feat(ci): extract task descriptions and embed in metrics artifact

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

.github/workflows/benchmark_tests.yml

@@ -0,0 +1,309 @@
+# 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.
+
+# Integration tests: build an isolated Docker image per benchmark and run a
+# 1-episode smoke eval. Each benchmark gets its own image so incompatible
+# dependency trees (e.g. hf-libero vs metaworld==3.0.0) can never collide.
+#
+# To add a new benchmark:
+#   1. Add docker/Dockerfile.benchmark.<name>  (install only lerobot[<name>])
+#   2. Copy one of the jobs below and adjust the image name and eval command.
+name: Benchmark Integration Tests
+
+on:
+  # Run manually from the Actions tab
+  workflow_dispatch:
+
+  # Run every Monday at 02:00 UTC.
+  schedule:
+    - cron: "0 2 * * 1"
+
+  push:
+    branches:
+      - feat/benchmark-ci
+      - main
+    paths:
+      - "src/lerobot/envs/**"
+      - "src/lerobot/scripts/lerobot_eval.py"
+      - "docker/Dockerfile.benchmark.*"
+      - ".github/workflows/benchmark_tests.yml"
+      - "pyproject.toml"
+
+  pull_request:
+    branches:
+      - main
+    paths:
+      - "src/lerobot/envs/**"
+      - "src/lerobot/scripts/lerobot_eval.py"
+      - "docker/Dockerfile.benchmark.*"
+      - ".github/workflows/benchmark_tests.yml"
+      - "pyproject.toml"
+
+permissions:
+  contents: read
+
+env:
+  UV_VERSION: "0.8.0"
+  PYTHON_VERSION: "3.12"
+
+# Cancel in-flight runs for the same branch/PR.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  # ── LIBERO ────────────────────────────────────────────────────────────────
+  # Isolated image: lerobot[libero] only (hf-libero, dm-control, mujoco chain)
+  libero-integration-test:
+    name: Libero — build image + 1-episode eval
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+          lfs: true
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          cache-binary: false
+
+      # Build the benchmark-specific image; layer cache lives in the runner's
+      # local Docker daemon — reused across re-runs on the same machine.
+      - name: Build Libero benchmark image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: docker/Dockerfile.benchmark.libero
+          push: false
+          load: true
+          tags: lerobot-benchmark-libero:ci
+          cache-from: type=local,src=/tmp/.buildx-cache-libero
+          cache-to: type=local,dest=/tmp/.buildx-cache-libero,mode=max
+
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          docker run --rm \
+            -e HF_HOME=/tmp/hf \
+            lerobot-benchmark-libero:ci \
+            bash -c "hf auth login --token '$HF_USER_TOKEN' --add-to-git-credential && hf auth whoami"
+
+      - name: Run Libero smoke eval (1 episode)
+        run: |
+          # Named container (no --rm) so we can docker cp artifacts out.
+          # Output to /tmp inside the container — user_lerobot cannot create
+          # root-level dirs like /artifacts.
+          docker run --name libero-eval --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            lerobot-benchmark-libero:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              lerobot-eval \
+                --policy.path=pepijn223/smolvla_libero \
+                --env.type=libero \
+                --env.task=libero_spatial \
+                --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
+              python3 /lerobot/scripts/ci/extract_task_descriptions.py \
+                --env libero --task libero_spatial \
+                --output /tmp/eval-artifacts/task_descriptions.json 2>/dev/null || true
+            "
+
+      - name: Copy Libero artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/libero-artifacts
+          docker cp libero-eval:/tmp/eval-artifacts/. /tmp/libero-artifacts/ 2>/dev/null || true
+          docker rm -f libero-eval || true
+
+      - name: Parse Libero eval metrics
+        if: always()
+        run: |
+          python3 scripts/ci/parse_eval_metrics.py \
+            --artifacts-dir /tmp/libero-artifacts \
+            --env libero \
+            --task libero_spatial \
+            --policy pepijn223/smolvla_libero
+
+      - name: Upload Libero rollout video
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: libero-rollout-video
+          path: /tmp/libero-artifacts/videos/
+          if-no-files-found: warn
+
+      - name: Upload Libero eval metrics
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: libero-metrics
+          path: /tmp/libero-artifacts/metrics.json
+          if-no-files-found: warn
+
+      # ── LIBERO TRAIN+EVAL SMOKE ──────────────────────────────────────────────
+      # Train SmolVLA for 1 step (batch_size=1, dataset episode 0 only) then
+      # immediately runs eval inside the training loop (eval_freq=1, 1 episode).
+      # Tests the full train→eval-within-training pipeline end-to-end.
+      - name: Run Libero train+eval smoke (1 step, eval_freq=1)
+        run: |
+          docker run --name libero-train-smoke --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            lerobot-benchmark-libero:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              accelerate launch --num_processes=1 \$(which lerobot-train) \
+                --policy.path=lerobot/smolvla_base \
+                --policy.load_vlm_weights=true \
+                --policy.scheduler_decay_steps=25000 \
+                --policy.freeze_vision_encoder=false \
+                --policy.train_expert_only=false \
+                --dataset.repo_id=lerobot/libero \
+                --dataset.episodes=[0] \
+                --dataset.use_imagenet_stats=false \
+                --env.type=libero \
+                --env.task=libero_spatial \
+                '--env.camera_name_mapping={\"agentview_image\": \"camera1\", \"robot0_eye_in_hand_image\": \"camera2\"}' \
+                --policy.empty_cameras=1 \
+                --output_dir=/tmp/train-smoke \
+                --steps=1 \
+                --batch_size=1 \
+                --eval_freq=1 \
+                --eval.n_episodes=1 \
+                --eval.batch_size=1 \
+                --eval.use_async_envs=false \
+                --save_freq=1 \
+                --policy.push_to_hub=false \
+                '--rename_map={\"observation.images.image\": \"observation.images.camera1\", \"observation.images.image2\": \"observation.images.camera2\"}'
+            "
+
+      - name: Copy Libero train-smoke artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/libero-train-smoke-artifacts
+          docker cp libero-train-smoke:/tmp/train-smoke/. /tmp/libero-train-smoke-artifacts/ 2>/dev/null || true
+          docker rm -f libero-train-smoke || true
+
+      - name: Upload Libero train-smoke eval video
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: libero-train-smoke-video
+          path: /tmp/libero-train-smoke-artifacts/eval/
+          if-no-files-found: warn
+
+  # ── METAWORLD ─────────────────────────────────────────────────────────────
+  # Isolated image: lerobot[metaworld] only (metaworld==3.0.0, mujoco>=3 chain)
+  metaworld-integration-test:
+    name: MetaWorld — build image + 1-episode eval
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+          lfs: true
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          cache-binary: false
+
+      - name: Build MetaWorld benchmark image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: docker/Dockerfile.benchmark.metaworld
+          push: false
+          load: true
+          tags: lerobot-benchmark-metaworld:ci
+          cache-from: type=local,src=/tmp/.buildx-cache-metaworld
+          cache-to: type=local,dest=/tmp/.buildx-cache-metaworld,mode=max
+
+      - name: Run MetaWorld smoke eval (1 episode)
+        run: |
+          docker run --name metaworld-eval --gpus all \
+            --shm-size=4g \
+            -e HF_HOME=/tmp/hf \
+            -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+            -e HF_HUB_DOWNLOAD_TIMEOUT=300 \
+            lerobot-benchmark-metaworld:ci \
+            bash -c "
+              hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+              lerobot-eval \
+                --policy.path=pepijn223/smolvla_metaworld \
+                --env.type=metaworld \
+                --env.task=metaworld-push-v3 \
+                --eval.batch_size=1 \
+                --eval.n_episodes=1 \
+                --eval.use_async_envs=false \
+                --policy.device=cuda \
+                '--rename_map={\"observation.image\": \"observation.images.camera1\"}' \
+                --policy.empty_cameras=2 \
+                --output_dir=/tmp/eval-artifacts
+              python3 /lerobot/scripts/ci/extract_task_descriptions.py \
+                --env metaworld --task metaworld-push-v3 \
+                --output /tmp/eval-artifacts/task_descriptions.json 2>/dev/null || true
+            "
+
+      - name: Copy MetaWorld artifacts from container
+        if: always()
+        run: |
+          mkdir -p /tmp/metaworld-artifacts
+          docker cp metaworld-eval:/tmp/eval-artifacts/. /tmp/metaworld-artifacts/ 2>/dev/null || true
+          docker rm -f metaworld-eval || true
+
+      - name: Parse MetaWorld eval metrics
+        if: always()
+        run: |
+          python3 scripts/ci/parse_eval_metrics.py \
+            --artifacts-dir /tmp/metaworld-artifacts \
+            --env metaworld \
+            --task metaworld-push-v3 \
+            --policy pepijn223/smolvla_metaworld
+
+      - name: Upload MetaWorld rollout video
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: metaworld-rollout-video
+          path: /tmp/metaworld-artifacts/videos/
+          if-no-files-found: warn
+
+      - name: Upload MetaWorld eval metrics
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: metaworld-metrics
+          path: /tmp/metaworld-artifacts/metrics.json
+          if-no-files-found: warn

.github/workflows/docker_publish.yml

@@ -12,8 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# This workflow handles nightly testing & docker images publishing.
-name: Nightly
+# This workflow handles Docker image publishing & testing.
+name: Docker Publish & Test
 permissions:
   contents: read
 
@@ -39,8 +39,8 @@ concurrency:
 
 jobs:
   # This job builds a CPU image for testing & distribution
-  build-docker-cpu-nightly:
-    name: Build CPU Docker for Nightly
+  build-docker-cpu:
+    name: Build CPU Docker
     runs-on:
       group: aws-general-8-plus
     if: github.repository == 'huggingface/lerobot'
@@ -74,8 +74,8 @@ jobs:
           tags: ${{ env.DOCKER_IMAGE_NAME_CPU }}
 
   # This job builds a GPU image for testing & distribution
-  build-docker-gpu-nightly:
-    name: Build GPU Docker for Nightly
+  build-docker-gpu:
+    name: Build GPU Docker
     runs-on:
       group: aws-general-8-plus
     if: github.repository == 'huggingface/lerobot'
@@ -109,9 +109,9 @@ jobs:
           tags: ${{ env.DOCKER_IMAGE_NAME_GPU }}
 
   # This job runs the E2E tests + pytest with all extras in the CPU image
-  nightly-cpu-tests:
-    name: Nightly CPU Tests
-    needs: [build-docker-cpu-nightly]
+  cpu-tests:
+    name: CPU Tests
+    needs: [build-docker-cpu]
     runs-on:
       group: aws-g6-4xlarge-plus
     env:
@@ -121,7 +121,7 @@ jobs:
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
       HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
-      image: ${{ needs.build-docker-cpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
+      image: ${{ needs.build-docker-cpu.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --shm-size "16gb"
       credentials:
         username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
@@ -142,9 +142,9 @@ jobs:
         run: make test-end-to-end
 
   # This job runs the E2E tests + pytest with all extras in the GPU image
-  nightly-gpu-tests:
-    name: Nightly GPU Tests
-    needs: [build-docker-gpu-nightly]
+  gpu-tests:
+    name: GPU Tests
+    needs: [build-docker-gpu]
     runs-on:
       group: aws-g6-4xlarge-plus
     env:
@@ -154,7 +154,7 @@ jobs:
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
       HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
-      image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
+      image: ${{ needs.build-docker-gpu.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
       credentials:
         username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
@@ -175,9 +175,9 @@ jobs:
         run: make test-end-to-end
 
   # This job runs multi-GPU training tests with 4 GPUs
-  nightly-multi-gpu-tests:
-    name: Nightly Multi-GPU Tests
-    needs: [build-docker-gpu-nightly]
+  multi-gpu-tests:
+    name: Multi-GPU Tests
+    needs: [build-docker-gpu]
     runs-on:
       group: aws-g4dn-12xlarge  # Instance with 4 GPUs
     env:
@@ -188,7 +188,7 @@ jobs:
       CUDA_VISIBLE_DEVICES: "0,1,2,3"
       HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
-      image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
+      image: ${{ needs.build-docker-gpu.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
       credentials:
         username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}

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

.github/workflows/documentation.yml

@@ -55,7 +55,7 @@ jobs:
       github.repository == 'huggingface/lerobot'
     permissions:
       contents: read
-    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@main
+    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
     with:
       commit_sha: ${{ github.sha }}
       package: lerobot
@@ -78,7 +78,7 @@ jobs:
     permissions:
       contents: read
       pull-requests: write
-    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@main
+    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@90b4ee2c10b81b5c1a6367c4e6fc9e2fb510a7e3  # main
     with:
       commit_sha: ${{ github.event.pull_request.head.sha }}
       pr_number: ${{ github.event.number }}

.github/workflows/fast_tests.yml

@@ -27,6 +27,7 @@ on:
       - "tests/**"
       - ".github/workflows/**"
       - "pyproject.toml"
+      - "uv.lock"
       - "Makefile"
   push:
     branches:
@@ -36,6 +37,7 @@ on:
       - "tests/**"
       - ".github/workflows/**"
       - "pyproject.toml"
+      - "uv.lock"
       - "Makefile"
 
 permissions:
@@ -63,7 +65,7 @@ jobs:
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
       HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
         with:
           persist-credentials: false
           lfs: true
@@ -81,14 +83,14 @@ jobs:
           libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev
 
       - name: Setup uv and Python
-        uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
+        uses: astral-sh/setup-uv@d0cc045d04ccac9d8b7881df0226f9e82c39688e  # v6
         with:
           enable-cache: true
           version: ${{ env.UV_VERSION }}
           python-version: ${{ env.PYTHON_VERSION }}
 
       - name: Install lerobot with test extras
-        run: uv sync --extra "test"
+        run: uv sync --locked --extra "test"
 
       - name: Login to Hugging Face
         if: env.HF_USER_TOKEN != ''

.github/workflows/full_tests.yml

@@ -29,6 +29,7 @@ on:
       - "tests/**"
       - ".github/workflows/**"
       - "pyproject.toml"
+      - "uv.lock"
       - "Makefile"
 
 permissions:
@@ -62,7 +63,7 @@ jobs:
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
       HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
         with:
           lfs: true
           persist-credentials: false
@@ -79,14 +80,14 @@ jobs:
           speech-dispatcher libgeos-dev portaudio19-dev
 
       - name: Setup uv and Python
-        uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
+        uses: astral-sh/setup-uv@d0cc045d04ccac9d8b7881df0226f9e82c39688e  # v6
         with:
           enable-cache: true
           version: ${{ env.UV_VERSION }}
           python-version: ${{ env.PYTHON_VERSION }}
 
       - name: Install lerobot with all extras
-        run: uv sync --extra all # TODO(Steven): Make flash-attn optional
+        run: uv sync --locked --extra all # TODO(Steven): Make flash-attn optional
 
       - name: Login to Hugging Face
         if: env.HF_USER_TOKEN != ''
@@ -136,21 +137,21 @@ jobs:
           sudo apt-get update
           sudo apt-get install git-lfs
           git lfs install
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
         with:
           lfs: true
           persist-credentials: false
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f  # v3
         with:
           cache-binary: false
       - name: Login to Docker Hub
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        uses: docker/login-action@c94ce9fb468520275223c153574b00df6fe4bcc9  # v3
         with:
           username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
           password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
       - name: Build and push Docker image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8  # v6
         with:
           context: .
           file: ./docker/Dockerfile.internal

.github/workflows/latest_deps_tests.yml

@@ -12,38 +12,81 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# This workflow handles full testing with unboud dependencies versions.
-name: Unbound Dependency Tests
+# This workflow tests the project against the latest upstream dependencies
+# (within pyproject.toml constraints) and opens a PR to update uv.lock
+# if the tests pass and the lockfile has changed.
+name: Latest Dependency Tests
 
 on:
   # Allows running this workflow manually from the Actions tab
   workflow_dispatch:
 
-  # Run on the 1st and 15th of every month at 09:00 UTC
-  # schedule:
-  #  - cron: '0 2 1,15 * *'
-
-permissions:
-  contents: read
+  # Runs at 03:00 UTC
+  schedule:
+    - cron: "0 3 * * *"
 
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
   PYTHON_VERSION: "3.12"
-  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound
+  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:latest-deps
 
-# Ensures that only the latest action is built, canceling older runs.
+# Ensures that only the latest run is active, canceling older runs.
 concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  group: ${{ github.workflow }}
   cancel-in-progress: true
 
 jobs:
 
-  # This job runs the E2E tests + pytest with all unbound extras
-  full-tests:
-    name: Full Unbound Tests
+  # This job upgrades the lockfile and checks if dependencies have changed
+  upgrade-lock:
+    name: Upgrade Lockfile
     runs-on: ubuntu-latest
     if: github.repository == 'huggingface/lerobot'
+    permissions:
+      contents: read
+    outputs:
+      changed: ${{ steps.diff.outputs.changed }}
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+
+      - name: Setup uv and Python
+        uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          version: ${{ env.UV_VERSION }}
+          python-version: ${{ env.PYTHON_VERSION }}
+
+      - name: Upgrade uv.lock
+        run: uv lock --upgrade
+
+      - name: Check for changes
+        id: diff
+        run: |
+          if git diff --quiet uv.lock; then
+            echo "changed=false" >> "$GITHUB_OUTPUT"
+            echo "uv.lock is up to date — no dependency changes."
+          else
+            echo "changed=true" >> "$GITHUB_OUTPUT"
+            echo "uv.lock has changed — running tests."
+          fi
+
+      - name: Upload updated lockfile
+        if: steps.diff.outputs.changed == 'true'
+        uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: uv-lock
+          path: uv.lock
+
+  # This job runs the full test suite with the upgraded dependencies
+  cpu-tests:
+    name: CPU Tests (Latest Deps)
+    needs: [upgrade-lock]
+    if: needs.upgrade-lock.outputs.changed == 'true'
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
     env:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
@@ -55,6 +98,11 @@ jobs:
           lfs: true
           persist-credentials: false
 
+      - name: Download updated lockfile
+        uses: actions/download-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: uv-lock
+
       # NOTE(Steven): Mount to `/mnt` to avoid the limited storage on `/home`. Consider cleaning default SDKs or using self-hosted runners for more space.
       # (As of 2024-06-10, the runner's `/home` has only 6.2 GB free—8% of its 72 GB total.)
       - name: Setup /mnt storage
@@ -73,34 +121,32 @@ jobs:
           version: ${{ env.UV_VERSION }}
           python-version: ${{ env.PYTHON_VERSION }}
 
-      - name: Unbound dependencies
-        run: |
-          sed -i 's/,[[:space:]]*<[0-9\.]*//g' pyproject.toml
-          echo "Dependencies unbound:" && cat pyproject.toml
-
       - name: Install lerobot with all extras
-        run: uv sync --extra all # TODO(Steven): Make flash-attn optional
+        run: uv sync --locked --extra all # TODO(Steven): Make flash-attn optional
+
       - name: Login to Hugging Face
         if: env.HF_USER_TOKEN != ''
         run: |
           uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
           uv run hf auth whoami
+
       - name: Run pytest (all extras)
-        run: uv run pytest tests -vv
+        run: uv run pytest tests -vv --maxfail=10
 
       - name: Run end-to-end tests
         run: uv run make test-end-to-end
 
-  # This job builds a GPU enabled image for testing
+  # This job builds a GPU-enabled Docker image with the upgraded dependencies
   build-and-push-docker:
     name: Build and Push Docker
+    needs: [upgrade-lock]
+    if: needs.upgrade-lock.outputs.changed == 'true'
+    permissions:
+      contents: read
     runs-on:
       group: aws-general-8-plus
-    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME }}
-    env:
-      GITHUB_REF: ${{ github.ref }}
     steps:
       - name: Install Git LFS
         run: |
@@ -111,6 +157,12 @@ jobs:
         with:
           lfs: true
           persist-credentials: false
+
+      - name: Download updated lockfile
+        uses: actions/download-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: uv-lock
+
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
         with:
@@ -127,14 +179,13 @@ jobs:
           file: ./docker/Dockerfile.internal
           push: true
           tags: ${{ env.DOCKER_IMAGE_NAME }}
-          build-args: |
-            UNBOUND_DEPS=true
 
-  # This job runs pytest with all unbound extras in a GPU enabled host
-  # It runs everytime a test image is created
+  # This job runs pytest with all extras on a GPU-enabled host
   gpu-tests:
-    name: GPU Unbound Tests
+    name: GPU Tests (Latest Deps)
     needs: [build-and-push-docker]
+    permissions:
+      contents: read
     runs-on:
       group: aws-g6-4xlarge-plus
     env:
@@ -159,17 +210,69 @@ jobs:
         run: |
           hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
           hf auth whoami
+      - name: Fix ptxas permissions
+        run: chmod +x /lerobot/.venv/lib/python3.12/site-packages/triton/backends/nvidia/bin/ptxas
       - name: Run pytest on GPU
-        run: pytest tests -vv
+        run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests
         run: make test-end-to-end
 
-  # This job deletes the test image recently created
-  # It runs everytime after the gpu-tests have finished
-  delete-unbound-image:
-    name: Delete Unbound Image
+  # This job creates or updates a PR with the upgraded lockfile
+  open-pr:
+    name: Open PR
+    needs: [cpu-tests, gpu-tests, upgrade-lock]
+    if: success() && needs.upgrade-lock.outputs.changed == 'true'
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+      pull-requests: write
+    env:
+      GH_TOKEN: ${{ secrets.UPDATE_LOCK_TOKEN }}
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+
+      - name: Download updated lockfile
+        uses: actions/download-artifact@v4 # zizmor: ignore[unpinned-uses]
+        with:
+          name: uv-lock
+
+      - name: Create or update PR
+        run: |
+          set -euo pipefail
+          BRANCH="auto/update-uv-lock"
+
+          git config user.name "github-actions[bot]"
+          git config user.email "github-actions[bot]@users.noreply.github.com"
+          git remote set-url origin "https://x-access-token:${GH_TOKEN}@github.com/${{ github.repository }}.git"
+
+          git checkout -B "$BRANCH"
+          git add uv.lock
+          git commit -m "chore(dependencies): update uv.lock"
+          git push --force origin "$BRANCH"
+
+          # Create PR only if one doesn't already exist for this branch
+          EXISTING_PR=$(gh pr list --head "$BRANCH" --state open --json number --jq '.[0].number')
+          if [ -z "$EXISTING_PR" ]; then
+            gh pr create \
+              --title "chore(dependencies): update uv.lock" \
+              --body "Automated update of \`uv.lock\` after successful latest dependency tests (CPU + GPU).
+
+          This PR upgrades all dependencies to their latest versions within the ranges specified in \`pyproject.toml\`." \
+              --head "$BRANCH" \
+              --base main
+          else
+            echo "PR #$EXISTING_PR already exists, branch has been updated."
+          fi
+
+  # This job deletes the temporary Docker image after tests complete
+  cleanup-docker:
+    name: Cleanup Docker Image
     needs: [gpu-tests, build-and-push-docker]
     if: always() && needs.build-and-push-docker.result == 'success'
+    permissions:
+      contents: read
     runs-on: ubuntu-latest
     steps:
       - name: Get Docker Hub Token and Delete Image
@@ -180,8 +283,7 @@ jobs:
           IMAGE_FULL: ${{ needs.build-and-push-docker.outputs.image_tag }}
         run: |
           IMAGE_NAME=$(echo "$IMAGE_FULL" | cut -d':' -f1)
-          IMAGE_TAG=$(echo "$IMAGE_FULL" | cut -d':' -f2)
-
+          IMAGE_TAG=$(echo "$IMAGE_FULL" | cut -d':' -f2-)
           echo "Attempting to delete image: $IMAGE_NAME:$IMAGE_TAG"
 
           TOKEN=$(curl -s -H "Content-Type: application/json" \

.github/workflows/quality.yml

@@ -43,16 +43,16 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout code
-        uses: actions/checkout@v6
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
         with:
           persist-credentials: false
 
       - name: Set up Python
-        uses: actions/setup-python@v6
+        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6
         with:
           python-version: '3.12'
 
       - name: Run pre-commit hooks
-        uses: pre-commit/action@v3.0.1 # zizmor: ignore[unpinned-uses]
+        uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd  # v3.0.1
         with:
           extra_args: --all-files --show-diff-on-failure --color=always

.github/workflows/release.yml

@@ -38,12 +38,12 @@ jobs:
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@v6
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
         with:
           persist-credentials: false
 
       - name: Set up Python
-        uses: actions/setup-python@v6
+        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6
         with:
           python-version: '3.12'
 
@@ -104,7 +104,7 @@ jobs:
       - name: Publish to TestPyPI for pre-releases
         # True for tags like 'v0.2.0-rc1'
         if: startsWith(github.ref, 'refs/tags/v') && contains(github.ref, '-')
-        uses: pypa/gh-action-pypi-publish@v1.13.0 # zizmor: ignore[unpinned-uses, use-trusted-publishing]
+        uses: pypa/gh-action-pypi-publish@ed0c53931b1dc9bd32cbe73a98c7f6766f8a527e  # v1.13.0
         with:
           repository-url: https://test.pypi.org/legacy/
           verbose: true
@@ -112,7 +112,7 @@ jobs:
 
       - name: Publish to PyPI
         if: startsWith(github.ref, 'refs/tags/v') && !contains(github.ref, '-')
-        uses: pypa/gh-action-pypi-publish@v1.13.0 # zizmor: ignore[unpinned-uses, use-trusted-publishing]
+        uses: pypa/gh-action-pypi-publish@ed0c53931b1dc9bd32cbe73a98c7f6766f8a527e  # v1.13.0
         with:
           verbose: true
           print-hash: true
@@ -127,7 +127,7 @@ jobs:
     env:
       MUJOCO_GL: egl
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
         with:
           lfs: true
           persist-credentials: false
@@ -137,7 +137,7 @@ jobs:
           git curl libglib2.0-0 libegl1-mesa-dev ffmpeg libusb-1.0-0-dev \
           speech-dispatcher libgeos-dev portaudio19-dev
       - name: Setup uv and Python
-        uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
+        uses: astral-sh/setup-uv@d0cc045d04ccac9d8b7881df0226f9e82c39688e  # v6
         with:
           enable-cache: true # zizmor: ignore[cache-poisoning]
           version: ${{ env.UV_VERSION }}

.github/workflows/security.yml

@@ -43,12 +43,12 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout code
-        uses: actions/checkout@v6 # zizmor: ignore[unpinned-uses]
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
         with:
           fetch-depth: 0
           persist-credentials: false
 
       - name: Secret Scanning
-        uses: trufflesecurity/trufflehog@v3.90.0  # zizmor: ignore[unpinned-uses]
+        uses: trufflesecurity/trufflehog@eafb8c5f6a06175141c27f17bcc17941853d0047  # v3.90.0
         with:
           extra_args: --only-verified

.gitignore

@@ -25,7 +25,6 @@ node_modules/
 
 # Lock files
 poetry.lock
-uv.lock
 Pipfile.lock
 
 ### Build & Distribution ###

README.md

@@ -4,7 +4,8 @@
 
 <div align="center">
 
-[![Tests](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml?query=branch%3Amain)
+[![Tests](https://github.com/huggingface/lerobot/actions/workflows/latest_deps_tests.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/latest_deps_tests.yml?query=branch%3Amain)
+[![Tests](https://github.com/huggingface/lerobot/actions/workflows/docker_publish.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/docker_publish.yml?query=branch%3Amain)
 [![Python versions](https://img.shields.io/pypi/pyversions/lerobot)](https://www.python.org/downloads/)
 [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://github.com/huggingface/lerobot/blob/main/LICENSE)
 [![Status](https://img.shields.io/pypi/status/lerobot)](https://pypi.org/project/lerobot/)

docker/Dockerfile.benchmark.libero

@@ -0,0 +1,89 @@
+# 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.
+
+# Isolated benchmark image for LIBERO integration tests.
+# Installs only lerobot[libero] so its dep tree (hf-libero, dm-control, mujoco)
+# cannot conflict with other benchmarks.
+#
+# Build:  docker build -f docker/Dockerfile.benchmark.libero -t lerobot-benchmark-libero .
+# Run:    docker run --gpus all --rm lerobot-benchmark-libero lerobot-eval ...
+
+ARG CUDA_VERSION=12.4.1
+ARG OS_VERSION=22.04
+FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
+
+ARG PYTHON_VERSION=3.12
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    MUJOCO_GL=egl \
+    PATH=/lerobot/.venv/bin:$PATH \
+    CUDA_VISIBLE_DEVICES=0 \
+    DEVICE=cuda
+
+# System deps — same set as Dockerfile.internal
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    software-properties-common build-essential git curl \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
+    cmake pkg-config ninja-build \
+    && add-apt-repository -y ppa:deadsnakes/ppa \
+    && apt-get update \
+    && apt-get install -y --no-install-recommends \
+       python${PYTHON_VERSION} \
+       python${PYTHON_VERSION}-venv \
+       python${PYTHON_VERSION}-dev \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && mv /root/.local/bin/uv /usr/local/bin/uv \
+    && useradd --create-home --shell /bin/bash user_lerobot \
+    && usermod -aG sudo user_lerobot \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /lerobot
+RUN chown -R user_lerobot:user_lerobot /lerobot
+USER user_lerobot
+
+ENV HOME=/home/user_lerobot \
+    HF_HOME=/home/user_lerobot/.cache/huggingface \
+    HF_LEROBOT_HOME=/home/user_lerobot/.cache/huggingface/lerobot \
+    TORCH_HOME=/home/user_lerobot/.cache/torch \
+    TRITON_CACHE_DIR=/home/user_lerobot/.cache/triton
+
+RUN uv venv --python python${PYTHON_VERSION}
+
+# Install only lerobot[libero] — completely isolated from metaworld's dep tree
+COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
+COPY --chown=user_lerobot:user_lerobot src/ src/
+
+RUN uv sync --locked --extra libero --extra smolvla --no-cache
+
+# Pre-download lerobot/libero-assets from HF Hub so nothing is fetched at
+# runtime (which times out on CI). Point the libero config at the cached path.
+# libero/libero/__init__.py calls input() when ~/.libero/config.yaml is missing,
+# so we write the config before any libero import can happen.
+RUN LIBERO_DIR=$(python${PYTHON_VERSION} -c \
+      "import importlib.util, os; s=importlib.util.find_spec('libero'); \
+       print(os.path.join(os.path.dirname(s.origin), 'libero'))") && \
+    mkdir -p /home/user_lerobot/.libero && \
+    python${PYTHON_VERSION} -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: ${LIBERO_DIR}/init_files\n" \
+    > /home/user_lerobot/.libero/config.yaml
+
+RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
+
+COPY --chown=user_lerobot:user_lerobot . .
+
+CMD ["/bin/bash"]

docker/Dockerfile.benchmark.metaworld

@@ -0,0 +1,74 @@
+# 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.
+
+# Isolated benchmark image for MetaWorld integration tests.
+# Installs only lerobot[metaworld] so its dep tree (metaworld==3.0.0, mujoco>=3)
+# cannot conflict with other benchmarks.
+#
+# Build:  docker build -f docker/Dockerfile.benchmark.metaworld -t lerobot-benchmark-metaworld .
+# Run:    docker run --gpus all --rm lerobot-benchmark-metaworld lerobot-eval ...
+
+ARG CUDA_VERSION=12.4.1
+ARG OS_VERSION=22.04
+FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
+
+ARG PYTHON_VERSION=3.12
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    MUJOCO_GL=egl \
+    PATH=/lerobot/.venv/bin:$PATH \
+    CUDA_VISIBLE_DEVICES=0 \
+    DEVICE=cuda
+
+# System deps — same set as Dockerfile.internal
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    software-properties-common build-essential git curl \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
+    cmake pkg-config ninja-build \
+    && add-apt-repository -y ppa:deadsnakes/ppa \
+    && apt-get update \
+    && apt-get install -y --no-install-recommends \
+       python${PYTHON_VERSION} \
+       python${PYTHON_VERSION}-venv \
+       python${PYTHON_VERSION}-dev \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && mv /root/.local/bin/uv /usr/local/bin/uv \
+    && useradd --create-home --shell /bin/bash user_lerobot \
+    && usermod -aG sudo user_lerobot \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /lerobot
+RUN chown -R user_lerobot:user_lerobot /lerobot
+USER user_lerobot
+
+ENV HOME=/home/user_lerobot \
+    HF_HOME=/home/user_lerobot/.cache/huggingface \
+    HF_LEROBOT_HOME=/home/user_lerobot/.cache/huggingface/lerobot \
+    TORCH_HOME=/home/user_lerobot/.cache/torch \
+    TRITON_CACHE_DIR=/home/user_lerobot/.cache/triton
+
+RUN uv venv --python python${PYTHON_VERSION}
+
+# Install only lerobot[metaworld] — completely isolated from libero's dep tree
+COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
+COPY --chown=user_lerobot:user_lerobot src/ src/
+
+RUN uv sync --locked --extra metaworld --extra smolvla --no-cache
+
+RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
+
+COPY --chown=user_lerobot:user_lerobot . .
+
+CMD ["/bin/bash"]

docker/Dockerfile.internal

@@ -73,17 +73,10 @@ ENV HOME=/home/user_lerobot \
 RUN uv venv --python python${PYTHON_VERSION}
 
