fix(image transforms): cleaning up image_transforms implementation in LeRobotDataset

chore(dependencies): Bump lerobot to 0.5.2 (#3307 )
* chore(dependencies): Bump lerobot to 0.5.2 * chore(dependecies): upgrade uv.lock
2026-05-31 10:51:35 +00:00 · 2026-04-07 22:15:34 +02:00 · 2026-04-07 17:17:33 +02:00 · 2026-04-07 16:11:14 +02:00 · 2026-04-07 15:59:11 +02:00 · 2026-04-07 15:50:18 +02:00
42 changed files with 7051 additions and 842 deletions
--- a/.github/workflows/docker_publish.yml
+++ b/.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 }}
--- a/.github/workflows/documentation-upload-pr.yml
+++ b/.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:
--- a/.github/workflows/documentation.yml
+++ b/.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 }}
--- a/.github/workflows/fast_tests.yml
+++ b/.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 != ''
--- a/.github/workflows/full_tests.yml
+++ b/.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
--- a/.github/workflows/unbound_deps_tests.yml
+++ b/.github/workflows/unbound_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" \
--- a/.github/workflows/quality.yml
+++ b/.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
--- a/.github/workflows/release.yml
+++ b/.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 }}
--- a/.github/workflows/security.yml
+++ b/.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
--- a/.gitignore
+++ b/.gitignore
@@ -25,7 +25,6 @@ node_modules/

 # Lock files
 poetry.lock
-uv.lock
 Pipfile.lock

 ### Build & Distribution ###
--- a/README.md
+++ b/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/)
--- a/docker/Dockerfile.internal
+++ b/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

--- a/docker/Dockerfile.user
+++ b/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
--- a/docker/README.md
+++ b/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
+```
--- a/docs/source/adding_benchmarks.mdx
+++ b/docs/source/adding_benchmarks.mdx
@@ -115,22 +115,23 @@ Each `EnvConfig` subclass declares two dicts that tell the policy what to expect
 ## 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.
+  At minimum, you need three files: a **gym.Env wrapper**, an **EnvConfig
+  subclass**, and a **factory dispatch branch**. Everything else is optional or
+  documentation.
 </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                           |
+| 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 for the CLI      |
+| `src/lerobot/envs/factory.py`            | Yes      | Tells `make_env()` how to build your envs |
+| `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        |

 ### 1. The gym.Env wrapper (`src/lerobot/envs/<benchmark>.py`)

@@ -178,10 +179,7 @@ See `create_libero_envs()` (multi-suite, multi-task) and `create_metaworld_envs(

 ### 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.
+Register a config dataclass so users can select your benchmark with `--env.type=<name>`:

 ```python
@EnvConfig.register_subclass("<benchmark_name>")
@@ -206,20 +204,6 @@ class MyBenchmarkEnvConfig(EnvConfig):
    @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 = False):
-        """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:
@@ -227,11 +211,36 @@ 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`)
+### 3. The factory dispatch (`src/lerobot/envs/factory.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):
+Add a branch in `make_env()` to call your factory function:
+
+```python
+elif "<benchmark_name>" in cfg.type:
+    from lerobot.envs.<benchmark> import create_<benchmark>_envs
+
+    if cfg.task is None:
+        raise ValueError("<BenchmarkName> requires a task to be specified")
+
+    return create_<benchmark>_envs(
+        task=cfg.task,
+        n_envs=n_envs,
+        gym_kwargs=cfg.gym_kwargs,
+        env_cls=env_cls,
+    )
+```
+
+If your benchmark needs an env processor, add it in `make_env_pre_post_processors()`:
+
+```python
+if isinstance(env_cfg, MyBenchmarkEnvConfig) or "<benchmark_name>" in env_cfg.type:
+    preprocessor_steps.append(MyBenchmarkProcessorStep())
+```
+
+### 4. 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):

 ```python
@dataclass
@@ -251,7 +260,7 @@ class MyBenchmarkProcessorStep(ObservationProcessorStep):

 See `LiberoProcessorStep` for a full example (image rotation, quaternion-to-axis-angle conversion).

