBATION_TAIL_RE.sub("", instruction).strip()
-
 
 def _libero_descriptions(task_suite: str) -> dict[str, str]:
     from libero.libero import benchmark  # type: ignore[import-untyped]
@@ -61,10 +47,7 @@ def _libero_descriptions(task_suite: str) -> dict[str, str]:
         )
         return {}
     suite = suite_dict[task_suite]()
-    return {
-        f"{task_suite}_{i}": _strip_libero_perturbation_tail(suite.get_task(i).language)
-        for i in range(suite.n_tasks)
-    }
+    return {f"{task_suite}_{i}": suite.get_task(i).language for i in range(suite.n_tasks)}
 
 
 def _metaworld_descriptions(task_name: str) -> dict[str, str]:
@@ -74,120 +57,19 @@ def _metaworld_descriptions(task_name: str) -> dict[str, str]:
     return {f"{task_name}_0": label}
 
 
-def _robotwin_descriptions(task_names: str) -> dict[str, str]:
-    """Return descriptions for each requested RoboTwin task. Reads
-    `description/task_instruction/<task>.json` from the RoboTwin clone
-    (cwd is /opt/robotwin in CI). Falls back to the task name if missing."""
-    out: dict[str, str] = {}
-    root = Path("description/task_instruction")
-    for name in (t.strip() for t in task_names.split(",") if t.strip()):
-        desc_file = root / f"{name}.json"
-        desc = name.replace("_", " ")
-        if desc_file.is_file():
-            data = json.loads(desc_file.read_text())
-            full = data.get("full_description") or desc
-            # Strip the schema placeholders ({A}, {a}) — keep the sentence readable.
-            desc = full.replace("<", "").replace(">", "")
-        out[f"{name}_0"] = desc
-    return out
-
-
-def _robocasa_descriptions(task_spec: str) -> dict[str, str]:
-    """For each task in the comma-separated list, emit a cleaned-name label.
-
-    RoboCasa episodes carry their language instruction in the env's
-    `ep_meta['lang']`, populated per reset. Pulling it requires spinning
-    up the full kitchen env per task (~seconds each); we use the task
-    name as the key here and let the eval's episode info carry the
-    actual instruction.
-    """
-    out: dict[str, str] = {}
-    for task in (t.strip() for t in task_spec.split(",") if t.strip()):
-        # Split CamelCase into words: "CloseFridge" → "close fridge".
-        label = "".join(f" {c.lower()}" if c.isupper() else c for c in task).strip()
-        out[f"{task}_0"] = label or task
-    return out
-
-
-_ROBOMME_DESCRIPTIONS = {
-    "BinFill": "Fill the target bin with the correct number of cubes",
-    "PickXtimes": "Pick the indicated cube the specified number of times",
-    "SwingXtimes": "Swing the object the specified number of times",
-    "StopCube": "Grasp and stop the moving cube",
-    "VideoUnmask": "Pick the cube shown in the reference video",
-    "VideoUnmaskSwap": "Pick the cube matching the reference video after a swap",
-    "ButtonUnmask": "Press the button indicated by the reference",
-    "ButtonUnmaskSwap": "Press the correct button after objects are swapped",
-    "PickHighlight": "Pick the highlighted cube",
-    "VideoRepick": "Repick the cube shown in the reference video",
-    "VideoPlaceButton": "Place the cube on the button shown in the video",
-    "VideoPlaceOrder": "Place cubes in the order shown in the video",
-    "MoveCube": "Move the cube to the target location",
-    "InsertPeg": "Insert the peg into the target hole",
-    "PatternLock": "Unlock the pattern by pressing buttons in sequence",
-    "RouteStick": "Route the stick through the required waypoints",
-}
-
-
-def _robomme_descriptions(task_names: str, task_ids: list[int] | None = None) -> dict[str, str]:
-    """Return descriptions for each requested RoboMME task. Keys match the
-    video filename pattern `<task>_<task_id>` used by the eval script."""
-    if task_ids is None:
-        task_ids = [0]
-    out: dict[str, str] = {}
-    for name in (t.strip() for t in task_names.split(",") if t.strip()):
-        desc = _ROBOMME_DESCRIPTIONS.get(name, name)
-        for tid in task_ids:
-            out[f"{name}_{tid}"] = desc
-    return out
-
-
-def _vlabench_descriptions(task_spec: str) -> dict[str, str]:
-    """For each task in the comma-separated list, emit a cleaned-name label.
-
-    VLABench tasks carry language instructions on their dm_control task
-    object, but pulling them requires loading the full env per task
-    (~seconds each). The CI smoke-eval already captures the instruction
-    inside its episode info; this mapping is just enough to key
-    `metrics.json` by `<task>_0`.
-    """
-    out: dict[str, str] = {}
-    for task in (t.strip() for t in task_spec.split(",") if t.strip()):
-        out[f"{task}_0"] = task.replace("_", " ").strip()
-    return out
-
-
 def main() -> int:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--env", required=True, help="Environment family (libero, metaworld, ...)")
     parser.add_argument("--task", required=True, help="Task/suite name (e.g. libero_spatial)")
-    parser.add_argument(
-        "--task-ids",
-        type=str,
-        default=None,
-        help="Comma-separated task IDs (e.g. '0,1,2'). Default: [0]",
-    )
     parser.add_argument("--output", required=True, help="Path to write task_descriptions.json")
     args = parser.parse_args()
 
-    task_ids: list[int] | None = None
-    if args.task_ids:
-        task_ids = [int(x.strip()) for x in args.task_ids.split(",")]
-
     descriptions: dict[str, str] = {}
     try:
-        if args.env == ("libero", "libero_plus"):
+        if args.env == "libero":
             descriptions = _libero_descriptions(args.task)
         elif args.env == "metaworld":
             descriptions = _metaworld_descriptions(args.task)
-        elif args.env == "robotwin":
-            descriptions = _robotwin_descriptions(args.task)
-        elif args.env == "robocasa":
-            descriptions = _robocasa_descriptions(args.task)
-        elif args.env == "robomme":
-            descriptions = _robomme_descriptions(args.task, task_ids=task_ids)
-        elif args.env == "vlabench":
-            descriptions = _vlabench_descriptions(args.task)
         else:
             print(
                 f"[extract_task_descriptions] No description extractor for env '{args.env}'.",

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -33,7 +33,7 @@ import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
 
 from lerobot.utils.decorators import check_if_not_connected
-from lerobot.utils.import_utils import _reachy2_sdk_available, require_package
+from lerobot.utils.import_utils import _reachy2_sdk_available
 
 if TYPE_CHECKING or _reachy2_sdk_available:
     from reachy2_sdk.media.camera import CameraView
@@ -76,7 +76,6 @@ class Reachy2Camera(Camera):
         Args:
             config: The configuration settings for the camera.
         """
-        require_package("reachy2_sdk", extra="reachy2")
         super().__init__(config)
 
         self.config = config

src/lerobot/cameras/realsense/camera_realsense.py

@@ -17,21 +17,18 @@ Provides the RealSenseCamera class for capturing frames from Intel RealSense cam
 """
 
 import logging
-import sys
 import time
 from threading import Event, Lock, Thread
-from typing import TYPE_CHECKING, Any
+from typing import Any
 
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
 from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
-from lerobot.utils.import_utils import _pyrealsense2_available, require_package
-
-if TYPE_CHECKING or _pyrealsense2_available:
-    import pyrealsense2 as rs
-else:
-    rs = None
+try:
+    import pyrealsense2 as rs  # type: ignore  # TODO: add type stubs for pyrealsense2
+except Exception as e:
+    logging.info(f"Could not import realsense: {e}")
 
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.errors import DeviceNotConnectedError
@@ -42,7 +39,6 @@ from ..utils import get_cv2_rotation
 from .configuration_realsense import RealSenseCameraConfig
 
 logger = logging.getLogger(__name__)
-pkg_name = "pyrealsense2-macosx" if sys.platform == "darwin" else "pyrealsense2"
 
 
 class RealSenseCamera(Camera):
@@ -116,7 +112,7 @@ class RealSenseCamera(Camera):
         Args:
             config: The configuration settings for the camera.
         """
-        require_package(pkg_name, extra="intelrealsense", import_name="pyrealsense2")
+
         super().__init__(config)
 
         self.config = config

src/lerobot/cameras/zmq/camera_zmq.py

@@ -28,19 +28,12 @@ import json
 import logging
 import time
 from threading import Event, Lock, Thread
-from typing import TYPE_CHECKING, Any
+from typing import Any
 
 import cv2
 import numpy as np
 from numpy.typing import NDArray
 
-from lerobot.utils.import_utils import _zmq_available, require_package
-
-if TYPE_CHECKING or _zmq_available:
-    import zmq
-else:
-    zmq = None
-
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.errors import DeviceNotConnectedError
 
@@ -81,8 +74,8 @@ class ZMQCamera(Camera):
     """
 
     def __init__(self, config: ZMQCameraConfig):
-        require_package("pyzmq", extra="pyzmq-dep", import_name="zmq")
         super().__init__(config)
+        import zmq
 
         self.config = config
         self.server_address = config.server_address
@@ -124,6 +117,8 @@ class ZMQCamera(Camera):
         logger.info(f"Connecting to {self}...")
 
         try:
+            import zmq
+
             self.context = zmq.Context()
             self.socket = self.context.socket(zmq.SUB)
             self.socket.setsockopt_string(zmq.SUBSCRIBE, "")
@@ -185,8 +180,11 @@ class ZMQCamera(Camera):
 
         try:
             message = self.socket.recv_string()
-        except zmq.Again as e:
-            raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
+        except Exception as e:
+            # zmq is lazy-imported in connect(), so check by name to avoid a top-level import
+            if type(e).__name__ == "Again":
+                raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
+            raise
 
         # Decode JSON message
         data = json.loads(message)

src/lerobot/common/control_utils.py

@@ -28,12 +28,6 @@ import numpy as np
 import torch
 
 from lerobot.policies import PreTrainedPolicy, prepare_observation_for_inference
-from lerobot.utils.import_utils import _deepdiff_available, require_package
-
-if TYPE_CHECKING or _deepdiff_available:
-    from deepdiff import DeepDiff
-else:
-    DeepDiff = None
 
 if TYPE_CHECKING:
     from lerobot.datasets import LeRobotDataset
@@ -223,7 +217,10 @@ def sanity_check_dataset_robot_compatibility(
     Raises:
         ValueError: If any of the checked metadata fields do not match.
     """
-    require_package("deepdiff", extra="deepdiff-dep")
+    from lerobot.utils.import_utils import require_package
+
+    require_package("deepdiff", extra="hardware")
+    from deepdiff import DeepDiff
 
     from lerobot.utils.constants import DEFAULT_FEATURES
 

src/lerobot/common/train_utils.py

@@ -99,6 +99,7 @@ def save_checkpoint(
         optimizer (Optimizer | None, optional): The optimizer to save the state from. Defaults to None.
         scheduler (LRScheduler | None, optional): The scheduler to save the state from. Defaults to None.
         preprocessor: The preprocessor/pipeline to save. Defaults to None.
+        postprocessor: The postprocessor/pipeline to save. Defaults to None.
     """
     pretrained_dir = checkpoint_dir / PRETRAINED_MODEL_DIR
     policy.save_pretrained(pretrained_dir)

src/lerobot/configs/default.py

@@ -35,9 +35,6 @@ class DatasetConfig:
     revision: str | None = None
     use_imagenet_stats: bool = True
     video_backend: str = field(default_factory=get_safe_default_codec)
-    # When True, video frames are returned as uint8 tensors (0-255) instead of float32 (0.0-1.0).
-    # This reduces memory and speeds up DataLoader IPC. The training pipeline handles the conversion.
-    return_uint8: bool = False
     streaming: bool = False
 
     def __post_init__(self) -> None:

src/lerobot/configs/train.py

@@ -56,8 +56,6 @@ class TrainPipelineConfig(HubMixin):
     # Number of workers for the dataloader.
     num_workers: int = 4
     batch_size: int = 8
-    prefetch_factor: int = 4
-    persistent_workers: bool = True
     steps: int = 100_000
     eval_freq: int = 20_000
     log_freq: int = 200
@@ -209,10 +207,3 @@ class TrainPipelineConfig(HubMixin):
         cli_args = kwargs.pop("cli_args", [])
         with draccus.config_type("json"):
             return draccus.parse(cls, config_file, args=cli_args)
-
-
-@dataclass(kw_only=True)
-class TrainRLServerPipelineConfig(TrainPipelineConfig):
-    # NOTE: In RL, we don't need an offline dataset
-    # TODO: Make `TrainPipelineConfig.dataset` optional
-    dataset: DatasetConfig | None = None  # type: ignore[assignment] # because the parent class has made it's type non-optional

src/lerobot/datasets/dataset_reader.py

@@ -16,7 +16,6 @@
 """Private reader component for LeRobotDataset. Handles random-access reading (HF dataset, delta indices, video decoding)."""
 
 from collections.abc import Callable
-from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
 
 import datasets
@@ -50,7 +49,6 @@ class DatasetReader:
         video_backend: str,
         delta_timestamps: dict[str, list[float]] | None,
         image_transforms: Callable | None,
-        return_uint8: bool = False,
     ):
         """Initialize the reader with metadata, filtering, and transform config.
 
@@ -75,7 +73,6 @@ class DatasetReader:
         self._tolerance_s = tolerance_s
         self._video_backend = video_backend
         self._image_transforms = image_transforms
-        self._return_uint8 = return_uint8
 
         self.hf_dataset: datasets.Dataset | None = None
         self._absolute_to_relative_idx: dict[int, int] | None = None
@@ -108,8 +105,10 @@ class DatasetReader:
         """Build absolute-to-relative index mapping from loaded hf_dataset."""
         self._absolute_to_relative_idx = None
         if self.episodes is not None and self.hf_dataset is not None:
-            indices = self.hf_dataset.data.column("index").to_numpy()
-            self._absolute_to_relative_idx = dict(zip(indices.tolist(), range(len(indices)), strict=True))
+            self._absolute_to_relative_idx = {
+                abs_idx.item() if isinstance(abs_idx, torch.Tensor) else abs_idx: rel_idx
+                for rel_idx, abs_idx in enumerate(self.hf_dataset["index"])
+            }
 
     @property
     def num_frames(self) -> int:
@@ -236,30 +235,16 @@ class DatasetReader:
         Segmentation Fault.
         """
         ep = self._meta.episodes[ep_idx]
-
-        def _decode_single(vid_key: str, query_ts: list[float]) -> tuple[str, torch.Tensor]:
+        item = {}
+        for vid_key, query_ts in query_timestamps.items():
             from_timestamp = ep[f"videos/{vid_key}/from_timestamp"]
             shifted_query_ts = [from_timestamp + ts for ts in query_ts]
+
             video_path = self.root / self._meta.get_video_file_path(ep_idx, vid_key)
-            frames = decode_video_frames(
-                video_path,
-                shifted_query_ts,
-                self._tolerance_s,
-                self._video_backend,
-                return_uint8=self._return_uint8,
-            )
-            return vid_key, frames.squeeze(0)
+            frames = decode_video_frames(video_path, shifted_query_ts, self._tolerance_s, self._video_backend)
+            item[vid_key] = frames.squeeze(0)
 
-        items = list(query_timestamps.items())
-
-        # Single camera: no threading overhead
-        if len(items) <= 1:
-            return {vid_key: _decode_single(vid_key, query_ts)[1] for vid_key, query_ts in items}
-
-        # Multi-camera: decode in parallel (video decoding releases the GIL)
-        with ThreadPoolExecutor(max_workers=len(items)) as pool:
-            futures = [pool.submit(_decode_single, k, ts) for k, ts in items]
-            return dict(f.result() for f in futures)
+        return item
 
     def get_item(self, idx) -> dict:
         """Core __getitem__ logic. Assumes hf_dataset is loaded.

src/lerobot/datasets/dataset_writer.py

@@ -597,7 +597,7 @@ class DatasetWriter:
 
     def cleanup_interrupted_episode(self, episode_index: int) -> None:
         """Remove temporary image directories for an interrupted episode."""
-        for key in self._meta.camera_keys:
+        for key in self._meta.video_keys:
             img_dir = self._get_image_file_path(
                 episode_index=episode_index, image_key=key, frame_index=0
             ).parent

src/lerobot/datasets/factory.py

@@ -92,7 +92,6 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 image_transforms=image_transforms,
                 revision=cfg.dataset.revision,
                 video_backend=cfg.dataset.video_backend,
-                return_uint8=True,
                 tolerance_s=cfg.tolerance_s,
             )
         else:
@@ -105,7 +104,6 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 revision=cfg.dataset.revision,
                 max_num_shards=cfg.num_workers,
                 tolerance_s=cfg.tolerance_s,
-                return_uint8=True,
             )
     else:
         raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")

src/lerobot/datasets/image_writer.py

@@ -30,13 +30,13 @@ def safe_stop_image_writer(func):
     def wrapper(*args, **kwargs):
         try:
             return func(*args, **kwargs)
-        except BaseException:
+        except Exception as e:
             dataset = kwargs.get("dataset")
             writer = getattr(dataset, "writer", None) if dataset else None
             if writer is not None and writer.image_writer is not None:
                 logger.warning("Waiting for image writer to terminate...")
                 writer.image_writer.stop()
-            raise
+            raise e
 
     return wrapper
 

src/lerobot/datasets/lerobot_dataset.py

@@ -56,7 +56,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         force_cache_sync: bool = False,
         download_videos: bool = True,
         video_backend: str | None = None,
-        return_uint8: bool = False,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
         streaming_encoding: bool = False,
@@ -203,7 +202,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.tolerance_s = tolerance_s
         self.revision = revision if revision else CODEBASE_VERSION
         self._video_backend = video_backend if video_backend else get_safe_default_codec()
-        self._return_uint8 = return_uint8
         self._batch_encoding_size = batch_encoding_size
         self._vcodec = resolve_vcodec(vcodec)
         self._encoder_threads = encoder_threads
@@ -227,7 +225,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
             video_backend=self._video_backend,
             delta_timestamps=delta_timestamps,
             image_transforms=image_transforms,
-            return_uint8=self._return_uint8,
         )
 
         # Load actual data
@@ -291,7 +288,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 video_backend=self._video_backend,
                 delta_timestamps=self.delta_timestamps,
                 image_transforms=self.image_transforms,
-                return_uint8=self._return_uint8,
             )
         return self.reader
 
@@ -687,7 +683,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.delta_timestamps = None
         obj.episodes = None
         obj._video_backend = video_backend if video_backend is not None else get_safe_default_codec()
-        obj._return_uint8 = False
         obj._batch_encoding_size = batch_encoding_size
         obj._vcodec = vcodec
         obj._encoder_threads = encoder_threads
@@ -780,7 +775,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.delta_timestamps = None
         obj.episodes = None
         obj._video_backend = video_backend if video_backend else get_safe_default_codec()
-        obj._return_uint8 = False
         obj._batch_encoding_size = batch_encoding_size
         obj._vcodec = vcodec
         obj._encoder_threads = encoder_threads

src/lerobot/datasets/streaming_dataset.py

@@ -251,7 +251,6 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
         seed: int = 42,
         rng: np.random.Generator | None = None,
         shuffle: bool = True,
-        return_uint8: bool = False,
     ):
         """Initialize a StreamingLeRobotDataset.
 
@@ -289,7 +288,6 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
 
         self.streaming = streaming
         self.buffer_size = buffer_size
-        self._return_uint8 = return_uint8
 
         # We cache the video decoders to avoid re-initializing them at each frame (avoiding a ~10x slowdown)
         self.video_decoder_cache = None
@@ -555,11 +553,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
             root = self.meta.url_root if self.streaming and not self.streaming_from_local else self.root
             video_path = f"{root}/{self.meta.get_video_file_path(ep_idx, video_key)}"
             frames = decode_video_frames_torchcodec(
-                video_path,
-                query_ts,
-                self.tolerance_s,
-                decoder_cache=self.video_decoder_cache,
-                return_uint8=self._return_uint8,
+                video_path, query_ts, self.tolerance_s, decoder_cache=self.video_decoder_cache
             )
 
             item[video_key] = frames.squeeze(0) if len(query_ts) == 1 else frames

src/lerobot/datasets/video_utils.py

@@ -123,7 +123,6 @@ def decode_video_frames(
     timestamps: list[float],
     tolerance_s: float,
     backend: str | None = None,
-    return_uint8: bool = False,
 ) -> torch.Tensor:
     """
     Decodes video frames using the specified backend.
@@ -132,23 +131,19 @@ def decode_video_frames(
         video_path (Path): Path to the video file.
         timestamps (list[float]): List of timestamps to extract frames.
         tolerance_s (float): Allowed deviation in seconds for frame retrieval.
-        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available in the platform; otherwise, defaults to "pyav".
-        return_uint8 (bool): If True, return raw uint8 frames without float32 normalization.
-            This reduces memory for DataLoader IPC; normalization can be done on GPU afterward.
+        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available in the platform; otherwise, defaults to "pyav"..
 
     Returns:
-        torch.Tensor: Decoded frames (float32 in [0,1] by default, or uint8 if return_uint8=True).
+        torch.Tensor: Decoded frames.
 
     Currently supports torchcodec on cpu and pyav.
     """
     if backend is None:
         backend = get_safe_default_codec()
     if backend == "torchcodec":
-        return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s, return_uint8=return_uint8)
+        return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s)
     elif backend in ["pyav", "video_reader"]:
-        return decode_video_frames_torchvision(
-            video_path, timestamps, tolerance_s, backend, return_uint8=return_uint8
-        )
+        return decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
     else:
         raise ValueError(f"Unsupported video backend: {backend}")
 
@@ -159,7 +154,6 @@ def decode_video_frames_torchvision(
     tolerance_s: float,
     backend: str = "pyav",
     log_loaded_timestamps: bool = False,
-    return_uint8: bool = False,
 ) -> torch.Tensor:
     """Loads frames associated to the requested timestamps of a video
 
@@ -246,17 +240,14 @@ def decode_video_frames_torchvision(
     if log_loaded_timestamps:
         logger.info(f"{closest_ts=}")
 
+    # convert to the pytorch format which is float32 in [0,1] range (and channel first)
+    closest_frames = closest_frames.type(torch.float32) / 255
+
     if len(timestamps) != len(closest_frames):
         raise FrameTimestampError(
             f"Number of retrieved frames ({len(closest_frames)}) does not match "
             f"number of queried timestamps ({len(timestamps)})"
         )
-
-    if return_uint8:
-        return closest_frames
-
-    # convert to the pytorch format which is float32 in [0,1] range (and channel first)
-    closest_frames = closest_frames.type(torch.float32) / 255
     return closest_frames
 
 
@@ -315,7 +306,6 @@ def decode_video_frames_torchcodec(
     tolerance_s: float,
     log_loaded_timestamps: bool = False,
     decoder_cache: VideoDecoderCache | None = None,
-    return_uint8: bool = False,
 ) -> torch.Tensor:
     """Loads frames associated with the requested timestamps of a video using torchcodec.
 
@@ -383,16 +373,14 @@ def decode_video_frames_torchcodec(
     if log_loaded_timestamps:
         logger.info(f"{closest_ts=}")
 
+    # convert to float32 in [0,1] range
+    closest_frames = (closest_frames / 255.0).type(torch.float32)
+
     if not len(timestamps) == len(closest_frames):
         raise FrameTimestampError(
             f"Retrieved timestamps differ from queried {set(closest_frames) - set(timestamps)}"
         )
 
-    if return_uint8:
-        return closest_frames
-
-    # convert to float32 in [0,1] range
-    closest_frames = (closest_frames / 255.0).type(torch.float32)
     return closest_frames
 
 

src/lerobot/envs/configs.py

@@ -331,7 +331,6 @@ class LiberoEnv(EnvConfig):
     camera_name_mapping: dict[str, str] | None = None
     observation_height: int = 360
     observation_width: int = 360
-    is_libero_plus: bool = False
     features: dict[str, PolicyFeature] = field(
         default_factory=lambda: {
             ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
@@ -433,7 +432,6 @@ class LiberoEnv(EnvConfig):
             control_mode=self.control_mode,
             episode_length=self.episode_length,
             camera_name_mapping=self.camera_name_mapping,
-            is_libero_plus=self.is_libero_plus,
         )
 
     def get_env_processors(self):
@@ -498,146 +496,6 @@ class MetaworldEnv(EnvConfig):
         )
 
 
-@EnvConfig.register_subclass("robocasa")
-@dataclass
-class RoboCasaEnv(EnvConfig):
-    task: str = "CloseFridge"
-    fps: int = 20
-    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"
-    observation_height: int = 256
-    observation_width: int = 256
-    visualization_height: int = 512
-    visualization_width: int = 512
-    split: str | None = None
-    # Object-mesh registries to sample from. Upstream default is
-    # ("objaverse", "lightwheel"), but objaverse is ~30GB and the CI image
-    # only ships the lightwheel pack. Override to include objaverse once
-    # you've run `python -m robocasa.scripts.download_kitchen_assets
-    # --type objaverse` locally.
-    obj_registries: list[str] = field(default_factory=lambda: ["lightwheel"])
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(12,))}
-    )
-    features_map: dict[str, str] = field(default_factory=lambda: {ACTION: ACTION, "agent_pos": OBS_STATE})
-
-    def __post_init__(self):
-        if self.obs_type not in ("pixels", "pixels_agent_pos"):
-            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
-
-        # Preserve raw RoboCasa camera names end-to-end (e.g.
-        # `observation.images.robot0_agentview_left`). This matches the
-        # naming convention used by the RoboCasa datasets on the Hub, so
-        # trained policies don't need a `--rename_map` at eval time.
-        cams = [c.strip() for c in self.camera_name.split(",") if c.strip()]
-        for cam in cams:
-            self.features[f"pixels/{cam}"] = PolicyFeature(
-                type=FeatureType.VISUAL,
-                shape=(self.observation_height, self.observation_width, 3),
-            )
-            self.features_map[f"pixels/{cam}"] = f"{OBS_IMAGES}.{cam}"
-
-        if self.obs_type == "pixels_agent_pos":
-            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(16,))
-
-    @property
-    def gym_kwargs(self) -> dict:
-        kwargs: dict[str, Any] = {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-            "observation_height": self.observation_height,
-            "observation_width": self.observation_width,
-            "visualization_height": self.visualization_height,
-            "visualization_width": self.visualization_width,
-        }
-        if self.split is not None:
-            kwargs["split"] = self.split
-        return kwargs
-
-    def create_envs(self, n_envs: int, use_async_envs: bool = False):
-        from .robocasa import create_robocasa_envs
-
-        if self.task is None:
-            raise ValueError("RoboCasaEnv requires a task to be specified")
-        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
-        return create_robocasa_envs(
-            task=self.task,
-            n_envs=n_envs,
-            camera_name=self.camera_name,
-            gym_kwargs=self.gym_kwargs,
-            env_cls=env_cls,
-            episode_length=self.episode_length,
-            obj_registries=tuple(self.obj_registries),
-        )
-
-
-@EnvConfig.register_subclass("vlabench")
-@dataclass
-class VLABenchEnv(EnvConfig):
-    task: str = "select_fruit"
-    fps: int = 10
-    episode_length: int = 500
-    obs_type: str = "pixels_agent_pos"
-    render_mode: str = "rgb_array"
-    render_resolution: tuple[int, int] = (480, 480)
-    robot: str = "franka"
-    action_mode: str = "eef"
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            ACTION: ACTION,
-            "agent_pos": OBS_STATE,
-            "pixels/image": f"{OBS_IMAGES}.image",
-            "pixels/second_image": f"{OBS_IMAGES}.second_image",
-            "pixels/wrist_image": f"{OBS_IMAGES}.wrist_image",
-        }
-    )
-
-    def __post_init__(self):
-        h, w = self.render_resolution
-        if self.obs_type == "pixels":
-            self.features["pixels/image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-            self.features["pixels/second_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-            self.features["pixels/wrist_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-        elif self.obs_type == "pixels_agent_pos":
-            self.features["pixels/image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-            self.features["pixels/second_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-            self.features["pixels/wrist_image"] = PolicyFeature(type=FeatureType.VISUAL, shape=(h, w, 3))
-            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(7,))
-        else:
-            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-            "render_resolution": self.render_resolution,
-            "robot": self.robot,
-            "max_episode_steps": self.episode_length,
-            "action_mode": self.action_mode,
-        }
-
-    def create_envs(self, n_envs: int, use_async_envs: bool = False):
-        from .vlabench import create_vlabench_envs
-
-        if self.task is None:
-            raise ValueError("VLABenchEnv requires a task to be specified")
-        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
-        return create_vlabench_envs(
-            task=self.task,
-            n_envs=n_envs,
-            gym_kwargs=self.gym_kwargs,
-            env_cls=env_cls,
-        )
-
-
 @EnvConfig.register_subclass("isaaclab_arena")
 @dataclass
 class IsaaclabArenaEnv(HubEnvConfig):
@@ -716,171 +574,3 @@ class IsaaclabArenaEnv(HubEnvConfig):
             ),
             PolicyProcessorPipeline(steps=[]),
         )
-
-
-@EnvConfig.register_subclass("libero_plus")
-@dataclass
-class LiberoPlusEnv(LiberoEnv):
-    """Config for LIBERO-plus robustness benchmark evaluation.
-
-    LIBERO-plus extends LIBERO with 7 perturbation dimensions (camera viewpoints,
-    object layouts, robot initial states, language instructions, lighting, background
-    textures, sensor noise) producing ~10k task variants.
-
-    The gym interface is identical to LIBERO so this class reuses ``LiberoEnv``
-    entirely — only the registered name and default task suite differ.
-
-    Install: see docker/Dockerfile.benchmark.libero_plus — LIBERO-plus ships
-    as a namespace package from a git fork and must be cloned + PYTHONPATH'd
-    rather than installed as a pyproject extra.
-
-    See Also:
-        https://github.com/sylvestf/LIBERO-plus
-    """
-
-    task: str = "libero_spatial"
-    is_libero_plus: bool = True
-
-
-@EnvConfig.register_subclass("robotwin")
-@dataclass
-class RoboTwinEnvConfig(EnvConfig):
-    """Configuration for RoboTwin 2.0 benchmark environments.
-
-    RoboTwin 2.0 is a dual-arm manipulation benchmark with 50 tasks built on the
-    SAPIEN simulator. The robot is an Aloha-AgileX bimanual platform with 14 DOF
-    (7 per arm). All three cameras are enabled by default.
-
-    See: https://robotwin-platform.github.io
-    Dataset: https://huggingface.co/datasets/lerobot/robotwin_unified
-    """
-
-    task: str = "beat_block_hammer"  # single task or comma-separated list
-    fps: int = 25
-    episode_length: int = 300
-    obs_type: str = "pixels_agent_pos"
-    render_mode: str = "rgb_array"
-    # Available cameras from RoboTwin's aloha-agilex embodiment: head_camera
-    # (torso-mounted) + left_camera / right_camera (wrists).
-    camera_names: str = "head_camera,left_camera,right_camera"
-    # Match the D435 dims in task_config/demo_clean.yml (_camera_config.yml).
-    # Gym's vector-env concatenate pre-allocates buffers of this shape, so it
-    # must equal what SAPIEN actually renders.
-    observation_height: int = 240
-    observation_width: int = 320
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            ACTION: ACTION,
-            "pixels/head_camera": f"{OBS_IMAGES}.head_camera",
-            "pixels/left_camera": f"{OBS_IMAGES}.left_camera",
-            "pixels/right_camera": f"{OBS_IMAGES}.right_camera",
-            "agent_pos": OBS_STATE,
-        }
-    )
-
-    def __post_init__(self):
-        cam_list = [c.strip() for c in self.camera_names.split(",") if c.strip()]
-        for cam in cam_list:
-            self.features[f"pixels/{cam}"] = PolicyFeature(
-                type=FeatureType.VISUAL,
-                shape=(self.observation_height, self.observation_width, 3),
-            )
-            # Keep features_map entry if already set (default_factory); add if missing.
-            key = f"pixels/{cam}"
-            if key not in self.features_map:
-                self.features_map[key] = f"{OBS_IMAGES}.{cam}"
-
-        if self.obs_type == "pixels_agent_pos":
-            self.features["agent_pos"] = PolicyFeature(
-                type=FeatureType.STATE,
-                shape=(14,),  # 14 DOF: 7 per arm
-            )
-        elif self.obs_type != "pixels":
-            raise ValueError(
-                f"Unsupported obs_type '{self.obs_type}'. "
-                "RoboTwinEnvConfig supports 'pixels' and 'pixels_agent_pos'."
-            )
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {}
-
-    def create_envs(self, n_envs: int, use_async_envs: bool = True):
-        from lerobot.envs.robotwin import create_robotwin_envs
-
-        if not self.task:
-            raise ValueError("RoboTwinEnvConfig requires `task` to be specified.")
-
-        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
-        cam_list = [c.strip() for c in self.camera_names.split(",") if c.strip()]
-        return create_robotwin_envs(
-            task=self.task,
-            n_envs=n_envs,
-            env_cls=env_cls,
-            camera_names=cam_list,
-            observation_height=self.observation_height,
-            observation_width=self.observation_width,
-            episode_length=self.episode_length,
-        )
-
-
-@EnvConfig.register_subclass("robomme")
-@dataclass
-class RoboMMEEnv(EnvConfig):
-    """RoboMME memory-augmented manipulation benchmark (ManiSkill/SAPIEN).
-
-    16 tasks across 4 suites: Counting, Permanence, Reference, Imitation.
-    Dataset: lerobot/robomme (LeRobot v3.0, 1,600 episodes).
-    Benchmark: https://github.com/RoboMME/robomme_benchmark
-
-    Requires the `robomme` git package installed separately (Linux only);
-    see docker/Dockerfile.benchmark.robomme for the canonical install.
-    """
-
-    task: str = "PickXtimes"
-    fps: int = 10
-    episode_length: int = 300
-    action_space: str = "joint_angle"  # or "ee_pose" (7-D)
-    dataset_split: str = "test"  # "train" | "val" | "test"
-    task_ids: list[int] | None = None
-    features: dict[str, PolicyFeature] = field(default_factory=dict)
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            ACTION: ACTION,
-            "pixels/image": f"{OBS_IMAGES}.image",
-            "pixels/wrist_image": f"{OBS_IMAGES}.wrist_image",
-            "agent_pos": OBS_STATE,
-        }
-    )
-
-    def __post_init__(self):
-        action_dim = 8 if self.action_space == "joint_angle" else 7
-        self.features = {
-            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_dim,)),
-            "pixels/image": PolicyFeature(type=FeatureType.VISUAL, shape=(256, 256, 3)),
-            "pixels/wrist_image": PolicyFeature(type=FeatureType.VISUAL, shape=(256, 256, 3)),
-            "agent_pos": PolicyFeature(type=FeatureType.STATE, shape=(8,)),
-        }
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {}
-
-    def create_envs(self, n_envs: int, use_async_envs: bool = True):
-        from lerobot.envs.robomme import create_robomme_envs
-
-        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
-        return create_robomme_envs(
-            task=self.task,
-            n_envs=n_envs,
-            action_space_type=self.action_space,
-            dataset=self.dataset_split,
-            episode_length=self.episode_length,
-            task_ids=self.task_ids,
-            env_cls=env_cls,
-        )

src/lerobot/envs/libero.py

@@ -16,7 +16,6 @@
 from __future__ import annotations
 
 import os
-import re
 from collections import defaultdict
 from collections.abc import Callable, Iterable, Mapping, Sequence
 from functools import partial
@@ -32,7 +31,20 @@ from libero.libero.envs import OffScreenRenderEnv
 
 from lerobot.types import RobotObservation
 
-from .utils import _LazyAsyncVectorEnv, parse_camera_names
+from .utils import _LazyAsyncVectorEnv
+
+
+def _parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
+    """Normalize camera_name into a non-empty list of strings."""
+    if isinstance(camera_name, str):
+        cams = [c.strip() for c in camera_name.split(",") if c.strip()]
+    elif isinstance(camera_name, (list | tuple)):
+        cams = [str(c).strip() for c in camera_name if str(c).strip()]
+    else:
+        raise TypeError(f"camera_name must be str or sequence[str], got {type(camera_name).__name__}")
+    if not cams:
+        raise ValueError("camera_name resolved to an empty list.")
+    return cams
 
 
 def _get_suite(name: str) -> benchmark.Benchmark:
@@ -57,34 +69,14 @@ def _select_task_ids(total_tasks: int, task_ids: Iterable[int] | None) -> list[i
     return ids
 