 # Install Python dependencies for caching
-COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml README.md MANIFEST.in ./
+COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
 COPY --chown=user_lerobot:user_lerobot src/ src/
 
-ARG UNBOUND_DEPS=false
-
-RUN if [ "$UNBOUND_DEPS" = "true" ]; then \
-    sed -i 's/,[[:space:]]*<[0-9\.]*//g' pyproject.toml; \
-    echo "Dependencies unbound:" && cat pyproject.toml; \
-    fi
-
-RUN uv pip install --no-cache ".[all]"
+RUN uv sync --locked --extra all --no-cache
 
 RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
 

docker/Dockerfile.user

@@ -61,17 +61,10 @@ ENV HOME=/home/user_lerobot \
 RUN uv venv
 
 # Install Python dependencies for caching
-COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml README.md MANIFEST.in ./
+COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
 COPY --chown=user_lerobot:user_lerobot src/ src/
 
-ARG UNBOUND_DEPS=false
-
-RUN if [ "$UNBOUND_DEPS" = "true" ]; then \
-    sed -i 's/,[[:space:]]*<[0-9\.]*//g' pyproject.toml; \
-    echo "Dependencies unbound:" && cat pyproject.toml; \
-    fi
-
-RUN uv pip install --no-cache ".[all]"
+RUN uv sync --locked --extra all --no-cache
 
 # Copy the rest of the application code
 # Make sure to have the git-LFS files for testing

docker/README.md

@@ -0,0 +1,77 @@
+# Docker
+
+This directory contains Dockerfiles for running LeRobot in containerized environments. Both images are **built nightly from `main`** and published to Docker Hub with the full environment pre-baked — no dependency setup required.
+
+## Pre-built Images
+
+```bash
+# CPU-only image (based on Dockerfile.user)
+docker pull huggingface/lerobot-cpu:latest
+
+# GPU image with CUDA support (based on Dockerfile.internal)
+docker pull huggingface/lerobot-gpu:latest
+```
+
+## Quick Start
+
+The fastest way to start training is to pull the GPU image and run `lerobot-train` directly. This is the same environment used for all of our CI, so it is a well-tested, batteries-included setup.
+
+```bash
+docker run -it --rm --gpus all --shm-size 16gb huggingface/lerobot-gpu:latest
+
+# inside the container:
+lerobot-train --policy.type=act --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human
+```
+
+## Dockerfiles
+
+### `Dockerfile.user` (CPU)
+
+A lightweight image based on `python:3.12-slim`. Includes all Python dependencies and system libraries but does not include CUDA — there is no GPU support. Useful for exploring the codebase, running scripts, or working with robots, but not practical for training.
+
+### `Dockerfile.internal` (GPU)
+
+A CUDA-enabled image based on `nvidia/cuda`. This is the image for training — mostly used for internal interactions with the GPU cluster.
+
+## Usage
+
+### Running a pre-built image
+
+```bash
+# CPU
+docker run -it --rm huggingface/lerobot-cpu:latest
+
+# GPU
+docker run -it --rm --gpus all --shm-size 16gb huggingface/lerobot-gpu:latest
+```
+
+### Building locally
+
+From the repo root:
+
+```bash
+# CPU
+docker build -f docker/Dockerfile.user -t lerobot-user .
+docker run -it --rm lerobot-user
+
+# GPU
+docker build -f docker/Dockerfile.internal -t lerobot-internal .
+docker run -it --rm --gpus all --shm-size 16gb lerobot-internal
+```
+
+### Multi-GPU training
+
+To select specific GPUs, set `CUDA_VISIBLE_DEVICES` when launching the container:
+
+```bash
+# Use 4 GPUs
+docker run -it --rm --gpus all --shm-size 16gb \
+  -e CUDA_VISIBLE_DEVICES=0,1,2,3 \
+  huggingface/lerobot-gpu:latest
+```
+
+### USB device access (e.g. robots, cameras)
+
+```bash
+docker run -it --device=/dev/ -v /dev/:/dev/ --rm huggingface/lerobot-cpu:latest
+```

docs/source/_toctree.yml

@@ -17,6 +17,8 @@
     title: Train RL in Simulation
   - local: multi_gpu_training
     title: Multi GPU training
+  - local: hil_data_collection
+    title: Human In the Loop Data Collection
   - local: peft_training
     title: Training with PEFT (e.g., LoRA)
   - local: rename_map
@@ -69,13 +71,19 @@
     title: Environments from the Hub
   - local: envhub_leisaac
     title: Control & Train Robots in Sim (LeIsaac)
+  title: "Simulation"
+- sections:
+  - local: evaluation
+    title: Evaluation (lerobot-eval)
+  - local: adding_benchmarks
+    title: Adding a New Benchmark
+  - local: libero
+    title: LIBERO
+  - local: metaworld
+    title: Meta-World
   - local: envhub_isaaclab_arena
     title: NVIDIA IsaacLab Arena Environments
-  - local: libero
-    title: Using Libero
-  - local: metaworld
-    title: Using MetaWorld
-  title: "Simulation"
+  title: "Benchmarks"
 - sections:
   - local: introduction_processors
     title: Introduction to Robot Processors

docs/source/adding_benchmarks.mdx

@@ -0,0 +1,397 @@
+# Adding a New Benchmark
+
+This guide walks you through adding a new simulation benchmark to LeRobot. Follow the steps in order and use the existing benchmarks as templates.
+
+A benchmark in LeRobot is a set of [Gymnasium](https://gymnasium.farama.org/) environments that wrap a third-party simulator (like LIBERO or Meta-World) behind a standard `gym.Env` interface. The `lerobot-eval` CLI then runs evaluation uniformly across all benchmarks.
+
+## Existing benchmarks at a glance
+
+Before diving in, here is what is already integrated:
+
+| Benchmark      | Env file            | Config class       | Tasks               | Action dim   | Processor                    |
+| -------------- | ------------------- | ------------------ | ------------------- | ------------ | ---------------------------- |
+| LIBERO         | `envs/libero.py`    | `LiberoEnv`        | 130 across 5 suites | 7            | `LiberoProcessorStep`        |
+| Meta-World     | `envs/metaworld.py` | `MetaworldEnv`     | 50 (MT50)           | 4            | None                         |
+| IsaacLab Arena | Hub-hosted          | `IsaaclabArenaEnv` | Configurable        | Configurable | `IsaaclabArenaProcessorStep` |
+
+Use `src/lerobot/envs/libero.py` and `src/lerobot/envs/metaworld.py` as reference implementations.
+
+## How it all fits together
+
+### Data flow
+
+During evaluation, data moves through four stages:
+
+```
+1. gym.Env  ──→  raw observations (numpy dicts)
+
+2. Preprocessing  ──→  standard LeRobot keys + task description
+   (preprocess_observation in envs/utils.py, env.call("task_description"))
+
+3. Processors  ──→  env-specific then policy-specific transforms
+   (env_preprocessor, policy_preprocessor)
+
+4. Policy  ──→  select_action()  ──→  action tensor
+   then reverse: policy_postprocessor → env_postprocessor → numpy action → env.step()
+```
+
+Most benchmarks only need to care about stage 1 (producing observations in the right format) and optionally stage 3 (if env-specific transforms are needed).
+
+### Environment structure
+
+`make_env()` returns a nested dict of vectorized environments:
+
+```python
+dict[str, dict[int, gym.vector.VectorEnv]]
+#    ^suite       ^task_id
+```
+
+A single-task env (e.g. PushT) looks like `{"pusht": {0: vec_env}}`.
+A multi-task benchmark (e.g. LIBERO) looks like `{"libero_spatial": {0: vec0, 1: vec1, ...}, ...}`.
+
+### How evaluation runs
+
+All benchmarks are evaluated the same way by `lerobot-eval`:
+
+1. `make_env()` builds the nested `{suite: {task_id: VectorEnv}}` dict.
+2. `eval_policy_all()` iterates over every suite and task.
+3. For each task, it runs `n_episodes` rollouts via `rollout()`.
+4. Results are aggregated hierarchically: episode, task, suite, overall.
+5. Metrics include `pc_success` (success rate), `avg_sum_reward`, and `avg_max_reward`.
+
+The critical piece: your env must return `info["is_success"]` on every `step()` call. This is how the eval loop knows whether a task was completed.
+
+## What your environment must provide
+
+LeRobot does not enforce a strict observation schema. Instead it relies on a set of conventions that all benchmarks follow.
+
+### Env attributes
+
+Your `gym.Env` must set these attributes:
+
+| Attribute            | Type  | Why                                                  |
+| -------------------- | ----- | ---------------------------------------------------- |
+| `_max_episode_steps` | `int` | `rollout()` uses this to cap episode length          |
+| `task_description`   | `str` | Passed to VLA policies as a language instruction     |
+| `task`               | `str` | Fallback identifier if `task_description` is not set |
+
+### Success reporting
+
+Your `step()` and `reset()` must include `"is_success"` in the `info` dict:
+
+```python
+info = {"is_success": True}   # or False
+return observation, reward, terminated, truncated, info
+```
+
+### Observations
+
+The simplest approach is to map your simulator's outputs to the standard keys that `preprocess_observation()` already understands. Do this inside your `gym.Env` (e.g. in a `_format_raw_obs()` helper):
+
+| Your env should output    | LeRobot maps it to         | What it is                            |
+| ------------------------- | -------------------------- | ------------------------------------- |
+| `"pixels"` (single array) | `observation.image`        | Single camera image, HWC uint8        |
+| `"pixels"` (dict)         | `observation.images.<cam>` | Multiple cameras, each HWC uint8      |
+| `"agent_pos"`             | `observation.state`        | Proprioceptive state vector           |
+| `"environment_state"`     | `observation.env_state`    | Full environment state (e.g. PushT)   |
+| `"robot_state"`           | `observation.robot_state`  | Nested robot state dict (e.g. LIBERO) |
+
+If your simulator uses different key names, you have two options:
+
+1. **Recommended:** Rename them to the standard keys inside your `gym.Env` wrapper.
+2. **Alternative:** Write an env processor to transform observations after `preprocess_observation()` runs (see step 4 below).
+
+### Actions
+
+Actions are continuous numpy arrays in a `gym.spaces.Box`. The dimensionality depends on your benchmark (7 for LIBERO, 4 for Meta-World, etc.). Policies adapt to different action dimensions through their `input_features` / `output_features` config.
+
+### Feature declaration
+
+Each `EnvConfig` subclass declares two dicts that tell the policy what to expect:
+
+- `features` — maps feature names to `PolicyFeature(type, shape)` (e.g. action dim, image shape).
+- `features_map` — maps raw observation keys to LeRobot convention keys (e.g. `"agent_pos"` to `"observation.state"`).
+
+## Step by step
+
+<Tip>
+  At minimum, you need two files: a **gym.Env wrapper** and an **EnvConfig
+  subclass** with a `create_envs()` override. Everything else is optional or
+  documentation. No changes to `factory.py` are needed.
+</Tip>
+
+### Checklist
+
+| File                                      | Required | Why                                                          |
+| ----------------------------------------- | -------- | ------------------------------------------------------------ |
+| `src/lerobot/envs/<benchmark>.py`         | Yes      | Wraps the simulator as a standard gym.Env                    |
+| `src/lerobot/envs/configs.py`             | Yes      | Registers your benchmark and its `create_envs()` for the CLI |
+| `src/lerobot/processor/env_processor.py`  | Optional | Custom observation/action transforms                         |
+| `src/lerobot/envs/utils.py`               | Optional | Only if you need new raw observation keys                    |
+| `pyproject.toml`                          | Yes      | Declares benchmark-specific dependencies                     |
+| `docs/source/<benchmark>.mdx`             | Yes      | User-facing documentation page                               |
+| `docs/source/_toctree.yml`                | Yes      | Adds your page to the docs sidebar                           |
+| `docker/Dockerfile.benchmark.<benchmark>` | Yes      | Isolated Docker image for CI smoke tests                     |
+| `.github/workflows/benchmark_tests.yml`   | Yes      | CI job that builds the image and runs a 1-episode smoke eval |
+
+### 1. The gym.Env wrapper (`src/lerobot/envs/<benchmark>.py`)
+
+Create a `gym.Env` subclass that wraps the third-party simulator:
+
+```python
+class MyBenchmarkEnv(gym.Env):
+    metadata = {"render_modes": ["rgb_array"], "render_fps": <fps>}
+
+    def __init__(self, task_suite, task_id, ...):
+        super().__init__()
+        self.task = <task_name_string>
+        self.task_description = <natural_language_instruction>
+        self._max_episode_steps = <max_steps>
+        self.observation_space = spaces.Dict({...})
+        self.action_space = spaces.Box(low=..., high=..., shape=(...,), dtype=np.float32)
+
+    def reset(self, seed=None, **kwargs):
+        ...  # return (observation, info) — info must contain {"is_success": False}
+
+    def step(self, action: np.ndarray):
+        ...  # return (obs, reward, terminated, truncated, info) — info must contain {"is_success": <bool>}
+
+    def render(self):
+        ...  # return RGB image as numpy array
+
+    def close(self):
+        ...
+```
+
+**GPU-based simulators (e.g. MuJoCo with EGL rendering):** If your simulator allocates GPU/EGL contexts during `__init__`, defer that allocation to a `_ensure_env()` helper called on first `reset()`/`step()`. This avoids inheriting stale GPU handles when `AsyncVectorEnv` spawns worker processes. See `LiberoEnv._ensure_env()` for the pattern.
+
+Also provide a factory function that returns the nested dict structure:
+
+```python
+def create_mybenchmark_envs(
+    task: str,
+    n_envs: int,
+    gym_kwargs: dict | None = None,
+    env_cls: type | None = None,
+) -> dict[str, dict[int, Any]]:
+    """Create {suite_name: {task_id: VectorEnv}} for MyBenchmark."""
+    ...
+```
+
+See `create_libero_envs()` (multi-suite, multi-task) and `create_metaworld_envs()` (difficulty-grouped tasks) for reference.
+
+### 2. The config (`src/lerobot/envs/configs.py`)
+
+Register a config dataclass so users can select your benchmark with `--env.type=<name>`. Each config owns its environment creation and processor logic via two methods:
+
+- **`create_envs(n_envs, use_async_envs)`** — Returns `{suite: {task_id: VectorEnv}}`. The base class default uses `gym.make()` for single-task envs. Multi-task benchmarks override this.
+- **`get_env_processors()`** — Returns `(preprocessor, postprocessor)`. The base class default returns identity (no-op) pipelines. Override if your benchmark needs observation/action transforms.
+
+```python
+@EnvConfig.register_subclass("<benchmark_name>")
+@dataclass
+class MyBenchmarkEnvConfig(EnvConfig):
+    task: str = "<default_task>"
+    fps: int = <fps>
+    obs_type: str = "pixels_agent_pos"
+
+    features: dict[str, PolicyFeature] = field(default_factory=lambda: {
+        ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(<action_dim>,)),
+    })
+    features_map: dict[str, str] = field(default_factory=lambda: {
+        ACTION: ACTION,
+        "agent_pos": OBS_STATE,
+        "pixels": OBS_IMAGE,
+    })
+
+    def __post_init__(self):
+        ...  # populate features based on obs_type
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {"obs_type": self.obs_type, "render_mode": self.render_mode}
+
+    def create_envs(self, n_envs: int, use_async_envs: bool = True):
+        """Override for multi-task benchmarks or custom env creation."""
+        from lerobot.envs.<benchmark> import create_<benchmark>_envs
+        return create_<benchmark>_envs(task=self.task, n_envs=n_envs, ...)
+
+    def get_env_processors(self):
+        """Override if your benchmark needs observation/action transforms."""
+        from lerobot.processor.pipeline import PolicyProcessorPipeline
+        from lerobot.processor.env_processor import MyBenchmarkProcessorStep
+        return (
+            PolicyProcessorPipeline(steps=[MyBenchmarkProcessorStep()]),
+            PolicyProcessorPipeline(steps=[]),
+        )
+```
+
+Key points:
+
+- The `register_subclass` name is what users pass on the CLI (`--env.type=<name>`).
+- `features` tells the policy what the environment produces.
+- `features_map` maps raw observation keys to LeRobot convention keys.
+- **No changes to `factory.py` needed** — the factory delegates to `cfg.create_envs()` and `cfg.get_env_processors()` automatically.
+
+### 3. Env processor (optional — `src/lerobot/processor/env_processor.py`)
+
+Only needed if your benchmark requires observation transforms beyond what `preprocess_observation()` handles (e.g. image flipping, coordinate conversion). Define the processor step here and return it from `get_env_processors()` in your config (see step 2):
+
+```python
+@dataclass
+@ProcessorStepRegistry.register(name="<benchmark>_processor")
+class MyBenchmarkProcessorStep(ObservationProcessorStep):
+    def _process_observation(self, observation):
+        processed = observation.copy()
+        # your transforms here
+        return processed
+
+    def transform_features(self, features):
+        return features  # update if shapes change
+
+    def observation(self, observation):
+        return self._process_observation(observation)
+```
+
+See `LiberoProcessorStep` for a full example (image rotation, quaternion-to-axis-angle conversion).
+
+### 4. Dependencies (`pyproject.toml`)
+
+Add a new optional-dependency group:
+
+```toml
+mybenchmark = ["my-benchmark-pkg==1.2.3", "lerobot[scipy-dep]"]
+```
+
+Pinning rules:
+
+- **Always pin** benchmark packages to exact versions for reproducibility (e.g. `metaworld==3.0.0`).
+- **Add platform markers** when needed (e.g. `; sys_platform == 'linux'`).
+- **Pin fragile transitive deps** if known (e.g. `gymnasium==1.1.0` for Meta-World).
+- **Document constraints** in your benchmark doc page.
+
+Users install with:
+
+```bash
+pip install -e ".[mybenchmark]"
+```
+
+### 5. Documentation (`docs/source/<benchmark>.mdx`)
+
+Write a user-facing page following the template in the next section. See `docs/source/libero.mdx` and `docs/source/metaworld.mdx` for full examples.
+
+### 6. Table of contents (`docs/source/_toctree.yml`)
+
+Add your benchmark to the "Benchmarks" section:
+
+```yaml
+- sections:
+    - local: libero
+      title: LIBERO
+    - local: metaworld
+      title: Meta-World
+    - local: envhub_isaaclab_arena
+      title: NVIDIA IsaacLab Arena Environments
+    - local: <your_benchmark>
+      title: <Your Benchmark Name>
+  title: "Benchmarks"
+```
+
+### 7. CI smoke test (`docker/` + `.github/workflows/benchmark_tests.yml`)
+
+Each benchmark must have an isolated Docker image and a CI job that runs a 1-episode eval. This catches install-time regressions (broken transitive deps, import errors, interactive prompts) before they reach users.
+
+**Create `docker/Dockerfile.benchmark.<benchmark>`** — copy an existing one and change only the extra name:
+
+```dockerfile
+# Isolated benchmark image — installs lerobot[<benchmark>] only.
+# Build: docker build -f docker/Dockerfile.benchmark.<benchmark> -t lerobot-benchmark-<benchmark> .
+ARG CUDA_VERSION=12.4.1
+ARG OS_VERSION=22.04
+FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
+ARG PYTHON_VERSION=3.12
+# ... (same system deps as Dockerfile.benchmark.libero) ...
+RUN uv sync --locked --extra <benchmark> --no-cache
+```
+
+Each benchmark gets its own image so its dependency tree (pinned simulator packages, specific mujoco/scipy versions) cannot conflict with other benchmarks.
+
+**Add a job to `.github/workflows/benchmark_tests.yml`** — copy an existing job block and adjust:
+
+```yaml
+<benchmark>-integration-test:
+  name: <Benchmark> — build image + 1-episode eval
+  runs-on:
+    group: aws-g6-4xlarge-plus
+  env:
+    HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
+  steps:
+    - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      with:
+        persist-credentials: false
+        lfs: true
+    - name: Set up Docker Buildx
+      uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+      with:
+        cache-binary: false
+    - name: Build <Benchmark> image
+      uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+      with:
+        context: .
+        file: docker/Dockerfile.benchmark.<benchmark>
+        push: false
+        load: true
+        tags: lerobot-benchmark-<benchmark>:ci
+        cache-from: type=local,src=/tmp/.buildx-cache-<benchmark>
+        cache-to: type=local,dest=/tmp/.buildx-cache-<benchmark>,mode=max
+    - name: Run <Benchmark> smoke eval (1 episode)
+      run: |
+        docker run --rm --gpus all \
+          --shm-size=4g \
+          -e HF_HOME=/tmp/hf \
+          -e HF_USER_TOKEN="${HF_USER_TOKEN}" \
+          lerobot-benchmark-<benchmark>:ci \
+          bash -c "
+            hf auth login --token \"\$HF_USER_TOKEN\" --add-to-git-credential 2>/dev/null || true
+            lerobot-eval \
+              --policy.path=<hub_policy_path> \
+              --env.type=<benchmark> \
+              --env.task=<task> \
+              --eval.batch_size=1 \
+              --eval.n_episodes=1 \
+              --eval.use_async_envs=false \
+              --policy.device=cuda
+          "
+```
+
+**Tips:**
+
+- If the benchmark library prompts for user input on import (like LIBERO asking for a dataset folder), pass the relevant env var in the `docker run` command (e.g. `-e LIBERO_DATA_FOLDER=/tmp/libero_data`).
+- The job is scoped to only trigger on changes to `src/lerobot/envs/**`, `src/lerobot/scripts/lerobot_eval.py`, and the Dockerfiles — it won't run on unrelated PRs.
+
+## Verifying your integration
+
+After completing the steps above, confirm that everything works:
+
+1. **Install** — `pip install -e ".[mybenchmark]"` and verify the dependency group installs cleanly.
+2. **Smoke test env creation** — call `make_env()` with your config in Python, check that the returned dict has the expected `{suite: {task_id: VectorEnv}}` shape, and that `reset()` returns observations with the right keys.
+3. **Run a full eval** — `lerobot-eval --env.type=<name> --env.task=<task> --eval.n_episodes=1 --policy.path=<any_compatible_policy>` to exercise the full pipeline end-to-end. (`batch_size` defaults to auto-tuning based on CPU cores; pass `--eval.batch_size=1` to force a single environment.)
+4. **Check success detection** — verify that `info["is_success"]` flips to `True` when the task is actually completed. This is what the eval loop uses to compute success rates.
+5. **Add CI smoke test** — follow step 7 above to add a Dockerfile and CI job. This ensures the install stays green as dependencies evolve.
+
+## Writing a benchmark doc page
+
+Each benchmark `.mdx` page should include:
+
+- **Title and description** — 1-2 paragraphs on what the benchmark tests and why it matters.
+- **Links** — paper, GitHub repo, project website (if available).
+- **Overview image or GIF.**
+- **Available tasks** — table of task suites with counts and brief descriptions.
+- **Installation** — `pip install -e ".[<benchmark>]"` plus any extra steps (env vars, system packages).
+- **Evaluation** — recommended `lerobot-eval` command with `n_episodes` for reproducible results. `batch_size` defaults to auto; only specify it if needed. Include single-task and multi-task examples if applicable. See the [Evaluation guide](evaluation) for details.
+- **Policy inputs and outputs** — observation keys with shapes, action space description.
+- **Recommended evaluation episodes** — how many episodes per task is standard.
+- **Training** — example `lerobot-train` command.
+- **Reproducing published results** — link to pretrained model, eval command, results table (if available).
+
+See `docs/source/libero.mdx` and `docs/source/metaworld.mdx` for complete examples.

docs/source/evaluation.mdx

@@ -0,0 +1,162 @@
+# Evaluation
+
+`lerobot-eval` runs a trained policy on a simulation benchmark and reports success rate, reward, and (optionally) episode videos. It handles environment creation, batched rollouts, and metric aggregation automatically.
+
+## Quick start
+
+Evaluate a Hub-hosted policy on LIBERO:
+
+```bash
+lerobot-eval \
+    --policy.path=pepijn223/smolvla_libero \
+    --env.type=libero \
+    --env.task=libero_spatial \
+    --eval.n_episodes=10 \
+    --policy.device=cuda
+```
+
+Evaluate a local checkpoint:
+
+```bash
+lerobot-eval \
+    --policy.path=outputs/train/act_pusht/checkpoints/005000/pretrained_model \
+    --env.type=pusht \
+    --eval.n_episodes=10
+```
+
+`batch_size` defaults to **auto** (based on CPU cores). The script picks the right number of parallel environments for your machine.
+
+## Key flags
+
+| Flag                    | Default        | Description                                                                           |
+| ----------------------- | -------------- | ------------------------------------------------------------------------------------- |
+| `--policy.path`         | required       | Hub repo ID or local path to a pretrained model                                       |
+| `--env.type`            | required       | Benchmark name (`pusht`, `libero`, `metaworld`, etc.)                                 |
+| `--env.task`            | varies         | Task or suite name (e.g. `libero_spatial`, `libero_10`)                               |
+| `--eval.n_episodes`     | `50`           | Total episodes to run (across all tasks)                                              |
+| `--eval.batch_size`     | `0` (auto)     | Number of parallel environments. `0` = auto-tune from CPU cores                       |
+| `--eval.use_async_envs` | `true`         | Use `AsyncVectorEnv` (parallel stepping). Auto-downgrades to sync when `batch_size=1` |
+| `--policy.device`       | `cuda`         | Inference device                                                                      |
+| `--policy.use_amp`      | `false`        | Mixed-precision inference (saves VRAM, faster on Ampere+)                             |
+| `--seed`                | `1000`         | Random seed for reproducibility                                                       |
+| `--output_dir`          | auto-generated | Where to write results and videos                                                     |
+
+### Environment-specific flags
+
+Some benchmarks accept additional flags through `--env.*`:
+
+```bash
+# LIBERO: map simulator camera names to policy feature names
+--env.camera_name_mapping='{"agentview_image": "camera1", "robot0_eye_in_hand_image": "camera2"}'
+
+# Fill unused camera slots with zeros
+--policy.empty_cameras=1
+```
+
+See each benchmark's documentation ([LIBERO](libero), [Meta-World](metaworld)) for benchmark-specific flags.
+
+## How batch_size works
+
+`batch_size` controls how many environments run in parallel within a single `VectorEnv`:
+
+| `batch_size`  | Behavior                                                             |
+| ------------- | -------------------------------------------------------------------- |
+| `0` (default) | Auto-tune: `floor(cpu_cores × 0.7)`, capped by `n_episodes` and `64` |
+| `1`           | Single environment, synchronous. Useful for debugging                |
+| `N`           | N environments step in parallel via `AsyncVectorEnv`                 |
+
+When `batch_size > 1` and `use_async_envs=true`, each environment runs in its own subprocess via Gymnasium's `AsyncVectorEnv`. This parallelizes the simulation stepping (the main bottleneck), while the policy runs a single batched forward pass on GPU.
+
+**Example:** On a 16-core machine with `n_episodes=100`:
+
+- Auto batch_size = `floor(16 × 0.7)` = `11`
+- 11 environments step simultaneously → ~11× faster than sequential
+
+## Performance
+
+### AsyncVectorEnv (default)
+
+`AsyncVectorEnv` spawns one subprocess per environment. Each subprocess has its own simulator instance. While the policy computes actions on GPU, all environments step in parallel on CPU:
+
+```
+GPU:  [inference]....[inference]....[inference]....
+CPU:  [step × N]....................[step × N]......
+      ↑ parallel                   ↑ parallel
+```
+
+For GPU-based simulators (LIBERO, Meta-World), the environments use **lazy initialization**: the GPU/EGL context is created inside the worker subprocess on first `reset()`, not in the parent process. This avoids `EGL_BAD_CONTEXT` crashes from inheriting stale GPU handles across `fork()`.
+
+### Lazy task loading
+
+For multi-task benchmarks (e.g. LIBERO with 10 tasks), environments are wrapped in `_LazyAsyncVectorEnv` which defers worker creation until the task is actually evaluated. This keeps peak process count = `batch_size` instead of `n_tasks × batch_size`. After each task completes, workers are closed to free resources.
+
+### Tuning for speed
+
+| Situation                      | Recommendation                                        |
+| ------------------------------ | ----------------------------------------------------- |
+| Slow eval, low GPU utilization | Increase `batch_size` (or leave at auto)              |
+| Out of memory (system RAM)     | Decrease `batch_size`                                 |
+| Out of GPU memory              | Decrease `batch_size`, or use `--policy.use_amp=true` |
+| Debugging / single-stepping    | `--eval.batch_size=1 --eval.use_async_envs=false`     |
+
+## Output
+
+Results are written to `output_dir` (default: `outputs/eval/<date>/<time>_<job_name>/`):
+
+- `eval_info.json` — full metrics: per-episode, per-task, per-group, and overall aggregates
+- `videos/` — episode recordings (when `--eval.n_episodes_to_render > 0`)
+
+### Metrics
+
+| Metric           | Description                                                          |
+| ---------------- | -------------------------------------------------------------------- |
+| `pc_success`     | Success rate (%). Based on `info["is_success"]` from the environment |
+| `avg_sum_reward` | Mean cumulative reward per episode                                   |
+| `avg_max_reward` | Mean peak reward per episode                                         |
+| `n_episodes`     | Total episodes evaluated                                             |
+| `eval_s`         | Total wall-clock time                                                |
+| `eval_ep_s`      | Mean wall-clock time per episode                                     |
+
+## Multi-task evaluation
+
+For benchmarks with multiple tasks (LIBERO suites, Meta-World MT50), `lerobot-eval` automatically:
+
+1. Creates environments for all tasks in the selected suite(s)
+2. Evaluates each task sequentially (one task's workers at a time)
+3. Aggregates metrics per-task, per-group (suite), and overall
+
+```bash
+# Evaluate all 10 tasks in libero_spatial
+lerobot-eval \
+    --policy.path=pepijn223/smolvla_libero \
+    --env.type=libero \
+    --env.task=libero_spatial \
+    --eval.n_episodes=10
+
+# Evaluate multiple suites
+lerobot-eval \
+    --policy.path=pepijn223/smolvla_libero \
+    --env.type=libero \
+    --env.task="libero_spatial,libero_object" \
+    --eval.n_episodes=10
+```
+
+## API usage
+
+You can call the eval functions directly from Python:
+
+```python
+from lerobot.envs.factory import make_env
+from lerobot.policies.factory import make_policy
+from lerobot.scripts.lerobot_eval import eval_policy
+
+envs = make_env(env_cfg, n_envs=10)
+policy = make_policy(cfg=policy_cfg, env_cfg=env_cfg)
+
+metrics = eval_policy(
+    env=envs["libero_spatial"][0],
+    policy=policy,
+    n_episodes=10,
+)
+print(metrics["pc_success"])
+```

docs/source/groot.mdx

@@ -131,4 +131,4 @@ lerobot-record \
 
 ## License
 
-This model follows the **Apache 2.0 License**, consistent with the original [GR00T repository](https://github.com/NVIDIA/Isaac-GR00T).
+This model follows NVIDIA's proprietary license, consistent with the original [GR00T repository](https://github.com/NVIDIA/Isaac-GR00T). Future versions (starting from N1.7) will follow **Apache 2.0 License**.