-### 4. Dependencies (`pyproject.toml`)
+### 5. Dependencies (`pyproject.toml`)

 Add a new optional-dependency group:

@@ -272,11 +281,11 @@ Users install with:
 pip install -e ".[mybenchmark]"
 ```

-### 5. Documentation (`docs/source/<benchmark>.mdx`)
+### 6. 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`)
+### 7. Table of contents (`docs/source/_toctree.yml`)

 Add your benchmark to the "Benchmarks" section:

--- a/docs/source/env_processor.mdx
+++ b/docs/source/env_processor.mdx
@@ -151,7 +151,7 @@ observation = {

 ### Factory Function

-The `make_env_pre_post_processors` function delegates to `env_cfg.get_env_processors()`:
+The `make_env_pre_post_processors` function follows the same pattern as `make_pre_post_processors` for policies:

 ```python
 from lerobot.envs.factory import make_env_pre_post_processors
@@ -159,30 +159,46 @@ from lerobot.envs.configs import LiberoEnv, PushtEnv

 # For LIBERO: Returns LiberoProcessorStep in preprocessor
 libero_cfg = LiberoEnv(task="libero_spatial", camera_name=["agentview"])
-env_preprocessor, env_postprocessor = make_env_pre_post_processors(libero_cfg, policy_cfg)
+env_preprocessor, env_postprocessor = make_env_pre_post_processors(libero_cfg)

 # For other environments: Returns identity processors (no-op)
 pusht_cfg = PushtEnv()
-env_preprocessor, env_postprocessor = make_env_pre_post_processors(pusht_cfg, policy_cfg)
+env_preprocessor, env_postprocessor = make_env_pre_post_processors(pusht_cfg)
 ```

-### How It Works
-
-Each `EnvConfig` subclass can override `get_env_processors()` to return benchmark-specific
-processor pipelines. The base class returns identity (no-op) processors by default.
+### Implementation in `envs/factory.py`

 ```python
-# In your EnvConfig subclass:
-def get_env_processors(self):
-    from lerobot.processor.pipeline import PolicyProcessorPipeline
-    return (
-        PolicyProcessorPipeline(steps=[MyProcessorStep()]),
-        PolicyProcessorPipeline(steps=[]),
-    )
-```
+def make_env_pre_post_processors(
+    env_cfg: EnvConfig,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+]:
+    """
+    Create preprocessor and postprocessor pipelines for environment observations.

-The factory function `make_env_pre_post_processors` simply delegates to this method,
-with a special case for `XVLAConfig` policies which override the env processors entirely.
+    Args:
+        env_cfg: The configuration of the environment.
+
+    Returns:
+        A tuple containing:
+            - preprocessor: Pipeline that processes environment observations
+            - postprocessor: Pipeline that processes environment outputs
+    """
+    # For LIBERO environments, add the LiberoProcessorStep to preprocessor
+    if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
+        preprocessor = PolicyProcessorPipeline(steps=[LiberoProcessorStep()])
+    else:
+        # For all other environments, return an identity preprocessor
+        preprocessor = PolicyProcessorPipeline(steps=[])
+
+    # Postprocessor is currently identity for all environments
+    # Future: Could add environment-specific action transformations
+    postprocessor = PolicyProcessorPipeline(steps=[])
+
+    return preprocessor, postprocessor
+```

 ### Integration in Evaluation

--- a/docs/source/groot.mdx
+++ b/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**.
--- a/docs/source/installation.mdx
+++ b/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 🤗

--- a/docs/source/multi_task_dit.mdx
+++ b/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:
--- a/docs/source/policy_pi05_README.md
+++ b/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).
--- a/docs/source/policy_pi0_README.md
+++ b/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).
--- a/docs/source/policy_rtc_README.md
+++ b/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.
--- a/docs/source/policy_sarm_README.md
+++ b/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}
+}
+```
--- a/pyproject.toml
+++ b/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",
--- a/src/lerobot/configs/default.py
+++ b/src/lerobot/configs/default.py
@@ -67,8 +67,7 @@ class EvalConfig:
    # `batch_size` specifies the number of environments to use in a gym.vector.VectorEnv.
    batch_size: int = 50
    # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
-    # Defaults to True; automatically downgraded to SyncVectorEnv when batch_size=1.
-    use_async_envs: bool = True
+    use_async_envs: bool = False

    def __post_init__(self) -> None:
        if self.batch_size > self.n_episodes:
--- a/src/lerobot/datasets/dataset_reader.py
+++ b/src/lerobot/datasets/dataset_reader.py
@@ -72,6 +72,8 @@ class DatasetReader:
        self.episodes = episodes
        self._tolerance_s = tolerance_s
        self._video_backend = video_backend
+        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

        self.hf_dataset: datasets.Dataset | None = None
@@ -83,6 +85,16 @@ class DatasetReader:
            check_delta_timestamps(delta_timestamps, meta.fps, tolerance_s)
            self.delta_indices = get_delta_indices(delta_timestamps, meta.fps)

