draft changes

* feat(ci): adds release versioning to docs

* chore(ci): remove TODO

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -12,57 +12,83 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-name: "\U0001F41B Bug Report"
-description: Submit a bug report to help us improve LeRobot
+name: "🚀 Issue / Bug / Request"
+description: Report a bug, suggest an improvement, or ask a technical question.
 body:
   - type: markdown
     attributes:
       value: |
-        Thanks for taking the time to submit a bug report! 🐛
-        If this is not a bug related to the LeRobot library directly, but instead a general question about your code or the library specifically please use our [discord](https://discord.gg/s3KuuzsPFb).
+        ### Thanks for contributing to LeRobot! 🙌
+        Please choose the most relevant sections below. If this is a general "how-to" question, consider our [Discord](https://discord.gg/s3KuuzsPFb) for faster community support.
+
+  - type: dropdown
+    id: issue-type
+    attributes:
+      label: Ticket Type
+      description: What kind of ticket are you opening?
+      options:
+        - "🐛 Bug Report (Something isn't working)"
+        - "💡 Feature Request / Improvement"
+        - "❓ Technical Question"
+        - "🧹 Maintenance / Documentation"
+    validations:
+      required: true
 
   - type: textarea
     id: system-info
     attributes:
-      label: System Info
-      description: Please share your LeRobot configuration by running `lerobot-info` (if installed) or `python -m lerobot.scripts.display_sys_info` (if not installed) and pasting the output below.
+      label: Environment & System Info
+      description: |
+        For bugs or technical questions, please run `lerobot-info` and paste the output.
+        (Optional for feature requests).
       render: Shell
-      placeholder: lerobot version, OS, python version, numpy version, torch version, and lerobot's configuration
+      placeholder: lerobot version, OS, python version, etc.
+
+  - type: textarea
+    id: description
     validations:
       required: true
+    attributes:
+      label: Description
+      description: |
+        Provide a clear summary of the issue or your proposal.
+        - **Bugs:** What is happening?
+        - **Features:** What is the goal/use case?
+        - **Questions:** What are you trying to achieve?
+      placeholder: |
+        A clear and concise description of the issue or suggestion.
+
+  - type: textarea
+    id: context-repro
+    attributes:
+      label: Context & Reproduction
+      description: |
+        Provide a code snippet, steps to reproduce a bug, or technical details about your proposal.
+        Please use code blocks for scripts and CLI commands.
+      placeholder: |
+        Steps to reproduce / Usage example:
+        1.
+        2.
+        3.
+
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant logs or stack trace
+      description: If applicable, paste relevant error logs here.
+      render: Shell
 
   - type: checkboxes
-    id: information-scripts-examples
+    id: extras
     attributes:
-      label: Information
-      description: 'The problem arises when using:'
+      label: Checklist
       options:
-        - label: "One of the scripts in the examples/ folder of LeRobot"
-        - label: "My own task or dataset (give details below)"
+        - label: I have searched existing tickets to ensure this isn't a duplicate.
+        - label: I am using the latest version of the `main` branch.
+        - label: I have verified this is not an environment-specific problem.
 
   - type: textarea
-    id: reproduction
-    validations:
-      required: true
+    id: workaround
     attributes:
-      label: Reproduction
-      description: |
-        If needed, provide a simple code sample that reproduces the problem you ran into. It can be a Colab link or just a code snippet.
-        Sharing error messages or stack traces could be useful as well!
-        Important! Use code tags to correctly format your code. See https://help.github.com/en/github/writing-on-github/creating-and-highlighting-code-blocks#syntax-highlighting
-        Try to avoid screenshots, as they are hard to read and don't allow copy-and-pasting.
-
-      placeholder: |
-        Steps to reproduce the behavior:
-
-          1.
-          2.
-          3.
-
-  - type: textarea
-    id: expected-behavior
-    validations:
-      required: true
-    attributes:
-      label: Expected behavior
-      description: "A clear and concise description of what you would expect to happen."
+      label: Additional Info / Workarounds
+      description: Anything else we should know? If you have a workaround, please share it!

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,41 +1,54 @@
-## What this does
+## Title
 
-Explain what this PR does. Feel free to tag your PR with the appropriate label(s).
+Short, imperative summary (e.g., "fix(robots): handle None in sensor parser"). See [CONTRIBUTING.md](../CONTRIBUTING.md) for PR conventions.
 
-Examples:
-| Title | Label |
-|----------------------|-----------------|
-| Fixes #[issue] | (🐛 Bug) |
-| Adds new dataset | (🗃️ Dataset) |
-| Optimizes something | (⚡️ Performance) |
+## Type / Scope
 
-## How it was tested
+- **Type**: (Bug | Feature | Docs | Performance | Test | CI | Chore)
+- **Scope**: (optional — name of module or package affected)
 
-Explain/show how you tested your changes.
+## Summary / Motivation
 
-Examples:
+- One-paragraph description of what changes and why.
+- Why this change is needed and any trade-offs or design notes.
 
-- Added `test_something` in `tests/test_stuff.py`.
-- Added `new_feature` and checked that training converges with policy X on dataset/environment Y.
-- Optimized `some_function`, it now runs X times faster than previously.
+## Related issues
 
-## How to checkout & try? (for the reviewer)
+- Fixes / Closes: # (if any)
+- Related: # (if any)
 
-Provide a simple way for the reviewer to try out your changes.
+## What changed
 
-Examples:
+- Short, concrete bullets of the modifications (files/behaviour).
+- Short note if this introduces breaking changes and migration steps.
 
-```bash
-pytest -sx tests/test_stuff.py::test_something
-```
+## How was this tested
 
-```bash
-lerobot-train --some.option=true
-```
+- Tests added: list new tests or test files.
+- Manual checks / dataset runs performed.
 
-## SECTION TO REMOVE BEFORE SUBMITTING YOUR PR
+## How to run locally (reviewer)
 
-**Note**: Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
-members/contributors who may be interested in your PR. Try to avoid tagging more than 3 people.
+- Run the relevant tests:
 
-**Note**: Before submitting this PR, please read the [contributor guideline](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md#submitting-a-pull-request-pr).
+  ```bash
+  pytest -q tests/ -k <keyword>
+  ```
+
+- Run a quick example or CLI (if applicable):
+
+  ```bash
+  lerobot-train --some.option=true
+  ```
+
+## Checklist (required before merge)
+
+- [ ] Linting/formatting run (`pre-commit run -a`)
+- [ ] All tests pass locally (`pytest`)
+- [ ] Documentation updated
+- [ ] CI is green
+
+## Reviewer notes
+
+- Anything the reviewer should focus on (performance, edge-cases, specific files) or general notes.
+- Anyone in the community is free to review the PR.

.github/labeler.yml

@@ -0,0 +1,69 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+CI:
+  - changed-files:
+      - any-glob-to-any-file:
+        - '.github/**'
+        - 'docker/**'
+
+github_actions:
+  - changed-files:
+      - any-glob-to-any-file: '.github/**'
+
+documentation:
+  - changed-files:
+      - any-glob-to-any-file:
+          - '**/*.md'
+          - '**/*.mdx'
+          - 'docs/**'
+
+examples:
+  - changed-files:
+      - any-glob-to-any-file: 'examples/**'
+
+tests:
+  - changed-files:
+      - any-glob-to-any-file: 'tests/**'
+
+sensors:
+  - changed-files:
+      - any-glob-to-any-file: 'src/lerobot/cameras/**'
+
+configuration:
+  - changed-files:
+      - any-glob-to-any-file: 'src/lerobot/configs/**'
+
+dataset:
+  - changed-files:
+      - any-glob-to-any-file: 'src/lerobot/datasets/**'
+
+evaluation:
+  - changed-files:
+      - any-glob-to-any-file: 'src/lerobot/envs/**'
+
+robots:
+  - changed-files:
+      - any-glob-to-any-file:
+          - 'src/lerobot/teleoperators/**'
+          - 'src/lerobot/robots/**'
+          - 'src/lerobot/motors/**'
+
+policies:
+  - changed-files:
+      - any-glob-to-any-file: 'src/lerobot/policies/**'
+
+processor:
+  - changed-files:
+      - any-glob-to-any-file: 'src/lerobot/processor/**'

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

@@ -31,7 +31,8 @@ jobs:
     name: Upload Preview and Comment
     if: >
       github.event.workflow_run.event == 'pull_request' &&
-      github.event.workflow_run.conclusion == 'success'
+      github.event.workflow_run.conclusion == 'success' &&
+      github.repository == 'huggingface/lerobot'
     uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@main
     with:
       package_name: lerobot

.github/workflows/documentation.yml

@@ -33,6 +33,9 @@ on:
     paths:
       - "docs/**"
 
+  release:
+    types: [published]
+
 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
@@ -42,14 +45,16 @@ jobs:
   # This job builds and deploys the official documentation.
   build_main_docs:
     name: Build Main Docs
-    if: github.event_name == 'push' || github.event_name == 'workflow_dispatch'
+    if: >
+      (github.event_name == 'push' || github.event_name == 'workflow_dispatch' || github.event_name == 'release') &&
+      github.repository == 'huggingface/lerobot'
     permissions:
       contents: read
     uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@main
     with:
       commit_sha: ${{ github.sha }}
       package: lerobot
-      additional_args: --not_python_module
+      additional_args: --not_python_module ${{ github.event_name == 'release' && format('--version {0}', github.event.release.tag_name) || '' }}
     secrets:
       token: ${{ secrets.HUGGINGFACE_PUSH }}
       hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}
@@ -58,7 +63,7 @@ jobs:
   # The result of this job triggers the 'Upload PR Documentation' workflow.
   build_pr_docs:
     name: Build PR Docs
-    if: github.event_name == 'pull_request'
+    if: github.event_name == 'pull_request' && github.repository == 'huggingface/lerobot'
     permissions:
       contents: read
       pull-requests: write

.github/workflows/fast_tests.yml

@@ -45,7 +45,6 @@ permissions:
 env:
   UV_VERSION: "0.8.0"
   PYTHON_VERSION: "3.10"
-  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu
 
 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
 concurrency:
@@ -60,12 +59,19 @@ jobs:
     runs-on: ubuntu-latest
     env:
       MUJOCO_GL: egl
+      HF_HOME: /mnt/cache/.cache/huggingface
+      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           persist-credentials: false
           lfs: true
 
+      # 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
+        run: sudo chown -R $USER:$USER /mnt
+
       # TODO(Steven): Evaluate the need of these dependencies
       - name: Install apt dependencies
         run: |

.github/workflows/full_tests.yml

@@ -58,12 +58,19 @@ jobs:
       github.event_name == 'workflow_dispatch'
     env:
       MUJOCO_GL: egl
+      HF_HOME: /mnt/cache/.cache/huggingface
+      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           lfs: true
           persist-credentials: false
 
+      # 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
+        run: sudo chown -R $USER:$USER /mnt
+
       - name: Install apt dependencies
         run: |
           sudo apt-get update && sudo apt-get install -y build-essential \
@@ -78,7 +85,7 @@ jobs:
           python-version: ${{ env.PYTHON_VERSION }}
 
       - name: Install lerobot with all extras
-        run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional
+        run: uv sync --extra all # TODO(Steven): Make flash-attn optional
 
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv --maxfail=10
@@ -120,7 +127,7 @@ jobs:
           sudo apt-get update
           sudo apt-get install git-lfs
           git lfs install
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           lfs: true
           persist-credentials: false

.github/workflows/issue_labeler.yml

@@ -0,0 +1,77 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This workflow automatically labels issues based on their content.
+name: Issue Labeler
+on:
+  # Trigger on new issues and edits to existing issues
+  issues:
+    types: [opened, edited]
+
+permissions:
+  contents: read
+  issues: write
+
+jobs:
+  label-issue:
+    name: Auto Label Issue
+    runs-on: ubuntu-latest
+    if: github.repository == 'huggingface/lerobot'
+    steps:
+      - uses: actions/github-script@v8
+        with:
+          script: |
+            // Setup Input Text
+            const body = (context.payload.issue.body || '');
+            const title = (context.payload.issue.title || '');
+            const cleanBody = body.replace(/```[\s\S]*?```/g, '');
+            const text = `${title}\n${cleanBody}`.toLowerCase();
+            const labelsToAdd = new Set();
+            const matches = (re) => re.test(text);
+
+            // Keyword Heuristics
+
+            if (matches(/\b(bug|error|crash|exception)\b/i)) labelsToAdd.add('bug');
+            if (matches(/\b(new feature|enhancement|improvement|proposal|feature request)\b/i)) labelsToAdd.add('enhancement');
+            if (matches(/\b(question|how to|clarify|explain|how do i|help me|question about)\b/i)) labelsToAdd.add('question');
+            if (matches(/\b(documentation|docs?|readme|tutorial|wiki|typo|docstring)\b/i)) labelsToAdd.add('documentation');
+            if (matches(/\b(example|sample|demo|notebook)s?\b/i)) labelsToAdd.add('examples');
+            if (matches(/\b(datasets?|data loader|data augmentation|data preprocessing)\b/i)) labelsToAdd.add('dataset');
+            if (matches(/\b(mujoco|isaac|simulation|sim)\b/i)) labelsToAdd.add('simulation');
+            if (matches(/\b(train|training|optimizer|gradient|wandb|sac)\b/i)) labelsToAdd.add('training');
+            if (matches(/\b(rerun|plot|render|rendering|visualizer)/i)) labelsToAdd.add('visualization');
+            if (matches(/\b(cameras?|opencv|realsense|lidars?|sensors?|imus?|microphones?|rgbd|encoders?)\b/i)) labelsToAdd.add('sensors');
+            if (matches(/\b(urdf|actuators?|calibration|end-effector|kinematics)\b/i)) labelsToAdd.add('robots');
+            if (matches(/\b(teleop|teleoperator|controller|leader|follower|joystick|gamepad)\b/i)) labelsToAdd.add('teleoperators');
+            if (matches(/\b(policy|policies|model?)\b/i)) labelsToAdd.add('policies');
+            if (matches(/\b(processor|pipeline|preprocessor|postprocessor)s?\b/i)) labelsToAdd.add('processor');
+            if (matches(/\b(eval|evaluate|evaluation|metrics?|score|benchmarks?)\b/i)) labelsToAdd.add('evaluation');
+            if (matches(/\b(tests?|pytest|unittest|failing test)\b/i)) labelsToAdd.add('tests');
+            if (matches(/\b(ci|github actions?|github workflows?|gha|docker|pypi)\b/i)) labelsToAdd.add('CI');
+            if (matches(/\b(perf|latency|throughput|fps|speed|performance|slow|fast|slower|faster|memory usage)\b/i)) labelsToAdd.add('performance');
+            if (matches(/\b(dependency|dependencies|pip|install error|importerror|package not found|pyproject)\b/i)) labelsToAdd.add('dependencies');
+            if (matches(/\b(configuration|config|arguments?|input feature|dracuss)\b/i)) labelsToAdd.add('configuration');
+
+            // Apply Labels
+            const labels = Array.from(labelsToAdd).filter(Boolean);
+
+            if (labels.length > 0) {
+              console.log(`Adding labels: ${labels.join(', ')}`);
+              await github.rest.issues.addLabels({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                issue_number: context.issue.number,
+                labels,
+              });
+            }

.github/workflows/nightly.yml

@@ -43,6 +43,7 @@ jobs:
     name: Build CPU Docker for Nightly
     runs-on:
       group: aws-general-8-plus
+    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME_CPU }}
     steps:
@@ -51,7 +52,7 @@ jobs:
           sudo apt-get update
           sudo apt-get install git-lfs
           git lfs install
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           lfs: true
           persist-credentials: false
@@ -77,6 +78,7 @@ jobs:
     name: Build GPU Docker for Nightly
     runs-on:
       group: aws-general-8-plus
+    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME_GPU }}
     steps:
@@ -85,7 +87,7 @@ jobs:
           sudo apt-get update
           sudo apt-get install git-lfs
           git lfs install
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           lfs: true
           persist-credentials: false

.github/workflows/pr_labeler.yml

@@ -0,0 +1,39 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This workflow labels pull requests based on the files that were changed.
+name: Pull Request Labeler
+
+on:
+  # Allows labeling pull requests when they are opened or updated
+  # zizmor: ignore[dangerous-triggers] Needed to label PRs from forks
+  pull_request_target:
+    branches:
+      - main
+    types: [opened, synchronize, reopened, ready_for_review]
+
+permissions:
+  contents: read
+  pull-requests: write
+
+jobs:
+  triage:
+    name: Label PR
+    runs-on: ubuntu-latest
+    if: github.repository == 'huggingface/lerobot' && !github.event.pull_request.draft
+    steps:
+      - uses: actions/labeler@v6
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
+          sync-labels: true # Removes labels if files are removed from the PR

.github/workflows/quality.yml

@@ -43,12 +43,12 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           persist-credentials: false
 
       - name: Set up Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: '3.10'
 

.github/workflows/release.yml

@@ -29,6 +29,7 @@ jobs:
   build-and-publish:
     name: Build and publish Python distributions
     runs-on: ubuntu-latest
+    if: github.repository == 'huggingface/lerobot'
     outputs:
       version: ${{ steps.extract_info.outputs.tag_version }}
     permissions:
@@ -37,12 +38,12 @@ jobs:
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           persist-credentials: false
 
       - name: Set up Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: '3.10'
 
@@ -134,7 +135,7 @@ jobs:
     env:
       MUJOCO_GL: egl
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           lfs: true
           persist-credentials: false
@@ -176,4 +177,3 @@ jobs:
 
 # TODO(Steven): Publish draft/pre-release and to test pypi weekly
 # TODO(Steven): Separate build and publish job
-# TODO(Steven): Tag documentation with the same version as the package

.github/workflows/security.yml

@@ -43,7 +43,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4 # zizmor: ignore[unpinned-uses]
+        uses: actions/checkout@v6 # zizmor: ignore[unpinned-uses]
         with:
           fetch-depth: 0
           persist-credentials: false

.github/workflows/stale.yml

@@ -45,6 +45,7 @@ jobs:
   stale:
     name: Close Stale Issues and PRs
     runs-on: ubuntu-latest
+    if: github.repository == 'huggingface/lerobot'
     permissions:
       actions: write
       contents: write # only for delete-branch option

.github/workflows/unbound_deps_tests.yml

@@ -43,14 +43,22 @@ jobs:
   full-tests:
     name: Full Unbound Tests
     runs-on: ubuntu-latest
+    if: github.repository == 'huggingface/lerobot'
     env:
       MUJOCO_GL: egl
+      HF_HOME: /mnt/cache/.cache/huggingface
+      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           lfs: true
           persist-credentials: false
 
+      # 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
+        run: sudo chown -R $USER:$USER /mnt
+
       - name: Install apt dependencies
         run: |
           sudo apt-get update && sudo apt-get install -y build-essential \
@@ -70,7 +78,7 @@ jobs:
           echo "Dependencies unbound:" && cat pyproject.toml
 
       - name: Install lerobot with all extras
-        run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional
+        run: uv sync --extra all # TODO(Steven): Make flash-attn optional
 
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv
@@ -93,7 +101,7 @@ jobs:
           sudo apt-get update
           sudo apt-get install git-lfs
           git lfs install
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           lfs: true
           persist-credentials: false

.pre-commit-config.yaml

@@ -87,7 +87,7 @@ repos:
   # TODO(Steven): Uncomment when ready to use
   ##### Static Analysis & Typing #####
   - repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.18.2
+    rev: v1.19.1
     hooks:
       - id: mypy
         args: [--config-file=pyproject.toml]

CODE_OF_CONDUCT.md

@@ -52,7 +52,7 @@ decisions when appropriate.
 
 This Code of Conduct applies within all community spaces, and also applies when
 an individual is officially representing the community in public spaces.
-Examples of representing our community include using an official email address,
+Examples of representing our community include using an official e-mail address,
 posting via an official social media account, or acting as an appointed
 representative at an online or offline event.
 
@@ -60,7 +60,7 @@ representative at an online or offline event.
 
 Instances of abusive, harassing, or otherwise unacceptable behavior may be
 reported to the community leaders responsible for enforcement at
-[feedback@huggingface.co](mailto:feedback@huggingface.co).
+feedback@huggingface.co.
 All complaints will be reviewed and investigated promptly and fairly.
 
 All community leaders are obligated to respect the privacy and security of the

CONTRIBUTING.md

@@ -1,323 +1,83 @@
-# How to contribute to 🤗 LeRobot?
+# How to contribute to 🤗 LeRobot
 
-Everyone is welcome to contribute, and we value everybody's contribution. Code
-is thus not the only way to help the community. Answering questions, helping
-others, reaching out and improving the documentations are immensely valuable to
-the community.
+Everyone is welcome to contribute, and we value everybody's contribution. Code is not the only way to help the community. Answering questions, helping others, reaching out, and improving the documentation are immensely valuable.
 
-It also helps us if you spread the word: reference the library from blog posts
-on the awesome projects it made possible, shout out on Twitter when it has
-helped you, or simply ⭐️ the repo to say "thank you".
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md).
 
-Whichever way you choose to contribute, please be mindful to respect our
-[code of conduct](https://github.com/huggingface/lerobot/blob/main/CODE_OF_CONDUCT.md).
+## Ways to Contribute
 
-## You can contribute in so many ways!
+You can contribute in many ways:
 
-Some of the ways you can contribute to 🤗 LeRobot:
+- **Fixing issues:** Resolve bugs or improve existing code.
+- **New features:** Develop new features.
+- **Extend:** Implement new models/policies, robots, or simulation environments and upload datasets to the Hugging Face Hub.
+- **Documentation:** Improve examples, guides, and docstrings.
+- **Feedback:** Submit tickets related to bugs or desired new features.
 
-- Fixing outstanding issues with the existing code.
-- Implementing new models, datasets or simulation environments.
-- Contributing to the examples or to the documentation.
-- Submitting issues related to bugs or desired new features.
+If you are unsure where to start, join our [Discord Channel](https://discord.gg/JkrYNdmw).
 
-Following the guides below, feel free to open issues and PRs and to coordinate your efforts with the community on our [Discord Channel](https://discord.gg/VjFz58wn3R). For specific inquiries, reach out to [Remi Cadene](mailto:remi.cadene@huggingface.co).
+## Development Setup
 
-If you are not sure how to contribute or want to know the next features we working on, look on this project page: [LeRobot TODO](https://github.com/orgs/huggingface/projects/46)
+To contribute code, you need to set up a development environment.
 