docs/source/hil_data_collection.mdx

@@ -0,0 +1,269 @@
+# Human-In-the-Loop Data Collection
+
+Human-In-the-Loop (HIL) data collection lets you improve a trained policy by deploying it on a real robot while a human operator monitors and intervenes when needed. The intervention data (recovery movements and corrections) is recorded alongside autonomous segments, producing a richer training dataset that teaches the policy how to handle failures.
+
+---
+
+## Why Human-In-the-Loop?
+
+Standard behavioral cloning trains policies on successful demonstrations only. During deployment, small errors can compound and push the robot into states never seen during training (distribution shift). HIL data collection addresses this by:
+
+- Running the trained policy on the real robot
+- Having a human intervene when the robot is about to fail
+- Recording the human's recovery and correction as training data
+- Fine-tuning the policy on the combined dataset
+
+This produces a policy that not only knows how to perform the task, but also how to recover when things go wrong.
+
+---
+
+## How It Works
+
+During a HIL session, the human operator follows this loop within each episode:
+
+1. **Watch** the policy run autonomously
+2. **Pause** when failure is imminent, the robot holds its position
+3. **Take control** and teleoperate the robot back to a good state (recovery), then correct the behavior
+4. **Return control to the policy**, the policy resumes autonomous execution
+5. Repeat steps 2–4 as many times as needed during the episode
+6. **End the episode** when the task is complete, save and move on to the next rollout
+
+Both autonomous and human-controlled segments are recorded. The policy and human can alternate control multiple times within a single episode, and the episode continues from the current state after each handoff (no reset required just because intervention happened). This captures autonomous execution, recovery, and correction in one continuous trajectory. After collection, the combined dataset (original demonstrations + HIL data) is used to fine-tune the policy.
+
+This process can be repeated iteratively: deploy, collect, fine-tune, repeat. Each round targets the current policy's failure modes.
+
+```
+┌─────────────────────────────────────────────────────────────────────────┐
+│  Policy v0 (trained on demos)                                           │
+│       ↓                                                                 │
+│  HIL Collection (target current failure modes) → Fine-tune → Policy v1  │
+│       ↓                                                                 │
+│  HIL Collection (target new failure modes) → Fine-tune → Policy v2      │
+│       ↓                                                                 │
+│  ... (repeat until satisfactory performance)                            │
+└─────────────────────────────────────────────────────────────────────────┘
+```
+
+---
+
+## Hardware Requirements
+
+### Teleoperator Requirements
+
+The `examples/hil` HIL scripts require **teleoperators with active motors** that can:
+
+- Enable/disable torque programmatically
+- Move to target positions (to mirror the robot state when pausing)
+
+**Compatible teleoperators in the current `examples/hil` scripts:**
+
+- `openarm_mini` - OpenArm Mini
+- `so_leader` - SO100 / SO101 leader arm
+
+> [!IMPORTANT]
+> The provided `examples/hil` commands default to `bi_openarm_follower` + `openarm_mini`.
+> `so_follower` + `so_leader` configs are also registered and can be used via CLI flags.
+
+---
+
+## Script
+
+A single script handles both synchronous and RTC-based inference. Toggle RTC with `--rtc.enabled=true`:
+
+| Mode                     | Flag                 | Models                |
+| ------------------------ | -------------------- | --------------------- |
+| Standard (default)       | _(no flag needed)_   | ACT, Diffusion Policy |
+| Real-Time Chunking (RTC) | `--rtc.enabled=true` | Pi0, Pi0.5, SmolVLA   |
+
+---
+
+## Step-by-Step Guide
+
+### Step 1: Pre-train a Base Policy
+
+First, train a policy on your demonstration dataset:
+
+```bash
+python src/lerobot/scripts/lerobot_train.py \
+    --dataset.repo_id=your-username/demo-dataset \
+    --policy.type=pi0 \
+    --output_dir=outputs/pretrain \
+    --batch_size=32 \
+    --steps=50000
+```
+
+### Step 2: Collect HIL Data
+
+**Standard inference (ACT, Diffusion Policy):**
+
+```bash
+python examples/hil/hil_data_collection.py \
+    --robot.type=bi_openarm_follower \
+    --robot.left_arm_config.port=can1 \
+    --robot.left_arm_config.side=left \
+    --robot.right_arm_config.port=can0 \
+    --robot.right_arm_config.side=right \
+    --robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}}' \
+    --teleop.type=openarm_mini \
+    --teleop.port_left=/dev/ttyACM0 \
+    --teleop.port_right=/dev/ttyACM1 \
+    --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --dataset.repo_id=your-username/hil-dataset \
+    --dataset.single_task="Fold the T-shirt properly" \
+    --dataset.fps=30 \
+    --dataset.episode_time_s=1000 \
+    --dataset.num_episodes=50 \
+    --interpolation_multiplier=2
+```
+
+**With RTC for large models (Pi0, Pi0.5, SmolVLA):**
+
+For models with high inference latency, enable RTC for smooth execution:
+
+```bash
+python examples/hil/hil_data_collection.py \
+    --rtc.enabled=true \
+    --rtc.execution_horizon=20 \
+    --rtc.max_guidance_weight=5.0 \
+    --rtc.prefix_attention_schedule=LINEAR \
+    --robot.type=bi_openarm_follower \
+    --robot.left_arm_config.port=can1 \
+    --robot.left_arm_config.side=left \
+    --robot.right_arm_config.port=can0 \
+    --robot.right_arm_config.side=right \
+    --robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}}' \
+    --teleop.type=openarm_mini \
+    --teleop.port_left=/dev/ttyACM0 \
+    --teleop.port_right=/dev/ttyACM1 \
+    --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --dataset.repo_id=your-username/hil-rtc-dataset \
+    --dataset.single_task="Fold the T-shirt properly" \
+    --dataset.fps=30 \
+    --dataset.episode_time_s=1000 \
+    --dataset.num_episodes=50 \
+    --interpolation_multiplier=3
+```
+
+**Controls (Conceptual):**
+
+The interaction model is:
+
+- **Pause input**: pause autonomous policy execution
+- **Takeover input**: transfer control to the human operator and record intervention data
+- **Return-to-policy input**: hand control back to the policy and continue the same episode
+- **Episode control inputs**: save/re-record/stop/reset as needed
+
+Exact key/pedal bindings can differ across scripts and hardware integrations. Use each script's printed controls as the source of truth for the concrete mapping on your setup.
+
+**The HIL Protocol:**
+
+1. Watch the policy run autonomously (teleop is idle/free)
+2. When you see imminent failure, trigger the **pause input**
+   - Policy stops
+   - Teleoperator moves to match robot position (torque enabled)
+   - No frames recorded during pause
+3. Trigger the **takeover input** to take control
+   - Teleoperator torque disabled, free to move
+   - **Recovery**: Teleoperate the robot back to a good state
+   - **Correction**: Correct the behavior
+   - All movements are recorded
+4. Trigger the **return-to-policy input**
+   - Policy resumes autonomous execution from the current state
+   - You can intervene again at any time (repeat steps 2–4)
+5. End and save the episode when the task is complete (or episode time limit is reached)
+6. **Reset**: Teleop moves to robot position, you can move the robot to the starting position
+7. Start the next episode
+
+**Foot Pedal Setup (Linux):**
+
+If using a USB foot pedal (PCsensor FootSwitch), ensure access:
+
+```bash
+sudo setfacl -m u:$USER:rw /dev/input/by-id/usb-PCsensor_FootSwitch-event-kbd
+```
+
+### Step 3: Fine-tune the Policy
+
+Fine-tune on the **combined** dataset (`demo-dataset` + `hil-dataset` merged together):
+
+```bash
+python src/lerobot/scripts/lerobot_train.py \
+    --dataset.repo_id=your-username/hil-dataset \
+    --policy.type=pi0 \
+    --policy.pretrained_path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --output_dir=outputs/hil_finetune \
+    --steps=20000
+```
+
+Then deploy the fine-tuned policy and repeat from Step 2 to target its remaining failure modes.
+
+---
+
+## Tips for Effective HIL Collection
+
+### When to Intervene
+
+Intervene when you see:
+
+- Robot about to make an irreversible mistake
+- Robot hesitating or showing uncertain behavior
+- Robot deviating from the expected trajectory
+
+### Recovery: Teleoperating Back to a Good State
+
+During recovery, teleoperate the robot back to a state where:
+
+- The robot is in a familiar, in-distribution configuration
+- The current subtask can still be completed
+- The recovery trajectory itself is informative training data
+
+### Quality of Corrections
+
+During correction:
+
+- Provide **confident, clean** trajectories
+- Complete the current subtask fully
+- Don't overcorrect or add unnecessary movements
+
+---
+
+## Related Work
+
+This HIL data collection approach builds on ideas from interactive imitation learning:
+
+- **DAgger** (Ross et al., 2011) introduced the core idea: instead of only training on expert demonstrations, query the expert for corrections on states the _learner_ visits. This breaks the compounding-error cycle of standard behavioral cloning by iteratively collecting on-policy data.
+
+- **HG-DAgger** (Kelly et al., 2019) made this practical for robotics: a human expert monitors the robot and only intervenes when needed, rather than labeling every state. The gating between autonomous and human control is exactly the pause → takeover → return-to-policy loop used in the scripts here.
+
+- **RaC** (Hu et al., 2025) scales this loop to long-horizon tasks by explicitly decomposing interventions into **recovery** (teleoperating back to a good state) and **correction** (demonstrating the right behavior from there). This decomposition is the protocol followed by the HIL scripts in `examples/hil`.
+
+- **π0.6/RECAP** (Physical Intelligence, 2025) applies the same iterative collect-and-finetune loop at scale with VLA models, showing that even large pretrained policies benefit substantially from targeted human corrections on their own failure modes. π0.6 is trained using RECAP.
+
+```bibtex
+@article{ross2011dagger,
+  title={A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning},
+  author={Ross, Stéphane and Gordon, Geoffrey and Bagnell, Drew},
+  journal={Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics},
+  year={2011}
+}
+
+@article{kelly2019hgdagger,
+  title={HG-DAgger: Interactive Imitation Learning with Human Experts},
+  author={Kelly, Michael and Sidrane, Chelsea and Driggs-Campbell, Katherine and Kochenderfer, Mykel J},
+  journal={arXiv preprint arXiv:1810.02890},
+  year={2019}
+}
+
+@article{hu2025rac,
+  title={RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction},
+  author={Hu, Zheyuan and Wu, Robyn and Enock, Naveen and Li, Jasmine and Kadakia, Riya and Erickson, Zackory and Kumar, Aviral},
+  journal={arXiv preprint arXiv:2509.07953},
+  year={2025}
+}
+
+@article{pi2025recap,
+  title={π0.6: a VLA That Learns From Experience},
+  author={Physical Intelligence},
+  year={2025}
+}
+```

docs/source/installation.mdx

@@ -1,6 +1,6 @@
 # Installation
 
-This guide uses `conda` (via miniforge) to manage environments (recommended). If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.12 and `ffmpeg` installed with the `libsvtav1` encoder, then skip ahead to [Environment Setup](#step-2-environment-setup).
+This guide uses `conda` (via miniforge) to manage environments (recommended). If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.12 and support PyTorch >= 2.10, then skip ahead to [Environment Setup](#step-2-environment-setup).
 
 ## Step 1 (`conda` only): Install [`miniforge`](https://conda-forge.org/download/)
 
@@ -20,7 +20,7 @@ Create a virtual environment with Python 3.12:
 conda create -y -n lerobot python=3.12
 ```
 </hfoption>
-<hfoption id="uv">
+<hfoption id="uv (PyTorch >= 2.10 only)">
 ```bash
 uv python install 3.12
 uv venv --python 3.12
@@ -32,48 +32,87 @@ uv venv --python 3.12
 Then activate your virtual environment, you have to do this each time you open a shell to use lerobot:
 
 <!-- prettier-ignore-start -->
+
 <hfoptions id="activate_venv">
-<hfoption id="conda">```bash
+<hfoption id="conda">
+```bash
 conda activate lerobot
-```</hfoption>
-<hfoption id="uv">
-```bash
-# Linux/macOSsource
-source .venv/bin/activate
-# Windows PowerShell
-source .venv\Scripts\Activate.ps1
 ```
-</hfoption>
-</hfoptions>
-<!-- prettier-ignore-end -->
-
-When using `conda`, install `ffmpeg` in your environment:
-
-```bash
-conda install ffmpeg -c conda-forge
-ffmpeg -version  # ffmpeg 8.X is not yet supported !
-```
-
-> [!TIP]
-> This usually installs `ffmpeg 7.X` for your platform compiled with the `libsvtav1` encoder. If `libsvtav1` is not supported (check supported encoders with `ffmpeg -encoders`), you can:
->
-> - _[On any platform]_ Explicitly install `ffmpeg 7.X` using:
->
-> ```bash
-> conda install ffmpeg=7.1.1 -c conda-forge
-> ```
->
-> - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
 
 > [!NOTE]
-> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to install `evdev` with the following command:
+> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to also install `evdev`:
 >
 > ```bash
 > conda install evdev -c conda-forge
 > ```
 
+</hfoption>
+<hfoption id="uv (PyTorch >= 2.10 only)">
+```bash
+# Linux/macOS
+source .venv/bin/activate
+# Windows PowerShell
+.venv\Scripts\activate
+```
+
+> [!NOTE]
+> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to also install `evdev`:
+>
+> ```bash
+> sudo apt install libevdev-dev
+> uv pip install evdev
+> ```
+
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->
+
+### Install `ffmpeg` (for video decoding)
+
+LeRobot uses [TorchCodec](https://github.com/meta-pytorch/torchcodec) for video decoding by default, which requires `ffmpeg`.
+
+> [!NOTE]
+> **Platform support:** TorchCodec is **not available** on macOS Intel (x86_64), Linux ARM (aarch64, arm64, armv7l), or Windows with PyTorch < 2.8. On these platforms, LeRobot automatically falls back to `pyav` — so you do not need to install `ffmpeg` and can skip to Step 3.
+
+If your platform supports TorchCodec, install `ffmpeg` using one of the methods below:
+
+<!-- prettier-ignore-start -->
+
+<hfoptions id="install_ffmpeg">
+<hfoption id="conda (any PyTorch version)">
+
+Install `ffmpeg` in your conda environment. This works with **all PyTorch versions** and is **required for PyTorch < 2.10**:
+
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+> [!TIP]
+> This usually installs `ffmpeg 8.X` with the `libsvtav1` encoder. If you run into issues (e.g. `libsvtav1` missing — check with `ffmpeg -encoders` — or a version mismatch with `torchcodec`), you can explicitly install `ffmpeg 7.1.1` using:
+>
+> ```bash
+> conda install ffmpeg=7.1.1 -c conda-forge
+> ```
+
+</hfoption>
+<hfoption id="uv (PyTorch >= 2.10 only)">
+
+Starting with **PyTorch >= 2.10** (TorchCodec ≥ 0.10), TorchCodec can dynamically link to a system-wide `ffmpeg` installation. This is useful when using `uv` or other non-`conda` environment managers:
+
+```bash
+# Ubuntu/Debian
+sudo apt install ffmpeg
+
+# macOS (Apple Silicon)
+brew install ffmpeg
+```
+
 > [!IMPORTANT]
-> If you are using `uv` you will have to install `ffmpeg` system-wide (outside of the virtual environment). You rely on `uv` and `torchcodec` ability to dynamically link to the system `ffmpeg`.
+> System-wide `ffmpeg` is **only supported with PyTorch >= 2.10** (TorchCodec ≥ 0.10). For older PyTorch versions, you **must** use `conda install ffmpeg -c conda-forge` instead.
+
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->
 
 ## Step 3: Install LeRobot 🤗
 

docs/source/libero.mdx

@@ -1,36 +1,61 @@
 # LIBERO
 
-**LIBERO** is a benchmark designed to study **lifelong robot learning**. The idea is that robots won’t just be pretrained once in a factory, they’ll need to keep learning and adapting with their human users over time. This ongoing adaptation is called **lifelong learning in decision making (LLDM)**, and it’s a key step toward building robots that become truly personalized helpers.
+LIBERO is a benchmark designed to study **lifelong robot learning** — the idea that robots need to keep learning and adapting with their users over time, not just be pretrained once. It provides a set of standardized manipulation tasks that focus on **knowledge transfer**: how well a robot can apply what it has already learned to new situations. By evaluating on LIBERO, different algorithms can be compared fairly and researchers can build on each other's work.
 
-- 📄 [LIBERO paper](https://arxiv.org/abs/2306.03310)
-- 💻 [Original LIBERO repo](https://github.com/Lifelong-Robot-Learning/LIBERO)
-
-To make progress on this challenge, LIBERO provides a set of standardized tasks that focus on **knowledge transfer**: how well a robot can apply what it has already learned to new situations. By evaluating on LIBERO, different algorithms can be compared fairly and researchers can build on each other’s work.
-
-LIBERO includes **five task suites**:
-
-- **LIBERO-Spatial (`libero_spatial`)** – tasks that require reasoning about spatial relations.
-- **LIBERO-Object (`libero_object`)** – tasks centered on manipulating different objects.
-- **LIBERO-Goal (`libero_goal`)** – goal-conditioned tasks where the robot must adapt to changing targets.
-- **LIBERO-90 (`libero_90`)** – 90 short-horizon tasks from the LIBERO-100 collection.
-- **LIBERO-Long (`libero_10`)** – 10 long-horizon tasks from the LIBERO-100 collection.
-
-Together, these suites cover **130 tasks**, ranging from simple object manipulations to complex multi-step scenarios. LIBERO is meant to grow over time, and to serve as a shared benchmark where the community can test and improve lifelong learning algorithms.
+- Paper: [Benchmarking Knowledge Transfer for Lifelong Robot Learning](https://arxiv.org/abs/2306.03310)
+- GitHub: [Lifelong-Robot-Learning/LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO)
+- Project website: [libero-project.github.io](https://libero-project.github.io)
 
 ![An overview of the LIBERO benchmark](https://libero-project.github.io/assets/img/libero/fig1.png)
 
-## Evaluating with LIBERO
+## Available tasks
 
-At **LeRobot**, we ported [LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO) into our framework and used it mainly to **evaluate [SmolVLA](https://huggingface.co/docs/lerobot/en/smolvla)**, our lightweight Vision-Language-Action model.
+LIBERO includes **five task suites** covering **130 tasks**, ranging from simple object manipulations to complex multi-step scenarios:
 
-LIBERO is now part of our **multi-eval supported simulation**, meaning you can benchmark your policies either on a **single suite of tasks** or across **multiple suites at once** with just a flag.
+| Suite          | CLI name         | Tasks | Description                                        |
+| -------------- | ---------------- | ----- | -------------------------------------------------- |
+| LIBERO-Spatial | `libero_spatial` | 10    | Tasks requiring reasoning about spatial relations  |
+| LIBERO-Object  | `libero_object`  | 10    | Tasks centered on manipulating different objects   |
+| LIBERO-Goal    | `libero_goal`    | 10    | Goal-conditioned tasks with changing targets       |
+| LIBERO-90      | `libero_90`      | 90    | Short-horizon tasks from the LIBERO-100 collection |
+| LIBERO-Long    | `libero_10`      | 10    | Long-horizon tasks from the LIBERO-100 collection  |
 
-To Install LIBERO, after following LeRobot official instructions, just do:
-`pip install -e ".[libero]"`
+## Installation
+
+After following the LeRobot installation instructions:
+
+```bash
+pip install -e ".[libero]"
+```
+
+<Tip>
+LIBERO requires Linux (`sys_platform == 'linux'`). LeRobot uses MuJoCo for simulation — set the rendering backend before training or evaluation:
+
+```bash
+export MUJOCO_GL=egl  # for headless servers (HPC, cloud)
+```
+
+</Tip>
+
+## 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 \
+  --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 a policy on one LIBERO suite:
+Evaluate on one LIBERO suite:
 
 ```bash
 lerobot-eval \
@@ -42,15 +67,13 @@ lerobot-eval \
 ```
 
 - `--env.task` picks the suite (`libero_object`, `libero_spatial`, etc.).
-- `--env.task_ids` picks task ids to run (`[0]`, `[1,2,3]`, etc.). Omit this flag (or set it to `null`) to run all tasks in the suite.
+- `--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 in total.
-
----
+- `--eval.n_episodes` sets how many episodes to run per task.
 
 ### Multi-suite evaluation
 
-Benchmark a policy across multiple suites at once:
+Benchmark a policy across multiple suites at once by passing a comma-separated list:
 
 ```bash
 lerobot-eval \
@@ -61,50 +84,49 @@ lerobot-eval \
   --eval.n_episodes=2
 ```
 
-- Pass a comma-separated list to `--env.task` for multi-suite evaluation.
+### Control mode
 
-### Control Mode
+LIBERO 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:
 
-LIBERO now supports two control modes: relative and absolute. This matters because different VLA checkpoints are trained with different mode of action to output hence control parameterizations.
-You can switch them with: `env.control_mode = "relative"` and `env.control_mode = "absolute"`
+```bash
+--env.control_mode=relative   # or "absolute"
+```
 
 ### Policy inputs and outputs
 
-When using LIBERO through LeRobot, policies interact with the environment via **observations** and **actions**:
+**Observations:**
 
-- **Observations**
-  - `observation.state` – proprioceptive features (agent state).
-  - `observation.images.image` – main camera view (`agentview_image`).
-  - `observation.images.image2` – wrist camera view (`robot0_eye_in_hand_image`).
+- `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
 
-  ⚠️ **Note:** LeRobot enforces the `.images.*` prefix for any multi-modal visual features. Always ensure that your policy config `input_features` use the same naming keys, and that your dataset metadata keys follow this convention during evaluation.
-  If your data contains different keys, you must rename the observations to match what the policy expects, since naming keys are encoded inside the normalization statistics layer.
-  This will be fixed with the upcoming Pipeline PR.
+<Tip warning={true}>
+  LeRobot enforces the `.images.*` prefix for visual features. Ensure your
+  policy config `input_features` use the same naming keys, and that your dataset
+  metadata keys follow this convention. If your data contains different keys,
+  you must rename the observations to match what the policy expects, since
+  naming keys are encoded inside the normalization statistics layer.
+</Tip>
 
-- **Actions**
-  - Continuous control values in a `Box(-1, 1, shape=(7,))` space.
+**Actions:**
 
-We also provide a notebook for quick testing:
-Training with LIBERO
+- Continuous control in `Box(-1, 1, shape=(7,))` — 6D end-effector delta + 1D gripper
 
-## Training with LIBERO
+### Recommended evaluation episodes
 
-When training on LIBERO tasks, make sure your dataset parquet and metadata keys follow the LeRobot convention.
+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.
 
-The environment expects:
+## Training
 
-- `observation.state` → 8-dim agent state
-- `observation.images.image` → main camera (`agentview_image`)
-- `observation.images.image2` → wrist camera (`robot0_eye_in_hand_image`)
+### Dataset
 
-⚠️ Cleaning the dataset upfront is **cleaner and more efficient** than remapping keys inside the code.
-To avoid potential mismatches and key errors, we provide a **preprocessed LIBERO dataset** that is fully compatible with the current LeRobot codebase and requires no additional manipulation:
-👉 [HuggingFaceVLA/libero](https://huggingface.co/datasets/HuggingFaceVLA/libero)
+We provide a preprocessed LIBERO dataset fully compatible with LeRobot:
 
-For reference, here is the **original dataset** published by Physical Intelligence:
-👉 [physical-intelligence/libero](https://huggingface.co/datasets/physical-intelligence/libero)
+- [HuggingFaceVLA/libero](https://huggingface.co/datasets/HuggingFaceVLA/libero)
 
----
+For reference, the original dataset published by Physical Intelligence:
+
+- [physical-intelligence/libero](https://huggingface.co/datasets/physical-intelligence/libero)
 
 ### Example training command
 
@@ -121,52 +143,39 @@ lerobot-train \
   --batch_size=4 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
-  --eval_freq=1000 \
+  --eval_freq=1000
 ```
 
----
+## Reproducing published results
 
-### Note on rendering
+We reproduce the results of Pi0.5 on the LIBERO benchmark. We take the Physical Intelligence LIBERO base model (`pi05_libero`) and finetune for an additional 6k steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
 
-LeRobot uses MuJoCo for simulation. You need to set the rendering backend before training or evaluation:
+The finetuned model: [lerobot/pi05_libero_finetuned](https://huggingface.co/lerobot/pi05_libero_finetuned)
 
-- `export MUJOCO_GL=egl` → for headless servers (e.g. HPC, cloud)
-
-## Reproducing π₀.₅ results
-
-We reproduce the results of π₀.₅ on the LIBERO benchmark using the LeRobot implementation. We take the Physical Intelligence LIBERO base model (`pi05_libero`) and finetune for an additional 6k steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
-
-The finetuned model can be found here:
-
-- **π₀.₅ LIBERO**: [lerobot/pi05_libero_finetuned](https://huggingface.co/lerobot/pi05_libero_finetuned)
-
-We then evaluate the finetuned model using the LeRobot LIBERO implementation, by running the following command:
+### Evaluation command
 
 ```bash
 lerobot-eval \
-  --output_dir=/logs/ \
+  --output_dir=./eval_logs/ \
   --env.type=libero \
   --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
   --eval.batch_size=1 \
   --eval.n_episodes=10 \
   --policy.path=pi05_libero_finetuned \
   --policy.n_action_steps=10 \
-  --output_dir=./eval_logs/ \
   --env.max_parallel_tasks=1
 ```
 
-**Note:** We set `n_action_steps=10`, similar to the original OpenPI implementation.
+We set `n_action_steps=10`, matching the original OpenPI implementation.
 
 ### Results
 
-We obtain the following results on the LIBERO benchmark:
+| Model               | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average  |
+| ------------------- | -------------- | ------------- | ----------- | --------- | -------- |
+| **Pi0.5 (LeRobot)** | 97.0           | 99.0          | 98.0        | 96.0      | **97.5** |
 
-| Model    | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average  |
-| -------- | -------------- | ------------- | ----------- | --------- | -------- |
-| **π₀.₅** | 97.0           | 99.0          | 98.0        | 96.0      | **97.5** |
+These results are consistent with the [original results](https://github.com/Physical-Intelligence/openpi/tree/main/examples/libero#results) reported by Physical Intelligence:
 
-These results are consistent with the original [results](https://github.com/Physical-Intelligence/openpi/tree/main/examples/libero#results) reported by Physical Intelligence:
-
-| Model    | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average   |
-| -------- | -------------- | ------------- | ----------- | --------- | --------- |
-| **π₀.₅** | 98.8           | 98.2          | 98.0        | 92.4      | **96.85** |
+| Model              | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average   |
+| ------------------ | -------------- | ------------- | ----------- | --------- | --------- |
+| **Pi0.5 (OpenPI)** | 98.8           | 98.2          | 98.0        | 92.4      | **96.85** |

docs/source/metaworld.mdx

@@ -1,32 +1,111 @@
 # Meta-World
 
-Meta-World is a well-designed, open-source simulation benchmark for multi-task and meta reinforcement learning in continuous-control robotic manipulation. It gives researchers a shared, realistic playground to test whether algorithms can _learn many different tasks_ and _generalize quickly to new ones_ — two central challenges for real-world robotics.
+Meta-World is an open-source simulation benchmark for **multi-task and meta reinforcement learning** in continuous-control robotic manipulation. It bundles 50 diverse manipulation tasks using everyday objects and a common tabletop Sawyer arm, providing a standardized playground to test whether algorithms can learn many different tasks and generalize quickly to new ones.
 
-- 📄 [MetaWorld paper](https://arxiv.org/pdf/1910.10897)
-- 💻 [Original MetaWorld repo](https://github.com/Farama-Foundation/Metaworld)
+- Paper: [Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning](https://arxiv.org/abs/1910.10897)
+- GitHub: [Farama-Foundation/Metaworld](https://github.com/Farama-Foundation/Metaworld)
+- Project website: [metaworld.farama.org](https://metaworld.farama.org)
 
 ![MetaWorld MT10 demo](https://meta-world.github.io/figures/ml45.gif)
 
-## Why Meta-World matters
+## Available tasks
 
-- **Diverse, realistic tasks.** Meta-World bundles a large suite of simulated manipulation tasks (50 in the MT50 suite) using everyday objects and a common tabletop Sawyer arm. This diversity exposes algorithms to a wide variety of dynamics, contacts and goal specifications while keeping a consistent control and observation structure.
-- **Focus on generalization and multi-task learning.** By evaluating across task distributions that share structure but differ in goals and objects, Meta-World reveals whether an agent truly learns transferable skills rather than overfitting to a narrow task.
-- **Standardized evaluation protocol.** It provides clear evaluation modes and difficulty splits, so different methods can be compared fairly across easy, medium, hard and very-hard regimes.
-- **Empirical insight.** Past evaluations on Meta-World show impressive progress on some fronts, but also highlight that current multi-task and meta-RL methods still struggle with large, diverse task sets. That gap points to important research directions.
+Meta-World provides 50 tasks organized into difficulty groups. In LeRobot, you can evaluate on individual tasks, difficulty groups, or the full MT50 suite:
 
-## What it enables in LeRobot
+| Group      | CLI name             | Tasks | Description                                            |
+| ---------- | -------------------- | ----- | ------------------------------------------------------ |
+| Easy       | `easy`               | 28    | Tasks with simple dynamics and single-step goals       |
+| Medium     | `medium`             | 11    | Tasks requiring multi-step reasoning                   |
+| Hard       | `hard`               | 6     | Tasks with complex contacts and precise manipulation   |
+| Very Hard  | `very_hard`          | 5     | The most challenging tasks in the suite                |
+| MT50 (all) | Comma-separated list | 50    | All 50 tasks — the most challenging multi-task setting |
 
-In LeRobot, you can evaluate any policy or vision-language-action (VLA) model on Meta-World tasks and get a clear success-rate measure. The integration is designed to be straightforward:
+You can also pass individual task names directly (e.g., `assembly-v3`, `dial-turn-v3`).
 
-- We provide a LeRobot-ready dataset for Meta-World (MT50) on the HF Hub: `https://huggingface.co/datasets/lerobot/metaworld_mt50`.
-  - This dataset is formatted for the MT50 evaluation that uses all 50 tasks (the most challenging multi-task setting).
-  - MT50 gives the policy a one-hot task vector and uses fixed object/goal positions for consistency.
+We provide a LeRobot-ready dataset for Meta-World MT50 on the HF Hub: [lerobot/metaworld_mt50](https://huggingface.co/datasets/lerobot/metaworld_mt50). This dataset is formatted for the MT50 evaluation that uses all 50 tasks with fixed object/goal positions and one-hot task vectors for consistency.
 