 
-# LIBERO-plus perturbation variants encode the perturbation in the filename
-# but on disk only the base `.pruned_init` exists — strip the suffix to match
-# LIBERO-plus's own suite.get_task_init_states() (we reimplement it here so we
-# can pass weights_only=False for PyTorch 2.6+ numpy pickles).
-_LIBERO_PERTURBATION_SUFFIX_RE = re.compile(r"_(?:language|view|light)_[^.]*|_(?:table|tb)_\d+")
-
-
-def get_task_init_states(task_suite: Any, i: int, is_libero_plus: bool = False) -> np.ndarray:
-    task = task_suite.tasks[i]
-    filename = Path(task.init_states_file)
-    root = Path(get_libero_path("init_states"))
-
-    if not is_libero_plus:
-        init_states_path = root / task.problem_folder / filename.name
-        return torch.load(init_states_path, weights_only=False)  # nosec B614
-
-    # LIBERO-plus: `_add_` / `_level` variants store extra-object layouts under
-    # libero_newobj/ as a flat array that must be reshaped to (1, -1).
-    if "_add_" in filename.name or "_level" in filename.name:
-        init_states_path = root / "libero_newobj" / task.problem_folder / filename.name
-        init_states = torch.load(init_states_path, weights_only=False)  # nosec B614
-        return init_states.reshape(1, -1)
-
-    # LIBERO-plus perturbation variants encode the perturbation in the filename
-    # but on disk only the base `.pruned_init` exists — strip the suffix to match.
-    stripped = _LIBERO_PERTURBATION_SUFFIX_RE.sub("", filename.stem) + filename.suffix
-    init_states_path = root / task.problem_folder / stripped
-    return torch.load(init_states_path, weights_only=False)  # nosec B614
+def get_task_init_states(task_suite: Any, i: int) -> np.ndarray:
+    init_states_path = (
+        Path(get_libero_path("init_states"))
+        / task_suite.tasks[i].problem_folder
+        / task_suite.tasks[i].init_states_file
+    )
+    init_states = torch.load(init_states_path, weights_only=False)  # nosec B614
+    return init_states
 
 
 def get_libero_dummy_action():
@@ -126,11 +118,9 @@ class LiberoEnv(gym.Env):
         camera_name_mapping: dict[str, str] | None = None,
         num_steps_wait: int = 10,
         control_mode: str = "relative",
-        is_libero_plus: bool = False,
     ):
         super().__init__()
         self.task_id = task_id
-        self.is_libero_plus = is_libero_plus
         self.obs_type = obs_type
         self.render_mode = render_mode
         self.observation_width = observation_width
@@ -138,7 +128,7 @@ class LiberoEnv(gym.Env):
         self.visualization_width = visualization_width
         self.visualization_height = visualization_height
         self.init_states = init_states
-        self.camera_name = parse_camera_names(
+        self.camera_name = _parse_camera_names(
             camera_name
         )  # agentview_image (main) or robot0_eye_in_hand_image (wrist)
 
@@ -157,11 +147,7 @@ class LiberoEnv(gym.Env):
         self.episode_index = episode_index
         self.episode_length = episode_length
         # Load once and keep
-        self._init_states = (
-            get_task_init_states(task_suite, self.task_id, is_libero_plus=self.is_libero_plus)
-            if self.init_states
-            else None
-        )
+        self._init_states = get_task_init_states(task_suite, self.task_id) if self.init_states else None
         self._reset_stride = n_envs  # when performing a reset, append `_reset_stride` to `init_state_id`.
 
         self.init_state_id = self.episode_index  # tie each sub-env to a fixed init state
@@ -394,7 +380,6 @@ def _make_env_fns(
     gym_kwargs: Mapping[str, Any],
     control_mode: str,
     camera_name_mapping: dict[str, str] | None = None,
-    is_libero_plus: bool = False,
 ) -> list[Callable[[], LiberoEnv]]:
     """Build n_envs factory callables for a single (suite, task_id)."""
 
@@ -411,7 +396,6 @@ def _make_env_fns(
             n_envs=n_envs,
             control_mode=control_mode,
             camera_name_mapping=camera_name_mapping,
-            is_libero_plus=is_libero_plus,
             **local_kwargs,
         )
 
@@ -434,7 +418,6 @@ def create_libero_envs(
     control_mode: str = "relative",
     episode_length: int | None = None,
     camera_name_mapping: dict[str, str] | None = None,
-    is_libero_plus: bool = False,
 ) -> dict[str, dict[int, Any]]:
     """
     Create vectorized LIBERO environments with a consistent return shape.
@@ -454,7 +437,7 @@ def create_libero_envs(
     gym_kwargs = dict(gym_kwargs or {})
     task_ids_filter = gym_kwargs.pop("task_ids", None)  # optional: limit to specific tasks
 
-    camera_names = parse_camera_names(camera_name)
+    camera_names = _parse_camera_names(camera_name)
     suite_names = [s.strip() for s in str(task).split(",") if s.strip()]
     if not suite_names:
         raise ValueError("`task` must contain at least one LIBERO suite name.")
@@ -479,7 +462,6 @@ def create_libero_envs(
         # Probe once and reuse to avoid creating a temp env per task.
         cached_obs_space: spaces.Space | None = None
         cached_act_space: spaces.Space | None = None
-        cached_metadata: dict[str, Any] | None = None
 
         for tid in selected:
             fns = _make_env_fns(
@@ -493,14 +475,12 @@ def create_libero_envs(
                 gym_kwargs=gym_kwargs,
                 control_mode=control_mode,
                 camera_name_mapping=camera_name_mapping,
-                is_libero_plus=is_libero_plus,
             )
             if is_async:
-                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
+                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space)
                 if cached_obs_space is None:
                     cached_obs_space = lazy.observation_space
                     cached_act_space = lazy.action_space
-                    cached_metadata = lazy.metadata
                 out[suite_name][tid] = lazy
             else:
                 out[suite_name][tid] = env_cls(fns)

src/lerobot/envs/metaworld.py

@@ -311,7 +311,6 @@ def create_metaworld_envs(
     is_async = env_cls is gym.vector.AsyncVectorEnv
     cached_obs_space = None
     cached_act_space = None
-    cached_metadata = None
     out: dict[str, dict[int, Any]] = defaultdict(dict)
 
     for group in task_groups:
@@ -325,11 +324,10 @@ def create_metaworld_envs(
             fns = [(lambda tn=task_name: MetaworldEnv(task=tn, **gym_kwargs)) for _ in range(n_envs)]
 
             if is_async:
-                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
+                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space)
                 if cached_obs_space is None:
                     cached_obs_space = lazy.observation_space
                     cached_act_space = lazy.action_space
-                    cached_metadata = lazy.metadata
                 out[group][tid] = lazy
             else:
                 out[group][tid] = env_cls(fns)

src/lerobot/envs/robocasa.py

@@ -1,425 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-from __future__ import annotations
-
-import logging
-from collections import defaultdict
-from collections.abc import Callable, Sequence
-from functools import partial
-from typing import Any
-
-import gymnasium as gym
-import numpy as np
-from gymnasium import spaces
-
-from lerobot.types import RobotObservation
-
-from .utils import _LazyAsyncVectorEnv, parse_camera_names
-
-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)
-ACTION_LOW = -1.0
-ACTION_HIGH = 1.0
-
-# Default PandaOmron cameras. We surface these raw names directly as
-# `observation.images.<name>` so the LeRobot dataset/policy keys match
-# RoboCasa's native convention (no implicit renaming).
-DEFAULT_CAMERAS = [
-    "robot0_agentview_left",
-    "robot0_eye_in_hand",
-    "robot0_agentview_right",
-]
-
-# Object-mesh registries to sample from. RoboCasa's upstream default is
-# ("objaverse", "lightwheel"), but the objaverse pack is huge (~30GB) and
-# most users — including our CI image — only download the lightwheel pack
-# (`--type objs_lw` in `download_kitchen_assets`). When a sampled object
-# category has zero candidates in every registry, robocasa crashes with
-# `ValueError: Probabilities contain NaN` (0/0 divide in the probability
-# normalization). Restricting to registries that are actually on disk
-# avoids the NaN and matches what the asset download provides.
-DEFAULT_OBJ_REGISTRIES: tuple[str, ...] = ("lightwheel",)
-
-# Task-group shortcuts accepted as `--env.task`. When the user passes one of
-# these names, we expand it to the upstream RoboCasa task list and auto-set
-# the dataset split. Individual task names (optionally comma-separated) still
-# take precedence; this only triggers on an exact group-name match.
-_TASK_GROUP_SPLITS = {
-    "atomic_seen": "target",
-    "composite_seen": "target",
-    "composite_unseen": "target",
-    "pretrain50": "pretrain",
-    "pretrain100": "pretrain",
-    "pretrain200": "pretrain",
-    "pretrain300": "pretrain",
-}
-
-
-def _resolve_tasks(task: str) -> tuple[list[str], str | None]:
-    """Resolve a `--env.task` value to (task_names, split_override).
-
-    If `task` is a known task-group name (e.g. `atomic_seen`, `pretrain100`),
-    expand it via `robocasa.utils.dataset_registry.{TARGET,PRETRAINING}_TASKS`
-    and return the matching split. Otherwise treat `task` as a single task or
-    comma-separated list and leave the split untouched (None).
-    """
-    key = task.strip()
-    if key in _TASK_GROUP_SPLITS:
-        from robocasa.utils.dataset_registry import PRETRAINING_TASKS, TARGET_TASKS
-
-        combined = {**TARGET_TASKS, **PRETRAINING_TASKS}
-        if key not in combined:
-            raise ValueError(
-                f"Task group '{key}' is not available in this version of robocasa. "
-                f"Known groups: {sorted(combined.keys())}."
-            )
-        return list(combined[key]), _TASK_GROUP_SPLITS[key]
-
-    names = [t.strip() for t in task.split(",") if t.strip()]
-    if not names:
-        raise ValueError("`task` must contain at least one RoboCasa task name.")
-    return names, None
-
-
-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)
-    """
-    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],
-    }
-
-
-class RoboCasaEnv(gym.Env):
-    """LeRobot gym.Env wrapper for RoboCasa365 kitchen environments.
-
-    Wraps RoboCasaGymEnv from the robocasa package and converts its
-    dict-based observations and actions into the flat arrays LeRobot expects.
-    Raw RoboCasa camera names are preserved verbatim under `pixels/<cam>`.
-    """
-
-    metadata = {"render_modes": ["rgb_array"], "render_fps": 20}
-
-    def __init__(
-        self,
-        task: str,
-        camera_name: str | Sequence[str] = ",".join(DEFAULT_CAMERAS),
-        obs_type: str = "pixels_agent_pos",
-        render_mode: str = "rgb_array",
-        observation_width: int = 256,
-        observation_height: int = 256,
-        visualization_width: int = 512,
-        visualization_height: int = 512,
-        split: str | None = None,
-        episode_length: int | None = None,
-        obj_registries: Sequence[str] = DEFAULT_OBJ_REGISTRIES,
-        episode_index: int = 0,
-    ):
-        super().__init__()
-        self.task = task
-        self.obs_type = obs_type
-        self.render_mode = render_mode
-        self.observation_width = observation_width
-        self.observation_height = observation_height
-        self.visualization_width = visualization_width
-        self.visualization_height = visualization_height
-        self.split = split
-        self.obj_registries = tuple(obj_registries)
-        # Per-worker index (0..n_envs-1) used to spread the user-provided
-        # seed across factories so each sub-env explores a distinct layout
-        # even when the same seed is passed to `reset()`.
-        self.episode_index = int(episode_index)
-
-        self.camera_name = parse_camera_names(camera_name)
-
-        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().
-        self._env: Any = None
-        self.task_description = ""
-
-        images = {
-            cam: spaces.Box(
-                low=0,
-                high=255,
-                shape=(self.observation_height, self.observation_width, 3),
-                dtype=np.uint8,
-            )
-            for cam in self.camera_name
-        }
-
-        if self.obs_type == "pixels":
-            self.observation_space = spaces.Dict({"pixels": spaces.Dict(images)})
-        elif self.obs_type == "pixels_agent_pos":
-            self.observation_space = spaces.Dict(
-                {
-                    "pixels": spaces.Dict(images),
-                    "agent_pos": spaces.Box(
-                        low=-np.inf,
-                        high=np.inf,
-                        shape=(OBS_STATE_DIM,),
-                        dtype=np.float32,
-                    ),
-                }
-            )
-        else:
-            raise ValueError(f"Unsupported obs_type '{self.obs_type}'. Use 'pixels' or 'pixels_agent_pos'.")
-
-        self.action_space = spaces.Box(
-            low=ACTION_LOW,
-            high=ACTION_HIGH,
-            shape=(ACTION_DIM,),
-            dtype=np.float32,
-        )
-
-    def _ensure_env(self) -> None:
-        """Create the underlying RoboCasaGymEnv on first use.
-
-        Called inside the worker subprocess after fork(), so each worker gets
-        its own clean rendering context rather than inheriting a stale one from
-        the parent process (which causes crashes with AsyncVectorEnv).
-        """
-        if self._env is not None:
-            return
-        from robocasa.wrappers.gym_wrapper import RoboCasaGymEnv
-
-        # RoboCasaGymEnv defaults split="test", which create_env rejects
-        # (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.
-        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,
-        )
-
-        ep_meta = self._env.env.get_ep_meta()
-        self.task_description = ep_meta.get("lang", self.task)
-
-    def _format_raw_obs(self, raw_obs: dict) -> RobotObservation:
-        """Convert RoboCasaGymEnv observation dict to LeRobot format."""
-        # RoboCasaGymEnv emits camera frames under "video.<cam>".
-        images = {cam: raw_obs[f"video.{cam}"] for cam in self.camera_name if f"video.{cam}" in raw_obs}
-
-        if self.obs_type == "pixels":
-            return {"pixels": images}
-
-        # `state.*` keys come from PandaOmronKeyConverter inside the wrapper.
-        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.gripper_qpos", np.zeros(2)),
-            ],
-            axis=-1,
-        ).astype(np.float32)
-
-        return {"pixels": images, "agent_pos": agent_pos}
-
-    def render(self) -> np.ndarray:
-        self._ensure_env()
-        assert self._env is not None
-        return self._env.render()
-
-    def reset(self, seed=None, **kwargs):
-        self._ensure_env()
-        assert self._env is not None
-        super().reset(seed=seed)
-        # Spread the seed across workers so n_envs factories don't all
-        # roll the same scene. With an explicit user seed we shift it by
-        # episode_index; with no seed we fall back to episode_index so
-        # each worker is still distinct rather than inheriting the same
-        # global RNG state.
-        worker_seed = seed + self.episode_index if seed is not None else self.episode_index
-        raw_obs, info = self._env.reset(seed=worker_seed)
-
-        ep_meta = self._env.env.get_ep_meta()
-        self.task_description = ep_meta.get("lang", self.task)
-
-        observation = self._format_raw_obs(raw_obs)
-        info = {"is_success": False}
-        return observation, info
-
-    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
-        self._ensure_env()
-        assert self._env is not None
-        if action.ndim != 1:
-            raise ValueError(
-                f"Expected action to be 1-D (shape (action_dim,)), "
-                f"but got shape {action.shape} with ndim={action.ndim}"
-            )
-
-        action_dict = convert_action(action)
-        raw_obs, reward, done, truncated, info = self._env.step(action_dict)
-
-        is_success = bool(info.get("success", False))
-        terminated = done or is_success
-        info.update({"task": self.task, "done": done, "is_success": is_success})
-
-        observation = self._format_raw_obs(raw_obs)
-        if terminated:
-            info["final_info"] = {
-                "task": self.task,
-                "done": bool(done),
-                "is_success": bool(is_success),
-            }
-            self.reset()
-
-        return observation, reward, terminated, truncated, info
-
-    def close(self):
-        if self._env is not None:
-            self._env.close()
-
-
-def _make_env_fns(
-    *,
-    task: str,
-    n_envs: int,
-    camera_names: list[str],
-    obs_type: str,
-    render_mode: str,
-    observation_width: int,
-    observation_height: int,
-    visualization_width: int,
-    visualization_height: int,
-    split: str | None,
-    episode_length: int | None,
-    obj_registries: Sequence[str],
-) -> list[Callable[[], RoboCasaEnv]]:
-    """Build n_envs factory callables for a single task.
-
-    Each factory carries a distinct ``episode_index`` (``0..n_envs-1``) so
-    ``RoboCasaEnv.reset()`` can derive a per-worker seed series from the
-    user-provided seed.
-    """
-
-    def _make_env(episode_index: int) -> RoboCasaEnv:
-        return RoboCasaEnv(
-            task=task,
-            camera_name=camera_names,
-            obs_type=obs_type,
-            render_mode=render_mode,
-            observation_width=observation_width,
-            observation_height=observation_height,
-            visualization_width=visualization_width,
-            visualization_height=visualization_height,
-            split=split,
-            episode_length=episode_length,
-            obj_registries=obj_registries,
-            episode_index=episode_index,
-        )
-
-    return [partial(_make_env, i) for i in range(n_envs)]
-
-
-def create_robocasa_envs(
-    task: str,
-    n_envs: int,
-    gym_kwargs: dict[str, Any] | None = None,
-    camera_name: str | Sequence[str] = ",".join(DEFAULT_CAMERAS),
-    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
-    episode_length: int | None = None,
-    obj_registries: Sequence[str] = DEFAULT_OBJ_REGISTRIES,
-) -> dict[str, dict[int, Any]]:
-    """Create vectorized RoboCasa365 environments with a consistent return shape.
-
-    Returns:
-        dict[task_name][task_id] -> vec_env (env_cls([...]) with exactly n_envs factories)
-
-    `task` can be:
-      - a single task name (e.g. `CloseFridge`)
-      - a comma-separated list of task names (e.g. `CloseFridge,PickPlaceCoffee`)
-      - a benchmark-group shortcut (`atomic_seen`, `composite_seen`,
-        `composite_unseen`, `pretrain50`, `pretrain100`, `pretrain200`,
-        `pretrain300`), which auto-expands to the upstream task list and
-        auto-sets the dataset `split` ("target" or "pretrain").
-    """
-    if env_cls is None or not callable(env_cls):
-        raise ValueError("env_cls must be a callable that wraps a list of environment factory callables.")
-    if not isinstance(n_envs, int) or n_envs <= 0:
-        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
-
-    gym_kwargs = dict(gym_kwargs or {})
-    obs_type = gym_kwargs.pop("obs_type", "pixels_agent_pos")
-    render_mode = gym_kwargs.pop("render_mode", "rgb_array")
-    observation_width = gym_kwargs.pop("observation_width", 256)
-    observation_height = gym_kwargs.pop("observation_height", 256)
-    visualization_width = gym_kwargs.pop("visualization_width", 512)
-    visualization_height = gym_kwargs.pop("visualization_height", 512)
-    split = gym_kwargs.pop("split", None)
-
-    camera_names = parse_camera_names(camera_name)
-    task_names, group_split = _resolve_tasks(str(task))
-    if group_split is not None and split is None:
-        split = group_split
-
-    logger.info(
-        "Creating RoboCasa envs | tasks=%s | split=%s | n_envs(per task)=%d",
-        task_names,
-        split,
-        n_envs,
-    )
-
-    is_async = env_cls is gym.vector.AsyncVectorEnv
-
-    cached_obs_space: spaces.Space | None = None
-    cached_act_space: spaces.Space | None = None
-    cached_metadata: dict[str, Any] | None = None
-    out: dict[str, dict[int, Any]] = defaultdict(dict)
-
-    for task_name in task_names:
-        fns = _make_env_fns(
-            task=task_name,
-            n_envs=n_envs,
-            camera_names=camera_names,
-            obs_type=obs_type,
-            render_mode=render_mode,
-            observation_width=observation_width,
-            observation_height=observation_height,
-            visualization_width=visualization_width,
-            visualization_height=visualization_height,
-            split=split,
-            episode_length=episode_length,
-            obj_registries=obj_registries,
-        )
-
-        if is_async:
-            lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
-            if cached_obs_space is None:
-                cached_obs_space = lazy.observation_space
-                cached_act_space = lazy.action_space
-                cached_metadata = lazy.metadata
-            out[task_name][0] = lazy
-        else:
-            out[task_name][0] = env_cls(fns)
-        logger.info("Built vec env | task=%s | n_envs=%d", task_name, n_envs)
-
-    return {name: dict(task_map) for name, task_map in out.items()}

src/lerobot/envs/robomme.py

@@ -1,245 +0,0 @@
-"""RoboMME environment wrapper for LeRobot evaluation.
-
-Wraps the RoboMME ``BenchmarkEnvBuilder`` into a Gymnasium-compatible
-``VectorEnv`` suitable for ``lerobot_eval``.
-
-RoboMME tasks:
-  Counting:    BinFill, PickXtimes, SwingXtimes, StopCube
-  Permanence:  VideoUnmask, VideoUnmaskSwap, ButtonUnmask, ButtonUnmaskSwap
-  Reference:   PickHighlight, VideoRepick, VideoPlaceButton, VideoPlaceOrder
-  Imitation:   MoveCube, InsertPeg, PatternLock, RouteStick
-
-Dataset: lerobot/robomme (LeRobot v3.0, 1,600 episodes)
-Install: see docker/Dockerfile.benchmark.robomme  (Linux only — mani-skill vs numpy pin conflict)
-Benchmark: https://github.com/RoboMME/robomme_benchmark
-"""
-
-from __future__ import annotations
-
-from collections.abc import Callable, Sequence
-from functools import partial
-from typing import Any
-
-import gymnasium as gym
-import numpy as np
-from gymnasium import spaces
-
-from .utils import _LazyAsyncVectorEnv
-
-ROBOMME_TASKS = [
-    "BinFill",
-    "PickXtimes",
-    "SwingXtimes",
-    "StopCube",
-    "VideoUnmask",
-    "VideoUnmaskSwap",
-    "ButtonUnmask",
-    "ButtonUnmaskSwap",
-    "PickHighlight",
-    "VideoRepick",
-    "VideoPlaceButton",
-    "VideoPlaceOrder",
-    "MoveCube",
-    "InsertPeg",
-    "PatternLock",
-    "RouteStick",
-]
-
-
-class RoboMMEGymEnv(gym.Env):
-    """Thin Gymnasium wrapper around a single RoboMME episode env."""
-
-    metadata = {"render_modes": ["rgb_array"], "render_fps": 10}
-
-    def __init__(
-        self,
-        task: str = "PickXtimes",
-        action_space_type: str = "joint_angle",
-        dataset: str = "test",
-        episode_idx: int = 0,
-        max_steps: int = 300,
-    ):
-        super().__init__()
-        from robomme.env_record_wrapper import BenchmarkEnvBuilder
-
-        self._task = task
-        self._action_space_type = action_space_type
-        self._dataset = dataset
-        self._episode_idx = episode_idx
-        self._max_steps = max_steps
-        self._max_episode_steps = max_steps
-
-        self._builder = BenchmarkEnvBuilder(
-            env_id=task,
-            dataset=dataset,
-            action_space=action_space_type,
-            gui_render=False,
-            max_steps=max_steps,
-        )
-        self._env = None
-        self._last_raw_obs: dict | None = None
-
-        action_dim = 8 if action_space_type == "joint_angle" else 7
-        self.action_space = spaces.Box(low=-1.0, high=1.0, shape=(action_dim,), dtype=np.float32)
-        # `pixels` must be a nested Dict so `preprocess_observation()` in
-        # envs/utils.py picks it up and maps each camera to
-        # `observation.images.<cam>`. A flat layout (`pixels/image`,
-        # `pixels/wrist_image`) silently drops every image from the batch.
-        self.observation_space = spaces.Dict(
-            {
-                "pixels": spaces.Dict(
-                    {
-                        "image": spaces.Box(0, 255, shape=(256, 256, 3), dtype=np.uint8),
-                        "wrist_image": spaces.Box(0, 255, shape=(256, 256, 3), dtype=np.uint8),
-                    }
-                ),
-                "agent_pos": spaces.Box(-np.inf, np.inf, shape=(8,), dtype=np.float32),
-            }
-        )
-
-    def reset(self, *, seed=None, options=None):
-        super().reset(seed=seed)
-        self._env = self._builder.make_env_for_episode(
-            episode_idx=self._episode_idx,
-            max_steps=self._max_steps,
-        )
-        obs, info = self._env.reset()
-        self._last_raw_obs = obs
-        return self._convert_obs(obs), self._convert_info(info)
-
-    def step(self, action):
-        obs, reward, terminated, truncated, info = self._env.step(action)
-        self._last_raw_obs = obs
-
-        terminated_bool = bool(terminated.item()) if hasattr(terminated, "item") else bool(terminated)
-        truncated_bool = bool(truncated.item()) if hasattr(truncated, "item") else bool(truncated)
-
-        status = info.get("status", "ongoing")
-        is_success = status == "success"
-        conv_info = self._convert_info(info)
-        conv_info["is_success"] = is_success
-
-        return self._convert_obs(obs), float(reward), terminated_bool, truncated_bool, conv_info
-
-    def render(self) -> np.ndarray | None:
-        """Return the front camera image from the last observation for video recording."""
-        if self._last_raw_obs is None:
-            return np.zeros((256, 256, 3), dtype=np.uint8)
-        front = self._last_raw_obs.get("front_rgb_list")
-        if front is None:
-            return np.zeros((256, 256, 3), dtype=np.uint8)
-        frame = front[-1] if isinstance(front, list) else front
-        return np.asarray(frame, dtype=np.uint8)
-
-    def _convert_obs(self, obs: dict) -> dict:
-        front_rgb = (
-            obs["front_rgb_list"][-1] if isinstance(obs["front_rgb_list"], list) else obs["front_rgb_list"]
-        )
-        wrist_rgb = (
-            obs["wrist_rgb_list"][-1] if isinstance(obs["wrist_rgb_list"], list) else obs["wrist_rgb_list"]
-        )
-        joint_state = (
-            obs["joint_state_list"][-1]
-            if isinstance(obs["joint_state_list"], list)
-            else obs["joint_state_list"]
-        )
-        gripper_state = (
-            obs["gripper_state_list"][-1]
-            if isinstance(obs["gripper_state_list"], list)
-            else obs["gripper_state_list"]
-        )
-
-        front_rgb = np.asarray(front_rgb, dtype=np.uint8)
-        wrist_rgb = np.asarray(wrist_rgb, dtype=np.uint8)
-        joint = np.asarray(joint_state, dtype=np.float32).flatten()[:7]
-        gripper = np.asarray(gripper_state, dtype=np.float32).flatten()[:1]
-        state = np.concatenate([joint, gripper])
-
-        return {
-            "pixels": {"image": front_rgb, "wrist_image": wrist_rgb},
-            "agent_pos": state,
-        }
-
-    def _convert_info(self, info: dict) -> dict:
-        return {
-            "status": info.get("status", "ongoing"),
-            "task_goal": info.get("task_goal", ""),
-        }
-
-
-def _make_env_fns(
-    *,
-    task: str,
-    n_envs: int,
-    action_space_type: str,
-    dataset: str,
-    episode_length: int,
-    task_id: int,
-) -> list[Callable[[], RoboMMEGymEnv]]:
-    """Build n_envs factory callables for one RoboMME task id."""
-
-    def _make_one(episode_index: int) -> RoboMMEGymEnv:
-        return RoboMMEGymEnv(
-            task=task,
-            action_space_type=action_space_type,
-            dataset=dataset,
-            episode_idx=episode_index,
-            max_steps=episode_length,
-        )
-
-    return [partial(_make_one, task_id + i) for i in range(n_envs)]
-
-
-def create_robomme_envs(
-    task: str,
-    n_envs: int = 1,
-    action_space_type: str = "joint_angle",
-    dataset: str = "test",
-    episode_length: int = 300,
-    task_ids: list[int] | None = None,
-    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
-) -> dict[str, dict[int, gym.vector.VectorEnv]]:
-    """Create vectorized RoboMME environments for evaluation.
-
-    `task` may be a single RoboMME task name (e.g. "PickXtimes") or a
-    comma-separated list (e.g. "PickXtimes,BinFill,StopCube"). Each task
-    becomes its own suite in the returned mapping.
-
-    Returns {suite_name: {task_id: VectorEnv}} matching lerobot's expected format.
-    """
-    if env_cls is None or not callable(env_cls):
-        raise ValueError("env_cls must be a callable that wraps a list of env factory callables.")
-    if not isinstance(n_envs, int) or n_envs <= 0:
-        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
-
-    if task_ids is None:
-        task_ids = [0]
-
-    task_names = [t.strip() for t in task.split(",") if t.strip()]
-    is_async = env_cls is gym.vector.AsyncVectorEnv
-    cached_obs_space: spaces.Space | None = None
-    cached_act_space: spaces.Space | None = None
-    cached_metadata: dict[str, Any] | None = None
-    out: dict[str, dict[int, gym.vector.VectorEnv]] = {}
-    for task_name in task_names:
-        envs_by_task: dict[int, gym.vector.VectorEnv] = {}
-        for task_id in task_ids:
-            fns = _make_env_fns(
-                task=task_name,
-                n_envs=n_envs,
-                action_space_type=action_space_type,
-                dataset=dataset,
-                episode_length=episode_length,
-                task_id=task_id,
-            )
-            if is_async:
-                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
-                if cached_obs_space is None:
-                    cached_obs_space = lazy.observation_space
-                    cached_act_space = lazy.action_space
-                    cached_metadata = lazy.metadata
-                envs_by_task[task_id] = lazy
-            else:
-                envs_by_task[task_id] = env_cls(fns)
-        out[task_name] = envs_by_task
-    return out