-## Submitting a new issue or feature request
+### 1. Fork and Clone
 
-Do your best to follow these guidelines when submitting an issue or a feature
-request. It will make it easier for us to come back to you quickly and with good
-feedback.
-
-### Did you find a bug?
-
-The 🤗 LeRobot library is robust and reliable thanks to the users who notify us of
-the problems they encounter. So thank you for reporting an issue.
-
-First, we would really appreciate it if you could **make sure the bug was not
-already reported** (use the search bar on Github under Issues).
-
-Did not find it? :( So we can act quickly on it, please follow these steps:
-
-- Include your **OS type and version**, the versions of **Python** and **PyTorch**.
-- A short, self-contained, code snippet that allows us to reproduce the bug in
-  less than 30s.
-- The full traceback if an exception is raised.
-- Attach any other additional information, like screenshots, you think may help.
-
-### Do you want a new feature?
-
-A good feature request addresses the following points:
-
-1. Motivation first:
-
-- Is it related to a problem/frustration with the library? If so, please explain
-  why. Providing a code snippet that demonstrates the problem is best.
-- Is it related to something you would need for a project? We'd love to hear
-  about it!
-- Is it something you worked on and think could benefit the community?
-  Awesome! Tell us what problem it solved for you.
-
-2. Write a _paragraph_ describing the feature.
-3. Provide a **code snippet** that demonstrates its future use.
-4. In case this is related to a paper, please attach a link.
-5. Attach any additional information (drawings, screenshots, etc.) you think may help.
-
-If your issue is well written we're already 80% of the way there by the time you
-post it.
-
-## Adding new policies, datasets or environments
-
-Look at our implementations for [datasets](./src/lerobot/datasets/), [policies](./src/lerobot/policies/),
-environments ([aloha](https://github.com/huggingface/gym-aloha),
-[pusht](https://github.com/huggingface/gym-pusht))
-and follow the same api design.
-
-When implementing a new dataset loadable with LeRobotDataset follow these steps:
-
-- Update `available_datasets_per_env` in `lerobot/__init__.py`
-
-When implementing a new environment (e.g. `gym_aloha`), follow these steps:
-
-- Update `available_tasks_per_env` and `available_datasets_per_env` in `lerobot/__init__.py`
-
-When implementing a new policy class (e.g. `DiffusionPolicy`) follow these steps:
-
-- Update `available_policies` and `available_policies_per_env`, in `lerobot/__init__.py`
-- Set the required `name` class attribute.
-- Update variables in `tests/test_available.py` by importing your new Policy class
-
-## Submitting a pull request (PR)
-
-Before writing code, we strongly advise you to search through the existing PRs or
-issues to make sure that nobody is already working on the same thing. If you are
-unsure, it is always a good idea to open an issue to get some feedback.
-
-You will need basic `git` proficiency to be able to contribute to
-🤗 LeRobot. `git` is not the easiest tool to use but it has the greatest
-manual. Type `git --help` in a shell and enjoy. If you prefer books, [Pro
-Git](https://git-scm.com/book/en/v2) is a very good reference.
-
-Follow these steps to start contributing:
-
-1. Fork the [repository](https://github.com/huggingface/lerobot) by
-   clicking on the 'Fork' button on the repository's page. This creates a copy of the code
-   under your GitHub user account.
-
-2. Clone your fork to your local disk, and add the base repository as a remote. The following command
-   assumes you have your public SSH key uploaded to GitHub. See the following guide for more
-   [information](https://docs.github.com/en/repositories/creating-and-managing-repositories/cloning-a-repository).
-
-   ```bash
-   git clone git@github.com:<your Github handle>/lerobot.git
-   cd lerobot
-   git remote add upstream https://github.com/huggingface/lerobot.git
-   ```
-
-3. Create a new branch to hold your development changes, and do this for every new PR you work on.
-
-   Start by synchronizing your `main` branch with the `upstream/main` branch (more details in the [GitHub Docs](https://docs.github.com/en/github/collaborating-with-issues-and-pull-requests/syncing-a-fork)):
-
-   ```bash
-   git checkout main
-   git fetch upstream
-   git rebase upstream/main
-   ```
-
-   Once your `main` branch is synchronized, create a new branch from it:
-
-   ```bash
-   git checkout -b a-descriptive-name-for-my-changes
-   ```
-
-   🚨 **Do not** work on the `main` branch.
-
-4. for development, we advise to use a tool like `poetry` or `uv` instead of just `pip` to easily track our dependencies.
-   Follow the instructions to [install poetry](https://python-poetry.org/docs/#installation) (use a version >=2.1.0) or to [install uv](https://docs.astral.sh/uv/getting-started/installation/#installation-methods) if you don't have one of them already.
-
-   Set up a development environment with conda:
-
-   ```bash
-   conda create -y -n lerobot-dev python=3.10 && conda activate lerobot-dev
-   ```
-
-   If you're using `uv`, it can manage python versions so you can instead do:
-
-   ```bash
-   uv venv --python 3.10 && source .venv/bin/activate
-   ```
-
-   To develop on 🤗 LeRobot, you will at least need to install the `dev` and `test` extras dependencies along with the core library:
-
-   using `poetry`
-
-   ```bash
-   poetry sync --extras "dev test"
-   ```
-
-   using `uv`
-
-   ```bash
-   uv sync --extra dev --extra test
-   ```
-
-   You can also install the project with all its dependencies (including environments):
-
-   using `poetry`
-
-   ```bash
-   poetry sync --all-extras
-   ```
-
-   using `uv`
-
-   ```bash
-   uv sync --all-extras
-   ```
-
-   > **Note:** If you don't install simulation environments with `--all-extras`, the tests that require them will be skipped when running the pytest suite locally. However, they _will_ be tested in the CI. In general, we advise you to install everything and test locally before pushing.
-
-   Whichever command you chose to install the project (e.g. `poetry sync --all-extras`), you should run it again when pulling code with an updated version of `pyproject.toml` and `poetry.lock` in order to synchronize your virtual environment with the new dependencies.
-
-   The equivalent of `pip install some-package`, would just be:
-
-   using `poetry`
-
-   ```bash
-   poetry add some-package
-   ```
-
-   using `uv`
-
-   ```bash
-   uv add some-package
-   ```
-
-   When making changes to the poetry sections of the `pyproject.toml`, you should run the following command to lock dependencies.
-   using `poetry`
-
-   ```bash
-   poetry lock
-   ```
-
-   using `uv`
-
-   ```bash
-   uv lock
-   ```
-
-5. Develop the features on your branch.
-
-   As you work on the features, you should make sure that the test suite
-   passes. You should run the tests impacted by your changes like this (see
-   below an explanation regarding the environment variable):
-
-   ```bash
-   pytest tests/<TEST_TO_RUN>.py
-   ```
-
-6. Follow our style.
-
-   `lerobot` relies on `ruff` to format its source code
-   consistently. Set up [`pre-commit`](https://pre-commit.com/) to run these checks
-   automatically as Git commit hooks.
-
-   Install `pre-commit` hooks:
-
-   ```bash
-   pre-commit install
-   ```
-
-   You can run these hooks whenever you need on staged files with:
-
-   ```bash
-   pre-commit
-   ```
-
-   Once you're happy with your changes, add changed files using `git add` and
-   make a commit with `git commit` to record your changes locally:
-
-   ```bash
-   git add modified_file.py
-   git commit
-   ```
-
-   Note, if you already committed some changes that have a wrong formatting, you can use:
-
-   ```bash
-   pre-commit run --all-files
-   ```
-
-   Please write [good commit messages](https://chris.beams.io/posts/git-commit/).
-
-   It is a good idea to sync your copy of the code with the original
-   repository regularly. This way you can quickly account for changes:
-
-   ```bash
-   git fetch upstream
-   git rebase upstream/main
-   ```
-
-   Push the changes to your account using:
-
-   ```bash
-   git push -u origin a-descriptive-name-for-my-changes
-   ```
-
-7. Once you are satisfied (**and the checklist below is happy too**), go to the
-   webpage of your fork on GitHub. Click on 'Pull request' to send your changes
-   to the project maintainers for review.
-
-8. It's ok if maintainers ask you for changes. It happens to core contributors
-   too! So everyone can see the changes in the Pull request, work in your local
-   branch and push the changes to your fork. They will automatically appear in
-   the pull request.
-
-### Checklist
-
-1. The title of your pull request should be a summary of its contribution;
-2. If your pull request addresses an issue, please mention the issue number in
-   the pull request description to make sure they are linked (and people
-   consulting the issue know you are working on it);
-3. To indicate a work in progress please prefix the title with `[WIP]`, or preferably mark
-   the PR as a draft PR. These are useful to avoid duplicated work, and to differentiate
-   it from PRs ready to be merged;
-4. Make sure existing tests pass;
-
-### Tests
-
-An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the [tests folder](https://github.com/huggingface/lerobot/tree/main/tests).
-
-Install [git lfs](https://git-lfs.com/) to retrieve test artifacts (if you don't have it already).
-
-On Mac:
+Fork the repository on GitHub, then clone your fork:
 
 ```bash
-brew install git-lfs
-git lfs install
+git clone https://github.com/<your-handle>/lerobot.git
+cd lerobot
+git remote add upstream https://github.com/huggingface/lerobot.git
 ```
 
-On Ubuntu:
+### 2. Environment Installation
+
+Please follow our [Installation Guide](./docs/source/installation.mdx) for the environment setup & installation from source.
+
+## Running Tests & Quality Checks
+
+### Code Style (Pre-commit)
+
+Install `pre-commit` hooks to run checks automatically before you commit:
 
 ```bash
-sudo apt-get install git-lfs
-git lfs install
+pre-commit install
 ```
 
-Pull artifacts if they're not in [tests/artifacts](tests/artifacts)
+To run checks manually on all files:
 
 ```bash
+pre-commit run --all-files
+```
+
+### Running Tests
+
+We use `pytest`. First, ensure you have test artifacts by installing **git-lfs**:
+
+```bash
+git lfs install
 git lfs pull
 ```
 
-We use `pytest` in order to run the tests. From the root of the
-repository, here's how to run tests with `pytest` for the library:
+Run the full suite (this may require extras installed):
 
 ```bash
-python -m pytest -sv ./tests
+pytest -sv ./tests
 ```
 
-You can specify a smaller set of tests in order to test only the feature
-you're working on.
+Or run a specific test file during development:
+
+```bash
+pytest -sv tests/test_specific_feature.py
+```
+
+## Submitting Issues & Pull Requests
+
+Use the templates for required fields and examples.
+
+- **Issues:** Follow the [ticket template](./.github/ISSUE_TEMPLATE/bug-report.yml).
+- **Pull requests:** Rebase on `upstream/main`, use a descriptive branch (don't work on `main`), run `pre-commit` and tests locally, and follow the [PR template](./.github/PULL_REQUEST_TEMPLATE.md).
+
+One member of the LeRobot team will then review your contribution.
+
+Thank you for contributing to LeRobot!

README.md

@@ -1,7 +1,5 @@
 <p align="center">
-  <img alt="LeRobot, Hugging Face Robotics Library" src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/lerobot-logo-thumbnail.png" width="100%">
-  <br/>
-  <br/>
+  <img alt="LeRobot, Hugging Face Robotics Library" src="./media/readme/lerobot-logo-thumbnail.png" width="100%">
 </p>
 
 <div align="center">
@@ -12,323 +10,130 @@
 [![Status](https://img.shields.io/pypi/status/lerobot)](https://pypi.org/project/lerobot/)
 [![Version](https://img.shields.io/pypi/v/lerobot)](https://pypi.org/project/lerobot/)
 [![Contributor Covenant](https://img.shields.io/badge/Contributor%20Covenant-v2.1-ff69b4.svg)](https://github.com/huggingface/lerobot/blob/main/CODE_OF_CONDUCT.md)
-[![Discord](https://dcbadge.vercel.app/api/server/C5P34WJ68S?style=flat)](https://discord.gg/s3KuuzsPFb)
-
-<!-- [![Coverage](https://codecov.io/gh/huggingface/lerobot/branch/main/graph/badge.svg?token=TODO)](https://codecov.io/gh/huggingface/lerobot) -->
 
 </div>
 
-<h2 align="center">
-    <p><a href="https://huggingface.co/docs/lerobot/hope_jr">
-        Build Your Own HopeJR Robot!</a></p>
-</h2>
+**LeRobot** aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier to entry so that everyone can contribute to and benefit from shared datasets and pretrained models.
 
-<div align="center">
-  <img
-    src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/hope_jr/hopejr.png"
-    alt="HopeJR robot"
-    title="HopeJR robot"
-    width="60%"
-  />
+🤗 A hardware-agnostic, Python-native interface that standardizes control across diverse platforms, from low-cost arms (SO-100) to humanoids.
 
-  <p><strong>Meet HopeJR – A humanoid robot arm and hand for dexterous manipulation!</strong></p>
-  <p>Control it with exoskeletons and gloves for precise hand movements.</p>
-  <p>Perfect for advanced manipulation tasks! 🤖</p>
+🤗 A standardized, scalable LeRobotDataset format (Parquet + MP4 or images) hosted on the Hugging Face Hub, enabling efficient storage, streaming and visualization of massive robotic datasets.
 
-  <p><a href="https://huggingface.co/docs/lerobot/hope_jr">
-      See the full HopeJR tutorial here.</a></p>
-</div>
+🤗 State-of-the-art policies that have been shown to transfer to the real-world ready for training and deployment.
 
-<br/>
+🤗 Comprehensive support for the open-source ecosystem to democratize physical AI.
 
-<h2 align="center">
-    <p><a href="https://huggingface.co/docs/lerobot/so101">
-        Build Your Own SO-101 Robot!</a></p>
-</h2>
+## Quick Start
 
-<div align="center">
-  <table>
-    <tr>
-      <td align="center"><img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/so101/so101.webp" alt="SO-101 follower arm" title="SO-101 follower arm" width="90%"/></td>
-      <td align="center"><img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/so101/so101-leader.webp" alt="SO-101 leader arm" title="SO-101 leader arm" width="90%"/></td>
-    </tr>
-  </table>
-
-  <p><strong>Meet the updated SO100, the SO-101 – Just €114 per arm!</strong></p>
-  <p>Train it in minutes with a few simple moves on your laptop.</p>
-  <p>Then sit back and watch your creation act autonomously! 🤯</p>
-
-  <p><a href="https://huggingface.co/docs/lerobot/so101">
-      See the full SO-101 tutorial here.</a></p>
-
-  <p>Want to take it to the next level? Make your SO-101 mobile by building LeKiwi!</p>
-  <p>Check out the <a href="https://huggingface.co/docs/lerobot/lekiwi">LeKiwi tutorial</a> and bring your robot to life on wheels.</p>
-
-  <img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/lekiwi/kiwi.webp" alt="LeKiwi mobile robot" title="LeKiwi mobile robot" width="50%">
-</div>
-
-<br/>
-
-<h3 align="center">
-    <p>LeRobot: State-of-the-art AI for real-world robotics</p>
-</h3>
-
----
-
-🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier to entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
-
-🤗 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real-world with a focus on imitation learning and reinforcement learning.
-
-🤗 LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulation environments to get started without assembling a robot. In the coming weeks, the plan is to add more and more support for real-world robotics on the most affordable and capable robots out there.
-
-🤗 LeRobot hosts pretrained models and datasets on this Hugging Face community page: [huggingface.co/lerobot](https://huggingface.co/lerobot)
-
-#### Examples of pretrained models on simulation environments
-
-<table>
-  <tr>
-    <td><img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/gym/aloha_act.gif" width="100%" alt="ACT policy on ALOHA env"/></td>
-    <td><img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/gym/simxarm_tdmpc.gif" width="100%" alt="TDMPC policy on SimXArm env"/></td>
-    <td><img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/gym/pusht_diffusion.gif" width="100%" alt="Diffusion policy on PushT env"/></td>
-  </tr>
-  <tr>
-    <td align="center">ACT policy on ALOHA env</td>
-    <td align="center">TDMPC policy on SimXArm env</td>
-    <td align="center">Diffusion policy on PushT env</td>
-  </tr>
-</table>
-
-## Installation
-
-LeRobot works with Python 3.10+ and PyTorch 2.2+.
-
-### Environment Setup
-
-Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniforge`](https://conda-forge.org/download/):
-
-```bash
-conda create -y -n lerobot python=3.10
-conda activate lerobot
-```
-
-When using `conda`, install `ffmpeg` in your environment:
-
-```bash
-conda install ffmpeg -c conda-forge
-```
-
-> **NOTE:** 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]_ 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`.
-
-### Install LeRobot 🤗
-
-#### From Source
-
-First, clone the repository and navigate into the directory:
-
-```bash
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-```
-
-Then, install the library in editable mode. This is useful if you plan to contribute to the code.
-
-```bash
-pip install -e .
-```
-
-> **NOTE:** If you encounter build errors, you may need to install additional dependencies (`cmake`, `build-essential`, and `ffmpeg libs`). On Linux, run:
-> `sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev`. For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
-
-For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
-
-- [aloha](https://github.com/huggingface/gym-aloha)
-- [xarm](https://github.com/huggingface/gym-xarm)
-- [pusht](https://github.com/huggingface/gym-pusht)
-
-For instance, to install 🤗 LeRobot with aloha and pusht, use:
-
-```bash
-pip install -e ".[aloha, pusht]"
-```
-
-### Installation from PyPI
-
-**Core Library:**
-Install the base package with:
+LeRobot can be installed directly from PyPI.
 
 ```bash
 pip install lerobot
+lerobot-info
 ```
 
-_This installs only the default dependencies._
+> [!IMPORTANT]
+> For detailed installation guide, please see the [Installation Documentation](https://huggingface.co/docs/lerobot/installation).
 
-**Extra Features:**
-To install additional functionality, use one of the following:
+## Robots & Control
+
+<div align="center">
+  <img src="./media/readme/robots_control_video.webp" width="640px" alt="Reachy 2 Demo">
+</div>
+
+LeRobot provides a unified `Robot` class interface that decouples control logic from hardware specifics. It supports a wide range of robots and teleoperation devices.
+
+```python
+from lerobot.robots.myrobot import MyRobot
+
+# Connect to a robot
+robot = MyRobot(config=...)
+robot.connect()
+
+# Read observation and send action
+obs = robot.get_observation()
+action = model.select_action(obs)
+robot.send_action(action)
+```
+
+**Supported Hardware:** SO100, LeKiwi, Koch, HopeJR, OMX, EarthRover, Reachy2, Gamepads, Keyboards, Phones, OpenARM, Unitree G1.
+
+While these devices are natively integrated into the LeRobot codebase, the library is designed to be extensible. You can easily implement the Robot interface to utilize LeRobot's data collection, training, and visualization tools for your own custom robot.
+
+For detailed hardware setup guides, see the [Hardware Documentation](https://huggingface.co/docs/lerobot/integrate_hardware).
+
+## LeRobot Dataset
+
+To solve the data fragmentation problem in robotics, we utilize the **LeRobotDataset** format.
+
+- **Structure:** Synchronized MP4 videos (or images) for vision and Parquet files for state/action data.
+- **HF Hub Integration:** Explore thousands of robotics datasets on the [Hugging Face Hub](https://huggingface.co/lerobot).
+- **Tools:** Seamlessly delete episodes, split by indices/fractions, add/remove features, and merge multiple datasets.
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Load a dataset from the Hub
+dataset = LeRobotDataset("lerobot/aloha_mobile_cabinet")
+
+# Access data (automatically handles video decoding)
+episode_index=0
+print(f"{dataset[episode_index]['action'].shape=}\n")
+```
+
+Learn more about it in the [LeRobotDataset Documentation](https://huggingface.co/docs/lerobot/lerobot-dataset-v3)
+
+## SoTA Models
+
+LeRobot implements state-of-the-art policies in pure PyTorch, covering Imitation Learning, Reinforcement Learning, and Vision-Language-Action (VLA) models, with more coming soon. It also provides you with the tools to instrument and inspect your training process.
+
+<p align="center">
+  <img alt="Gr00t Architecture" src="./media/readme/VLA_architecture.jpg" width="640px">
+</p>
+
+Training a policy is as simple as running a script configuration:
 
 ```bash
-pip install 'lerobot[all]'          # All available features
-pip install 'lerobot[aloha,pusht]'  # Specific features (Aloha & Pusht)
-pip install 'lerobot[feetech]'      # Feetech motor support
+lerobot-train \
+  --policy=act \
+  --dataset.repo_id=lerobot/aloha_mobile_cabinet
 ```
 
-_Replace `[...]` with your desired features._
+| Category                   | Models                                                                                                                                                                 |
+| -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md)                       |
+| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                            |
+| **VLAs Models**            | [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx) |
 
-**Available Tags:**
-For a full list of optional dependencies, see:
-https://pypi.org/project/lerobot/
+Similarly to the hardware, you can easily implement your own policy & leverage LeRobot's data collection, training, and visualization tools, and share your model to the HF Hub
 
-> [!NOTE]
-> For lerobot 0.4.0, if you want to install pi tags, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
->
-> This will be solved in the next patch release
+For detailed policy setup guides, see the [Policy Documentation](https://huggingface.co/docs/lerobot/bring_your_own_policies).
 
-### Weights & Biases
+## Inference & Evaluation
 
-To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
+Evaluate your policies in simulation or on real hardware using the unified evaluation script. LeRobot supports standard benchmarks like **LIBERO**, **MetaWorld** and more to come.
 
 ```bash
-wandb login
+# Evaluate a policy on the LIBERO benchmark
+lerobot-eval \
+  --policy.path=lerobot/pi0_libero_finetuned \
+  --env.type=libero \
+  --env.task=libero_object \
+  --eval.n_episodes=10
 ```
 
-(note: you will also need to enable WandB in the configuration. See below.)
+Learn how to implement your own simulation environment or benchmark and distribute it from the HF Hub by following the [EnvHub Documentation](https://huggingface.co/docs/lerobot/envhub)
 
-### Visualize datasets
+## Resources
 
-Check out [example 1](https://github.com/huggingface/lerobot/blob/main/examples/dataset/load_lerobot_dataset.py) that illustrates how to use our dataset class which automatically downloads data from the Hugging Face hub.
-
-You can also locally visualize episodes from a dataset on the hub by executing our script from the command line:
-
-```bash
-lerobot-dataset-viz \
-    --repo-id lerobot/pusht \
-    --episode-index 0
-```
-
-or from a dataset in a local folder with the `root` option and the `--mode local` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
-
-```bash
-lerobot-dataset-viz \
-    --repo-id lerobot/pusht \
-    --root ./my_local_data_dir \
-    --mode local \
-    --episode-index 0
-```
-
-It will open `rerun.io` and display the camera streams, robot states and actions, like this:
-
-https://github-production-user-asset-6210df.s3.amazonaws.com/4681518/328035972-fd46b787-b532-47e2-bb6f-fd536a55a7ed.mov?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20240505%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240505T172924Z&X-Amz-Expires=300&X-Amz-Signature=d680b26c532eeaf80740f08af3320d22ad0b8a4e4da1bcc4f33142c15b509eda&X-Amz-SignedHeaders=host&actor_id=24889239&key_id=0&repo_id=748713144
-
-Our script can also visualize datasets stored on a distant server. See `lerobot-dataset-viz --help` for more instructions.
-
-### The `LeRobotDataset` format
-
-A dataset in `LeRobotDataset` format is very simple to use. It can be loaded from a repository on the Hugging Face hub or a local folder simply with e.g. `dataset = LeRobotDataset("lerobot/aloha_static_coffee")` and can be indexed into like any Hugging Face and PyTorch dataset. For instance `dataset[0]` will retrieve a single temporal frame from the dataset containing observation(s) and an action as PyTorch tensors ready to be fed to a model.
-
-A specificity of `LeRobotDataset` is that, rather than retrieving a single frame by its index, we can retrieve several frames based on their temporal relationship with the indexed frame, by setting `delta_timestamps` to a list of relative times with respect to the indexed frame. For example, with `delta_timestamps = {"observation.image": [-1, -0.5, -0.2, 0]}` one can retrieve, for a given index, 4 frames: 3 "previous" frames 1 second, 0.5 seconds, and 0.2 seconds before the indexed frame, and the indexed frame itself (corresponding to the 0 entry). See example [1_load_lerobot_dataset.py](https://github.com/huggingface/lerobot/blob/main/examples/dataset/load_lerobot_dataset.py) for more details on `delta_timestamps`.
-
-Under the hood, the `LeRobotDataset` format makes use of several ways to serialize data which can be useful to understand if you plan to work more closely with this format. We tried to make a flexible yet simple dataset format that would cover most type of features and specificities present in reinforcement learning and robotics, in simulation and in real-world, with a focus on cameras and robot states but easily extended to other types of sensory inputs as long as they can be represented by a tensor.
-
-Here are the important details and internal structure organization of a typical `LeRobotDataset` instantiated with `dataset = LeRobotDataset("lerobot/aloha_static_coffee")`. The exact features will change from dataset to dataset but not the main aspects:
-
-```
-dataset attributes:
-  ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
-  │  ├ observation.images.cam_high (VideoFrame):
-  │  │   VideoFrame = {'path': path to a mp4 video, 'timestamp' (float32): timestamp in the video}
-  │  ├ observation.state (list of float32): position of an arm joints (for instance)
-  │  ... (more observations)
-  │  ├ action (list of float32): goal position of an arm joints (for instance)
-  │  ├ episode_index (int64): index of the episode for this sample
-  │  ├ frame_index (int64): index of the frame for this sample in the episode ; starts at 0 for each episode
-  │  ├ timestamp (float32): timestamp in the episode
-  │  ├ next.done (bool): indicates the end of an episode ; True for the last frame in each episode
-  │  └ index (int64): general index in the whole dataset
-  ├ meta: a LeRobotDatasetMetadata object containing:
-  │  ├ info: a dictionary of metadata on the dataset
-  │  │  ├ codebase_version (str): this is to keep track of the codebase version the dataset was created with
-  │  │  ├ fps (int): frame per second the dataset is recorded/synchronized to
-  │  │  ├ features (dict): all features contained in the dataset with their shapes and types
-  │  │  ├ total_episodes (int): total number of episodes in the dataset
-  │  │  ├ total_frames (int): total number of frames in the dataset
-  │  │  ├ robot_type (str): robot type used for recording
-  │  │  ├ data_path (str): formattable string for the parquet files
-  │  │  └ video_path (str): formattable string for the video files (if using videos)
-  │  ├ episodes: a DataFrame containing episode metadata with columns:
-  │  │  ├ episode_index (int): index of the episode
-  │  │  ├ tasks (list): list of tasks for this episode
-  │  │  ├ length (int): number of frames in this episode
-  │  │  ├ dataset_from_index (int): start index of this episode in the dataset
-  │  │  └ dataset_to_index (int): end index of this episode in the dataset
-  │  ├ stats: a dictionary of statistics (max, mean, min, std) for each feature in the dataset, for instance
-  │  │  ├ observation.images.front_cam: {'max': tensor with same number of dimensions (e.g. `(c, 1, 1)` for images, `(c,)` for states), etc.}
-  │  │  └ ...
-  │  └ tasks: a DataFrame containing task information with task names as index and task_index as values
-  ├ root (Path): local directory where the dataset is stored
-  ├ image_transforms (Callable): optional image transformations to apply to visual modalities
-  └ delta_timestamps (dict): optional delta timestamps for temporal queries
-```
-
-A `LeRobotDataset` is serialised using several widespread file formats for each of its parts, namely:
-
-- hf_dataset stored using Hugging Face datasets library serialization to parquet
-- videos are stored in mp4 format to save space
-- metadata are stored in plain json/jsonl files
-
-Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can specify its location with the `root` argument if it's not in the default `~/.cache/huggingface/lerobot` location.
-
-#### Reproduce state-of-the-art (SOTA)
-
-We provide some pretrained policies on our [hub page](https://huggingface.co/lerobot) that can achieve state-of-the-art performances.
-You can reproduce their training by loading the config from their run. Simply running:
-
-```bash
-lerobot-train --config_path=lerobot/diffusion_pusht
-```
-
-reproduces SOTA results for Diffusion Policy on the PushT task.
-
-## Contribute
-
-If you would like to contribute to 🤗 LeRobot, please check out our [contribution guide](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md).
-
-### Add a pretrained policy
-
-Once you have trained a policy you may upload it to the Hugging Face hub using a hub id that looks like `${hf_user}/${repo_name}` (e.g. [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)).
-
-You first need to find the checkpoint folder located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). Within that there is a `pretrained_model` directory which should contain:
-
-- `config.json`: A serialized version of the policy configuration (following the policy's dataclass config).
-- `model.safetensors`: A set of `torch.nn.Module` parameters, saved in [Hugging Face Safetensors](https://huggingface.co/docs/safetensors/index) format.
-- `train_config.json`: A consolidated configuration containing all parameters used for training. The policy configuration should match `config.json` exactly. This is useful for anyone who wants to evaluate your policy or for reproducibility.
-
-To upload these to the hub, run the following:
-
-```bash
-huggingface-cli upload ${hf_user}/${repo_name} path/to/pretrained_model
-```
-
-See [lerobot_eval.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_eval.py) for an example of how other people may use your policy.
-
-### Acknowledgment
-
-- The LeRobot team 🤗 for building SmolVLA [Paper](https://arxiv.org/abs/2506.01844), [Blog](https://huggingface.co/blog/smolvla).
-- Thanks to Tony Zhao, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
-- Thanks to Cheng Chi, Zhenjia Xu and colleagues for open sourcing Diffusion policy, Pusht environment and datasets, as well as UMI datasets. Ours are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu) and [UMI Gripper](https://umi-gripper.github.io).
-- Thanks to Nicklas Hansen, Yunhai Feng and colleagues for open sourcing TDMPC policy, Simxarm environments and datasets. Ours are adapted from [TDMPC](https://github.com/nicklashansen/tdmpc) and [FOWM](https://www.yunhaifeng.com/FOWM).
-- Thanks to Antonio Loquercio and Ashish Kumar for their early support.
-- Thanks to [Seungjae (Jay) Lee](https://sjlee.cc/), [Mahi Shafiullah](https://mahis.life/) and colleagues for open sourcing [VQ-BeT](https://sjlee.cc/vq-bet/) policy and helping us adapt the codebase to our repository. The policy is adapted from [VQ-BeT repo](https://github.com/jayLEE0301/vq_bet_official).
+- **[Documentation](https://huggingface.co/docs/lerobot/index):** The complete guide to tutorials & API.
+- **[Discord](https://discord.gg/3gxM6Avj):** Join the `LeRobot` server to discuss with the community.
+- **[X](https://x.com/LeRobotHF):** Follow us on X to stay up-to-date with the latest developments.
+- **[Robot Learning Tutorial](https://huggingface.co/spaces/lerobot/robot-learning-tutorial):** A free, hands-on course to learn robot learning using LeRobot.
 