-- Task descriptions and the exact keys required for evaluation are available in the repo/dataset — use these to ensure your policy outputs the right success signals.
+## Installation
 
-## Quick start, train a SmolVLA policy on Meta-World
+After following the LeRobot installation instructions:
 
-Example command to train a SmolVLA policy on a subset of tasks:
+```bash
+pip install -e ".[metaworld]"
+```
+
+<Tip warning={true}>
+If you encounter an `AssertionError: ['human', 'rgb_array', 'depth_array']` when running Meta-World environments, this is a mismatch between Meta-World and your Gymnasium version. Fix it with:
+
+```bash
+pip install "gymnasium==1.1.0"
+```
+
+</Tip>
+
+## Evaluation
+
+### Default evaluation (recommended)
+
+Evaluate on the medium difficulty split (a good balance of coverage and compute):
+
+```bash
+lerobot-eval \
+  --policy.path="your-policy-id" \
+  --env.type=metaworld \
+  --env.task=medium \
+  --eval.batch_size=1 \
+  --eval.n_episodes=10
+```
+
+### Single-task evaluation
+
+Evaluate on a specific task:
+
+```bash
+lerobot-eval \
+  --policy.path="your-policy-id" \
+  --env.type=metaworld \
+  --env.task=assembly-v3 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=10
+```
+
+### Multi-task evaluation
+
+Evaluate across multiple tasks or difficulty groups:
+
+```bash
+lerobot-eval \
+  --policy.path="your-policy-id" \
+  --env.type=metaworld \
+  --env.task=assembly-v3,dial-turn-v3,handle-press-side-v3 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=10
+```
+
+- `--env.task` accepts explicit task lists (comma-separated) or difficulty groups (e.g., `easy`, `medium`, `hard`, `very_hard`).
+- `--eval.batch_size` controls how many environments run in parallel.
+- `--eval.n_episodes` sets how many episodes to run per task.
+
+### Policy inputs and outputs
+
+**Observations:**
+
+- `observation.image` — single camera view (`corner2`), 480x480 HWC uint8
+- `observation.state` — 4-dim proprioceptive state (end-effector position + gripper)
+
+**Actions:**
+
+- Continuous control in `Box(-1, 1, shape=(4,))` — 3D end-effector delta + 1D gripper
+
+### Recommended evaluation episodes
+
+For reproducible benchmarking, use **10 episodes per task**. For the full MT50 suite this gives 500 total episodes. If you care about generalization, run on the full MT50 — it is intentionally challenging and reveals strengths/weaknesses better than a few narrow tasks.
+
+## Training
+
+### Example training command
+
+Train a SmolVLA policy on a subset of Meta-World tasks:
 
 ```bash
 lerobot-train \
@@ -44,37 +123,8 @@ lerobot-train \
   --eval_freq=1000
 ```
 
-Notes:
-
-- `--env.task` accepts explicit task lists (comma separated) or difficulty groups (e.g., `env.task="hard"`).
-- Adjust `batch_size`, `steps`, and `eval_freq` to match your compute budget.
-- **Gymnasium Assertion Error**: if you encounter an error like
-  `AssertionError: ['human', 'rgb_array', 'depth_array']` when running MetaWorld environments, this comes from a mismatch between MetaWorld and your Gymnasium version.
-  We recommend using:
-
-```bash
-  pip install "gymnasium==1.1.0"
-```
-
-to ensure proper compatibility.
-
-## Quick start — evaluate a trained policy
-
-To evaluate a trained policy on the Meta-World medium difficulty split:
-
-```bash
-lerobot-eval \
-  --policy.path="your-policy-id" \
-  --env.type=metaworld \
-  --env.task=medium \
-  --eval.batch_size=1 \
-  --eval.n_episodes=2
-```
-
-This will run episodes and return per-task success rates using the standard Meta-World evaluation keys.
-
 ## Practical tips
 
-- If you care about generalization, run on the full MT50 suite — it’s intentionally challenging and reveals strengths/weaknesses better than a few narrow tasks.
-- Use the one-hot task conditioning for multi-task training (MT10 / MT50 conventions) so policies have explicit task context.
+- Use the one-hot task conditioning for multi-task training (MT10/MT50 conventions) so policies have explicit task context.
 - Inspect the dataset task descriptions and the `info["is_success"]` keys when writing post-processing or logging so your success metrics line up with the benchmark.
+- Adjust `batch_size`, `steps`, and `eval_freq` to match your compute budget.