+    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
+
+    def clear_image_transforms(self) -> None:
+        """Remove the transform applied to visual observations."""
+        self._image_transforms = None
+
    def try_load(self) -> bool:
        """Attempt to load from local cache. Returns True if data is sufficient."""
        try:
--- a/src/lerobot/datasets/lerobot_dataset.py
+++ b/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 +194,6 @@ 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.delta_timestamps = delta_timestamps
        self.episodes = episodes
        self.tolerance_s = tolerance_s
@@ -222,6 +223,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
            delta_timestamps=delta_timestamps,
            image_transforms=image_transforms,
        )
+        self.image_transforms = image_transforms

        # Load actual data
        if force_cache_sync or not self.reader.try_load():
@@ -475,6 +477,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
            f"}})"
        )

+    def set_image_transforms(self, image_transforms: Callable | None) -> None:
+        """Replace the transform applied to visual observations."""
+        self._ensure_reader().set_image_transforms(image_transforms)
+        self.image_transforms = image_transforms
+
+    def clear_image_transforms(self) -> None:
+        """Remove the transform applied to visual observations."""
+        if self.reader is not None:
+            self.reader.set_image_transforms(None)
+        self.image_transforms = None
+
    # ── Hub methods (stay on facade) ──────────────────────────────────

    def push_to_hub(
--- a/src/lerobot/datasets/multi_dataset.py
+++ b/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):
--- a/src/lerobot/envs/configs.py
+++ b/src/lerobot/envs/configs.py
@@ -12,16 +12,11 @@
 # 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
@@ -72,45 +67,6 @@ 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)
-
-        vec = env_cls([_make_one for _ in range(n_envs)], autoreset_mode=gym.vector.AutoresetMode.SAME_STEP)
-        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):
@@ -389,32 +345,6 @@ class LiberoEnv(EnvConfig):
            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 = gym.vector.AsyncVectorEnv if (use_async_envs and n_envs > 1) else gym.vector.SyncVectorEnv
-        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,
-        )
-
-    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
@@ -457,19 +387,6 @@ 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 = gym.vector.AsyncVectorEnv if (use_async_envs and n_envs > 1) else gym.vector.SyncVectorEnv
-        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
@@ -537,18 +454,3 @@ 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=[]),
-        )
--- a/src/lerobot/envs/factory.py
+++ b/src/lerobot/envs/factory.py
@@ -13,52 +13,96 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from __future__ import annotations
-
+import importlib
 from typing import Any

 import gymnasium as gym
+from gymnasium.envs.registration import registry as gym_registry

-from lerobot.envs.configs import EnvConfig, HubEnvConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.envs.configs import AlohaEnv, EnvConfig, HubEnvConfig, IsaaclabArenaEnv, LiberoEnv, PushtEnv
 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:
-    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)
+    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.")


 def make_env_pre_post_processors(
    env_cfg: EnvConfig,
-    policy_cfg: Any,
-) -> tuple[Any, Any]:
+    policy_cfg: PreTrainedConfig,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+]:
    """
    Create preprocessor and postprocessor pipelines for environment observations.

-    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.
-    """
-    from lerobot.policies.xvla.configuration_xvla import XVLAConfig
+    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)
+    """
+    # Preprocessor and Postprocessor steps are Identity for most environments
+    preprocessor_steps: list[ProcessorStep] = []
+    postprocessor_steps: list[ProcessorStep] = []
    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()

-    return env_cfg.get_env_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


 def make_env(
    cfg: EnvConfig | str,
    n_envs: int = 1,
-    use_async_envs: bool = True,
+    use_async_envs: bool = False,
    hub_cache_dir: str | None = None,
    trust_remote_code: bool = False,
 ) -> dict[str, dict[int, gym.vector.VectorEnv]]:
@@ -119,4 +163,57 @@ def make_env(
    if n_envs < 1:
        raise ValueError("`n_envs` must be at least 1")