 ## Citation
 
-If you want, you can cite this work with:
+If you use LeRobot in your research, please cite:
 
 ```bibtex
 @misc{cadene2024lerobot,
@@ -339,6 +144,14 @@ If you want, you can cite this work with:
 }
 ```
 
-## Star History
+## Contribute
 
-[![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)
+We welcome contributions from everyone in the community! To get started, please read our [CONTRIBUTING.md](./CONTRIBUTING.md) guide. Whether you're adding a new feature, improving documentation, or fixing a bug, your help and feedback are invaluable. We're incredibly excited about the future of open-source robotics and can't wait to work with you on what's next—thank you for your support!
+
+<p align="center">
+  <img alt="SO101 Video" src="./media/readme/so100_video.webp" width="640px">
+</p>
+
+<div align="center">
+<sub>Built by the <a href="https://huggingface.co/lerobot">LeRobot</a> team at <a href="https://huggingface.co">Hugging Face</a> with ❤️</sub>
+</div>

benchmarks/video/benchmark.py

@@ -1,94 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-import threading
-import time
-from contextlib import ContextDecorator
-
-
-class TimeBenchmark(ContextDecorator):
-    """
-    Measures execution time using a context manager or decorator.
-
-    This class supports both context manager and decorator usage, and is thread-safe for multithreaded
-    environments.
-
-    Args:
-        print: If True, prints the elapsed time upon exiting the context or completing the function. Defaults
-        to False.
-
-    Examples:
-
-        Using as a context manager:
-
-        >>> benchmark = TimeBenchmark()
-        >>> with benchmark:
-        ...     time.sleep(1)
-        >>> print(f"Block took {benchmark.result:.4f} seconds")
-        Block took approximately 1.0000 seconds
-
-        Using with multithreading:
-
-        ```python
-        import threading
-
-        benchmark = TimeBenchmark()
-
-
-        def context_manager_example():
-            with benchmark:
-                time.sleep(0.01)
-            print(f"Block took {benchmark.result_ms:.2f} milliseconds")
-
-
-        threads = []
-        for _ in range(3):
-            t1 = threading.Thread(target=context_manager_example)
-            threads.append(t1)
-
-        for t in threads:
-            t.start()
-
-        for t in threads:
-            t.join()
-        ```
-        Expected output:
-        Block took approximately 10.00 milliseconds
-        Block took approximately 10.00 milliseconds
-        Block took approximately 10.00 milliseconds
-    """
-
-    def __init__(self, print=False):
-        self.local = threading.local()
-        self.print_time = print
-
-    def __enter__(self):
-        self.local.start_time = time.perf_counter()
-        return self
-
-    def __exit__(self, *exc):
-        self.local.end_time = time.perf_counter()
-        self.local.elapsed_time = self.local.end_time - self.local.start_time
-        if self.print_time:
-            print(f"Elapsed time: {self.local.elapsed_time:.4f} seconds")
-        return False
-
-    @property
-    def result(self):
-        return getattr(self.local, "elapsed_time", None)
-
-    @property
-    def result_ms(self):
-        return self.result * 1e3

benchmarks/video/capture_camera_feed.py

@@ -1,102 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Capture video feed from a camera as raw images."""
-
-import argparse
-import datetime as dt
-import os
-import time
-from pathlib import Path
-
-import cv2
-import rerun as rr
-
-# see https://rerun.io/docs/howto/visualization/limit-ram
-RERUN_MEMORY_LIMIT = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "5%")
-
-
-def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int, duration: int):
-    rr.init("lerobot_capture_camera_feed")
-    rr.spawn(memory_limit=RERUN_MEMORY_LIMIT)
-
-    now = dt.datetime.now()
-    capture_dir = output_dir / f"{now:%Y-%m-%d}" / f"{now:%H-%M-%S}"
-    if not capture_dir.exists():
-        capture_dir.mkdir(parents=True, exist_ok=True)
-
-    # Opens the default webcam
-    cap = cv2.VideoCapture(0)
-    if not cap.isOpened():
-        print("Error: Could not open video stream.")
-        return
-
-    cap.set(cv2.CAP_PROP_FPS, fps)
-    cap.set(cv2.CAP_PROP_FRAME_WIDTH, width)
-    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
-
-    frame_index = 0
-    start_time = time.time()
-    while time.time() - start_time < duration:
-        ret, frame = cap.read()
-
-        if not ret:
-            print("Error: Could not read frame.")
-            break
-        rr.log("video/stream", rr.Image(frame), static=True)
-        cv2.imwrite(str(capture_dir / f"frame_{frame_index:06d}.png"), frame)
-        frame_index += 1
-
-    # Release the capture
-    cap.release()
-
-    # TODO(Steven): Add a graceful shutdown via a close() method for the Viewer context, though not currently supported in the Rerun API.
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--output-dir",
-        type=Path,
-        default=Path("outputs/cam_capture/"),
-        help="Directory where the capture images are written. A subfolder named with the current date & time will be created inside it for each capture.",
-    )
-    parser.add_argument(
-        "--fps",
-        type=int,
-        default=30,
-        help="Frames Per Second of the capture.",
-    )
-    parser.add_argument(
-        "--width",
-        type=int,
-        default=1280,
-        help="Width of the captured images.",
-    )
-    parser.add_argument(
-        "--height",
-        type=int,
-        default=720,
-        help="Height of the captured images.",
-    )
-    parser.add_argument(
-        "--duration",
-        type=int,
-        default=20,
-        help="Duration in seconds for which the video stream should be captured.",
-    )
-    args = parser.parse_args()
-    display_and_save_video_stream(**vars(args))

benchmarks/video/run_video_benchmark.py

@@ -21,11 +21,13 @@ See the provided README.md or run `python benchmark/video/run_video_benchmark.py
 
 import argparse
 import datetime as dt
+import itertools
 import random
 import shutil
 from collections import OrderedDict
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path
+from threading import Lock
 
 import einops
 import numpy as np
@@ -35,13 +37,13 @@ import torch
 from skimage.metrics import mean_squared_error, peak_signal_noise_ratio, structural_similarity
 from tqdm import tqdm
 
-from benchmarks.video.benchmark import TimeBenchmark
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.video_utils import (
-    decode_video_frames_torchvision,
+    decode_video_frames,
     encode_video_frames,
 )
 from lerobot.utils.constants import OBS_IMAGE
+from lerobot.utils.utils import TimerManager
 
 BASE_ENCODING = OrderedDict(
     [
@@ -86,7 +88,7 @@ def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> t
     frames = []
     for ts in timestamps:
         idx = int(ts * fps)
-        frame = PIL.Image.open(imgs_dir / f"frame_{idx:06d}.png")
+        frame = PIL.Image.open(imgs_dir / f"frame-{idx:06d}.png")
         frame = torch.from_numpy(np.array(frame))
         frame = frame.type(torch.float32) / 255
         frame = einops.rearrange(frame, "h w c -> c h w")
@@ -97,21 +99,21 @@ def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> t
 def save_decoded_frames(
     imgs_dir: Path, save_dir: Path, frames: torch.Tensor, timestamps: list[float], fps: int
 ) -> None:
-    if save_dir.exists() and len(list(save_dir.glob("frame_*.png"))) == len(timestamps):
+    if save_dir.exists() and len(list(save_dir.glob("frame-*.png"))) == len(timestamps):
         return
 
     save_dir.mkdir(parents=True, exist_ok=True)
     for i, ts in enumerate(timestamps):
         idx = int(ts * fps)
         frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
-        PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame_{idx:06d}_decoded.png")
-        shutil.copyfile(imgs_dir / f"frame_{idx:06d}.png", save_dir / f"frame_{idx:06d}_original.png")
+        PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame-{idx:06d}_decoded.png")
+        shutil.copyfile(imgs_dir / f"frame-{idx:06d}.png", save_dir / f"frame-{idx:06d}_original.png")
 
 
 def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
     episode_index = 0
     ep_num_images = dataset.meta.episodes["length"][episode_index]
-    if imgs_dir.exists() and len(list(imgs_dir.glob("frame_*.png"))) == ep_num_images:
+    if imgs_dir.exists() and len(list(imgs_dir.glob("frame-*.png"))) == ep_num_images:
         return
 
     imgs_dir.mkdir(parents=True, exist_ok=True)
@@ -125,7 +127,7 @@ def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
         tqdm(imgs_dataset, desc=f"saving {dataset.repo_id} first episode images", leave=False)
     ):
         img = item[img_keys[0]]
-        img.save(str(imgs_dir / f"frame_{i:06d}.png"), quality=100)
+        img.save(str(imgs_dir / f"frame-{i:06d}.png"), quality=100)
 
         if i >= ep_num_images - 1:
             break
@@ -149,18 +151,6 @@ def sample_timestamps(timestamps_mode: str, ep_num_images: int, fps: int) -> lis
     return [idx / fps for idx in frame_indexes]
 
 
-def decode_video_frames(
-    video_path: str,
-    timestamps: list[float],
-    tolerance_s: float,
-    backend: str,
-) -> torch.Tensor:
-    if backend in ["pyav", "video_reader"]:
-        return decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
-    else:
-        raise NotImplementedError(backend)
-
-
 def benchmark_decoding(
     imgs_dir: Path,
     video_path: Path,
@@ -172,8 +162,8 @@ def benchmark_decoding(
     num_workers: int = 4,
     save_frames: bool = False,
 ) -> dict:
-    def process_sample(sample: int):
-        time_benchmark = TimeBenchmark()
+    def process_sample(sample: int, lock: Lock):
+        time_benchmark = TimerManager(log=False)
         timestamps = sample_timestamps(timestamps_mode, ep_num_images, fps)
         num_frames = len(timestamps)
         result = {
@@ -182,13 +172,13 @@ def benchmark_decoding(
             "mse_values": [],
         }
 
-        with time_benchmark:
+        with time_benchmark, lock:
             frames = decode_video_frames(video_path, timestamps=timestamps, tolerance_s=5e-1, backend=backend)
-        result["load_time_video_ms"] = time_benchmark.result_ms / num_frames
+        result["load_time_video_ms"] = (time_benchmark.last * 1000) / num_frames
 
         with time_benchmark:
             original_frames = load_original_frames(imgs_dir, timestamps, fps)
-        result["load_time_images_ms"] = time_benchmark.result_ms / num_frames
+        result["load_time_images_ms"] = (time_benchmark.last * 1000) / num_frames
 
         frames_np, original_frames_np = frames.numpy(), original_frames.numpy()
         for i in range(num_frames):
@@ -215,8 +205,10 @@ def benchmark_decoding(
     # A sample is a single set of decoded frames specified by timestamps_mode (e.g. a single frame, 2 frames, etc.).
     # For each sample, we record metrics (loading time and quality metrics) which are then averaged over all samples.
     # As these samples are independent, we run them in parallel threads to speed up the benchmark.
+    # Use a single shared lock for all worker threads
+    shared_lock = Lock()
     with ThreadPoolExecutor(max_workers=num_workers) as executor:
-        futures = [executor.submit(process_sample, i) for i in range(num_samples)]
+        futures = [executor.submit(process_sample, i, shared_lock) for i in range(num_samples)]
         for future in tqdm(as_completed(futures), total=num_samples, desc="samples", leave=False):
             result = future.result()
             load_times_video_ms.append(result["load_time_video_ms"])
@@ -358,24 +350,27 @@ def main(
                 imgs_dir = output_dir / "images" / dataset.repo_id.replace("/", "_")
                 # We only use the first episode
                 save_first_episode(imgs_dir, dataset)
-                for key, values in tqdm(encoding_benchmarks.items(), desc="encodings (g, crf)", leave=False):
-                    for value in tqdm(values, desc=f"encodings ({key})", leave=False):
-                        encoding_cfg = BASE_ENCODING.copy()
-                        encoding_cfg["vcodec"] = video_codec
-                        encoding_cfg["pix_fmt"] = pixel_format
+                for duet in [
+                    dict(zip(encoding_benchmarks.keys(), unique_combination, strict=False))
+                    for unique_combination in itertools.product(*encoding_benchmarks.values())
+                ]:
+                    encoding_cfg = BASE_ENCODING.copy()
+                    encoding_cfg["vcodec"] = video_codec
+                    encoding_cfg["pix_fmt"] = pixel_format
+                    for key, value in duet.items():
                         encoding_cfg[key] = value
-                        args_path = Path("_".join(str(value) for value in encoding_cfg.values()))
-                        video_path = output_dir / "videos" / args_path / f"{repo_id.replace('/', '_')}.mp4"
-                        benchmark_table += benchmark_encoding_decoding(
-                            dataset,
-                            video_path,
-                            imgs_dir,
-                            encoding_cfg,
-                            decoding_benchmarks,
-                            num_samples,
-                            num_workers,
-                            save_frames,
-                        )
+                    args_path = Path("_".join(str(value) for value in encoding_cfg.values()))
+                    video_path = output_dir / "videos" / args_path / f"{repo_id.replace('/', '_')}.mp4"
+                    benchmark_table += benchmark_encoding_decoding(
+                        dataset,
+                        video_path,
+                        imgs_dir,
+                        encoding_cfg,
+                        decoding_benchmarks,
+                        num_samples,
+                        num_workers,
+                        save_frames,
+                    )
 
             # Save intermediate results
             benchmark_df = pd.DataFrame(benchmark_table, columns=headers)
@@ -409,9 +404,9 @@ if __name__ == "__main__":
         nargs="*",
         default=[
             "lerobot/pusht_image",
-            "aliberts/aloha_mobile_shrimp_image",
-            "aliberts/paris_street",
-            "aliberts/kitchen",
+            "lerobot/aloha_mobile_shrimp_image",
+            "lerobot/paris_street",
+            "lerobot/kitchen",
         ],
         help="Datasets repo-ids to test against. First episodes only are used. Must be images.",
     )
@@ -419,7 +414,7 @@ if __name__ == "__main__":
         "--vcodec",
         type=str,
         nargs="*",
-        default=["libx264", "hevc", "libsvtav1"],
+        default=["h264", "hevc", "libsvtav1"],
         help="Video codecs to be tested",
     )
     parser.add_argument(
@@ -468,7 +463,7 @@ if __name__ == "__main__":
         "--backends",
         type=str,
         nargs="*",
-        default=["pyav", "video_reader"],
+        default=["torchcodec", "pyav"],
         help="Torchvision decoding backend to be tested.",
     )
     parser.add_argument(

docs/source/_toctree.yml

@@ -9,6 +9,8 @@
     title: Imitation Learning for Robots
   - local: cameras
     title: Cameras
+  - local: bring_your_own_policies
+    title: Bring Your Own Policies
   - local: integrate_hardware
     title: Bring Your Own Hardware
   - local: hilserl
@@ -37,7 +39,15 @@
     title: π₀.₅ (Pi05)
   - local: groot
     title: NVIDIA GR00T N1.5
+  - local: xvla
+    title: X-VLA
+  - local: walloss
+    title: WALL-OSS
   title: "Policies"
+- sections:
+  - local: sarm
+    title: SARM
+  title: "Reward Models"
 - sections:
   - local: async
     title: Use Async Inference
@@ -79,11 +89,19 @@
     title: Hope Jr
   - local: reachy2
     title: Reachy 2
+  - local: unitree_g1
+    title: Unitree G1
+  - local: earthrover_mini_plus
+    title: Earth Rover Mini
   title: "Robots"
 - sections:
   - local: phone_teleop
     title: Phone
   title: "Teleoperators"
+- sections:
+  - local: torch_accelerators
+    title: PyTorch accelerators
+  title: "Supported Hardware"
 - sections:
   - local: notebooks
     title: Notebooks

docs/source/async.mdx

@@ -278,7 +278,7 @@ We found the default values of `actions_per_chunk` and `chunk_size_threshold` to
 2. **Adjust your `fps` based on inference latency.** While the server generates a new action chunk, the client is not idle and is stepping through its current action queue. If the two processes happen at fundamentally different speeds, the client might end up with an empty queue. As such, you should reduce your fps if you consistently run out of actions in queue.
 3. **Adjust `chunk_size_threshold`**.
    - Values closer to `0.0` result in almost sequential behavior. Values closer to `1.0` → send observation every step (more bandwidth, relies on good world-model).
-   - We found values around 0.5-0.6 to work well. If you want to tweak this, spin up a `RobotClient` setting the `--debug-visualize-queue-size` to `True`. This will plot the action queue size evolution at runtime, and you can use it to find the value of `chunk_size_threshold` that works best for your setup.
+   - We found values around 0.5-0.6 to work well. If you want to tweak this, spin up a `RobotClient` setting the `--debug_visualize_queue_size` to `True`. This will plot the action queue size evolution at runtime, and you can use it to find the value of `chunk_size_threshold` that works best for your setup.
 
 <p align="center">
   <img
@@ -289,7 +289,7 @@ We found the default values of `actions_per_chunk` and `chunk_size_threshold` to
 <p align="center">
   <i>
     The action queue size is plotted at runtime when the
-    `--debug-visualize-queue-size` flag is passed, for various levels of
+    `--debug_visualize_queue_size` flag is passed, for various levels of
     `chunk_size_threshold` (`g` in the SmolVLA paper).
   </i>
 </p>

docs/source/bring_your_own_policies.mdx

@@ -0,0 +1,175 @@
+# Bring Your Own Policies
+
+This tutorial explains how to integrate your own custom policy implementations into the LeRobot ecosystem, allowing you to leverage all LeRobot tools for training, evaluation, and deployment while using your own algorithms.
+
+## Step 1: Create a Policy Package
+
+Your custom policy should be organized as an installable Python package following LeRobot's plugin conventions.
+
+### Package Structure
+
+Create a package with the prefix `lerobot_policy_` (IMPORTANT!) followed by your policy name:
+
+```bash
+lerobot_policy_my_custom_policy/
+├── pyproject.toml
+└── src/
+    └── lerobot_policy_my_custom_policy/
+        ├── __init__.py
+        ├── configuration_my_custom_policy.py
+        ├── modeling_my_custom_policy.py
+        └── processor_my_custom_policy.py
+```
+
+### Package Configuration
+
+Set up your `pyproject.toml`:
+
+```toml
+[project]
+name = "lerobot_policy_my_custom_policy"
+version = "0.1.0"
+dependencies = [
+    # your policy-specific dependencies
+]
+requires-python = ">= 3.11"
+
+[build-system]
+build-backend = # your-build-backend
+requires = # your-build-system
+```
+
+## Step 2: Define the Policy Configuration
+
+Create a configuration class that inherits from `PreTrainedConfig` and registers your policy type:
+
+```python
+# configuration_my_custom_policy.py
+from dataclasses import dataclass, field
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+
+@PreTrainedConfig.register_subclass("my_custom_policy")
+@dataclass
+class MyCustomPolicyConfig(PreTrainedConfig):
+    """Configuration class for MyCustomPolicy.
+
+    Args:
+        n_obs_steps: Number of observation steps to use as input
+        horizon: Action prediction horizon
+        n_action_steps: Number of action steps to execute
+        hidden_dim: Hidden dimension for the policy network
+        # Add your policy-specific parameters here
+    """
+    # ...PreTrainedConfig fields...
+    pass
+
+    def __post_init__(self):
+        super().__post_init__()
+        # Add any validation logic here
+
+    def validate_features(self) -> None:
+        """Validate input/output feature compatibility."""
+        # Implement validation logic for your policy's requirements
+        pass
+```
+
+## Step 3: Implement the Policy Class
+
+Create your policy implementation by inheriting from LeRobot's base `PreTrainedPolicy` class:
+
+```python
+# modeling_my_custom_policy.py
+import torch
+import torch.nn as nn
+from typing import Dict, Any
+
+from lerobot.policies.pretrained import PreTrainedPolicy
+from .configuration_my_custom_policy import MyCustomPolicyConfig
+
+class MyCustomPolicy(PreTrainedPolicy):
+    config_class = MyCustomPolicyConfig
+    name = "my_custom_policy"
+
+    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: Dict[str, Any] = None):
+        super().__init__(config, dataset_stats)
+        ...
+```
+
+## Step 4: Add Data Processors
+
+Create processor functions:
+
+```python
+# processor_my_custom_policy.py
+from typing import Dict, Any
+import torch
+
+
+def make_my_custom_policy_pre_post_processors(
+    config,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Create preprocessing and postprocessing functions for your policy."""
+    pass  # Define your preprocessing and postprocessing logic here
+
+```
+
+## Step 5: Package Initialization
+
+Expose your classes in the package's `__init__.py`:
+
+```python
+# __init__.py
+"""Custom policy package for LeRobot."""
+
+try:
+    import lerobot  # noqa: F401
+except ImportError:
+    raise ImportError(
+        "lerobot is not installed. Please install lerobot to use this policy package."
+    )
+
+from .configuration_my_custom_policy import MyCustomPolicyConfig
+from .modeling_my_custom_policy import MyCustomPolicy
+from .processor_my_custom_policy import make_my_custom_policy_pre_post_processors
+
+__all__ = [
+    "MyCustomPolicyConfig",
+    "MyCustomPolicy",
+    "make_my_custom_policy_pre_post_processors",
+]
+```
+
+## Step 6: Installation and Usage
+
+### Install Your Policy Package
+
+```bash
+cd lerobot_policy_my_custom_policy
+pip install -e .
+
+# Or install from PyPI if published
+pip install lerobot_policy_my_custom_policy
+```
+
+### Use Your Policy
+
+Once installed, your policy automatically integrates with LeRobot's training and evaluation tools:
+
+```bash
+lerobot-train \
+    --policy.type my_custom_policy \
+    --env.type pusht \
+    --steps 200000
+```
+
+## Examples and Community Contributions
+
+Check out these example policy implementations:
+
+- [DiTFlow Policy](https://github.com/danielsanjosepro/lerobot_policy_ditflow) - Diffusion Transformer policy with flow-matching objective. Try it out in this example: [DiTFlow Example](https://github.com/danielsanjosepro/test_lerobot_policy_ditflow)
+
+Share your policy implementations with the community! 🤗

docs/source/earthrover_mini_plus.mdx

@@ -0,0 +1,206 @@
+# EarthRover Mini Plus
+
+The EarthRover Mini Plus is a fully open source mobile robot that connects through the cloud using the Frodobots SDK. This lets you control the robot and record datasets for training AI models.
+
+## What You Need
+
+### Hardware
+
+- EarthRover Mini robot
+- Computer with Python 3.10 or newer
+- Internet connection
+
+### Setting Up the Frodobots SDK
+
+The robot needs the [Frodobots SDK](https://github.com/Frodobots/earth-rovers-sdk) running on your computer. Here's how:
+
+1. Download and install the SDK:
+
+```bash
+git clone https://github.com/Frodobots/earth-rovers-sdk.git
+cd earth-rovers-sdk
+pip install -r requirements.txt
+```
+
+2. Start the SDK:
+
+```bash
+hypercorn main:app --reload
+```
+
+3. Open your web browser and go to `http://localhost:8000`, then click "Join"
+
+The SDK gives you:
+
+- Live video from front and rear cameras
+
+> [!IMPORTANT]
+> The SDK must be running before you can use the robot.
+
+## Install LeRobot
+
+Follow our [Installation Guide](./installation) to install LeRobot.
+
+In addition to the base installation, install the EarthRover Mini dependencies:
+
+```bash
+pip install -e .
+```
+
+## How It Works
+
+The robot uses the internet to communicate:
+
+- **Movement commands**: Sent through the SDK
+- **Camera video**: Received from the SDK
+- **Robot info**: Battery, location, speed from the SDK
+
+You don't need to plug anything in - it all works through the SDK.
+
+## Calibration
+
+No calibration needed! The robot is ready to use as soon as the SDK is running.
+
+## Controlling the Robot
+
+You control the robot using your keyboard - just like playing a video game with WASD keys.
+
+### Keyboard Controls
+
+| Key | Action                           |
+| --- | -------------------------------- |
+| W   | Move forward                     |
+| S   | Move backward                    |
+| A   | Turn left (with forward motion)  |
+| D   | Turn right (with forward motion) |
+| Q   | Rotate left in place             |
+| E   | Rotate right in place            |
+| X   | Stop all movement                |
+| +/= | Increase speed                   |
+| -   | Decrease speed                   |
+| ESC | Disconnect                       |
+
+### Speed Settings
+
+You can adjust how fast the robot moves:
+
+- **Forward/backward speed**: Default is full speed (1.0)
+- **Turning speed**: Default is full speed (1.0)
+- **Speed changes**: Use +/- keys to adjust by 0.1 each time
+
+### Try It Out
+
+Test driving the robot before recording data:
+
+```python
+from lerobot.robots.earthrover_mini_plus import EarthRoverMiniPlus, EarthRoverMiniPlusConfig
+from lerobot.teleoperators.keyboard import KeyboardRoverTeleop, KeyboardRoverTeleopConfig
+
+# Initialize robot
+robot_config = EarthRoverMiniPlusConfig()
+robot = EarthRoverMiniPlus(robot_config)
+
+# Initialize teleoperator
+teleop_config = KeyboardRoverTeleopConfig(
+    linear_speed=1.0,
+    angular_speed=1.0,
+    speed_increment=0.1
+)
+teleop = KeyboardRoverTeleop(teleop_config)
+
+# Connect
+robot.connect()
+teleop.connect()
+
+# Teleoperate (use keyboard controls)
+try:
+    while True:
+        action = teleop.get_action()
+        robot.send_action(action)
+except KeyboardInterrupt:
+    pass
+finally:
+    robot.disconnect()
+    teleop.disconnect()
+```
+
+> [!TIP]
+> If you're using a Mac, you might need to give Terminal permission to access your keyboard for teleoperation. Go to System Preferences > Security & Privacy > Input Monitoring and check the box for Terminal.
+
+## Recording Data
+
+Once you can drive the robot well, you can start recording data to train AI models. The system records:
+
+- **What you do**: How you move the robot (forward, backward, turning)
+- **What the robot sees**:
+  - Videos from both cameras
+  - Robot speed and direction
+  - Battery level and location
+  - GPS position and signal
+  - Other sensor data
+- **When it happened**: Timestamps for everything
+
+### Setting Up Hugging Face
+
+We use Hugging Face to store your data online. First, log in with your token from [Hugging Face settings](https://huggingface.co/settings/tokens):
+
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face username:
+
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+### Start Recording
+
+Use the standard recording command:
+
+```bash
+python src/lerobot/scripts/lerobot_record.py \
+    --robot.type=earthrover_mini_plus \
+    --teleop.type=keyboard_rover \
+    --dataset.repo_id=your_username/dataset_name \
+    --dataset.num_episodes=2 \
+    --dataset.fps=10 \
+    --dataset.single_task="Navigate around obstacles" \
+    --display_data=true
+```
+
+Replace `your_username/dataset_name` with your Hugging Face username and a name for your dataset.
+
+### What Gets Saved
+
+Your dataset includes:
+
+**Your Actions (2 things)**:
+
+- How much you moved forward/backward
+- How much you turned left/right
+
+**Robot Observations (12 things)**:
+
+- Front camera video
+- Rear camera video
+- Current speed
+- Battery level
+- Which way the robot is facing
+- GPS location (latitude, longitude, signal strength)
+- Network signal strength
+- Vibration level
+- Lamp status (on/off)
+
+### Where Your Data Goes
+
+On your computer: `~/.cache/huggingface/lerobot/{repo-id}`
+
+After recording, your data automatically uploads to your Hugging Face page:
+
+```bash
+echo https://huggingface.co/datasets/${HF_USER}/earthrover-navigation
+```
+
+Your dataset will be tagged with `LeRobot` for community discovery.