src/lerobot/envs/robotwin.py

@@ -1,488 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-from __future__ import annotations
-
-import importlib
-import logging
-from collections import defaultdict
-from collections.abc import Callable, Sequence
-from functools import partial
-from typing import Any
-
-import gymnasium as gym
-import numpy as np
-import torch
-from gymnasium import spaces
-
-from lerobot.types import RobotObservation
-
-from .utils import _LazyAsyncVectorEnv
-
-logger = logging.getLogger(__name__)
-
-# Camera names as used by RoboTwin 2.0. The wrapper appends "_rgb" when looking
-# up keys in get_obs() output (e.g. "head_camera" → "head_camera_rgb").
-ROBOTWIN_CAMERA_NAMES: tuple[str, ...] = (
-    "head_camera",
-    "left_camera",
-    "right_camera",
-)
-
-ACTION_DIM = 14  # 7 DOF × 2 arms
-ACTION_LOW = -1.0
-ACTION_HIGH = 1.0
-DEFAULT_EPISODE_LENGTH = 300
-# D435 dims from task_config/_camera_config.yml (what demo_clean.yml selects).
-DEFAULT_CAMERA_H = 240
-DEFAULT_CAMERA_W = 320
-
-# Task list from RoboTwin 2.0's `envs/` directory — mirrors upstream exactly
-# (50 tasks as of main; earlier revisions had 60 with a different split).
-# Keep this in sync with:
-#   gh api /repos/RoboTwin-Platform/RoboTwin/contents/envs --paginate \
-#     | jq -r '.[].name' | grep -E '\.py$' | grep -v '^_' | sed 's/\.py$//'
-ROBOTWIN_TASKS: tuple[str, ...] = (
-    "adjust_bottle",
-    "beat_block_hammer",
-    "blocks_ranking_rgb",
-    "blocks_ranking_size",
-    "click_alarmclock",
-    "click_bell",
-    "dump_bin_bigbin",
-    "grab_roller",
-    "handover_block",
-    "handover_mic",
-    "hanging_mug",
-    "lift_pot",
-    "move_can_pot",
-    "move_pillbottle_pad",
-    "move_playingcard_away",
-    "move_stapler_pad",
-    "open_laptop",
-    "open_microwave",
-    "pick_diverse_bottles",
-    "pick_dual_bottles",
-    "place_a2b_left",
-    "place_a2b_right",
-    "place_bread_basket",
-    "place_bread_skillet",
-    "place_burger_fries",
-    "place_can_basket",
-    "place_cans_plasticbox",
-    "place_container_plate",
-    "place_dual_shoes",
-    "place_empty_cup",
-    "place_fan",
-    "place_mouse_pad",
-    "place_object_basket",
-    "place_object_scale",
-    "place_object_stand",
-    "place_phone_stand",
-    "place_shoe",
-    "press_stapler",
-    "put_bottles_dustbin",
-    "put_object_cabinet",
-    "rotate_qrcode",
-    "scan_object",
-    "shake_bottle",
-    "shake_bottle_horizontally",
-    "stack_blocks_three",
-    "stack_blocks_two",
-    "stack_bowls_three",
-    "stack_bowls_two",
-    "stamp_seal",
-    "turn_switch",
-)
-
-
-_ROBOTWIN_SETUP_CACHE: dict[str, dict[str, Any]] = {}
-
-
-def _load_robotwin_setup_kwargs(task_name: str) -> dict[str, Any]:
-    """Build the kwargs dict RoboTwin's setup_demo expects.
-
-    Mirrors the config loading done by RoboTwin's ``script/eval_policy.py``:
-    reads ``task_config/demo_clean.yml``, resolves the embodiment file from
-    ``_embodiment_config.yml``, loads the robot's own ``config.yml``, and
-    reads camera dimensions from ``_camera_config.yml``.
-
-    Uses ``aloha-agilex`` single-robot dual-arm by default (the only embodiment
-    used by beat_block_hammer and most smoke-test tasks).
-    """
-    if task_name in _ROBOTWIN_SETUP_CACHE:
-        return dict(_ROBOTWIN_SETUP_CACHE[task_name])
-
-    import os
-
-    import yaml  # type: ignore[import-untyped]
-    from envs import CONFIGS_PATH  # type: ignore[import-not-found]
-
-    task_config = "demo_clean"
-    with open(os.path.join(CONFIGS_PATH, f"{task_config}.yml"), encoding="utf-8") as f:
-        args = yaml.safe_load(f)
-
-    # Resolve embodiment — demo_clean.yml uses [aloha-agilex] (dual-arm single robot)
-    with open(os.path.join(CONFIGS_PATH, "_embodiment_config.yml"), encoding="utf-8") as f:
-        embodiment_types = yaml.safe_load(f)
-    embodiment = args.get("embodiment", ["aloha-agilex"])
-    if len(embodiment) == 1:
-        robot_file = embodiment_types[embodiment[0]]["file_path"]
-        args["left_robot_file"] = robot_file
-        args["right_robot_file"] = robot_file
-        args["dual_arm_embodied"] = True
-    elif len(embodiment) == 3:
-        args["left_robot_file"] = embodiment_types[embodiment[0]]["file_path"]
-        args["right_robot_file"] = embodiment_types[embodiment[1]]["file_path"]
-        args["embodiment_dis"] = embodiment[2]
-        args["dual_arm_embodied"] = False
-    else:
-        raise ValueError(f"embodiment must have 1 or 3 items, got {len(embodiment)}")
-
-    with open(os.path.join(args["left_robot_file"], "config.yml"), encoding="utf-8") as f:
-        args["left_embodiment_config"] = yaml.safe_load(f)
-    with open(os.path.join(args["right_robot_file"], "config.yml"), encoding="utf-8") as f:
-        args["right_embodiment_config"] = yaml.safe_load(f)
-
-    # Camera dimensions
-    with open(os.path.join(CONFIGS_PATH, "_camera_config.yml"), encoding="utf-8") as f:
-        camera_config = yaml.safe_load(f)
-    head_cam = args["camera"]["head_camera_type"]
-    args["head_camera_h"] = camera_config[head_cam]["h"]
-    args["head_camera_w"] = camera_config[head_cam]["w"]
-
-    # Headless overrides
-    args["render_freq"] = 0
-    args["task_name"] = task_name
-    args["task_config"] = task_config
-
-    _ROBOTWIN_SETUP_CACHE[task_name] = args
-    return dict(args)
-
-
-def _load_robotwin_task(task_name: str) -> type:
-    """Dynamically import and return a RoboTwin 2.0 task class.
-
-    RoboTwin tasks live in ``envs/<task_name>.py`` relative to the repository
-    root and are expected to be on ``sys.path`` after installation.
-    """
-    try:
-        module = importlib.import_module(f"envs.{task_name}")
-    except ModuleNotFoundError as e:
-        raise ModuleNotFoundError(
-            f"Could not import RoboTwin task '{task_name}'. "
-            "Ensure RoboTwin 2.0 is installed and its 'envs/' directory is on PYTHONPATH. "
-            "See the RoboTwin installation guide: https://robotwin-platform.github.io/doc/usage/robotwin-install.html"
-        ) from e
-    task_cls = getattr(module, task_name, None)
-    if task_cls is None:
-        raise AttributeError(f"Task class '{task_name}' not found in envs/{task_name}.py")
-    return task_cls
-
-
-class RoboTwinEnv(gym.Env):
-    """Gymnasium wrapper around a single RoboTwin 2.0 task.
-
-    RoboTwin uses a custom SAPIEN-based API (``setup_demo`` / ``get_obs`` /
-    ``take_action`` / ``check_success``) rather than the standard gym interface.
-    This class bridges that API to Gymnasium so that ``lerobot-eval`` can drive
-    RoboTwin exactly like LIBERO or Meta-World.
-
-    The underlying SAPIEN environment is created lazily on the first ``reset()``
-    call *inside the worker process*.  This is required for
-    ``gym.vector.AsyncVectorEnv`` compatibility: SAPIEN allocates EGL/GPU
-    contexts that must not be forked from the parent process.
-
-    Observations
-    ------------
-    The ``pixels`` dict uses the raw RoboTwin camera names as keys (e.g.
-    ``"head_camera"``, ``"left_camera"``). ``preprocess_observation`` in
-    ``envs/utils.py`` then converts these to ``observation.images.<cam>``.
-
-    Actions
-    -------
-    14-dim float32 array in ``[-1, 1]`` (joint-space, 7 DOF per arm).
-
-    Autograd
-    --------
-    ``setup_demo`` and ``take_action`` drive CuRobo's Newton trajectory
-    optimizer, which calls ``cost.backward()`` internally. lerobot_eval wraps
-    the rollout in ``torch.no_grad()``, so both call sites re-enable grad.
-    """
-
-    metadata = {"render_modes": ["rgb_array"], "render_fps": 25}
-
-    def __init__(
-        self,
-        task_name: str,
-        episode_index: int = 0,
-        n_envs: int = 1,
-        camera_names: Sequence[str] = ROBOTWIN_CAMERA_NAMES,
-        observation_height: int | None = None,
-        observation_width: int | None = None,
-        episode_length: int = DEFAULT_EPISODE_LENGTH,
-        render_mode: str = "rgb_array",
-    ):
-        super().__init__()
-        self.task_name = task_name
-        self.task = task_name  # used by add_envs_task() in utils.py
-        self.task_description = task_name.replace("_", " ")
-        self.episode_index = episode_index
-        self._reset_stride = n_envs
-        self.camera_names = list(camera_names)
-        # Default to D435 dims (the camera type baked into task_config/demo_clean.yml).
-        # The YAML-driven lookup is deferred to reset() so construction doesn't
-        # import RoboTwin's `envs` module — fast-tests run without RoboTwin installed.
-        self.observation_height = observation_height or DEFAULT_CAMERA_H
-        self.observation_width = observation_width or DEFAULT_CAMERA_W
-        self.episode_length = episode_length
-        self._max_episode_steps = episode_length  # lerobot_eval.rollout reads this
-        self.render_mode = render_mode
-
-        self._env: Any | None = None  # deferred — created on first reset() inside worker
-        self._step_count: int = 0
-        self._black_frame = np.zeros((self.observation_height, self.observation_width, 3), dtype=np.uint8)
-
-        image_spaces = {
-            cam: spaces.Box(
-                low=0,
-                high=255,
-                shape=(self.observation_height, self.observation_width, 3),
-                dtype=np.uint8,
-            )
-            for cam in self.camera_names
-        }
-        self.observation_space = spaces.Dict(
-            {
-                "pixels": spaces.Dict(image_spaces),
-                "agent_pos": spaces.Box(low=-np.inf, high=np.inf, shape=(ACTION_DIM,), dtype=np.float32),
-            }
-        )
-        self.action_space = spaces.Box(
-            low=ACTION_LOW, high=ACTION_HIGH, shape=(ACTION_DIM,), dtype=np.float32
-        )
-
-    def _ensure_env(self) -> None:
-        """Create the SAPIEN environment on first use.
-
-        Called inside the worker subprocess after fork(), so each worker gets
-        its own EGL/GPU context rather than inheriting a stale one from the
-        parent process (which causes crashes with AsyncVectorEnv).
-        """
-        if self._env is not None:
-            return
-        task_cls = _load_robotwin_task(self.task_name)
-        self._env = task_cls()
-
-    def _get_obs(self) -> RobotObservation:
-        assert self._env is not None, "_get_obs called before _ensure_env()"
-        raw = self._env.get_obs()
-        cameras_raw = raw.get("observation", {})
-
-        images: dict[str, np.ndarray] = {}
-        for cam in self.camera_names:
-            cam_data = cameras_raw.get(cam)
-            img = cam_data.get("rgb") if cam_data else None
-            if img is None:
-                images[cam] = self._black_frame
-                continue
-            img = np.asarray(img, dtype=np.uint8)
-            if img.ndim == 2:
-                img = np.stack([img, img, img], axis=-1)
-            elif img.shape[-1] != 3:
-                img = img[..., :3]
-            images[cam] = img
-
-        ja = raw.get("joint_action") or {}
-        vec = ja.get("vector")
-        if vec is not None:
-            arr = np.asarray(vec, dtype=np.float32).ravel()
-            joint_state = (
-                arr[:ACTION_DIM] if arr.size >= ACTION_DIM else np.zeros(ACTION_DIM, dtype=np.float32)
-            )
-        else:
-            joint_state = np.zeros(ACTION_DIM, dtype=np.float32)
-
-        return {"pixels": images, "agent_pos": joint_state}
-
-    def reset(self, seed: int | None = None, **kwargs) -> tuple[RobotObservation, dict]:
-        self._ensure_env()
-        super().reset(seed=seed)
-        assert self._env is not None  # set by _ensure_env() above
-
-        actual_seed = self.episode_index if seed is None else seed
-        setup_kwargs = _load_robotwin_setup_kwargs(self.task_name)
-        setup_kwargs.update(seed=actual_seed, is_test=True)
-        with torch.enable_grad():
-            self._env.setup_demo(**setup_kwargs)
-        self.episode_index += self._reset_stride
-        self._step_count = 0
-
-        obs = self._get_obs()
-        return obs, {"is_success": False, "task": self.task_name}
-
-    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
-        assert self._env is not None, "step() called before reset()"
-        if action.ndim != 1 or action.shape[0] != ACTION_DIM:
-            raise ValueError(f"Expected 1-D action of shape ({ACTION_DIM},), got {action.shape}")
-
-        with torch.enable_grad():
-            if hasattr(self._env, "take_action"):
-                self._env.take_action(action)
-            else:
-                self._env.step(action)
-
-        self._step_count += 1
-
-        is_success = bool(getattr(self._env, "eval_success", False))
-        if not is_success and hasattr(self._env, "check_success"):
-            is_success = bool(self._env.check_success())
-
-        obs = self._get_obs()
-        reward = float(is_success)
-        terminated = is_success
-        truncated = self._step_count >= self.episode_length
-
-        info: dict[str, Any] = {
-            "task": self.task_name,
-            "is_success": is_success,
-            "step": self._step_count,
-        }
-        if terminated or truncated:
-            info["final_info"] = {
-                "task": self.task_name,
-                "is_success": is_success,
-            }
-            self.reset()
-
-        return obs, reward, terminated, truncated, info
-
-    def render(self) -> np.ndarray:
-        self._ensure_env()
-        obs = self._get_obs()
-        # Prefer head camera for rendering; fall back to first available.
-        if "head_camera" in obs["pixels"]:
-            return obs["pixels"]["head_camera"]
-        return next(iter(obs["pixels"].values()))
-
-    def close(self) -> None:
-        if self._env is not None:
-            if hasattr(self._env, "close_env"):
-                import contextlib
-
-                with contextlib.suppress(TypeError):
-                    self._env.close_env()
-            self._env = None
-
-
-# ---- Multi-task factory --------------------------------------------------------
-
-
-def _make_env_fns(
-    *,
-    task_name: str,
-    n_envs: int,
-    camera_names: list[str],
-    observation_height: int,
-    observation_width: int,
-    episode_length: int,
-) -> list[Callable[[], RoboTwinEnv]]:
-    """Return n_envs factory callables for a single task."""
-
-    def _make_one(episode_index: int) -> RoboTwinEnv:
-        return RoboTwinEnv(
-            task_name=task_name,
-            episode_index=episode_index,
-            n_envs=n_envs,
-            camera_names=camera_names,
-            observation_height=observation_height,
-            observation_width=observation_width,
-            episode_length=episode_length,
-        )
-
-    return [partial(_make_one, i) for i in range(n_envs)]
-
-
-def create_robotwin_envs(
-    task: str,
-    n_envs: int,
-    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
-    camera_names: Sequence[str] = ROBOTWIN_CAMERA_NAMES,
-    observation_height: int = DEFAULT_CAMERA_H,
-    observation_width: int = DEFAULT_CAMERA_W,
-    episode_length: int = DEFAULT_EPISODE_LENGTH,
-) -> dict[str, dict[int, Any]]:
-    """Create vectorized RoboTwin 2.0 environments.
-
-    Returns:
-        ``dict[task_name][0] -> VectorEnv`` — one entry per task, each wrapping
-        ``n_envs`` parallel rollouts.
-
-    Args:
-        task: Comma-separated list of task names (e.g. ``"beat_block_hammer"``
-            or ``"beat_block_hammer,click_bell"``).
-        n_envs: Number of parallel rollouts per task.
-        env_cls: Vector env constructor (e.g. ``gym.vector.AsyncVectorEnv``).
-        camera_names: Cameras to include in observations.
-        observation_height: Pixel height for all cameras.
-        observation_width: Pixel width for all cameras.
-        episode_length: Max steps before truncation.
-    """
-    if env_cls is None or not callable(env_cls):
-        raise ValueError("env_cls must be callable (e.g. gym.vector.AsyncVectorEnv).")
-    if not isinstance(n_envs, int) or n_envs <= 0:
-        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
-
-    task_names = [t.strip() for t in str(task).split(",") if t.strip()]
-    if not task_names:
-        raise ValueError("`task` must contain at least one RoboTwin task name.")
-
-    unknown = [t for t in task_names if t not in ROBOTWIN_TASKS]
-    if unknown:
-        raise ValueError(f"Unknown RoboTwin tasks: {unknown}. Available tasks: {sorted(ROBOTWIN_TASKS)}")
-
-    logger.info(
-        "Creating RoboTwin envs | tasks=%s | n_envs(per task)=%d",
-        task_names,
-        n_envs,
-    )
-
-    is_async = env_cls is gym.vector.AsyncVectorEnv
-    cached_obs_space: spaces.Space | None = None
-    cached_act_space: spaces.Space | None = None
-    cached_metadata: dict[str, Any] | None = None
-
-    out: dict[str, dict[int, Any]] = defaultdict(dict)
-    for task_name in task_names:
-        fns = _make_env_fns(
-            task_name=task_name,
-            n_envs=n_envs,
-            camera_names=list(camera_names),
-            observation_height=observation_height,
-            observation_width=observation_width,
-            episode_length=episode_length,
-        )
-        if is_async:
-            lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
-            if cached_obs_space is None:
-                cached_obs_space = lazy.observation_space
-                cached_act_space = lazy.action_space
-                cached_metadata = lazy.metadata
-            out[task_name][0] = lazy
-        else:
-            out[task_name][0] = env_cls(fns)
-        logger.info("Built vec env | task=%s | n_envs=%d", task_name, n_envs)
-
-    return {k: dict(v) for k, v in out.items()}

src/lerobot/envs/utils.py

@@ -34,25 +34,6 @@ from lerobot.utils.utils import get_channel_first_image_shape
 from .configs import EnvConfig
 
 
-def parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
-    """Normalize ``camera_name`` into a non-empty list of strings.
-
-    Accepts a comma-separated string (``"cam_a,cam_b"``) or a sequence of
-    strings (tuples/lists). Whitespace is stripped; empty entries are
-    dropped. Raises ``TypeError`` for unsupported input types and
-    ``ValueError`` when the normalized list is empty.
-    """
-    if isinstance(camera_name, str):
-        cams = [c.strip() for c in camera_name.split(",") if c.strip()]
-    elif isinstance(camera_name, (list | tuple)):
-        cams = [str(c).strip() for c in camera_name if str(c).strip()]
-    else:
-        raise TypeError(f"camera_name must be str or sequence[str], got {type(camera_name).__name__}")
-    if not cams:
-        raise ValueError("camera_name resolved to an empty list.")
-    return cams
-
-
 def _convert_nested_dict(d):
     result = {}
     for k, v in d.items():
@@ -172,20 +153,17 @@ class _LazyAsyncVectorEnv:
         env_fns: list[Callable],
         observation_space=None,
         action_space=None,
-        metadata=None,
     ):
         self._env_fns = env_fns
         self._env: gym.vector.AsyncVectorEnv | None = None
         self.num_envs = len(env_fns)
-        if observation_space is not None and action_space is not None and metadata is not None:
+        if observation_space is not None and action_space is not None:
             self.observation_space = observation_space
             self.action_space = action_space
-            self.metadata = metadata
         else:
             tmp = env_fns[0]()
             self.observation_space = tmp.observation_space
             self.action_space = tmp.action_space
-            self.metadata = tmp.metadata
             tmp.close()
         self.single_observation_space = self.observation_space
         self.single_action_space = self.action_space
@@ -194,10 +172,6 @@ class _LazyAsyncVectorEnv:
         if self._env is None:
             self._env = gym.vector.AsyncVectorEnv(self._env_fns, context="forkserver", shared_memory=True)
 
-    @property
-    def unwrapped(self):
-        return self
-
     def reset(self, **kwargs):
         self._ensure()
         return self._env.reset(**kwargs)