-    return cfg.create_envs(n_envs=n_envs, use_async_envs=use_async_envs)
+    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}}
--- a/src/lerobot/envs/libero.py
+++ b/src/lerobot/envs/libero.py
@@ -150,17 +150,7 @@ class LiberoEnv(gym.Env):

        self.init_state_id = self.episode_index  # tie each sub-env to a fixed init state

-        # 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
-        )
-
+        self._env = self._make_envs_task(task_suite, self.task_id)
        default_steps = 500
        self._max_episode_steps = (
            TASK_SUITE_MAX_STEPS.get(task_suite_name, default_steps)
@@ -231,32 +221,28 @@ 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"]
        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]
@@ -308,7 +294,6 @@ 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()
@@ -335,8 +320,6 @@ 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,)), "
@@ -367,8 +350,7 @@ class LiberoEnv(gym.Env):
        return observation, reward, terminated, truncated, info

    def close(self):
-        if self._env is not None:
-            self._env.close()
+        self._env.close()


 def _make_env_fns(
--- a/src/lerobot/envs/metaworld.py
+++ b/src/lerobot/envs/metaworld.py
@@ -97,9 +97,8 @@ class MetaworldEnv(gym.Env):
        self.visualization_height = visualization_height
        self.camera_name = camera_name

-        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._env = self._make_envs_task(self.task)
+        self._max_episode_steps = self._env.max_path_length
        self.task_description = TASK_DESCRIPTIONS[self.task]

        self.expert_policy = TASK_POLICY_MAPPING[self.task]()
@@ -137,24 +136,6 @@ 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.
@@ -162,13 +143,26 @@ 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:
@@ -215,7 +209,6 @@ 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)
@@ -239,7 +232,6 @@ 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,)), "
@@ -271,8 +263,7 @@ class MetaworldEnv(gym.Env):
        return observation, reward, terminated, truncated, info

    def close(self):
-        if self._env is not None:
-            self._env.close()
+        self._env.close()


 # ---- Main API ----------------------------------------------------------------
--- a/src/lerobot/policies/multi_task_dit/README.md
+++ b/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}
-}
-```
--- a/src/lerobot/policies/multi_task_dit/README.md
+++ b/src/lerobot/policies/multi_task_dit/README.md
@@ -0,0 +1 @@
+../../../../docs/source/policy_multi_task_dit_README.md
--- a/src/lerobot/policies/pi0/README.md
+++ b/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).
--- a/src/lerobot/policies/pi0/README.md
+++ b/src/lerobot/policies/pi0/README.md
@@ -0,0 +1 @@
+../../../../docs/source/policy_pi0_README.md
--- a/src/lerobot/policies/pi05/README.md
+++ b/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).
--- a/src/lerobot/policies/pi05/README.md
+++ b/src/lerobot/policies/pi05/README.md
@@ -0,0 +1 @@
+../../../../docs/source/policy_pi05_README.md
--- a/src/lerobot/policies/rtc/README.md
+++ b/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.
--- a/src/lerobot/policies/rtc/README.md
+++ b/src/lerobot/policies/rtc/README.md
@@ -0,0 +1 @@
+../../../../docs/source/policy_rtc_README.md
--- a/src/lerobot/policies/sarm/README.md
+++ b/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}
-}
-```
--- a/src/lerobot/policies/sarm/README.md
+++ b/src/lerobot/policies/sarm/README.md
@@ -0,0 +1 @@
+../../../../docs/source/policy_sarm_README.md
--- a/src/lerobot/scripts/lerobot_eval.py
+++ b/src/lerobot/scripts/lerobot_eval.py
@@ -47,7 +47,6 @@ You can learn about the CLI options for this script in the `EvalPipelineConfig`
 """

 import concurrent.futures as cf
-import copy
 import json
 import logging
 import threading
@@ -57,6 +56,7 @@ from collections.abc import Callable
 from contextlib import nullcontext
 from copy import deepcopy
 from dataclasses import asdict
+from functools import partial
 from pathlib import Path
 from pprint import pformat
 from typing import Any, TypedDict
@@ -73,6 +73,7 @@ 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,
@@ -165,9 +166,9 @@ def rollout(
        if return_observations:
            all_observations.append(deepcopy(observation))