docs/source/envhub_leisaac.mdx

@@ -139,7 +139,7 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     so101_leader,
 )
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import init_logging
 from lerobot.envs.factory import make_env
 
@@ -196,7 +196,7 @@ def teleop_loop(teleop: Teleoperator, env: gym.Env, fps: int):
             obs, info = env.reset()
 
         dt_s = time.perf_counter() - loop_start
-        busy_wait(1 / fps - dt_s)
+        precise_sleep(1 / fps - dt_s)
         loop_s = time.perf_counter() - loop_start
         print(f"\ntime: {loop_s * 1e3:.2f}ms ({1 / loop_s:.0f} Hz)")
 

docs/source/il_robots.mdx

@@ -201,7 +201,8 @@ from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
-from lerobot.record import record_loop
+from lerobot.scripts.lerobot_record import record_loop
+from lerobot.processor import make_default_processors
 
 NUM_EPISODES = 5
 FPS = 30
@@ -209,12 +210,19 @@ EPISODE_TIME_SEC = 60
 RESET_TIME_SEC = 10
 TASK_DESCRIPTION = "My task description"
 
-# Create the robot and teleoperator configurations
-camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+# Create robot configuration
 robot_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem58760434471", id="my_awesome_follower_arm", cameras=camera_config
+    id="my_awesome_follower_arm",
+    cameras={
+        "front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS) # Optional: fourcc="MJPG" for troubleshooting OpenCV async error.
+    },
+    port="/dev/tty.usbmodem58760434471",
+)
+
+teleop_config = SO100LeaderConfig(
+    id="my_awesome_leader_arm",
+    port="/dev/tty.usbmodem585A0077581",
 )
-teleop_config = SO100LeaderConfig(port="/dev/tty.usbmodem585A0077581", id="my_awesome_leader_arm")
 
 # Initialize the robot and teleoperator
 robot = SO100Follower(robot_config)
@@ -243,6 +251,9 @@ init_rerun(session_name="recording")
 robot.connect()
 teleop.connect()
 
+# Create the required processors
+teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+
 episode_idx = 0
 while episode_idx < NUM_EPISODES and not events["stop_recording"]:
     log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
@@ -251,6 +262,9 @@ while episode_idx < NUM_EPISODES and not events["stop_recording"]:
         robot=robot,
         events=events,
         fps=FPS,
+        teleop_action_processor=teleop_action_processor,
+        robot_action_processor=robot_action_processor,
+        robot_observation_processor=robot_observation_processor,
         teleop=teleop,
         dataset=dataset,
         control_time_s=EPISODE_TIME_SEC,
@@ -265,6 +279,9 @@ while episode_idx < NUM_EPISODES and not events["stop_recording"]:
             robot=robot,
             events=events,
             fps=FPS,
+            teleop_action_processor=teleop_action_processor,
+            robot_action_processor=robot_action_processor,
+            robot_observation_processor=robot_observation_processor,
             teleop=teleop,
             control_time_s=RESET_TIME_SEC,
             single_task=TASK_DESCRIPTION,
@@ -393,7 +410,7 @@ import time
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
 episode_idx = 0
@@ -415,7 +432,7 @@ for idx in range(dataset.num_frames):
     }
     robot.send_action(action)
 
-    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+    precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
 
 robot.disconnect()
 ```
@@ -428,7 +445,7 @@ Your robot should replicate movements similar to those you recorded. For example
 
 ## Train a policy
 
-To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_train.py) script. A few arguments are required. Here is an example command:
 
 ```bash
 lerobot-train \
@@ -485,7 +502,7 @@ huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
 
 ## Run inference and evaluate your policy
 
-You can use the `record` script from [`lerobot/record.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/record.py) with a policy checkpoint as input, to run inference and evaluate your policy. For instance, run this command or API example to run inference and record 10 evaluation episodes:
+You can use the `record` script from [`lerobot-record`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_record.py) with a policy checkpoint as input, to run inference and evaluate your policy. For instance, run this command or API example to run inference and record 10 evaluation episodes:
 
 <hfoptions id="eval">
 <hfoption id="Command">

docs/source/installation.mdx

@@ -90,7 +90,7 @@ If you encounter build errors, you may need to install additional dependencies:
 To install these for linux run:
 
 ```bash
-sudo apt-get install cmake build-essential python-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config
+sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev
 ```
 
 For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)

docs/source/libero.mdx

@@ -62,6 +62,11 @@ lerobot-eval \
 
 - Pass a comma-separated list to `--env.task` for multi-suite evaluation.
 
+### Control Mode
+
+LIBERO now supports two control modes: relative and absolute. This matters because different VLA checkpoints are trained with different mode of action to output hence control parameterizations.
+You can switch them with: `env.control_mode = "relative"` and `env.control_mode = "absolute"`
+
 ### Policy inputs and outputs
 
 When using LIBERO through LeRobot, policies interact with the environment via **observations** and **actions**:

docs/source/policy_walloss_README.md

@@ -0,0 +1,35 @@
+# WALL-OSS
+
+This repository contains the Hugging Face port of **WALL-OSS**, a Vision-Language-Action model for cross-embodiment robotic control based on Qwen2.5-VL with flow matching/FAST action prediction.
+
+---
+
+## Model Overview
+
+| Feature            | Description                                           |
+| ------------------ | ----------------------------------------------------- | --- |
+| Base Model         | Qwen2.5-VL (Vision-Language Model)                    |
+| Action Prediction  | Flow Matching (diffusion) or FAST (discrete tokens)   |
+| Architecture       | Mixture of Experts (MoE) with action-specific routing |     |
+| Multi-Modal Inputs | Vision (images/videos), Language, Proprioception      |
+
+---
+
+## Citation
+
+If you use this work, please cite:
+
+```bibtex
+@article{zhai2025igniting,
+    title   = {Igniting VLMs Toward the Embodied Space},
+    author  = {Zhai, Andy and Liu, Brae and Fang, Bruno and Cai, Chalse and Ma, Ellie and Yin, Ethan and Wang, Hao and Zhou, Hugo and Wang, James and Shi, Lights and Liang, Lucy and Wang, Make and Wang, Qian and Gan, Roy and Yu, Ryan and Li, Shalfun and Liu, Starrick and Chen, Sylas and Chen, Vincent and Xu, Zach},
+    journal = {arXiv preprint arXiv:2509.11766},
+    year    = {2025}
+}
+```
+
+---
+
+## License
+
+This port follows the **Apache 2.0 License**.

docs/source/sarm.mdx

@@ -0,0 +1,586 @@
+# SARM: Stage-Aware Reward Modeling
+
+SARM (Stage-Aware Reward Modeling) is a video-based reward modeling framework for long-horizon robot manipulation tasks. This guide covers how to train SARM reward models and optionally use them with Reward-Aligned Behavior Cloning (RA-BC).
+
+**Paper**: [SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation](https://arxiv.org/abs/2509.25358)
+
+## Why Reward Models?
+
+Standard behavior cloning treats all demonstration frames equally, but real-world robot datasets are messy. They contain hesitations, corrections, and variable-quality trajectories. Reward models solve this by learning a generalizable notion of **task progress** from demonstrations: given video frames and a task description, they predict how close the robot is to completing the task (0→1). This learned "progress signal" can be used in multiple ways, two promising applications are: (1) **weighted imitation learning** (RA-BC), where high-progress frames receive more weight during policy training, and (2) **reinforcement learning**, where the reward model provides dense rewards for online or offline policy improvement.
+
+## Overview
+
+SARM has following features:
+
+1. **Stage-aware architecture**: Jointly predicts the high-level task stage and fine-grained progress within each stage
+2. **Subtask annotations**: Uses natural language subtask annotations to derive consistent progress labels
+3. **Temporal proportions**: Computes dataset-level priors (α̅\_k) for each subtask to normalize progress across variable-length demonstrations
+
+SARM trains on a compact **stage+tau** target for each frame:
+
+- **stage**: integer stage index `k ∈ {0, ..., K-1}`
+- **τ (tau)**: within-stage progress `τ ∈ [0, 1]`
+- **target encoding**: `y = k + τ` (this is what the dataset processor produces)
+
+At inference time (and in downstream RA-BC), SARM converts the raw `k + τ` value into a **normalized progress** in `[0, 1]` using dataset-level **temporal proportions** `α̅_k` (stored in `meta/temporal_proportions_*.json`).
+
+This matches **Formula (2)** from the paper:
+
+```
+progress_t = P_{k-1} + α̅_k × τ_t
+```
+
+Where:
+
+- `τ_t = (t - s_k) / (e_k - s_k)` is within-subtask normalized time
+- `P_{k-1}` is cumulative prior (sum of previous subtask proportions)
+- `α̅_k` is the temporal proportion for subtask k
+
+This ensures identical task states map to consistent progress values, even across demonstrations of different lengths.
+
+## Inputs and Targets (What the new code expects)
+
+SARM is trained through its processor (`src/lerobot/policies/sarm/processor_sarm.py`), which:
+
+- **Encodes** images and task text with CLIP (ViT-B/32) into `video_features` and `text_features`
+- **Pads/truncates** robot state into `state_features` (up to `max_state_dim`)
+- **Builds targets** as `sparse_targets` (and `dense_targets` in `dense_only`/`dual`) using the stage+tau encoding `y = k + τ`
+- **Masks rewind frames** using a per-sample `lengths` tensor (rewind is a training-time augmentation)
+
+At minimum, each training sample needs:
+
+- `task` (string): task description
+- `policy.image_key` images and `policy.state_key` states from the dataset
+
+---
+
+## Annotation Modes
+
+You can choose from **3 annotation modes** that determine how progress labels are computed:
+
+| Mode           | Annotations Required | Heads                        | Use Case                                                     |
+| -------------- | -------------------- | ---------------------------- | ------------------------------------------------------------ |
+| `single_stage` | None                 | Sparse only                  | Simple tasks, quick experiments, no VLM needed               |
+| `dense_only`   | Dense (VLM)          | Dual (sparse auto-generated) | Detailed subtask tracking without defining high-level stages |
+| `dual`         | Sparse + Dense (VLM) | Dual                         | Full SARM paper setup with both granularities                |
+
+### Mode Details
+
+<hfoptions id="mode_explanation">
+<hfoption id="single_stage">
+
+**No annotations required.** The entire episode is treated as a single stage called `"task"`, and progress is linear from 0 to 1 over the episode duration.
+
+- **Sparse head**: 1 stage ("task"), linear progress
+- **Dense head**: Not used
+- **Best for**: Simple tasks, quick experiments, or when VLM annotation is not available
+
+## Set Up Your Environment
+
+1. Install LeRobot by following our [Installation Guide](./installation).
+2. Install SARM dependencies by running:
+
+```bash
+pip install -e ".[sarm]"
+```
+
+Workflow:
+
+```
+1. Train SARM → 2. Visualize predictions → 3. (Optional) Train policy with RA-BC
+```
+
+</hfoption>
+<hfoption id="dense_only">
+
+**Only dense (fine-grained) annotations from a VLM.** The sparse head automatically uses a single `"task"` stage covering the full episode, while the dense head learns detailed subtask progression.
+
+- **Sparse head**: 1 stage ("task"), linear progress (auto-generated)
+- **Dense head**: Multiple fine-grained stages from VLM annotations
+- **Best for**: When you want detailed subtask tracking but don't need to define high-level stages
+
+Workflow:
+
+```
+1. Annotate (dense) → 2. Verify → 3. Train SARM → 4. Visualize → 5. (Optional) Train policy with RA-BC
+```
+
+</hfoption>
+<hfoption id="dual">
+
+**Both sparse and dense annotations from VLM.** Full dual-head mode as described in the SARM paper, with both high-level (sparse) and fine-grained (dense) stage predictions.
+
+- **Sparse head**: High-level stages from VLM annotations
+- **Dense head**: Fine-grained stages from VLM annotations
+- **Best for**: Complex multi-stage tasks where both granularities are useful
+
+Workflow:
+
+```
+1. Annotate (sparse+dense) → 2. Verify → 3. Train SARM → 4. Visualize → 5. (Optional) Train policy with RA-BC
+```
+
+</hfoption>
+</hfoptions>
+
+---
+
+## Step 1: Subtask Annotation
+
+<hfoptions id="annotation_mode">
+<hfoption id="single_stage">
+
+**No annotation required!** Skip this step entirely. The model will use the episode's task description and compute linear progress automatically.
+
+</hfoption>
+<hfoption id="dense_only">
+
+Generate **dense (fine-grained) annotations only** using a VLM. The sparse stage will be auto-generated.
+
+```bash
+python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
+  --repo-id your-username/your-dataset \
+  --dense-only \
+  --dense-subtasks "Bring robot arms up from starting position,Grab near side and do 1st fold,Grab side and do 2nd fold,Grab side and do 3rd fold to finish folding" \
+  --video-key observation.images.base \
+  --num-workers 4 \
+  --push-to-hub
+```
+
+**What gets saved:**
+
+- `meta/temporal_proportions_sparse.json` - Auto-generated sparse proportions (`{"task": 1.0}`)
+- `meta/temporal_proportions_dense.json` - Dense temporal proportions
+- Per-episode columns in `episodes/*.parquet`:
+  - `dense_subtask_names`, `dense_subtask_start_frames`, `dense_subtask_end_frames`
+  - (also time-based columns: `dense_subtask_start_times`, `dense_subtask_end_times`)
+
+</hfoption>
+<hfoption id="dual">
+
+Generate **both sparse (high-level) and dense (fine-grained) annotations** using a VLM.
+
+```bash
+python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
+  --repo-id your-username/your-dataset \
+  --sparse-subtasks "Bring arms up from starting position,Fold the towel (3 folds in total)" \
+  --dense-subtasks "Bring robot arms up from starting position,Grab near side and do 1st fold,Grab side and do 2nd fold,Grab side and do 3rd fold to finish folding" \
+  --video-key observation.images.base \
+  --num-workers 4 \
+  --push-to-hub
+```
+
+**What gets saved:**
+
+- `meta/temporal_proportions_sparse.json` - Sparse temporal proportions
+- `meta/temporal_proportions_dense.json` - Dense temporal proportions
+- Per-episode columns in `episodes/*.parquet`:
+  - `sparse_subtask_names`, `sparse_subtask_start_frames`, `sparse_subtask_end_frames`
+  - `dense_subtask_names`, `dense_subtask_start_frames`, `dense_subtask_end_frames`
+  - (also time-based columns: `*_subtask_start_times`, `*_subtask_end_times`)
+
+</hfoption>
+</hfoptions>
+
+### Annotation Arguments
+
+| Argument               | Description                                                                     |
+| ---------------------- | ------------------------------------------------------------------------------- |
+| `--repo-id`            | HuggingFace dataset repository ID                                               |
+| `--sparse-subtasks`    | Comma-separated list of high-level subtask names                                |
+| `--dense-subtasks`     | Comma-separated list of fine-grained subtask names                              |
+| `--dense-only`         | Generate only dense annotations (auto-creates sparse "task" stage)              |
+| `--video-key`          | Camera/video key to use (e.g., `observation.images.top`)                        |
+| `--num-workers`        | Number of parallel GPU workers (default: 1)                                     |
+| `--episodes`           | Specific episode indices to annotate (default: all)                             |
+| `--skip-existing`      | Skip episodes that already have annotations                                     |
+| `--model`              | VLM model (default: `Qwen/Qwen3-VL-30B-A3B-Instruct`)                           |
+| `--num-visualizations` | Number of episodes to visualize after annotation (default: 5, set to 0 to skip) |
+
+> **Note**: After annotation completes, 5 episodes are automatically visualized by default. Use `--num-visualizations 0` to skip this step.
+
+---
+
+## Step 2: Verify Annotations
+
+<hfoptions id="verify_mode">
+<hfoption id="single_stage">
+
+**No verification needed!** Skip this step.
+
+</hfoption>
+<hfoption id="dense_only">
+
+Visualize annotations using the `--visualize-only` flag:
+
+```bash
+python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
+  --repo-id your-username/your-dataset \
+  --visualize-only \
+  --visualize-type dense \
+  --num-visualizations 5 \
+  --video-key observation.images.base \
+  --output-dir ./subtask_viz
+```
+
+</hfoption>
+<hfoption id="dual">
+
+Visualize annotations using the `--visualize-only` flag:
+
+```bash
+python src/lerobot/data_processing/sarm_annotations/subtask_annotation.py \
+  --repo-id your-username/your-dataset \
+  --visualize-only \
+  --visualize-type both \
+  --num-visualizations 5 \
+  --video-key observation.images.base \
+  --output-dir ./subtask_viz
+```
+
+</hfoption>
+</hfoptions>
+
+This generates visualizations showing video frames with subtask boundaries overlaid and timeline of subtasks.
+
+### Visualization Arguments
+
+| Argument               | Description                                                    |
+| ---------------------- | -------------------------------------------------------------- |
+| `--visualize-only`     | Only visualize existing annotations (no generation)            |
+| `--num-visualizations` | Number of episodes to visualize (default: 5)                   |
+| `--visualize-type`     | Type of annotations to visualize: `sparse`, `dense`, or `both` |
+
+**Tip**: If annotations are inaccurate, adjust your subtask descriptions to be more specific and re-run.
+
+---
+
+## Step 3: Train SARM
+
+<hfoptions id="train_mode">
+<hfoption id="single_stage">
+
+Train with **no annotations** - uses linear progress from 0 to 1:
+
+```bash
+python src/lerobot/scripts/lerobot_train.py \
+  --dataset.repo_id=your-username/your-dataset \
+  --policy.type=sarm \
+  --policy.annotation_mode=single_stage \
+  --policy.image_key=observation.images.base \
+  --output_dir=outputs/train/sarm_single \
+  --batch_size=32 \
+  --steps=5000 \
+  --wandb.enable=true \
+  --wandb.project=sarm \
+  --policy.repo_id=your-username/your-model-name
+```
+
+</hfoption>
+<hfoption id="dense_only">
+
+Train with **dense annotations only** (sparse auto-generated):
+
+```bash
+python src/lerobot/scripts/lerobot_train.py \
+  --dataset.repo_id=your-username/your-dataset \
+  --policy.type=sarm \
+  --policy.annotation_mode=dense_only \
+  --policy.image_key=observation.images.base \
+  --output_dir=outputs/train/sarm_dense \
+  --batch_size=32 \
+  --steps=5000 \
+  --wandb.enable=true \
+  --wandb.project=sarm \
+  --policy.repo_id=your-username/your-model-name
+```
+
+</hfoption>
+<hfoption id="dual">
+
+Train with **both sparse and dense annotations**:
+
+```bash
+python src/lerobot/scripts/lerobot_train.py \
+  --dataset.repo_id=your-username/your-dataset \
+  --policy.type=sarm \
+  --policy.annotation_mode=dual \
+  --policy.image_key=observation.images.base \
+  --output_dir=outputs/train/sarm_dual \
+  --batch_size=32 \
+  --steps=5000 \
+  --wandb.enable=true \
+  --wandb.project=sarm \
+  --policy.repo_id=your-username/your-model-name
+```
+
+</hfoption>
+</hfoptions>
+
+### Multi-GPU Training
+
+Add `accelerate launch --multi_gpu --num_processes=4` to use multiple GPUs for training.
+
+### Training Arguments
+
+| Argument                   | Description                                                       | Default                  |
+| -------------------------- | ----------------------------------------------------------------- | ------------------------ |
+| `--policy.annotation_mode` | `single_stage`, `dense_only`, or `dual`                           | `single_stage`           |
+| `--policy.image_key`       | Camera key for images                                             | `observation.images.top` |
+| `--policy.state_key`       | Key for joint states                                              | `observation.state`      |
+| `--policy.n_obs_steps`     | Observation history steps (total obs frames = `n_obs_steps + 1`)  | `8`                      |
+| `--policy.frame_gap`       | Gap (in frames) between sampled observations (at 30 fps: 30 ≈ 1s) | `30`                     |
+
+---
+
+## Step 4: Visualize Predictions
+
+Use `compute_rabc_weights.py` with `--visualize-only` to visualize model predictions (and, if available, annotation-derived targets) without writing a parquet file.
+
+<hfoptions id="viz_mode">
+<hfoption id="single_stage">
+
+```bash
+python src/lerobot/policies/sarm/compute_rabc_weights.py \
+  --dataset-repo-id your-username/your-dataset \
+  --reward-model-path your-username/sarm-model \
+  --visualize-only \
+  --num-visualizations 5 \
+  --head-mode sparse \
+  --output-dir ./sarm_viz
+```
+
+</hfoption>
+<hfoption id="dense_only">
+
+```bash
+python src/lerobot/policies/sarm/compute_rabc_weights.py \
+  --dataset-repo-id your-username/your-dataset \
+  --reward-model-path your-username/sarm-model \
+  --visualize-only \
+  --num-visualizations 5 \
+  --head-mode dense \
+  --output-dir ./sarm_viz
+```
+
+</hfoption>
+<hfoption id="dual">
+
+```bash
+python src/lerobot/policies/sarm/compute_rabc_weights.py \
+  --dataset-repo-id your-username/your-dataset \
+  --reward-model-path your-username/sarm-model \
+  --visualize-only \
+  --num-visualizations 5 \
+  --head-mode both \
+  --output-dir ./sarm_viz
+```
+
+</hfoption>
+</hfoptions>
+
+The visualization shows:
+
+- **Progress plot**: Predicted progress (and optional annotation-derived “GT” when available and `--stride 1`)
+- **Stage probabilities**: Stacked area plot of predicted stage probabilities
+- **Sample frames**: Key frames from the episode with progress/stage labels
+
+### Visualization Arguments
+
+| Argument               | Description                                               |
+| ---------------------- | --------------------------------------------------------- |
+| `--visualize-only`     | Only visualize predictions (no RABC computation)          |
+| `--num-visualizations` | Number of episodes to visualize (default: 5)              |
+| `--head-mode`          | SARM head to use: `sparse`, `dense`, or `both`            |
+| `--stride`             | Compute every N frames, interpolate the rest (default: 1) |
+
+---
+
+## Step 5 (Optional): Train Policy with RA-BC
+
+Reward-Aligned Behavior Cloning (RA-BC) uses the trained SARM model to weight training samples based on predicted progress improvement. This requires two steps:
+
+1. **Precompute progress values** for all frames using the trained SARM model
+2. **Train policy** with RA-BC weighting using the precomputed values
+
+### How RA-BC Works
+
+For each training sample, RA-BC computes the progress delta:
+
+```
+r_i = φ(o_{t+Δ}) - φ(o_t)
+```
+
+Where `φ` is the SARM progress prediction and `Δ` is the policy's `chunk_size`. Samples with positive progress (good demonstrations) get higher weights, while samples with negative or zero progress get down-weighted.
+
+The weighting follows **Equations 8-9** from the paper:
+
+- **Soft weight**: `w̃_i = clip((r_i − (μ − 2σ)) / (4σ + ε), 0, 1)`
+- **Final weight**: `w_i = 𝟙{r_i > κ} + 𝟙{0 ≤ r_i ≤ κ} × w̃_i`
+
+### Step 5a: Compute SARM Progress Values
+
+First, run the SARM model on all frames in your dataset to compute progress values:
+
+```bash
+python src/lerobot/policies/sarm/compute_rabc_weights.py \
+  --dataset-repo-id your-username/your-dataset \
+  --reward-model-path your-username/sarm-model \
+  --head-mode sparse \
+  --num-visualizations 5 \
+  --push-to-hub
+```
+
+This script:
+
+- Processes all frames and computes progress values
+- Saves progress values to a parquet file next to the dataset on disk (defaults to `<dataset_root>/sarm_progress.parquet`)
+- Generates visualizations of the first N episodes (default: 5)
+
+**Arguments:**
+
+| Argument               | Description                                                    | Default    |
+| ---------------------- | -------------------------------------------------------------- | ---------- |
+| `--reward-model-path`  | Path to trained SARM model                                     | (required) |
+| `--head-mode`          | SARM head to use: `sparse`, `dense`, or `both`                 | `sparse`   |
+| `--device`             | Device for inference                                           | `cuda`     |
+| `--visualize-only`     | Only visualize predictions (no RA-BC computation)              | `false`    |
+| `--num-visualizations` | Number of episodes to visualize (default: 5, set to 0 to skip) | `5`        |
+
+**Output format** (`sarm_progress.parquet`):
+
+| Column            | Description                                    |
+| ----------------- | ---------------------------------------------- |
+| `index`           | Global frame index in dataset                  |
+| `episode_index`   | Episode number                                 |
+| `frame_index`     | Local frame index within episode               |
+| `progress_sparse` | Sparse head progress value [0, 1]              |
+| `progress_dense`  | Dense head progress value [0, 1] (if computed) |
+
+### Step 5b: Train Policy with RA-BC
+
+Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`). Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
+
+```bash
+python src/lerobot/scripts/lerobot_train.py \
+  --dataset.repo_id=your-username/your-dataset \
+  --policy.type=pi0 \
+  --use_rabc=true \
+  --rabc_head_mode=sparse \
+  --rabc_kappa=0.01 \
+  --output_dir=outputs/train/policy_rabc \
+  --batch_size=32 \
+  --steps=40000
+```
+
+The training script automatically:
+
+- Loads the precomputed progress values from the parquet file
+- Uses the policy's `chunk_size` to compute progress deltas (Δ)
+- Computes sample weights based on progress improvement
+- Applies weighted loss during training
+
+**RA-BC Arguments:**
+
+| Argument               | Description                                                | Default                            |
+| ---------------------- | ---------------------------------------------------------- | ---------------------------------- |
+| `--use_rabc`           | Enable RA-BC sample weighting                              | `false`                            |
+| `--rabc_progress_path` | Path to progress parquet file (auto-detected from dataset) | `sarm_progress.parquet` in dataset |
+| `--rabc_head_mode`     | Which SARM head's progress to use: `sparse` or `dense`     | `sparse`                           |
+| `--rabc_kappa`         | Threshold κ for high-quality samples                       | `0.01`                             |
+
+### Tuning RA-BC Kappa
+
+The `kappa` parameter is the threshold that determines which samples get full weight (w=1). Understanding how to tune it is critical for RA-BC to work effectively.
+
+**How the weighting works:**
+
+| Condition           | Weight                  |
+| ------------------- | ----------------------- |
+| `delta > kappa`     | 1.0 (hard threshold)    |
+| `0 ≤ delta ≤ kappa` | Soft weight from Eq. 8  |
+| `delta < 0`         | 0.0 (negative progress) |
+
+**Diagnosing kappa issues:**
+
+Monitor these WandB metrics during training:
+
+| Metric             | Healthy Range | Problem Indicator         |
+| ------------------ | ------------- | ------------------------- |
+| `rabc_mean_weight` | 0.3 - 0.8     | ≈ 1.0 means kappa too low |
+| `rabc_delta_mean`  | > 0           | Should be positive        |
+| `rabc_delta_std`   | > 0           | Variance in data quality  |
+
+**If `rabc_mean_weight ≈ 1.0`:** Your kappa is too low. Most samples have `delta > kappa` and bypass the soft-weighting entirely. RA-BC becomes equivalent to vanilla BC.
+
+**Setting kappa based on your data:**
+
+The default `kappa=0.01` was tuned for the paper's T-shirt folding task (~90s episodes at 30fps). For your dataset, check the logged `rabc_delta_mean` and `rabc_delta_std`:
+
+```
+# If delta_mean ≈ 0.03 and delta_std ≈ 0.02:
+# Most deltas fall in range [0.01, 0.05]
+
+# Option 1: Set kappa = delta_mean (medium selectivity)
+--rabc_kappa=0.03
+
+# Option 2: Set kappa = delta_mean + delta_std (high selectivity)
+--rabc_kappa=0.05
+
+# Option 3: Set kappa = delta_mean + 2*delta_std (very selective)
+--rabc_kappa=0.07
+```
+
+**When RA-BC may not help:**
+
+If your dataset is already high quality (consistent progress across all demonstrations), RA-BC won't provide much benefit since there's nothing to filter.
+
+### Multi-GPU Training with RA-BC
+
+```bash
+accelerate launch \
+  --multi_gpu \
+  --num_processes=4 \
+  src/lerobot/scripts/lerobot_train.py \
+  --dataset.repo_id=your-username/your-dataset \
+  --policy.type=pi0 \
+  --use_rabc=true \
+  --rabc_kappa=0.01 \
+  --output_dir=outputs/train/policy_rabc \
+  --batch_size=32 \
+  --steps=40000
+```
+
+---
+
+## Tips & Best Practices
+
+### Choosing a Mode
+
+- **Start with `single_stage`** for quick experiments - no annotation overhead
+- Use **`dense_only`** when you want detailed progress tracking but tasks don't have clear high-level stages
+- Use **`dual`** for complex tasks where both coarse and fine-grained progress is meaningful
+
+### Annotation Quality
+
+1. **Be specific with subtask names**: Instead of "fold", use "grab near side and fold toward center"
+2. **Verify with visualization**: Always check a few episodes before training
+3. **Consistent naming**: Use the same subtask names across all episodes
+
+### RA-BC
+
+1. **Train SARM first**: RA-BC quality depends entirely on SARM quality
+2. **Monitor `rabc_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))
+
+---
+
+## 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}
+}
+```