docs/source/multi_task_dit.mdx

@@ -331,6 +331,54 @@ lerobot-train \
   --wandb.project=multitask_dit
 ```
 
+## Libero Results
+
+```
+python -m lerobot.scripts.lerobot_train \
+  --dataset.repo_id=HuggingFaceVLA/libero \
+  --policy.type=multi_task_dit \
+  --policy.push_to_hub=false \
+  --output_dir="./outputs/multitask_dit_libero" \
+  --job_name="multitask-dit-libero" \
+  --wandb.enable=true \
+  --wandb.project=multitask_dit_libero \
+  --dataset.image_transforms.enable=true \
+  --dataset.image_transforms.max_num_transforms=4 \
+  --dataset.image_transforms.tfs='{"brightness":{"type":"ColorJitter","kwargs":{"brightness":[0.75,1.25]}},"contrast":{"type":"ColorJitter","kwargs":{"contrast":[0.6,1.4]}},"saturation":{"type":"ColorJitter","kwargs":{"saturation":[0.8,1.2]}},"hue":{"type":"ColorJitter","kwargs":{"hue":[-0.05,0.05]}},"sharpness":{"type":"SharpnessJitter","kwargs":{"sharpness":[0.6,1.4]}},"rotation":{"type":"RandomRotation","kwargs":{"degrees":[-5,5]}},"translation":{"type":"RandomAffine","kwargs":{"degrees":0,"translate":[0.1,0.1]}}}' \
+  --dataset.video_backend=torchcodec \
+  --policy.use_amp=true \
+  --policy.horizon=48 \
+  --policy.n_obs_steps=2 \
+  --policy.use_rope=true \
+  --policy.use_positional_encoding=false \
+  --policy.hidden_dim=768 \
+  --policy.num_layers=8 \
+  --policy.num_heads=12 \
+  --policy.dropout=0.1 \
+  --policy.timestep_embed_dim=256 \
+  --policy.objective=diffusion \
+  --policy.optimizer_lr=3e-4 \
+  --policy.optimizer_weight_decay=0 \
+  --policy.scheduler_warmup_steps=0 \
+  --policy.vision_encoder_name=openai/clip-vit-base-patch16 \
+  --policy.image_resize_shape=[256,256] \
+  --policy.image_crop_is_random=true \
+  --policy.text_encoder_name=openai/clip-vit-base-patch16 \
+  --policy.vision_encoder_lr_multiplier=0.1 \
+  --policy.device=cuda \
+  --num_workers=8 \
+  --save_freq=4000 \
+  --log_freq=100 \
+  --steps=100000 \
+  --batch_size=320
+```
+
+Results:
+
+| LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average |
+| -------------- | ------------- | ----------- | --------- | ------- |
+| 87.0           | 98.2          | 93.8        | 83.2      | 90.6    |
+
 ## References
 
 For more details on the technical implementation and architecture, see:

docs/source/policy_pi05_README.md

@@ -0,0 +1,91 @@
+# π₀.₅ (pi05)
+
+This repository contains the Hugging Face port of **π₀.₅**, adapted from [OpenPI](https://github.com/Physical-Intelligence/openpi) by the Physical Intelligence.
+It is designed as a **Vision-Language-Action model with open-world generalization**.
+
+---
+
+## Model Overview
+
+| Feature              | π₀                                                     | π₀.₅                                      |
+| -------------------- | ------------------------------------------------------ | ----------------------------------------- |
+| Time Conditioning    | Concatenates time with actions via `action_time_mlp_*` | Uses `time_mlp_*` for AdaRMS conditioning |
+| AdaRMS               | Not used                                               | Used in action expert                     |
+| Tokenizer Length     | 48 tokens                                              | 200 tokens                                |
+| Discrete State Input | False (Uses `state_proj` layer)                        | True                                      |
+| Parameter Count      | Higher (includes state embedding)                      | Lower (no state embedding)                |
+
+---
+
+## Relative Actions
+
+π₀.₅ supports training with **relative actions**, where the model learns relative offsets
+from the current robot state instead of absolute joint positions. This mirrors the
+relative-action transform in OpenPI (`DeltaActions`) and can improve performance.
+
+### How it works
+
+1. **During preprocessing**, absolute actions are converted to relative offsets:
+   `relative = action - state` (for selected joints).
+2. The relative actions are normalized using statistics computed from the relative distribution.
+3. **During postprocessing**, predicted relative actions are converted back to absolute:
+   `absolute = relative + state`.
+
+Joints listed in `relative_exclude_joints` (e.g., gripper) are kept absolute.
+
+### Configuration
+
+| Parameter                 | Type        | Default       | Description                                                      |
+| ------------------------- | ----------- | ------------- | ---------------------------------------------------------------- |
+| `use_relative_actions`    | `bool`      | `False`       | Enable relative-action training                                  |
+| `relative_exclude_joints` | `list[str]` | `["gripper"]` | Joint names to keep absolute (matched by substring)              |
+| `action_feature_names`    | `list[str]` | `None`        | Auto-populated from dataset metadata at runtime by `make_policy` |
+
+### Training example
+
+```bash
+python -m lerobot.scripts.lerobot_train \
+  --policy.type=pi05 \
+  --dataset.repo_id=your_org/your_dataset \
+  --policy.use_relative_actions=true \
+  --policy.relative_exclude_joints='["gripper"]'
+```
+
+When `use_relative_actions=true`, the training script automatically:
+
+- Computes relative action statistics from the dataset (sampled chunk-level relative actions)
+- Replaces the standard action stats with relative stats for normalization
+- Broadcasts these stats across all ranks in distributed training
+
+---
+
+## Citation
+
+If you use this work, please cite both **OpenPI** and the π₀.₅ paper:
+
+```bibtex
+@misc{openpi2024,
+  author       = {Physical Intelligence Lab},
+  title        = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
+  year         = {2024},
+  publisher    = {GitHub},
+  howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
+  license      = {Apache-2.0}
+}
+
+@misc{intelligence2025pi05visionlanguageactionmodelopenworld,
+  title        = {π₀.₅: a Vision-Language-Action Model with Open-World Generalization},
+  author       = {Physical Intelligence and Kevin Black and Noah Brown and James Darpinian and Karan Dhabalia and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Manuel Y. Galliker and Dibya Ghosh and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Devin LeBlanc and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Allen Z. Ren and Lucy Xiaoyang Shi and Laura Smith and Jost Tobias Springenberg and Kyle Stachowicz and James Tanner and Quan Vuong and Homer Walke and Anna Walling and Haohuan Wang and Lili Yu and Ury Zhilinsky},
+  year         = {2025},
+  eprint       = {2504.16054},
+  archivePrefix= {arXiv},
+  primaryClass = {cs.LG},
+  url          = {https://arxiv.org/abs/2504.16054},
+}
+```
+
+---
+
+## License
+
+This port follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).

docs/source/policy_pi0_README.md

@@ -0,0 +1,108 @@
+# π₀ (pi0)
+
+This repository contains the Hugging Face port of **π₀**, adapted from [OpenPI](https://github.com/Physical-Intelligence/openpi) by the Physical Intelligence.
+It is designed as a **Vision-Language-Action model for general robot control**.
+
+---
+
+## Model Overview
+
+| Feature              | π₀                                                     | π₀.₅                                      |
+| -------------------- | ------------------------------------------------------ | ----------------------------------------- |
+| Time Conditioning    | Concatenates time with actions via `action_time_mlp_*` | Uses `time_mlp_*` for AdaRMS conditioning |
+| AdaRMS               | Not used                                               | Used in action expert                     |
+| Tokenizer Length     | 48 tokens                                              | 200 tokens                                |
+| Discrete State Input | False (Uses `state_proj` layer)                        | True                                      |
+| Parameter Count      | Higher (includes state embedding)                      | Lower (no state embedding)                |
+
+---
+
+## Relative Actions
+
+π₀ supports training with **relative actions**, where the model learns relative offsets
+from the current robot state instead of absolute joint positions. This mirrors the
+relative-action transform in OpenPI (`DeltaActions`) and can improve performance.
+
+### How it works
+
+1. **During preprocessing**, absolute actions are converted to relative offsets:
+   `relative = action - state` (for selected joints).
+2. The relative actions are normalized using statistics computed from the relative distribution.
+3. **During postprocessing**, predicted relative actions are converted back to absolute:
+   `absolute = relative + state`.
+
+Joints listed in `relative_exclude_joints` (e.g., gripper) are kept absolute.
+
+### Configuration
+
+| Parameter                 | Type        | Default       | Description                                                      |
+| ------------------------- | ----------- | ------------- | ---------------------------------------------------------------- |
+| `use_relative_actions`    | `bool`      | `False`       | Enable relative-action training                                  |
+| `relative_exclude_joints` | `list[str]` | `["gripper"]` | Joint names to keep absolute (matched by substring)              |
+| `action_feature_names`    | `list[str]` | `None`        | Auto-populated from dataset metadata at runtime by `make_policy` |
+
+### Training example
+
+```bash
+python -m lerobot.scripts.lerobot_train \
+  --policy.type=pi0 \
+  --dataset.repo_id=your_org/your_dataset \
+  --policy.use_relative_actions=true \
+  --policy.relative_exclude_joints='["gripper"]'
+```
+
+When `use_relative_actions=true`, the training script automatically:
+
+- Computes relative action statistics from the dataset (sampled chunk-level relative actions)
+- Replaces the standard action stats with relative stats for normalization
+- Broadcasts these stats across all ranks in distributed training
+
+### Recomputing stats for an existing dataset
+
+If you want to precompute relative action stats offline, use `recompute_stats` from
+`lerobot.datasets.dataset_tools`:
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.dataset_tools import recompute_stats
+
+dataset = LeRobotDataset("your_org/your_dataset")
+dataset = recompute_stats(
+    dataset,
+    relative_action=True,
+    relative_exclude_joints=["gripper"],
+)
+```
+
+---
+
+## Citation
+
+If you use this work, please cite both **OpenPI** and the π₀ paper:
+
+```bibtex
+@misc{openpi2024,
+  author       = {Physical Intelligence Lab},
+  title        = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
+  year         = {2024},
+  publisher    = {GitHub},
+  howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
+  license      = {Apache-2.0}
+}
+
+@misc{black2024pi0visionlanguageactionflowmodel,
+  title        = {π₀: A Vision-Language-Action Flow Model for General Robot Control},
+  author       = {Kevin Black and Noah Brown and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Lucy Xiaoyang Shi and James Tanner and Quan Vuong and Anna Walling and Haohuan Wang and Ury Zhilinsky},
+  year         = {2024},
+  eprint       = {2410.24164},
+  archivePrefix= {arXiv},
+  primaryClass = {cs.LG},
+  url          = {https://arxiv.org/abs/2410.24164},
+}
+```
+
+---
+
+## License
+
+This port follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).

docs/source/policy_rtc_README.md

@@ -0,0 +1,38 @@
+# Real-Time Chunking (RTC)
+
+This module contains the LeRobot implementation of **Real-Time Chunking (RTC)**, an inference-time technique for flow-matching based policies.
+
+**Note**: RTC is not a policy itself, but rather an inference enhancement that works with flow-matching based policies including [π₀](../pi0/), [π₀.₅](../pi05/), and [SmolVLA](../smolvla/).
+
+---
+
+## Citation
+
+If you use Real-Time Chunking in your work, please cite:
+
+```bibtex
+@misc{openpi2024,
+  author       = {Physical Intelligence Lab},
+  title        = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
+  year         = {2024},
+  publisher    = {GitHub},
+  howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
+  license      = {Apache-2.0}
+}
+
+@misc{black2025realtimeexecutionactionchunking,
+      title={Real-Time Execution of Action Chunking Flow Policies},
+      author={Kevin Black and Manuel Y. Galliker and Sergey Levine},
+      year={2025},
+      eprint={2506.07339},
+      archivePrefix={arXiv},
+      primaryClass={cs.RO},
+      url={https://arxiv.org/abs/2506.07339},
+}
+```
+
+---
+
+## License
+
+This implementation follows the **Apache 2.0 License**, consistent with the LeRobot project.

docs/source/policy_sarm_README.md

@@ -0,0 +1,14 @@
+## Paper
+
+https://arxiv.org/abs/2509.25358
+
+## Citation
+
+```bibtex
+@article{chen2025sarm,
+  title={SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation},
+  author={Chen, Qianzhong and Yu, Justin and Schwager, Mac and Abbeel, Pieter and Shentu, Yide and Wu, Philipp},
+  journal={arXiv preprint arXiv:2509.25358},
+  year={2025}
+}
+```

examples/dataset/create_progress_videos.py

@@ -0,0 +1,680 @@
+#!/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.
+
+"""
+Create MP4 (or GIF) videos with sarm_progress overlay for specified episodes.
+
+Downloads datasets from HuggingFace, seeks directly into the episode segment
+of the source video, draws a progress line on each frame, and writes the result.
+
+Usage:
+    python examples/dataset/create_progress_videos.py \
+        --repo-id lerobot-data-collection/level2_final_quality3 \
+        --episode 1100
+
+    python examples/dataset/create_progress_videos.py \
+        --repo-id lerobot-data-collection/level2_final_quality3 \
+        --episode 1100 \
+        --camera-key observation.images.top \
+        --output-dir ./my_videos \
+        --gif
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import logging
+import subprocess
+from pathlib import Path
+
+import cv2
+import numpy as np
+import pandas as pd
+from huggingface_hub import snapshot_download
+
+GRAPH_Y_TOP_FRAC = 0.01
+GRAPH_Y_BOT_FRAC = 0.99
+LINE_THICKNESS = 3
+SHADOW_THICKNESS = 6
+REF_ALPHA = 0.45
+FILL_ALPHA = 0.55
+SCORE_FONT_SCALE = 0.8
+TASK_FONT_SCALE = 0.55
+
+
+def download_episode_metadata(repo_id: str, episode: int) -> Path:
+    """Download only the metadata and sarm_progress files for a dataset.
+
+    Args:
+        repo_id: HuggingFace dataset repository ID.
+        episode: Episode index (used for logging only; all meta is fetched).
+
+    Returns:
+        Local cache path for the downloaded snapshot.
+    """
+    logging.info("[1/4] Downloading metadata for %s (episode %d) ...", repo_id, episode)
+    local_path = Path(
+        snapshot_download(
+            repo_id=repo_id,
+            repo_type="dataset",
+            allow_patterns=["meta/**", "sarm_progress.parquet"],
+            ignore_patterns=["*.mp4"],
+        )
+    )
+    return local_path
+
+
+def load_episode_meta(local_path: Path, episode: int, camera_key: str | None) -> dict:
+    """Read info.json and episode parquet to resolve fps, video path, and timestamps.
+
+    Args:
+        local_path: Local cache directory containing meta/.
+        episode: Episode index to look up.
+        camera_key: Camera observation key (e.g. "observation.images.base").
+            If None, the first available video key is used.
+
+    Returns:
+        Dict with keys: fps, camera, video_rel, chunk_index, file_index,
+        from_ts, to_ts, task_name.
+    """
+    info = json.loads((local_path / "meta" / "info.json").read_text())
+    fps = info["fps"]
+    features = info["features"]
+
+    video_keys = [k for k, v in features.items() if v.get("dtype") == "video"]
+    if not video_keys:
+        raise RuntimeError("No video keys found in dataset features")
+
+    if camera_key is not None:
+        if camera_key not in video_keys:
+            raise RuntimeError(f"camera_key='{camera_key}' not found. Available: {video_keys}")
+        selected_camera = camera_key
+    else:
+        selected_camera = video_keys[0]
+    logging.info("   fps=%d  camera='%s'  all_cams=%s", fps, selected_camera, video_keys)
+
+    episode_rows = []
+    for parquet_file in sorted((local_path / "meta" / "episodes").glob("**/*.parquet")):
+        episode_rows.append(pd.read_parquet(parquet_file))
+    episode_df = pd.concat(episode_rows, ignore_index=True)
+    row = episode_df[episode_df["episode_index"] == episode]
+    if row.empty:
+        raise RuntimeError(f"Episode {episode} not found in episode metadata")
+    row = row.iloc[0]
+
+    chunk_col = f"videos/{selected_camera}/chunk_index"
+    file_col = f"videos/{selected_camera}/file_index"
+    ts_from_col = f"videos/{selected_camera}/from_timestamp"
+    ts_to_col = f"videos/{selected_camera}/to_timestamp"
+
+    if chunk_col not in row.index:
+        chunk_col = f"{selected_camera}/chunk_index"
+        file_col = f"{selected_camera}/file_index"
+        ts_from_col = f"{selected_camera}/from_timestamp"
+        ts_to_col = f"{selected_camera}/to_timestamp"
+    if chunk_col not in row.index:
+        raise RuntimeError(
+            f"Cannot find video metadata columns for {selected_camera}.\nAvailable: {list(row.index)}"
+        )
+
+    chunk_index = int(row[chunk_col])
+    file_index = int(row[file_col])
+    from_timestamp = float(row[ts_from_col])
+    to_timestamp = float(row[ts_to_col])
+
+    video_template = info.get(
+        "video_path", "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4"
+    )
+    video_rel = video_template.format(
+        video_key=selected_camera,
+        chunk_index=chunk_index,
+        file_index=file_index,
+    )
+
+    task_name = _resolve_task_name(row, local_path)
+
+    return {
+        "fps": fps,
+        "camera": selected_camera,
+        "video_rel": video_rel,
+        "chunk_index": chunk_index,
+        "file_index": file_index,
+        "from_ts": from_timestamp,
+        "to_ts": to_timestamp,
+        "task_name": task_name,
+    }
+
+
+def _resolve_task_name(row: pd.Series, local_path: Path) -> str:
+    """Best-effort extraction of the task name for an episode row.
+
+    Args:
+        row: Single-episode row from the episodes parquet.
+        local_path: Dataset cache root.
+
+    Returns:
+        Task name string, or empty string if unavailable.
+    """
+    try:
+        if "tasks" in row.index and row["tasks"] is not None:
+            tasks_val = row["tasks"]
+            if isinstance(tasks_val, (list, tuple, np.ndarray)) and len(tasks_val) > 0:
+                return str(tasks_val[0])
+            return str(tasks_val).strip("[]'")
+
+        tasks_parquet = local_path / "meta" / "tasks.parquet"
+        if tasks_parquet.exists():
+            tasks_df = pd.read_parquet(tasks_parquet)
+            task_idx = int(row.get("task_index", 0)) if "task_index" in row.index else 0
+            match = tasks_df[tasks_df["task_index"] == task_idx]
+            if not match.empty:
+                return str(match.index[0])
+    except Exception as exc:
+        logging.warning("Could not load task name: %s", exc)
+    return ""
+
+
+def download_video_file(repo_id: str, local_path: Path, video_rel: str) -> Path:
+    """Download the specific video file if not already cached.
+
+    Args:
+        repo_id: HuggingFace dataset repository ID.
+        local_path: Local cache directory.
+        video_rel: Relative path to the video file within the dataset.
+
+    Returns:
+        Absolute path to the downloaded video file.
+    """
+    video_path = local_path / video_rel
+    if video_path.exists():
+        logging.info("   Video already cached: %s", video_path)
+        return video_path
+    logging.info("[2/4] Downloading video file %s ...", video_rel)
+    snapshot_download(
+        repo_id=repo_id,
+        repo_type="dataset",
+        local_dir=str(local_path),
+        allow_patterns=[video_rel],
+    )
+    if not video_path.exists():
+        raise RuntimeError(f"Video not found after download: {video_path}")
+    return video_path
+
+
+def load_progress_data(local_path: Path, episode: int) -> np.ndarray | None:
+    """Load sarm_progress values for an episode.
+
+    Args:
+        local_path: Dataset cache root.
+        episode: Episode index.
+
+    Returns:
+        Sorted (N, 2) array of (frame_index, progress), or None if unavailable.
+    """
+    parquet_path = local_path / "sarm_progress.parquet"
+    if not parquet_path.exists():
+        logging.warning("sarm_progress.parquet not found")
+        return None
+    df = pd.read_parquet(parquet_path)
+    logging.info("   sarm_progress.parquet columns: %s", list(df.columns))
+    episode_df = df[df["episode_index"] == episode].copy()
+    if episode_df.empty:
+        logging.warning("No sarm_progress rows for episode %d", episode)
+        return None
+    episode_df = episode_df.sort_values("frame_index")
+
+    if "progress_dense" in episode_df.columns and episode_df["progress_dense"].notna().any():
+        progress_column = "progress_dense"
+    elif "progress_sparse" in episode_df.columns:
+        progress_column = "progress_sparse"
+    else:
+        progress_columns = [c for c in episode_df.columns if "progress" in c.lower()]
+        if not progress_columns:
+            return None
+        progress_column = progress_columns[0]
+
+    logging.info("   Using progress column: '%s'", progress_column)
+    return episode_df[["frame_index", progress_column]].rename(columns={progress_column: "progress"}).values
+
+
+def _precompute_pixel_coords(
+    progress_data: np.ndarray,
+    num_frames: int,
+    frame_width: int,
+    frame_height: int,
+) -> np.ndarray:
+    """Map progress samples to pixel coordinates for overlay drawing.
+
+    Args:
+        progress_data: (N, 2) array of (frame_index, progress).
+        num_frames: Total number of video frames.
+        frame_width: Video width in pixels.
+        frame_height: Video height in pixels.
+
+    Returns:
+        (N, 2) array of (x, y) pixel coordinates.
+    """
+    frame_indices = progress_data[:, 0].astype(float)
+    progress_values = np.clip(progress_data[:, 1].astype(float), 0.0, 1.0)
+
+    y_top = int(frame_height * GRAPH_Y_TOP_FRAC)
+    y_bot = int(frame_height * GRAPH_Y_BOT_FRAC)
+    graph_height = y_bot - y_top
+
+    x_coords = (frame_indices / (num_frames - 1) * (frame_width - 1)).astype(int)
+    y_coords = (y_bot - progress_values * graph_height).astype(int)
+
+    return np.stack([x_coords, y_coords], axis=1)
+
+
+def _progress_color(normalized_position: float) -> tuple[int, int, int]:
+    """Interpolate BGR color from red to green based on position in [0, 1].
+
+    Args:
+        normalized_position: Value in [0, 1] indicating how far along the episode.
+
+    Returns:
+        BGR color tuple.
+    """
+    red = int(255 * (1.0 - normalized_position))
+    green = int(255 * normalized_position)
+    return (0, green, red)
+
+
+def _prerender_fill_polygon(
+    pixel_coords: np.ndarray,
+    frame_width: int,
+    frame_height: int,
+) -> np.ndarray:
+    """Pre-render the grey fill polygon under the progress curve as a BGRA image.
+
+    Args:
+        pixel_coords: (N, 2) array of (x, y) pixel coordinates.
+        frame_width: Video width in pixels.
+        frame_height: Video height in pixels.
+
+    Returns:
+        BGRA image array of shape (frame_height, frame_width, 4).
+    """
+    y_bot = int(frame_height * GRAPH_Y_BOT_FRAC)
+    fill_image = np.zeros((frame_height, frame_width, 4), dtype=np.uint8)
+    polygon = np.concatenate(
+        [
+            pixel_coords,
+            [[pixel_coords[-1][0], y_bot], [pixel_coords[0][0], y_bot]],
+        ],
+        axis=0,
+    ).astype(np.int32)
+    cv2.fillPoly(fill_image, [polygon], color=(128, 128, 128, int(255 * FILL_ALPHA)))
+    return fill_image
+
+
+def _alpha_composite_region(base: np.ndarray, overlay_bgra: np.ndarray, x_limit: int) -> None:
+    """Blend BGRA overlay onto BGR base in-place, up to x_limit columns.
+
+    Args:
+        base: BGR frame to draw on (modified in-place).
+        overlay_bgra: BGRA overlay image.
+        x_limit: Only blend columns [0, x_limit).
+    """
+    if x_limit <= 0:
+        return
+    region_base = base[:, :x_limit]
+    region_overlay = overlay_bgra[:, :x_limit]
+    alpha = region_overlay[:, :, 3:4].astype(np.float32) / 255.0
+    region_base[:] = np.clip(
+        region_overlay[:, :, :3].astype(np.float32) * alpha + region_base.astype(np.float32) * (1.0 - alpha),
+        0,
+        255,
+    ).astype(np.uint8)
+
+
+def _draw_text_outlined(
+    frame: np.ndarray,
+    text: str,
+    position: tuple[int, int],
+    font_scale: float,
+    thickness: int = 1,
+) -> None:
+    """Draw white text with a dark outline for readability on any background.
+
+    Args:
+        frame: BGR image to draw on (modified in-place).
+        text: String to render.
+        position: (x, y) bottom-left corner of the text.
+        font_scale: OpenCV font scale.
+        thickness: Text stroke thickness.
+    """
+    font = cv2.FONT_HERSHEY_SIMPLEX
+    cv2.putText(frame, text, position, font, font_scale, (0, 0, 0), thickness + 2, cv2.LINE_AA)
+    cv2.putText(frame, text, position, font, font_scale, (255, 255, 255), thickness, cv2.LINE_AA)
+
+
+def composite_progress_video(
+    video_path: Path,
+    from_timestamp: float,
+    to_timestamp: float,
+    progress_data: np.ndarray,
+    output_path: Path,
+    fps: float,
+    task_name: str = "",
+) -> Path:
+    """Read episode frames by seeking into the source video, draw progress overlay, write output.
+
+    Uses cv2.CAP_PROP_POS_MSEC to seek directly into the source video,
+    eliminating the need for an intermediate clip file.
+
+    Args:
+        video_path: Path to the full source video file.
+        from_timestamp: Start timestamp of the episode in seconds.
+        to_timestamp: End timestamp of the episode in seconds.
+        progress_data: (N, 2) array of (frame_index, progress).
+        output_path: Path to write the output MP4.
+        fps: Frames per second for the output video.
+        task_name: Optional task name to display at the top of the video.
+
+    Returns:
+        Path to the written output file (MP4).
+    """
+    capture = cv2.VideoCapture(str(video_path))
+    try:
+        capture.set(cv2.CAP_PROP_POS_MSEC, from_timestamp * 1000)
+
+        frame_width = int(capture.get(cv2.CAP_PROP_FRAME_WIDTH))
+        frame_height = int(capture.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        duration_seconds = to_timestamp - from_timestamp
+        num_frames = int(round(duration_seconds * fps))
+
+        logging.info(
+            "   Video: %dx%d, %d frames @ %.1f fps (%.2fs)",
+            frame_width,
+            frame_height,
+            num_frames,
+            fps,
+            duration_seconds,
+        )
+
+        pixel_coords = _precompute_pixel_coords(progress_data, num_frames, frame_width, frame_height)
+        y_ref = int(frame_height * GRAPH_Y_TOP_FRAC)
+
+        fill_image = _prerender_fill_polygon(pixel_coords, frame_width, frame_height)
+
+        ref_line_image = np.zeros((frame_height, frame_width, 4), dtype=np.uint8)
+        cv2.line(
+            ref_line_image,
+            (0, y_ref),
+            (frame_width - 1, y_ref),
+            (200, 200, 200, int(255 * REF_ALPHA)),
+            1,
+            cv2.LINE_AA,
+        )
+
+        frame_indices = progress_data[:, 0].astype(int)
+        progress_values = progress_data[:, 1].astype(float)
+
+        logging.info("[3/4] Compositing %d frames ...", num_frames)
+        fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+        writer = cv2.VideoWriter(str(output_path), fourcc, fps, (frame_width, frame_height))
+
+        for frame_idx in range(num_frames):
+            ret, frame = capture.read()
+            if not ret:
+                break
+
+            drawn_count = int(np.searchsorted(frame_indices, frame_idx, side="right"))
+            x_current = (
+                int(pixel_coords[min(drawn_count, len(pixel_coords)) - 1][0]) + 1 if drawn_count > 0 else 0
+            )
+
+            _alpha_composite_region(frame, ref_line_image, frame_width)
+            _alpha_composite_region(frame, fill_image, x_current)
+
+            if drawn_count >= 2:
+                time_position = (drawn_count - 1) / max(len(progress_values) - 1, 1)
+                line_color = _progress_color(time_position)
+                points = pixel_coords[:drawn_count].reshape(-1, 1, 2).astype(np.int32)
+                cv2.polylines(
+                    frame,
+                    [points],
+                    isClosed=False,
+                    color=(255, 255, 255),
+                    thickness=SHADOW_THICKNESS,
+                    lineType=cv2.LINE_AA,
+                )
+                cv2.polylines(
+                    frame,
+                    [points],
+                    isClosed=False,
+                    color=line_color,
+                    thickness=LINE_THICKNESS,
+                    lineType=cv2.LINE_AA,
+                )
+
+            if drawn_count > 0:
+                score = float(progress_values[min(drawn_count, len(progress_values)) - 1])
+                score_text = f"{score:.2f}"
+                (text_width, _), _ = cv2.getTextSize(
+                    score_text, cv2.FONT_HERSHEY_SIMPLEX, SCORE_FONT_SCALE, 2
+                )
+                score_x = frame_width - text_width - 12
+                score_y = frame_height - 12
+                time_position = (drawn_count - 1) / max(len(progress_values) - 1, 1)
+                score_color = _progress_color(time_position)
+                cv2.putText(
+                    frame,
+                    score_text,
+                    (score_x, score_y),
+                    cv2.FONT_HERSHEY_SIMPLEX,
+                    SCORE_FONT_SCALE,
+                    (0, 0, 0),
+                    4,
+                    cv2.LINE_AA,
+                )
+                cv2.putText(
+                    frame,
+                    score_text,
+                    (score_x, score_y),
+                    cv2.FONT_HERSHEY_SIMPLEX,
+                    SCORE_FONT_SCALE,
+                    score_color,
+                    2,
+                    cv2.LINE_AA,
+                )
+
+            if task_name:
+                (text_width, _), _ = cv2.getTextSize(task_name, cv2.FONT_HERSHEY_SIMPLEX, TASK_FONT_SCALE, 1)
+                task_x = max((frame_width - text_width) // 2, 4)
+                _draw_text_outlined(frame, task_name, (task_x, 22), TASK_FONT_SCALE)
+
+            writer.write(frame)
+            if frame_idx % 100 == 0:
+                logging.info("   Frame %d/%d ...", frame_idx, num_frames)
+
+        writer.release()
+    finally:
+        capture.release()
+
+    logging.info("   MP4 written: %s", output_path)
+    return output_path
+
+
+def convert_mp4_to_gif(mp4_path: Path) -> Path:
+    """Convert an MP4 to an optimized GIF using ffmpeg palette generation.
+
+    Args:
+        mp4_path: Path to the source MP4 file.
+
+    Returns:
+        Path to the generated GIF file.
+    """
+    capture = cv2.VideoCapture(str(mp4_path))
+    frame_width = int(capture.get(cv2.CAP_PROP_FRAME_WIDTH))
+    capture.release()
+
+    gif_path = mp4_path.with_suffix(".gif")
+    palette_path = mp4_path.parent / "_palette.png"
+
+    logging.info("[4/4] Converting to GIF ...")
+    result_palette = subprocess.run(  # nosec B607
+        [
+            "ffmpeg",
+            "-y",
+            "-i",
+            str(mp4_path),
+            "-vf",
+            f"fps=10,scale={frame_width}:-1:flags=lanczos,palettegen=max_colors=128:stats_mode=diff",
+            "-update",
+            "1",
+            str(palette_path),
+        ],
+        capture_output=True,
+        text=True,
+    )
+    if result_palette.returncode != 0:
+        logging.warning("palettegen failed:\n%s", result_palette.stderr[-500:])
+
+    result_gif = subprocess.run(  # nosec B607
+        [
+            "ffmpeg",
+            "-y",
+            "-i",
+            str(mp4_path),
+            "-i",
+            str(palette_path),
+            "-filter_complex",
+            f"fps=10,scale={frame_width}:-1:flags=lanczos[v];[v][1:v]paletteuse=dither=bayer:bayer_scale=3",
+            str(gif_path),
+        ],
+        capture_output=True,
+        text=True,
+    )
+    if result_gif.returncode != 0:
+        logging.warning("GIF encode failed:\n%s", result_gif.stderr[-500:])
+
+    palette_path.unlink(missing_ok=True)
+    logging.info("   GIF written: %s", gif_path)
+    return gif_path
+
+
+def process_dataset(
+    repo_id: str,
+    episode: int,
+    camera_key: str | None,
+    output_dir: Path,
+    create_gif: bool = False,
+) -> Path | None:
+    """Full pipeline: download, extract metadata, composite progress, write output.
+
+    Args:
+        repo_id: HuggingFace dataset repository ID.
+        episode: Episode index.
+        camera_key: Camera key to use, or None for auto-selection.
+        output_dir: Directory to write output files.
+        create_gif: If True, also generate a GIF from the MP4.
+
+    Returns:
+        Path to the final output file, or None on failure.
+    """
+    safe_name = repo_id.replace("/", "_")
+    logging.info("Processing: %s  |  episode %d", repo_id, episode)
+
+    local_path = download_episode_metadata(repo_id, episode)
+    logging.info("   Local cache: %s", local_path)
+
+    episode_meta = load_episode_meta(local_path, episode, camera_key)
+    logging.info("   Episode meta: %s", episode_meta)
+
+    video_path = download_video_file(repo_id, local_path, episode_meta["video_rel"])
+
+    progress_data = load_progress_data(local_path, episode)
+    if progress_data is None:
+        logging.error("Could not load sarm_progress data. Skipping overlay.")
+        return None
+
+    logging.info("   Progress frames: %d", len(progress_data))
+
+    output_path = output_dir / f"{safe_name}_ep{episode}_progress.mp4"
+    final_path = composite_progress_video(
+        video_path=video_path,
+        from_timestamp=episode_meta["from_ts"],
+        to_timestamp=episode_meta["to_ts"],
+        progress_data=progress_data,
+        output_path=output_path,
+        fps=episode_meta["fps"],
+        task_name=episode_meta.get("task_name", ""),
+    )
+
+    if create_gif:
+        final_path = convert_mp4_to_gif(final_path)
+
+    logging.info("Done: %s", final_path)
+    return final_path
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Create MP4/GIF videos with sarm_progress overlay for dataset episodes."
+    )
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="HuggingFace dataset repository ID (e.g. 'lerobot-data-collection/level2_final_quality3').",
+    )
+    parser.add_argument(
+        "--episode",
+        type=int,
+        required=True,
+        help="Episode index to visualize.",
+    )
+    parser.add_argument(
+        "--camera-key",
+        type=str,
+        default=None,
+        help="Camera observation key (e.g. 'observation.images.base'). Auto-selects first camera if omitted.",
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=Path("progress_videos"),
+        help="Directory to write output files (default: ./progress_videos).",
+    )
+    parser.add_argument(
+        "--gif",
+        action="store_true",
+        help="Also generate a GIF from the MP4 output.",
+    )
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
+
+    args.output_dir.mkdir(parents=True, exist_ok=True)
+
+    result = process_dataset(
+        repo_id=args.repo_id,
+        episode=args.episode,
+        camera_key=args.camera_key,
+        output_dir=args.output_dir,
+        create_gif=args.gif,
+    )
+
+    if result:
+        logging.info("Output: %s", result)
+
+
+if __name__ == "__main__":
+    main()

examples/hil/hil_utils.py

@@ -0,0 +1,228 @@
+# 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.
+
+"""Shared utilities for Human-in-the-Loop data collection scripts."""
+
+import logging
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+
+from lerobot.processor import (
+    IdentityProcessorStep,
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+)
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots import Robot
+from lerobot.teleoperators import Teleoperator
+from lerobot.utils.control_utils import is_headless
+from lerobot.utils.robot_utils import precise_sleep
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class HILDatasetConfig:
+    repo_id: str
+    single_task: str
+    root: str | Path | None = None
+    fps: int = 30
+    episode_time_s: float = 120
+    num_episodes: int = 50
+    video: bool = True
+    push_to_hub: bool = True
+    private: bool = False
+    tags: list[str] | None = None
+    num_image_writer_processes: int = 0
+    num_image_writer_threads_per_camera: int = 4
+    video_encoding_batch_size: int = 1
+    vcodec: str = "auto"
+    streaming_encoding: bool = True
+    encoder_queue_maxsize: int = 30
+    encoder_threads: int | None = None
+    rename_map: dict[str, str] = field(default_factory=dict)
+
+
+def teleop_has_motor_control(teleop: Teleoperator) -> bool:
+    """Check if teleoperator has motor control capabilities."""
+    return all(hasattr(teleop, attr) for attr in ("enable_torque", "disable_torque", "write_goal_positions"))
+
+
+def teleop_disable_torque(teleop: Teleoperator) -> None:
+    """Disable teleop torque if supported."""
+    if hasattr(teleop, "disable_torque"):
+        teleop.disable_torque()
+
+
+def teleop_enable_torque(teleop: Teleoperator) -> None:
+    """Enable teleop torque if supported."""
+    if hasattr(teleop, "enable_torque"):
+        teleop.enable_torque()
+
+
+def teleop_smooth_move_to(teleop: Teleoperator, target_pos: dict, duration_s: float = 2.0, fps: int = 50):
+    """Smoothly move teleop to target position if motor control is available."""
+    if not teleop_has_motor_control(teleop):
+        logger.warning("Teleop does not support motor control - cannot mirror robot position")
+        return
+
+    teleop_enable_torque(teleop)
+    current = teleop.get_action()
+    steps = max(int(duration_s * fps), 1)
+
+    for step in range(steps + 1):
+        t = step / steps
+        interp = {}
+        for k in current:
+            if k in target_pos:
+                interp[k] = current[k] * (1 - t) + target_pos[k] * t
+            else:
+                interp[k] = current[k]
+        teleop.write_goal_positions(interp)
+        time.sleep(1 / fps)
+
+
+def init_keyboard_listener():
+    """Initialize keyboard listener with HIL controls."""
+    events = {
+        "exit_early": False,
+        "rerecord_episode": False,
+        "stop_recording": False,
+        "policy_paused": False,
+        "correction_active": False,
+        "resume_policy": False,
+        "in_reset": False,
+        "start_next_episode": False,
+    }
+
+    if is_headless():
+        logger.warning("Headless environment - keyboard controls unavailable")
+        return None, events
+
+    from pynput import keyboard
+
+    def on_press(key):
+        try:
+            if events["in_reset"]:
+                if key in [keyboard.Key.space, keyboard.Key.right]:
+                    logger.info("[HIL] Starting next episode...")
+                    events["start_next_episode"] = True
+                elif hasattr(key, "char") and key.char == "c":
+                    events["start_next_episode"] = True
+                elif key == keyboard.Key.esc:
+                    logger.info("[HIL] ESC - Stop recording, pushing to hub...")
+                    events["stop_recording"] = True
+                    events["start_next_episode"] = True
+            else:
+                if key == keyboard.Key.space:
+                    if not events["policy_paused"] and not events["correction_active"]:
+                        logger.info("[HIL] PAUSED - Press 'c' to take control or 'p' to resume policy")
+                        events["policy_paused"] = True