-        # Infer "task" from sub-environments.
-        # env.call() works with both SyncVectorEnv and AsyncVectorEnv.
-        observation["task"] = env.call("task")
+        # Infer "task" from attributes of environments.
+        # TODO: works with SyncVectorEnv but not AsyncVectorEnv
+        observation = add_envs_task(env, observation)

        # Apply environment-specific preprocessing (e.g., LiberoProcessorStep for LIBERO)
        observation = env_preprocessor(observation)
@@ -733,48 +734,34 @@ def eval_policy_all(
            group_acc[group]["video_paths"].extend(paths)
            overall["video_paths"].extend(paths)

-    def _make_thread_policy(p: PreTrainedPolicy) -> PreTrainedPolicy:
-        """Shallow copy sharing weight tensors, with independent per-thread state.
-
-        copy.copy() gives a new Python object whose _parameters dict is a shared
-        reference (same tensor storage, zero extra VRAM). reset() then rebinds
-        mutable state (action queues etc.) to fresh per-thread objects.
-
-        Note: does NOT work for ACT with temporal_ensemble_coeff — that policy's
-        reset() mutates a shared sub-object. Use max_parallel_tasks=1 for that config.
-        """
-        thread_p = copy.copy(p)
-        thread_p.reset()
-        return thread_p
-
    # Choose runner (sequential vs threaded)
-    _runner_kwargs = {
-        "env_preprocessor": env_preprocessor,
-        "env_postprocessor": env_postprocessor,
-        "preprocessor": preprocessor,
-        "postprocessor": postprocessor,
-        "n_episodes": n_episodes,
-        "max_episodes_rendered": max_episodes_rendered,
-        "videos_dir": videos_dir,
-        "return_episode_data": return_episode_data,
-        "start_seed": start_seed,
-    }
+    task_runner = partial(
+        run_one,
+        policy=policy,
+        env_preprocessor=env_preprocessor,
+        env_postprocessor=env_postprocessor,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
+        n_episodes=n_episodes,
+        max_episodes_rendered=max_episodes_rendered,
+        videos_dir=videos_dir,
+        return_episode_data=return_episode_data,
+        start_seed=start_seed,
+    )