docs/source/so101.mdx

@@ -30,131 +30,6 @@ The follower arm uses 6x STS3215 motors with 1/345 gearing. The leader, however,
 | Wrist Roll          |   5   |  1 / 147   |
 | Gripper             |   6   |  1 / 147   |
 
-### Clean Parts
-
-Remove all support material from the 3D-printed parts. The easiest way to do this is using a small screwdriver to get underneath the support material.
-
-It is advisable to install one 3-pin cable in the motor after placing them before continuing assembly.
-
-### Joint 1
-
-- Place the first motor into the base.
-- Fasten the motor with 4 M2x6mm screws (smallest screws). Two from the top and two from the bottom.
-- Slide over the first motor holder and fasten it using two M2x6mm screws (one on each side).
-- Install both motor horns, securing the top horn with a M3x6mm screw.
-- Attach the shoulder part.
-- Tighten the shoulder part with 4 M3x6mm screws on top and 4 M3x6mm screws on the bottom
-- Add the shoulder motor holder.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint1_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-### Joint 2
-
-- Slide the second motor in from the top.
-- Fasten the second motor with 4 M2x6mm screws.
-- Attach both motor horns to motor 2, again use the M3x6mm horn screw.
-- Attach the upper arm with 4 M3x6mm screws on each side.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint2_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-### Joint 3
-
-- Insert motor 3 and fasten using 4 M2x6mm screws
-- Attach both motor horns to motor 3 and secure one again with a M3x6mm horn screw.
-- Connect the forearm to motor 3 using 4 M3x6mm screws on each side.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint3_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-### Joint 4
-
-- Slide over motor holder 4.
-- Slide in motor 4.
-- Fasten motor 4 with 4 M2x6mm screws and attach its motor horns, use a M3x6mm horn screw.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint4_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-### Joint 5
-
-- Insert motor 5 into the wrist holder and secure it with 2 M2x6mm front screws.
-- Install only one motor horn on the wrist motor and secure it with a M3x6mm horn screw.
-- Secure the wrist to motor 4 using 4 M3x6mm screws on both sides.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint5_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-### Gripper / Handle
-
-<hfoptions id="assembly">
-<hfoption id="Follower">
-
-- Attach the gripper to motor 5, attach it to the motor horn on the wrist using 4 M3x6mm screws.
-- Insert the gripper motor and secure it with 2 M2x6mm screws on each side.
-- Attach the motor horns and again use a M3x6mm horn screw.
-- Install the gripper claw and secure it with 4 M3x6mm screws on both sides.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Gripper_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-</hfoption>
-<hfoption id="Leader">
-
-- Mount the leader holder onto the wrist and secure it with 4 M3x6mm screws.
-- Attach the handle to motor 5 using 1 M2x6mm screw.
-- Insert the gripper motor, secure it with 2 M2x6mm screws on each side, attach a motor horn using a M3x6mm horn screw.
-- Attach the follower trigger with 4 M3x6mm screws.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Leader_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-</hfoption>
-</hfoptions>
-
 ## Configure the motors
 
 ### 1. Find the USB ports associated with each arm
@@ -340,6 +215,131 @@ leader.setup_motors()
 </hfoption>
 </hfoptions>
 
+### Clean Parts
+
+Remove all support material from the 3D-printed parts. The easiest way to do this is using a small screwdriver to get underneath the support material.
+
+It is advisable to install one 3-pin cable in the motor after placing them before continuing assembly.
+
+### Joint 1
+
+- Place the first motor into the base.
+- Fasten the motor with 4 M2x6mm screws (smallest screws). Two from the top and two from the bottom.
+- Slide over the first motor holder and fasten it using two M2x6mm screws (one on each side).
+- Install both motor horns, securing the top horn with a M3x6mm screw.
+- Attach the shoulder part.
+- Tighten the shoulder part with 4 M3x6mm screws on top and 4 M3x6mm screws on the bottom
+- Add the shoulder motor holder.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint1_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+### Joint 2
+
+- Slide the second motor in from the top.
+- Fasten the second motor with 4 M2x6mm screws.
+- Attach both motor horns to motor 2, again use the M3x6mm horn screw.
+- Attach the upper arm with 4 M3x6mm screws on each side.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint2_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+### Joint 3
+
+- Insert motor 3 and fasten using 4 M2x6mm screws
+- Attach both motor horns to motor 3 and secure one again with a M3x6mm horn screw.
+- Connect the forearm to motor 3 using 4 M3x6mm screws on each side.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint3_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+### Joint 4
+
+- Slide over motor holder 4.
+- Slide in motor 4.
+- Fasten motor 4 with 4 M2x6mm screws and attach its motor horns, use a M3x6mm horn screw.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint4_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+### Joint 5
+
+- Insert motor 5 into the wrist holder and secure it with 2 M2x6mm front screws.
+- Install only one motor horn on the wrist motor and secure it with a M3x6mm horn screw.
+- Secure the wrist to motor 4 using 4 M3x6mm screws on both sides.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint5_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+### Gripper / Handle
+
+<hfoptions id="assembly">
+<hfoption id="Follower">
+
+- Attach the gripper to motor 5, attach it to the motor horn on the wrist using 4 M3x6mm screws.
+- Insert the gripper motor and secure it with 2 M2x6mm screws on each side.
+- Attach the motor horns and again use a M3x6mm horn screw.
+- Install the gripper claw and secure it with 4 M3x6mm screws on both sides.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Gripper_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+</hfoption>
+<hfoption id="Leader">
+
+- Mount the leader holder onto the wrist and secure it with 4 M3x6mm screws.
+- Attach the handle to motor 5 using 1 M2x6mm screw.
+- Insert the gripper motor, secure it with 2 M2x6mm screws on each side, attach a motor horn using a M3x6mm horn screw.
+- Attach the follower trigger with 4 M3x6mm screws.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Leader_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+</hfoption>
+</hfoptions>
+
 ## Calibrate
 
 Next, you'll need to calibrate your robot to ensure that the leader and follower arms have the same position values when they are in the same physical position.

docs/source/torch_accelerators.mdx

@@ -0,0 +1,42 @@
+# PyTorch accelerators
+
+LeRobot supports multiple hardware acceleration options for both training and inference.
+
+These options include:
+
+- **CPU**: CPU executes all computations, no dedicated accelerator is used
+- **CUDA**: acceleration with NVIDIA & AMD GPUs
+- **MPS**: acceleration with Apple Silicon GPUs
+- **XPU**: acceleration with Intel integrated and discrete GPUs
+
+## Getting Started
+
+To use particular accelerator, a suitable version of PyTorch should be installed.
+
+For CPU, CUDA, and MPS backends follow instructions provided on [PyTorch installation page](https://pytorch.org/get-started/locally).
+For XPU backend, follow instructions from [PyTorch documentation](https://docs.pytorch.org/docs/stable/notes/get_start_xpu.html).
+
+### Verifying the installation
+
+After installation, accelerator availability can be verified by running
+
+```python
+import torch
+print(torch.<backend_name>.is_available())  # <backend_name> is cuda, mps, or xpu
+```
+
+## How to run training or evaluation
+
+To select the desired accelerator, use the `--policy.device` flag when running `lerobot-train` or `lerobot-eval`. For example, to use MPS on Apple Silicon, run:
+
+```bash
+lerobot-train
+    --policy.device=mps ...
+```
+
+```bash
+lerobot-eval \
+    --policy.device=mps ...
+```
+
+However, in most cases, presence of an accelerator is detected automatically and `policy.device` parameter can be omitted from CLI commands.

docs/source/unitree_g1.mdx

@@ -0,0 +1,208 @@
+# Unitree G1 Robot Setup and Control
+
+This guide covers the complete setup process for the Unitree G1 humanoid, from initial connection to running gr00t_wbc locomotion.
+
+## About the Unitree G1
+
+We offer support for both 29 and 23 DOF G1. We introduce:
+
+- **`unitree g1` robot class, handling low level communication with the humanoid**
+- **ZMQ socket bridge** for remote communication over WiFi, allowing one to deploy policies remotely instead of over ethernet or directly on the Orin
+- **GR00T locomotion policy** for bipedal walking and balance
+- **MuJoCo simulation mode** for testing policies without the physical robot
+
+---
+
+## Part 1: Connect to Robot over Ethernet
+
+### Step 1: Configure Your Computer's Ethernet Interface
+
+Set a static IP on the same subnet as the robot:
+
+```bash
+# Replace 'enp131s0' with your ethernet interface name (check with `ip a`)
+sudo ip addr flush dev enp131s0
+sudo ip addr add 192.168.123.200/24 dev enp131s0
+sudo ip link set enp131s0 up
+```
+
+**Note**: The robot's Ethernet IP is fixed at `192.168.123.164`. Your computer must use `192.168.123.x` where x ≠ 164.
+
+### Step 2: SSH into the Robot
+
+```bash
+ssh unitree@192.168.123.164
+# Password: 123
+```
+
+You should now be connected to the robot's onboard computer.
+
+---
+
+## Part 2: Enable WiFi on the Robot
+
+Once connected via Ethernet, follow these steps to enable WiFi:
+
+### Step 1: Enable WiFi Hardware
+
+```bash
+# Unblock WiFi radio
+sudo rfkill unblock wifi
+sudo rfkill unblock all
+
+# Bring up WiFi interface
+sudo ip link set wlan0 up
+
+# Enable NetworkManager control
+sudo nmcli radio wifi on
+sudo nmcli device set wlan0 managed yes
+sudo systemctl restart NetworkManager
+```
+
+### Step 2: Enable Internet Forwarding
+
+**On your laptop:**
+
+```bash
+# Enable IP forwarding
+sudo sysctl -w net.ipv4.ip_forward=1
+
+# Set up NAT (replace wlp132s0f0 with your WiFi interface)
+sudo iptables -t nat -A POSTROUTING -o wlp132s0f0 -s 192.168.123.0/24 -j MASQUERADE
+sudo iptables -A FORWARD -i wlp132s0f0 -o enp131s0 -m state --state RELATED,ESTABLISHED -j ACCEPT
+sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
+```
+
+**On the robot:**
+
+```bash
+# Add laptop as default gateway
+sudo ip route del default 2>/dev/null || true
+sudo ip route add default via 192.168.123.200 dev eth0
+echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf
+
+# Test connection
+ping -c 3 8.8.8.8
+```
+
+### Step 3: Connect to WiFi Network
+
+```bash
+# List available networks
+nmcli device wifi list
+
+# Connect to your WiFi (example)
+sudo nmcli connection add type wifi ifname wlan0 con-name "YourNetwork" ssid "YourNetwork"
+sudo nmcli connection modify "YourNetwork" wifi-sec.key-mgmt wpa-psk
+sudo nmcli connection modify "YourNetwork" wifi-sec.psk "YourPassword"
+sudo nmcli connection modify "YourNetwork" connection.autoconnect yes
+sudo nmcli connection up "YourNetwork"
+
+# Check WiFi IP address
+ip a show wlan0
+```
+
+### Step 4: SSH Over WiFi
+
+Once connected to WiFi, note the robot's IP address and disconnect the Ethernet cable. You can now SSH over WiFi:
+
+```bash
+ssh unitree@<YOUR_ROBOT_IP>
+# Password: 123
+```
+
+Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address (e.g., `172.18.129.215`).
+
+---
+
+## Part 3: Robot Server Setup
+
+### Step 1: Install LeRobot on the Orin
+
+SSH into the robot and install LeRobot:
+
+```bash
+ssh unitree@<YOUR_ROBOT_IP>
+
+conda create -y -n lerobot python=3.10
+conda activate lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python  && pip install -e .
+```
+
+**Note**: The Unitree SDK requires CycloneDDS v0.10.2 to be installed. See the [Unitree SDK documentation](https://github.com/unitreerobotics/unitree_sdk2_python) for details.
+
+### Step 2: Run the Robot Server
+
+On the robot:
+
+```bash
+python src/lerobot/robots/unitree_g1/run_g1_server.py
+```
+
+**Important**: Keep this terminal running. The server must be active for remote control.
+
+---
+
+## Part 4: Running GR00T Locomotion
+
+With the robot server running, you can now control the robot from your laptop.
+
+### Step 1: Install LeRobot on your machine
+
+```bash
+conda create -y -n lerobot python=3.10
+conda activate lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python  && pip install -e .
+```
+
+### Step 2: Update Robot IP in Config
+
+Edit the config file to match your robot's WiFi IP:
+
+```python
+# In src/lerobot/robots/unitree_g1/config_unitree_g1.py
+robot_ip: str = "<YOUR_ROBOT_IP>"  # Replace with your robot's WiFi IP.
+```
+
+**Note**: When running directly on the G1 (not remotely), set `robot_ip: str = "127.0.0.1"` instead.
+
+### Step 3: Run the Locomotion Policy
+
+```bash
+# Run GR00T locomotion controller
+python examples/unitree_g1/gr00t_locomotion.py --repo-id "nepyope/GR00T-WholeBodyControl_g1"
+```
+
+### Step 4: Control with Remote
+
+- **Left stick**: Forward/backward and left/right movement
+- **Right stick**: Rotation
+- **R1 button**: Raise waist height
+- **R2 button**: Lower waist height
+
+Press `Ctrl+C` to stop the policy.
+
+---
+
+## Extra: Running in Simulation Mode (MuJoCo)
+
+You can now test and develop policies without a physical robot using MuJoCo. to do so set `is_simulation=True` in config.
+
+## Additional Resources
+
+- [Unitree SDK Documentation](https://github.com/unitreerobotics/unitree_sdk2_python)
+- [GR00T Policy Repository](https://huggingface.co/nepyope/GR00T-WholeBodyControl_g1)
+- [LeRobot Documentation](https://github.com/huggingface/lerobot)
+- [Unitree_IL_Lerobot](https://github.com/unitreerobotics/unitree_IL_lerobot)
+
+---
+
+_Last updated: December 2025_

docs/source/using_dataset_tools.mdx

@@ -11,13 +11,14 @@ LeRobot provides several utilities for manipulating datasets:
 3. **Merge Datasets** - Combine multiple datasets into one. The datasets must have identical features, and episodes are concatenated in the order specified in `repo_ids`
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
+6. **Convert to Video** - Convert image-based datasets to video format for efficient storage
 
 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
 
 ## Command-Line Tool: lerobot-edit-dataset
 
-`lerobot-edit-dataset` is a command-line script for editing datasets. It can be used to delete episodes, split datasets, merge datasets, add features, and remove features.
+`lerobot-edit-dataset` is a command-line script for editing datasets. It can be used to delete episodes, split datasets, merge datasets, add features, remove features, and convert image datasets to video format.
 
 Run `lerobot-edit-dataset --help` for more information on the configuration of each operation.
 
@@ -86,9 +87,71 @@ lerobot-edit-dataset \
     --operation.feature_names "['observation.images.top']"
 ```
 
+#### Convert to Video
+
+Convert an image-based dataset to video format, creating a new LeRobotDataset where images are stored as videos. This is useful for reducing storage requirements and improving data loading performance. The new dataset will have the exact same structure as the original, but with images encoded as MP4 videos in the proper LeRobot format.
+
+```bash
+# Local-only: Save to a custom output directory (no hub push)
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_to_video \
+    --operation.output_dir /path/to/output/pusht_video
+
+# Save with new repo_id (local storage)
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --new_repo_id lerobot/pusht_video \
+    --operation.type convert_to_video
+
+# Convert and push to Hugging Face Hub
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --new_repo_id lerobot/pusht_video \
+    --operation.type convert_to_video \
+    --push_to_hub true
+
+# Convert with custom video codec and quality settings
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_to_video \
+    --operation.output_dir outputs/pusht_video \
+    --operation.vcodec libsvtav1 \
+    --operation.pix_fmt yuv420p \
+    --operation.g 2 \
+    --operation.crf 30
+
+# Convert only specific episodes
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_to_video \
+    --operation.output_dir outputs/pusht_video \
+    --operation.episode_indices "[0, 1, 2, 5, 10]"
+
+# Convert with multiple workers for parallel processing
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_to_video \
+    --operation.output_dir outputs/pusht_video \
+    --operation.num_workers 8
+```
+
+**Parameters:**
+
+- `output_dir`: Custom output directory (optional - by default uses `new_repo_id` or `{repo_id}_video`)
+- `vcodec`: Video codec to use - options: `h264`, `hevc`, `libsvtav1` (default: `libsvtav1`)
+- `pix_fmt`: Pixel format - options: `yuv420p`, `yuv444p` (default: `yuv420p`)
+- `g`: Group of pictures (GOP) size - lower values give better quality but larger files (default: 2)
+- `crf`: Constant rate factor - lower values give better quality but larger files, 0 is lossless (default: 30)
+- `fast_decode`: Fast decode tuning option (default: 0)
+- `episode_indices`: List of specific episodes to convert (default: all episodes)
+- `num_workers`: Number of parallel workers for processing (default: 4)
+
+**Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved.
+
 ### Push to Hub
 
-Add the `--push_to_hub` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:
+Add the `--push_to_hub true` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:
 