+                elif hasattr(key, "char") and key.char == "c":
+                    if events["policy_paused"] and not events["correction_active"]:
+                        logger.info("[HIL] Taking control...")
+                        events["start_next_episode"] = True
+                elif hasattr(key, "char") and key.char == "p":
+                    if events["policy_paused"] or events["correction_active"]:
+                        logger.info("[HIL] Resuming policy...")
+                        events["resume_policy"] = True
+                elif key == keyboard.Key.right:
+                    logger.info("[HIL] End episode")
+                    events["exit_early"] = True
+                elif key == keyboard.Key.left:
+                    logger.info("[HIL] Re-record episode")
+                    events["rerecord_episode"] = True
+                    events["exit_early"] = True
+                elif key == keyboard.Key.esc:
+                    logger.info("[HIL] ESC - Stop recording...")
+                    events["stop_recording"] = True
+                    events["exit_early"] = True
+        except Exception as e:
+            logger.info(f"Key error: {e}")
+
+    listener = keyboard.Listener(on_press=on_press)
+    listener.start()
+    return listener, events
+
+
+def make_identity_processors():
+    """Create identity processors for recording."""
+    teleop_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[IdentityProcessorStep()],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+    obs_proc = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[IdentityProcessorStep()],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+    return teleop_proc, obs_proc
+
+
+def reset_loop(robot: Robot, teleop: Teleoperator, events: dict, fps: int):
+    """Reset period where human repositions environment."""
+    logger.info("[HIL] RESET")
+
+    events["in_reset"] = True
+    events["start_next_episode"] = False
+
+    obs = robot.get_observation()
+    robot_pos = {k: v for k, v in obs.items() if k.endswith(".pos") and k in robot.observation_features}
+    teleop_smooth_move_to(teleop, robot_pos, duration_s=2.0, fps=50)
+
+    logger.info("Press any key to enable teleoperation")
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        precise_sleep(0.05)
+
+    if events["stop_recording"]:
+        return
+
+    events["start_next_episode"] = False
+    teleop_disable_torque(teleop)
+    logger.info("Teleop enabled - press any key to start episode")
+
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        loop_start = time.perf_counter()
+        action = teleop.get_action()
+        robot.send_action(action)
+        precise_sleep(1 / fps - (time.perf_counter() - loop_start))
+
+    events["in_reset"] = False
+    events["start_next_episode"] = False
+    events["exit_early"] = False
+    events["policy_paused"] = False
+    events["correction_active"] = False
+    events["resume_policy"] = False
+
+
+def print_controls(rtc: bool = False):
+    """Print control instructions."""
+    mode = "Human-in-the-Loop Data Collection" + (" (RTC)" if rtc else "")
+    logger.info(
+        "%s\n  Controls:\n"
+        "    SPACE  - Pause policy\n"
+        "    c      - Take control\n"
+        "    p      - Resume policy after pause/correction\n"
+        "    →      - End episode\n"
+        "    ESC    - Stop and push to hub",
+        mode,
+    )

examples/rtc/eval_with_real_robot.py

@@ -69,15 +69,20 @@ Usage:
         --policy.path=lerobot-data-collection/folding_final \
         --robot.type=bi_openarm_follower \
         --robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}}' \
-        --robot.left_arm_config.port=can1 \
+        --robot.left_arm_config.port=can0 \
         --robot.left_arm_config.side=left \
         --robot.left_arm_config.can_interface=socketcan \
-        --robot.right_arm_config.port=can0 \
+        --robot.left_arm_config.disable_torque_on_disconnect=true \
+        --robot.left_arm_config.max_relative_target=8.0 \
+        --robot.right_arm_config.port=can1 \
         --robot.right_arm_config.side=right \
         --robot.right_arm_config.can_interface=socketcan \
+        --robot.right_arm_config.disable_torque_on_disconnect=true \
+        --robot.right_arm_config.max_relative_target=8.0 \
         --task="Fold the T-shirt properly" \
         --fps=30 \
         --duration=2000 \
+        --interpolation_multiplier=3 \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
         --rtc.max_guidance_weight=5.0 \
@@ -104,9 +109,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import RTCAttentionSchedule
 from lerobot.datasets.feature_utils import build_dataset_frame, hw_to_dataset_features
 from lerobot.policies.factory import get_policy_class, make_pre_post_processors
-from lerobot.policies.rtc.action_queue import ActionQueue
-from lerobot.policies.rtc.configuration_rtc import RTCConfig
-from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.policies.rtc import ActionInterpolator, ActionQueue, LatencyTracker, RTCConfig
 from lerobot.processor import (
     NormalizerProcessorStep,
     RelativeActionsProcessorStep,
@@ -181,6 +184,7 @@ class RTCDemoConfig(HubMixin):
     # Demo parameters
     duration: float = 30.0  # Duration to run the demo (seconds)
     fps: float = 10.0  # Action execution frequency (Hz)
+    interpolation_multiplier: int = 1  # Control rate multiplier (1=off, 2=2x, 3=3x)
 
     # Compute device
     device: str | None = None  # Device to run on (cuda, cpu, auto)
@@ -461,20 +465,23 @@ def actor_control(
         action_keys = [k for k in robot.action_features() if k.endswith(".pos")]
 
         action_count = 0
-        action_interval = 1.0 / cfg.fps
+        interpolator = ActionInterpolator(multiplier=cfg.interpolation_multiplier)
+        action_interval = interpolator.get_control_interval(cfg.fps)
 
         while not shutdown_event.is_set():
             start_time = time.perf_counter()
 
-            # Try to get an action from the queue with timeout
-            action = action_queue.get()
+            if interpolator.needs_new_action():
+                new_action = action_queue.get()
+                if new_action is not None:
+                    interpolator.add(new_action.cpu())
 
+            action = interpolator.get()
             if action is not None:
                 action = action.cpu()
                 action_dict = {key: action[i].item() for i, key in enumerate(action_keys)}
                 action_processed = robot_action_processor((action_dict, None))
                 robot.send_action(action_processed)
-
                 action_count += 1
 
             dt_s = time.perf_counter() - start_time

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.5.1"
+version = "0.5.2"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
@@ -71,9 +71,9 @@ dependencies = [
     "cmake>=3.29.0.1,<4.2.0",
     "packaging>=24.2,<26.0",
 
-    "torch>=2.2.1,<2.11.0",
-    "torchcodec>=0.2.1,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')",
-    "torchvision>=0.21.0,<0.26.0",
+    "torch>=2.7,<2.11.0",
+    "torchcodec>=0.3.0,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # NOTE: Windows support starts at version 0.7 (needs torch==2.8), ffmpeg>=8 support starts at version 0.8.1 (needs torch==2.9), system-wide ffmpeg support starts at version 0.10 (needs torch==2.10).
+    "torchvision>=0.22.0,<0.26.0",
 
     "einops>=0.8.0,<0.9.0",
     "opencv-python-headless>=4.9.0,<4.14.0",

scripts/ci/extract_task_descriptions.py

@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+# 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.
+
+"""Extract natural-language task descriptions for a benchmark suite.
+
+Runs inside the benchmark Docker container (where the env library is installed)
+immediately after lerobot-eval, writing a JSON file that parse_eval_metrics.py
+picks up and embeds in metrics.json.
+
+Output format: {"<suite>_<task_idx>": "<nl instruction>", ...}
+
+Usage:
+    python scripts/ci/extract_task_descriptions.py \\
+        --env libero --task libero_spatial \\
+        --output /tmp/eval-artifacts/task_descriptions.json
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+
+def _libero_descriptions(task_suite: str) -> dict[str, str]:
+    from libero.libero import benchmark  # type: ignore[import-untyped]
+
+    suite_dict = benchmark.get_benchmark_dict()
+    if task_suite not in suite_dict:
+        print(
+            f"[extract_task_descriptions] Unknown LIBERO suite '{task_suite}'. "
+            f"Available: {list(suite_dict.keys())}",
+            file=sys.stderr,
+        )
+        return {}
+    suite = suite_dict[task_suite]()
+    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]:
+    # MetaWorld tasks don't expose a separate NL description attribute;
+    # use a cleaned version of the task name as the description.
+    label = task_name.removeprefix("metaworld-").replace("-", " ").strip()
+    return {f"{task_name}_0": label}
+
+
+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("--output", required=True, help="Path to write task_descriptions.json")
+    args = parser.parse_args()
+
+    descriptions: dict[str, str] = {}
+    try:
+        if args.env == "libero":
+            descriptions = _libero_descriptions(args.task)
+        elif args.env == "metaworld":
+            descriptions = _metaworld_descriptions(args.task)
+        else:
+            print(
+                f"[extract_task_descriptions] No description extractor for env '{args.env}'.",
+                file=sys.stderr,
+            )
+    except Exception as exc:
+        print(f"[extract_task_descriptions] Warning: {exc}", file=sys.stderr)
+
+    out_path = Path(args.output)
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    out_path.write_text(json.dumps(descriptions, indent=2))
+    print(f"[extract_task_descriptions] {len(descriptions)} descriptions → {out_path}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

scripts/ci/parse_eval_metrics.py

@@ -0,0 +1,129 @@
+#!/usr/bin/env python3
+# 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.
+
+"""Parse lerobot-eval output into a small metrics.json artifact.
+
+Reads eval_info.json written by lerobot-eval --output_dir and extracts the
+key metrics needed by the health dashboard. Handles both single-task and
+multi-task eval output formats.
+
+Usage:
+    python scripts/ci/parse_eval_metrics.py \\
+        --artifacts-dir /tmp/libero-artifacts \\
+        --env libero \\
+        --task libero_spatial \\
+        --policy pepijn223/smolvla_libero
+
+Writes <artifacts-dir>/metrics.json. The CI workflow then uploads this file
+as a GitHub Actions artifact named "<env>-metrics".
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import sys
+from pathlib import Path
+
+
+def _extract_metrics(info: dict) -> tuple[float | None, int | None, float | None, float | None]:
+    """Extract (pc_success, n_episodes, avg_sum_reward, eval_s) from eval_info.json.
+
+    Handles two output shapes:
+      - Single-task: {"aggregated": {"pc_success": 80.0, ...}}
+      - Multi-task:  {"overall": {"pc_success": 80.0, "n_episodes": 5, ...}}
+    """
+    for key in ("aggregated", "overall"):
+        if key not in info:
+            continue
+        agg = info[key]
+        pc = agg.get("pc_success")
+        n = agg.get("n_episodes")
+        reward = agg.get("avg_sum_reward")
+        eval_s = agg.get("eval_s")
+        if pc is not None and not math.isnan(pc):
+            return (
+                float(pc),
+                int(n) if n is not None else None,
+                float(reward) if reward is not None else None,
+                float(eval_s) if eval_s is not None else None,
+            )
+
+    return None, None, None, None
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    parser.add_argument("--artifacts-dir", required=True, help="Path to the mounted artifacts volume")
+    parser.add_argument("--env", required=True, help="Environment name (e.g. libero)")
+    parser.add_argument("--task", required=True, help="Task name (e.g. libero_spatial)")
+    parser.add_argument("--policy", required=True, help="Policy hub path (e.g. pepijn223/smolvla_libero)")
+    args = parser.parse_args()
+
+    artifacts_dir = Path(args.artifacts_dir)
+    eval_info_path = artifacts_dir / "eval_info.json"
+
+    pc_success: float | None = None
+    n_episodes: int | None = None
+    avg_sum_reward: float | None = None
+    eval_s: float | None = None
+
+    if eval_info_path.exists():
+        try:
+            info = json.loads(eval_info_path.read_text())
+            pc_success, n_episodes, avg_sum_reward, eval_s = _extract_metrics(info)
+        except (json.JSONDecodeError, KeyError, TypeError) as exc:
+            print(f"[parse_eval_metrics] Warning: could not parse eval_info.json: {exc}", file=sys.stderr)
+    else:
+        print(
+            f"[parse_eval_metrics] Warning: {eval_info_path} not found — eval may have failed.",
+            file=sys.stderr,
+        )
+
+    task_descriptions: dict[str, str] = {}
+    task_desc_path = artifacts_dir / "task_descriptions.json"
+    if task_desc_path.exists():
+        try:
+            task_descriptions = json.loads(task_desc_path.read_text())
+        except json.JSONDecodeError as exc:
+            print(
+                f"[parse_eval_metrics] Warning: could not parse task_descriptions.json: {exc}",
+                file=sys.stderr,
+            )
+
+    metrics = {
+        "env": args.env,
+        "task": args.task,
+        "policy": args.policy,
+        "pc_success": pc_success,
+        "n_episodes": n_episodes,
+        "avg_sum_reward": avg_sum_reward,
+        "eval_s": eval_s,
+        "task_descriptions": task_descriptions,
+    }
+
+    out_path = artifacts_dir / "metrics.json"
+    out_path.write_text(json.dumps(metrics, indent=2))
+    print(f"[parse_eval_metrics] Written: {out_path}")
+    print(json.dumps(metrics, indent=2))
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

src/lerobot/configs/default.py

@@ -65,20 +65,27 @@ class WandBConfig:
 class EvalConfig:
     n_episodes: int = 50
     # `batch_size` specifies the number of environments to use in a gym.vector.VectorEnv.
-    batch_size: int = 50
+    # Set to 0 for auto-tuning based on available CPU cores and n_episodes.
+    batch_size: int = 0
     # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
-    use_async_envs: bool = False
+    # Defaults to True; automatically downgraded to SyncVectorEnv when batch_size=1.
+    use_async_envs: bool = True
 
     def __post_init__(self) -> None:
+        if self.batch_size == 0:
+            self.batch_size = self._auto_batch_size()
         if self.batch_size > self.n_episodes:
-            raise ValueError(
-                "The eval batch size is greater than the number of eval episodes "
-                f"({self.batch_size} > {self.n_episodes}). As a result, {self.batch_size} "
-                f"eval environments will be instantiated, but only {self.n_episodes} will be used. "
-                "This might significantly slow down evaluation. To fix this, you should update your command "
-                f"to increase the number of episodes to match the batch size (e.g. `eval.n_episodes={self.batch_size}`), "
-                f"or lower the batch size (e.g. `eval.batch_size={self.n_episodes}`)."
-            )
+            self.batch_size = self.n_episodes
+
+    def _auto_batch_size(self) -> int:
+        """Pick batch_size based on CPU cores, capped by n_episodes."""
+        import math
+        import os
+
+        cpu_cores = os.cpu_count() or 4
+        # Each async env worker needs ~1 core; leave headroom for main process + inference.
+        by_cpu = max(1, math.floor(cpu_cores * 0.7))
+        return min(by_cpu, self.n_episodes, 64)
 
 
 @dataclass

src/lerobot/datasets/lerobot_dataset.py

@@ -151,9 +151,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 ``$HF_LEROBOT_HOME/hub``.
             episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                 their episode_index in this list. Defaults to None.
-            image_transforms (Callable | None, optional): You can pass standard v2 image transforms from
-                torchvision.transforms.v2 here which will be applied to visual modalities (whether they come
-                from videos or images). Defaults to None.
+            image_transforms (Callable | None, optional):
+                Transform applied to visual modalities inside `__getitem__` after image decoding / tensor
+                conversion. This works for both image-backed and video-backed observations and can later be
+                updated with `set_image_transforms()` or cleared with `clear_image_transforms()`.
+                Defaults to None.
             delta_timestamps (dict[list[float]] | None, optional): _description_. Defaults to None.
             tolerance_s (float, optional): Tolerance in seconds used to ensure data timestamps are actually in
                 sync with the fps value. It is used at the init of the dataset to make sure that each
@@ -192,7 +194,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         super().__init__()
         self.repo_id = repo_id
         self._requested_root = Path(root) if root else None
-        self.image_transforms = image_transforms
+        self.reader = None
+        self.set_image_transforms(image_transforms)
         self.delta_timestamps = delta_timestamps
         self.episodes = episodes
         self.tolerance_s = tolerance_s
@@ -475,6 +478,18 @@ class LeRobotDataset(torch.utils.data.Dataset):
             f"}})"
         )
 
+    def set_image_transforms(self, image_transforms: Callable | None) -> None:
+        """Replace the transform applied to visual observations."""
+        if image_transforms is not None and not callable(image_transforms):
+            raise TypeError("image_transforms must be callable or None.")
+        self.image_transforms = image_transforms
+        if self.reader is not None:
+            self.reader._image_transforms = image_transforms
+
+    def clear_image_transforms(self) -> None:
+        """Remove the transform applied to visual observations."""
+        self.set_image_transforms(None)
+
     # ── Hub methods (stay on facade) ──────────────────────────────────
 
     def push_to_hub(

src/lerobot/datasets/multi_dataset.py

@@ -89,12 +89,24 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
                 )
                 self.disabled_features.update(extra_keys)
 
-        self.image_transforms = image_transforms
         self.delta_timestamps = delta_timestamps
         # TODO(rcadene, aliberts): We should not perform this aggregation for datasets
         # with multiple robots of different ranges. Instead we should have one normalization
         # per robot.
         self.stats = aggregate_stats([dataset.meta.stats for dataset in self._datasets])
+        self.set_image_transforms(image_transforms)
+
+    def set_image_transforms(self, image_transforms: Callable | None) -> None:
+        """Replace the transform for this dataset and its children."""
+        if image_transforms is not None and not callable(image_transforms):
+            raise TypeError("image_transforms must be callable or None.")
+        self.image_transforms = image_transforms
+        for dataset in getattr(self, "_datasets", []):
+            dataset.set_image_transforms(self.image_transforms)
+
+    def clear_image_transforms(self) -> None:
+        """Remove the transform from this dataset and its children."""
+        self.set_image_transforms(None)
 
     @property
     def repo_id_to_index(self):

src/lerobot/envs/configs.py

@@ -12,11 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from __future__ import annotations
+
 import abc
+import importlib
 from dataclasses import dataclass, field, fields
 from typing import Any
 
 import draccus
+import gymnasium as gym
+from gymnasium.envs.registration import registry as gym_registry
 
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.robots import RobotConfig
@@ -39,6 +44,13 @@ from lerobot.utils.constants import (
 )
 
 
+def _make_vec_env_cls(use_async: bool, n_envs: int):
+    """Return the right VectorEnv constructor."""
+    if use_async and n_envs > 1:
+        return gym.vector.AsyncVectorEnv
+    return gym.vector.SyncVectorEnv
+
+
 @dataclass
 class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
     task: str | None = None
@@ -67,6 +79,50 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
     def gym_kwargs(self) -> dict:
         raise NotImplementedError()
 
+    def create_envs(
+        self,
+        n_envs: int,
+        use_async_envs: bool = True,
+    ) -> dict[str, dict[int, gym.vector.VectorEnv]]:
+        """Create {suite: {task_id: VectorEnv}}.
+
+        Default: single-task env via gym.make(). Multi-task benchmarks override.
+        AsyncVectorEnv is the default for n_envs > 1; auto-downgraded to Sync for n_envs=1.
+        """
+        env_cls = gym.vector.AsyncVectorEnv if (use_async_envs and n_envs > 1) else gym.vector.SyncVectorEnv
+
+        if self.gym_id not in gym_registry:
+            print(f"gym id '{self.gym_id}' not found, attempting to import '{self.package_name}'...")
+            try:
+                importlib.import_module(self.package_name)
+            except ModuleNotFoundError as e:
+                raise ModuleNotFoundError(
+                    f"Package '{self.package_name}' required for env '{self.type}' not found. "
+                    f"Please install it or check PYTHONPATH."
+                ) from e
+
+            if self.gym_id not in gym_registry:
+                raise gym.error.NameNotFound(
+                    f"Environment '{self.gym_id}' not registered even after importing '{self.package_name}'."
+                )
+
+        def _make_one():
+            return gym.make(self.gym_id, disable_env_checker=self.disable_env_checker, **self.gym_kwargs)
+
+        try:
+            from gymnasium.vector import AutoresetMode
+
+            vec = env_cls([_make_one for _ in range(n_envs)], autoreset_mode=AutoresetMode.SAME_STEP)
+        except ImportError:
+            vec = env_cls([_make_one for _ in range(n_envs)])
+        return {self.type: {0: vec}}
+
+    def get_env_processors(self):
+        """Return (preprocessor, postprocessor) for this env. Default: identity."""
+        from lerobot.processor.pipeline import PolicyProcessorPipeline
+
+        return PolicyProcessorPipeline(steps=[]), PolicyProcessorPipeline(steps=[])
+
 
 @dataclass
 class HubEnvConfig(EnvConfig):
@@ -338,13 +394,51 @@ class LiberoEnv(EnvConfig):
         else:
             raise ValueError(f"Unsupported obs_type: {self.obs_type}")
 
+        if self.camera_name_mapping is not None:
+            mapped_agentview = self.camera_name_mapping.get("agentview_image", "image")
+            mapped_eye_in_hand = self.camera_name_mapping.get("robot0_eye_in_hand_image", "image2")
+            self.features_map[LIBERO_KEY_PIXELS_AGENTVIEW] = f"{OBS_IMAGES}.{mapped_agentview}"
+            self.features_map[LIBERO_KEY_PIXELS_EYE_IN_HAND] = f"{OBS_IMAGES}.{mapped_eye_in_hand}"
+
     @property
     def gym_kwargs(self) -> dict:
-        kwargs: dict[str, Any] = {"obs_type": self.obs_type, "render_mode": self.render_mode}
+        kwargs: dict[str, Any] = {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+            "observation_height": self.observation_height,
+            "observation_width": self.observation_width,
+        }
         if self.task_ids is not None:
             kwargs["task_ids"] = self.task_ids
         return kwargs
 
+    def create_envs(self, n_envs: int, use_async_envs: bool = True):
+        from lerobot.envs.libero import create_libero_envs
+
+        if self.task is None:
+            raise ValueError("LiberoEnv requires a task to be specified")
+        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
+        return create_libero_envs(
+            task=self.task,
+            n_envs=n_envs,
+            camera_name=self.camera_name,
+            init_states=self.init_states,
+            gym_kwargs=self.gym_kwargs,
+            env_cls=env_cls,
+            control_mode=self.control_mode,
+            episode_length=self.episode_length,
+            camera_name_mapping=self.camera_name_mapping,
+        )
+
+    def get_env_processors(self):
+        from lerobot.processor.env_processor import LiberoProcessorStep
+        from lerobot.processor.pipeline import PolicyProcessorPipeline
+
+        return (
+            PolicyProcessorPipeline(steps=[LiberoProcessorStep()]),
+            PolicyProcessorPipeline(steps=[]),
+        )
+
 
 @EnvConfig.register_subclass("metaworld")
 @dataclass
@@ -387,6 +481,19 @@ class MetaworldEnv(EnvConfig):
             "render_mode": self.render_mode,
         }
 
+    def create_envs(self, n_envs: int, use_async_envs: bool = True):
+        from lerobot.envs.metaworld import create_metaworld_envs
+
+        if self.task is None:
+            raise ValueError("MetaWorld requires a task to be specified")
+        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
+        return create_metaworld_envs(
+            task=self.task,
+            n_envs=n_envs,
+            gym_kwargs=self.gym_kwargs,
+            env_cls=env_cls,
+        )
+
 
 @EnvConfig.register_subclass("isaaclab_arena")
 @dataclass
@@ -454,3 +561,18 @@ class IsaaclabArenaEnv(HubEnvConfig):
     @property
     def gym_kwargs(self) -> dict:
         return {}
+
+    def get_env_processors(self):
+        from lerobot.processor.env_processor import IsaaclabArenaProcessorStep
+        from lerobot.processor.pipeline import PolicyProcessorPipeline
+
+        state_keys = tuple(k.strip() for k in (self.state_keys or "").split(",") if k.strip())
+        camera_keys = tuple(k.strip() for k in (self.camera_keys or "").split(",") if k.strip())
+        if not state_keys and not camera_keys:
+            raise ValueError("At least one of state_keys or camera_keys must be specified.")
+        return (
+            PolicyProcessorPipeline(
+                steps=[IsaaclabArenaProcessorStep(state_keys=state_keys, camera_keys=camera_keys)]
+            ),
+            PolicyProcessorPipeline(steps=[]),
+        )

src/lerobot/envs/factory.py

@@ -13,96 +13,52 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import importlib
+from __future__ import annotations
+
 from typing import Any
 
 import gymnasium as gym
-from gymnasium.envs.registration import registry as gym_registry
 
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.envs.configs import AlohaEnv, EnvConfig, HubEnvConfig, IsaaclabArenaEnv, LiberoEnv, PushtEnv
+from lerobot.envs.configs import EnvConfig, HubEnvConfig
 from lerobot.envs.utils import _call_make_env, _download_hub_file, _import_hub_module, _normalize_hub_result
-from lerobot.policies.xvla.configuration_xvla import XVLAConfig
-from lerobot.processor import ProcessorStep
-from lerobot.processor.env_processor import IsaaclabArenaProcessorStep, LiberoProcessorStep
-from lerobot.processor.pipeline import PolicyProcessorPipeline
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:
-    if env_type == "aloha":
-        return AlohaEnv(**kwargs)
-    elif env_type == "pusht":
-        return PushtEnv(**kwargs)
-    elif env_type == "libero":
-        return LiberoEnv(**kwargs)
-    else:
-        raise ValueError(f"Policy type '{env_type}' is not available.")
+    try:
+        cls = EnvConfig.get_choice_class(env_type)
+    except KeyError as err:
+        raise ValueError(
+            f"Environment type '{env_type}' is not registered. "
+            f"Available: {list(EnvConfig.get_known_choices().keys())}"
+        ) from err
+    return cls(**kwargs)
 
 
 def make_env_pre_post_processors(
     env_cfg: EnvConfig,
-    policy_cfg: PreTrainedConfig,
-) -> tuple[
-    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
-    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
-]:
+    policy_cfg: Any,
+) -> tuple[Any, Any]:
     """
     Create preprocessor and postprocessor pipelines for environment observations.
 
-    This function creates processor pipelines that transform raw environment
-    observations and actions. By default, it returns identity processors that do nothing.
-    For specific environments like LIBERO, it adds environment-specific processing steps.
-
-    Args:
-        env_cfg: The configuration of the environment.
-
-    Returns:
-        A tuple containing:
-            - preprocessor: Pipeline that processes environment observations
-            - postprocessor: Pipeline that processes environment outputs (currently identity)
+    Returns a tuple of (preprocessor, postprocessor). By default, delegates to
+    ``env_cfg.get_env_processors()``.  The XVLAConfig policy-specific override
+    stays here because it depends on the *policy* config, not the env config.
     """
-    # Preprocessor and Postprocessor steps are Identity for most environments
-    preprocessor_steps: list[ProcessorStep] = []
-    postprocessor_steps: list[ProcessorStep] = []
+    from lerobot.policies.xvla.configuration_xvla import XVLAConfig
+
     if isinstance(policy_cfg, XVLAConfig):
         from lerobot.policies.xvla.processor_xvla import make_xvla_libero_pre_post_processors
 
         return make_xvla_libero_pre_post_processors()
 
-    # For LIBERO environments, add the LiberoProcessorStep to preprocessor
-    if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
-        preprocessor_steps.append(LiberoProcessorStep())
-
-    # For Isaaclab Arena environments, add the IsaaclabArenaProcessorStep
-    if isinstance(env_cfg, IsaaclabArenaEnv) or "isaaclab_arena" in env_cfg.type:
-        # Parse comma-separated keys (handle None for state-based policies)
-        if env_cfg.state_keys:
-            state_keys = tuple(k.strip() for k in env_cfg.state_keys.split(",") if k.strip())
-        else:
-            state_keys = ()
-        if env_cfg.camera_keys:
-            camera_keys = tuple(k.strip() for k in env_cfg.camera_keys.split(",") if k.strip())
-        else:
-            camera_keys = ()
-        if not state_keys and not camera_keys:
-            raise ValueError("At least one of state_keys or camera_keys must be specified.")
-        preprocessor_steps.append(
-            IsaaclabArenaProcessorStep(
-                state_keys=state_keys,
-                camera_keys=camera_keys,
-            )
-        )
-
-    preprocessor = PolicyProcessorPipeline(steps=preprocessor_steps)
-    postprocessor = PolicyProcessorPipeline(steps=postprocessor_steps)
-
-    return preprocessor, postprocessor
+    return env_cfg.get_env_processors()
 
 
 def make_env(
     cfg: EnvConfig | str,
     n_envs: int = 1,
-    use_async_envs: bool = False,
+    use_async_envs: bool = True,
     hub_cache_dir: str | None = None,
     trust_remote_code: bool = False,
 ) -> dict[str, dict[int, gym.vector.VectorEnv]]:
@@ -163,57 +119,4 @@ def make_env(
     if n_envs < 1:
         raise ValueError("`n_envs` must be at least 1")
 
-    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
-
-    if "libero" in cfg.type:
-        from lerobot.envs.libero import create_libero_envs
-
-        if cfg.task is None:
-            raise ValueError("LiberoEnv requires a task to be specified")
-
-        return create_libero_envs(
-            task=cfg.task,
-            n_envs=n_envs,
-            camera_name=cfg.camera_name,
-            init_states=cfg.init_states,
-            gym_kwargs=cfg.gym_kwargs,
-            env_cls=env_cls,
-            control_mode=cfg.control_mode,
-            episode_length=cfg.episode_length,
-        )
-    elif "metaworld" in cfg.type:
-        from lerobot.envs.metaworld import create_metaworld_envs
-
-        if cfg.task is None:
-            raise ValueError("MetaWorld requires a task to be specified")
-
-        return create_metaworld_envs(
-            task=cfg.task,
-            n_envs=n_envs,
-            gym_kwargs=cfg.gym_kwargs,
-            env_cls=env_cls,
-        )
-
-    if cfg.gym_id not in gym_registry:
-        print(f"gym id '{cfg.gym_id}' not found, attempting to import '{cfg.package_name}'...")
-        try:
-            importlib.import_module(cfg.package_name)
-        except ModuleNotFoundError as e:
-            raise ModuleNotFoundError(
-                f"Package '{cfg.package_name}' required for env '{cfg.type}' not found. "
-                f"Please install it or check PYTHONPATH."
-            ) from e
-
-        if cfg.gym_id not in gym_registry:
-            raise gym.error.NameNotFound(
-                f"Environment '{cfg.gym_id}' not registered even after importing '{cfg.package_name}'."
-            )
-
-    def _make_one():
-        return gym.make(cfg.gym_id, disable_env_checker=cfg.disable_env_checker, **(cfg.gym_kwargs or {}))
-
-    vec = env_cls([_make_one for _ in range(n_envs)], autoreset_mode=gym.vector.AutoresetMode.SAME_STEP)
-
-    # normalize to {suite: {task_id: vec_env}} for consistency
-    suite_name = cfg.type  # e.g., "pusht", "aloha"
-    return {suite_name: {0: vec}}
+    return cfg.create_envs(n_envs=n_envs, use_async_envs=use_async_envs)

src/lerobot/envs/libero.py

@@ -29,6 +29,7 @@ from gymnasium import spaces
 from libero.libero import benchmark, get_libero_path
 from libero.libero.envs import OffScreenRenderEnv
 
+from lerobot.envs.utils import _LazyAsyncVectorEnv
 from lerobot.types import RobotObservation
 
 
@@ -150,7 +151,17 @@ class LiberoEnv(gym.Env):
 
         self.init_state_id = self.episode_index  # tie each sub-env to a fixed init state
 
-        self._env = self._make_envs_task(task_suite, self.task_id)
+        # Extract task metadata without allocating GPU resources (safe before fork).
+        task = task_suite.get_task(task_id)
+        self.task = task.name
+        self.task_description = task.language
+        self._task_bddl_file = os.path.join(
+            get_libero_path("bddl_files"), task.problem_folder, task.bddl_file
+        )
+        self._env: OffScreenRenderEnv | None = (
+            None  # deferred — created on first reset() inside the worker subprocess
+        )
+
         default_steps = 500
         self._max_episode_steps = (
             TASK_SUITE_MAX_STEPS.get(task_suite_name, default_steps)
@@ -221,28 +232,33 @@ class LiberoEnv(gym.Env):
             low=ACTION_LOW, high=ACTION_HIGH, shape=(ACTION_DIM,), dtype=np.float32
         )
 
+    def _ensure_env(self) -> None:
+        """Create the underlying OffScreenRenderEnv on first use.
+
+        Called inside the worker subprocess after fork(), so each worker gets
+        its own clean EGL context rather than inheriting a stale one from the
+        parent process (which causes EGL_BAD_CONTEXT crashes with AsyncVectorEnv).
+        """
+        if self._env is not None:
+            return
+        env = OffScreenRenderEnv(
+            bddl_file_name=self._task_bddl_file,
+            camera_heights=self.observation_height,
+            camera_widths=self.observation_width,
+        )
+        env.reset()
+        self._env = env
+
     def render(self):
+        self._ensure_env()
         raw_obs = self._env.env._get_observations()
-        image = self._format_raw_obs(raw_obs)["pixels"]["image"]
+        pixels = self._format_raw_obs(raw_obs)["pixels"]
+        image = next(iter(pixels.values()))
         image = image[::-1, ::-1]  # flip both H and W for visualization
         return image
 
-    def _make_envs_task(self, task_suite: Any, task_id: int = 0):
-        task = task_suite.get_task(task_id)
-        self.task = task.name
-        self.task_description = task.language
-        task_bddl_file = os.path.join(get_libero_path("bddl_files"), task.problem_folder, task.bddl_file)
-
-        env_args = {
-            "bddl_file_name": task_bddl_file,
-            "camera_heights": self.observation_height,
-            "camera_widths": self.observation_width,
-        }
-        env = OffScreenRenderEnv(**env_args)
-        env.reset()
-        return env
-
     def _format_raw_obs(self, raw_obs: RobotObservation) -> RobotObservation:
+        assert self._env is not None, "_format_raw_obs called before _ensure_env()"
         images = {}
         for camera_name in self.camera_name:
             image = raw_obs[camera_name]
@@ -294,6 +310,7 @@ class LiberoEnv(gym.Env):
         )
 
     def reset(self, seed=None, **kwargs):
+        self._ensure_env()
         super().reset(seed=seed)
         self._env.seed(seed)
         raw_obs = self._env.reset()
@@ -320,6 +337,8 @@ class LiberoEnv(gym.Env):
         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,)), "
@@ -339,18 +358,13 @@ class LiberoEnv(gym.Env):
         )
         observation = self._format_raw_obs(raw_obs)
         if terminated:
-            info["final_info"] = {
-                "task": self.task,
-                "task_id": self.task_id,
-                "done": bool(done),
-                "is_success": bool(is_success),
-            }
             self.reset()
         truncated = False
         return observation, reward, terminated, truncated, info
 
     def close(self):
-        self._env.close()
+        if self._env is not None:
+            self._env.close()
 
 
 def _make_env_fns(
@@ -364,6 +378,7 @@ def _make_env_fns(
     init_states: bool,
     gym_kwargs: Mapping[str, Any],
     control_mode: str,
+    camera_name_mapping: dict[str, str] | None = None,
 ) -> list[Callable[[], LiberoEnv]]:
     """Build n_envs factory callables for a single (suite, task_id)."""
 
@@ -379,6 +394,7 @@ def _make_env_fns(
             episode_index=episode_index,
             n_envs=n_envs,
             control_mode=control_mode,
+            camera_name_mapping=camera_name_mapping,
             **local_kwargs,
         )
 
@@ -400,6 +416,7 @@ def create_libero_envs(
     env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
     control_mode: str = "relative",
     episode_length: int | None = None,
+    camera_name_mapping: dict[str, str] | None = None,
 ) -> dict[str, dict[int, Any]]:
     """
     Create vectorized LIBERO environments with a consistent return shape.
@@ -430,6 +447,8 @@ def create_libero_envs(
     if task_ids_filter is not None:
         print(f"Restricting to task_ids={task_ids_filter}")
 
+    is_async = env_cls is gym.vector.AsyncVectorEnv
+
     out: dict[str, dict[int, Any]] = defaultdict(dict)
     for suite_name in suite_names:
         suite = _get_suite(suite_name)
@@ -438,6 +457,11 @@ def create_libero_envs(
         if not selected:
             raise ValueError(f"No tasks selected for suite '{suite_name}' (available: {total}).")
 
+        # All tasks in a suite share identical observation/action spaces.
+        # 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
+
         for tid in selected:
             fns = _make_env_fns(
                 suite=suite,
@@ -449,9 +473,16 @@ def create_libero_envs(
                 init_states=init_states,
                 gym_kwargs=gym_kwargs,
                 control_mode=control_mode,
+                camera_name_mapping=camera_name_mapping,
             )
-            out[suite_name][tid] = env_cls(fns)
+            if is_async:
+                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
+                out[suite_name][tid] = lazy
+            else:
+                out[suite_name][tid] = env_cls(fns)
             print(f"Built vec env | suite={suite_name} | task_id={tid} | n_envs={n_envs}")
 
-    # return plain dicts for predictability
     return {suite: dict(task_map) for suite, task_map in out.items()}

src/lerobot/envs/metaworld.py

@@ -25,6 +25,7 @@ import metaworld.policies as policies
 import numpy as np
 from gymnasium import spaces
 
+from lerobot.envs.utils import _LazyAsyncVectorEnv
 from lerobot.types import RobotObservation
 
 # ---- Load configuration data from the external JSON file ----
@@ -97,8 +98,9 @@ class MetaworldEnv(gym.Env):
         self.visualization_height = visualization_height
         self.camera_name = camera_name
 
-        self._env = self._make_envs_task(self.task)