    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 = run_one(task_group, task_id, env, policy=policy, **_runner_kwargs)
+            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})
    else:
-        # threaded path: each thread gets a shallow policy copy (shared weights, independent state)
+        # 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(
-                    run_one, task_group, task_id, env, policy=_make_thread_policy(policy), **_runner_kwargs
-                )
+                fut = executor.submit(task_runner, task_group, task_id, env)
                fut2meta[fut] = (task_group, task_id)
            for fut in cf.as_completed(fut2meta):
                tg, tid, metrics = fut.result()
--- a/src/lerobot/scripts/lerobot_record.py
+++ b/src/lerobot/scripts/lerobot_record.py
@@ -421,6 +421,7 @@ def record_loop(
                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()
--- a/tests/datasets/test_datasets.py
+++ b/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):
--- a/tests/envs/test_dispatch.py
+++ b/tests/envs/test_dispatch.py
@@ -1,143 +0,0 @@
-"""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)
-        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()."""
-    from lerobot.configs.policies import PreTrainedConfig
-
-    cfg = make_env_config("aloha")
-    pre, post = make_env_pre_post_processors(cfg, PreTrainedConfig())
-    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.SyncVectorEnv)
-        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 PolicyProcessorPipeline
-
-    @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 PolicyProcessorPipeline(steps=[]), PolicyProcessorPipeline(steps=[])
-
-    pre, post = _Env().get_env_processors()
-    assert isinstance(pre, PolicyProcessorPipeline)
--- a/uv.lock
+++ b/uv.lock
Author	SHA1	Message	Date
CarolinePascal	2c78c46fcd	fix(image transforms): cleaning up image_transforms implementation in LeRobotDataset	2026-04-07 22:15:34 +02:00
Steven Palma	4eecbad32b	chore(dependencies): Bump lerobot to 0.5.2 (#3307 ) * chore(dependencies): Bump lerobot to 0.5.2 * chore(dependecies): upgrade uv.lock	2026-04-07 17:17:33 +02:00
Pauline Bailly-Masson	1396b9fab7	🔒 Pin GitHub Actions to commit SHAs (#3265 ) * 🔒 pin quality.yml actions to commit SHAs * 🔒 pin fast_tests.yml actions to commit SHAs * 🔒 pin full_tests.yml actions to commit SHAs * 🔒 pin documentation.yml actions to commit SHAs * 🔒 pin documentation-upload-pr.yml actions to commit SHAs * 🔒 pin release.yml actions to commit SHAs * 🔒 pin security.yml actions to commit SHAs --------- Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>	2026-04-07 16:11:14 +02:00
Francesco Capuano	7c032f19fc	feat(dataset): registering torchvision transforms (#3153 ) * add: a flexible transformation registry * fix: image transforms can be set both at init and after * add: tests * fix: take in review * feat(datasets): add image transform setters * fix: pre-commit * fix: CI --------- Signed-off-by: Francesco Capuano <74058581+fracapuano@users.noreply.github.com>	2026-04-07 15:59:11 +02:00
Anthony Chan	e2f27bf71b	Fix lerobot_train script without interpolation (#3281 )	2026-04-07 15:50:18 +02:00
Steven Palma	ea36a4a176	chore(docs): new badge for readme (#3303 )	2026-04-07 10:47:03 +02:00
Steven Palma	399b3c9ba5	chore(dependencies): update uv.lock (#3302 ) Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>	2026-04-07 09:49:00 +02:00
Steven Palma	913041e753	fix(ci): latest deps tests permissions (#3296 ) * fix(ci): latest deps tests permissions * fix(ci): force push dep update branch * fix(ci): change secret for permissions & Ci trigger	2026-04-06 14:56:05 +02:00
Steven Palma	2b541ddd4c	docs(ci): add readme for dockerfile (#3295 )	2026-04-06 13:22:45 +02:00
Steven Palma	50a1e67e94	feat(ci): add `uv.lock` (#3292 ) * feat(ci): add uv.lock * feat(ci): use uv.lock in CI PR testing * chore(ci): rename nightly to docker publish and test * feat(ci): automated update of uv.lock + remove unbound check + docker images now use uv.lock * fix(ci): add --force-with-lease + set -e for silent erros	2026-04-06 12:23:37 +02:00
Steven Palma	d60a700d2b	chore(policy): multi dit docs (#3285 ) * docs(policy): add libero results multi task dit + remove readme in src code * docs(policy): add hyperlink to doc file in src code * chore(style): pre-commit	2026-04-05 21:23:13 +02:00
Steven Palma	8c3d4cf900	chore(docs): no policy readme in src code (#3286 ) * chore(docs): move policies readme out of src code * chore(docs): create symlink for policy readme	2026-04-05 19:25:38 +02:00
Caroline Pascal	b6e60a6e30	chore(dependencies): bump minimum torch from 2.2.1 to 2.7 (#3156 ) * feat(ffmpeg): updating ffmpeg verion to 8.X * Revert "feat(ffmpeg): updating ffmpeg verion to 8.X" This reverts commit `bb0f03185c`. * chore(pyproject): updating pyproject to fit the minimally required version of torchcodec * chore(docs): updating doc with specific instructions for ffmpeg/torchcodec installation * fix(typo): reverting ceiling bound on pytorch to 2.11.0 * chore(format): removing empty line * chore(typo): fixing typo * chore(docs): adding warning in case of torchcodec/ffmpeg version mismatch * chore(docs): applying comments * chore(docs): adding uv commands for evdev on WSL * fix(typo): fixing typo * fix(typo): fixing typos again * chore(ruff): format * fix(evdev install): splitting evdev install instructions between conda and uv * chore(ruff): format --------- Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>	2026-04-05 19:24:43 +02:00
Steven Palma	3596681d94	docs(policy): fix gr00t license docs (#3284 )	2026-04-05 19:09:15 +02:00
				`@@ -0,0 +1 @@`
				`../../../../docs/source/policy_multi_task_dit_README.md`
				`@@ -0,0 +1 @@`
				`../../../../docs/source/policy_pi0_README.md`
				`@@ -0,0 +1 @@`
				`../../../../docs/source/policy_pi05_README.md`
				`@@ -0,0 +1 @@`
				`../../../../docs/source/policy_rtc_README.md`
				`@@ -0,0 +1 @@`
				`../../../../docs/source/policy_sarm_README.md`