 ```bash
 lerobot-edit-dataset \
@@ -96,7 +159,45 @@ lerobot-edit-dataset \
     --new_repo_id lerobot/pusht_after_deletion \
     --operation.type delete_episodes \
     --operation.episode_indices "[0, 2, 5]" \
-    --push_to_hub
+    --push_to_hub true
 ```
 
 There is also a tool for adding features to a dataset that is not yet covered in `lerobot-edit-dataset`.
+
+# Dataset Visualization
+
+## Online Visualization
+
+When you record a dataset using `lerobot`, it automatically uploads to the Hugging Face Hub unless you specify otherwise. To view the dataset online, use our **LeRobot Dataset Visualizer**, available at:
+https://huggingface.co/spaces/lerobot/visualize_dataset
+
+## Local Visualization
+
+You can also visualize episodes from a dataset locally using our command-line tool.
+
+**From the Hugging Face Hub:**
+
+```bash
+lerobot-dataset-viz \
+    --repo-id lerobot/pusht \
+    --episode-index 0
+```
+
+**From a local folder:**
+Add the `--root` option and set `--mode local`. For example, to search in `./my_local_data_dir/lerobot/pusht`:
+
+```bash
+lerobot-dataset-viz \
+    --repo-id lerobot/pusht \
+    --root ./my_local_data_dir \
+    --mode local \
+    --episode-index 0
+```
+
+Once executed, the tool opens `rerun.io` and displays the camera streams, robot states, and actions for the selected episode.
+
+For advanced usage—including visualizing datasets stored on a remote server—run:
+
+```bash
+lerobot-dataset-viz --help
+```

docs/source/walloss.mdx

@@ -0,0 +1,74 @@
+# WALL-OSS
+
+WALL-OSS is an open-source foundation model for embodied intelligence, proposed by the [XSquare Robot](https://x2robot.com/en/research/68bc2cde8497d7f238dde690) team in 2025. The LeRobot implementation is adapted from their open-source [WallX](https://github.com/X-Square-Robot/wall-x) repository.
+
+X Square Robot’s WALL-OSS is now integrated into Hugging Face’s LeRobot ecosystem. This is an exciting collaborative project between the LeRobot and X Square Robot teams. You can now post-train, evaluate, and deploy WALL-OSS directly through LeRobot. With this, we’re aiming to make it easier for the open-source robotics community to customize and deploy WALL-OSS foundation models. Read and explore WALL-OSS [paper](https://arxiv.org/pdf/2509.11766) and [code](https://github.com/X-Square-Robot/wall-x).
+
+## Model Overview
+
+The WALL-OSS team is building the embodied foundation model to capture and compress the world's most valuable data: the continuous, high-fidelity stream of physical interaction. By creating a direct feedback loop between the model's decisions and the body's lived experience, the emergence of a truly generalizable intelligence is enabled—one that understands not just how the world works, but how to act effectively within it.
+
+Technically, WALL-OSS introduces a tightly coupled multimodal architecture (tightly-coupled MoE structure) that integrates both discrete and continuous action modeling strategies. Through a two-stage training pipeline (Inspiration → Integration), the model gradually unifies semantic reasoning and high-frequency action generation. Its core innovations include:
+
+- **Embodied perception–enhanced multimodal pretraining**: Large-scale training on unified vision–language–action data to strengthen spatial, causal, and manipulation understanding.
+- **Unified Cross-Level Chain-of-Thought (Uni-CoT)**: A single differentiable framework that unifies high-level instruction reasoning, sub-task decomposition, and fine-grained action synthesis, forming a continuous chain from “understanding” to “execution.”
+- **Mixture-of-Experts (MoE) action heads**: Dynamically activating experts depending on the task phase and modeling actions in discrete or continuous space to maintain stable VLM priors.
+- **Two-stage training paradigm**:
+  - **Inspiration stage**: Injecting discrete action priors to strengthen spatial understanding and semantic-action alignment.
+  - **Integration stage**: Using flow matching to achieve high-frequency continuous control.
+
+## Installation Requirements
+
+1. Install LeRobot by following our [Installation Guide](./installation).
+2. Install WallX dependencies by running:
+
+   ```bash
+   pip install -e ".[wallx]"
+   ```
+
+## Usage
+
+To use WallX in LeRobot, specify the policy type as:
+
+```python
+policy.type=wall_x
+```
+
+## Training
+
+For training WallX, you can use the standard LeRobot training script with the appropriate configuration:
+
+```bash
+python src/lerobot/scripts/lerobot_train.py \
+    --dataset.repo_id=your_dataset \
+    --policy.type=wall_x \
+    --output_dir=./outputs/wallx_training \
+    --job_name=wallx_training \
+    --policy.repo_id=your_repo_id \
+    --policy.pretrained_name_or_path=x-square-robot/wall-oss-flow \
+    --policy.prediction_mode=diffusion \
+    --policy.attn_implementation=eager \
+    --steps=3000 \
+    --policy.device=cuda \
+    --batch_size=32
+```
+
+### Training Arguments
+
+| Argument                       | Description                                                                                                                                                   |
+| ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `--dataset.repo_id`            | The Hugging Face Hub repository ID for your training dataset (e.g., `lerobot/aloha_sim_insertion_human`)                                                      |
+| `--policy.type`                | Specifies using the WallX policy architecture                                                                                                                 |
+| `--output_dir`                 | Local directory where training checkpoints and logs will be saved                                                                                             |
+| `--job_name`                   | A name identifier for this training run (used in logging/tracking)                                                                                            |
+| `--policy.repo_id`             | Your Hugging Face Hub repo ID where the trained model will be pushed                                                                                          |
+| `--policy.pretrained_path`     | Path to pretrained WallX weights to initialize from (the official WALL-OSS checkpoint)                                                                        |
+| `--policy.prediction_mode`     | The action prediction strategy: `diffusion` or `fast` - `diffusion` uses iterative denoising for action generation, `fast` uses next token prediction instead |
+| `--policy.attn_implementation` | Attention implementation backend - `eager` uses standard PyTorch attention (alternatives include `flash_attention_2` or `sdpa`)                               |
+| `--steps`                      | Total number of training steps to run                                                                                                                         |
+| `--policy.device`              | Device to train on (`cuda` for GPU, `cpu` for CPU)                                                                                                            |
+| `--batch_size`                 | Number of samples per training batch                                                                                                                          |
+
+## License
+
+This model follows the **Apache 2.0 License**, consistent with the original [WallX repository](https://github.com/X-Square-Robot/wall-x).

docs/source/xvla.mdx

@@ -0,0 +1,528 @@
+# X-VLA: The First Soft-Prompted Robot Foundation Model for Any Robot, Any Task
+
+## Overview
+
+For years, robotics has aspired to build agents that can follow natural human instructions and operate dexterously across many environments and robot bodies. Recent breakthroughs in LLMs and VLMs suggest a path forward: extend these foundation-model architectures to embodied control by grounding them in actions. This has led to the rise of Vision-Language-Action (VLA) models, with the hope that a single generalist model could combine broad semantic understanding with robust manipulation skills.
+
+But training such models is difficult. Robot data is fragmented across platforms, sensors, embodiments, and collection protocols. Heterogeneity appears everywhere: different arm configurations, different action spaces, different camera setups, different visual domains, and different task distributions. These inconsistencies create major distribution shifts that make pretraining unstable and adaptation unreliable.
+
+Inspired by meta-learning and prompt learning, we ask: **"What if a VLA model could learn the structure of each robot and dataset the same way LLMs learn tasks, through prompts?"**
+
+**X-VLA** is a soft-prompted, flow-matching VLA framework that treats each hardware setup as a "task" and encodes it using a small set of learnable embeddings. These **Soft Prompts** capture embodiment and domain-specific variations, guiding the Transformer from the earliest stages of multimodal fusion. With this mechanism, X-VLA can reconcile diverse robot morphologies, data types, and sensor setups within a single unified architecture.
+
+<p align="center">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-architecture.png"
+    alt="XVLA Architecture"
+    style="max-width: 100%; height: auto; width: 800px;"
+  />
+</p>
+
+Built from pure Transformer encoders, X-VLA scales naturally with model size and dataset diversity. Across 6 simulation benchmarks and 3 real robots, Soft Prompts consistently outperform existing methods in handling hardware and domain differences. X-VLA-0.9B, trained on 290K episodes spanning seven robotic platforms, learns an embodiment-agnostic generalist policy in Phase I, and adapts efficiently to new robots in Phase II simply by learning a new set of prompts, while keeping the backbone frozen.
+
+<p align="center">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-architecture2.png"
+    alt="XVLA Architecture 2"
+    style="width: 60%; height: auto;"
+  />
+</p>
+
+With only 1% of parameters tuned (9M), X-VLA-0.9B achieves near-π₀ performance on LIBERO and Simpler-WidowX, despite using **300× fewer trainable parameters**. It also demonstrates strong real-world dexterity with minimal demonstrations, including folding cloths in under two minutes.
+
+<p align="center">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-fold.png"
+    alt="XVLA fold visualization"
+    style="width: 95%; max-width: 1100px; height: auto;"
+  />
+</p>
+
+X-VLA shows that generalist robot intelligence does not require increasingly complex architectures, only the right way to absorb heterogeneity. Soft Prompts offer a simple, scalable mechanism for unifying diverse robotic data, paving the way toward adaptable, cross-embodiment robot foundation models.
+
+## Installation
+
+After installing LeRobot, install the X-VLA dependencies:
+
+```bash
+pip install -e .[xvla]
+```
+
+After the new release, you'll be able to do:
+
+```bash
+pip install lerobot[xvla]
+```
+
+## Quick Start
+
+### Basic Usage
+
+To use X-VLA in your LeRobot configuration, specify the policy type as:
+
+```bash
+policy.type=xvla
+```
+
+### Evaluating Pre-trained Checkpoints
+
+Example evaluation with LIBERO:
+
+```bash
+lerobot-eval \
+  --policy.path="lerobot/xvla-libero" \
+  --env.type=libero \
+  --env.task=libero_spatial,libero_goal,libero_10 \
+  --env.control_mode=absolute \
+  --eval.batch_size=1 \
+  --eval.n_episodes=1 \
+  --env.episode_length=800 \
+  --seed=142
+```
+
+## Available Checkpoints
+
+### 🎯 Base Model
+
+**[lerobot/xvla-base](https://huggingface.co/lerobot/xvla-base)**
+
+A 0.9B parameter instantiation of X-VLA, trained with a carefully designed data processing and learning recipe. The training pipeline consists of two phases:
+
+- **Phase I: Pretraining** - Pretrained on 290K episodes from Droid, Robomind, and Agibot, spanning seven platforms across five types of robotic arms (single-arm to bi-manual setups). By leveraging soft prompts to absorb embodiment-specific variations, the model learns an embodiment-agnostic generalist policy.
+
+- **Phase II: Domain Adaptation** - Adapted to deployable policies for target domains. A new set of soft prompts is introduced and optimized to encode the hardware configuration of the novel domain, while the pretrained backbone remains frozen.
+
+### Simulation Checkpoints
+
+**[lerobot/xvla-libero](https://huggingface.co/lerobot/xvla-libero)**
+
+Achieves 93% success rate on LIBERO benchmarks. Fine-tuned from the base model for simulation tasks.
+
+**[lerobot/xvla-widowx](https://huggingface.co/lerobot/xvla-widowx)**
+
+Fine-tuned on BridgeData for pick-and-place experiments on compact WidowX platforms. Demonstrates robust manipulation capabilities.
+
+### 🤖 Real-World Checkpoints
+
+**[lerobot/xvla-folding](https://huggingface.co/lerobot/xvla-folding)**
+
+A fine-tuned dexterous manipulation model trained on the high-quality Soft-FOLD cloth folding dataset. Achieves 100% success rate over 2 hours of continuous cloth folding.
+
+**[lerobot/xvla-agibot-world](https://huggingface.co/lerobot/xvla-agibot-world)**
+
+Optimized for AgileX robot dexterous manipulation tasks.
+
+**[lerobot/xvla-google-robot](https://huggingface.co/lerobot/xvla-google-robot)**
+
+Adapted for Google Robot platforms.
+
+## Training X-VLA
+
+### Recommended Training Configuration
+
+When fine-tuning X-VLA for a new embodiment or task, we recommend not freezing the VLM, and also setting the `policy.dtype=bfloat16` to not hit OOM errors.
+
+```bash
+lerobot-train \
+  --dataset.repo_id=YOUR_DATASET \
+  --output_dir=./outputs/xvla_training \
+  --job_name=xvla_training \
+  --policy.path="lerobot/xvla-base" \
+  --policy.repo_id="HF_USER/xvla-your-robot" \
+  --policy.dtype=bfloat16 \
+  --policy.action_mode=auto \
+  --steps=20000 \
+  --policy.device=cuda \
+  --policy.freeze_vision_encoder=false \
+  --policy.freeze_language_encoder=false \
+  --policy.train_policy_transformer=true \
+  --policy.train_soft_prompts=true \
+```
+
+### Training Parameters Explained
+
+| Parameter                  | Default | Description                                    |
+| -------------------------- | ------- | ---------------------------------------------- |
+| `freeze_vision_encoder`    | `false` | Do not freeze the VLM vision encoder weights   |
+| `freeze_language_encoder`  | `false` | Do not freeze the VLM language encoder weights |
+| `train_policy_transformer` | `true`  | Allow policy transformer layers to train       |
+| `train_soft_prompts`       | `true`  | Allow soft prompts to train                    |
+
+**💡 Best Practice**: For Phase II adaptation to new embodiments, do not freeze the VLM encoders and also train the policy transformer and soft prompts.
+
+### Example: Training on Bimanual Robot
+
+```bash
+lerobot-train \
+  --dataset.repo_id=pepijn223/bimanual-so100-handover-cube \
+  --output_dir=./outputs/xvla_bimanual \
+  --job_name=xvla_so101_training \
+  --policy.path="lerobot/xvla-base" \
+  --policy.dtype=bfloat16 \
+  --policy.repo_id="YOUR_USERNAME/xvla-biso101" \
+  --steps=3000 \
+  --policy.device=cuda \
+  --policy.action_mode=so101_bimanual \
+  --policy.freeze_vision_encoder=false \
+  --policy.freeze_language_encoder=false \
+  --policy.train_policy_transformer=true \
+  --policy.train_soft_prompts=true
+```
+
+💡 **Best Performance:** If you have sufficient computational resources and want to achieve best X-VLA finetuning performance, you should follow the official finetuning strategy:
+
+**🔥 Full-finetune all components with a custom learning-rate scheme**
+
+To ensure stable optimization, the Vision-Language Model (VLM) must be trained with only 1/10 of the base learning rate, while all other components use the full LR.
+This LR ratio is crucial for achieving strong and stable finetuning performance. This is already done for you by default.
+❕Note
+
+Completely matching the official reported performance may require an additional warm-up LR schedule for soft-prompts, which can bring minor improvements.
+We encourage implementing this in your customized training pipeline for optimal results.
+
+## Core Concepts
+
+### 1. Action Modes
+
+X-VLA uses an **Action Registry** system to handle different action spaces and embodiments. The `action_mode` parameter defines how actions are processed, what loss functions are used, and how predictions are post-processed.
+
+#### Available Action Modes
+
+| Action Mode      | Action Dim              | Description                                 | Use Case                             |
+| ---------------- | ----------------------- | ------------------------------------------- | ------------------------------------ |
+| `ee6d`           | 20                      | End-effector with xyz, 6D rotation, gripper | Dual-arm setups with spatial control |
+| `joint`          | 14                      | Joint-space with gripper                    | Direct joint control robots          |
+| `agibot_ee6d`    | 20                      | AGI-bot variant with MSE loss               | AGI-bot platforms                    |
+| `so101_bimanual` | 20 (model), 12 (real)   | SO101 bimanual robot                        | Bimanual manipulation tasks          |
+| `auto`           | 20 (model), auto (real) | Auto-detects action dim from dataset        | **Recommended** for new robots       |
+
+#### Why Action Modes Matter
+
+When you have a pretrained checkpoint like `lerobot/xvla-base` trained with `action_dim=20`, and you want to train on a dataset with a different action dimension (e.g., 14 for bimanual arms), you can't simply trim the action dimension. The action mode orchestrates:
+
+1. **Loss Computation**: Different loss functions for different action components (MSE for joints, BCE for grippers, etc.)
+2. **Preprocessing**: Zeroing out gripper channels, padding dimensions
+3. **Postprocessing**: Applying sigmoid to gripper logits, trimming padding
+
+#### Example: BimanualSO101 Action Space
+
+The `so101_bimanual` action mode handles the mismatch between model output (20D) and real robot control (12D):
+
+```python
+# Model outputs 20 dimensions for compatibility
+dim_action = 20
+
+# Real robot only needs 12 dimensions
+# [left_arm (6), right_arm (6)] = [joints (5) + gripper (1)] × 2
+REAL_DIM = 12
+
+# Preprocessing: Pad 12D actions to 20D for training
+# Postprocessing: Trim 20D predictions to 12D for deployment
+```
+
+See the [action_hub.py](/home/jade_choghari/robot/lerobot/src/lerobot/policies/xvla/action_hub.py) implementation for details.
+
+#### Auto Action Mode (Recommended)
+
+The `auto` action mode is the easiest way to use X-VLA with any robot. It automatically detects your dataset's action dimension and handles padding/trimming:
+
+```bash
+lerobot-train \
+  --policy.path="lerobot/xvla-base" \
+  --policy.action_mode=auto \
+  --policy.max_action_dim=20 \
+  ...
+```
+
+**How it works:**
+
+- Reads `action_feature.shape[-1]` from your dataset (e.g., 7 for Franka)
+- Model outputs `max_action_dim` (default 20) for pretrained compatibility
+- Loss is computed **only on the real dimensions**: `MSE(pred[:,:,:real_dim], target[:,:,:real_dim])`
+- Postprocess trims output back to `real_dim` for robot control
+
+This eliminates the need to create custom action modes for most robots.
+
+### 2. Domain IDs
+
+Domain IDs are learnable identifiers for different robot configurations and camera setups. They allow X-VLA to distinguish between:
+
+- Different robots (Robot 1 vs Robot 2)
+- Different camera configurations (cam1 vs cam2)
+- Different combinations (Robot1-cam1-cam2 vs Robot1-cam1 vs Robot2-cam1)
+
+#### Setting Domain IDs
+
+**During Training**: By default, domain_id is set to 0 for general training.
+
+**During Evaluation**: Specify the domain_id that matches your checkpoint's training configuration.
+
+```python
+# Example: LIBERO checkpoint uses domain_id=3
+domain_id = 3
+```
+
+The domain_id is automatically added to observations by the `XVLAAddDomainIdProcessorStep` in the preprocessing pipeline.
+
+The `lerobot/xvla-base` model has been trained on the following domain IDs. It is recommended to choose one that most resembles your robot/configuration:
+
+#### Fine-tuning Datasets
+
+| Dataset Name     | Domain ID |
+| ---------------- | --------- |
+| Bridge           | 0         |
+| RT1              | 1         |
+| Calvin           | 2         |
+| libero           | 3         |
+| widowx-air       | 4         |
+| AIR-AGILEX-HQ    | 5         |
+| robotwin2_abs_ee | 6         |
+| robotwin2_clean  | 6         |
+| robocasa-human   | 7         |
+| VLABench         | 8         |
+| AGIBOT-challenge | 9         |
+| AIR-AGILEX       | 10        |
+| AIRBOT           | 18        |
+
+### 3. Processor Steps
+
+X-VLA requires specific preprocessing and postprocessing steps for proper operation.
+
+#### Required Preprocessing Steps
+
+1. **XVLAImageToFloatProcessorStep**: Converts images from [0, 255] to [0, 1] range
+2. **XVLAImageNetNormalizeProcessorStep**: Applies ImageNet normalization (required for VLM backbone)
+3. **XVLAAddDomainIdProcessorStep**: Adds domain_id to observations
+
+#### Example Custom Processor
+
+For LIBERO environments, a custom processor handles the specific observation format:
+
+```python
+from lerobot.policies.xvla.processor_xvla import LiberoProcessorStep
+
+processor = LiberoProcessorStep()
+# Handles robot_state dictionary, converts rotation matrices to 6D representation
+# Applies 180° image rotation for camera convention
+```
+
+### 4. Configuration Parameters
+
+Key configuration parameters for X-VLA:
+
+```python
+# Observation and action
+n_obs_steps: int = 1          # Number of observation timesteps
+chunk_size: int = 32           # Action sequence length
+n_action_steps: int = 32       # Number of action steps to execute
+
+# Model architecture
+hidden_size: int = 1024        # Transformer hidden dimension
+depth: int = 24                # Number of transformer layers
+num_heads: int = 16            # Number of attention heads
+num_domains: int = 30          # Maximum number of domain IDs
+len_soft_prompts: int = 32     # Length of soft prompt embeddings
+
+# Action space
+action_mode: str = "ee6d"      # Action space type (use "auto" for auto-detection)
+use_proprio: bool = True       # Use proprioceptive state
+max_state_dim: int = 32        # Maximum state dimension
+max_action_dim: int = 20       # Max action dim for padding (used by "auto" mode)
+
+# Vision
+num_image_views: int | None    # Number of camera views
+resize_imgs_with_padding: tuple[int, int] | None  # Target image size with padding
+
+# Training
+num_denoising_steps: int = 10  # Flow matching denoising steps
+```
+
+## Creating Custom Action Modes
+
+If your robot has a unique action space, you can create a custom action mode:
+
+### Step 1: Define Your Action Space
+
+```python
+from lerobot.policies.xvla.action_hub import BaseActionSpace, register_action
+import torch.nn as nn
+
+@register_action("my_custom_robot")
+class MyCustomActionSpace(BaseActionSpace):
+    """Custom action space for my robot."""
+
+    dim_action = 15  # Your robot's action dimension
+    gripper_idx = (7, 14)  # Gripper channel indices
+
+    def __init__(self):
+        super().__init__()
+        self.mse = nn.MSELoss()
+        self.bce = nn.BCEWithLogitsLoss()
+
+    def compute_loss(self, pred, target):
+        """Define your loss computation."""
+        # Example: MSE for joints, BCE for grippers
+        joints_loss = self.mse(pred[:, :, :7], target[:, :, :7])
+        gripper_loss = self.bce(pred[:, :, self.gripper_idx],
+                                target[:, :, self.gripper_idx])
+
+        return {
+            "joints_loss": joints_loss,
+            "gripper_loss": gripper_loss,
+        }
+
+    def preprocess(self, proprio, action, mode="train"):
+        """Preprocess actions before training."""
+        # Example: Zero out grippers in proprioception
+        proprio_m = proprio.clone()
+        action_m = action.clone() if action is not None else None
+        proprio_m[..., self.gripper_idx] = 0.0
+        if action_m is not None:
+            action_m[..., self.gripper_idx] = 0.0
+        return proprio_m, action_m
+
+    def postprocess(self, action):
+        """Post-process predictions for deployment."""
+        # Example: Apply sigmoid to gripper logits
+        action[..., self.gripper_idx] = torch.sigmoid(action[..., self.gripper_idx])
+        return action
+```
+
+### Step 2: Use Your Custom Action Mode
+
+```bash
+lerobot-train \
+  --policy.action_mode=my_custom_robot \
+  --dataset.repo_id=YOUR_DATASET \
+  --policy.path="lerobot/xvla-base" \
+  ...
+```
+
+## Advanced Topics
+
+### Multi-Camera Support
+
+X-VLA supports multiple camera views through the `num_image_views` parameter:
+
+```python
+# Configure for 3 camera views
+policy.num_image_views=3
+
+# Add empty cameras if you have fewer physical cameras
+policy.empty_cameras=1  # Adds 1 zero-padded camera view
+```
+
+### Custom Preprocessing Pipeline
+
+Create a custom preprocessing pipeline for your environment:
+
+```python
+from lerobot.processor import PolicyProcessorPipeline
+from lerobot.policies.xvla.processor_xvla import (
+    XVLAImageToFloatProcessorStep,
+    XVLAImageNetNormalizeProcessorStep,
+    XVLAAddDomainIdProcessorStep,
+)
+
+# Build custom pipeline
+preprocessor = PolicyProcessorPipeline(
+    steps=[
+        YourCustomProcessorStep(),  # Your custom processing
+        XVLAImageToFloatProcessorStep(),  # Required: convert to float
+        XVLAImageNetNormalizeProcessorStep(),  # Required: ImageNet norm
+        XVLAAddDomainIdProcessorStep(domain_id=5),  # Your domain ID
+    ]
+)
+```
+
+### Handling Different Action Dimensions
+
+When your dataset has fewer action dimensions than the pretrained model:
+
+**Option 1 (Recommended)**: Use `auto` action mode
+
+```bash
+# Automatically detects your dataset's action dimension
+# Works with any robot without custom code
+policy.action_mode=auto
+policy.max_action_dim=20  # Match pretrained model
+```
+
+**Option 2**: Use a predefined action mode with built-in padding
+
+```python
+# Model expects 20D, dataset has 12D
+# Action mode handles padding internally
+action_mode = "so101_bimanual"  # Pads 12 → 20
+```
+
+**Option 2**: Create a custom action mode that maps dimensions explicitly
+
+```python
+@register_action("my_mapped_action")
+class MappedActionSpace(BaseActionSpace):
+    dim_action = 20
+    REAL_DIM = 12
+
+    def _pad_to_model_dim(self, x):
+        # Custom padding logic
+        ...
+```
+
+## Troubleshooting
+
+### Common Issues
+
+**Issue**: "Action dimension mismatch"
+
+- **Solution**: Check that your `action_mode` matches your robot's action space. Create a custom action mode if needed.
+
+**Issue**: "Image values outside [0, 1] range"
+
+- **Solution**: Ensure images are preprocessed with `XVLAImageToFloatProcessorStep` before normalization.
+
+**Issue**: "Domain ID not found"
+
+- **Solution**: Make sure `XVLAAddDomainIdProcessorStep` is in your preprocessing pipeline with the correct domain_id.
+
+**Issue**: "Low success rate on new embodiment"
+
+- **Solution**:
+  1. Verify your action_mode is correct
+  2. Check that soft prompts are being trained (`train_soft_prompts=True`)
+  3. Ensure proper preprocessing (ImageNet normalization, domain_id)
+  4. Consider increasing training steps
+
+**Issue**: "Out of memory during training"
+
+- **Solution**:
+  1. Reduce `chunk_size` (e.g., from 32 to 16)
+  2. Enable gradient checkpointing
+  3. Reduce batch size
+  4. Freeze more components
+
+## Citation
+
+If you use X-VLA in your research, please cite:
+
+```bibtex
+@article{zheng2025x,
+  title   = {X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model},