-        self._max_episode_steps = self._env.max_path_length
+        self._env_name = self.task  # already stripped of "metaworld-" prefix above
+        self._env = None  # deferred — created on first reset() inside the worker subprocess
+        self._max_episode_steps = 500  # MT1 environments always have max_path_length=500
         self.task_description = TASK_DESCRIPTIONS[self.task]
 
         self.expert_policy = TASK_POLICY_MAPPING[self.task]()
@@ -136,6 +138,24 @@ class MetaworldEnv(gym.Env):
 
         self.action_space = spaces.Box(low=-1, high=1, shape=(ACTION_DIM,), dtype=np.float32)
 
+    def _ensure_env(self) -> None:
+        """Create the underlying MetaWorld 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).
+        """
+        if self._env is not None:
+            return
+        mt1 = metaworld.MT1(self._env_name, seed=42)
+        env = mt1.train_classes[self._env_name](render_mode="rgb_array", camera_name=self.camera_name)
+        env.set_task(mt1.train_tasks[0])
+        if self.camera_name == "corner2":
+            env.model.cam_pos[2] = [0.75, 0.075, 0.7]
+        env.reset()
+        env._freeze_rand_vec = False  # otherwise no randomization
+        self._env = env
+
     def render(self) -> np.ndarray:
         """
         Render the current environment frame.
@@ -143,26 +163,13 @@ class MetaworldEnv(gym.Env):
         Returns:
             np.ndarray: The rendered RGB image from the environment.
         """
+        self._ensure_env()
         image = self._env.render()
         if self.camera_name == "corner2":
             # Images from this camera are flipped — correct them
             image = np.flip(image, (0, 1))
         return image
 
-    def _make_envs_task(self, env_name: str):
-        mt1 = metaworld.MT1(env_name, seed=42)
-        env = mt1.train_classes[env_name](render_mode="rgb_array", camera_name=self.camera_name)
-        env.set_task(mt1.train_tasks[0])
-        if self.camera_name == "corner2":
-            env.model.cam_pos[2] = [
-                0.75,
-                0.075,
-                0.7,
-            ]  # corner2 position, similar to https://arxiv.org/pdf/2206.14244
-        env.reset()
-        env._freeze_rand_vec = False  # otherwise no randomization
-        return env
-
     def _format_raw_obs(self, raw_obs: np.ndarray) -> RobotObservation:
         image = None
         if self._env is not None:
@@ -209,6 +216,7 @@ class MetaworldEnv(gym.Env):
             observation (RobotObservation): The initial formatted observation.
             info (Dict[str, Any]): Additional info about the reset state.
         """
+        self._ensure_env()
         super().reset(seed=seed)
 
         raw_obs, info = self._env.reset(seed=seed)
@@ -232,6 +240,7 @@ class MetaworldEnv(gym.Env):
             truncated (bool): Whether the episode was truncated due to a time limit.
             info (Dict[str, Any]): Additional environment info.
         """
+        self._ensure_env()
         if action.ndim != 1:
             raise ValueError(
                 f"Expected action to be 1-D (shape (action_dim,)), "
@@ -263,7 +272,8 @@ class MetaworldEnv(gym.Env):
         return observation, reward, terminated, truncated, info
 
     def close(self):
-        self._env.close()
+        if self._env is not None:
+            self._env.close()
 
 
 # ---- Main API ----------------------------------------------------------------
@@ -297,6 +307,9 @@ def create_metaworld_envs(
 
     print(f"Creating Meta-World envs | task_groups={task_groups} | n_envs(per task)={n_envs}")
 
+    is_async = env_cls is gym.vector.AsyncVectorEnv
+    cached_obs_space = None
+    cached_act_space = None
     out: dict[str, dict[int, Any]] = defaultdict(dict)
 
     for group in task_groups:
@@ -309,7 +322,14 @@ def create_metaworld_envs(
             # build n_envs factories
             fns = [(lambda tn=task_name: MetaworldEnv(task=tn, **gym_kwargs)) for _ in range(n_envs)]
 
-            out[group][tid] = env_cls(fns)
+            if is_async:
+                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
+                out[group][tid] = lazy
+            else:
+                out[group][tid] = env_cls(fns)
 
     # return a plain dict for consistency
     return {group: dict(task_map) for group, task_map in out.items()}

src/lerobot/envs/utils.py

@@ -16,7 +16,7 @@
 import importlib.util
 import os
 import warnings
-from collections.abc import Mapping, Sequence
+from collections.abc import Callable, Mapping, Sequence
 from functools import singledispatch
 from typing import Any
 
@@ -130,56 +130,99 @@ def env_to_policy_features(env_cfg: EnvConfig) -> dict[str, PolicyFeature]:
     return policy_features
 
 
-def are_all_envs_same_type(env: gym.vector.VectorEnv) -> bool:
-    first_type = type(env.envs[0])  # Get type of first env
-    return all(type(e) is first_type for e in env.envs)  # Fast type check
+def _sub_env_has_attr(env: gym.vector.VectorEnv, attr: str) -> bool:
+    try:
+        env.get_attr(attr)
+        return True
+    except (AttributeError, Exception):
+        return False
+
+
+class _LazyAsyncVectorEnv:
+    """Defers AsyncVectorEnv creation until first use.
+
+    Creating all tasks' AsyncVectorEnvs upfront spawns N_tasks × n_envs worker
+    processes, all of which allocate EGL/GPU resources immediately. Since tasks
+    are evaluated sequentially, only one task's workers need to be alive at a
+    time. This wrapper stores the factory functions and creates the real
+    AsyncVectorEnv on first reset()/step()/call(), keeping peak process count = n_envs.
+    """
+
+    def __init__(
+        self,
+        env_fns: list[Callable],
+        observation_space=None,
+        action_space=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:
+            self.observation_space = observation_space
+            self.action_space = action_space
+        else:
+            tmp = env_fns[0]()
+            self.observation_space = tmp.observation_space
+            self.action_space = tmp.action_space
+            tmp.close()
+        self.single_observation_space = self.observation_space
+        self.single_action_space = self.action_space
+
+    def _ensure(self) -> None:
+        if self._env is None:
+            self._env = gym.vector.AsyncVectorEnv(self._env_fns, context="forkserver", shared_memory=True)
+
+    def reset(self, **kwargs):
+        self._ensure()
+        return self._env.reset(**kwargs)
+
+    def step(self, actions):
+        self._ensure()
+        return self._env.step(actions)
+
+    def call(self, name, *args, **kwargs):
+        self._ensure()
+        return self._env.call(name, *args, **kwargs)
+
+    def get_attr(self, name):
+        self._ensure()
+        return self._env.get_attr(name)
+
+    def close(self) -> None:
+        if self._env is not None:
+            self._env.close()
+            self._env = None
 
 
 def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
     with warnings.catch_warnings():
-        warnings.simplefilter("once", UserWarning)  # Apply filter only in this function
+        warnings.simplefilter("once", UserWarning)
 
-        if not (hasattr(env.envs[0], "task_description") and hasattr(env.envs[0], "task")):
+        if not (_sub_env_has_attr(env, "task_description") and _sub_env_has_attr(env, "task")):
             warnings.warn(
                 "The environment does not have 'task_description' and 'task'. Some policies require these features.",
                 UserWarning,
                 stacklevel=2,
             )
-        if not are_all_envs_same_type(env):
-            warnings.warn(
-                "The environments have different types. Make sure you infer the right task from each environment. Empty task will be passed instead.",
-                UserWarning,
-                stacklevel=2,
-            )
 
 
 def add_envs_task(env: gym.vector.VectorEnv, observation: RobotObservation) -> RobotObservation:
     """Adds task feature to the observation dict with respect to the first environment attribute."""
-    if hasattr(env.envs[0], "task_description"):
-        task_result = env.call("task_description")
+    if _sub_env_has_attr(env, "task_description"):
+        task_result = list(env.call("task_description"))
 
-        if isinstance(task_result, tuple):
-            task_result = list(task_result)
-
-        if not isinstance(task_result, list):
-            raise TypeError(f"Expected task_description to return a list, got {type(task_result)}")
         if not all(isinstance(item, str) for item in task_result):
             raise TypeError("All items in task_description result must be strings")
 
         observation["task"] = task_result
-    elif hasattr(env.envs[0], "task"):
-        task_result = env.call("task")
+    elif _sub_env_has_attr(env, "task"):
+        task_result = list(env.call("task"))
 
-        if isinstance(task_result, tuple):
-            task_result = list(task_result)
-
-        if not isinstance(task_result, list):
-            raise TypeError(f"Expected task to return a list, got {type(task_result)}")
         if not all(isinstance(item, str) for item in task_result):
             raise TypeError("All items in task result must be strings")
 
         observation["task"] = task_result
-    else:  #  For envs without language instructions, e.g. aloha transfer cube and etc.
+    else:
         num_envs = observation[list(observation.keys())[0]].shape[0]
         observation["task"] = ["" for _ in range(num_envs)]
     return observation

src/lerobot/policies/multi_task_dit/README.md

@@ -1,37 +0,0 @@
-# Multitask DiT Policy
-
-## Citation
-
-If you use this work, please cite the following works:
-
-```bibtex
-@misc{jones2025multitaskditpolicy,
-  author = {Bryson Jones},
-  title = {Dissecting and Open-Sourcing Multitask Diffusion Transformer Policy},
-  year = {2025},
-  url = {https://brysonkjones.substack.com/p/dissecting-and-open-sourcing-multitask-diffusion-transformer-policy},
-  note = {Blog post}
-}
-```
-
-```bibtex
-@misc{trilbmteam2025carefulexaminationlargebehaviormodels,
-  author       = {TRI LBM Team},
-  title        = {A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation},
-  year         = {2025},
-  eprint       = {arXiv:2507.05331},
-  archivePrefix = {arXiv},
-  primaryClass = {cs.RO},
-  url          = {https://arxiv.org/abs/2507.05331}
-}
-```
-
-```bibtex
-@misc{bostondynamics2025largebehaviormodelsatlas,
-  author       = {Boston Dynamics and TRI Research Team},
-  title        = {Large Behavior Models and Atlas Find New Footing},
-  year         = {2025},
-  url          = {https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/},
-  note         = {Blog post}
-}
-```

src/lerobot/policies/multi_task_dit/README.md

@@ -0,0 +1 @@
+../../../../docs/source/policy_multi_task_dit_README.md

src/lerobot/policies/pi0/README.md

@@ -1,108 +0,0 @@
-# π₀ (pi0)
-
-This repository contains the Hugging Face port of **π₀**, adapted from [OpenPI](https://github.com/Physical-Intelligence/openpi) by the Physical Intelligence.
-It is designed as a **Vision-Language-Action model for general robot control**.
-
----
-
-## Model Overview
-
-| Feature              | π₀                                                     | π₀.₅                                      |
-| -------------------- | ------------------------------------------------------ | ----------------------------------------- |
-| Time Conditioning    | Concatenates time with actions via `action_time_mlp_*` | Uses `time_mlp_*` for AdaRMS conditioning |
-| AdaRMS               | Not used                                               | Used in action expert                     |
-| Tokenizer Length     | 48 tokens                                              | 200 tokens                                |
-| Discrete State Input | False (Uses `state_proj` layer)                        | True                                      |
-| Parameter Count      | Higher (includes state embedding)                      | Lower (no state embedding)                |
-
----
-
-## Relative Actions
-
-π₀ supports training with **relative actions**, where the model learns relative offsets
-from the current robot state instead of absolute joint positions. This mirrors the
-relative-action transform in OpenPI (`DeltaActions`) and can improve performance.
-
-### How it works
-
-1. **During preprocessing**, absolute actions are converted to relative offsets:
-   `relative = action - state` (for selected joints).
-2. The relative actions are normalized using statistics computed from the relative distribution.
-3. **During postprocessing**, predicted relative actions are converted back to absolute:
-   `absolute = relative + state`.
-
-Joints listed in `relative_exclude_joints` (e.g., gripper) are kept absolute.
-
-### Configuration
-
-| Parameter                 | Type        | Default       | Description                                                      |
-| ------------------------- | ----------- | ------------- | ---------------------------------------------------------------- |
-| `use_relative_actions`    | `bool`      | `False`       | Enable relative-action training                                  |
-| `relative_exclude_joints` | `list[str]` | `["gripper"]` | Joint names to keep absolute (matched by substring)              |
-| `action_feature_names`    | `list[str]` | `None`        | Auto-populated from dataset metadata at runtime by `make_policy` |
-
-### Training example
-
-```bash
-python -m lerobot.scripts.lerobot_train \
-  --policy.type=pi0 \
-  --dataset.repo_id=your_org/your_dataset \
-  --policy.use_relative_actions=true \
-  --policy.relative_exclude_joints='["gripper"]'
-```
-
-When `use_relative_actions=true`, the training script automatically:
-
-- Computes relative action statistics from the dataset (sampled chunk-level relative actions)
-- Replaces the standard action stats with relative stats for normalization
-- Broadcasts these stats across all ranks in distributed training
-
-### Recomputing stats for an existing dataset
-
-If you want to precompute relative action stats offline, use `recompute_stats` from
-`lerobot.datasets.dataset_tools`:
-
-```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.dataset_tools import recompute_stats
-
-dataset = LeRobotDataset("your_org/your_dataset")
-dataset = recompute_stats(
-    dataset,
-    relative_action=True,
-    relative_exclude_joints=["gripper"],
-)
-```
-
----
-
-## Citation
-
-If you use this work, please cite both **OpenPI** and the π₀ paper:
-
-```bibtex
-@misc{openpi2024,
-  author       = {Physical Intelligence Lab},
-  title        = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
-  year         = {2024},
-  publisher    = {GitHub},
-  howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
-  license      = {Apache-2.0}
-}
-
-@misc{black2024pi0visionlanguageactionflowmodel,
-  title        = {π₀: A Vision-Language-Action Flow Model for General Robot Control},
-  author       = {Kevin Black and Noah Brown and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Lucy Xiaoyang Shi and James Tanner and Quan Vuong and Anna Walling and Haohuan Wang and Ury Zhilinsky},
-  year         = {2024},
-  eprint       = {2410.24164},
-  archivePrefix= {arXiv},
-  primaryClass = {cs.LG},
-  url          = {https://arxiv.org/abs/2410.24164},
-}
-```
-
----
-
-## License
-
-This port follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).

src/lerobot/policies/pi0/README.md

@@ -0,0 +1 @@
+../../../../docs/source/policy_pi0_README.md

src/lerobot/policies/pi05/README.md

@@ -1,91 +0,0 @@
-# π₀.₅ (pi05)
-
-This repository contains the Hugging Face port of **π₀.₅**, adapted from [OpenPI](https://github.com/Physical-Intelligence/openpi) by the Physical Intelligence.
-It is designed as a **Vision-Language-Action model with open-world generalization**.
-
----
-
-## Model Overview
-
-| Feature              | π₀                                                     | π₀.₅                                      |
-| -------------------- | ------------------------------------------------------ | ----------------------------------------- |
-| Time Conditioning    | Concatenates time with actions via `action_time_mlp_*` | Uses `time_mlp_*` for AdaRMS conditioning |
-| AdaRMS               | Not used                                               | Used in action expert                     |
-| Tokenizer Length     | 48 tokens                                              | 200 tokens                                |
-| Discrete State Input | False (Uses `state_proj` layer)                        | True                                      |
-| Parameter Count      | Higher (includes state embedding)                      | Lower (no state embedding)                |
-
----
-
-## Relative Actions
-
-π₀.₅ supports training with **relative actions**, where the model learns relative offsets
-from the current robot state instead of absolute joint positions. This mirrors the
-relative-action transform in OpenPI (`DeltaActions`) and can improve performance.
-
-### How it works
-
-1. **During preprocessing**, absolute actions are converted to relative offsets:
-   `relative = action - state` (for selected joints).
-2. The relative actions are normalized using statistics computed from the relative distribution.
-3. **During postprocessing**, predicted relative actions are converted back to absolute:
-   `absolute = relative + state`.
-
-Joints listed in `relative_exclude_joints` (e.g., gripper) are kept absolute.
-
-### Configuration
-
-| Parameter                 | Type        | Default       | Description                                                      |
-| ------------------------- | ----------- | ------------- | ---------------------------------------------------------------- |
-| `use_relative_actions`    | `bool`      | `False`       | Enable relative-action training                                  |
-| `relative_exclude_joints` | `list[str]` | `["gripper"]` | Joint names to keep absolute (matched by substring)              |
-| `action_feature_names`    | `list[str]` | `None`        | Auto-populated from dataset metadata at runtime by `make_policy` |
-
-### Training example
-
-```bash
-python -m lerobot.scripts.lerobot_train \
-  --policy.type=pi05 \
-  --dataset.repo_id=your_org/your_dataset \
-  --policy.use_relative_actions=true \
-  --policy.relative_exclude_joints='["gripper"]'
-```
-
-When `use_relative_actions=true`, the training script automatically:
-
-- Computes relative action statistics from the dataset (sampled chunk-level relative actions)
-- Replaces the standard action stats with relative stats for normalization
-- Broadcasts these stats across all ranks in distributed training
-
----
-
-## Citation
-
-If you use this work, please cite both **OpenPI** and the π₀.₅ paper:
-
-```bibtex
-@misc{openpi2024,
-  author       = {Physical Intelligence Lab},
-  title        = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
-  year         = {2024},
-  publisher    = {GitHub},
-  howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
-  license      = {Apache-2.0}
-}
-
-@misc{intelligence2025pi05visionlanguageactionmodelopenworld,
-  title        = {π₀.₅: a Vision-Language-Action Model with Open-World Generalization},
-  author       = {Physical Intelligence and Kevin Black and Noah Brown and James Darpinian and Karan Dhabalia and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Manuel Y. Galliker and Dibya Ghosh and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Devin LeBlanc and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Allen Z. Ren and Lucy Xiaoyang Shi and Laura Smith and Jost Tobias Springenberg and Kyle Stachowicz and James Tanner and Quan Vuong and Homer Walke and Anna Walling and Haohuan Wang and Lili Yu and Ury Zhilinsky},
-  year         = {2025},
-  eprint       = {2504.16054},
-  archivePrefix= {arXiv},
-  primaryClass = {cs.LG},
-  url          = {https://arxiv.org/abs/2504.16054},
-}
-```
-
----
-
-## License
-
-This port follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).

src/lerobot/policies/pi05/README.md

@@ -0,0 +1 @@
+../../../../docs/source/policy_pi05_README.md

src/lerobot/policies/rtc/README.md

@@ -1,38 +0,0 @@
-# Real-Time Chunking (RTC)
-
-This module contains the LeRobot implementation of **Real-Time Chunking (RTC)**, an inference-time technique for flow-matching based policies.
-
-**Note**: RTC is not a policy itself, but rather an inference enhancement that works with flow-matching based policies including [π₀](../pi0/), [π₀.₅](../pi05/), and [SmolVLA](../smolvla/).
-
----
-
-## Citation
-
-If you use Real-Time Chunking in your work, please cite:
-
-```bibtex
-@misc{openpi2024,
-  author       = {Physical Intelligence Lab},
-  title        = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
-  year         = {2024},
-  publisher    = {GitHub},
-  howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
-  license      = {Apache-2.0}
-}
-
-@misc{black2025realtimeexecutionactionchunking,
-      title={Real-Time Execution of Action Chunking Flow Policies},
-      author={Kevin Black and Manuel Y. Galliker and Sergey Levine},
-      year={2025},
-      eprint={2506.07339},
-      archivePrefix={arXiv},
-      primaryClass={cs.RO},
-      url={https://arxiv.org/abs/2506.07339},
-}
-```
-
----
-
-## License
-
-This implementation follows the **Apache 2.0 License**, consistent with the LeRobot project.

src/lerobot/policies/rtc/README.md

@@ -0,0 +1 @@
+../../../../docs/source/policy_rtc_README.md

src/lerobot/policies/rtc/__init__.py

@@ -0,0 +1,29 @@
+# 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.
+
+"""Real-Time Chunking (RTC) utilities for action-chunking policies."""
+
+from lerobot.policies.rtc.action_interpolator import ActionInterpolator
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.policies.rtc.modeling_rtc import RTCProcessor
+
+__all__ = [
+    "ActionInterpolator",
+    "ActionQueue",
+    "LatencyTracker",
+    "RTCConfig",
+    "RTCProcessor",
+]

src/lerobot/policies/rtc/action_interpolator.py

@@ -0,0 +1,116 @@
+# 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.
+
+"""Action interpolation for smoother robot control.
+
+Provides configurable Nx control rate by interpolating between consecutive actions.
+Useful with RTC and action-chunking policies to reduce jerkiness.
+"""
+
+from torch import Tensor
+
+
+class ActionInterpolator:
+    """Interpolates between consecutive actions for smoother control.
+
+    When enabled with multiplier N, produces N actions per policy action
+    by linearly interpolating between the previous and current action.
+
+    Example with multiplier=3:
+        prev_action -> [1/3 interpolated, 2/3 interpolated, current_action]
+
+    This effectively multiplies the control rate for smoother motion.
+
+    Usage:
+        interpolator = ActionInterpolator(multiplier=2)  # 2x control rate
+
+        # In control loop:
+        if interpolator.needs_new_action():
+            new_action = queue.get()
+            if new_action:
+                interpolator.add(new_action.cpu())
+
+        action = interpolator.get()
+        if action:
+            robot.send_action(action)
+    """
+
+    def __init__(self, multiplier: int = 1):
+        """Initialize the interpolator.
+
+        Args:
+            multiplier: Control rate multiplier (1 = no interpolation, 2 = 2x, 3 = 3x, etc.)
+        """
+        if multiplier < 1:
+            raise ValueError(f"multiplier must be >= 1, got {multiplier}")
+        self.multiplier = multiplier
+        self._prev: Tensor | None = None
+        self._buffer: list[Tensor] = []
+        self._idx = 0
+
+    @property
+    def enabled(self) -> bool:
+        """Whether interpolation is active (multiplier > 1)."""
+        return self.multiplier > 1
+
+    def reset(self):
+        """Reset interpolation state (call between episodes)."""
+        self._prev = None
+        self._buffer = []
+        self._idx = 0
+
+    def needs_new_action(self) -> bool:
+        """Check if a new action is needed from the queue."""
+        return self._idx >= len(self._buffer)
+
+    def add(self, action: Tensor) -> None:
+        """Add a new action and compute interpolated sequence.
+
+        Args:
+            action: New action tensor from policy/queue (already on CPU).
+        """
+        if self.multiplier > 1 and self._prev is not None:
+            self._buffer = []
+            for i in range(1, self.multiplier + 1):
+                t = i / self.multiplier
+                interp = self._prev + t * (action - self._prev)
+                self._buffer.append(interp)
+        else:
+            # First step: no previous action yet, so run at base FPS without interpolation.
+            self._buffer = [action.clone()]
+        self._prev = action.clone()
+        self._idx = 0
+
+    def get(self) -> Tensor | None:
+        """Get the next interpolated action.
+
+        Returns:
+            Next action tensor, or None if buffer is exhausted.
+        """
+        if self._idx >= len(self._buffer):
+            return None
+        action = self._buffer[self._idx]
+        self._idx += 1
+        return action
+
+    def get_control_interval(self, fps: float) -> float:
+        """Get the control interval based on interpolation multiplier.
+
+        Args:
+            fps: Base frames per second.
+
+        Returns:
+            Control interval in seconds (divided by multiplier).
+        """
+        return 1.0 / (fps * self.multiplier)

src/lerobot/policies/rtc/action_queue.py

@@ -79,6 +79,13 @@ class ActionQueue:
             self.last_index += 1
             return action.clone()
 
+    def clear(self) -> None:
+        """Clear queued actions and reset consumption index."""
+        with self.lock:
+            self.queue = None
+            self.original_queue = None
+            self.last_index = 0
+
     def qsize(self) -> int:
         """Get the number of remaining actions in the queue.
 
@@ -123,14 +130,26 @@ class ActionQueue:
         with self.lock:
             if self.original_queue is None:
                 return None
-            return self.original_queue[self.last_index :]
+            return self.original_queue[self.last_index :].clone()
+
+    def get_processed_left_over(self) -> Tensor | None:
+        """Get leftover processed actions (the actions currently executed by the robot).
+
+        Returns:
+            Tensor | None: Remaining processed actions (remaining_steps, action_dim),
+                or None if no processed queue exists.
+        """
+        with self.lock:
+            if self.queue is None:
+                return None
+            return self.queue[self.last_index :].clone()
 
     def merge(
         self,
         original_actions: Tensor,
         processed_actions: Tensor,
         real_delay: int,
-        action_index_before_inference: int | None = 0,
+        action_index_before_inference: int | None = None,
     ):
         """Merge new actions into the queue.
 
@@ -145,10 +164,10 @@ class ActionQueue:
             action_index_before_inference: Index before inference started, for validation.
         """
         with self.lock:
-            self._check_delays(real_delay, action_index_before_inference)
+            delay = self._check_and_resolve_delays(real_delay, action_index_before_inference)
 
             if self.cfg.enabled:
-                self._replace_actions_queue(original_actions, processed_actions, real_delay)
+                self._replace_actions_queue(original_actions, processed_actions, delay)
                 return
 
             self._append_actions_queue(original_actions, processed_actions)
@@ -164,12 +183,13 @@ class ActionQueue:
             processed_actions: Post-processed actions for robot.
             real_delay: Number of time steps to skip due to inference delay.
         """
-        self.original_queue = original_actions[real_delay:].clone()
-        self.queue = processed_actions[real_delay:].clone()
+        clamped_delay = max(0, min(real_delay, len(original_actions), len(processed_actions)))
+        self.original_queue = original_actions[clamped_delay:].clone()
+        self.queue = processed_actions[clamped_delay:].clone()
 
         logger.debug(f"original_actions shape: {self.original_queue.shape}")
         logger.debug(f"processed_actions shape: {self.queue.shape}")
-        logger.debug(f"real_delay: {real_delay}")
+        logger.debug(f"real_delay: {real_delay}, clamped_delay: {clamped_delay}")
 
         self.last_index = 0
 
@@ -196,7 +216,9 @@ class ActionQueue:
 
         self.last_index = 0
 
-    def _check_delays(self, real_delay: int, action_index_before_inference: int | None = None):
+    def _check_and_resolve_delays(
+        self, real_delay: int, action_index_before_inference: int | None = None
+    ) -> int:
         """Validate that computed delays match expectations.
 
         Compares the delay computed from inference latency with the actual
@@ -205,15 +227,20 @@ class ActionQueue:
         Args:
             real_delay: Delay computed from inference latency.
             action_index_before_inference: Action index when inference started.
-        """
-        if action_index_before_inference is None:
-            return
 
-        indexes_diff = self.last_index - action_index_before_inference
-        if indexes_diff != real_delay:
-            # Let's check that action index difference (real delay calculated based on action queue)
-            # is the same as delay calculated based on inference latency
-            logger.warning(
-                f"[ACTION_QUEUE] Indexes diff is not equal to real delay. "
-                f"Indexes diff: {indexes_diff}, real delay: {real_delay}"
-            )
+        Returns:
+            int: Delay to use.
+        """
+        effective_delay = max(0, real_delay)
+
+        if action_index_before_inference is not None:
+            indexes_diff = max(0, self.last_index - action_index_before_inference)
+            if indexes_diff != real_delay:
+                logger.warning(
+                    "Indexes diff is not equal to real delay. indexes_diff=%d, real_delay=%d",
+                    indexes_diff,
+                    real_delay,
+                )
+                return real_delay
+
+        return effective_delay

src/lerobot/policies/sarm/README.md

@@ -1,14 +0,0 @@
-## Paper
-
-https://arxiv.org/abs/2509.25358
-
-## Citation
-
-```bibtex
-@article{chen2025sarm,
-  title={SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation},
-  author={Chen, Qianzhong and Yu, Justin and Schwager, Mac and Abbeel, Pieter and Shentu, Yide and Wu, Philipp},
-  journal={arXiv preprint arXiv:2509.25358},
-  year={2025}
-}
-```

src/lerobot/policies/sarm/README.md

@@ -0,0 +1 @@
+../../../../docs/source/policy_sarm_README.md

src/lerobot/processor/tokenizer_processor.py

@@ -136,8 +136,8 @@ class TokenizerProcessorStep(ObservationProcessorStep):
         # Standardize to a list of strings for the tokenizer
         if isinstance(task, str):
             return [task]
-        elif isinstance(task, list) and all(isinstance(t, str) for t in task):
-            return task
+        elif isinstance(task, (list, tuple)) and all(isinstance(t, str) for t in task):
+            return list(task)
 
         return None
 

src/lerobot/robots/bi_openarm_follower/bi_openarm_follower.py

@@ -96,9 +96,11 @@ class BiOpenArmFollower(Robot):
         left_arm_motors_ft = self.left_arm._motors_ft
         right_arm_motors_ft = self.right_arm._motors_ft
 
+        # Right first, then left — matches the teleoperator (OpenArmMini) ordering
+        # and the dataset feature names recorded during data collection.
         return {
-            **{f"left_{k}": v for k, v in left_arm_motors_ft.items()},
             **{f"right_{k}": v for k, v in right_arm_motors_ft.items()},
+            **{f"left_{k}": v for k, v in left_arm_motors_ft.items()},
         }
 
     @property
@@ -150,14 +152,16 @@ class BiOpenArmFollower(Robot):
         left_cam_keys = set(self.left_arm.cameras.keys())
         right_cam_keys = set(self.right_arm.cameras.keys())
 
-        left_obs = self.left_arm.get_observation()
-        for key, value in left_obs.items():
-            obs_dict[key if key in left_cam_keys else f"left_{key}"] = value
-
+        # Right first, then left — matches the teleoperator (OpenArmMini) ordering
+        # and the dataset feature names recorded during data collection.
         right_obs = self.right_arm.get_observation()
         for key, value in right_obs.items():
             obs_dict[key if key in right_cam_keys else f"right_{key}"] = value
 
+        left_obs = self.left_arm.get_observation()
+        for key, value in left_obs.items():
+            obs_dict[key if key in left_cam_keys else f"left_{key}"] = value
+
         return obs_dict
 
     @check_if_not_connected
@@ -183,7 +187,7 @@ class BiOpenArmFollower(Robot):
         prefixed_sent_action_left = {f"left_{key}": value for key, value in sent_action_left.items()}
         prefixed_sent_action_right = {f"right_{key}": value for key, value in sent_action_right.items()}
 
-        return {**prefixed_sent_action_left, **prefixed_sent_action_right}
+        return {**prefixed_sent_action_right, **prefixed_sent_action_left}
 
     @check_if_not_connected
     def disconnect(self):

src/lerobot/robots/bi_openarm_follower/config_bi_openarm_follower.py

@@ -23,10 +23,12 @@ from ..config import RobotConfig
 
 
 @RobotConfig.register_subclass("bi_openarm_follower")
-@dataclass
+@dataclass(kw_only=True)
 class BiOpenArmFollowerConfig(RobotConfig):
     """Configuration class for Bi OpenArm Follower robots."""
 
+    id: str | None = "bi_openarm_follower"
+
     left_arm_config: OpenArmFollowerConfigBase
     right_arm_config: OpenArmFollowerConfigBase
 

src/lerobot/scripts/lerobot_eval.py

@@ -73,7 +73,6 @@ from lerobot.configs import parser
 from lerobot.configs.eval import EvalPipelineConfig
 from lerobot.envs.factory import make_env, make_env_pre_post_processors
 from lerobot.envs.utils import (
-    add_envs_task,
     check_env_attributes_and_types,
     close_envs,
     preprocess_observation,
@@ -166,9 +165,15 @@ def rollout(
         if return_observations:
             all_observations.append(deepcopy(observation))
 
-        # Infer "task" from attributes of environments.
-        # TODO: works with SyncVectorEnv but not AsyncVectorEnv
-        observation = add_envs_task(env, observation)
+        # Infer "task" from sub-environments (prefer natural language description).
+        # env.call() works with both SyncVectorEnv and AsyncVectorEnv.
+        try:
+            observation["task"] = list(env.call("task_description"))
+        except Exception:
+            try:
+                observation["task"] = list(env.call("task"))
+            except Exception:
+                observation["task"] = [""] * env.num_envs
 
         # Apply environment-specific preprocessing (e.g., LiberoProcessorStep for LIBERO)
         observation = env_preprocessor(observation)
@@ -201,6 +206,11 @@ def rollout(
                     "You're likely using an older version of gymnasium (< 1.0). Please upgrade."
                 )
             successes = final_info["is_success"].tolist()
+        elif "is_success" in info:
+            is_success = info["is_success"]
+            successes = (
+                is_success.tolist() if hasattr(is_success, "tolist") else [bool(is_success)] * env.num_envs
+            )
         else:
             successes = [False] * env.num_envs
 
@@ -313,8 +323,9 @@ def eval_policy(
         n_to_render_now = min(max_episodes_rendered - n_episodes_rendered, env.num_envs)
         if isinstance(env, gym.vector.SyncVectorEnv):
             ep_frames.append(np.stack([env.envs[i].render() for i in range(n_to_render_now)]))  # noqa: B023
-        elif isinstance(env, gym.vector.AsyncVectorEnv):
+        elif hasattr(env, "call"):
             # Here we must render all frames and discard any we don't need.
+            # Covers AsyncVectorEnv and _LazyAsyncVectorEnv (which wraps one).
             ep_frames.append(np.stack(env.call("render")[:n_to_render_now]))
 
     if max_episodes_rendered > 0:
@@ -516,7 +527,7 @@ def eval_main(cfg: EvalPipelineConfig):
 
     logging.info(colored("Output dir:", "yellow", attrs=["bold"]) + f" {cfg.output_dir}")
 
-    logging.info("Making environment.")
+    logging.info(f"Making environment (batch_size={cfg.eval.batch_size}, async={cfg.eval.use_async_envs}).")
     envs = make_env(
         cfg.env,
         n_envs=cfg.eval.batch_size,
@@ -750,23 +761,39 @@ def eval_policy_all(
     )
 
     if max_parallel_tasks <= 1:
-        # sequential path (single accumulator path on the main thread)
-        # NOTE: keeping a single-threaded accumulator avoids concurrent list appends or locks
-        for task_group, task_id, env in tasks:
-            tg, tid, metrics = task_runner(task_group, task_id, env)
-            _accumulate_to(tg, metrics)
-            per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
+        prefetch_thread: threading.Thread | None = None
+        for i, (task_group, task_id, env) in enumerate(tasks):
+            if prefetch_thread is not None:
+                prefetch_thread.join()
+                prefetch_thread = None
+
+            try:
+                tg, tid, metrics = task_runner(task_group, task_id, env)
+                _accumulate_to(tg, metrics)
+                per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
+            finally:
+                env.close()
+                # Prefetch next task's workers *after* closing current env to prevent
+                # GPU memory overlap between consecutive tasks.
+                if i + 1 < len(tasks):
+                    next_env = tasks[i + 1][2]
+                    if hasattr(next_env, "_ensure"):
+                        prefetch_thread = threading.Thread(target=next_env._ensure, daemon=True)
+                        prefetch_thread.start()
     else:
-        # threaded path: submit all tasks, consume completions on main thread and accumulate there
         with cf.ThreadPoolExecutor(max_workers=max_parallel_tasks) as executor:
             fut2meta = {}
             for task_group, task_id, env in tasks:
                 fut = executor.submit(task_runner, task_group, task_id, env)
-                fut2meta[fut] = (task_group, task_id)
+                fut2meta[fut] = (task_group, task_id, env)
             for fut in cf.as_completed(fut2meta):
-                tg, tid, metrics = fut.result()
-                _accumulate_to(tg, metrics)
-                per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
+                tg, tid, env = fut2meta[fut]
+                try:
+                    tg, tid, metrics = fut.result()
+                    _accumulate_to(tg, metrics)
+                    per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
+                finally:
+                    env.close()
 
     # compute aggregated metrics helper (robust to lists/scalars)
     def _agg_from_list(xs):

src/lerobot/scripts/lerobot_record.py

@@ -74,6 +74,8 @@ from pathlib import Path
 from pprint import pformat
 from typing import Any
 
+import torch
+
 from lerobot.cameras import (  # noqa: F401
     CameraConfig,  # noqa: F401
 )
@@ -90,6 +92,7 @@ from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_featur
 from lerobot.datasets.video_utils import VideoEncodingManager
 from lerobot.policies.factory import make_policy, make_pre_post_processors
 from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rtc import ActionInterpolator
 from lerobot.policies.utils import make_robot_action
 from lerobot.processor import (
     PolicyAction,
@@ -226,6 +229,9 @@ class RecordConfig:
     play_sounds: bool = True
     # Resume recording on an existing dataset.
     resume: bool = False
+    # Action interpolation multiplier for smoother policy control (1=off, 2=2x, 3=3x)
+    # Only applies when using a policy (not teleop)