src/lerobot/envs/vlabench.py

@@ -1,589 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""VLABench environment wrapper for LeRobot.
-
-VLABench is a large-scale benchmark for language-conditioned robotic manipulation
-with long-horizon reasoning, built on MuJoCo/dm_control.
-
-- Paper: https://arxiv.org/abs/2412.18194
-- GitHub: https://github.com/OpenMOSS/VLABench
-- Website: https://vlabench.github.io
-"""
-
-from __future__ import annotations
-
-import contextlib
-import logging
-from collections import defaultdict
-from collections.abc import Callable, Sequence
-from typing import Any
-
-import cv2
-import gymnasium as gym
-import numpy as np
-from gymnasium import spaces
-from scipy.spatial.transform import Rotation
-
-from lerobot.types import RobotObservation
-
-from .utils import _LazyAsyncVectorEnv
-
-logger = logging.getLogger(__name__)
-
-ACTION_DIM = 7  # pos(3) + euler(3) + gripper(1)
-ACTION_LOW = np.array([-1.0, -1.0, -1.0, -1.0, -1.0, -1.0, 0.0], dtype=np.float32)
-ACTION_HIGH = np.array([1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0], dtype=np.float32)
-
-# Default max episode steps per task type
-DEFAULT_MAX_EPISODE_STEPS = 500
-
-# VLABench task suites
-PRIMITIVE_TASKS = [
-    "select_fruit",
-    "select_toy",
-    "select_chemistry_tube",
-    "add_condiment",
-    "select_book",
-    "select_painting",
-    "select_drink",
-    "insert_flower",
-    "select_billiards",
-    "select_ingredient",
-    "select_mahjong",
-    "select_poker",
-    # Physical series
-    "density_qa",
-    "friction_qa",
-    "magnetism_qa",
-    "reflection_qa",
-    "simple_cuestick_usage",
-    "simple_seesaw_usage",
-    "sound_speed_qa",
-    "thermal_expansion_qa",
-    "weight_qa",
-]
-
-COMPOSITE_TASKS = [
-    "cluster_billiards",
-    "cluster_book",
-    "cluster_drink",
-    "cluster_toy",
-    "cook_dishes",
-    "cool_drink",
-    "find_unseen_object",
-    "get_coffee",
-    "hammer_nail",
-    "heat_food",
-    "make_juice",
-    "play_mahjong",
-    "play_math_game",
-    "play_poker",
-    "play_snooker",
-    "rearrange_book",
-    "rearrange_chemistry_tube",
-    "set_dining_table",
-    "set_study_table",
-    "store_food",
-    "take_chemistry_experiment",
-    "use_seesaw_complex",
-]
-
-SUITE_TASKS: dict[str, list[str]] = {
-    "primitive": PRIMITIVE_TASKS,
-    "composite": COMPOSITE_TASKS,
-}
-
-
-class VLABenchEnv(gym.Env):
-    """Gymnasium wrapper for VLABench environments.
-
-    Wraps the dm_control-based VLABench simulator behind a standard gym.Env interface.
-    Supports multiple cameras (front, second, wrist) and end-effector control.
-    """
-
-    metadata = {"render_modes": ["rgb_array"], "render_fps": 10}
-
-    def __init__(
-        self,
-        task: str = "select_fruit",
-        obs_type: str = "pixels_agent_pos",
-        render_mode: str = "rgb_array",
-        render_resolution: tuple[int, int] = (480, 480),
-        robot: str = "franka",
-        max_episode_steps: int = DEFAULT_MAX_EPISODE_STEPS,
-        action_mode: str = "eef",
-    ):
-        super().__init__()
-        self.task = task
-        self.obs_type = obs_type
-        self.render_mode = render_mode
-        self.render_resolution = render_resolution
-        self.robot = robot
-        self._max_episode_steps = max_episode_steps
-        self.action_mode = action_mode
-
-        # Deferred — created on first reset() inside worker subprocess to avoid
-        # inheriting stale GPU/EGL contexts when AsyncVectorEnv spawns workers.
-        # We never cache `env.physics`: dm_control exposes it as a weakref
-        # proxy that goes stale across resets (rebuilds the sim), so we always
-        # refetch it via `self._env.physics` at the call site.
-        self._env = None
-        self.task_description = ""  # populated on first reset
-        # Cached world-frame XYZ of the robot base link. The VLABench datasets
-        # log both `observation.state` positions and `actions` positions in
-        # robot-base frame (see VLABench/scripts/convert_to_lerobot.py which
-        # subtracts `robot_frame_pos` from ee_pos). The robot is attached at a
-        # fixed offset per task so this is safe to cache once per env build.
-        self._robot_base_xyz: np.ndarray | None = None
-
-        h, w = self.render_resolution
-
-        if self.obs_type == "state":
-            raise NotImplementedError(
-                "The 'state' observation type is not supported in VLABenchEnv. "
-                "Please use 'pixels' or 'pixels_agent_pos'."
-            )
-        elif self.obs_type == "pixels":
-            self.observation_space = spaces.Dict(
-                {
-                    "pixels": spaces.Dict(
-                        {
-                            "image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
-                            "second_image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
-                            "wrist_image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
-                        }
-                    ),
-                }
-            )
-        elif self.obs_type == "pixels_agent_pos":
-            self.observation_space = spaces.Dict(
-                {
-                    "pixels": spaces.Dict(
-                        {
-                            "image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
-                            "second_image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
-                            "wrist_image": spaces.Box(low=0, high=255, shape=(h, w, 3), dtype=np.uint8),
-                        }
-                    ),
-                    "agent_pos": spaces.Box(low=-np.inf, high=np.inf, shape=(7,), dtype=np.float64),
-                }
-            )
-        else:
-            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
-
-        self.action_space = spaces.Box(low=ACTION_LOW, high=ACTION_HIGH, dtype=np.float32)
-
-    # Max attempts to rebuild the underlying env when MuJoCo throws
-    # `PhysicsError` (e.g. mjWARN_BADQACC) during VLABench's 20-step
-    # reset warm-up. Some random task/layout samples land in unstable
-    # initial configurations; re-sampling the layout almost always
-    # gives a stable one. A handful of upstream tasks (notably
-    # `select_mahjong`) have layout samplers that diverge often enough
-    # to need >>5 retries, so we pick a generous ceiling.
-    _ENSURE_ENV_MAX_ATTEMPTS = 20
-
-    def _ensure_env(self) -> None:
-        """Create the underlying VLABench env on first use.
-
-        Called inside the worker subprocess after fork(), so each worker gets
-        its own clean rendering context rather than inheriting a stale one from
-        the parent process (which causes crashes with AsyncVectorEnv).
-
-        Retries on `PhysicsError`: VLABench's `LM4ManipDMEnv.reset()` runs 20
-        warm-up `step()` calls while toggling gravity/fluids to let the scene
-        settle; for some random layouts MuJoCo's integrator diverges and
-        raises `mjWARN_BADQACC`. Re-sampling the layout almost always yields
-        a stable one, so we retry a number of times before giving up. Between
-        attempts we reseed NumPy's global RNG from OS entropy so the upstream
-        task sampler explores fresh initial states — without this, retries
-        can replay the same diverging configuration when the sampler is
-        deterministic given the current RNG state.
-        """
-        if self._env is not None:
-            return
-
-        import VLABench.robots  # noqa: F401  # type: ignore[import-untyped]
-        import VLABench.tasks  # noqa: F401  # type: ignore[import-untyped]
-        from dm_control.rl.control import PhysicsError  # type: ignore[import-untyped]
-        from VLABench.envs import load_env  # type: ignore[import-untyped]
-
-        h, w = self.render_resolution
-        last_exc: PhysicsError | None = None
-        for attempt in range(1, self._ENSURE_ENV_MAX_ATTEMPTS + 1):
-            try:
-                env = load_env(task=self.task, robot=self.robot, render_resolution=(h, w))
-                self._env = env
-                break
-            except PhysicsError as exc:
-                last_exc = exc
-                logger.warning(
-                    "PhysicsError on attempt %d/%d while building task '%s': %s. Retrying with fresh layout…",
-                    attempt,
-                    self._ENSURE_ENV_MAX_ATTEMPTS,
-                    self.task,
-                    exc,
-                )
-                np.random.seed(None)
-        if self._env is None:
-            assert last_exc is not None
-            raise RuntimeError(
-                f"VLABench task '{self.task}' failed to produce a stable "
-                f"initial layout after {self._ENSURE_ENV_MAX_ATTEMPTS} "
-                f"attempts. This task's upstream sampler diverges too "
-                f"often for the configured robot; consider removing it "
-                f"from the eval set. Last physics error: {last_exc}"
-            ) from last_exc
-
-        # Extract task description from the dm_control task
-        task_obj = self._env.task
-        if hasattr(task_obj, "task_description"):
-            self.task_description = task_obj.task_description
-        elif hasattr(task_obj, "language_instruction"):
-            self.task_description = task_obj.language_instruction
-        else:
-            self.task_description = self.task
-
-        # Cache robot base world position so `_build_ctrl_from_action` and
-        # `_get_obs` can translate between robot-frame (dataset) and
-        # world-frame (dm_control) without hitting physics every call.
-        try:
-            self._robot_base_xyz = np.asarray(self._env.get_robot_frame_position(), dtype=np.float64).reshape(
-                3
-            )
-        except Exception:
-            # Fallback to VLABench's default Franka base position.
-            self._robot_base_xyz = np.array([0.0, -0.4, 0.78], dtype=np.float64)
-
-    def _get_obs(self) -> dict:
-        """Get current observation from the environment."""
-        assert self._env is not None
-
-        obs = self._env.get_observation()
-        h, w = self.render_resolution
-
-        def _to_hwc3(arr: np.ndarray) -> np.ndarray:
-            """Coerce any camera array to the declared (h, w, 3) uint8 shape."""
-            a = np.asarray(arr)
-            # Drop a leading singleton batch dim if present.
-            while a.ndim > 3 and a.shape[0] == 1:
-                a = a[0]
-            if a.ndim == 3 and a.shape[0] in (1, 3, 4) and a.shape[-1] not in (1, 3, 4):
-                # CHW → HWC
-                a = np.transpose(a, (1, 2, 0))
-            if a.ndim == 2:
-                a = np.stack([a] * 3, axis=-1)
-            if a.ndim != 3:
-                return np.zeros((h, w, 3), dtype=np.uint8)
-            # Force 3 channels.
-            if a.shape[-1] == 1:
-                a = np.repeat(a, 3, axis=-1)
-            elif a.shape[-1] == 4:
-                a = a[..., :3]
-            elif a.shape[-1] != 3:
-                return np.zeros((h, w, 3), dtype=np.uint8)
-            if a.shape[:2] != (h, w):
-                a = cv2.resize(a, (w, h), interpolation=cv2.INTER_AREA)
-            return a.astype(np.uint8)
-
-        # Extract camera images — VLABench returns (n_cameras, C, H, W) or individual arrays
-        raw_frames: list[np.ndarray] = []
-        if "rgb" in obs:
-            rgb = obs["rgb"]
-            if isinstance(rgb, np.ndarray):
-                if rgb.ndim == 4:
-                    raw_frames = [rgb[i] for i in range(rgb.shape[0])]
-                elif rgb.ndim == 3:
-                    raw_frames = [rgb]
-
-        image_keys = ["image", "second_image", "wrist_image"]
-        images: dict[str, np.ndarray] = {}
-        for i, key in enumerate(image_keys):
-            if i < len(raw_frames):
-                images[key] = _to_hwc3(raw_frames[i])
-            else:
-                images[key] = np.zeros((h, w, 3), dtype=np.uint8)
-
-        # Convert VLABench's raw ee_state `[pos_world(3), quat_wxyz(4), open(1)]`
-        # to the dataset's observation.state layout `[pos_robot(3), euler_xyz(3),
-        # gripper(1)]`. See VLABench/scripts/convert_to_lerobot.py — positions
-        # are stored in robot-base frame and orientations as scipy extrinsic
-        # 'xyz' euler angles.
-        raw = np.asarray(obs.get("ee_state", np.zeros(8)), dtype=np.float64).ravel()
-        pos_world = raw[:3] if raw.size >= 3 else np.zeros(3, dtype=np.float64)
-        quat_wxyz = raw[3:7] if raw.size >= 7 else np.array([1.0, 0.0, 0.0, 0.0], dtype=np.float64)
-        gripper = float(raw[7]) if raw.size >= 8 else 0.0
-
-        base = self._robot_base_xyz if self._robot_base_xyz is not None else np.zeros(3, dtype=np.float64)
-        pos_robot = pos_world - base
-        euler_xyz = Rotation.from_quat([quat_wxyz[1], quat_wxyz[2], quat_wxyz[3], quat_wxyz[0]]).as_euler(
-            "xyz", degrees=False
-        )
-
-        ee_state = np.concatenate([pos_robot, euler_xyz, [gripper]]).astype(np.float64)
-
-        if self.obs_type == "pixels":
-            return {"pixels": images}
-        elif self.obs_type == "pixels_agent_pos":
-            return {
-                "pixels": images,
-                "agent_pos": ee_state.astype(np.float64),
-            }
-        else:
-            raise ValueError(f"Unknown obs_type: {self.obs_type}")
-
-    # ---- Action adaptation (EEF → joint ctrl) --------------------------------
-    #
-    # The HF vlabench datasets log 7D actions
-    # `[x, y, z (robot frame), rx, ry, rz (scipy extrinsic xyz), gripper]`,
-    # exactly matching VLABench's own eval pipeline (evaluator.base):
-    #   pos, euler, g = policy(...)
-    #   quat = euler_to_quaternion(*euler)      # extrinsic xyz -> wxyz
-    #   _, qpos = robot.get_qpos_from_ee_pos(physics, pos=pos + base, quat=quat)
-    #   env.step(np.concatenate([qpos, [g, g]]))
-    #
-    # VLABench's dm_control task writes `data.ctrl[:] = action` directly — for
-    # Franka that's 9 entries (7 arm joints + 2 gripper fingers). We mirror the
-    # above conversion so the policy's EEF commands actually drive the robot.
-
-    _FRANKA_FINGER_OPEN = 0.04  # qpos when gripper fully open
-
-    def _build_ctrl_from_action(self, action: np.ndarray, ctrl_dim: int) -> np.ndarray:
-        """Convert a 7D EEF action into the `ctrl_dim`-sized joint command vector.
-
-        For the Franka default (ctrl_dim=9): 7 arm joint qposes (via IK) +
-        2 gripper finger qposes (open/closed based on the gripper scalar).
-        If the action is already joint-space (shape matches ctrl_dim), pass
-        through.
-        """
-        if action.shape[0] == ctrl_dim:
-            return action.astype(np.float64, copy=False)
-
-        if action.shape[0] != 7:
-            # Unknown layout — fall back to zero-pad so the sim doesn't crash.
-            padded = np.zeros(ctrl_dim, dtype=np.float64)
-            padded[: min(action.shape[0], ctrl_dim)] = action[:ctrl_dim]
-            return padded
-
-        from dm_control.utils.inverse_kinematics import qpos_from_site_pose
-
-        # Action position is in robot-base frame (see convert_to_lerobot.py);
-        # dm_control's IK expects a world-frame target.
-        base = self._robot_base_xyz if self._robot_base_xyz is not None else np.zeros(3, dtype=np.float64)
-        pos_world = np.asarray(action[:3], dtype=np.float64) + base
-        rx, ry, rz = float(action[3]), float(action[4]), float(action[5])
-        gripper = float(np.clip(action[6], 0.0, 1.0))
-
-        # Dataset euler is scipy extrinsic 'xyz' (same as VLABench's
-        # `euler_to_quaternion`). scipy emits `[x, y, z, w]`; dm_control's IK
-        # and MuJoCo use `[w, x, y, z]`, so reorder.
-        qxyzw = Rotation.from_euler("xyz", [rx, ry, rz], degrees=False).as_quat()
-        quat = np.array([qxyzw[3], qxyzw[0], qxyzw[1], qxyzw[2]], dtype=np.float64)
-
-        assert self._env is not None
-        robot = self._env.task.robot
-        site_name = robot.end_effector_site.full_identifier
-
-        # inplace=False so IK doesn't mutate physics state mid-step — we only
-        # want the solved qpos. Fetch a fresh physics handle — caching it can
-        # yield a stale weakref after a reset.
-        ik_result = qpos_from_site_pose(
-            self._env.physics,
-            site_name=site_name,
-            target_pos=pos_world,
-            target_quat=quat,
-            inplace=False,
-            max_steps=100,
-        )
-        n_dof = robot.n_dof  # 7 for Franka
-        arm_qpos = ik_result.qpos[:n_dof]
-
-        # Dataset gripper convention: 1 = open (finger qpos = 0.04),
-        # 0 = closed (finger qpos = 0.0). See VLABench/scripts/convert_to_lerobot.py
-        # where `trajectory[i][-1] > 0.03` is encoded as `1`.
-        finger_qpos = gripper * self._FRANKA_FINGER_OPEN
-
-        ctrl = np.zeros(ctrl_dim, dtype=np.float64)
-        ctrl[:n_dof] = arm_qpos
-        # Remaining entries are gripper fingers (usually 2 for Franka).
-        ctrl[n_dof:] = finger_qpos
-        return ctrl
-
-    def reset(self, seed=None, **kwargs) -> tuple[RobotObservation, dict[str, Any]]:
-        self._ensure_env()
-        assert self._env is not None
-        super().reset(seed=seed)
-
-        if seed is not None:
-            self._seed_inner_env(int(self.np_random.integers(0, 2**31 - 1)))
-
-        self._env.reset()
-
-        observation = self._get_obs()
-        info = {"is_success": False}
-        return observation, info
-
-    def _seed_inner_env(self, seed: int) -> None:
-        """Propagate `seed` to the inner dm_control env. `Environment.reset()`
-        doesn't accept a seed, so we re-seed the task and environment
-        `RandomState`s directly. Best-effort: silently skipped when the
-        expected attributes are absent on a given VLABench version.
-        """
-        for owner_attr, rng_attr in (("task", "random"), (None, "_random_state")):
-            owner = getattr(self._env, owner_attr) if owner_attr else self._env
-            rng = getattr(owner, rng_attr, None)
-            rng_seed = getattr(rng, "seed", None)
-            if callable(rng_seed):
-                rng_seed(seed)
-
-    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
-        from dm_control.rl.control import PhysicsError  # type: ignore[import-untyped]
-
-        self._ensure_env()
-        assert self._env is not None
-
-        if action.ndim != 1:
-            raise ValueError(
-                f"Expected action to be 1-D (shape (action_dim,)), "
-                f"but got shape {action.shape} with ndim={action.ndim}"
-            )
-
-        if self.action_mode not in ("eef", "joint", "delta_eef"):
-            raise ValueError(f"Unknown action_mode: {self.action_mode}")
-
-        # Always refetch physics — dm_control returns a weakref proxy that can
-        # go stale across resets.
-        physics = self._env.physics
-        ctrl_dim = int(physics.data.ctrl.shape[0])
-        ctrl = self._build_ctrl_from_action(action, ctrl_dim)
-        try:
-            timestep = self._env.step(ctrl)
-        except PhysicsError as exc:
-            # Physics integrator diverged (e.g. mjWARN_BADQACC). Treat it as
-            # a graceful failed termination rather than a hard crash — the
-            # rest of the multi-task eval should still run.
-            logger.warning(
-                "PhysicsError during step on task '%s': %s. Terminating episode.",
-                self.task,
-                exc,
-            )
-            observation = self._get_obs()
-            info = {"task": self.task, "is_success": False, "physics_error": True}
-            # Drop the stale env so the next reset() rebuilds it cleanly.
-            with contextlib.suppress(Exception):
-                self._env.close()
-            self._env = None
-            return observation, 0.0, True, False, info
-
-        # Extract reward from dm_control timestep
-        reward = float(timestep.reward) if timestep.reward is not None else 0.0
-
-        # Check success via the task's termination condition
-        is_success = False
-        if hasattr(self._env, "task") and hasattr(self._env.task, "should_terminate_episode"):
-            is_success = bool(self._env.task.should_terminate_episode(self._env.physics))
-
-        terminated = is_success
-        truncated = False
-        info = {
-            "task": self.task,
-            "is_success": is_success,
-        }
-
-        observation = self._get_obs()
-
-        if terminated:
-            self.reset()
-
-        return observation, reward, terminated, truncated, info
-
-    def render(self) -> np.ndarray:
-        self._ensure_env()
-        obs = self._get_obs()
-        return obs["pixels"]["image"]
-
-    def close(self):
-        if self._env is not None:
-            self._env.close()
-            self._env = None
-
-
-# ---- Main API ----------------------------------------------------------------
-
-
-def create_vlabench_envs(
-    task: str,
-    n_envs: int,
-    gym_kwargs: dict[str, Any] | None = None,
-    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
-) -> dict[str, dict[int, Any]]:
-    """
-    Create vectorized VLABench environments with a consistent return shape.
-
-    Returns:
-        dict[suite_name][task_id] -> vec_env (env_cls([...]) with exactly n_envs factories)
-
-    Notes:
-        - n_envs is the number of rollouts *per task*.
-        - `task` can be a suite name ("primitive", "composite"), a comma-separated list of
-          suite names, or individual task names (e.g. "select_fruit,heat_food").
-    """
-    if env_cls is None or not callable(env_cls):
-        raise ValueError("env_cls must be a callable that wraps a list of environment factory callables.")
-    if not isinstance(n_envs, int) or n_envs <= 0:
-        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
-
-    gym_kwargs = dict(gym_kwargs or {})
-    task_groups = [t.strip() for t in task.split(",") if t.strip()]
-    if not task_groups:
-        raise ValueError("`task` must contain at least one VLABench task or suite name.")
-
-    logger.info(
-        "Creating VLABench envs | task_groups=%s | n_envs(per task)=%d",
-        task_groups,
-        n_envs,
-    )
-
-    is_async = env_cls is gym.vector.AsyncVectorEnv
-    cached_obs_space = None
-    cached_act_space = None
-    cached_metadata = None
-    out: dict[str, dict[int, Any]] = defaultdict(dict)
-
-    for group in task_groups:
-        # Check if it's a suite name, otherwise treat as individual task
-        tasks = SUITE_TASKS.get(group, [group])
-
-        for tid, task_name in enumerate(tasks):
-            logger.info(
-                "Building vec env | group=%s | task_id=%d | task=%s",
-                group,
-                tid,
-                task_name,
-            )
-
-            fns = [(lambda tn=task_name: VLABenchEnv(task=tn, **gym_kwargs)) for _ in range(n_envs)]
-
-            if is_async:
-                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
-                if cached_obs_space is None:
-                    cached_obs_space = lazy.observation_space
-                    cached_act_space = lazy.action_space
-                    cached_metadata = lazy.metadata
-                out[group][tid] = lazy
-            else:
-                out[group][tid] = env_cls(fns)
-
-    return {group: dict(task_map) for group, task_map in out.items()}

src/lerobot/model/kinematics.py

@@ -12,19 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
 import numpy as np
 
-from lerobot.utils.import_utils import _placo_available, require_package
-
-if TYPE_CHECKING or _placo_available:
-    import placo  # type: ignore[import-not-found]
-else:
-    placo = None
-
 
 class RobotKinematics:
     """Robot kinematics using placo library for forward and inverse kinematics."""
@@ -43,7 +32,13 @@ class RobotKinematics:
             target_frame_name (str): Name of the end-effector frame in the URDF
             joint_names (list[str] | None): List of joint names to use for the kinematics solver
         """
-        require_package("placo", extra="placo-dep")
+        try:
+            import placo  # type: ignore[import-not-found] # C++ library with Python bindings, no type stubs available. TODO: Create stub file or request upstream typing support.
+        except ImportError as e:
+            raise ImportError(
+                "placo is required for RobotKinematics. "
+                "Please install the optional dependencies of `kinematics` in the package."
+            ) from e
 
         self.robot = placo.RobotWrapper(urdf_path)
         self.solver = placo.KinematicsSolver(self.robot)

src/lerobot/motors/damiao/damiao.py

@@ -24,7 +24,7 @@ from functools import cached_property
 from typing import TYPE_CHECKING, Any, TypedDict
 
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.import_utils import _can_available, require_package
+from lerobot.utils.import_utils import _can_available
 
 if TYPE_CHECKING or _can_available:
     import can
@@ -111,7 +111,6 @@ class DamiaoMotorsBus(MotorsBusBase):
             bitrate: Nominal bitrate in bps (default: 1000000 = 1 Mbps)
             data_bitrate: Data bitrate for CAN FD in bps (default: 5000000 = 5 Mbps), ignored if use_can_fd is False
         """
-        require_package("python-can", extra="damiao", import_name="can")
         super().__init__(port, motors, calibration)
         self.port = port
         self.can_interface = can_interface

src/lerobot/motors/motors_bus.py

@@ -356,8 +356,8 @@ class SerialMotorsBus(MotorsBusBase):
         motors: dict[str, Motor],
         calibration: dict[str, MotorCalibration] | None = None,
     ):
-        require_package("pyserial", extra="pyserial-dep", import_name="serial")
-        require_package("deepdiff", extra="deepdiff-dep")
+        require_package("pyserial", extra="hardware", import_name="serial")
+        require_package("deepdiff", extra="hardware")
         super().__init__(port, motors, calibration)
 
         self.port_handler: PortHandler

src/lerobot/motors/robstride/robstride.py

@@ -23,12 +23,12 @@ from types import SimpleNamespace
 from typing import TYPE_CHECKING, Any, TypedDict
 
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.import_utils import _can_available, require_package
+from lerobot.utils.import_utils import _can_available
 
 if TYPE_CHECKING or _can_available:
     import can
 else:
-    can = SimpleNamespace(Message=object, interface=None, BusABC=object)
+    can = SimpleNamespace(Message=object, interface=None)
 import numpy as np
 
 from lerobot.utils.errors import DeviceNotConnectedError
@@ -106,7 +106,6 @@ class RobstrideMotorsBus(MotorsBusBase):
             bitrate: Nominal bitrate in bps (default: 1000000 = 1 Mbps)
             data_bitrate: Data bitrate for CAN FD in bps (default: 5000000 = 5 Mbps), ignored if use_can_fd is False
         """
-        require_package("python-can", extra="robstride", import_name="can")
         super().__init__(port, motors, calibration)
         self.port = port
         self.can_interface = can_interface

src/lerobot/optim/schedulers.py

@@ -18,21 +18,14 @@ import logging
 import math
 from dataclasses import asdict, dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING
 
 import draccus
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LambdaLR, LRScheduler
 
 from lerobot.utils.constants import SCHEDULER_STATE
-from lerobot.utils.import_utils import _diffusers_available, require_package
 from lerobot.utils.io_utils import deserialize_json_into_object, write_json
 
-if TYPE_CHECKING or _diffusers_available:
-    from diffusers.optimization import get_scheduler
-else:
-    get_scheduler = None
-
 
 @dataclass
 class LRSchedulerConfig(draccus.ChoiceRegistry, abc.ABC):
@@ -54,7 +47,10 @@ class DiffuserSchedulerConfig(LRSchedulerConfig):
     num_warmup_steps: int | None = None
 
     def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
+        from lerobot.utils.import_utils import require_package
+
         require_package("diffusers", extra="diffusion")
+        from diffusers.optimization import get_scheduler
 
         kwargs = {**asdict(self), "num_training_steps": num_training_steps, "optimizer": optimizer}
         return get_scheduler(**kwargs)

src/lerobot/policies/__init__.py

@@ -15,6 +15,10 @@
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .factory import get_policy_class, make_policy, make_policy_config, make_pre_post_processors
+from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig as GaussianActorConfig
+from .gaussian_actor.reward_model.configuration_classifier import (
+    RewardClassifierConfig as RewardClassifierConfig,
+)
 from .groot.configuration_groot import GrootConfig as GrootConfig
 from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig as MultiTaskDiTConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
@@ -22,8 +26,6 @@ from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
 from .pi05.configuration_pi05 import PI05Config as PI05Config
 from .pretrained import PreTrainedPolicy as PreTrainedPolicy
 from .rtc import ActionInterpolator as ActionInterpolator
-from .sac.configuration_sac import SACConfig as SACConfig
-from .sac.reward_model.configuration_classifier import RewardClassifierConfig as RewardClassifierConfig
 from .sarm.configuration_sarm import SARMConfig as SARMConfig
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
@@ -32,21 +34,21 @@ from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
 from .wall_x.configuration_wall_x import WallXConfig as WallXConfig
 from .xvla.configuration_xvla import XVLAConfig as XVLAConfig
 
-# NOTE: Policy modeling classes (e.g., SACPolicy) are intentionally NOT re-exported here.
+# NOTE: Policy modeling classes (e.g., GaussianActorPolicy) are intentionally NOT re-exported here.
 # They have heavy optional dependencies and are loaded lazily via get_policy_class().
-# Import directly: ``from lerobot.policies.sac.modeling_sac import SACPolicy``
+# Import directly: ``from lerobot.policies.gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy``
 
 __all__ = [
     # Configuration classes
     "ACTConfig",
     "DiffusionConfig",
+    "GaussianActorConfig",
     "GrootConfig",
     "MultiTaskDiTConfig",
     "PI0Config",
     "PI0FastConfig",
     "PI05Config",
     "RewardClassifierConfig",
-    "SACConfig",
     "SARMConfig",
     "SmolVLAConfig",
     "TDMPCConfig",

src/lerobot/policies/act/modeling_act.py

@@ -142,10 +142,9 @@ class ACTPolicy(PreTrainedPolicy):
 
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
-        abs_err = F.l1_loss(batch[ACTION], actions_hat, reduction="none")
-        valid_mask = ~batch["action_is_pad"].unsqueeze(-1)
-        num_valid = valid_mask.sum() * abs_err.shape[-1]
-        l1_loss = (abs_err * valid_mask).sum() / num_valid.clamp_min(1)
+        l1_loss = (
+            F.l1_loss(batch[ACTION], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
+        ).mean()
 
         loss_dict = {"l1_loss": l1_loss.item()}
         if self.config.use_vae:

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -23,7 +23,6 @@ TODO(alexander-soare):
 import math
 from collections import deque
 from collections.abc import Callable
-from typing import TYPE_CHECKING
 
 import einops
 import numpy as np
@@ -33,14 +32,6 @@ import torchvision
 from torch import Tensor, nn
 
 from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
-from lerobot.utils.import_utils import _diffusers_available, require_package
-
-if TYPE_CHECKING or _diffusers_available:
-    from diffusers.schedulers.scheduling_ddim import DDIMScheduler
-    from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
-else:
-    DDIMScheduler = None
-    DDPMScheduler = None
 
 from ..pretrained import PreTrainedPolicy
 from ..utils import (
@@ -73,7 +64,6 @@ class DiffusionPolicy(PreTrainedPolicy):
             dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
                 that they will be passed with a call to `load_state_dict` before the policy is used.
         """
-        require_package("diffusers", extra="diffusion")
         super().__init__(config)
         config.validate_features()
         self.config = config
@@ -165,7 +155,11 @@ def _make_noise_scheduler(name: str, **kwargs: dict):
     Factory for noise scheduler instances of the requested type. All kwargs are passed
     to the scheduler.
     """
+    from lerobot.utils.import_utils import require_package
+
     require_package("diffusers", extra="diffusion")
+    from diffusers.schedulers.scheduling_ddim import DDIMScheduler
+    from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
 
     if name == "DDPM":
         return DDPMScheduler(**kwargs)
@@ -380,9 +374,7 @@ class DiffusionModel(nn.Module):
                     f"{self.config.do_mask_loss_for_padding=}."
                 )
             in_episode_bound = ~batch["action_is_pad"]
-            mask = in_episode_bound.unsqueeze(-1)
-            num_valid = mask.sum() * loss.shape[-1]
-            return (loss * mask).sum() / num_valid.clamp_min(1)
+            loss = loss * in_episode_bound.unsqueeze(-1)
 
         return loss.mean()
 

src/lerobot/policies/factory.py

@@ -46,13 +46,13 @@ from lerobot.utils.feature_utils import dataset_to_policy_features
 
 from .act.configuration_act import ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig
+from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig
+from .gaussian_actor.reward_model.configuration_classifier import RewardClassifierConfig
 from .groot.configuration_groot import GrootConfig
 from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig
 from .pi0.configuration_pi0 import PI0Config
 from .pi05.configuration_pi05 import PI05Config
 from .pretrained import PreTrainedPolicy
-from .sac.configuration_sac import SACConfig
-from .sac.reward_model.configuration_classifier import RewardClassifierConfig
 from .sarm.configuration_sarm import SARMConfig
 from .smolvla.configuration_smolvla import SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig
@@ -89,7 +89,7 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
 
     Args:
         name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
-            "multi_task_dit", "vqbet", "pi0", "pi05", "sac", "reward_classifier", "smolvla", "wall_x".
+            "multi_task_dit", "vqbet", "pi0", "pi05", "gaussian_actor", "reward_classifier", "smolvla", "wall_x".
     Returns:
         The policy class corresponding to the given name.
 
@@ -128,12 +128,12 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from .pi05.modeling_pi05 import PI05Policy
 
         return PI05Policy
-    elif name == "sac":
-        from .sac.modeling_sac import SACPolicy
+    elif name == "gaussian_actor":
+        from .gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy
 
-        return SACPolicy
+        return GaussianActorPolicy
     elif name == "reward_classifier":
-        from .sac.reward_model.modeling_classifier import Classifier
+        from .gaussian_actor.reward_model.modeling_classifier import Classifier
 
         return Classifier
     elif name == "smolvla":
@@ -172,7 +172,7 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
 
     Args:
         policy_type: The type of the policy. Supported types include "tdmpc",
-                     "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "sac",
+                     "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "gaussian_actor",
                      "smolvla", "reward_classifier", "wall_x".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
@@ -196,8 +196,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return PI0Config(**kwargs)
     elif policy_type == "pi05":
         return PI05Config(**kwargs)
-    elif policy_type == "sac":
-        return SACConfig(**kwargs)
+    elif policy_type == "gaussian_actor":
+        return GaussianActorConfig(**kwargs)
     elif policy_type == "smolvla":
         return SmolVLAConfig(**kwargs)
     elif policy_type == "reward_classifier":
@@ -370,16 +370,16 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
-    elif isinstance(policy_cfg, SACConfig):
-        from .sac.processor_sac import make_sac_pre_post_processors
+    elif isinstance(policy_cfg, GaussianActorConfig):
+        from .gaussian_actor.processor_gaussian_actor import make_gaussian_actor_pre_post_processors
 
-        processors = make_sac_pre_post_processors(
+        processors = make_gaussian_actor_pre_post_processors(
             config=policy_cfg,
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
     elif isinstance(policy_cfg, RewardClassifierConfig):
-        from .sac.reward_model.processor_classifier import make_classifier_processor
+        from .gaussian_actor.reward_model.processor_classifier import make_classifier_processor
 
         processors = make_classifier_processor(
             config=policy_cfg,

src/lerobot/policies/gaussian_actor/__init__.py

@@ -12,8 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .configuration_sac import SACConfig
-from .modeling_sac import SACPolicy
-from .processor_sac import make_sac_pre_post_processors
+from .configuration_gaussian_actor import GaussianActorConfig
+from .modeling_gaussian_actor import GaussianActorPolicy
+from .processor_gaussian_actor import make_gaussian_actor_pre_post_processors
 
-__all__ = ["SACConfig", "SACPolicy", "make_sac_pre_post_processors"]
+__all__ = ["GaussianActorConfig", "GaussianActorPolicy", "make_gaussian_actor_pre_post_processors"]

src/lerobot/policies/gaussian_actor/configuration_gaussian_actor.py

@@ -75,18 +75,19 @@ class PolicyConfig:
     init_final: float = 0.05
 
 
-@PreTrainedConfig.register_subclass("sac")
+@PreTrainedConfig.register_subclass("gaussian_actor")
 @dataclass
-class SACConfig(PreTrainedConfig):
-    """Soft Actor-Critic (SAC) configuration.
+class GaussianActorConfig(PreTrainedConfig):
+    """Gaussian actor configuration.
 
-    SAC is an off-policy actor-critic deep RL algorithm based on the maximum entropy
-    reinforcement learning framework. It learns a policy and a Q-function simultaneously
-    using experience collected from the environment.
+    This configures the policy-side (actor + observation encoder) of a Gaussian
+    policy, as used by SAC and related maximum-entropy continuous-control algorithms.
+    By default the actor output is a tanh-squashed diagonal Gaussian
+    (``TanhMultivariateNormalDiag``); the tanh squashing can be disabled via
+    ``policy_kwargs.use_tanh_squash``. The critics, temperature, and Bellman-update
+    logic live on the algorithm side (see ``lerobot.rl.algorithms.sac``).
 
-    This configuration class contains all the parameters needed to define a SAC agent,
-    including network architectures, optimization settings, and algorithm-specific
-    hyperparameters.
+    CLI: ``--policy.type=gaussian_actor``.
     """
 
     # Mapping of feature types to normalization modes
@@ -122,7 +123,7 @@ class SACConfig(PreTrainedConfig):
     device: str = "cpu"
     # Device to store the model on
     storage_device: str = "cpu"
-    # Name of the vision encoder model (Set to "helper2424/resnet10" for hil serl resnet10)
+    # Name of the vision encoder model (Set to "lerobot/resnet10" for hil serl resnet10)
     vision_encoder_name: str | None = None
     # Whether to freeze the vision encoder during training
     freeze_vision_encoder: bool = True
@@ -135,78 +136,41 @@ class SACConfig(PreTrainedConfig):
     # Dimension of the image embedding pooling
     image_embedding_pooling_dim: int = 8
 
-    # Training parameter
-    # Number of steps for online training
-    online_steps: int = 1000000
-    # Capacity of the online replay buffer
-    online_buffer_capacity: int = 100000
-    # Capacity of the offline replay buffer
-    offline_buffer_capacity: int = 100000
-    # Whether to use asynchronous prefetching for the buffers
-    async_prefetch: bool = False
-    # Number of steps before learning starts
-    online_step_before_learning: int = 100
-    # Frequency of policy updates
-    policy_update_freq: int = 1
-
-    # SAC algorithm parameters
-    # Discount factor for the SAC algorithm
-    discount: float = 0.99
-    # Initial temperature value
-    temperature_init: float = 1.0
-    # Number of critics in the ensemble
-    num_critics: int = 2
-    # Number of subsampled critics for training
-    num_subsample_critics: int | None = None
-    # Learning rate for the critic network
-    critic_lr: float = 3e-4
-    # Learning rate for the actor network
-    actor_lr: float = 3e-4
-    # Learning rate for the temperature parameter
-    temperature_lr: float = 3e-4
-    # Weight for the critic target update
-    critic_target_update_weight: float = 0.005
-    # Update-to-data ratio for the UTD algorithm (If you want enable utd_ratio, you need to set it to >1)
-    utd_ratio: int = 1
+    # Encoder architecture
     # Hidden dimension size for the state encoder
     state_encoder_hidden_dim: int = 256
     # Dimension of the latent space
     latent_dim: int = 256
-    # Target entropy for the SAC algorithm
-    target_entropy: float | None = None
-    # Whether to use backup entropy for the SAC algorithm
-    use_backup_entropy: bool = True
-    # Gradient clipping norm for the SAC algorithm
-    grad_clip_norm: float = 40.0
 
-    # Network configuration
-    # Configuration for the critic network architecture
-    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    # Configuration for the actor network architecture
-    actor_network_kwargs: ActorNetworkConfig = field(default_factory=ActorNetworkConfig)
-    # Configuration for the policy parameters
-    policy_kwargs: PolicyConfig = field(default_factory=PolicyConfig)
-    # Configuration for the discrete critic network
-    discrete_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    # Configuration for actor-learner architecture
+    # Online training (TODO(Khalil): relocate to TrainRLServerPipelineConfig)
+    online_steps: int = 1000000
+    online_buffer_capacity: int = 100000
+    offline_buffer_capacity: int = 100000
+    async_prefetch: bool = False
+    online_step_before_learning: int = 100
+
+    # Actor-learner transport (TODO(Khalil): relocate to TrainRLServerPipelineConfig).
     actor_learner_config: ActorLearnerConfig = field(default_factory=ActorLearnerConfig)
-    # Configuration for concurrency settings (you can use threads or processes for the actor and learner)
     concurrency: ConcurrencyConfig = field(default_factory=ConcurrencyConfig)
 
-    # Optimizations
-    use_torch_compile: bool = True
+    # Network architecture
+    # Actor network
+    actor_network_kwargs: ActorNetworkConfig = field(default_factory=ActorNetworkConfig)
+    # Gaussian head parameters
+    policy_kwargs: PolicyConfig = field(default_factory=PolicyConfig)
+    # Discrete critic
+    discrete_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
 
     def __post_init__(self):
         super().__post_init__()
-        # Any validation specific to SAC configuration
 
     def get_optimizer_preset(self) -> MultiAdamConfig:
         return MultiAdamConfig(
             weight_decay=0.0,
             optimizer_groups={
-                "actor": {"lr": self.actor_lr},
-                "critic": {"lr": self.critic_lr},
-                "temperature": {"lr": self.temperature_lr},
+                "actor": {"lr": 3e-4},
+                "critic": {"lr": 3e-4},
+                "temperature": {"lr": 3e-4},
             },
         )
 

src/lerobot/policies/gaussian_actor/modeling_gaussian_actor.py

@@ -15,16 +15,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import math
 from collections.abc import Callable
 from dataclasses import asdict
-from typing import Literal
+from typing import Any
 
-import einops
-import numpy as np
 import torch
 import torch.nn as nn
-import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 from torch.distributions import MultivariateNormal, TanhTransform, Transform, TransformedDistribution
 
@@ -32,20 +28,29 @@ from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_STATE
 
 from ..pretrained import PreTrainedPolicy
 from ..utils import get_device_from_parameters
-from .configuration_sac import SACConfig, is_image_feature
+from .configuration_gaussian_actor import GaussianActorConfig, is_image_feature
 
 DISCRETE_DIMENSION_INDEX = -1  # Gripper is always the last dimension
 
 
-class SACPolicy(
+class GaussianActorPolicy(
     PreTrainedPolicy,
 ):
-    config_class = SACConfig
-    name = "sac"
+    """Gaussian actor + observation encoder.
+
+    Policy-side ``nn.Module`` used by SAC and related maximum-entropy continuous
+    control algorithms. It owns the actor network (``Policy``) and the observation
+    encoder (``GaussianActorObservationEncoder``); the critics, temperature, and
+    Bellman-update logic live on the algorithm side
+    (see ``lerobot.rl.algorithms.sac``).
+    """
+
+    config_class = GaussianActorConfig
+    name = "gaussian_actor"
 
     def __init__(
         self,
-        config: SACConfig | None = None,
+        config: GaussianActorConfig | None = None,
     ):
         super().__init__(config)
         config.validate_features()
@@ -54,9 +59,8 @@ class SACPolicy(
         # Determine action dimension and initialize all components
         continuous_action_dim = config.output_features[ACTION].shape[0]
         self._init_encoders()
-        self._init_critics(continuous_action_dim)
         self._init_actor(continuous_action_dim)
-        self._init_temperature()
+        self._init_discrete_critic()
 
     def get_optim_params(self) -> dict:
         optim_params = {
@@ -65,11 +69,7 @@ class SACPolicy(
                 for n, p in self.actor.named_parameters()
                 if not n.startswith("encoder") or not self.shared_encoder
             ],
-            "critic": self.critic_ensemble.parameters(),
-            "temperature": self.log_alpha,
         }
-        if self.config.num_discrete_actions is not None:
-            optim_params["discrete_critic"] = self.discrete_critic.parameters()
         return optim_params
 
     def reset(self):
@@ -79,7 +79,9 @@ class SACPolicy(
     @torch.no_grad()
     def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
         """Predict a chunk of actions given environment observations."""
-        raise NotImplementedError("SACPolicy does not support action chunking. It returns single actions!")
+        raise NotImplementedError(
+            "GaussianActorPolicy does not support action chunking. It returns single actions!"
+        )
 
     @torch.no_grad()
     def select_action(self, batch: dict[str, Tensor]) -> Tensor:
@@ -92,360 +94,55 @@ class SACPolicy(
         actions, _, _ = self.actor(batch, observations_features)
 
         if self.config.num_discrete_actions is not None:
-            discrete_action_value = self.discrete_critic(batch, observations_features)
-            discrete_action = torch.argmax(discrete_action_value, dim=-1, keepdim=True)
+            if self.discrete_critic is not None:
+                discrete_action_value = self.discrete_critic(batch, observations_features)
+                discrete_action = torch.argmax(discrete_action_value, dim=-1, keepdim=True)
+            else:
+                discrete_action = torch.ones(
+                    (*actions.shape[:-1], 1), device=actions.device, dtype=actions.dtype
+                )
             actions = torch.cat([actions, discrete_action], dim=-1)
 
         return actions
 
-    def critic_forward(
-        self,
-        observations: dict[str, Tensor],
-        actions: Tensor,
-        use_target: bool = False,
-        observation_features: Tensor | None = None,
-    ) -> Tensor:
-        """Forward pass through a critic network ensemble
+    def forward(self, batch: dict[str, Tensor | dict[str, Tensor]]) -> dict[str, Tensor]:
+        """Actor forward pass: sample actions and return log-probabilities.
 