+  author  = {Zheng, Jinliang and Li, Jianxiong and Wang, Zhihao and Liu, Dongxiu and Kang, Xirui
+             and Feng, Yuchun and Zheng, Yinan and Zou, Jiayin and Chen, Yilun and Zeng, Jia and others},
+  journal = {arXiv preprint arXiv:2510.10274},
+  year    = {2025}
+}
+```
+
+## Additional Resources
+
+- [X-VLA Paper](https://arxiv.org/pdf/2510.10274)
+- [LeRobot Documentation](https://github.com/huggingface/lerobot)
+- [Action Registry Implementation](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/action_hub.py)
+- [Processor Implementation](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/processor_xvla.py)
+- [Model Configuration](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/configuration_xvla.py)
+
+## Contributing
+
+We welcome contributions! If you've implemented a new action mode or processor for your robot, please consider submitting a PR to help the community.

examples/backward_compatibility/replay.py

@@ -45,7 +45,7 @@ from lerobot.robots import (  # noqa: F401
     so101_follower,
 )
 from lerobot.utils.constants import ACTION
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import (
     init_logging,
     log_say,
@@ -97,7 +97,7 @@ def replay(cfg: ReplayConfig):
         robot.send_action(action)
 
         dt_s = time.perf_counter() - start_episode_t
-        busy_wait(1 / dataset.fps - dt_s)
+        precise_sleep(1 / dataset.fps - dt_s)
 
     robot.disconnect()
 

examples/lekiwi/replay.py

@@ -20,7 +20,7 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
 from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient
 from lerobot.utils.constants import ACTION
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
 EPISODE_IDX = 0
@@ -58,7 +58,7 @@ def main():
         # Send action to robot
         _ = robot.send_action(action)
 
-        busy_wait(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 
     robot.disconnect()
 

examples/lekiwi/teleoperate.py

@@ -19,7 +19,7 @@ import time
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.teleoperators.keyboard.teleop_keyboard import KeyboardTeleop, KeyboardTeleopConfig
 from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 FPS = 30
@@ -71,7 +71,7 @@ def main():
         # Visualize
         log_rerun_data(observation=observation, action=action)
 
-        busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
 
 
 if __name__ == "__main__":

examples/phone_to_so100/replay.py

@@ -29,7 +29,7 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.utils.constants import ACTION
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
 EPISODE_IDX = 0
@@ -96,7 +96,7 @@ def main():
         # Send action to robot
         _ = robot.send_action(joint_action)
 
-        busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
 
     # Clean up
     robot.disconnect()

examples/phone_to_so100/teleoperate.py

@@ -32,7 +32,7 @@ from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.teleoperators.phone.teleop_phone import Phone
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 FPS = 30
@@ -114,7 +114,7 @@ def main():
         # Visualize
         log_rerun_data(observation=phone_obs, action=joint_action)
 
-        busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
 
 
 if __name__ == "__main__":

examples/so100_to_so100_EE/replay.py

@@ -30,7 +30,7 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.utils.constants import ACTION
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
 EPISODE_IDX = 0
@@ -97,7 +97,7 @@ def main():
         # Send action to robot
         _ = robot.send_action(joint_action)
 
-        busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
 
     # Clean up
     robot.disconnect()

examples/so100_to_so100_EE/teleoperate.py

@@ -32,7 +32,7 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
 from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 FPS = 30
@@ -120,7 +120,7 @@ def main():
         # Visualize
         log_rerun_data(observation=leader_ee_act, action=follower_joints_act)
 
-        busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
 
 
 if __name__ == "__main__":

examples/unitree_g1/gr00t_locomotion.py

@@ -0,0 +1,347 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Example: GR00T Locomotion with Pre-loaded Policies
+
+This example demonstrates the NEW pattern for loading GR00T policies externally
+and passing them to the robot class.
+"""
+
+import argparse
+import logging
+import threading
+import time
+from collections import deque
+
+import numpy as np
+import onnxruntime as ort
+from huggingface_hub import hf_hub_download
+
+from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
+from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
+
+logger = logging.getLogger(__name__)
+
+GROOT_DEFAULT_ANGLES = np.zeros(29, dtype=np.float32)
+GROOT_DEFAULT_ANGLES[[0, 6]] = -0.1  # hip pitch
+GROOT_DEFAULT_ANGLES[[3, 9]] = 0.3  # knee
+GROOT_DEFAULT_ANGLES[[4, 10]] = -0.2  # ankle pitch
+
+MISSING_JOINTS = []
+G1_MODEL = "g1_23"  # or "g1_29"
+if G1_MODEL == "g1_23":
+    MISSING_JOINTS = [12, 14, 20, 21, 27, 28]  # waist yaw/pitch, wrist pitch/yaw
+
+LOCOMOTION_ACTION_SCALE = 0.25
+
+LOCOMOTION_CONTROL_DT = 0.02
+
+ANG_VEL_SCALE: float = 0.25
+DOF_POS_SCALE: float = 1.0
+DOF_VEL_SCALE: float = 0.05
+CMD_SCALE: list = [2.0, 2.0, 0.25]
+
+
+DEFAULT_GROOT_REPO_ID = "nepyope/GR00T-WholeBodyControl_g1"
+
+
+def load_groot_policies(
+    repo_id: str = DEFAULT_GROOT_REPO_ID,
+) -> tuple[ort.InferenceSession, ort.InferenceSession]:
+    """Load GR00T dual-policy system (Balance + Walk) from Hugging Face Hub.
+
+    Args:
+        repo_id: Hugging Face Hub repository ID containing the ONNX policies.
+    """
+    logger.info(f"Loading GR00T dual-policy system from Hugging Face Hub ({repo_id})...")
+
+    # Download ONNX policies from Hugging Face Hub
+    balance_path = hf_hub_download(
+        repo_id=repo_id,
+        filename="GR00T-WholeBodyControl-Balance.onnx",
+    )
+    walk_path = hf_hub_download(
+        repo_id=repo_id,
+        filename="GR00T-WholeBodyControl-Walk.onnx",
+    )
+
+    # Load ONNX policies
+    policy_balance = ort.InferenceSession(balance_path)
+    policy_walk = ort.InferenceSession(walk_path)
+
+    logger.info("GR00T policies loaded successfully")
+
+    return policy_balance, policy_walk
+
+
+class GrootLocomotionController:
+    """
+    Handles GR00T-style locomotion control for the Unitree G1 robot.
+
+    This controller manages:
+    - Dual-policy system (Balance + Walk)
+    - 29-joint observation processing
+    - 15D action output (legs + waist)
+    - Policy inference and motor command generation
+    """
+
+    def __init__(self, policy_balance, policy_walk, robot, config):
+        self.policy_balance = policy_balance
+        self.policy_walk = policy_walk
+        self.robot = robot
+        self.config = config
+
+        self.locomotion_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)  # vx, vy, theta_dot
+
+        # GR00T-specific state
+        self.groot_qj_all = np.zeros(29, dtype=np.float32)
+        self.groot_dqj_all = np.zeros(29, dtype=np.float32)
+        self.groot_action = np.zeros(15, dtype=np.float32)
+        self.groot_obs_single = np.zeros(86, dtype=np.float32)
+        self.groot_obs_history = deque(maxlen=6)
+        self.groot_obs_stacked = np.zeros(516, dtype=np.float32)
+        self.groot_height_cmd = 0.74  # Default base height
+        self.groot_orientation_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)
+
+        # input to gr00t is 6 frames (6*86D=516)
+        for _ in range(6):
+            self.groot_obs_history.append(np.zeros(86, dtype=np.float32))
+
+        # Thread management
+        self.locomotion_running = False
+        self.locomotion_thread = None
+
+        logger.info("GrootLocomotionController initialized")
+
+    def groot_locomotion_run(self):
+        # get current observation
+        robot_state = self.robot.get_observation()
+
+        if robot_state is None:
+            return
+
+        # get command from remote controller
+        if robot_state.wireless_remote is not None:
+            self.robot.remote_controller.set(robot_state.wireless_remote)
+            if self.robot.remote_controller.button[0]:  # R1 - raise waist
+                self.groot_height_cmd += 0.001
+                self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
+            if self.robot.remote_controller.button[4]:  # R2 - lower waist
+                self.groot_height_cmd -= 0.001
+                self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
+        else:
+            self.robot.remote_controller.lx = 0.0
+            self.robot.remote_controller.ly = 0.0
+            self.robot.remote_controller.rx = 0.0
+            self.robot.remote_controller.ry = 0.0
+
+        self.locomotion_cmd[0] = self.robot.remote_controller.ly  # forward/backward
+        self.locomotion_cmd[1] = self.robot.remote_controller.lx * -1  # left/right
+        self.locomotion_cmd[2] = self.robot.remote_controller.rx * -1  # rotation rate
+
+        for i in range(29):
+            self.groot_qj_all[i] = robot_state.motor_state[i].q
+            self.groot_dqj_all[i] = robot_state.motor_state[i].dq
+
+        # adapt observation for g1_23dof
+        for idx in MISSING_JOINTS:
+            self.groot_qj_all[idx] = 0.0
+            self.groot_dqj_all[idx] = 0.0
+
+        # Scale joint positions and velocities
+        qj_obs = self.groot_qj_all.copy()
+        dqj_obs = self.groot_dqj_all.copy()
+
+        # express imu data in gravity frame of reference
+        quat = robot_state.imu_state.quaternion
+        ang_vel = np.array(robot_state.imu_state.gyroscope, dtype=np.float32)
+        gravity_orientation = self.robot.get_gravity_orientation(quat)
+
+        # scale joint positions and velocities before policy inference
+        qj_obs = (qj_obs - GROOT_DEFAULT_ANGLES) * DOF_POS_SCALE
+        dqj_obs = dqj_obs * DOF_VEL_SCALE
+        ang_vel_scaled = ang_vel * ANG_VEL_SCALE
+
+        # build single frame observation
+        self.groot_obs_single[:3] = self.locomotion_cmd * np.array(CMD_SCALE)
+        self.groot_obs_single[3] = self.groot_height_cmd
+        self.groot_obs_single[4:7] = self.groot_orientation_cmd
+        self.groot_obs_single[7:10] = ang_vel_scaled
+        self.groot_obs_single[10:13] = gravity_orientation
+        self.groot_obs_single[13:42] = qj_obs
+        self.groot_obs_single[42:71] = dqj_obs
+        self.groot_obs_single[71:86] = self.groot_action  # 15D previous actions
+
+        # Add to history and stack observations (6 frames × 86D = 516D)
+        self.groot_obs_history.append(self.groot_obs_single.copy())
+
+        # Stack all 6 frames into 516D vector
+        for i, obs_frame in enumerate(self.groot_obs_history):
+            start_idx = i * 86
+            end_idx = start_idx + 86
+            self.groot_obs_stacked[start_idx:end_idx] = obs_frame
+
+        # Run policy inference (ONNX) with 516D stacked observation
+
+        cmd_magnitude = np.linalg.norm(self.locomotion_cmd)
+
+        selected_policy = (
+            self.policy_balance if cmd_magnitude < 0.05 else self.policy_walk
+        )  # balance/standing policy for small commands, walking policy for movement commands
+
+        # run policy inference
+        ort_inputs = {selected_policy.get_inputs()[0].name: np.expand_dims(self.groot_obs_stacked, axis=0)}
+        ort_outs = selected_policy.run(None, ort_inputs)
+        self.groot_action = ort_outs[0].squeeze()
+
+        # transform action back to target joint positions
+        target_dof_pos_15 = GROOT_DEFAULT_ANGLES[:15] + self.groot_action * LOCOMOTION_ACTION_SCALE
+
+        # command motors
+        for i in range(15):
+            motor_idx = i
+            self.robot.msg.motor_cmd[motor_idx].q = target_dof_pos_15[i]
+            self.robot.msg.motor_cmd[motor_idx].qd = 0
+            self.robot.msg.motor_cmd[motor_idx].kp = self.robot.kp[motor_idx]
+            self.robot.msg.motor_cmd[motor_idx].kd = self.robot.kd[motor_idx]
+            self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+        # adapt action for g1_23dof
+        for joint_idx in MISSING_JOINTS:
+            self.robot.msg.motor_cmd[joint_idx].q = 0.0
+            self.robot.msg.motor_cmd[joint_idx].qd = 0
+            self.robot.msg.motor_cmd[joint_idx].kp = self.robot.kp[joint_idx]
+            self.robot.msg.motor_cmd[joint_idx].kd = self.robot.kd[joint_idx]
+            self.robot.msg.motor_cmd[joint_idx].tau = 0
+
+        # send action to robot
+        self.robot.send_action(self.robot.msg)
+
+    def _locomotion_thread_loop(self):
+        """Background thread that runs the locomotion policy at specified rate."""
+        logger.info("Locomotion thread started")
+        while self.locomotion_running:
+            start_time = time.time()
+            try:
+                self.groot_locomotion_run()
+            except Exception as e:
+                logger.error(f"Error in locomotion loop: {e}")
+
+            # Sleep to maintain control rate
+            elapsed = time.time() - start_time
+            sleep_time = max(0, LOCOMOTION_CONTROL_DT - elapsed)
+            time.sleep(sleep_time)
+        logger.info("Locomotion thread stopped")
+
+    def start_locomotion_thread(self):
+        if self.locomotion_running:
+            logger.warning("Locomotion thread already running")
+            return
+
+        logger.info("Starting locomotion control thread...")
+        self.locomotion_running = True
+        self.locomotion_thread = threading.Thread(target=self._locomotion_thread_loop, daemon=True)
+        self.locomotion_thread.start()
+
+        logger.info("Locomotion control thread started!")
+
+    def stop_locomotion_thread(self):
+        if not self.locomotion_running:
+            return
+
+        logger.info("Stopping locomotion control thread...")
+        self.locomotion_running = False
+        if self.locomotion_thread:
+            self.locomotion_thread.join(timeout=2.0)
+        logger.info("Locomotion control thread stopped")
+
+    def reset_robot(self):
+        """Move robot legs to default standing position over 2 seconds (arms are not moved)."""
+        total_time = 3.0
+        num_step = int(total_time / self.robot.control_dt)
+
+        # Only control legs, not arms (first 12 joints)
+        default_pos = GROOT_DEFAULT_ANGLES  # First 12 values are leg angles
+        dof_size = len(default_pos)
+
+        # Get current lowstate
+        robot_state = self.robot.get_observation()
+
+        # Record the current leg positions
+        init_dof_pos = np.zeros(dof_size, dtype=np.float32)
+        for i in range(dof_size):
+            init_dof_pos[i] = robot_state.motor_state[i].q
+
+        # Move legs to default pos
+        for i in range(num_step):
+            alpha = i / num_step
+            for motor_idx in range(dof_size):
+                target_pos = default_pos[motor_idx]
+                self.robot.msg.motor_cmd[motor_idx].q = (
+                    init_dof_pos[motor_idx] * (1 - alpha) + target_pos * alpha
+                )
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = self.robot.kp[motor_idx]
+                self.robot.msg.motor_cmd[motor_idx].kd = self.robot.kd[motor_idx]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+            self.robot.msg.crc = self.robot.crc.Crc(self.robot.msg)
+            self.robot.lowcmd_publisher.Write(self.robot.msg)
+            time.sleep(self.robot.control_dt)
+        logger.info("Reached default position (legs only)")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="GR00T Locomotion Controller for Unitree G1")
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        default=DEFAULT_GROOT_REPO_ID,
+        help=f"Hugging Face Hub repo ID for GR00T policies (default: {DEFAULT_GROOT_REPO_ID})",
+    )
+    args = parser.parse_args()
+
+    # load policies
+    policy_balance, policy_walk = load_groot_policies(repo_id=args.repo_id)
+
+    # initialize robot
+    config = UnitreeG1Config()
+    robot = UnitreeG1(config)
+
+    # initialize gr00t locomotion controller
+    groot_controller = GrootLocomotionController(
+        policy_balance=policy_balance,
+        policy_walk=policy_walk,
+        robot=robot,
+        config=config,
+    )
+
+    # reset legs and start locomotion thread
+    try:
+        groot_controller.reset_robot()
+        groot_controller.start_locomotion_thread()
+
+        # log status
+        logger.info("Robot initialized with GR00T locomotion policies")
+        logger.info("Locomotion controller running in background thread")
+        logger.info("Press Ctrl+C to stop")
+
+        # keep robot alive
+        while True:
+            time.sleep(1.0)
+    except KeyboardInterrupt:
+        print("\nStopping locomotion...")
+        groot_controller.stop_locomotion_thread()
+        print("Done!")

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.2"
+version = "0.4.3"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -96,7 +96,7 @@ dependencies = [
 # Common
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.10.0"]
-transformers-dep = ["transformers>=4.53.0,<5.0.0"]
+transformers-dep = ["transformers>=4.57.1,<5.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"] # TODO: Bumb dependency (compatible with wandb)
 
 # Motors
@@ -107,6 +107,10 @@ dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
+unitree_g1 = [
+    "pyzmq>=26.2.1,<28.0.0",
+    "onnxruntime>=1.16.0"
+]
 reachy2 = ["reachy2_sdk>=1.0.14,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
@@ -116,6 +120,13 @@ intelrealsense = [
 phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
 
 # Policies
+wallx = [
+    "transformers==4.49.0",
+    "peft==0.17.1",
+    "scipy==1.15.3",
+    "torchdiffeq==0.2.5",
+    "qwen_vl_utils==0.0.11"
+]
 pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0", "safetensors>=0.4.3,<1.0.0"]
 groot = [
@@ -129,13 +140,15 @@ groot = [
     "ninja>=1.11.1,<2.0.0",
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
+sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "matplotlib>=3.10.3,<4.0.0", "qwen-vl-utils>=0.0.14"]
+xvla = ["lerobot[transformers-dep]"]
 hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # Features
 async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
 
 # Development
-dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1"]
+dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1"]
 test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
@@ -154,9 +167,11 @@ all = [
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
-    "lerobot[pi]",
+    # "lerobot[wallx]",
+    # "lerobot[pi]", TODO(Pepijn): Update pi to transformers v5
     "lerobot[smolvla]",
     # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
+    "lerobot[xvla]",
     "lerobot[hilserl]",
     "lerobot[async]",
     "lerobot[dev]",
@@ -167,6 +182,7 @@ all = [
     "lerobot[phone]",
     "lerobot[libero]",
     "lerobot[metaworld]",
+    "lerobot[sarm]"
 ]
 
 [project.scripts]
@@ -221,6 +237,7 @@ ignore = [
 
 [tool.ruff.lint.per-file-ignores]
 "__init__.py" = ["F401", "F403"]
+"src/lerobot/policies/wall_x/**" = ["N801", "N812", "SIM102", "SIM108", "SIM210", "SIM211", "B006", "B007", "SIM118"] # Supprese these as they are coming from original Qwen2_5_vl code TODO(pepijn): refactor original
 
 [tool.ruff.lint.isort]
 combine-as-imports = true
@@ -257,6 +274,7 @@ default.extend-ignore-identifiers-re = [
     "ein",
     "thw",
     "inpt",
+    "ROBOTIS",
 ]
 
 # TODO: Uncomment when ready to use
@@ -311,9 +329,9 @@ disallow_untyped_defs = true
 disallow_incomplete_defs = true
 check_untyped_defs = true
 
-# [[tool.mypy.overrides]]
-# module = "lerobot.optim.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.optim.*"
+ignore_errors = false
 
 [[tool.mypy.overrides]]
 module = "lerobot.model.*"
@@ -356,10 +374,84 @@ ignore_errors = false
 # module = "lerobot.async_inference.*"
 # ignore_errors = false
 
-# [[tool.mypy.overrides]]
-# module = "lerobot.transport.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.transport.*"
+ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.scripts.*"
 # ignore_errors = false
+
+[tool.uv]
+# wallx requires transformers==4.49.0 which conflicts with other extras that need >=4.53.0
+conflicts = [
+    [
+        { extra = "wallx" },
+        { extra = "transformers-dep" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "pi" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "smolvla" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "groot" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "xvla" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "sarm" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "hilserl" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "libero" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "all" },
+    ],
+    # pi uses custom branch which conflicts with transformers-dep
+    [
+        { extra = "pi" },
+        { extra = "transformers-dep" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "smolvla" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "groot" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "xvla" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "sarm" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "hilserl" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "libero" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "all" },
+    ],
+]

src/lerobot/async_inference/constants.py

@@ -26,4 +26,4 @@ DEFAULT_OBS_QUEUE_TIMEOUT = 2
 SUPPORTED_POLICIES = ["act", "smolvla", "diffusion", "tdmpc", "vqbet", "pi0", "pi05"]
 
 # TODO: Add all other robots
-SUPPORTED_ROBOTS = ["so100_follower", "so101_follower", "bi_so100_follower"]
+SUPPORTED_ROBOTS = ["so100_follower", "so101_follower", "bi_so100_follower", "omx_follower"]

src/lerobot/async_inference/robot_client.py

@@ -54,6 +54,7 @@ from lerobot.robots import (  # noqa: F401
     bi_so100_follower,
     koch_follower,
     make_robot_from_config,
+    omx_follower,
     so100_follower,
     so101_follower,
 )

src/lerobot/configs/train.py

@@ -56,6 +56,7 @@ class TrainPipelineConfig(HubMixin):
     steps: int = 100_000
     eval_freq: int = 20_000
     log_freq: int = 200
+    tolerance_s: float = 1e-4
     save_checkpoint: bool = True
     # Checkpoint is saved every `save_freq` training iterations and after the last training step.
     save_freq: int = 20_000
@@ -64,9 +65,17 @@ class TrainPipelineConfig(HubMixin):
     scheduler: LRSchedulerConfig | None = None
     eval: EvalConfig = field(default_factory=EvalConfig)
     wandb: WandBConfig = field(default_factory=WandBConfig)
-    checkpoint_path: Path | None = field(init=False, default=None)
+
+    # RA-BC (Reward-Aligned Behavior Cloning) parameters
+    use_rabc: bool = False  # Enable reward-weighted training
+    rabc_progress_path: str | None = None  # Path to precomputed SARM progress parquet file
+    rabc_kappa: float = 0.01  # Hard threshold for high-quality samples
+    rabc_epsilon: float = 1e-6  # Small constant for numerical stability
+    rabc_head_mode: str | None = "sparse"  # For dual-head models: "sparse" or "dense"
+
     # Rename map for the observation to override the image and state keys
     rename_map: dict[str, str] = field(default_factory=dict)
+    checkpoint_path: Path | None = field(init=False, default=None)
 
     def validate(self) -> None:
         # HACK: We parse again the cli args here to get the pretrained paths if there was some.
@@ -130,6 +139,14 @@ class TrainPipelineConfig(HubMixin):
                 "'policy.repo_id' argument missing. Please specify it to push the model to the hub."
             )
 
+        if self.use_rabc and not self.rabc_progress_path:
+            # Auto-detect from dataset path
+            repo_id = self.dataset.repo_id
+            if self.dataset.root:
+                self.rabc_progress_path = str(Path(self.dataset.root) / "sarm_progress.parquet")
+            else:
+                self.rabc_progress_path = f"hf://datasets/{repo_id}/sarm_progress.parquet"
+
     @classmethod
     def __get_path_fields__(cls) -> list[str]:
         """This enables the parser to load config from the policy using `--policy.path=local/dir`"""

src/lerobot/data_processing/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

src/lerobot/data_processing/sarm_annotations/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

src/lerobot/datasets/aggregate.py

@@ -136,21 +136,40 @@ def update_meta_data(
         df["_orig_chunk"] = df[orig_chunk_col].copy()
         df["_orig_file"] = df[orig_file_col].copy()
 
-        # Update chunk and file indices to point to destination
-        df[orig_chunk_col] = video_idx["chunk"]
-        df[orig_file_col] = video_idx["file"]
-
-        # Apply per-source-file timestamp offsets
+        # Get mappings for this video key
         src_to_offset = video_idx.get("src_to_offset", {})
-        if src_to_offset:
-            # Apply offset based on original source file
+        src_to_dst = video_idx.get("src_to_dst", {})
+
+        # Apply per-source-file mappings
+        if src_to_dst:
+            # Map each episode to its correct destination file and apply offset
             for idx in df.index:
-                src_key = (df.at[idx, "_orig_chunk"], df.at[idx, "_orig_file"])
+                # Convert to Python int to avoid numpy type mismatch in dict lookup
+                src_key = (int(df.at[idx, "_orig_chunk"]), int(df.at[idx, "_orig_file"]))
+
+                # Get destination chunk/file for this source file
+                dst_chunk, dst_file = src_to_dst.get(src_key, (video_idx["chunk"], video_idx["file"]))
+                df.at[idx, orig_chunk_col] = dst_chunk
+                df.at[idx, orig_file_col] = dst_file
+
+                # Apply timestamp offset
+                offset = src_to_offset.get(src_key, 0)
+                df.at[idx, f"videos/{key}/from_timestamp"] += offset
+                df.at[idx, f"videos/{key}/to_timestamp"] += offset
+        elif src_to_offset:
+            # Fallback: use same destination for all, but apply per-file offsets
+            df[orig_chunk_col] = video_idx["chunk"]
+            df[orig_file_col] = video_idx["file"]
+            for idx in df.index:
+                # Convert to Python int to avoid numpy type mismatch in dict lookup
+                src_key = (int(df.at[idx, "_orig_chunk"]), int(df.at[idx, "_orig_file"]))
                 offset = src_to_offset.get(src_key, 0)
                 df.at[idx, f"videos/{key}/from_timestamp"] += offset
                 df.at[idx, f"videos/{key}/to_timestamp"] += offset
         else:
             # Fallback to simple offset (for backward compatibility)
+            df[orig_chunk_col] = video_idx["chunk"]
+            df[orig_file_col] = video_idx["file"]
             df[f"videos/{key}/from_timestamp"] = (
                 df[f"videos/{key}/from_timestamp"] + video_idx["latest_duration"]
             )
@@ -268,6 +287,12 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
         videos_idx[key]["episode_duration"] = 0
         # Track offset for each source (chunk, file) pair
         videos_idx[key]["src_to_offset"] = {}
+        # Track destination (chunk, file) for each source (chunk, file) pair
+        videos_idx[key]["src_to_dst"] = {}
+        # Initialize dst_file_durations if not present
+        # dst_file_durations tracks duration of each destination file
+        if "dst_file_durations" not in videos_idx[key]:
+            videos_idx[key]["dst_file_durations"] = {}
 
     for key, video_idx in videos_idx.items():
         unique_chunk_file_pairs = {
@@ -282,9 +307,13 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
 
         chunk_idx = video_idx["chunk"]
         file_idx = video_idx["file"]