+    interpolation_multiplier: int = 1
 
     def __post_init__(self):
         # HACK: We parse again the cli args here to get the pretrained path if there was one.
@@ -298,6 +304,7 @@ def record_loop(
     control_time_s: int | None = None,
     single_task: str | None = None,
     display_data: bool = False,
+    interpolator: ActionInterpolator | None = None,
     display_compressed_images: bool = False,
 ):
     if dataset is not None and dataset.fps != fps:
@@ -334,6 +341,16 @@ def record_loop(
         preprocessor.reset()
         postprocessor.reset()
 
+    # Reset interpolator if provided
+    if interpolator is not None:
+        interpolator.reset()
+
+    # Calculate control interval based on interpolation
+    use_interpolation = interpolator is not None and interpolator.enabled and policy is not None
+    control_interval = interpolator.get_control_interval(fps) if interpolator else 1 / fps
+    # Pre-compute action key order outside the hot loop — it won't change mid-episode.
+    action_keys = sorted(robot.action_features) if use_interpolation else []
+
     no_action_count = 0
     timestamp = 0
     start_episode_t = time.perf_counter()
@@ -353,28 +370,68 @@ def record_loop(
         if policy is not None or dataset is not None:
             observation_frame = build_dataset_frame(dataset.features, obs_processed, prefix=OBS_STR)
 
+        # Track whether this iteration should be recorded to the dataset.
+        # Interpolated-only iterations send actions to the robot but don't record frames,
+        # keeping the dataset at the original fps while the robot moves at the higher rate.
+        is_record_frame = True
+
         # Get action from either policy or teleop
         if policy is not None and preprocessor is not None and postprocessor is not None:
-            action_values = predict_action(
-                observation=observation_frame,
-                policy=policy,
-                device=get_safe_torch_device(policy.config.device),
-                preprocessor=preprocessor,
-                postprocessor=postprocessor,
-                use_amp=policy.config.use_amp,
-                task=single_task,
-                robot_type=robot.robot_type,
-            )
+            # With interpolation: only call policy when interpolator needs new action
+            if use_interpolation:
+                ran_inference = False
 
-            act_processed_policy: RobotAction = make_robot_action(action_values, dataset.features)
+                if interpolator.needs_new_action():
+                    action_values = predict_action(
+                        observation=observation_frame,
+                        policy=policy,
+                        device=get_safe_torch_device(policy.config.device),
+                        preprocessor=preprocessor,
+                        postprocessor=postprocessor,
+                        use_amp=policy.config.use_amp,
+                        task=single_task,
+                        robot_type=robot.robot_type,
+                    )
+                    act_processed_policy = make_robot_action(action_values, dataset.features)
+                    robot_action_to_send = robot_action_processor((act_processed_policy, obs))
+
+                    action_tensor = torch.tensor([robot_action_to_send[k] for k in action_keys])
+                    interpolator.add(action_tensor)
+                    ran_inference = True
+
+                interp_action = interpolator.get()
+                if interp_action is not None:
+                    robot_action_to_send = {k: interp_action[i].item() for i, k in enumerate(action_keys)}
+                    action_values = robot_action_to_send
+                else:
+                    continue
+
+                is_record_frame = ran_inference
+            else:
+                action_values = predict_action(
+                    observation=observation_frame,
+                    policy=policy,
+                    device=get_safe_torch_device(policy.config.device),
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    use_amp=policy.config.use_amp,
+                    task=single_task,
+                    robot_type=robot.robot_type,
+                )
+                act_processed_policy: RobotAction = make_robot_action(action_values, dataset.features)
+                # Applies a pipeline to the action, default is IdentityProcessor
+                robot_action_to_send = robot_action_processor((act_processed_policy, obs))
+                action_values = robot_action_to_send
 
         elif policy is None and isinstance(teleop, Teleoperator):
+            act = teleop.get_action()
             if robot.name == "unitree_g1":
                 teleop.send_feedback(obs)
-            act = teleop.get_action()
 
             # Applies a pipeline to the raw teleop action, default is IdentityProcessor
             act_processed_teleop = teleop_action_processor((act, obs))
+            action_values = act_processed_teleop
+            robot_action_to_send = robot_action_processor((act_processed_teleop, obs))
 
         elif policy is None and isinstance(teleop, list):
             arm_action = teleop_arm.get_action()
@@ -383,6 +440,8 @@ def record_loop(
             base_action = robot._from_keyboard_to_base_action(keyboard_action)
             act = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
             act_processed_teleop = teleop_action_processor((act, obs))
+            action_values = act_processed_teleop
+            robot_action_to_send = robot_action_processor((act_processed_teleop, obs))
         else:
             no_action_count += 1
             if no_action_count == 1 or no_action_count % 10 == 0:
@@ -393,22 +452,14 @@ def record_loop(
                 )
             continue
 
-        # Applies a pipeline to the action, default is IdentityProcessor
-        if policy is not None and act_processed_policy is not None:
-            action_values = act_processed_policy
-            robot_action_to_send = robot_action_processor((act_processed_policy, obs))
-        else:
-            action_values = act_processed_teleop
-            robot_action_to_send = robot_action_processor((act_processed_teleop, obs))
-
         # Send action to robot
         # Action can eventually be clipped using `max_relative_target`,
         # so action actually sent is saved in the dataset. action = postprocessor.process(action)
         # TODO(steven, pepijn, adil): we should use a pipeline step to clip the action, so the sent action is the action that we input to the robot.
         _sent_action = robot.send_action(robot_action_to_send)
 
-        # Write to dataset
-        if dataset is not None:
+        # Write to dataset (only on real policy frames, not interpolated-only iterations)
+        if dataset is not None and is_record_frame:
             action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
             frame = {**observation_frame, **action_frame, "task": single_task}
             dataset.add_frame(frame)
@@ -420,7 +471,7 @@ def record_loop(
 
         dt_s = time.perf_counter() - start_loop_t
 
-        sleep_time_s: float = 1 / fps - dt_s
+        sleep_time_s: float = control_interval - dt_s
         if sleep_time_s < 0:
             logging.warning(
                 f"Record loop is running slower ({1 / dt_s:.1f} Hz) than the target FPS ({fps} Hz). Dataset frames might be dropped and robot control might be unstable. Common causes are: 1) Camera FPS not keeping up 2) Policy inference taking too long 3) CPU starvation"
@@ -506,6 +557,7 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
         policy = None if cfg.policy is None else make_policy(cfg.policy, ds_meta=dataset.meta)
         preprocessor = None
         postprocessor = None
+        interpolator = None
         if cfg.policy is not None:
             preprocessor, postprocessor = make_pre_post_processors(
                 policy_cfg=cfg.policy,
@@ -516,6 +568,10 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                     "rename_observations_processor": {"rename_map": cfg.dataset.rename_map},
                 },
             )
+            # Create interpolator for smoother policy control
+            if cfg.interpolation_multiplier > 1:
+                interpolator = ActionInterpolator(multiplier=cfg.interpolation_multiplier)
+                logging.info(f"Action interpolation enabled: {cfg.interpolation_multiplier}x control rate")
 
         robot.connect()
         if teleop is not None:
@@ -547,6 +603,7 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                     control_time_s=cfg.dataset.episode_time_s,
                     single_task=cfg.dataset.single_task,
                     display_data=cfg.display_data,
+                    interpolator=interpolator,
                     display_compressed_images=display_compressed_images,
                 )
 

src/lerobot/teleoperators/openarm_mini/openarm_mini.py

@@ -32,9 +32,15 @@ from .config_openarm_mini import OpenArmMiniConfig
 logger = logging.getLogger(__name__)
 
 # Motors whose direction is inverted during readout
-RIGHT_MOTORS_TO_FLIP = ["joint_1", "joint_2", "joint_3", "joint_4", "joint_5"]
+RIGHT_MOTORS_TO_FLIP = ["joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_7"]
 LEFT_MOTORS_TO_FLIP = ["joint_1", "joint_3", "joint_4", "joint_5", "joint_6", "joint_7"]
 
+# Leader joint 6 maps to follower joint 7 and vice versa
+JOINT_REMAP = {"joint_6": "joint_7", "joint_7": "joint_6"}
+JOINT_REMAP_REVERSE = {"joint_7": "joint_6", "joint_6": "joint_7"}
+
+GRIPPER_TELEOP_TO_DEGREES = -0.65
+
 
 class OpenArmMini(Teleoperator):
     """
@@ -95,6 +101,8 @@ class OpenArmMini(Teleoperator):
 
     @property
     def action_features(self) -> dict[str, type]:
+        # Right first, then left — matches the robot (BiOpenArmFollower) ordering
+        # and the dataset feature names recorded during data collection.
         features: dict[str, type] = {}
         for motor in self.bus_right.motors:
             features[f"right_{motor}.pos"] = float
@@ -276,16 +284,70 @@ class OpenArmMini(Teleoperator):
         right_positions = self.bus_right.sync_read("Present_Position")
         left_positions = self.bus_left.sync_read("Present_Position")
 
+        # Right first, then left — matches the robot (BiOpenArmFollower) ordering
+        # and the dataset feature names recorded during data collection.
+        # Joint 6↔7 remap: leader joint_6 → follower joint_7 and vice versa.
         action: dict[str, Any] = {}
         for motor, val in right_positions.items():
-            action[f"right_{motor}.pos"] = -val if motor in RIGHT_MOTORS_TO_FLIP else val
+            target = JOINT_REMAP.get(motor, motor)
+            if motor == "gripper":
+                # Convert gripper from teleop 0-100 to openarms degrees: 0→0°, 100→-65°
+                action[f"right_{target}.pos"] = val * GRIPPER_TELEOP_TO_DEGREES
+            else:
+                action[f"right_{target}.pos"] = -val if motor in RIGHT_MOTORS_TO_FLIP else val
         for motor, val in left_positions.items():
-            action[f"left_{motor}.pos"] = -val if motor in LEFT_MOTORS_TO_FLIP else val
+            target = JOINT_REMAP.get(motor, motor)
+            if motor == "gripper":
+                action[f"left_{target}.pos"] = val * GRIPPER_TELEOP_TO_DEGREES
+            else:
+                action[f"left_{target}.pos"] = -val if motor in LEFT_MOTORS_TO_FLIP else val
 
         dt_ms = (time.perf_counter() - start) * 1e3
         logger.debug(f"{self} read action: {dt_ms:.1f}ms")
         return action
 
+    def enable_torque(self) -> None:
+        """Enable torque on both arms for position control."""
+        self.bus_right.enable_torque()
+        self.bus_left.enable_torque()
+
+    def disable_torque(self) -> None:
+        """Disable torque on both arms for free movement."""
+        self.bus_right.disable_torque()
+        self.bus_left.disable_torque()
+
+    def write_goal_positions(self, positions: dict[str, float]) -> None:
+        """Write goal positions to motors (inverse of get_action flip/gripper/remap logic)."""
+        right_goals: dict[str, float] = {}
+        left_goals: dict[str, float] = {}
+
+        for key, val in positions.items():
+            if not key.endswith(".pos"):
+                continue
+            motor_name = key.removesuffix(".pos")
+            if motor_name.startswith("right_"):
+                base = motor_name.removeprefix("right_")
+                # Reverse remap: follower joint_7 → leader joint_6 and vice versa
+                target = JOINT_REMAP_REVERSE.get(base, base)
+                if base == "gripper":
+                    # Convert robot degrees to teleop 0-100: 0°→0, -65°→100
+                    right_goals[target] = val / GRIPPER_TELEOP_TO_DEGREES
+                else:
+                    # Un-flip using the ORIGINAL motor name (target = leader motor)
+                    right_goals[target] = -val if target in RIGHT_MOTORS_TO_FLIP else val
+            elif motor_name.startswith("left_"):
+                base = motor_name.removeprefix("left_")
+                target = JOINT_REMAP_REVERSE.get(base, base)
+                if base == "gripper":
+                    left_goals[target] = val / GRIPPER_TELEOP_TO_DEGREES
+                else:
+                    left_goals[target] = -val if target in LEFT_MOTORS_TO_FLIP else val
+
+        if right_goals:
+            self.bus_right.sync_write("Goal_Position", right_goals)
+        if left_goals:
+            self.bus_left.sync_write("Goal_Position", left_goals)
+
     def send_feedback(self, feedback: dict[str, float]) -> None:
         raise NotImplementedError("Feedback is not yet implemented for OpenArm Mini.")
 

tests/datasets/test_datasets.py

@@ -24,6 +24,7 @@ import torch
 from huggingface_hub import HfApi
 from PIL import Image
 from safetensors.torch import load_file
+from torchvision.transforms import v2
 
 import lerobot
 from lerobot.configs.default import DatasetConfig
@@ -34,6 +35,7 @@ from lerobot.datasets.image_writer import image_array_to_pil_image
 from lerobot.datasets.io_utils import hf_transform_to_torch
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.multi_dataset import MultiLeRobotDataset
+from lerobot.datasets.transforms import ImageTransforms, ImageTransformsConfig
 from lerobot.datasets.utils import (
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
@@ -355,6 +357,62 @@ def test_add_frame_image_pil(image_dataset):
     assert dataset[0]["image"].shape == torch.Size(DUMMY_CHW)
 
 
+def test_set_image_transforms_applies_transparently(image_dataset):
+    dataset = image_dataset
+    dataset.add_frame({"image": np.random.rand(*DUMMY_CHW), "task": "Dummy task"})
+    dataset.save_episode()
+    dataset.finalize()
+
+    dataset.set_image_transforms(v2.Resize((224, 224)))
+    assert dataset[0]["image"].shape == torch.Size((3, 224, 224))
+
+    dataset.set_image_transforms(v2.Resize((128, 128)))
+    assert dataset[0]["image"].shape == torch.Size((3, 128, 128))
+
+    dataset.clear_image_transforms()
+    assert dataset[0]["image"].shape == torch.Size(DUMMY_CHW)
+
+
+def test_set_image_transforms_supports_lerobot_image_transforms(image_dataset):
+    dataset = image_dataset
+    dataset.add_frame({"image": np.random.rand(*DUMMY_CHW), "task": "Dummy task"})
+    dataset.save_episode()
+    dataset.finalize()
+
+    image_transforms = ImageTransforms(ImageTransformsConfig(enable=False))
+    dataset.set_image_transforms(image_transforms)
+
+    assert dataset.image_transforms is image_transforms
+    assert dataset[0]["image"].shape == torch.Size(DUMMY_CHW)
+
+
+def test_set_image_transforms_supports_loaded_dataset(tmp_path, lerobot_dataset_factory):
+    dataset = lerobot_dataset_factory(root=tmp_path / "test", use_videos=False)
+    dataset.set_image_transforms(v2.Compose([v2.Resize((224, 224)), v2.Resize((112, 112))]))
+
+    camera_key = dataset.meta.camera_keys[0]
+    assert dataset[0][camera_key].shape == torch.Size((3, 112, 112))
+
+
+def test_multilerobot_dataset_set_image_transforms_propagates(tmp_path, lerobot_dataset_factory):
+    root = tmp_path / "multi"
+    repo_ids = ["lerobot/test_multi_a", "lerobot/test_multi_b"]
+
+    for repo_id in repo_ids:
+        lerobot_dataset_factory(root=root / repo_id, repo_id=repo_id, use_videos=False)
+
+    dataset = MultiLeRobotDataset(repo_ids, root=root, download_videos=False)
+    dataset.set_image_transforms(v2.Resize((96, 96)))
+
+    camera_key = dataset.camera_keys[0]
+    assert dataset[0][camera_key].shape == torch.Size((3, 96, 96))
+    assert all(child.image_transforms is dataset.image_transforms for child in dataset._datasets)
+
+    dataset.clear_image_transforms()
+    assert dataset.image_transforms is None
+    assert all(child.image_transforms is None for child in dataset._datasets)
+
+
 def test_image_array_to_pil_image_wrong_range_float_0_255():
     image = np.random.rand(*DUMMY_HWC) * 255
     with pytest.raises(ValueError):

tests/envs/test_dispatch.py

@@ -0,0 +1,143 @@
+"""Tests for the benchmark dispatch refactor (create_envs / get_env_processors on EnvConfig)."""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+
+import gymnasium as gym
+import pytest
+from gymnasium.envs.registration import register, registry as gym_registry
+
+from lerobot.configs.types import PolicyFeature
+from lerobot.envs.configs import EnvConfig
+from lerobot.envs.factory import make_env, make_env_config, make_env_pre_post_processors
+
+logger = logging.getLogger(__name__)
+
+
+def test_registry_all_types():
+    """make_env_config should resolve every registered EnvConfig subclass via the registry."""
+    known = list(EnvConfig.get_known_choices().keys())
+    assert len(known) >= 6
+    for t in known:
+        cfg = make_env_config(t)
+        if not isinstance(cfg, EnvConfig):
+            continue
+        assert cfg.type == t
+
+
+def test_unknown_type():
+    with pytest.raises(ValueError, match="not registered"):
+        make_env_config("nonexistent")
+
+
+def test_identity_processors():
+    """Base class get_env_processors() returns identity pipelines."""
+    cfg = make_env_config("aloha")
+    pre, post = cfg.get_env_processors()
+    assert len(pre.steps) == 0 and len(post.steps) == 0
+
+
+def test_delegation():
+    """make_env() should call cfg.create_envs(), not use if/elif dispatch."""
+    sentinel = {"delegated": {0: "marker"}}
+    fake = type(
+        "Fake",
+        (),
+        {
+            "hub_path": None,
+            "create_envs": lambda self, n_envs, use_async_envs=False: sentinel,
+        },
+    )()
+    result = make_env(fake, n_envs=1)
+    assert result is sentinel
+
+
+def test_processors_delegation():
+    """make_env_pre_post_processors delegates to cfg.get_env_processors()."""
+    cfg = make_env_config("aloha")
+    pre, post = make_env_pre_post_processors(cfg, policy_cfg=None)
+    assert len(pre.steps) == 0
+
+
+def test_base_create_envs():
+    """Base class create_envs() should build a single-task VectorEnv via gym.make()."""
+    gym_id = "_dispatch_test/CartPole-v99"
+    if gym_id not in gym_registry:
+        register(id=gym_id, entry_point="gymnasium.envs.classic_control:CartPoleEnv")
+
+    @EnvConfig.register_subclass("_dispatch_base_test")
+    @dataclass
+    class _Env(EnvConfig):
+        task: str = "CartPole-v99"
+        fps: int = 10
+        features: dict[str, PolicyFeature] = field(default_factory=dict)
+
+        @property
+        def package_name(self):
+            return "_dispatch_test"
+
+        @property
+        def gym_id(self):
+            return gym_id
+
+        @property
+        def gym_kwargs(self):
+            return {}
+
+    try:
+        envs = _Env().create_envs(n_envs=2)
+        assert "_dispatch_base_test" in envs
+        env = envs["_dispatch_base_test"][0]
+        assert isinstance(env, gym.vector.VectorEnv)
+        assert env.num_envs == 2
+        env.close()
+    finally:
+        if gym_id in gym_registry:
+            del gym_registry[gym_id]
+
+
+def test_custom_create_envs_override():
+    """A custom EnvConfig subclass can override create_envs()."""
+    mock_vec = gym.vector.SyncVectorEnv([lambda: gym.make("CartPole-v1")])
+
+    @EnvConfig.register_subclass("_dispatch_custom_test")
+    @dataclass
+    class _Env(EnvConfig):
+        task: str = "x"
+        features: dict[str, PolicyFeature] = field(default_factory=dict)
+
+        @property
+        def gym_kwargs(self):
+            return {}
+
+        def create_envs(self, n_envs, use_async_envs=False):
+            return {"custom_suite": {0: mock_vec}}
+
+    try:
+        result = make_env(_Env(), n_envs=1)
+        assert "custom_suite" in result
+    finally:
+        mock_vec.close()
+
+
+def test_custom_get_env_processors_override():
+    """A custom EnvConfig subclass can override get_env_processors()."""
+    from lerobot.processor.pipeline import DataProcessorPipeline
+
+    @EnvConfig.register_subclass("_dispatch_proc_test")
+    @dataclass
+    class _Env(EnvConfig):
+        task: str = "x"
+        features: dict[str, PolicyFeature] = field(default_factory=dict)
+
+        @property
+        def gym_kwargs(self):
+            return {}
+
+        def get_env_processors(self):
+            return DataProcessorPipeline(steps=[]), DataProcessorPipeline(steps=[])
+
+    pre, post = _Env().get_env_processors()
+    assert isinstance(pre, DataProcessorPipeline)

tests/policies/rtc/test_action_interpolator.py

@@ -0,0 +1,559 @@
+# 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.
+
+"""Tests for ActionInterpolator and its interaction with ActionQueue (RTC)."""
+
+import pytest
+import torch
+
+from lerobot.policies.rtc.action_interpolator import ActionInterpolator
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+
+# ====================== Fixtures ======================
+
+
+@pytest.fixture
+def interp2():
+    """Create an ActionInterpolator with multiplier=2."""
+    return ActionInterpolator(multiplier=2)
+
+
+@pytest.fixture
+def interp3():
+    """Create an ActionInterpolator with multiplier=3."""
+    return ActionInterpolator(multiplier=3)
+
+
+# ====================== Initialization Tests ======================
+
+
+def test_interpolator_multiplier_1_no_interpolation():
+    """Test multiplier=1 creates a disabled interpolator."""
+    interp = ActionInterpolator(multiplier=1)
+    assert interp.multiplier == 1
+    assert not interp.enabled
+
+
+def test_interpolator_multiplier_2_enabled():
+    """Test multiplier=2 creates an enabled interpolator."""
+    interp = ActionInterpolator(multiplier=2)
+    assert interp.multiplier == 2
+    assert interp.enabled
+
+
+def test_interpolator_multiplier_0_raises():
+    """Test multiplier=0 raises ValueError."""
+    with pytest.raises(ValueError, match="multiplier must be >= 1"):
+        ActionInterpolator(multiplier=0)
+
+
+def test_interpolator_negative_multiplier_raises():
+    """Test negative multiplier raises ValueError."""
+    with pytest.raises(ValueError, match="multiplier must be >= 1"):
+        ActionInterpolator(multiplier=-1)
+
+
+def test_interpolator_default_multiplier_is_1():
+    """Test default multiplier is 1 (disabled)."""
+    interp = ActionInterpolator()
+    assert interp.multiplier == 1
+    assert not interp.enabled
+
+
+# ====================== needs_new_action Tests ======================
+
+
+def test_needs_new_action_true_initially(interp2):
+    """Test needs_new_action() returns True before any action is added."""
+    assert interp2.needs_new_action()
+
+
+def test_needs_new_action_false_after_add(interp2):
+    """Test needs_new_action() returns False right after add()."""
+    interp2.add(torch.tensor([1.0, 2.0]))
+    assert not interp2.needs_new_action()
+
+
+def test_needs_new_action_true_after_buffer_exhausted(interp2):
+    """Test needs_new_action() returns True after consuming all buffered actions."""
+    interp2.add(torch.tensor([1.0, 2.0]))
+    interp2.get()
+    assert interp2.needs_new_action()
+
+
+def test_needs_new_action_true_after_all_interpolated_consumed(interp2):
+    """Test needs_new_action() tracks interpolated sub-steps correctly."""
+    interp2.add(torch.tensor([0.0, 0.0]))
+    interp2.get()
+    assert interp2.needs_new_action()
+
+    interp2.add(torch.tensor([2.0, 4.0]))
+    interp2.get()
+    assert not interp2.needs_new_action()
+    interp2.get()
+    assert interp2.needs_new_action()
+
+
+# ====================== Passthrough Tests (multiplier=1) ======================
+
+
+def test_passthrough_single_action_returned_as_is():
+    """Test multiplier=1 returns the action unchanged."""
+    interp = ActionInterpolator(multiplier=1)
+    action = torch.tensor([3.0, 5.0])
+    interp.add(action)
+
+    result = interp.get()
+    assert result is not None
+    torch.testing.assert_close(result, action)
+
+
+def test_passthrough_none_after_single_get():
+    """Test multiplier=1 returns None after consuming the single action."""
+    interp = ActionInterpolator(multiplier=1)
+    interp.add(torch.tensor([1.0]))
+    interp.get()
+    assert interp.get() is None
+
+
+def test_passthrough_sequential_actions():
+    """Test multiplier=1 passes through consecutive actions one at a time."""
+    interp = ActionInterpolator(multiplier=1)
+    for val in [1.0, 2.0, 3.0]:
+        action = torch.tensor([val])
+        interp.add(action)
+        result = interp.get()
+        torch.testing.assert_close(result, action)
+        assert interp.get() is None
+
+
+# ====================== Interpolation Tests (multiplier=2) ======================
+
+
+def test_interpolation_2x_first_action_no_interpolation(interp2):
+    """Test first action has no previous, so buffer is just [action]."""
+    interp2.add(torch.tensor([0.0, 0.0]))
+    result = interp2.get()
+    torch.testing.assert_close(result, torch.tensor([0.0, 0.0]))
+    assert interp2.get() is None
+
+
+def test_interpolation_2x_second_action_produces_two_steps(interp2):
+    """Test second action produces 2 interpolated sub-steps."""
+    interp2.add(torch.tensor([0.0, 0.0]))
+    interp2.get()
+
+    interp2.add(torch.tensor([2.0, 4.0]))
+    step1 = interp2.get()
+    step2 = interp2.get()
+
+    torch.testing.assert_close(step1, torch.tensor([1.0, 2.0]))
+    torch.testing.assert_close(step2, torch.tensor([2.0, 4.0]))
+    assert interp2.get() is None
+
+
+def test_interpolation_2x_three_consecutive_actions(interp2):
+    """Test interpolation across three consecutive actions."""
+    a0 = torch.tensor([0.0])
+    a1 = torch.tensor([4.0])
+    a2 = torch.tensor([10.0])
+
+    interp2.add(a0)
+    torch.testing.assert_close(interp2.get(), a0)
+
+    interp2.add(a1)
+    torch.testing.assert_close(interp2.get(), torch.tensor([2.0]))
+    torch.testing.assert_close(interp2.get(), torch.tensor([4.0]))
+
+    interp2.add(a2)
+    torch.testing.assert_close(interp2.get(), torch.tensor([7.0]))
+    torch.testing.assert_close(interp2.get(), torch.tensor([10.0]))
+
+
+# ====================== Interpolation Tests (multiplier=3) ======================
+
+
+def test_interpolation_3x_produces_three_steps(interp3):
+    """Test multiplier=3 produces 3 interpolated sub-steps."""
+    interp3.add(torch.tensor([0.0, 0.0]))
+    interp3.get()
+
+    interp3.add(torch.tensor([3.0, 6.0]))
+    s1 = interp3.get()
+    s2 = interp3.get()
+    s3 = interp3.get()
+
+    torch.testing.assert_close(s1, torch.tensor([1.0, 2.0]))
+    torch.testing.assert_close(s2, torch.tensor([2.0, 4.0]))
+    torch.testing.assert_close(s3, torch.tensor([3.0, 6.0]))
+    assert interp3.get() is None
+
+
+def test_interpolation_3x_last_step_equals_target(interp3):
+    """Test last interpolated step equals the target action exactly."""
+    interp3.add(torch.tensor([10.0]))
+    interp3.get()
+
+    target = torch.tensor([100.0])
+    interp3.add(target)
+    interp3.get()
+    interp3.get()
+    last = interp3.get()
+    torch.testing.assert_close(last, target)
+
+
+# ====================== Reset Tests ======================
+
+
+def test_reset_clears_buffer(interp2):
+    """Test reset() clears the action buffer."""
+    interp2.add(torch.tensor([1.0]))
+    interp2.reset()
+    assert interp2.needs_new_action()
+    assert interp2.get() is None
+
+
+def test_reset_clears_prev(interp2):
+    """Test after reset, next add produces single-element buffer (no prev)."""
+    interp2.add(torch.tensor([0.0]))
+    interp2.get()
+    interp2.add(torch.tensor([10.0]))
+    interp2.get()
+    interp2.get()
+
+    interp2.reset()
+    interp2.add(torch.tensor([5.0]))
+    result = interp2.get()
+    torch.testing.assert_close(result, torch.tensor([5.0]))
+    assert interp2.get() is None
+
+
+def test_reset_episode_boundary(interp2):
+    """Test reset between two simulated episodes."""
+    interp2.add(torch.tensor([0.0]))
+    interp2.get()
+    interp2.add(torch.tensor([10.0]))
+    interp2.get()
+    interp2.get()
+
+    interp2.reset()
+
+    interp2.add(torch.tensor([100.0]))
+    result = interp2.get()
+    torch.testing.assert_close(result, torch.tensor([100.0]))
+    assert interp2.get() is None
+
+
+# ====================== get_control_interval Tests ======================
+
+
+def test_control_interval_30fps_multiplier_1():
+    """Test control interval at 30fps with no interpolation."""
+    interp = ActionInterpolator(multiplier=1)
+    assert interp.get_control_interval(30.0) == pytest.approx(1.0 / 30.0)
+
+
+def test_control_interval_30fps_multiplier_2(interp2):
+    """Test control interval at 30fps with 2x interpolation."""
+    assert interp2.get_control_interval(30.0) == pytest.approx(1.0 / 60.0)
+
+
+def test_control_interval_30fps_multiplier_3(interp3):
+    """Test control interval at 30fps with 3x interpolation."""
+    assert interp3.get_control_interval(30.0) == pytest.approx(1.0 / 90.0)
+
+
+def test_control_interval_60fps_multiplier_2(interp2):
+    """Test control interval at 60fps with 2x interpolation."""
+    assert interp2.get_control_interval(60.0) == pytest.approx(1.0 / 120.0)
+
+
+# ====================== get() on Empty Tests ======================
+
+
+def test_get_returns_none_before_any_add():
+    """Test get() returns None when no action has been added."""
+    interp = ActionInterpolator(multiplier=2)
+    assert interp.get() is None
+
+
+def test_get_returns_none_after_reset(interp2):
+    """Test get() returns None after reset."""
+    interp2.add(torch.tensor([1.0]))
+    interp2.reset()
+    assert interp2.get() is None
+
+
+# ====================== Multi-Dimensional Action Tests ======================
+
+
+def test_6dof_interpolation(interp2):
+    """Test interpolation works correctly with 6-dimensional actions."""
+    prev = torch.zeros(6)
+    target = torch.tensor([1.0, 2.0, 3.0, 4.0, 5.0, 6.0])
+
+    interp2.add(prev)
+    interp2.get()
+
+    interp2.add(target)
+    mid = interp2.get()
+    end = interp2.get()
+
+    torch.testing.assert_close(mid, target / 2)
+    torch.testing.assert_close(end, target)
+
+
+# ====================== Simulated Control Loop Tests ======================
+
+
+def test_control_loop_produces_correct_action_count():
+    """Test N policy actions with multiplier M yields 1 + (N-1)*M robot commands."""
+    multiplier = 3
+    n_policy_actions = 5
+    interp = ActionInterpolator(multiplier=multiplier)
+
+    robot_commands = 0
+    for i in range(n_policy_actions):
+        action = torch.tensor([float(i)])
+        if interp.needs_new_action():
+            interp.add(action)
+        while True:
+            a = interp.get()
+            if a is None:
+                break
+            robot_commands += 1
+
+    expected = 1 + (n_policy_actions - 1) * multiplier
+    assert robot_commands == expected
+
+
+def test_control_loop_monotonic_increase():
+    """Test actions [0, 1, 2, 3] with multiplier=2 produce monotonically increasing values."""
+    interp = ActionInterpolator(multiplier=2)
+    all_values = []
+
+    for i in range(4):
+        interp.add(torch.tensor([float(i)]))
+        while True:
+            a = interp.get()
+            if a is None:
+                break
+            all_values.append(a.item())
+
+    for i in range(1, len(all_values)):
+        assert all_values[i] >= all_values[i - 1]
+
+
+# ====================== ActionQueue + ActionInterpolator Integration Tests ======================
+
+
+def _make_chunk(n_steps: int, action_dim: int = 2, offset: float = 0.0) -> torch.Tensor:
+    """Create a simple action chunk: each row is [offset + step_idx, offset + step_idx]."""
+    return torch.arange(n_steps, dtype=torch.float32).unsqueeze(1).expand(-1, action_dim) + offset
+
+
+def test_queue_interpolator_consumption_rate_matches_base_fps():
+    """Test queue.get() is called at base fps rate, not multiplied fps."""
+    cfg = RTCConfig(enabled=True, execution_horizon=10)
+    queue = ActionQueue(cfg)
+    interp = ActionInterpolator(multiplier=3)
+
+    chunk = _make_chunk(10)
+    queue.merge(chunk, chunk.clone(), real_delay=0)
+
+    queue_gets = 0
+    control_ticks = 0
+
+    while True:
+        if interp.needs_new_action():
+            if queue.empty():
+                break
+            action = queue.get()
+            if action is None:
+                break
+            interp.add(action)
+            queue_gets += 1
+
+        result = interp.get()
+        if result is not None:
+            control_ticks += 1
+
+    assert queue_gets == 10
+    assert control_ticks == 1 + 9 * 3
+
+
+def test_queue_interpolator_leftover_decreases_only_on_queue_get():
+    """Test get_left_over() shrinks only on queue.get(), not on interpolator sub-steps."""
+    cfg = RTCConfig(enabled=True, execution_horizon=10)
+    queue = ActionQueue(cfg)
+    interp = ActionInterpolator(multiplier=3)
+
+    chunk = _make_chunk(6)
+    queue.merge(chunk, chunk.clone(), real_delay=0)
+
+    assert interp.needs_new_action()
+    interp.add(queue.get())
+    leftover_after_first_get = queue.get_left_over()
+    assert leftover_after_first_get is not None
+    assert len(leftover_after_first_get) == 5
+
+    interp.get()
+    assert len(queue.get_left_over()) == 5
+
+    interp.add(queue.get())
+    assert len(queue.get_left_over()) == 4
+
+    for _ in range(3):
+        assert interp.get() is not None
+    assert len(queue.get_left_over()) == 4
+
+
+def test_queue_interpolator_processed_leftover_tracks_queue_index():
+    """Test get_processed_left_over() reflects queue's last_index, not interpolator state."""
+    cfg = RTCConfig(enabled=True, execution_horizon=10)
+    queue = ActionQueue(cfg)
+    interp = ActionInterpolator(multiplier=2)
+
+    original = _make_chunk(8, offset=0.0)
+    processed = _make_chunk(8, offset=100.0)
+    queue.merge(original, processed, real_delay=0)
+
+    left = queue.get_processed_left_over()
+    assert len(left) == 8
+
+    for _ in range(3):
+        if interp.needs_new_action():
+            action = queue.get()
+            if action is not None:
+                interp.add(action)
+        interp.get()
+
+    proc_left = queue.get_processed_left_over()
+    orig_left = queue.get_left_over()
+    assert proc_left is not None and orig_left is not None
+    assert len(proc_left) == len(orig_left)
+    assert proc_left[0, 0].item() >= 100.0
+    assert orig_left[0, 0].item() < 100.0
+
+
+def test_queue_interpolator_merge_resets_queue_but_interpolator_keeps_prev():
+    """Test queue merge doesn't affect interpolator's prev, enabling smooth transitions."""
+    cfg = RTCConfig(enabled=True, execution_horizon=10)
+    queue = ActionQueue(cfg)
+    interp = ActionInterpolator(multiplier=2)
+
+    chunk1 = torch.tensor([[0.0], [2.0], [4.0], [6.0], [8.0]])
+    queue.merge(chunk1, chunk1.clone(), real_delay=0)
+
+    consumed = []
+    for _ in range(5):
+        if interp.needs_new_action():
+            a = queue.get()
+            if a is not None:
+                interp.add(a)
+        r = interp.get()
+        if r is not None:
+            consumed.append(r.item())
+
+    assert interp.needs_new_action()
+    assert consumed[-1] == pytest.approx(4.0)
+
+    idx_before = queue.get_action_index()
+
+    chunk2 = torch.tensor([[10.0], [12.0], [14.0]])
+    queue.merge(chunk2, chunk2.clone(), real_delay=0, action_index_before_inference=idx_before)
+
+    first_action = queue.get()
+    assert first_action is not None
+    interp.add(first_action)
+    first_from_new = interp.get()
+    assert first_from_new is not None
+    assert first_from_new.item() == pytest.approx(7.0)
+
+
+def test_queue_interpolator_reset_does_not_affect_queue():
+    """Test interpolator reset leaves queue state untouched."""
+    cfg = RTCConfig(enabled=True, execution_horizon=10)
+    queue = ActionQueue(cfg)
+    interp = ActionInterpolator(multiplier=2)
+
+    chunk = _make_chunk(5)
+    queue.merge(chunk, chunk.clone(), real_delay=0)
+
+    interp.add(queue.get())
+    interp.get()
+    interp.add(queue.get())
+    interp.get()
+    interp.get()
+
+    assert queue.qsize() == 3
+
+    interp.reset()
+
+    assert queue.qsize() == 3
+    assert len(queue.get_left_over()) == 3
+
+    interp.add(queue.get())
+    result = interp.get()
+    assert result is not None
+    assert queue.qsize() == 2
+
+
+def test_queue_interpolator_no_interpolation_1_to_1():
+    """Test multiplier=1 produces exactly 1 robot command per queue.get()."""
+    cfg = RTCConfig(enabled=True, execution_horizon=10)
+    queue = ActionQueue(cfg)
+    interp = ActionInterpolator(multiplier=1)
+
+    chunk = _make_chunk(5)
+    queue.merge(chunk, chunk.clone(), real_delay=0)
+
+    robot_commands = 0
+    while not queue.empty():
+        if interp.needs_new_action():
+            action = queue.get()
+            if action is not None:
+                interp.add(action)
+        result = interp.get()
+        if result is not None:
+            robot_commands += 1
+
+    assert robot_commands == 5
+
+
+def test_queue_interpolator_delay_skips_stale_actions():
+    """Test merge with delay correctly skips stale actions for the interpolator."""
+    cfg = RTCConfig(enabled=True, execution_horizon=10)
+    queue = ActionQueue(cfg)
+    interp = ActionInterpolator(multiplier=2)
+
+    chunk1 = _make_chunk(10)
+    queue.merge(chunk1, chunk1.clone(), real_delay=0)
+
+    for _ in range(5):
+        if interp.needs_new_action():
+            a = queue.get()
+            if a is not None:
+                interp.add(a)
+        interp.get()
+
+    assert queue.get_action_index() == 3
+
+    chunk2 = _make_chunk(10, offset=100.0)
+    queue.merge(chunk2, chunk2.clone(), real_delay=3, action_index_before_inference=0)
+
+    first_action = queue.get()
+    assert first_action is not None
+    torch.testing.assert_close(first_action, torch.tensor([103.0, 103.0]))

tests/processor/test_tokenizer_processor.py

@@ -189,6 +189,30 @@ def test_list_of_strings_tokenization(mock_auto_tokenizer):
     assert attention_mask.shape == (2, 8)
 
 
+@require_package("transformers")
+@patch("lerobot.processor.tokenizer_processor.AutoTokenizer")
+def test_tuple_of_strings_tokenization(mock_auto_tokenizer):
+    """Test tokenization of a tuple of strings (returned by VectorEnv.call())."""
+    mock_tokenizer = MockTokenizer(vocab_size=100)
+    mock_auto_tokenizer.from_pretrained.return_value = mock_tokenizer
+
+    processor = TokenizerProcessorStep(tokenizer_name="test-tokenizer", max_length=8)
+
+    transition = create_transition(
+        observation={"state": torch.tensor([1.0, 2.0])},
+        action=torch.tensor([0.1, 0.2]),
+        complementary_data={"task": ("pick up cube", "place on table")},
+    )
+
+    result = processor(transition)
+
+    observation = result[TransitionKey.OBSERVATION]
+    tokens = observation[f"{OBS_LANGUAGE}.tokens"]
+    attention_mask = observation[f"{OBS_LANGUAGE}.attention_mask"]
+    assert tokens.shape == (2, 8)
+    assert attention_mask.shape == (2, 8)
+
+
 @require_package("transformers")
 @patch("lerobot.processor.tokenizer_processor.AutoTokenizer")
 def test_custom_keys(mock_auto_tokenizer):