         Args:
-            observations: Dictionary of observations
-            actions: Action tensor
-            use_target: If True, use target critics, otherwise use ensemble critics
+            batch: A flat observation dict, or a training dict containing
+                ``"state"`` (observations) and optionally ``"observation_feature"``
+                (pre-computed encoder features).
 
         Returns:
-            Tensor of Q-values from all critics
+            Dict with ``"action"``, ``"log_prob"``, and ``"action_mean"`` tensors.
         """
+        observations = batch.get("state", batch)
+        observation_features = batch.get("observation_feature") if isinstance(batch, dict) else None
+        actions, log_probs, means = self.actor(observations, observation_features)
+        return {"action": actions, "log_prob": log_probs, "action_mean": means}
 
-        critics = self.critic_target if use_target else self.critic_ensemble
-        q_values = critics(observations, actions, observation_features)
-        return q_values
+    def load_actor_weights(self, state_dicts: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        from lerobot.utils.transition import move_state_dict_to_device
 
-    def discrete_critic_forward(
-        self, observations, use_target=False, observation_features=None
-    ) -> torch.Tensor:
-        """Forward pass through a discrete critic network
+        actor_state_dict = move_state_dict_to_device(state_dicts["policy"], device=device)
+        self.actor.load_state_dict(actor_state_dict)
 
-        Args:
-            observations: Dictionary of observations
-            use_target: If True, use target critics, otherwise use ensemble critics
-            observation_features: Optional pre-computed observation features to avoid recomputing encoder output
-
-        Returns:
-            Tensor of Q-values from the discrete critic network
-        """
-        discrete_critic = self.discrete_critic_target if use_target else self.discrete_critic
-        q_values = discrete_critic(observations, observation_features)
-        return q_values
-
-    def forward(
-        self,
-        batch: dict[str, Tensor | dict[str, Tensor]],
-        model: Literal["actor", "critic", "temperature", "discrete_critic"] = "critic",
-    ) -> dict[str, Tensor]:
-        """Compute the loss for the given model
-
-        Args:
-            batch: Dictionary containing:
-                - action: Action tensor
-                - reward: Reward tensor
-                - state: Observations tensor dict
-                - next_state: Next observations tensor dict
-                - done: Done mask tensor
-                - observation_feature: Optional pre-computed observation features
-                - next_observation_feature: Optional pre-computed next observation features
-            model: Which model to compute the loss for ("actor", "critic", "discrete_critic", or "temperature")
-
-        Returns:
-            The computed loss tensor
-        """
-        # Extract common components from batch
-        actions: Tensor = batch[ACTION]
-        observations: dict[str, Tensor] = batch["state"]
-        observation_features: Tensor = batch.get("observation_feature")
-
-        if model == "critic":
-            # Extract critic-specific components
-            rewards: Tensor = batch["reward"]
-            next_observations: dict[str, Tensor] = batch["next_state"]
-            done: Tensor = batch["done"]
-            next_observation_features: Tensor = batch.get("next_observation_feature")
-
-            loss_critic = self.compute_loss_critic(
-                observations=observations,
-                actions=actions,
-                rewards=rewards,
-                next_observations=next_observations,
-                done=done,
-                observation_features=observation_features,
-                next_observation_features=next_observation_features,
+        if "discrete_critic" in state_dicts and self.discrete_critic is not None:
+            discrete_critic_state_dict = move_state_dict_to_device(
+                state_dicts["discrete_critic"], device=device
             )
-
-            return {"loss_critic": loss_critic}
-
-        if model == "discrete_critic" and self.config.num_discrete_actions is not None:
-            # Extract critic-specific components
-            rewards: Tensor = batch["reward"]
-            next_observations: dict[str, Tensor] = batch["next_state"]
-            done: Tensor = batch["done"]
-            next_observation_features: Tensor = batch.get("next_observation_feature")
-            complementary_info = batch.get("complementary_info")
-            loss_discrete_critic = self.compute_loss_discrete_critic(
-                observations=observations,
-                actions=actions,
-                rewards=rewards,
-                next_observations=next_observations,
-                done=done,
-                observation_features=observation_features,
-                next_observation_features=next_observation_features,
-                complementary_info=complementary_info,
-            )
-            return {"loss_discrete_critic": loss_discrete_critic}
-        if model == "actor":
-            return {
-                "loss_actor": self.compute_loss_actor(
-                    observations=observations,
-                    observation_features=observation_features,
-                )
-            }
-
-        if model == "temperature":
-            return {
-                "loss_temperature": self.compute_loss_temperature(
-                    observations=observations,
-                    observation_features=observation_features,
-                )
-            }
-
-        raise ValueError(f"Unknown model type: {model}")
-
-    def update_target_networks(self):
-        """Update target networks with exponential moving average"""
-        for target_param, param in zip(
-            self.critic_target.parameters(),
-            self.critic_ensemble.parameters(),
-            strict=True,
-        ):
-            target_param.data.copy_(
-                param.data * self.config.critic_target_update_weight
-                + target_param.data * (1.0 - self.config.critic_target_update_weight)
-            )
-        if self.config.num_discrete_actions is not None:
-            for target_param, param in zip(
-                self.discrete_critic_target.parameters(),
-                self.discrete_critic.parameters(),
-                strict=True,
-            ):
-                target_param.data.copy_(
-                    param.data * self.config.critic_target_update_weight
-                    + target_param.data * (1.0 - self.config.critic_target_update_weight)
-                )
-
-    @property
-    def temperature(self) -> float:
-        """Return the current temperature value, always in sync with log_alpha."""
-        return self.log_alpha.exp().item()
-
-    def compute_loss_critic(
-        self,
-        observations,
-        actions,
-        rewards,
-        next_observations,
-        done,
-        observation_features: Tensor | None = None,
-        next_observation_features: Tensor | None = None,
-    ) -> Tensor:
-        with torch.no_grad():
-            next_action_preds, next_log_probs, _ = self.actor(next_observations, next_observation_features)
-
-            # 2- compute q targets
-            q_targets = self.critic_forward(
-                observations=next_observations,
-                actions=next_action_preds,
-                use_target=True,
-                observation_features=next_observation_features,
-            )
-
-            # subsample critics to prevent overfitting if use high UTD (update to date)
-            # TODO: Get indices before forward pass to avoid unnecessary computation
-            if self.config.num_subsample_critics is not None:
-                indices = torch.randperm(self.config.num_critics)
-                indices = indices[: self.config.num_subsample_critics]
-                q_targets = q_targets[indices]
-
-            # critics subsample size
-            min_q, _ = q_targets.min(dim=0)  # Get values from min operation
-            if self.config.use_backup_entropy:
-                min_q = min_q - (self.temperature * next_log_probs)
-
-            td_target = rewards + (1 - done) * self.config.discount * min_q
-
-        # 3- compute predicted qs
-        if self.config.num_discrete_actions is not None:
-            # NOTE: We only want to keep the continuous action part
-            # In the buffer we have the full action space (continuous + discrete)
-            # We need to split them before concatenating them in the critic forward
-            actions: Tensor = actions[:, :DISCRETE_DIMENSION_INDEX]
-        q_preds = self.critic_forward(
-            observations=observations,
-            actions=actions,
-            use_target=False,
-            observation_features=observation_features,
-        )
-
-        # 4- Calculate loss
-        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
-        td_target_duplicate = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
-        # You compute the mean loss of the batch for each critic and then to compute the final loss you sum them up
-        critics_loss = (
-            F.mse_loss(
-                input=q_preds,
-                target=td_target_duplicate,
-                reduction="none",
-            ).mean(dim=1)
-        ).sum()
-        return critics_loss
-
-    def compute_loss_discrete_critic(
-        self,
-        observations,
-        actions,
-        rewards,
-        next_observations,
-        done,
-        observation_features=None,
-        next_observation_features=None,
-        complementary_info=None,
-    ):
-        # NOTE: We only want to keep the discrete action part
-        # In the buffer we have the full action space (continuous + discrete)
-        # We need to split them before concatenating them in the critic forward
-        actions_discrete: Tensor = actions[:, DISCRETE_DIMENSION_INDEX:].clone()
-        actions_discrete = torch.round(actions_discrete)
-        actions_discrete = actions_discrete.long()
-
-        discrete_penalties: Tensor | None = None
-        if complementary_info is not None:
-            discrete_penalties: Tensor | None = complementary_info.get("discrete_penalty")
-
-        with torch.no_grad():
-            # For DQN, select actions using online network, evaluate with target network
-            next_discrete_qs = self.discrete_critic_forward(
-                next_observations, use_target=False, observation_features=next_observation_features
-            )
-            best_next_discrete_action = torch.argmax(next_discrete_qs, dim=-1, keepdim=True)
-
-            # Get target Q-values from target network
-            target_next_discrete_qs = self.discrete_critic_forward(
-                observations=next_observations,
-                use_target=True,
-                observation_features=next_observation_features,
-            )
-
-            # Use gather to select Q-values for best actions
-            target_next_discrete_q = torch.gather(
-                target_next_discrete_qs, dim=1, index=best_next_discrete_action
-            ).squeeze(-1)
-
-            # Compute target Q-value with Bellman equation
-            rewards_discrete = rewards
-            if discrete_penalties is not None:
-                rewards_discrete = rewards + discrete_penalties
-            target_discrete_q = rewards_discrete + (1 - done) * self.config.discount * target_next_discrete_q
-
-        # Get predicted Q-values for current observations
-        predicted_discrete_qs = self.discrete_critic_forward(
-            observations=observations, use_target=False, observation_features=observation_features
-        )
-
-        # Use gather to select Q-values for taken actions
-        predicted_discrete_q = torch.gather(predicted_discrete_qs, dim=1, index=actions_discrete).squeeze(-1)
-
-        # Compute MSE loss between predicted and target Q-values
-        discrete_critic_loss = F.mse_loss(input=predicted_discrete_q, target=target_discrete_q)
-        return discrete_critic_loss
-
-    def compute_loss_temperature(self, observations, observation_features: Tensor | None = None) -> Tensor:
-        """Compute the temperature loss"""
-        # calculate temperature loss
-        with torch.no_grad():
-            _, log_probs, _ = self.actor(observations, observation_features)
-        temperature_loss = (-self.log_alpha.exp() * (log_probs + self.target_entropy)).mean()
-        return temperature_loss
-
-    def compute_loss_actor(
-        self,
-        observations,
-        observation_features: Tensor | None = None,
-    ) -> Tensor:
-        actions_pi, log_probs, _ = self.actor(observations, observation_features)
-
-        q_preds = self.critic_forward(
-            observations=observations,
-            actions=actions_pi,
-            use_target=False,
-            observation_features=observation_features,
-        )
-        min_q_preds = q_preds.min(dim=0)[0]
-
-        actor_loss = ((self.temperature * log_probs) - min_q_preds).mean()
-        return actor_loss
+            self.discrete_critic.load_state_dict(discrete_critic_state_dict)
 
     def _init_encoders(self):
         """Initialize shared or separate encoders for actor and critic."""
         self.shared_encoder = self.config.shared_encoder
-        self.encoder_critic = SACObservationEncoder(self.config)
+        self.encoder_critic = GaussianActorObservationEncoder(self.config)
         self.encoder_actor = (
-            self.encoder_critic if self.shared_encoder else SACObservationEncoder(self.config)
+            self.encoder_critic if self.shared_encoder else GaussianActorObservationEncoder(self.config)
         )
 
-    def _init_critics(self, continuous_action_dim):
-        """Build critic ensemble, targets, and optional discrete critic."""
-        heads = [
-            CriticHead(
-                input_dim=self.encoder_critic.output_dim + continuous_action_dim,
-                **asdict(self.config.critic_network_kwargs),
-            )
-            for _ in range(self.config.num_critics)
-        ]
-        self.critic_ensemble = CriticEnsemble(encoder=self.encoder_critic, ensemble=heads)
-        target_heads = [
-            CriticHead(
-                input_dim=self.encoder_critic.output_dim + continuous_action_dim,
-                **asdict(self.config.critic_network_kwargs),
-            )
-            for _ in range(self.config.num_critics)
-        ]
-        self.critic_target = CriticEnsemble(encoder=self.encoder_critic, ensemble=target_heads)
-        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
-
-        if self.config.use_torch_compile:
-            self.critic_ensemble = torch.compile(self.critic_ensemble)
-            self.critic_target = torch.compile(self.critic_target)
-
-        if self.config.num_discrete_actions is not None:
-            self._init_discrete_critics()
-
-    def _init_discrete_critics(self):
-        """Build discrete discrete critic ensemble and target networks."""
-        self.discrete_critic = DiscreteCritic(
-            encoder=self.encoder_critic,
-            input_dim=self.encoder_critic.output_dim,
-            output_dim=self.config.num_discrete_actions,
-            **asdict(self.config.discrete_critic_network_kwargs),
-        )
-        self.discrete_critic_target = DiscreteCritic(
-            encoder=self.encoder_critic,
-            input_dim=self.encoder_critic.output_dim,
-            output_dim=self.config.num_discrete_actions,
-            **asdict(self.config.discrete_critic_network_kwargs),
-        )
-
-        # TODO: (maractingi, azouitine) Compile the discrete critic
-        self.discrete_critic_target.load_state_dict(self.discrete_critic.state_dict())
-
     def _init_actor(self, continuous_action_dim):
-        """Initialize policy actor network and default target entropy."""
+        """Initialize policy actor network."""
         # NOTE: The actor select only the continuous action part
         self.actor = Policy(
             encoder=self.encoder_actor,
@@ -455,21 +152,25 @@ class SACPolicy(
             **asdict(self.config.policy_kwargs),
         )
 
-        self.target_entropy = self.config.target_entropy
-        if self.target_entropy is None:
-            dim = continuous_action_dim + (1 if self.config.num_discrete_actions is not None else 0)
-            self.target_entropy = -np.prod(dim) / 2
+    def _init_discrete_critic(self) -> None:
+        """Initialize discrete critic network."""
+        if self.config.num_discrete_actions is None:
+            self.discrete_critic = None
+            return
 
-    def _init_temperature(self) -> None:
-        """Set up temperature parameter (log_alpha)."""
-        temp_init = self.config.temperature_init
-        self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
+        # TODO(Khalil): Compile the discrete critic
+        self.discrete_critic = DiscreteCritic(
+            encoder=self.encoder_critic,
+            input_dim=self.encoder_critic.output_dim,
+            output_dim=self.config.num_discrete_actions,
+            **asdict(self.config.discrete_critic_network_kwargs),
+        )
 
 
-class SACObservationEncoder(nn.Module):
+class GaussianActorObservationEncoder(nn.Module):
     """Encode image and/or state vector observations."""
 
-    def __init__(self, config: SACConfig) -> None:
+    def __init__(self, config: GaussianActorConfig) -> None:
         super().__init__()
         self.config = config
         self._init_image_layers()
@@ -677,84 +378,6 @@ class MLP(nn.Module):
         return self.net(x)
 
 
-class CriticHead(nn.Module):
-    def __init__(
-        self,
-        input_dim: int,
-        hidden_dims: list[int],
-        activations: Callable[[torch.Tensor], torch.Tensor] | str = nn.SiLU(),
-        activate_final: bool = False,
-        dropout_rate: float | None = None,
-        init_final: float | None = None,
-        final_activation: Callable[[torch.Tensor], torch.Tensor] | str | None = None,
-    ):
-        super().__init__()
-        self.net = MLP(
-            input_dim=input_dim,
-            hidden_dims=hidden_dims,
-            activations=activations,
-            activate_final=activate_final,
-            dropout_rate=dropout_rate,
-            final_activation=final_activation,
-        )
-        self.output_layer = nn.Linear(in_features=hidden_dims[-1], out_features=1)
-        if init_final is not None:
-            nn.init.uniform_(self.output_layer.weight, -init_final, init_final)
-            nn.init.uniform_(self.output_layer.bias, -init_final, init_final)
-        else:
-            orthogonal_init()(self.output_layer.weight)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        return self.output_layer(self.net(x))
-
-
-class CriticEnsemble(nn.Module):
-    """
-    CriticEnsemble wraps multiple CriticHead modules into an ensemble.
-
-    Args:
-        encoder (SACObservationEncoder): encoder for observations.
-        ensemble (List[CriticHead]): list of critic heads.
-        init_final (float | None): optional initializer scale for final layers.
-
-    Forward returns a tensor of shape (num_critics, batch_size) containing Q-values.
-    """
-
-    def __init__(
-        self,
-        encoder: SACObservationEncoder,
-        ensemble: list[CriticHead],
-        init_final: float | None = None,
-    ):
-        super().__init__()
-        self.encoder = encoder
-        self.init_final = init_final
-        self.critics = nn.ModuleList(ensemble)
-
-    def forward(
-        self,
-        observations: dict[str, torch.Tensor],
-        actions: torch.Tensor,
-        observation_features: torch.Tensor | None = None,
-    ) -> torch.Tensor:
-        device = get_device_from_parameters(self)
-        # Move each tensor in observations to device
-        observations = {k: v.to(device) for k, v in observations.items()}
-
-        obs_enc = self.encoder(observations, cache=observation_features)
-
-        inputs = torch.cat([obs_enc, actions], dim=-1)
-
-        # Loop through critics and collect outputs
-        q_values = []
-        for critic in self.critics:
-            q_values.append(critic(inputs))
-
-        # Stack outputs to match expected shape [num_critics, batch_size]
-        q_values = torch.stack([q.squeeze(-1) for q in q_values], dim=0)
-        return q_values
-
-
 class DiscreteCritic(nn.Module):
     def __init__(
         self,
@@ -800,7 +423,7 @@ class DiscreteCritic(nn.Module):
 class Policy(nn.Module):
     def __init__(
         self,
-        encoder: SACObservationEncoder,
+        encoder: GaussianActorObservationEncoder,
         network: nn.Module,
         action_dim: int,
         std_min: float = -5,
@@ -811,7 +434,7 @@ class Policy(nn.Module):
         encoder_is_shared: bool = False,
     ):
         super().__init__()
-        self.encoder: SACObservationEncoder = encoder
+        self.encoder: GaussianActorObservationEncoder = encoder
         self.network = network
         self.action_dim = action_dim
         self.std_min = std_min
@@ -885,7 +508,7 @@ class Policy(nn.Module):
 
 
 class DefaultImageEncoder(nn.Module):
-    def __init__(self, config: SACConfig):
+    def __init__(self, config: GaussianActorConfig):
         super().__init__()
         image_key = next(key for key in config.input_features if is_image_feature(key))
         self.image_enc_layers = nn.Sequential(
@@ -931,12 +554,12 @@ def freeze_image_encoder(image_encoder: nn.Module):
 
 
 class PretrainedImageEncoder(nn.Module):
-    def __init__(self, config: SACConfig):
+    def __init__(self, config: GaussianActorConfig):
         super().__init__()
 
         self.image_enc_layers, self.image_enc_out_shape = self._load_pretrained_vision_encoder(config)
 
-    def _load_pretrained_vision_encoder(self, config: SACConfig):
+    def _load_pretrained_vision_encoder(self, config: GaussianActorConfig):
         """Set up CNN encoder"""
         from transformers import AutoModel
 

src/lerobot/policies/gaussian_actor/processor_gaussian_actor.py

@@ -32,18 +32,18 @@ from lerobot.processor import (
 )
 from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
-from .configuration_sac import SACConfig
+from .configuration_gaussian_actor import GaussianActorConfig
 
 
-def make_sac_pre_post_processors(
-    config: SACConfig,
+def make_gaussian_actor_pre_post_processors(
+    config: GaussianActorConfig,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
 ) -> tuple[
     PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     PolicyProcessorPipeline[PolicyAction, PolicyAction],
 ]:
     """
-    Constructs pre-processor and post-processor pipelines for the SAC policy.
+    Constructs pre-processor and post-processor pipelines for the Gaussian actor policy.
 
     The pre-processing pipeline prepares input data for the model by:
     1. Renaming features to match pretrained configurations.
@@ -56,7 +56,7 @@ def make_sac_pre_post_processors(
     2. Unnormalizing the output features to their original scale.
 
     Args:
-        config: The configuration object for the SAC policy.
+        config: The configuration object for the tanh-Gaussian policy.
         dataset_stats: A dictionary of statistics for normalization.
 
     Returns:

src/lerobot/policies/gaussian_actor/reward_model/configuration_classifier.py

@@ -31,7 +31,7 @@ class RewardClassifierConfig(PreTrainedConfig):
     latent_dim: int = 256
     image_embedding_pooling_dim: int = 8
     dropout_rate: float = 0.1
-    model_name: str = "helper2424/resnet10"  # TODO: This needs to be updated. The model on the Hub doesn't call self.post_init() in its __init__, which is required by transformers v5 to set all_tied_weights_keys. The from_pretrained call fails when it tries to access this attribute during _finalize_model_loading.
+    model_name: str = "lerobot/resnet10"
     device: str = "cpu"
     model_type: str = "cnn"  # "transformer" or "cnn"
     num_cameras: int = 2

src/lerobot/policies/gaussian_actor/reward_model/modeling_classifier.py

@@ -108,6 +108,7 @@ class Classifier(PreTrainedPolicy):
     def __init__(
         self,
         config: RewardClassifierConfig,
+        **kwargs,
     ):
         from transformers import AutoModel
 
@@ -269,10 +270,6 @@ class Classifier(PreTrainedPolicy):
 
     def predict_reward(self, batch, threshold=0.5):
         """Eval method. Returns predicted reward with the decision threshold as argument."""
-        # Check for both OBS_IMAGE and OBS_IMAGES prefixes
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
         # Extract images from batch dict
         images = [batch[key] for key in self.config.input_features if key.startswith(OBS_IMAGE)]
 

src/lerobot/policies/groot/groot_n1.py

@@ -13,6 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from dataclasses import dataclass, field
 from pathlib import Path
 from typing import TYPE_CHECKING
 
@@ -173,14 +174,17 @@ N_COLOR_CHANNELS = 3
 
 
 # config
+@dataclass
 class GR00TN15Config(PretrainedConfig):
     model_type = "gr00t_n1_5"
+    backbone_cfg: dict = field(init=False, metadata={"help": "Backbone configuration."})
 
-    backbone_cfg: dict
-    action_head_cfg: dict
-    action_horizon: int
-    action_dim: int
-    compute_dtype: str = "float32"
+    action_head_cfg: dict = field(init=False, metadata={"help": "Action head configuration."})
+
+    action_horizon: int = field(init=False, metadata={"help": "Action horizon."})
+
+    action_dim: int = field(init=False, metadata={"help": "Action dimension."})
+    compute_dtype: str = field(default="float32", metadata={"help": "Compute dtype."})
 
     def __init__(self, **kwargs):
         super().__init__(**kwargs)

src/lerobot/policies/groot/modeling_groot.py

@@ -43,7 +43,6 @@ from torch import Tensor
 
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.utils.constants import ACTION, OBS_IMAGES
-from lerobot.utils.import_utils import require_package
 
 from ..pretrained import PreTrainedPolicy
 from .configuration_groot import GrootConfig
@@ -60,7 +59,6 @@ class GrootPolicy(PreTrainedPolicy):
 
     def __init__(self, config: GrootConfig, **kwargs):
         """Initialize Groot policy wrapper."""
-        require_package("transformers", extra="groot")
         super().__init__(config)
         config.validate_features()
         self.config = config

src/lerobot/policies/multi_task_dit/modeling_multi_task_dit.py

@@ -36,7 +36,7 @@ import torch.nn.functional as F  # noqa: N812
 import torchvision
 from torch import Tensor
 
-from lerobot.utils.import_utils import _diffusers_available, _transformers_available, require_package
+from lerobot.utils.import_utils import _transformers_available
 
 from .configuration_multi_task_dit import MultiTaskDiTConfig
 
@@ -46,13 +46,6 @@ if TYPE_CHECKING or _transformers_available:
 else:
     CLIPTextModel = None
     CLIPVisionModel = None
-
-if TYPE_CHECKING or _diffusers_available:
-    from diffusers.schedulers.scheduling_ddim import DDIMScheduler
-    from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
-else:
-    DDIMScheduler = None
-    DDPMScheduler = None
 from lerobot.utils.constants import (
     ACTION,
     OBS_IMAGES,
@@ -72,8 +65,6 @@ class MultiTaskDiTPolicy(PreTrainedPolicy):
     name = "multi_task_dit"
 
     def __init__(self, config: MultiTaskDiTConfig, **kwargs):
-        require_package("transformers", extra="multi_task_dit")
-        require_package("diffusers", extra="multi_task_dit")
         super().__init__(config)
         config.validate_features()
         self.config = config
@@ -652,6 +643,12 @@ class DiffusionObjective(nn.Module):
             "prediction_type": config.prediction_type,
         }
 
+        from lerobot.utils.import_utils import require_package
+
+        require_package("diffusers", extra="multi_task_dit")
+        from diffusers.schedulers.scheduling_ddim import DDIMScheduler
+        from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
+
         if config.noise_scheduler_type == "DDPM":
             self.noise_scheduler: DDPMScheduler | DDIMScheduler = DDPMScheduler(**scheduler_kwargs)
         elif config.noise_scheduler_type == "DDIM":
@@ -688,9 +685,8 @@ class DiffusionObjective(nn.Module):
         loss = F.mse_loss(predicted, target, reduction="none")
 
         if self.do_mask_loss_for_padding and "action_is_pad" in batch:
-            mask = ~batch["action_is_pad"].unsqueeze(-1)
-            num_valid = mask.sum() * loss.shape[-1]
-            return (loss * mask).sum() / num_valid.clamp_min(1)
+            valid_actions = ~batch["action_is_pad"]
+            loss = loss * valid_actions.unsqueeze(-1)
 
         return loss.mean()
 
@@ -753,9 +749,8 @@ class FlowMatchingObjective(nn.Module):
         loss = F.mse_loss(predicted_velocity, target_velocity, reduction="none")
 
         if self.do_mask_loss_for_padding and "action_is_pad" in batch:
-            mask = ~batch["action_is_pad"].unsqueeze(-1)
-            num_valid = mask.sum() * loss.shape[-1]
-            return (loss * mask).sum() / num_valid.clamp_min(1)
+            valid_mask = ~batch["action_is_pad"]
+            loss = loss * valid_mask.unsqueeze(-1)
 
         return loss.mean()
 

src/lerobot/policies/pi0/modeling_pi0.py

@@ -26,7 +26,7 @@ import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
 
-from lerobot.utils.import_utils import _transformers_available, require_package
+from lerobot.utils.import_utils import _transformers_available
 
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _transformers_available:
@@ -947,7 +947,6 @@ class PI0Policy(PreTrainedPolicy):
         Args:
             config: Policy configuration class instance.
         """
-        require_package("transformers", extra="pi")
         super().__init__(config)
         config.validate_features()
         self.config = config

src/lerobot/policies/pi05/modeling_pi05.py

@@ -26,7 +26,7 @@ import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
 
-from lerobot.utils.import_utils import _transformers_available, require_package
+from lerobot.utils.import_utils import _transformers_available
 
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _transformers_available:
@@ -918,7 +918,6 @@ class PI05Policy(PreTrainedPolicy):
         Args:
             config: Policy configuration class instance.
         """
-        require_package("transformers", extra="pi")
         super().__init__(config)
         config.validate_features()
         self.config = config

src/lerobot/policies/pi0_fast/modeling_pi0_fast.py

@@ -26,7 +26,7 @@ import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
 
-from lerobot.utils.import_utils import _scipy_available, _transformers_available, require_package
+from lerobot.utils.import_utils import _scipy_available, _transformers_available
 
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _scipy_available:
@@ -35,7 +35,7 @@ else:
     idct = None
 
 if TYPE_CHECKING or _transformers_available:
-    from transformers import AutoProcessor, AutoTokenizer
+    from transformers import AutoTokenizer
     from transformers.models.auto import CONFIG_MAPPING
 
     from ..pi_gemma import (
@@ -44,7 +44,6 @@ if TYPE_CHECKING or _transformers_available:
     )
 else:
     CONFIG_MAPPING = None
-    AutoProcessor = None
     AutoTokenizer = None
     PiGemmaModel = None
     PaliGemmaForConditionalGenerationWithPiGemma = None
@@ -827,14 +826,14 @@ class PI0FastPolicy(PreTrainedPolicy):
         Args:
             config: Policy configuration class instance.
         """
-        require_package("transformers", extra="pi")
-        require_package("scipy", extra="pi")
         super().__init__(config)
         config.validate_features()
         self.config = config
 
         # Load tokenizers first
         try:
+            from transformers import AutoProcessor, AutoTokenizer
+
             # Load FAST tokenizer
             self.action_tokenizer = AutoProcessor.from_pretrained(
                 config.action_tokenizer_name, trust_remote_code=True

src/lerobot/policies/sarm/processor_sarm.py

@@ -455,13 +455,7 @@ class SARMEncodingProcessorStep(ProcessorStep):
             inputs = {k: v.to(self.device) for k, v in inputs.items()}
 
             # Get image embeddings
-            # transformers 5.x returns BaseModelOutputWithPooling instead of a plain tensor
-            output = self.clip_model.get_image_features(**inputs)
-            if not isinstance(output, torch.Tensor):
-                output = output.pooler_output
-                if output is None:
-                    raise ValueError("pooler_output should not be None for CLIP models.")
-            embeddings = output.detach().cpu()
+            embeddings = self.clip_model.get_image_features(**inputs).detach().cpu()
 
             # Handle single frame case
             if embeddings.dim() == 1:
@@ -488,13 +482,7 @@ class SARMEncodingProcessorStep(ProcessorStep):
         inputs = self.clip_processor.tokenizer([text], return_tensors="pt", padding=True, truncation=True)
         inputs = {k: v.to(self.device) for k, v in inputs.items()}
 
-        # transformers 5.x returns BaseModelOutputWithPooling instead of a plain tensor
-        output = self.clip_model.get_text_features(**inputs)
-        if not isinstance(output, torch.Tensor):
-            output = output.pooler_output
-            if output is None:
-                raise ValueError("pooler_output should not be None for CLIP models.")
-        text_embedding = output.detach().cpu()
+        text_embedding = self.clip_model.get_text_features(**inputs).detach().cpu()
         text_embedding = text_embedding.expand(batch_size, -1)
 
         return text_embedding

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -62,7 +62,6 @@ from torch import Tensor, nn
 
 from lerobot.utils.constants import ACTION, OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS, OBS_STATE
 from lerobot.utils.device_utils import get_safe_dtype
-from lerobot.utils.import_utils import require_package
 
 from ..pretrained import PreTrainedPolicy
 from ..rtc.modeling_rtc import RTCProcessor
@@ -240,7 +239,6 @@ class SmolVLAPolicy(PreTrainedPolicy):
                     the configuration class is used.
         """
 
-        require_package("transformers", extra="smolvla")
         super().__init__(config)
         config.validate_features()
         self.config = config
@@ -394,21 +392,13 @@ class SmolVLAPolicy(PreTrainedPolicy):
         loss_dict["losses_after_rm_padding"] = losses.clone().mean().item()
 
         if reduction == "none":
-            # Return per-sample losses (B,) by averaging over valid (time, action) entries
-            if actions_is_pad is None:
-                per_sample_loss = losses.mean(dim=(1, 2))
-            else:
-                num_valid = ((~actions_is_pad).sum(dim=1) * losses.shape[-1]).clamp_min(1)
-                per_sample_loss = losses.sum(dim=(1, 2)) / num_valid
+            # Return per-sample losses (B,) by averaging over time and action dims
+            per_sample_loss = losses.mean(dim=(1, 2))
             loss_dict["loss"] = per_sample_loss.mean().item()
             return per_sample_loss, loss_dict
         else:
-            # Default: return scalar mean loss over valid (time, action) entries
-            if actions_is_pad is None:
-                loss = losses.mean()
-            else:
-                num_valid = ((~actions_is_pad).sum() * losses.shape[-1]).clamp_min(1)
-                loss = losses.sum() / num_valid
+            # Default: return scalar mean loss
+            loss = losses.mean()
             loss_dict["loss"] = loss.item()
             return loss, loss_dict
 

src/lerobot/policies/vqbet/vqbet_utils.py

@@ -27,7 +27,7 @@ import torch.distributed as distributed
 import torch.nn.functional as F  # noqa: N812
 from einops import pack, rearrange, reduce, repeat, unpack
 from torch import einsum, nn
-from torch.amp import autocast
+from torch.cuda.amp import autocast
 from torch.optim import Optimizer
 
 from .configuration_vqbet import VQBeTConfig
@@ -1370,7 +1370,7 @@ class EuclideanCodebook(nn.Module):
         batch_samples = rearrange(batch_samples, "h ... d -> h (...) d")
         self.replace(batch_samples, batch_mask=expired_codes)
 
-    @autocast("cuda", enabled=False)
+    @autocast(enabled=False)
     def forward(self, x, sample_codebook_temp=None, mask=None, freeze_codebook=False):
         needs_codebook_dim = x.ndim < 4
         sample_codebook_temp = (

src/lerobot/processor/__init__.py

@@ -61,6 +61,7 @@ from .hil_processor import (
     RewardClassifierProcessorStep,
     TimeLimitProcessorStep,
 )
+from .leader_follower_processor import LeaderArmInterventionStep
 from .newline_task_processor import NewLineTaskProcessorStep
 from .normalize_processor import NormalizerProcessorStep, UnnormalizerProcessorStep, hotswap_stats
 from .observation_processor import VanillaObservationProcessorStep
@@ -122,6 +123,7 @@ __all__ = [
     "ImageCropResizeProcessorStep",
     "InfoProcessorStep",
     "InterventionActionProcessorStep",
+    "LeaderArmInterventionStep",
     "make_default_processors",
     "make_default_teleop_action_processor",
     "make_default_robot_action_processor",

src/lerobot/processor/hil_processor.py

@@ -574,7 +574,7 @@ class RewardClassifierProcessorStep(ProcessorStep):
     def __post_init__(self):
         """Initializes the reward classifier model after the dataclass is created."""
         if self.pretrained_path is not None:
-            from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
+            from lerobot.policies.gaussian_actor.reward_model.modeling_classifier import Classifier
 
             self.reward_classifier = Classifier.from_pretrained(self.pretrained_path)
             self.reward_classifier.to(self.device)

src/lerobot/processor/leader_follower_processor.py

@@ -0,0 +1,270 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Processor step for using a leader arm as the HIL-SERL intervention device.
+
+Position-only port of the leader/follower control mode (no rotation): the leader
+arm acts as a 4-D end-effector delta source ``[dx, dy, dz, gripper]`` for the
+existing ``InterventionActionProcessorStep`` overriding pipeline.
+
+The teleop_action returned by the leader is a flat dictionary of joint angles
+(degrees) like ``{"shoulder_pan.pos": ..., ..., "gripper.pos": ...}``. This step
+converts that into a normalised EE-delta dictionary by:
+
+1. Running forward kinematics on the leader joints -> ``p_leader`` (xyz, m).
+2. Running forward kinematics on the follower joints (read from the env
+   transition's observation / complementary data) -> ``p_follower`` (xyz, m).
+3. Normalising ``p_leader - p_follower`` by ``end_effector_step_sizes`` and
+   clipping to ``[-1, 1]`` (matches the gamepad / keyboard EE convention).
+4. Mapping the leader gripper position ``[0, 100]`` to the discrete
+   ``{0=close, 1=stay, 2=open}`` action used by the SO follower.
+
+The output is written back to ``complementary_data["teleop_action"]`` so the
+rest of the action pipeline (``InterventionActionProcessorStep`` ->
+``MapTensorToDeltaActionDictStep`` -> IK) is unchanged.
+
+Additionally, when an optional ``teleop_device`` reference is provided, this
+step also pushes the follower's raw joint positions back to the leader via
+``teleop_device.send_action(follower_joints)`` every tick. Combined with
+:class:`SOLeaderFollower.send_action`, this implements the **haptic follow**
+behaviour from https://github.com/huggingface/lerobot/pull/2596: the leader
+mimics the follower while the human is hands-off, then drops torque the
+moment intervention is toggled so the user can grab and steer it.
+"""
+
+import logging
+from dataclasses import dataclass, field
+from typing import Any
+
+import numpy as np
+
+from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.model import RobotKinematics
+from lerobot.types import EnvTransition, TransitionKey
+
+from .pipeline import ProcessorStep, ProcessorStepRegistry
+
+logger = logging.getLogger(__name__)
+
+TELEOP_ACTION_KEY = "teleop_action"
+RAW_JOINT_POSITIONS_KEY = "raw_joint_positions"
+GRIPPER_KEY = "gripper"
+
+# Leader gripper is in [0, 100] when calibrated.
+LEADER_GRIPPER_OPEN_DEFAULT = 60.0
+LEADER_GRIPPER_CLOSE_DEFAULT = 30.0
+
+# Discrete gripper command convention (matches GripperVelocityToJoint).
+GRIPPER_CLOSE = 0.0
+GRIPPER_STAY = 1.0
+GRIPPER_OPEN = 2.0
+
+
+def _joint_dict_to_array(joint_dict: dict[str, float], motor_names: list[str]) -> np.ndarray | None:
+    """Pull joint positions in ``motor_names`` order from a ``"<motor>.pos"`` dict.
+
+    Returns ``None`` if any motor is missing.
+    """
+    out = np.zeros(len(motor_names), dtype=float)
+    for i, name in enumerate(motor_names):
+        v = joint_dict.get(f"{name}.pos")
+        if v is None:
+            return None
+        out[i] = float(v)
+    return out
+
+
+@ProcessorStepRegistry.register("leader_arm_intervention")
+@dataclass
+class LeaderArmInterventionStep(ProcessorStep):
+    """Convert leader joint positions in ``teleop_action`` into a 4-D EE-delta dict.
+
+    This step is intended to run **between** ``AddTeleopActionAsComplimentaryDataStep``
+    (which populates ``complementary_data["teleop_action"]`` with raw leader joint
+    angles) and ``InterventionActionProcessorStep`` (which expects a delta dict).
+
+    Attributes:
+        kinematics: Robot kinematic model shared with the follower; used for FK
+            on both the leader arm and the follower arm. Both arms must use the
+            same URDF joint order.
+        motor_names: Ordered joint names matching ``kinematics.joint_names``,
+            used to slice joint dicts.
+        end_effector_step_sizes: Per-axis normalisation in metres, e.g.
+            ``{"x": 0.025, "y": 0.025, "z": 0.025}``. The clamped delta is
+            ``(p_leader - p_follower) / step_size``.
+        use_gripper: When ``True``, append a discrete gripper command derived from
+            the leader gripper joint to the output dict.
+        leader_gripper_open: Threshold (>= ) above which the leader gripper is
+            considered ``open`` -> command ``2``.
+        leader_gripper_close: Threshold (<= ) below which the leader gripper is
+            considered ``closed`` -> command ``0``.
+        teleop_device: Optional reference to the leader teleoperator. When set
+            and the device implements ``send_action(action_dict)``, this step
+            pushes the follower's raw joints to it every tick to drive haptic
+            follow. The teleop is responsible for gating actual motor writes on
+            its own intervention state (see :class:`SOLeaderFollower`).
+    """
+
+    kinematics: RobotKinematics
+    motor_names: list[str]
+    end_effector_step_sizes: dict[str, float]
+    use_gripper: bool = True
+    leader_gripper_open: float = LEADER_GRIPPER_OPEN_DEFAULT
+    leader_gripper_close: float = LEADER_GRIPPER_CLOSE_DEFAULT
+    teleop_device: Any = None
+
+    _initial_follower_joints: np.ndarray | None = field(default=None, init=False, repr=False)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        new_transition = transition.copy()
+        complementary_data = dict(new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {}) or {})
+
+        # Haptic follow: push follower joints to the leader every step (whether
+        # or not we have a usable leader action this tick). The leader's own
+        # send_action gates writes on its intervention state.
+        follower_joints_dict = self._read_follower_joints_dict(transition, complementary_data)
+        if follower_joints_dict is not None:
+            self._push_haptic_follow(follower_joints_dict)
+
+        leader_joints_dict = complementary_data.get(TELEOP_ACTION_KEY)
+        if not isinstance(leader_joints_dict, dict):
+            # Nothing to convert (e.g. teleop disconnected). Leave transition untouched.
+            return new_transition
+
+        if not any(k.endswith(".pos") for k in leader_joints_dict):
+            # Already in EE-delta form (or unrecognised); skip.
+            return new_transition
+
+        follower_joints = (
+            _joint_dict_to_array(follower_joints_dict, self.motor_names)
+            if follower_joints_dict is not None
+            else None
+        )
+        leader_joints = _joint_dict_to_array(leader_joints_dict, self.motor_names)
+
+        if follower_joints is None or leader_joints is None:
+            # Cannot compute delta this step; expose a zero-action so downstream
+            # InterventionActionProcessorStep does not propagate stale joints.
+            complementary_data[TELEOP_ACTION_KEY] = self._zero_action()
+            new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
+            return new_transition
+
+        p_leader = self.kinematics.forward_kinematics(leader_joints)[:3, 3]
+        p_follower = self.kinematics.forward_kinematics(follower_joints)[:3, 3]
+
+        delta = p_leader - p_follower
+        delta_norm = np.array(
+            [
+                delta[0] / max(self.end_effector_step_sizes.get("x", 1.0), 1e-6),
+                delta[1] / max(self.end_effector_step_sizes.get("y", 1.0), 1e-6),
+                delta[2] / max(self.end_effector_step_sizes.get("z", 1.0), 1e-6),
+            ],
+            dtype=float,
+        )
+        delta_norm = np.clip(delta_norm, -1.0, 1.0)
+
+        teleop_action: dict[str, float] = {
+            "delta_x": float(delta_norm[0]),
+            "delta_y": float(delta_norm[1]),
+            "delta_z": float(delta_norm[2]),
+        }
+
+        if self.use_gripper:
+            leader_gripper = float(leader_joints_dict.get(f"{GRIPPER_KEY}.pos", 50.0))
+            teleop_action[GRIPPER_KEY] = self._discretise_gripper(leader_gripper)
+
+        complementary_data[TELEOP_ACTION_KEY] = teleop_action
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
+        return new_transition
+
+    def _read_follower_joints_dict(
+        self, transition: EnvTransition, complementary_data: dict[str, Any]
+    ) -> dict[str, float] | None:
+        """Best-effort read of the follower joints from the transition.
+
+        Tries (in order):
+        1. ``complementary_data["raw_joint_positions"]`` (set after env.step).
+        2. ``transition[OBSERVATION]`` if it is a flat ``"<motor>.pos"`` dict
+           (this is the convention used by ``step_env_and_process_transition``
+           when staging an action transition).
+
+        Returns the source dict if all expected motors are present, else
+        ``None``. We return the *dict* (not the array) because we want to feed
+        it back to ``teleop_device.send_action`` for haptic follow.
+        """
+        raw = complementary_data.get(RAW_JOINT_POSITIONS_KEY)
+        if isinstance(raw, dict) and all(f"{m}.pos" in raw for m in self.motor_names):
+            return raw  # type: ignore[return-value]
+
+        observation = transition.get(TransitionKey.OBSERVATION)
+        if isinstance(observation, dict) and all(f"{m}.pos" in observation for m in self.motor_names):
+            return observation  # type: ignore[return-value]
+
+        return None
+
+    def _push_haptic_follow(self, follower_joints_dict: dict[str, float]) -> None:
+        """Send the follower's joints back to the leader for haptic follow.
+
+        Errors are logged once and swallowed -- a failed haptic update must
+        never break the policy / learner loop.
+        """
+        if self.teleop_device is None:
+            return
+        send_action = getattr(self.teleop_device, "send_action", None)
+        if send_action is None:
+            return
+        try:
+            send_action(follower_joints_dict)
+        except NotImplementedError:
+            # Plain SOLeader / unsupported teleop -- silently disable haptic follow.
+            self.teleop_device = None
+        except Exception as e:  # pragma: no cover - hardware path
+            logger.warning(f"[LeaderArmInterventionStep] haptic follow failed: {e}")
+
+    def _discretise_gripper(self, leader_gripper_pos: float) -> float:
+        """Map a leader gripper position in ``[0, 100]`` to ``{0, 1, 2}``."""
+        if leader_gripper_pos >= self.leader_gripper_open:
+            return GRIPPER_OPEN
+        if leader_gripper_pos <= self.leader_gripper_close:
+            return GRIPPER_CLOSE
+        return GRIPPER_STAY
+
+    def _zero_action(self) -> dict[str, float]:
+        out: dict[str, float] = {"delta_x": 0.0, "delta_y": 0.0, "delta_z": 0.0}
+        if self.use_gripper:
+            out[GRIPPER_KEY] = GRIPPER_STAY
+        return out
+
+    def get_config(self) -> dict[str, Any]:
+        # `kinematics` and `teleop_device` are runtime objects (not JSON-serializable)
+        # and are re-injected by `gym_manipulator.make_processors`, so they are
+        # intentionally omitted from the saved config.
+        return {
+            "motor_names": list(self.motor_names),
+            "end_effector_step_sizes": dict(self.end_effector_step_sizes),
+            "use_gripper": self.use_gripper,
+            "leader_gripper_open": self.leader_gripper_open,
+            "leader_gripper_close": self.leader_gripper_close,
+        }
+
+    def reset(self) -> None:
+        self._initial_follower_joints = None
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features

src/lerobot/rl/__init__.py

@@ -12,23 +12,33 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""
-Reinforcement learning modules.
+"""Reinforcement learning modules.
 
-Requires: ``pip install 'lerobot[hilserl]'``
-
-Available modules (import directly)::
-
-    from lerobot.rl.actor import ...
-    from lerobot.rl.learner import ...
-    from lerobot.rl.learner_service import ...
-    from lerobot.rl.buffer import ...
-    from lerobot.rl.eval_policy import ...
-    from lerobot.rl.gym_manipulator import ...
+Distributed actor / learner entry points (``actor``, ``learner``,
+``learner_service``) require ``pip install 'lerobot[hilserl]'``. Algorithms,
+buffer, data sources and trainer are gRPC-free and usable standalone.
 """
 
-from lerobot.utils.import_utils import require_package
+from .algorithms.base import RLAlgorithm as RLAlgorithm
+from .algorithms.configs import RLAlgorithmConfig as RLAlgorithmConfig, TrainingStats as TrainingStats
+from .algorithms.factory import (
+    make_algorithm as make_algorithm,
+    make_algorithm_config as make_algorithm_config,
+)
+from .algorithms.sac.configuration_sac import SACAlgorithmConfig as SACAlgorithmConfig
+from .buffer import ReplayBuffer as ReplayBuffer
+from .data_sources import DataMixer as DataMixer, OnlineOfflineMixer as OnlineOfflineMixer
+from .trainer import RLTrainer as RLTrainer
 
-require_package("grpcio", extra="hilserl", import_name="grpc")
-
-__all__: list[str] = []
+__all__ = [
+    "RLAlgorithm",
+    "RLAlgorithmConfig",
+    "TrainingStats",
+    "make_algorithm",
+    "make_algorithm_config",
+    "SACAlgorithmConfig",
+    "RLTrainer",
+    "ReplayBuffer",
+    "DataMixer",
+    "OnlineOfflineMixer",
+]

src/lerobot/rl/actor.py

@@ -51,17 +51,20 @@ import os
 import time
 from functools import lru_cache
 from queue import Empty
+from typing import Any
 
 import grpc
 import torch
 from torch import nn
-from torch.multiprocessing import Event, Queue
+from torch.multiprocessing import Queue
 
 from lerobot.cameras import opencv  # noqa: F401
 from lerobot.configs import parser
-from lerobot.configs.train import TrainRLServerPipelineConfig
-from lerobot.policies import make_policy, make_pre_post_processors
-from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.policies import PreTrainedPolicy, make_policy, make_pre_post_processors
+from lerobot.processor import TransitionKey
+from lerobot.rl.process import ProcessSignalHandler
+from lerobot.rl.queue import get_last_item_from_queue
+from lerobot.rl.train_rl import TrainRLServerPipelineConfig
 from lerobot.robots import so_follower  # noqa: F401
 from lerobot.teleoperators import gamepad, so_leader  # noqa: F401
 from lerobot.teleoperators.utils import TeleopEvents
@@ -74,13 +77,11 @@ from lerobot.transport.utils import (
     send_bytes_in_chunks,
     transitions_to_bytes,
 )
-from lerobot.types import TransitionKey
 from lerobot.utils.device_utils import get_safe_torch_device
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.transition import (
     Transition,
-    move_state_dict_to_device,
     move_transition_to_device,
 )
 from lerobot.utils.utils import (
@@ -94,8 +95,6 @@ from .gym_manipulator import (
     reset_and_build_transition,
     step_env_and_process_transition,
 )
-from .process import ProcessSignalHandler
-from .queue import get_last_item_from_queue
 
 # Main entry point
 
@@ -212,7 +211,7 @@ def actor_cli(cfg: TrainRLServerPipelineConfig):
 
 def act_with_policy(
     cfg: TrainRLServerPipelineConfig,
-    shutdown_event: any,  # Event,
+    shutdown_event: Any,  # Event
     parameters_queue: Queue,
     transitions_queue: Queue,
     interactions_queue: Queue,
@@ -252,13 +251,13 @@ def act_with_policy(
     logging.info("make_policy")
 
     ### Instantiate the policy in both the actor and learner processes
-    ### To avoid sending a SACPolicy object through the port, we create a policy instance
+    ### To avoid sending a policy object through the port, we create a policy instance
     ### on both sides, the learner sends the updated parameters every n steps to update the actor's parameters
-    policy: SACPolicy = make_policy(
+    policy = make_policy(
         cfg=cfg.policy,
         env_cfg=cfg.env,
     )
-    policy = policy.eval()
+    policy = policy.to(device).eval()
     assert isinstance(policy, nn.Module)
 
     preprocessor, postprocessor = make_pre_post_processors(
@@ -419,7 +418,7 @@ def act_with_policy(
 
 def establish_learner_connection(
     stub: services_pb2_grpc.LearnerServiceStub,
-    shutdown_event: Event,  # type: ignore
+    shutdown_event: Any,  # Event
     attempts: int = 30,
 ):
     """Establish a connection with the learner.
@@ -471,7 +470,7 @@ def learner_service_client(
 def receive_policy(
     cfg: TrainRLServerPipelineConfig,
     parameters_queue: Queue,
-    shutdown_event: Event,  # type: ignore
+    shutdown_event: Any,  # Event
     learner_client: services_pb2_grpc.LearnerServiceStub | None = None,
     grpc_channel: grpc.Channel | None = None,
 ):
@@ -523,7 +522,7 @@ def receive_policy(
 def send_transitions(
     cfg: TrainRLServerPipelineConfig,
     transitions_queue: Queue,
-    shutdown_event: any,  # Event,
+    shutdown_event: Any,  # Event
     learner_client: services_pb2_grpc.LearnerServiceStub | None = None,
     grpc_channel: grpc.Channel | None = None,
 ) -> services_pb2.Empty:
@@ -573,7 +572,7 @@ def send_transitions(
 def send_interactions(
     cfg: TrainRLServerPipelineConfig,
     interactions_queue: Queue,
-    shutdown_event: Event,  # type: ignore
+    shutdown_event: Any,  # Event
     learner_client: services_pb2_grpc.LearnerServiceStub | None = None,
     grpc_channel: grpc.Channel | None = None,
 ) -> services_pb2.Empty:
@@ -623,7 +622,11 @@ def send_interactions(
     logging.info("[ACTOR] Interactions process stopped")
 
 
-def transitions_stream(shutdown_event: Event, transitions_queue: Queue, timeout: float) -> services_pb2.Empty:  # type: ignore
+def transitions_stream(
+    shutdown_event: Any,  # Event
+    transitions_queue: Queue,
+    timeout: float,
+) -> services_pb2.Empty:
     while not shutdown_event.is_set():
         try:
             message = transitions_queue.get(block=True, timeout=timeout)
@@ -639,9 +642,9 @@ def transitions_stream(shutdown_event: Event, transitions_queue: Queue, timeout:
 
 
 def interactions_stream(
-    shutdown_event: Event,
+    shutdown_event: Any,  # Event
     interactions_queue: Queue,
-    timeout: float,  # type: ignore
+    timeout: float,
 ) -> services_pb2.Empty:
     while not shutdown_event.is_set():
         try:
@@ -662,7 +665,7 @@ def interactions_stream(
 #  Policy functions
 
 
-def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device):
+def update_policy_parameters(policy: PreTrainedPolicy, parameters_queue: Queue, device):
     bytes_state_dict = get_last_item_from_queue(parameters_queue, block=False)
     if bytes_state_dict is not None:
         logging.info("[ACTOR] Load new parameters from Learner.")
@@ -677,18 +680,7 @@ def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device)
         # - Send critic's encoder state when shared_encoder=True
         # - Skip encoder params entirely when freeze_vision_encoder=True
         # - Ensure discrete_critic gets correct encoder state (currently uses encoder_critic)
-
-        # Load actor state dict
-        actor_state_dict = move_state_dict_to_device(state_dicts["policy"], device=device)
-        policy.actor.load_state_dict(actor_state_dict)
-
-        # Load discrete critic if present
-        if hasattr(policy, "discrete_critic") and "discrete_critic" in state_dicts:
-            discrete_critic_state_dict = move_state_dict_to_device(
-                state_dicts["discrete_critic"], device=device
-            )
-            policy.discrete_critic.load_state_dict(discrete_critic_state_dict)
-            logging.info("[ACTOR] Loaded discrete critic parameters from Learner.")
+        policy.load_actor_weights(state_dicts, device=device)
 
 
 #  Utilities functions

src/lerobot/rl/algorithms/__init__.py

@@ -0,0 +1,20 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .sac import SACAlgorithm as SACAlgorithm, SACAlgorithmConfig as SACAlgorithmConfig
+
+__all__ = [
+    "SACAlgorithm",
+    "SACAlgorithmConfig",
+]

src/lerobot/rl/algorithms/base.py

@@ -0,0 +1,106 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import abc
+from collections.abc import Iterator
+from typing import TYPE_CHECKING, Any
+
+import torch
+from torch.optim import Optimizer
+
+from lerobot.rl.algorithms.configs import RLAlgorithmConfig, TrainingStats
+
+if TYPE_CHECKING:
+    from lerobot.rl.data_sources.data_mixer import DataMixer
+
+BatchType = dict[str, Any]
+
+
+class RLAlgorithm(abc.ABC):
+    """Base for all RL algorithms."""
+
+    config_class: type[RLAlgorithmConfig] | None = None
+    name: str | None = None
+
+    def __init_subclass__(cls, **kwargs):
+        super().__init_subclass__(**kwargs)
+        if not getattr(cls, "config_class", None):
+            raise TypeError(f"Class {cls.__name__} must define 'config_class'")
+        if not getattr(cls, "name", None):
+            raise TypeError(f"Class {cls.__name__} must define 'name'")
+
+    @abc.abstractmethod
+    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """One complete training step.
+
+        The algorithm calls ``next(batch_iterator)`` as many times as it
+        needs (e.g. ``utd_ratio`` times for SAC) to obtain fresh batches.
+        The iterator is owned by the trainer; the algorithm just consumes
+        from it.
+        """
+        ...
+
+    def configure_data_iterator(
+        self,
+        data_mixer: DataMixer,
+        batch_size: int,
+        *,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ) -> Iterator[BatchType]:
+        """Create the data iterator this algorithm needs.
+
+        The default implementation uses the standard ``data_mixer.get_iterator()``.
+        Algorithms that need specialised sampling should override this method.
+        """
+        return data_mixer.get_iterator(
+            batch_size=batch_size,
+            async_prefetch=async_prefetch,
+            queue_size=queue_size,
+        )
+
+    def make_optimizers_and_scheduler(self) -> dict[str, Optimizer]:
+        """Create, store, and return the optimizers needed for training.
+
+        Called on the **learner** side after construction.  Subclasses must
+        override this with algorithm-specific optimizer setup.
+        """
+        return {}
+
+    def get_optimizers(self) -> dict[str, Optimizer]:
+        """Return optimizers for checkpointing / external scheduling."""
+        return {}
+
+    @property
+    def optimization_step(self) -> int:
+        """Current learner optimization step.
+
+        Part of the stable contract for checkpoint/resume. Algorithms can
+        either use this default storage or override for custom behavior.
+        """
+        return getattr(self, "_optimization_step", 0)
+
+    @optimization_step.setter
+    def optimization_step(self, value: int) -> None:
+        self._optimization_step = int(value)
+
+    def get_weights(self) -> dict[str, Any]:
+        """Policy state-dict to push to actors."""
+        return {}
+
+    @abc.abstractmethod
+    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        """Load policy state-dict received from the learner."""

src/lerobot/rl/algorithms/configs.py

@@ -0,0 +1,76 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import abc
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+import draccus
+import torch
+
+if TYPE_CHECKING:
+    from lerobot.rl.algorithms.base import RLAlgorithm
+
+
+@dataclass
+class TrainingStats:
+    """Returned by ``algorithm.update()`` for logging and checkpointing."""
+
+    losses: dict[str, float] = field(default_factory=dict)
+    grad_norms: dict[str, float] = field(default_factory=dict)
+    extra: dict[str, float] = field(default_factory=dict)
+
+    def to_log_dict(self) -> dict[str, float]:
+        """Flatten all stats into a single dict for logging."""
+
+        d: dict[str, float] = {}
+        for name, val in self.losses.items():
+            d[name] = val
+        for name, val in self.grad_norms.items():
+            d[f"{name}_grad_norm"] = val
+        for name, val in self.extra.items():
+            d[name] = val
+        return d
+
+
+@dataclass
+class RLAlgorithmConfig(draccus.ChoiceRegistry, abc.ABC):
+    """Registry for algorithm configs."""
+
+    @property
+    def type(self) -> str:
+        """Registered name of this algorithm config (e.g. ``"sac"``)."""
+        choice_name = self.get_choice_name(self.__class__)
+        if not isinstance(choice_name, str):
+            raise TypeError(f"Expected string from get_choice_name, got {type(choice_name)}")
+        return choice_name
+
+    @abc.abstractmethod
+    def build_algorithm(self, policy: torch.nn.Module) -> RLAlgorithm:
+        """Construct the :class:`RLAlgorithm` for this config.
+
+        Must be overridden by every registered config subclass.
+        """
+        raise NotImplementedError(f"{type(self).__name__} must implement build_algorithm()")
+
+    @classmethod
+    @abc.abstractmethod
+    def from_policy_config(cls, policy_cfg: Any) -> RLAlgorithmConfig:
+        """Build an algorithm config from a policy config.
+
+        Must be overridden by every registered config subclass.
+        """
+        raise NotImplementedError(f"{cls.__name__} must implement from_policy_config()")

src/lerobot/rl/algorithms/factory.py

@@ -0,0 +1,47 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import torch
+
+from lerobot.rl.algorithms.base import RLAlgorithm
+from lerobot.rl.algorithms.configs import RLAlgorithmConfig
+
+
+def make_algorithm_config(algorithm_type: str, **kwargs) -> RLAlgorithmConfig:
+    """Instantiate an :class:`RLAlgorithmConfig` from its registered type name.
+
+    Args:
+        algorithm_type: Registry key of the algorithm (e.g. ``"sac"``).
+        **kwargs: Keyword arguments forwarded to the config class constructor.
+
+    Returns:
+        An instance of the matching ``RLAlgorithmConfig`` subclass.
+
+    Raises:
+        ValueError: If ``algorithm_type`` is not registered.
+    """
+    try:
+        cls = RLAlgorithmConfig.get_choice_class(algorithm_type)
+    except KeyError as err:
+        raise ValueError(
+            f"Algorithm type '{algorithm_type}' is not registered. "
+            f"Available: {list(RLAlgorithmConfig.get_known_choices().keys())}"
+        ) from err
+    return cls(**kwargs)
+
+
+def make_algorithm(cfg: RLAlgorithmConfig, policy: torch.nn.Module) -> RLAlgorithm:
+    return cfg.build_algorithm(policy)

src/lerobot/rl/algorithms/sac/__init__.py

@@ -0,0 +1,18 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from lerobot.rl.algorithms.sac.configuration_sac import SACAlgorithmConfig
+from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
+
+__all__ = ["SACAlgorithm", "SACAlgorithmConfig"]

src/lerobot/rl/algorithms/sac/configuration_sac.py

@@ -0,0 +1,90 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING
+
+import torch
+
+from lerobot.policies.gaussian_actor.configuration_gaussian_actor import (
+    CriticNetworkConfig,
+    GaussianActorConfig,
+)
+from lerobot.rl.algorithms.configs import RLAlgorithmConfig
+
+if TYPE_CHECKING:
+    from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
+
+
+@RLAlgorithmConfig.register_subclass("sac")
+@dataclass
+class SACAlgorithmConfig(RLAlgorithmConfig):
+    """SAC algorithm hyperparameters."""
+
+    # Optimizer learning rates
+    actor_lr: float = 3e-4
+    critic_lr: float = 3e-4
+    temperature_lr: float = 3e-4
+
+    # Bellman update
+    discount: float = 0.99
+    use_backup_entropy: bool = True
+    critic_target_update_weight: float = 0.005
+
+    # Critic ensemble
+    num_critics: int = 2
+    num_subsample_critics: int | None = None
+    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+    discrete_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+
+    # Temperature / entropy
+    temperature_init: float = 1.0
+    # Target entropy for automatic temperature tuning. If ``None``, defaults to
+    # ``-|A|/2`` where ``|A|`` is the total action dimension (continuous + 1 if
+    # there is a discrete action head).
+    target_entropy: float | None = None
+
+    # Update loop
+    utd_ratio: int = 1
+    policy_update_freq: int = 1
+    grad_clip_norm: float = 40.0
+
+    # Optimizations
+    # torch.compile is currently disabled by default
+    use_torch_compile: bool = False
+
+    # Policy config
+    policy_config: GaussianActorConfig | None = None
+
+    @classmethod
+    def from_policy_config(cls, policy_cfg: GaussianActorConfig) -> SACAlgorithmConfig:
+        """Build an algorithm config with default hyperparameters for a given policy."""
+        return cls(
+            policy_config=policy_cfg,
+            discrete_critic_network_kwargs=policy_cfg.discrete_critic_network_kwargs,
+        )
+
+    def build_algorithm(self, policy: torch.nn.Module) -> SACAlgorithm:
+        if self.policy_config is None:
+            raise ValueError(
+                "SACAlgorithmConfig.policy_config is None. "
+                "It must be populated (typically by TrainRLServerPipelineConfig.validate) "
+                "before calling build_algorithm()."
+            )
+
+        from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
+
+        return SACAlgorithm(policy=policy, config=self)

src/lerobot/rl/algorithms/sac/sac_algorithm.py

@@ -0,0 +1,595 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import math
+from collections.abc import Callable, Iterator
+from dataclasses import asdict
+from typing import Any
+
+import einops
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor
+from torch.optim import Optimizer
+
+from lerobot.policies.gaussian_actor.modeling_gaussian_actor import (
+    DISCRETE_DIMENSION_INDEX,
+    MLP,
+    DiscreteCritic,
+    GaussianActorObservationEncoder,
+    GaussianActorPolicy,
+    orthogonal_init,
+)
+from lerobot.policies.utils import get_device_from_parameters
+from lerobot.rl.algorithms.base import BatchType, RLAlgorithm
+from lerobot.rl.algorithms.configs import TrainingStats
+from lerobot.rl.algorithms.sac.configuration_sac import SACAlgorithmConfig
+from lerobot.utils.constants import ACTION
+from lerobot.utils.transition import move_state_dict_to_device
+
+
+class SACAlgorithm(RLAlgorithm):
+    """Soft Actor-Critic. Owns critics, targets, temperature, and loss computation."""
+
+    config_class = SACAlgorithmConfig
+    name = "sac"
+
+    def __init__(
+        self,
+        policy: GaussianActorPolicy,
+        config: SACAlgorithmConfig,
+    ):
+        self.config = config
+        self.policy_config = config.policy_config
+        self.policy = policy
+        self.optimizers: dict[str, Optimizer] = {}
+        self._optimization_step: int = 0
+
+        action_dim = self.policy.config.output_features[ACTION].shape[0]
+        self._init_critics(action_dim)
+        self._init_temperature(action_dim)
+
+        self._device = torch.device(self.policy.config.device)
+        self._move_to_device()
+
+    def _init_critics(self, action_dim) -> None:
+        """Build critic ensemble, targets."""
+        encoder = self.policy.encoder_critic
+
+        heads = [
+            CriticHead(
+                input_dim=encoder.output_dim + action_dim,
+                **asdict(self.config.critic_network_kwargs),
+            )
+            for _ in range(self.config.num_critics)
+        ]
+        self.critic_ensemble = CriticEnsemble(encoder=encoder, ensemble=heads)
+        target_heads = [
+            CriticHead(
+                input_dim=encoder.output_dim + action_dim,
+                **asdict(self.config.critic_network_kwargs),
+            )
+            for _ in range(self.config.num_critics)
+        ]
+        self.critic_target = CriticEnsemble(encoder=encoder, ensemble=target_heads)
+        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
+
+        # TODO(Khalil): Investigate and fix torch.compile
+        # NOTE: torch.compile is disabled, policy does not converge when enabled.
+        if self.config.use_torch_compile:
+            self.critic_ensemble = torch.compile(self.critic_ensemble)
+            self.critic_target = torch.compile(self.critic_target)
+
+        self.discrete_critic_target = None
+        if self.policy_config.num_discrete_actions is not None:
+            self.discrete_critic_target = self._init_discrete_critic_target(encoder)
+
+    def _init_discrete_critic_target(self, encoder: GaussianActorObservationEncoder) -> DiscreteCritic:
+        """Build target discrete critic (main network is owned by the policy)."""
+        discrete_critic_target = DiscreteCritic(
+            encoder=encoder,
+            input_dim=encoder.output_dim,
+            output_dim=self.policy_config.num_discrete_actions,
+            **asdict(self.config.discrete_critic_network_kwargs),
+        )
+        # TODO(Khalil): Compile the discrete critic
+        discrete_critic_target.load_state_dict(self.policy.discrete_critic.state_dict())
+        return discrete_critic_target
+
+    def _init_temperature(self, continuous_action_dim: int) -> None:
+        """Set up temperature parameter (log_alpha) and target entropy."""
+        temp_init = self.config.temperature_init
+        self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
+
+        self.target_entropy = self.config.target_entropy
+        if self.target_entropy is None:
+            total_action_dim = continuous_action_dim + (
+                1 if self.policy_config.num_discrete_actions is not None else 0
+            )
+            self.target_entropy = -total_action_dim / 2
+
+    def _move_to_device(self) -> None:
+        self.policy.to(self._device)
+        self.critic_ensemble.to(self._device)
+        self.critic_target.to(self._device)
+        self.log_alpha = nn.Parameter(self.log_alpha.data.to(self._device))
+        if self.discrete_critic_target is not None:
+            self.discrete_critic_target.to(self._device)
+
+    @property
+    def temperature(self) -> float:
+        """Return the current temperature value, always in sync with log_alpha."""
+        return self.log_alpha.exp().item()
+
+    def _critic_forward(
+        self,
+        observations: dict[str, Tensor],
+        actions: Tensor,
+        use_target: bool = False,
+        observation_features: Tensor | None = None,
+    ) -> Tensor:
+        """Forward pass through a critic network ensemble
+
+        Args:
+            observations: Dictionary of observations
+            actions: Action tensor
+            use_target: If True, use target critics, otherwise use ensemble critics
+
+        Returns:
+            Tensor of Q-values from all critics
+        """
+
+        critics = self.critic_target if use_target else self.critic_ensemble
+        q_values = critics(observations, actions, observation_features)
+        return q_values
+
+    def _discrete_critic_forward(
+        self, observations, use_target=False, observation_features=None
+    ) -> torch.Tensor:
+        """Forward pass through a discrete critic network
+
+        Args:
+            observations: Dictionary of observations
+            use_target: If True, use target critics, otherwise use ensemble critics
+            observation_features: Optional pre-computed observation features to avoid recomputing encoder output
+
+        Returns:
+            Tensor of Q-values from the discrete critic network
+        """
+        discrete_critic = self.discrete_critic_target if use_target else self.policy.discrete_critic
+        q_values = discrete_critic(observations, observation_features)
+        return q_values
+
+    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        clip = self.config.grad_clip_norm
+
+        for _ in range(self.config.utd_ratio - 1):
+            batch = next(batch_iterator)
+            fb = self._prepare_forward_batch(batch, include_complementary_info=True)
+
+            loss_critic = self._compute_loss_critic(fb)
+            self.optimizers["critic"].zero_grad()
+            loss_critic.backward()
+            torch.nn.utils.clip_grad_norm_(self.critic_ensemble.parameters(), max_norm=clip)
+            self.optimizers["critic"].step()
+
+            if self.policy_config.num_discrete_actions is not None:
+                loss_dc = self._compute_loss_discrete_critic(fb)
+                self.optimizers["discrete_critic"].zero_grad()
+                loss_dc.backward()
+                torch.nn.utils.clip_grad_norm_(self.policy.discrete_critic.parameters(), max_norm=clip)
+                self.optimizers["discrete_critic"].step()
+
+            self._update_target_networks()
+
+        batch = next(batch_iterator)
+        fb = self._prepare_forward_batch(batch, include_complementary_info=False)
+
+        loss_critic = self._compute_loss_critic(fb)
+        self.optimizers["critic"].zero_grad()
+        loss_critic.backward()
+        critic_grad = torch.nn.utils.clip_grad_norm_(self.critic_ensemble.parameters(), max_norm=clip).item()
+        self.optimizers["critic"].step()
+
+        stats = TrainingStats(
+            losses={"loss_critic": loss_critic.item()},
+            grad_norms={"critic": critic_grad},
+        )
+
+        if self.policy_config.num_discrete_actions is not None:
+            loss_dc = self._compute_loss_discrete_critic(fb)
+            self.optimizers["discrete_critic"].zero_grad()
+            loss_dc.backward()
+            dc_grad = torch.nn.utils.clip_grad_norm_(
+                self.policy.discrete_critic.parameters(), max_norm=clip
+            ).item()
+            self.optimizers["discrete_critic"].step()
+            stats.losses["loss_discrete_critic"] = loss_dc.item()
+            stats.grad_norms["discrete_critic"] = dc_grad
+
+        if self._optimization_step % self.config.policy_update_freq == 0:
+            for _ in range(self.config.policy_update_freq):
+                loss_actor = self._compute_loss_actor(fb)
+                self.optimizers["actor"].zero_grad()
+                loss_actor.backward()
+                actor_grad = torch.nn.utils.clip_grad_norm_(
+                    self.policy.actor.parameters(), max_norm=clip
+                ).item()
+                self.optimizers["actor"].step()
+
+                loss_temp = self._compute_loss_temperature(fb)
+                self.optimizers["temperature"].zero_grad()
+                loss_temp.backward()
+                temp_grad = torch.nn.utils.clip_grad_norm_([self.log_alpha], max_norm=clip).item()
+                self.optimizers["temperature"].step()
+
+            stats.losses["loss_actor"] = loss_actor.item()
+            stats.losses["loss_temperature"] = loss_temp.item()
+            stats.grad_norms["actor"] = actor_grad
+            stats.grad_norms["temperature"] = temp_grad
+            stats.extra["temperature"] = self.temperature
+
+        self._update_target_networks()
+        self._optimization_step += 1
+        return stats
+
+    def _compute_loss_critic(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        actions = batch[ACTION]
+        rewards = batch["reward"]
+        next_observations = batch["next_state"]
+        done = batch["done"]
+        observation_features = batch.get("observation_feature")
+        next_observation_features = batch.get("next_observation_feature")
+
+        with torch.no_grad():
+            next_action_preds, next_log_probs, _ = self.policy.actor(
+                next_observations, next_observation_features
+            )
+
+            # 2- compute q targets
+            q_targets = self._critic_forward(
+                observations=next_observations,
+                actions=next_action_preds,
+                use_target=True,
+                observation_features=next_observation_features,
+            )
+
+            # subsample critics to prevent overfitting if use high UTD (update to date)
+            # TODO: Get indices before forward pass to avoid unnecessary computation
+            if self.config.num_subsample_critics is not None:
+                indices = torch.randperm(self.config.num_critics)
+                indices = indices[: self.config.num_subsample_critics]
+                q_targets = q_targets[indices]
+
+            # critics subsample size
+            min_q, _ = q_targets.min(dim=0)  # Get values from min operation
+            if self.config.use_backup_entropy:
+                min_q = min_q - (self.temperature * next_log_probs)
+
+            td_target = rewards + (1 - done) * self.config.discount * min_q
+
+        # 3- compute predicted qs
+        if self.policy_config.num_discrete_actions is not None:
+            # NOTE: We only want to keep the continuous action part
+            # In the buffer we have the full action space (continuous + discrete)
+            # We need to split them before concatenating them in the critic forward
+            actions: Tensor = actions[:, :DISCRETE_DIMENSION_INDEX]
+        q_preds = self._critic_forward(
+            observations=observations,
+            actions=actions,
+            use_target=False,
+            observation_features=observation_features,
+        )
+
+        # 4- Calculate loss
+        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
+        td_target_duplicate = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
+        # You compute the mean loss of the batch for each critic and then to compute the final loss you sum them up
+        critics_loss = (
+            F.mse_loss(
+                input=q_preds,
+                target=td_target_duplicate,
+                reduction="none",
+            ).mean(dim=1)
+        ).sum()
+        return critics_loss
+
+    def _compute_loss_discrete_critic(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        actions = batch[ACTION]
+        rewards = batch["reward"]
+        next_observations = batch["next_state"]
+        done = batch["done"]
+        observation_features = batch.get("observation_feature")
+        next_observation_features = batch.get("next_observation_feature")
+        complementary_info = batch.get("complementary_info")
+
+        # NOTE: We only want to keep the discrete action part
+        # In the buffer we have the full action space (continuous + discrete)
+        # We need to split them before concatenating them in the critic forward
+        actions_discrete: Tensor = actions[:, DISCRETE_DIMENSION_INDEX:].clone()
+        actions_discrete = torch.round(actions_discrete)
+        actions_discrete = actions_discrete.long()
+
+        discrete_penalties: Tensor | None = None
+        if complementary_info is not None:
+            discrete_penalties = complementary_info.get("discrete_penalty")
+
+        with torch.no_grad():
+            # For DQN, select actions using online network, evaluate with target network
+            next_discrete_qs = self._discrete_critic_forward(
+                next_observations, use_target=False, observation_features=next_observation_features
+            )
+            best_next_discrete_action = torch.argmax(next_discrete_qs, dim=-1, keepdim=True)
+
+            # Get target Q-values from target network
+            target_next_discrete_qs = self._discrete_critic_forward(
+                observations=next_observations,
+                use_target=True,
+                observation_features=next_observation_features,
+            )
+
+            # Use gather to select Q-values for best actions
+            target_next_discrete_q = torch.gather(
+                target_next_discrete_qs, dim=1, index=best_next_discrete_action
+            ).squeeze(-1)
+
+            # Compute target Q-value with Bellman equation
+            rewards_discrete = rewards
+            if discrete_penalties is not None:
+                rewards_discrete = rewards + discrete_penalties
+            target_discrete_q = rewards_discrete + (1 - done) * self.config.discount * target_next_discrete_q
+
+        # Get predicted Q-values for current observations
+        predicted_discrete_qs = self._discrete_critic_forward(
+            observations=observations, use_target=False, observation_features=observation_features
+        )
+
+        # Use gather to select Q-values for taken actions
+        predicted_discrete_q = torch.gather(predicted_discrete_qs, dim=1, index=actions_discrete).squeeze(-1)
+
+        # Compute MSE loss between predicted and target Q-values
+        discrete_critic_loss = F.mse_loss(input=predicted_discrete_q, target=target_discrete_q)
+        return discrete_critic_loss
+
+    def _compute_loss_actor(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        observation_features = batch.get("observation_feature")
+
+        actions_pi, log_probs, _ = self.policy.actor(observations, observation_features)
+
+        q_preds = self._critic_forward(
+            observations=observations,
+            actions=actions_pi,
+            use_target=False,
+            observation_features=observation_features,
+        )
+        min_q_preds = q_preds.min(dim=0)[0]
+
+        actor_loss = ((self.temperature * log_probs) - min_q_preds).mean()
+        return actor_loss
+
+    def _compute_loss_temperature(self, batch: dict[str, Any]) -> Tensor:
+        """Compute the temperature loss"""
+        observations = batch["state"]
+        observation_features = batch.get("observation_feature")
+
+        # calculate temperature loss
+        with torch.no_grad():
+            _, log_probs, _ = self.policy.actor(observations, observation_features)
+
+        temperature_loss = (-self.log_alpha.exp() * (log_probs + self.target_entropy)).mean()
+        return temperature_loss
+
+    def _update_target_networks(self) -> None:
+        """Update target networks with exponential moving average"""
+        for target_p, p in zip(
+            self.critic_target.parameters(), self.critic_ensemble.parameters(), strict=True
+        ):
+            target_p.data.copy_(
+                p.data * self.config.critic_target_update_weight
+                + target_p.data * (1.0 - self.config.critic_target_update_weight)
+            )
+        if self.policy_config.num_discrete_actions is not None:
+            for target_p, p in zip(
+                self.discrete_critic_target.parameters(),
+                self.policy.discrete_critic.parameters(),
+                strict=True,
+            ):
+                target_p.data.copy_(
+                    p.data * self.config.critic_target_update_weight
+                    + target_p.data * (1.0 - self.config.critic_target_update_weight)
+                )
+
+    def _prepare_forward_batch(
+        self, batch: BatchType, *, include_complementary_info: bool = True
+    ) -> dict[str, Any]:
+        observations = batch["state"]
+        next_observations = batch["next_state"]
+        observation_features, next_observation_features = self.get_observation_features(
+            observations, next_observations
+        )
+        forward_batch: dict[str, Any] = {
+            ACTION: batch[ACTION],
+            "reward": batch["reward"],
+            "state": observations,
+            "next_state": next_observations,
+            "done": batch["done"],
+            "observation_feature": observation_features,
+            "next_observation_feature": next_observation_features,
+        }
+        if include_complementary_info and "complementary_info" in batch:
+            forward_batch["complementary_info"] = batch["complementary_info"]
+        return forward_batch
+
+    def make_optimizers_and_scheduler(self) -> dict[str, Optimizer]:
+        """
+        Creates and returns optimizers for the actor, critic, and temperature components of a reinforcement learning policy.
+
+        This function sets up Adam optimizers for:
+        - The **actor network**, ensuring that only relevant parameters are optimized.
+        - The **critic ensemble**, which evaluates the value function.
+        - The **temperature parameter**, which controls the entropy in soft actor-critic (SAC)-like methods.
+
+        It also initializes a learning rate scheduler, though currently, it is set to `None`.
+
+        NOTE:
+        - If the encoder is shared, its parameters are excluded from the actor's optimization process.
+        - The policy's log temperature (`log_alpha`) is wrapped in a list to ensure proper optimization as a standalone tensor.
+
+        Args:
+            cfg: Configuration object containing hyperparameters.
+            policy (nn.Module): The policy model containing the actor, critic, and temperature components.
+
+        Returns:
+            A dictionary mapping component names ("actor", "critic", "temperature")
+            to their respective Adam optimizers.
+        """
+        actor_params = self.policy.get_optim_params()["actor"]
+        self.optimizers = {
+            "actor": torch.optim.Adam(actor_params, lr=self.config.actor_lr),
+            "critic": torch.optim.Adam(self.critic_ensemble.parameters(), lr=self.config.critic_lr),
+            "temperature": torch.optim.Adam([self.log_alpha], lr=self.config.temperature_lr),
+        }
+        if self.policy_config.num_discrete_actions is not None:
+            self.optimizers["discrete_critic"] = torch.optim.Adam(
+                self.policy.discrete_critic.parameters(), lr=self.config.critic_lr
+            )
+        return self.optimizers
+
+    def get_optimizers(self) -> dict[str, Optimizer]:
+        return self.optimizers
+
+    def get_weights(self) -> dict[str, Any]:
+        """Send actor + discrete-critic state dicts."""
+        state_dicts: dict[str, Any] = {
+            "policy": move_state_dict_to_device(self.policy.actor.state_dict(), device="cpu"),
+        }
+        if self.policy_config.num_discrete_actions is not None:
+            state_dicts["discrete_critic"] = move_state_dict_to_device(
+                self.policy.discrete_critic.state_dict(), device="cpu"
+            )
+        return state_dicts
+
+    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        """Load actor + discrete-critic weights into the policy."""
+        self.policy.load_actor_weights(weights, device=device)
+
+    def get_observation_features(
+        self, observations: Tensor, next_observations: Tensor
+    ) -> tuple[Tensor | None, Tensor | None]:
+        """
+        Get observation features from the policy encoder. It act as cache for the observation features.
+        when the encoder is frozen, the observation features are not updated.
+        We can save compute by caching the observation features.
+
+        Args:
+            policy: The policy model
+            observations: The current observations
+            next_observations: The next observations
+
+        Returns:
+            tuple: observation_features, next_observation_features
+        """
+
+        if self.policy.config.vision_encoder_name is None or not self.policy.config.freeze_vision_encoder:
+            return None, None
+
+        with torch.no_grad():
+            observation_features = self.policy.actor.encoder.get_cached_image_features(observations)
+            next_observation_features = self.policy.actor.encoder.get_cached_image_features(next_observations)
+
+        return observation_features, next_observation_features
+
+
+class CriticHead(nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dims: list[int],
+        activations: Callable[[torch.Tensor], torch.Tensor] | str = nn.SiLU(),
+        activate_final: bool = False,
+        dropout_rate: float | None = None,
+        init_final: float | None = None,
+        final_activation: Callable[[torch.Tensor], torch.Tensor] | str | None = None,
+    ):
+        super().__init__()
+        self.net = MLP(
+            input_dim=input_dim,
+            hidden_dims=hidden_dims,
+            activations=activations,
+            activate_final=activate_final,
+            dropout_rate=dropout_rate,
+            final_activation=final_activation,
+        )
+        self.output_layer = nn.Linear(in_features=hidden_dims[-1], out_features=1)
+        if init_final is not None:
+            nn.init.uniform_(self.output_layer.weight, -init_final, init_final)
+            nn.init.uniform_(self.output_layer.bias, -init_final, init_final)
+        else:
+            orthogonal_init()(self.output_layer.weight)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.output_layer(self.net(x))
+
+
+class CriticEnsemble(nn.Module):
+    """
+    CriticEnsemble wraps multiple CriticHead modules into an ensemble.
+
+    Args:
+        encoder (GaussianActorObservationEncoder): encoder for observations.
+        ensemble (List[CriticHead]): list of critic heads.
+        init_final (float | None): optional initializer scale for final layers.
+
+    Forward returns a tensor of shape (num_critics, batch_size) containing Q-values.
+    """
+
+    def __init__(
+        self,
+        encoder: GaussianActorObservationEncoder,
+        ensemble: list[CriticHead],
+        init_final: float | None = None,
+    ):
+        super().__init__()
+        self.encoder = encoder
+        self.init_final = init_final
+        self.critics = nn.ModuleList(ensemble)
+
+    def forward(
+        self,
+        observations: dict[str, torch.Tensor],
+        actions: torch.Tensor,
+        observation_features: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        device = get_device_from_parameters(self)
+        # Move each tensor in observations to device
+        observations = {k: v.to(device) for k, v in observations.items()}
+
+        obs_enc = self.encoder(observations, cache=observation_features)
+
+        inputs = torch.cat([obs_enc, actions], dim=-1)
+
+        # Loop through critics and collect outputs
+        q_values = []
+        for critic in self.critics:
+            q_values.append(critic(inputs))
+
+        # Stack outputs to match expected shape [num_critics, batch_size]
+        q_values = torch.stack([q.squeeze(-1) for q in q_values], dim=0)
+        return q_values

src/lerobot/rl/buffer.py

@@ -97,8 +97,8 @@ class ReplayBuffer:
         Args:
             capacity (int): Maximum number of transitions to store in the buffer.
             device (str): The device where the tensors will be moved when sampling ("cuda:0" or "cpu").
-            state_keys (List[str]): The list of keys that appear in `state` and `next_state`.
-            image_augmentation_function (Optional[Callable]): A function that takes a batch of images
+            state_keys (list[str]): The list of keys that appear in `state` and `next_state`.
+            image_augmentation_function (Callable | None): A function that takes a batch of images
                 and returns a batch of augmented images. If None, a default augmentation function is used.
             use_drq (bool): Whether to use the default DRQ image augmentation style, when sampling in the buffer.
             storage_device: The device (e.g. "cpu" or "cuda:0") where the data will be stored.
@@ -634,7 +634,7 @@ class ReplayBuffer:
                 If None, you must handle or define default keys.
 
         Returns:
-            transitions (List[Transition]):
+            transitions (list[Transition]):
                 A list of Transition dictionaries with the same length as `dataset`.
         """
         if state_keys is None:

src/lerobot/rl/crop_dataset_roi.py

@@ -176,11 +176,11 @@ def convert_lerobot_dataset_to_cropped_lerobot_dataset(
 
     Args:
         original_dataset (LeRobotDataset): The source dataset.
-        crop_params_dict (Dict[str, Tuple[int, int, int, int]]):
+        crop_params_dict (dict[str, Tuple[int, int, int, int]]):
             A dictionary mapping observation keys to crop parameters (top, left, height, width).
         new_repo_id (str): Repository id for the new dataset.
         new_dataset_root (str): The root directory where the new dataset will be written.
-        resize_size (Tuple[int, int], optional): The target size (height, width) after cropping.
+        resize_size (tuple[int, int], optional): The target size (height, width) after cropping.
             Defaults to (128, 128).
 
     Returns:

src/lerobot/rl/data_sources/__init__.py

@@ -0,0 +1,17 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .data_mixer import BatchType, DataMixer, OnlineOfflineMixer
+
+__all__ = ["BatchType", "DataMixer", "OnlineOfflineMixer"]

src/lerobot/rl/data_sources/data_mixer.py

@@ -0,0 +1,112 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import abc
+
+from lerobot.rl.algorithms.base import BatchType
+from lerobot.rl.buffer import ReplayBuffer, concatenate_batch_transitions
+
+
+class DataMixer(abc.ABC):
+    """Abstract interface for all data mixing strategies.
+
+    Subclasses must implement ``sample(batch_size)`` and may override
+    ``get_iterator`` for specialised iteration.
+    """
+
+    @abc.abstractmethod
+    def sample(self, batch_size: int) -> BatchType:
+        """Draw one batch of ``batch_size`` transitions."""
+        ...
+
+    def get_iterator(
+        self,
+        batch_size: int,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ):
+        """Infinite iterator that yields batches.
+
+        The default implementation repeatedly calls ``self.sample()``.
+        Subclasses with underlying buffer iterators (async prefetch)
+        should override this for better throughput.
+        """
+        while True:
+            yield self.sample(batch_size)
+
+
+class OnlineOfflineMixer(DataMixer):
+    """Mixes transitions from an online and an optional offline replay buffer.
+
+    When both buffers are present, each batch is constructed by sampling
+    ``ceil(batch_size * online_ratio)`` from the online buffer and the
+    remainder from the offline buffer, then concatenating.
+
+    This mixer assumes both online and offline buffers are present.
+    """
+
+    def __init__(
+        self,
+        online_buffer: ReplayBuffer,
+        offline_buffer: ReplayBuffer | None = None,
+        online_ratio: float = 1.0,
+    ):
+        if not 0.0 <= online_ratio <= 1.0:
+            raise ValueError(f"online_ratio must be in [0, 1], got {online_ratio}")
+        self.online_buffer = online_buffer
+        self.offline_buffer = offline_buffer
+        self.online_ratio = online_ratio
+
+    def sample(self, batch_size: int) -> BatchType:
+        if self.offline_buffer is None:
+            return self.online_buffer.sample(batch_size)
+
+        n_online = max(1, int(batch_size * self.online_ratio))
+        n_offline = batch_size - n_online
+
+        online_batch = self.online_buffer.sample(n_online)
+        offline_batch = self.offline_buffer.sample(n_offline)
+        return concatenate_batch_transitions(online_batch, offline_batch)
+
+    def get_iterator(
+        self,
+        batch_size: int,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ):
+        """Yield batches by composing buffer async iterators."""
+
+        n_online = max(1, int(batch_size * self.online_ratio))
+
+        online_iter = self.online_buffer.get_iterator(
+            batch_size=n_online,
+            async_prefetch=async_prefetch,
+            queue_size=queue_size,
+        )
+
+        if self.offline_buffer is None:
+            yield from online_iter
+            return
+
+        n_offline = batch_size - n_online
+        offline_iter = self.offline_buffer.get_iterator(
+            batch_size=n_offline,
+            async_prefetch=async_prefetch,
+            queue_size=queue_size,
+        )
+
+        while True:
+            yield concatenate_batch_transitions(next(online_iter), next(offline_iter))

src/lerobot/rl/eval_policy.py

@@ -17,9 +17,9 @@ import logging
 
 from lerobot.cameras import opencv  # noqa: F401
 from lerobot.configs import parser
-from lerobot.configs.train import TrainRLServerPipelineConfig
 from lerobot.datasets import LeRobotDataset
 from lerobot.policies import make_policy
+from lerobot.rl.train_rl import TrainRLServerPipelineConfig
 from lerobot.robots import (  # noqa: F401
     RobotConfig,
     make_robot_from_config,

src/lerobot/rl/gym_manipulator.py

@@ -39,6 +39,7 @@ from lerobot.processor import (
     GymHILAdapterProcessorStep,
     ImageCropResizeProcessorStep,
     InterventionActionProcessorStep,
+    LeaderArmInterventionStep,
     MapDeltaActionToRobotActionStep,
     MapTensorToDeltaActionDictStep,
     Numpy2TorchActionProcessorStep,
@@ -481,15 +482,41 @@ def make_processors(
     env_pipeline_steps.append(AddBatchDimensionProcessorStep())
     env_pipeline_steps.append(DeviceProcessorStep(device=device))
 
-    action_pipeline_steps = [
+    action_pipeline_steps: list = [
         AddTeleopActionAsComplimentaryDataStep(teleop_device=teleop_device),
         AddTeleopEventsAsInfoStep(teleop_device=teleop_device),
-        InterventionActionProcessorStep(
-            use_gripper=cfg.processor.gripper.use_gripper if cfg.processor.gripper is not None else False,
-            terminate_on_success=terminate_on_success,
-        ),
     ]
 
+    use_gripper_for_intervention = (
+        cfg.processor.gripper.use_gripper if cfg.processor.gripper is not None else False
+    )
+
+    # Leader-arm intervention: convert raw leader joints in `teleop_action`
+    # into a 4-D EE-delta dict before the override step consumes it. The same
+    # step also drives haptic follow on the leader (when `teleop_device` is a
+    # `SOLeaderFollower`) by pushing the follower joints back via send_action.
+    if (
+        getattr(cfg.processor, "control_mode", "gamepad") == "leader"
+        and cfg.processor.inverse_kinematics is not None
+        and kinematics_solver is not None
+    ):
+        action_pipeline_steps.append(
+            LeaderArmInterventionStep(
+                kinematics=kinematics_solver,
+                motor_names=motor_names,
+                end_effector_step_sizes=cfg.processor.inverse_kinematics.end_effector_step_sizes,
+                teleop_device=teleop_device,
+                use_gripper=use_gripper_for_intervention,
+            )
+        )
+
+    action_pipeline_steps.append(
+        InterventionActionProcessorStep(
+            use_gripper=use_gripper_for_intervention,
+            terminate_on_success=terminate_on_success,
+        )
+    )
+
     # Replace InverseKinematicsProcessor with new kinematic processors
     if cfg.processor.inverse_kinematics is not None and kinematics_solver is not None:
         # Add EE bounds and safety processor
@@ -689,74 +716,81 @@ def control_loop(
     episode_step = 0
     episode_start_time = time.perf_counter()
 
-    while episode_idx < cfg.dataset.num_episodes_to_record:
-        step_start_time = time.perf_counter()
+    try:
+        while episode_idx < cfg.dataset.num_episodes_to_record:
+            step_start_time = time.perf_counter()
 
-        # Create a neutral action (no movement)
-        neutral_action = torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32)
-        if use_gripper:
-            neutral_action = torch.cat([neutral_action, torch.tensor([1.0])])  # Gripper stay
-
-        # Use the new step function
-        transition = step_env_and_process_transition(
-            env=env,
-            transition=transition,
-            action=neutral_action,
-            env_processor=env_processor,
-            action_processor=action_processor,
-        )
-        terminated = transition.get(TransitionKey.DONE, False)
-        truncated = transition.get(TransitionKey.TRUNCATED, False)
-
-        if cfg.mode == "record":
-            observations = {
-                k: v.squeeze(0).cpu()
-                for k, v in transition[TransitionKey.OBSERVATION].items()
-                if isinstance(v, torch.Tensor)
-            }
-            # Use teleop_action if available, otherwise use the action from the transition
-            action_to_record = transition[TransitionKey.COMPLEMENTARY_DATA].get(
-                "teleop_action", transition[TransitionKey.ACTION]
-            )
-            frame = {
-                **observations,
-                ACTION: action_to_record.cpu(),
-                REWARD: np.array([transition[TransitionKey.REWARD]], dtype=np.float32),
-                DONE: np.array([terminated or truncated], dtype=bool),
-            }
+            # Create a neutral action (no movement)
+            neutral_action = torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32)
             if use_gripper:
-                discrete_penalty = transition[TransitionKey.COMPLEMENTARY_DATA].get("discrete_penalty", 0.0)
-                frame["complementary_info.discrete_penalty"] = np.array([discrete_penalty], dtype=np.float32)
+                neutral_action = torch.cat([neutral_action, torch.tensor([1.0])])  # Gripper stay
 
-            if dataset is not None:
-                frame["task"] = cfg.dataset.task
-                dataset.add_frame(frame)
-
-        episode_step += 1
-
-        # Handle episode termination
-        if terminated or truncated:
-            episode_time = time.perf_counter() - episode_start_time
-            logging.info(
-                f"Episode ended after {episode_step} steps in {episode_time:.1f}s with reward {transition[TransitionKey.REWARD]}"
+            transition = step_env_and_process_transition(
+                env=env,
+                transition=transition,
+                action=neutral_action,
+                env_processor=env_processor,
+                action_processor=action_processor,
             )
-            episode_step = 0
-            episode_idx += 1
+            terminated = transition.get(TransitionKey.DONE, False)
+            truncated = transition.get(TransitionKey.TRUNCATED, False)
 
-            if dataset is not None:
-                if transition[TransitionKey.INFO].get(TeleopEvents.RERECORD_EPISODE, False):
-                    logging.info(f"Re-recording episode {episode_idx}")
-                    dataset.clear_episode_buffer()
-                    episode_idx -= 1
-                else:
-                    logging.info(f"Saving episode {episode_idx}")
-                    dataset.save_episode()
+            if cfg.mode == "record":
+                observations = {
+                    k: v.squeeze(0).cpu()
+                    for k, v in transition[TransitionKey.OBSERVATION].items()
+                    if isinstance(v, torch.Tensor)
+                }
+                action_to_record = transition[TransitionKey.COMPLEMENTARY_DATA].get(
+                    "teleop_action", transition[TransitionKey.ACTION]
+                )
+                frame = {
+                    **observations,
+                    ACTION: action_to_record.cpu(),
+                    REWARD: np.array([transition[TransitionKey.REWARD]], dtype=np.float32),
+                    DONE: np.array([terminated or truncated], dtype=bool),
+                }
+                if use_gripper:
+                    discrete_penalty = transition[TransitionKey.COMPLEMENTARY_DATA].get(
+                        "discrete_penalty", 0.0
+                    )
+                    frame["complementary_info.discrete_penalty"] = np.array(
+                        [discrete_penalty], dtype=np.float32
+                    )
 
-            # Reset for new episode
-            transition = reset_and_build_transition(env, env_processor, action_processor)
+                if dataset is not None:
+                    frame["task"] = cfg.dataset.task
+                    dataset.add_frame(frame)
 
-        # Maintain fps timing
-        precise_sleep(max(dt - (time.perf_counter() - step_start_time), 0.0))
+            episode_step += 1
+
+            # Handle episode termination
+            if terminated or truncated:
+                episode_time = time.perf_counter() - episode_start_time
+                logging.info(
+                    f"Episode ended after {episode_step} steps in {episode_time:.1f}s with reward {transition[TransitionKey.REWARD]}"
+                )
+                episode_step = 0
+                episode_idx += 1
+
+                if dataset is not None:
+                    if transition[TransitionKey.INFO].get(TeleopEvents.RERECORD_EPISODE, False):
+                        logging.info(f"Re-recording episode {episode_idx}")
+                        dataset.clear_episode_buffer()
+                        episode_idx -= 1
+                    else:
+                        logging.info(f"Saving episode {episode_idx}")
+                        dataset.save_episode()
+
+                # Reset for new episode
+                transition = reset_and_build_transition(env, env_processor, action_processor)
+
+            # Maintain fps timing
+            precise_sleep(max(dt - (time.perf_counter() - step_start_time), 0.0))
+    finally:
+        if dataset is not None and dataset.writer is not None and dataset.writer.image_writer is not None:
+            logging.info("Waiting for image writer to finish...")
+            dataset.writer.image_writer.stop()
 
     if dataset is not None and cfg.dataset.push_to_hub:
         logging.info("Finalizing dataset before pushing to hub")

src/lerobot/rl/learner.py

@@ -51,6 +51,7 @@ import time
 from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
 from pprint import pformat
+from typing import Any
 
 import grpc
 import torch
@@ -68,10 +69,15 @@ from lerobot.common.train_utils import (
 )
 from lerobot.common.wandb_utils import WandBLogger
 from lerobot.configs import parser
-from lerobot.configs.train import TrainRLServerPipelineConfig
 from lerobot.datasets import LeRobotDataset, make_dataset
 from lerobot.policies import make_policy, make_pre_post_processors
-from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.rl.algorithms.base import RLAlgorithm
+from lerobot.rl.algorithms.factory import make_algorithm
+from lerobot.rl.buffer import ReplayBuffer
+from lerobot.rl.data_sources import OnlineOfflineMixer
+from lerobot.rl.process import ProcessSignalHandler
+from lerobot.rl.train_rl import TrainRLServerPipelineConfig
+from lerobot.rl.trainer import RLTrainer
 from lerobot.robots import so_follower  # noqa: F401
 from lerobot.teleoperators import gamepad, so_leader  # noqa: F401
 from lerobot.teleoperators.utils import TeleopEvents
@@ -91,15 +97,12 @@ from lerobot.utils.constants import (
 )
 from lerobot.utils.device_utils import get_safe_torch_device
 from lerobot.utils.random_utils import set_seed
-from lerobot.utils.transition import move_state_dict_to_device, move_transition_to_device
 from lerobot.utils.utils import (
     format_big_number,
     init_logging,
 )
 
-from .buffer import ReplayBuffer, concatenate_batch_transitions
 from .learner_service import MAX_WORKERS, SHUTDOWN_TIMEOUT, LearnerService
-from .process import ProcessSignalHandler
 
 
 @parser.wrap()
@@ -179,7 +182,7 @@ def train(cfg: TrainRLServerPipelineConfig, job_name: str | None = None):
 def start_learner_threads(
     cfg: TrainRLServerPipelineConfig,
     wandb_logger: WandBLogger | None,
-    shutdown_event: any,  # Event,
+    shutdown_event: Any,  # Event
 ) -> None:
     """
     Start the learner threads for training.
@@ -253,7 +256,7 @@ def start_learner_threads(
 def add_actor_information_and_train(
     cfg: TrainRLServerPipelineConfig,
     wandb_logger: WandBLogger | None,
-    shutdown_event: any,  # Event,
+    shutdown_event: Any,  # Event
     transition_queue: Queue,
     interaction_message_queue: Queue,
     parameters_queue: Queue,
@@ -266,8 +269,8 @@ def add_actor_information_and_train(
     - Transfers transitions from the actor to the replay buffer.
     - Logs received interaction messages.
     - Ensures training begins only when the replay buffer has a sufficient number of transitions.
-    - Samples batches from the replay buffer and performs multiple critic updates.
-    - Periodically updates the actor, critic, and temperature optimizers.
+    - Delegates training updates to an ``RLAlgorithm``.
+    - Periodically pushes updated weights to actors.
     - Logs training statistics, including loss values and optimization frequency.
 
     NOTE: This function doesn't have a single responsibility, it should be split into multiple functions
@@ -286,17 +289,13 @@ def add_actor_information_and_train(
     # of 7%
     device = get_safe_torch_device(try_device=cfg.policy.device, log=True)
     storage_device = get_safe_torch_device(try_device=cfg.policy.storage_device)
-    clip_grad_norm_value = cfg.policy.grad_clip_norm
     online_step_before_learning = cfg.policy.online_step_before_learning
-    utd_ratio = cfg.policy.utd_ratio
     fps = cfg.env.fps
     log_freq = cfg.log_freq
     save_freq = cfg.save_freq
-    policy_update_freq = cfg.policy.policy_update_freq
     policy_parameters_push_frequency = cfg.policy.actor_learner_config.policy_parameters_push_frequency
     saving_checkpoint = cfg.save_checkpoint
     online_steps = cfg.policy.online_steps
-    async_prefetch = cfg.policy.async_prefetch
 
     # Initialize logging for multiprocessing
     if not use_threads(cfg):
@@ -308,7 +307,7 @@ def add_actor_information_and_train(
 
     logging.info("Initializing policy")
 
-    policy: SACPolicy = make_policy(
+    policy = make_policy(
         cfg=cfg.policy,
         env_cfg=cfg.env,
     )
@@ -317,20 +316,17 @@ def add_actor_information_and_train(
 
     policy.train()
 
-    preprocessor, _postprocessor = make_pre_post_processors(
+    algorithm = make_algorithm(cfg=cfg.algorithm, policy=policy)
+
+    preprocessor, postprocessor = make_pre_post_processors(
         policy_cfg=cfg.policy,
         dataset_stats=cfg.policy.dataset_stats,
     )
 
-    push_actor_policy_to_queue(parameters_queue=parameters_queue, policy=policy)
-
+    # Push initial policy weights to actors
+    push_actor_policy_to_queue(parameters_queue=parameters_queue, algorithm=algorithm)
     last_time_policy_pushed = time.time()
 
-    optimizers, lr_scheduler = make_optimizers_and_scheduler(cfg=cfg, policy=policy)
-
-    # If we are resuming, we need to load the training state
-    resume_optimization_step, resume_interaction_step = load_training_state(cfg=cfg, optimizers=optimizers)
-
     log_training_info(cfg=cfg, policy=policy)
 
     replay_buffer = initialize_replay_buffer(cfg, device, storage_device)
@@ -343,21 +339,35 @@ def add_actor_information_and_train(
             device=device,
             storage_device=storage_device,
         )
-        batch_size: int = batch_size // 2  # We will sample from both replay buffer
+
+    # DataMixer: online-only or online/offline 50-50 mix
+    data_mixer = OnlineOfflineMixer(
+        online_buffer=replay_buffer,
+        offline_buffer=offline_replay_buffer,
+        online_ratio=cfg.online_ratio,
+    )
+    # RLTrainer owns the iterator, preprocessor, and creates optimizers.
+    trainer = RLTrainer(
+        algorithm=algorithm,
+        data_mixer=data_mixer,
+        batch_size=batch_size,
+        preprocessor=preprocessor,
+    )
+
+    # If we are resuming, we need to load the training state
+    optimizers = algorithm.get_optimizers()
+    resume_optimization_step, resume_interaction_step = load_training_state(cfg=cfg, optimizers=optimizers)
 
     logging.info("Starting learner thread")
     interaction_message = None
     optimization_step = resume_optimization_step if resume_optimization_step is not None else 0
+    algorithm.optimization_step = optimization_step
     interaction_step_shift = resume_interaction_step if resume_interaction_step is not None else 0
 
     dataset_repo_id = None
     if cfg.dataset is not None:
         dataset_repo_id = cfg.dataset.repo_id
 
-    # Initialize iterators
-    online_iterator = None
-    offline_iterator = None
-
     # NOTE: THIS IS THE MAIN LOOP OF THE LEARNER
     while True:
         # Exit the training loop if shutdown is requested
@@ -370,7 +380,6 @@ def add_actor_information_and_train(
             transition_queue=transition_queue,
             replay_buffer=replay_buffer,
             offline_replay_buffer=offline_replay_buffer,
-            device=device,
             dataset_repo_id=dataset_repo_id,
             shutdown_event=shutdown_event,
         )
@@ -387,180 +396,20 @@ def add_actor_information_and_train(
         if len(replay_buffer) < online_step_before_learning:
             continue
 
-        if online_iterator is None:
-            online_iterator = replay_buffer.get_iterator(
-                batch_size=batch_size, async_prefetch=async_prefetch, queue_size=2
-            )
-
-        if offline_replay_buffer is not None and offline_iterator is None:
-            offline_iterator = offline_replay_buffer.get_iterator(
-                batch_size=batch_size, async_prefetch=async_prefetch, queue_size=2
-            )
-
         time_for_one_optimization_step = time.time()
-        for _ in range(utd_ratio - 1):
-            # Sample from the iterators
-            batch = next(online_iterator)
 
-            if dataset_repo_id is not None:
-                batch_offline = next(offline_iterator)
-                batch = concatenate_batch_transitions(
-                    left_batch_transitions=batch, right_batch_transition=batch_offline
-                )
-
-            actions = batch[ACTION]
-            rewards = batch["reward"]
-            observations = preprocessor.process_observation(batch["state"])
-            next_observations = preprocessor.process_observation(batch["next_state"])
-            done = batch["done"]
-            check_nan_in_transition(observations=observations, actions=actions, next_state=next_observations)
-
-            observation_features, next_observation_features = get_observation_features(
-                policy=policy, observations=observations, next_observations=next_observations
-            )
-
-            # Create a batch dictionary with all required elements for the forward method
-            forward_batch = {
-                ACTION: actions,
-                "reward": rewards,
-                "state": observations,
-                "next_state": next_observations,
-                "done": done,
-                "observation_feature": observation_features,
-                "next_observation_feature": next_observation_features,
-                "complementary_info": batch["complementary_info"],
-            }
-
-            # Use the forward method for critic loss
-            critic_output = policy.forward(forward_batch, model="critic")
-
-            # Main critic optimization
-            loss_critic = critic_output["loss_critic"]
-            optimizers["critic"].zero_grad()
-            loss_critic.backward()
-            critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-                parameters=policy.critic_ensemble.parameters(), max_norm=clip_grad_norm_value
-            )
-            optimizers["critic"].step()
-
-            # Discrete critic optimization (if available)
-            if policy.config.num_discrete_actions is not None:
-                discrete_critic_output = policy.forward(forward_batch, model="discrete_critic")
-                loss_discrete_critic = discrete_critic_output["loss_discrete_critic"]
-                optimizers["discrete_critic"].zero_grad()
-                loss_discrete_critic.backward()
-                discrete_critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-                    parameters=policy.discrete_critic.parameters(), max_norm=clip_grad_norm_value
-                )
-                optimizers["discrete_critic"].step()
-
-            # Update target networks (main and discrete)
-            policy.update_target_networks()
-
-        # Sample for the last update in the UTD ratio
-        batch = next(online_iterator)
-
-        if dataset_repo_id is not None:
-            batch_offline = next(offline_iterator)
-            batch = concatenate_batch_transitions(
-                left_batch_transitions=batch, right_batch_transition=batch_offline
-            )
-
-        actions = batch[ACTION]
-        rewards = batch["reward"]
-        observations = preprocessor.process_observation(batch["state"])
-        next_observations = preprocessor.process_observation(batch["next_state"])
-        done = batch["done"]
-
-        check_nan_in_transition(observations=observations, actions=actions, next_state=next_observations)
-
-        observation_features, next_observation_features = get_observation_features(
-            policy=policy, observations=observations, next_observations=next_observations
-        )
-
-        # Create a batch dictionary with all required elements for the forward method
-        forward_batch = {
-            ACTION: actions,
-            "reward": rewards,
-            "state": observations,
-            "next_state": next_observations,
-            "done": done,
-            "observation_feature": observation_features,
-            "next_observation_feature": next_observation_features,
-        }
-
-        critic_output = policy.forward(forward_batch, model="critic")
-
-        loss_critic = critic_output["loss_critic"]
-        optimizers["critic"].zero_grad()
-        loss_critic.backward()
-        critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-            parameters=policy.critic_ensemble.parameters(), max_norm=clip_grad_norm_value
-        ).item()
-        optimizers["critic"].step()
-
-        # Initialize training info dictionary
-        training_infos = {
-            "loss_critic": loss_critic.item(),
-            "critic_grad_norm": critic_grad_norm,
-        }
-
-        # Discrete critic optimization (if available)
-        if policy.config.num_discrete_actions is not None:
-            discrete_critic_output = policy.forward(forward_batch, model="discrete_critic")
-            loss_discrete_critic = discrete_critic_output["loss_discrete_critic"]
-            optimizers["discrete_critic"].zero_grad()
-            loss_discrete_critic.backward()
-            discrete_critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-                parameters=policy.discrete_critic.parameters(), max_norm=clip_grad_norm_value
-            ).item()
-            optimizers["discrete_critic"].step()
-
-            # Add discrete critic info to training info
-            training_infos["loss_discrete_critic"] = loss_discrete_critic.item()
-            training_infos["discrete_critic_grad_norm"] = discrete_critic_grad_norm
-
-        # Actor and temperature optimization (at specified frequency)
-        if optimization_step % policy_update_freq == 0:
-            for _ in range(policy_update_freq):
-                # Actor optimization
-                actor_output = policy.forward(forward_batch, model="actor")
-                loss_actor = actor_output["loss_actor"]
-                optimizers["actor"].zero_grad()
-                loss_actor.backward()
-                actor_grad_norm = torch.nn.utils.clip_grad_norm_(
-                    parameters=policy.actor.parameters(), max_norm=clip_grad_norm_value
-                ).item()
-                optimizers["actor"].step()
-
-                # Add actor info to training info
-                training_infos["loss_actor"] = loss_actor.item()
-                training_infos["actor_grad_norm"] = actor_grad_norm
-
-                # Temperature optimization
-                temperature_output = policy.forward(forward_batch, model="temperature")
-                loss_temperature = temperature_output["loss_temperature"]
-                optimizers["temperature"].zero_grad()
-                loss_temperature.backward()
-                temp_grad_norm = torch.nn.utils.clip_grad_norm_(
-                    parameters=[policy.log_alpha], max_norm=clip_grad_norm_value
-                ).item()
-                optimizers["temperature"].step()
-
-                # Add temperature info to training info
-                training_infos["loss_temperature"] = loss_temperature.item()
-                training_infos["temperature_grad_norm"] = temp_grad_norm
-                training_infos["temperature"] = policy.temperature
+        # One training step (trainer owns data_mixer iterator; algorithm owns UTD loop)
+        stats = trainer.training_step()
 
         # Push policy to actors if needed
         if time.time() - last_time_policy_pushed > policy_parameters_push_frequency:
-            push_actor_policy_to_queue(parameters_queue=parameters_queue, policy=policy)
+            push_actor_policy_to_queue(parameters_queue=parameters_queue, algorithm=algorithm)
             last_time_policy_pushed = time.time()
 
-        # Update target networks (main and discrete)
-        policy.update_target_networks()
+        training_infos = stats.to_log_dict()
 
         # Log training metrics at specified intervals
+        optimization_step = algorithm.optimization_step
         if optimization_step % log_freq == 0:
             training_infos["replay_buffer_size"] = len(replay_buffer)
             if offline_replay_buffer is not None:
@@ -588,7 +437,6 @@ def add_actor_information_and_train(
                 custom_step_key="Optimization step",
             )
 
-        optimization_step += 1
         if optimization_step % log_freq == 0:
             logging.info(f"[LEARNER] Number of optimization step: {optimization_step}")
 
@@ -605,6 +453,8 @@ def add_actor_information_and_train(
                 offline_replay_buffer=offline_replay_buffer,
                 dataset_repo_id=dataset_repo_id,
                 fps=fps,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
             )
 
 
@@ -612,7 +462,7 @@ def start_learner(
     parameters_queue: Queue,
     transition_queue: Queue,
     interaction_message_queue: Queue,
-    shutdown_event: any,  # Event,
+    shutdown_event: Any,  # Event
     cfg: TrainRLServerPipelineConfig,
 ):
     """
@@ -689,6 +539,8 @@ def save_training_checkpoint(
     offline_replay_buffer: ReplayBuffer | None = None,
     dataset_repo_id: str | None = None,
     fps: int = 30,
+    preprocessor=None,
+    postprocessor=None,
 ) -> None:
     """
     Save training checkpoint and associated data.
@@ -712,6 +564,8 @@ def save_training_checkpoint(
         offline_replay_buffer: Optional offline replay buffer to save
         dataset_repo_id: Repository ID for dataset
         fps: Frames per second for dataset
+        preprocessor: Optional preprocessor pipeline to save
+        postprocessor: Optional postprocessor pipeline to save
     """
     logging.info(f"Checkpoint policy after step {optimization_step}")
     _num_digits = max(6, len(str(online_steps)))
@@ -728,6 +582,8 @@ def save_training_checkpoint(
         policy=policy,
         optimizer=optimizers,
         scheduler=None,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
     )
 
     # Save interaction step manually
@@ -765,58 +621,6 @@ def save_training_checkpoint(
     logging.info("Resume training")
 
 
-def make_optimizers_and_scheduler(cfg: TrainRLServerPipelineConfig, policy: nn.Module):
-    """
-    Creates and returns optimizers for the actor, critic, and temperature components of a reinforcement learning policy.
-
-    This function sets up Adam optimizers for:
-    - The **actor network**, ensuring that only relevant parameters are optimized.
-    - The **critic ensemble**, which evaluates the value function.
-    - The **temperature parameter**, which controls the entropy in soft actor-critic (SAC)-like methods.
-
-    It also initializes a learning rate scheduler, though currently, it is set to `None`.
-
-    NOTE:
-    - If the encoder is shared, its parameters are excluded from the actor's optimization process.
-    - The policy's log temperature (`log_alpha`) is wrapped in a list to ensure proper optimization as a standalone tensor.
-
-    Args:
-        cfg: Configuration object containing hyperparameters.
-        policy (nn.Module): The policy model containing the actor, critic, and temperature components.
-
-    Returns:
-        Tuple[Dict[str, torch.optim.Optimizer], Optional[torch.optim.lr_scheduler._LRScheduler]]:
-        A tuple containing:
-        - `optimizers`: A dictionary mapping component names ("actor", "critic", "temperature") to their respective Adam optimizers.
-        - `lr_scheduler`: Currently set to `None` but can be extended to support learning rate scheduling.
-
-    """
-    optimizer_actor = torch.optim.Adam(
-        params=[
-            p
-            for n, p in policy.actor.named_parameters()
-            if not policy.config.shared_encoder or not n.startswith("encoder")
-        ],
-        lr=cfg.policy.actor_lr,
-    )
-    optimizer_critic = torch.optim.Adam(params=policy.critic_ensemble.parameters(), lr=cfg.policy.critic_lr)
-
-    if cfg.policy.num_discrete_actions is not None:
-        optimizer_discrete_critic = torch.optim.Adam(
-            params=policy.discrete_critic.parameters(), lr=cfg.policy.critic_lr
-        )
-    optimizer_temperature = torch.optim.Adam(params=[policy.log_alpha], lr=cfg.policy.critic_lr)
-    lr_scheduler = None
-    optimizers = {
-        "actor": optimizer_actor,
-        "critic": optimizer_critic,
-        "temperature": optimizer_temperature,
-    }
-    if cfg.policy.num_discrete_actions is not None:
-        optimizers["discrete_critic"] = optimizer_discrete_critic
-    return optimizers, lr_scheduler
-
-
 # Training setup functions
 
 
@@ -1021,33 +825,6 @@ def initialize_offline_replay_buffer(
 # Utilities/Helpers functions
 
 
-def get_observation_features(
-    policy: SACPolicy, observations: torch.Tensor, next_observations: torch.Tensor
-) -> tuple[torch.Tensor | None, torch.Tensor | None]:
-    """
-    Get observation features from the policy encoder. It act as cache for the observation features.
-    when the encoder is frozen, the observation features are not updated.
-    We can save compute by caching the observation features.
-
-    Args:
-        policy: The policy model
-        observations: The current observations
-        next_observations: The next observations
-
-    Returns:
-        tuple: observation_features, next_observation_features
-    """
-
-    if policy.config.vision_encoder_name is None or not policy.config.freeze_vision_encoder:
-        return None, None
-
-    with torch.no_grad():
-        observation_features = policy.actor.encoder.get_cached_image_features(observations)
-        next_observation_features = policy.actor.encoder.get_cached_image_features(next_observations)
-
-    return observation_features, next_observation_features
-
-
 def use_threads(cfg: TrainRLServerPipelineConfig) -> bool:
     return cfg.policy.concurrency.learner == "threads"
 
@@ -1098,19 +875,11 @@ def check_nan_in_transition(
     return nan_detected
 
 
-def push_actor_policy_to_queue(parameters_queue: Queue, policy: nn.Module):
+def push_actor_policy_to_queue(parameters_queue: Queue, algorithm: RLAlgorithm) -> None:
     logging.debug("[LEARNER] Pushing actor policy to the queue")
 
     # Create a dictionary to hold all the state dicts
-    state_dicts = {"policy": move_state_dict_to_device(policy.actor.state_dict(), device="cpu")}
-
-    # Add discrete critic if it exists
-    if hasattr(policy, "discrete_critic") and policy.discrete_critic is not None:
-        state_dicts["discrete_critic"] = move_state_dict_to_device(
-            policy.discrete_critic.state_dict(), device="cpu"
-        )
-        logging.debug("[LEARNER] Including discrete critic in state dict push")
-
+    state_dicts = algorithm.get_weights()
     state_bytes = state_to_bytes(state_dicts)
     parameters_queue.put(state_bytes)
 
@@ -1134,9 +903,8 @@ def process_transitions(
     transition_queue: Queue,
     replay_buffer: ReplayBuffer,
     offline_replay_buffer: ReplayBuffer,
-    device: str,
     dataset_repo_id: str | None,
-    shutdown_event: any,
+    shutdown_event: Any,  # Event
 ):
     """Process all available transitions from the queue.
 
@@ -1144,7 +912,6 @@ def process_transitions(
         transition_queue: Queue for receiving transitions from the actor
         replay_buffer: Replay buffer to add transitions to
         offline_replay_buffer: Offline replay buffer to add transitions to
-        device: Device to move transitions to
         dataset_repo_id: Repository ID for dataset
         shutdown_event: Event to signal shutdown
     """
@@ -1153,8 +920,6 @@ def process_transitions(
         transition_list = bytes_to_transitions(buffer=transition_list)
 
         for transition in transition_list:
-            transition = move_transition_to_device(transition=transition, device=device)
-
             # Skip transitions with NaN values
             if check_nan_in_transition(
                 observations=transition["state"],
@@ -1177,7 +942,7 @@ def process_interaction_messages(
     interaction_message_queue: Queue,
     interaction_step_shift: int,
     wandb_logger: WandBLogger | None,
-    shutdown_event: any,
+    shutdown_event: Any,  # Event
 ) -> dict | None:
     """Process all available interaction messages from the queue.
 

src/lerobot/rl/train_rl.py

@@ -0,0 +1,54 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Top-level pipeline config for distributed RL training (actor / learner)."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from lerobot.configs.default import DatasetConfig
+from lerobot.configs.train import TrainPipelineConfig
+from lerobot.rl.algorithms.configs import RLAlgorithmConfig
+from lerobot.rl.algorithms.factory import make_algorithm_config
+from lerobot.rl.algorithms.sac import SACAlgorithmConfig  # noqa: F401
+
+
+@dataclass(kw_only=True)
+class TrainRLServerPipelineConfig(TrainPipelineConfig):
+    # NOTE: In RL, we don't need an offline dataset
+    # TODO: Make `TrainPipelineConfig.dataset` optional
+    dataset: DatasetConfig | None = None  # type: ignore[assignment] # because the parent class has made it's type non-optional
+
+    # Algorithm config (a `draccus.ChoiceRegistry` subclass selected by `type`,
+    # e.g. ``"type": "sac"``). When omitted, defaults to a SAC config with
+    # default hyperparameters. The top-level `policy` is injected into
+    # ``algorithm.policy_config`` at validation time.
+    algorithm: RLAlgorithmConfig | None = None
+
+    # Data mixer strategy name. Currently supports "online_offline".
+    mixer: str = "online_offline"
+    # Fraction sampled from online replay when using OnlineOfflineMixer.
+    online_ratio: float = 0.5
+
+    def validate(self) -> None:
+        super().validate()
+
+        if self.algorithm is None:
+            self.algorithm = make_algorithm_config("sac")
+
+        # The pipeline owns the policy config; inject it so the algorithm can
+        # introspect policy architecture (e.g. ``num_discrete_actions``).
+        if getattr(self.algorithm, "policy_config", None) is None:
+            self.algorithm.policy_config = self.policy

src/lerobot/rl/trainer.py

@@ -0,0 +1,103 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from collections.abc import Iterator
+from typing import Any
+
+from lerobot.rl.algorithms.base import BatchType, RLAlgorithm
+from lerobot.rl.algorithms.configs import TrainingStats
+from lerobot.rl.data_sources.data_mixer import DataMixer
+
+
+class RLTrainer:
+    """Unified training step orchestrator.
+
+    Holds the algorithm, a DataMixer, and an optional preprocessor.
+    """
+
+    def __init__(
+        self,
+        algorithm: RLAlgorithm,
+        data_mixer: DataMixer,
+        batch_size: int,
+        *,
+        preprocessor: Any | None = None,
+    ):
+        self.algorithm = algorithm
+        self.data_mixer = data_mixer
+        self.batch_size = batch_size
+        self._preprocessor = preprocessor
+
+        self._iterator: Iterator[BatchType] | None = None
+
+        self.algorithm.make_optimizers_and_scheduler()
+
+    def _build_data_iterator(self) -> Iterator[BatchType]:
+        """Create a fresh algorithm-configured iterator (optionally preprocessed)."""
+        raw = self.algorithm.configure_data_iterator(
+            data_mixer=self.data_mixer,
+            batch_size=self.batch_size,
+        )
+        if self._preprocessor is not None:
+            return _PreprocessedIterator(raw, self._preprocessor)
+        return raw
+
+    def reset_data_iterator(self) -> None:
+        """Discard the current iterator so it will be rebuilt lazily next step."""
+        self._iterator = None
+
+    def set_data_mixer(self, data_mixer: DataMixer, *, reset: bool = True) -> None:
+        """Swap the active data mixer, optionally resetting the iterator."""
+        self.data_mixer = data_mixer
+        if reset:
+            self.reset_data_iterator()
+
+    def training_step(self) -> TrainingStats:
+        """Run one training step (algorithm-agnostic)."""
+        if self._iterator is None:
+            self._iterator = self._build_data_iterator()
+        return self.algorithm.update(self._iterator)
+
+
+def preprocess_rl_batch(preprocessor: Any, batch: BatchType) -> BatchType:
+    """Apply policy preprocessing to RL observations only.
+
+    This mirrors the pre-refactor SAC learner behavior where actions are left
+    unchanged and only state/next_state observations are normalized.
+    """
+    observations = batch["state"]
+    next_observations = batch["next_state"]
+    batch["state"] = preprocessor.process_observation(observations)
+    batch["next_state"] = preprocessor.process_observation(next_observations)
+
+    return batch
+
+
+class _PreprocessedIterator:
+    """Iterator wrapper that preprocesses each sampled RL batch."""
+
+    __slots__ = ("_raw", "_preprocessor")
+
+    def __init__(self, raw_iterator: Iterator[BatchType], preprocessor: Any) -> None:
+        self._raw = raw_iterator
+        self._preprocessor = preprocessor
+
+    def __iter__(self) -> _PreprocessedIterator:
+        return self
+
+    def __next__(self) -> BatchType:
+        batch = next(self._raw)
+        return preprocess_rl_batch(self._preprocessor, batch)

src/lerobot/robots/reachy2/robot_reachy2.py

@@ -20,7 +20,7 @@ from typing import TYPE_CHECKING, Any
 
 from lerobot.cameras import make_cameras_from_configs
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.import_utils import _reachy2_sdk_available, require_package
+from lerobot.utils.import_utils import _reachy2_sdk_available
 
 from ..robot import Robot
 from ..utils import ensure_safe_goal_position
@@ -81,7 +81,6 @@ class Reachy2Robot(Robot):
     name = "reachy2"
 
     def __init__(self, config: Reachy2RobotConfig):
-        require_package("reachy2_sdk", extra="reachy2")
         super().__init__(config)
 
         self.config = config

src/lerobot/robots/unitree_g1/unitree_g1.py

@@ -27,7 +27,7 @@ import numpy as np
 
 from lerobot.cameras import make_cameras_from_configs
 from lerobot.types import RobotAction, RobotObservation
-from lerobot.utils.import_utils import _unitree_sdk_available, require_package
+from lerobot.utils.import_utils import _unitree_sdk_available
 
 from ..robot import Robot
 from .config_unitree_g1 import UnitreeG1Config
@@ -111,7 +111,6 @@ class UnitreeG1(Robot):
     name = "unitree_g1"
 
     def __init__(self, config: UnitreeG1Config):
-        require_package("unitree-sdk2py", extra="unitree_g1", import_name="unitree_sdk2py")
         super().__init__(config)
 
         logger.info("Initialize UnitreeG1...")

src/lerobot/scripts/convert_dataset_v21_to_v30.py

@@ -286,7 +286,7 @@ def convert_videos(root: Path, new_root: Path, video_file_size_in_mb: int):
     if len(set(num_eps_per_cam)) != 1:
         raise ValueError(f"All cams dont have same number of episodes ({num_eps_per_cam}).")
 
-    episodes_metadata = []
+    episods_metadata = []
     num_cameras = len(video_keys)
     num_episodes = num_eps_per_cam[0]
     for ep_idx in tqdm.tqdm(range(num_episodes), desc="convert videos"):
@@ -299,9 +299,9 @@ def convert_videos(root: Path, new_root: Path, video_file_size_in_mb: int):
         ep_dict = {}
         for cam_idx in range(num_cameras):
             ep_dict.update(eps_metadata_per_cam[cam_idx][ep_idx])
-        episodes_metadata.append(ep_dict)
+        episods_metadata.append(ep_dict)
 
-    return episodes_metadata
+    return episods_metadata
 
 
 def convert_videos_of_camera(root: Path, new_root: Path, video_key: str, video_file_size_in_mb: int):