-        current_offset = video_idx["latest_duration"]
+        dst_file_durations = video_idx["dst_file_durations"]
 
         for src_chunk_idx, src_file_idx in unique_chunk_file_pairs:
+            # Convert to Python int to ensure consistent dict keys
+            src_chunk_idx = int(src_chunk_idx)
+            src_file_idx = int(src_file_idx)
+
             src_path = src_meta.root / DEFAULT_VIDEO_PATH.format(
                 video_key=key,
                 chunk_index=src_chunk_idx,
@@ -298,14 +327,17 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
             )
 
             src_duration = get_video_duration_in_s(src_path)
+            dst_key = (chunk_idx, file_idx)
 
             if not dst_path.exists():
-                # Store offset before incrementing
-                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_offset
+                # New destination file: offset is 0
+                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = 0
+                videos_idx[key]["src_to_dst"][(src_chunk_idx, src_file_idx)] = dst_key
                 dst_path.parent.mkdir(parents=True, exist_ok=True)
                 shutil.copy(str(src_path), str(dst_path))
+                # Track duration of this destination file
+                dst_file_durations[dst_key] = src_duration
                 videos_idx[key]["episode_duration"] += src_duration
-                current_offset += src_duration
                 continue
 
             # Check file sizes before appending
@@ -313,10 +345,11 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
             dst_size = get_file_size_in_mb(dst_path)
 
             if dst_size + src_size >= video_files_size_in_mb:
-                # Rotate to a new file, this source becomes start of new destination
-                # So its offset should be 0
-                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = 0
+                # Rotate to a new file - offset is 0
                 chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, chunk_size)
+                dst_key = (chunk_idx, file_idx)
+                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = 0
+                videos_idx[key]["src_to_dst"][(src_chunk_idx, src_file_idx)] = dst_key
                 dst_path = dst_meta.root / DEFAULT_VIDEO_PATH.format(
                     video_key=key,
                     chunk_index=chunk_idx,
@@ -324,16 +357,20 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
                 )
                 dst_path.parent.mkdir(parents=True, exist_ok=True)
                 shutil.copy(str(src_path), str(dst_path))
-                # Reset offset for next file
-                current_offset = src_duration
+                # Track duration of this new destination file
+                dst_file_durations[dst_key] = src_duration
             else:
-                # Append to existing video file - use current accumulated offset
-                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_offset
+                # Append to existing destination file
+                # Offset is the current duration of this destination file
+                current_dst_duration = dst_file_durations.get(dst_key, 0)
+                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_dst_duration
+                videos_idx[key]["src_to_dst"][(src_chunk_idx, src_file_idx)] = dst_key
                 concatenate_video_files(
                     [dst_path, src_path],
                     dst_path,
                 )
-                current_offset += src_duration
+                # Update duration of this destination file
+                dst_file_durations[dst_key] = current_dst_duration + src_duration
 
             videos_idx[key]["episode_duration"] += src_duration
 

src/lerobot/datasets/factory.py

@@ -98,6 +98,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 image_transforms=image_transforms,
                 revision=cfg.dataset.revision,
                 video_backend=cfg.dataset.video_backend,
+                tolerance_s=cfg.tolerance_s,
             )
         else:
             dataset = StreamingLeRobotDataset(
@@ -108,6 +109,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 image_transforms=image_transforms,
                 revision=cfg.dataset.revision,
                 max_num_shards=cfg.num_workers,
+                tolerance_s=cfg.tolerance_s,
             )
     else:
         raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")

src/lerobot/envs/configs.py

@@ -245,7 +245,7 @@ class HILSerlRobotEnvConfig(EnvConfig):
 class LiberoEnv(EnvConfig):
     task: str = "libero_10"  # can also choose libero_spatial, libero_object, etc.
     fps: int = 30
-    episode_length: int = 520
+    episode_length: int | None = None
     obs_type: str = "pixels_agent_pos"
     render_mode: str = "rgb_array"
     camera_name: str = "agentview_image,robot0_eye_in_hand_image"
@@ -272,6 +272,7 @@ class LiberoEnv(EnvConfig):
             LIBERO_KEY_PIXELS_EYE_IN_HAND: f"{OBS_IMAGES}.image2",
         }
     )
+    control_mode: str = "relative"  # or "absolute"
 
     def __post_init__(self):
         if self.obs_type == "pixels":

src/lerobot/envs/factory.py

@@ -19,8 +19,10 @@ from typing import Any
 import gymnasium as gym
 from gymnasium.envs.registration import registry as gym_registry
 
+from lerobot.configs.policies import PreTrainedConfig
 from lerobot.envs.configs import AlohaEnv, EnvConfig, 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 LiberoProcessorStep
 from lerobot.processor.pipeline import PolicyProcessorPipeline
@@ -39,6 +41,7 @@ def make_env_config(env_type: str, **kwargs) -> EnvConfig:
 
 def make_env_pre_post_processors(
     env_cfg: EnvConfig,
+    policy_cfg: PreTrainedConfig,
 ) -> tuple[
     PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
@@ -61,6 +64,10 @@ def make_env_pre_post_processors(
     # 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()
 
     # For LIBERO environments, add the LiberoProcessorStep to preprocessor
     if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
@@ -136,6 +143,8 @@ def make_env(
             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

src/lerobot/envs/libero.py

@@ -80,10 +80,7 @@ def get_libero_dummy_action():
     return [0, 0, 0, 0, 0, 0, -1]
 
 
-OBS_STATE_DIM = 8
 ACTION_DIM = 7
-AGENT_POS_LOW = -1000.0
-AGENT_POS_HIGH = 1000.0
 ACTION_LOW = -1.0
 ACTION_HIGH = 1.0
 TASK_SUITE_MAX_STEPS: dict[str, int] = {
@@ -103,6 +100,7 @@ class LiberoEnv(gym.Env):
         task_suite: Any,
         task_id: int,
         task_suite_name: str,
+        episode_length: int | None = None,
         camera_name: str | Sequence[str] = "agentview_image,robot0_eye_in_hand_image",
         obs_type: str = "pixels",
         render_mode: str = "rgb_array",
@@ -114,6 +112,7 @@ class LiberoEnv(gym.Env):
         episode_index: int = 0,
         camera_name_mapping: dict[str, str] | None = None,
         num_steps_wait: int = 10,
+        control_mode: str = "relative",
     ):
         super().__init__()
         self.task_id = task_id
@@ -141,14 +140,19 @@ class LiberoEnv(gym.Env):
         self.camera_name_mapping = camera_name_mapping
         self.num_steps_wait = num_steps_wait
         self.episode_index = episode_index
+        self.episode_length = episode_length
         # Load once and keep
         self._init_states = get_task_init_states(task_suite, self.task_id) if self.init_states else None
         self._init_state_id = self.episode_index  # tie each sub-env to a fixed init state
 
         self._env = self._make_envs_task(task_suite, self.task_id)
         default_steps = 500
-        self._max_episode_steps = TASK_SUITE_MAX_STEPS.get(task_suite_name, default_steps)
-
+        self._max_episode_steps = (
+            TASK_SUITE_MAX_STEPS.get(task_suite_name, default_steps)
+            if self.episode_length is None
+            else self.episode_length
+        )
+        self.control_mode = control_mode
         images = {}
         for cam in self.camera_name:
             images[self.camera_name_mapping[cam]] = spaces.Box(
@@ -296,6 +300,15 @@ class LiberoEnv(gym.Env):
         # Increasing this value can improve determinism and reproducibility across resets.
         for _ in range(self.num_steps_wait):
             raw_obs, _, _, _ = self._env.step(get_libero_dummy_action())
+
+        if self.control_mode == "absolute":
+            for robot in self._env.robots:
+                robot.controller.use_delta = False
+        elif self.control_mode == "relative":
+            for robot in self._env.robots:
+                robot.controller.use_delta = True
+        else:
+            raise ValueError(f"Invalid control mode: {self.control_mode}")
         observation = self._format_raw_obs(raw_obs)
         info = {"is_success": False}
         return observation, info
@@ -341,8 +354,10 @@ def _make_env_fns(
     task_id: int,
     n_envs: int,
     camera_names: list[str],
+    episode_length: int | None,
     init_states: bool,
     gym_kwargs: Mapping[str, Any],
+    control_mode: str,
 ) -> list[Callable[[], LiberoEnv]]:
     """Build n_envs factory callables for a single (suite, task_id)."""
 
@@ -354,7 +369,9 @@ def _make_env_fns(
             task_suite_name=suite_name,
             camera_name=camera_names,
             init_states=init_states,
+            episode_length=episode_length,
             episode_index=episode_index,
+            control_mode=control_mode,
             **local_kwargs,
         )
 
@@ -374,6 +391,8 @@ def create_libero_envs(
     camera_name: str | Sequence[str] = "agentview_image,robot0_eye_in_hand_image",
     init_states: bool = True,
     env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
+    control_mode: str = "relative",
+    episode_length: int | None = None,
 ) -> dict[str, dict[int, Any]]:
     """
     Create vectorized LIBERO environments with a consistent return shape.
@@ -415,12 +434,14 @@ def create_libero_envs(
         for tid in selected:
             fns = _make_env_fns(
                 suite=suite,
+                episode_length=episode_length,
                 suite_name=suite_name,
                 task_id=tid,
                 n_envs=n_envs,
                 camera_names=camera_names,
                 init_states=init_states,
                 gym_kwargs=gym_kwargs,
+                control_mode=control_mode,
             )
             out[suite_name][tid] = env_cls(fns)
             print(f"Built vec env | suite={suite_name} | task_id={tid} | n_envs={n_envs}")

src/lerobot/optim/factory.py

@@ -35,6 +35,8 @@ def make_optimizer_and_scheduler(
         tuple[Optimizer, LRScheduler | None]: The couple (Optimizer, Scheduler). Scheduler can be `None`.
     """
     params = policy.get_optim_params() if cfg.use_policy_training_preset else policy.parameters()
+    if cfg.optimizer is None:
+        raise ValueError("Optimizer config is required but not provided in TrainPipelineConfig")
     optimizer = cfg.optimizer.build(params)
     lr_scheduler = cfg.scheduler.build(optimizer, cfg.steps) if cfg.scheduler is not None else None
     return optimizer, lr_scheduler

src/lerobot/optim/optimizers.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import abc
+from collections.abc import Iterable
 from dataclasses import asdict, dataclass, field
 from pathlib import Path
 from typing import Any
@@ -29,6 +30,17 @@ from lerobot.utils.constants import (
 )
 from lerobot.utils.io_utils import deserialize_json_into_object
 
+# Type alias for parameters accepted by optimizer build() methods.
+# This matches PyTorch's optimizer signature while also supporting:
+# - dict[str, Parameter]: Named parameters for differential LR by name (e.g., XVLA)
+# - dict[str, Iterable]: Multiple parameter groups for multi-optimizer configs (e.g., SAC)
+OptimizerParams = (
+    Iterable[torch.nn.Parameter]  # From model.parameters()
+    | Iterable[dict[str, Any]]  # List of param groups with lr/weight_decay overrides
+    | dict[str, torch.nn.Parameter]  # From dict(model.named_parameters()) for name-based LR
+    | dict[str, Any]  # For multi-optimizer configs (SAC) with multiple param groups
+)
+
 
 @dataclass
 class OptimizerConfig(draccus.ChoiceRegistry, abc.ABC):
@@ -45,13 +57,24 @@ class OptimizerConfig(draccus.ChoiceRegistry, abc.ABC):
         return "adam"
 
     @abc.abstractmethod
-    def build(self) -> torch.optim.Optimizer | dict[str, torch.optim.Optimizer]:
+    def build(self, params: OptimizerParams) -> torch.optim.Optimizer | dict[str, torch.optim.Optimizer]:
         """
         Build the optimizer. It can be a single optimizer or a dictionary of optimizers.
+
         NOTE: Multiple optimizers are useful when you have different models to optimize.
         For example, you can have one optimizer for the policy and another one for the value function
         in reinforcement learning settings.
 
+        Args:
+            params: Parameters to optimize. Accepts multiple formats depending on the optimizer:
+                - Iterable[Parameter]: From model.parameters() - standard PyTorch usage
+                - Iterable[dict]: List of param groups with 'params' key and optional
+                  'lr', 'weight_decay' overrides (e.g., ACT, VQBeT policies)
+                - dict[str, Parameter]: From dict(model.named_parameters()) for optimizers
+                  that apply differential learning rates by parameter name (e.g., XVLA)
+                - dict[str, Iterable]: For multi-optimizer configs where each key maps to
+                  a separate optimizer's parameters (e.g., SAC with actor/critic/temperature)
+
         Returns:
             The optimizer or a dictionary of optimizers.
         """
@@ -67,7 +90,7 @@ class AdamConfig(OptimizerConfig):
     weight_decay: float = 0.0
     grad_clip_norm: float = 10.0
 
-    def build(self, params: dict) -> torch.optim.Optimizer:
+    def build(self, params: OptimizerParams) -> torch.optim.Optimizer:
         kwargs = asdict(self)
         kwargs.pop("grad_clip_norm")
         return torch.optim.Adam(params, **kwargs)
@@ -82,7 +105,7 @@ class AdamWConfig(OptimizerConfig):
     weight_decay: float = 1e-2
     grad_clip_norm: float = 10.0
 
-    def build(self, params: dict) -> torch.optim.Optimizer:
+    def build(self, params: OptimizerParams) -> torch.optim.Optimizer:
         kwargs = asdict(self)
         kwargs.pop("grad_clip_norm")
         return torch.optim.AdamW(params, **kwargs)
@@ -98,12 +121,111 @@ class SGDConfig(OptimizerConfig):
     weight_decay: float = 0.0
     grad_clip_norm: float = 10.0
 
-    def build(self, params: dict) -> torch.optim.Optimizer:
+    def build(self, params: OptimizerParams) -> torch.optim.Optimizer:
         kwargs = asdict(self)
         kwargs.pop("grad_clip_norm")
         return torch.optim.SGD(params, **kwargs)
 
 
+@OptimizerConfig.register_subclass("xvla-adamw")
+@dataclass
+class XVLAAdamWConfig(OptimizerConfig):
+    """Custom AdamW optimizer for XVLA with differential learning rates.
+
+    The Vision-Language Model (VLM) is trained with 1/10 of the base learning rate
+    for stable optimization, while all other components use the full LR.
+
+    This LR ratio is crucial for achieving strong and stable finetuning performance.
+
+    Soft-prompts can optionally use a separate learning rate with warm-up support.
+    Set `soft_prompt_lr_scale` to a value < 1.0 (e.g., 0.1) to start soft-prompts
+    at a lower LR. Combine with a warmup scheduler for optimal results.
+
+    Note:
+        Completely matching official reported performance may require an additional
+        warm-up LR schedule for soft-prompts, which can bring minor improvements.
+        When `soft_prompt_warmup_lr_scale` is set, soft-prompts start at
+        `lr * soft_prompt_warmup_lr_scale` and should be warmed up via the scheduler.
+
+    Parameter Groups:
+        - Group 0 (vlm): VLM parameters at lr * 0.1, weight_decay * 0.1
+        - Group 1 (soft_prompts): Soft-prompt parameters at lr * soft_prompt_lr_scale
+        - Group 2 (other): All other parameters at full lr
+    """
+
+    lr: float = 1e-4
+    betas: tuple[float, float] = (0.9, 0.99)
+    eps: float = 1e-8
+    weight_decay: float = 0.0
+    grad_clip_norm: float = 10.0
+    # Soft-prompt specific settings
+    soft_prompt_lr_scale: float = 1.0  # Scale factor for soft-prompt LR (1.0 = same as base LR)
+    soft_prompt_warmup_lr_scale: float | None = None  # If set, start soft-prompts at this scale (e.g., 0.01)
+
+    def build(self, params: OptimizerParams) -> torch.optim.Optimizer:
+        """
+        Build AdamW optimizer with differential learning rates.
+
+        Args:
+            params: Must be a dict[str, Parameter] from dict(model.named_parameters())
+                or equivalent.
+
+        Returns:
+            AdamW optimizer with parameter groups for VLM, soft-prompts, and other components
+
+        Raises:
+            AssertionError: If params is not a dict (e.g., from model.parameters())
+        """
+        assert isinstance(params, dict), "Custom LR optimizer requires `named_parameters()` as inputs."
+
+        vlm_group, soft_prompt_group, other_group = [], [], []
+        for name, p in params.items():
+            if not p.requires_grad:
+                continue
+            if "vlm" in name.lower():
+                vlm_group.append(p)
+            elif "soft_prompt" in name.lower():
+                soft_prompt_group.append(p)
+            else:
+                other_group.append(p)
+
+        # Determine soft-prompt LR
+        soft_prompt_lr = self.lr * self.soft_prompt_lr_scale
+        if self.soft_prompt_warmup_lr_scale is not None:
+            # Start at warmup scale, scheduler will warm up to soft_prompt_lr
+            soft_prompt_lr = self.lr * self.soft_prompt_warmup_lr_scale
+
+        param_groups: list[dict[str, Any]] = [
+            {
+                "params": vlm_group,
+                "lr": self.lr * 0.1,
+                "weight_decay": self.weight_decay * 0.1,
+                "name": "vlm",
+            },
+            {
+                "params": soft_prompt_group,
+                "lr": soft_prompt_lr,
+                "weight_decay": self.weight_decay,
+                "name": "soft_prompts",
+            },
+            {
+                "params": other_group,
+                "lr": self.lr,
+                "weight_decay": self.weight_decay,
+                "name": "other",
+            },
+        ]
+
+        # Filter out empty groups
+        param_groups = [g for g in param_groups if len(g["params"]) > 0]
+
+        return torch.optim.AdamW(
+            param_groups,
+            betas=self.betas,
+            eps=self.eps,
+        )
+
+
 @OptimizerConfig.register_subclass("multi_adam")
 @dataclass
 class MultiAdamConfig(OptimizerConfig):
@@ -123,19 +245,25 @@ class MultiAdamConfig(OptimizerConfig):
     grad_clip_norm: float = 10.0
     optimizer_groups: dict[str, dict[str, Any]] = field(default_factory=dict)
 
-    def build(self, params_dict: dict[str, list]) -> dict[str, torch.optim.Optimizer]:
+    def build(self, params: OptimizerParams) -> dict[str, torch.optim.Optimizer]:
         """Build multiple Adam optimizers.
 
         Args:
-            params_dict: Dictionary mapping parameter group names to lists of parameters
-                         The keys should match the keys in optimizer_groups
+            params: Must be a dict[str, Iterable[Parameter]] mapping parameter group names
+                to iterables of parameters. The keys should match the keys in optimizer_groups.
+                Typically from policies that need separate optimizers (e.g., SAC with
+                actor/critic/temperature).
 
         Returns:
             Dictionary mapping parameter group names to their optimizers
+
+        Raises:
+            AssertionError: If params is not a dict
         """
+        assert isinstance(params, dict), "MultiAdamConfig requires a dict of parameter groups as inputs."
         optimizers = {}
 
-        for name, params in params_dict.items():
+        for name, group_params in params.items():
             # Get group-specific hyperparameters or use defaults
             group_config = self.optimizer_groups.get(name, {})
 
@@ -147,7 +275,7 @@ class MultiAdamConfig(OptimizerConfig):
                 "weight_decay": group_config.get("weight_decay", self.weight_decay),
             }
 
-            optimizers[name] = torch.optim.Adam(params, **optimizer_kwargs)
+            optimizers[name] = torch.optim.Adam(group_params, **optimizer_kwargs)
 
         return optimizers
 

src/lerobot/optim/schedulers.py

@@ -30,7 +30,7 @@ from lerobot.utils.io_utils import deserialize_json_into_object
 
 @dataclass
 class LRSchedulerConfig(draccus.ChoiceRegistry, abc.ABC):
-    num_warmup_steps: int
+    num_warmup_steps: int | None
 
     @property
     def type(self) -> str:

src/lerobot/policies/__init__.py

@@ -21,6 +21,8 @@ from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .smolvla.processor_smolvla import SmolVLANewLineProcessor
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
+from .wall_x.configuration_wall_x import WallXConfig as WallXConfig
+from .xvla.configuration_xvla import XVLAConfig as XVLAConfig
 
 __all__ = [
     "ACTConfig",
@@ -28,7 +30,10 @@ __all__ = [
     "PI0Config",
     "PI05Config",
     "SmolVLAConfig",
+    "SARMConfig",
     "TDMPCConfig",
     "VQBeTConfig",
     "GrootConfig",
+    "XVLAConfig",
+    "WallXConfig",
 ]

src/lerobot/policies/act/modeling_act.py

@@ -50,6 +50,7 @@ class ACTPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: ACTConfig,
+        **kwargs,
     ):
         """
         Args:

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -56,6 +56,7 @@ class DiffusionPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: DiffusionConfig,
+        **kwargs,
     ):
         """
         Args:

src/lerobot/policies/factory.py

@@ -16,6 +16,7 @@
 
 from __future__ import annotations
 
+import importlib
 import logging
 from typing import Any, TypedDict
 
@@ -36,10 +37,13 @@ from lerobot.policies.pi05.configuration_pi05 import PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.sac.configuration_sac import SACConfig
 from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
+from lerobot.policies.sarm.configuration_sarm import SARMConfig
 from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
 from lerobot.policies.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.policies.utils import validate_visual_features_consistency
 from lerobot.policies.vqbet.configuration_vqbet import VQBeTConfig
+from lerobot.policies.wall_x.configuration_wall_x import WallXConfig
+from lerobot.policies.xvla.configuration_xvla import XVLAConfig
 from lerobot.processor import PolicyAction, PolicyProcessorPipeline
 from lerobot.processor.converters import (
     batch_to_transition,
@@ -59,7 +63,7 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
 
     Args:
         name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
-              "vqbet", "pi0", "pi05", "sac", "reward_classifier", "smolvla".
+              "vqbet", "pi0", "pi05", "sac", "reward_classifier", "smolvla", "wall_x".
 
     Returns:
         The policy class corresponding to the given name.
@@ -103,12 +107,27 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
 
         return SmolVLAPolicy
+    elif name == "sarm":
+        from lerobot.policies.sarm.modeling_sarm import SARMRewardModel
+
+        return SARMRewardModel
     elif name == "groot":
         from lerobot.policies.groot.modeling_groot import GrootPolicy
 
         return GrootPolicy
+    elif name == "xvla":
+        from lerobot.policies.xvla.modeling_xvla import XVLAPolicy
+
+        return XVLAPolicy
+    elif name == "wall_x":
+        from lerobot.policies.wall_x.modeling_wall_x import WallXPolicy
+
+        return WallXPolicy
     else:
-        raise NotImplementedError(f"Policy with name {name} is not implemented.")
+        try:
+            return _get_policy_cls_from_policy_name(name=name)
+        except Exception as e:
+            raise ValueError(f"Policy type '{name}' is not available.") from e
 
 
 def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
@@ -121,7 +140,7 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
     Args:
         policy_type: The type of the policy. Supported types include "tdmpc",
                      "diffusion", "act", "vqbet", "pi0", "pi05", "sac", "smolvla",
-                     "reward_classifier".
+                     "reward_classifier", "wall_x".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
     Returns:
@@ -150,8 +169,16 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return RewardClassifierConfig(**kwargs)
     elif policy_type == "groot":
         return GrootConfig(**kwargs)
+    elif policy_type == "xvla":
+        return XVLAConfig(**kwargs)
+    elif policy_type == "wall_x":
+        return WallXConfig(**kwargs)
     else:
-        raise ValueError(f"Policy type '{policy_type}' is not available.")
+        try:
+            config_cls = PreTrainedConfig.get_choice_class(policy_type)
+            return config_cls(**kwargs)
+        except Exception as e:
+            raise ValueError(f"Policy type '{policy_type}' is not available.") from e
 
 
 class ProcessorConfigKwargs(TypedDict, total=False):
@@ -322,6 +349,14 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
+    elif isinstance(policy_cfg, SARMConfig):
+        from lerobot.policies.sarm.processor_sarm import make_sarm_pre_post_processors
+
+        processors = make_sarm_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            dataset_meta=kwargs.get("dataset_meta"),
+        )
     elif isinstance(policy_cfg, GrootConfig):
         from lerobot.policies.groot.processor_groot import make_groot_pre_post_processors
 
@@ -330,8 +365,32 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
+    elif isinstance(policy_cfg, XVLAConfig):
+        from lerobot.policies.xvla.processor_xvla import (
+            make_xvla_pre_post_processors,
+        )
+
+        processors = make_xvla_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
+    elif isinstance(policy_cfg, WallXConfig):
+        from lerobot.policies.wall_x.processor_wall_x import make_wall_x_pre_post_processors
+
+        processors = make_wall_x_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
     else:
-        raise NotImplementedError(f"Processor for policy type '{policy_cfg.type}' is not implemented.")
+        try:
+            processors = _make_processors_from_policy_config(
+                config=policy_cfg,
+                dataset_stats=kwargs.get("dataset_stats"),
+            )
+        except Exception as e:
+            raise ValueError(f"Processor for policy type '{policy_cfg.type}' is not implemented.") from e
 
     return processors
 
@@ -400,12 +459,18 @@ def make_policy(
             raise ValueError("env_cfg cannot be None when ds_meta is not provided")
         features = env_to_policy_features(env_cfg)
 
-    if not cfg.output_features:
-        cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+    cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
     if not cfg.input_features:
         cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
     kwargs["config"] = cfg
 
+    # Pass dataset_stats to the policy if available (needed for some policies like SARM)
+    if ds_meta is not None and hasattr(ds_meta, "stats"):
+        kwargs["dataset_stats"] = ds_meta.stats
+
+    if ds_meta is not None:
+        kwargs["dataset_meta"] = ds_meta
+
     if cfg.pretrained_path:
         # Load a pretrained policy and override the config if needed (for example, if there are inference-time
         # hyperparameters that we want to vary).
@@ -425,3 +490,65 @@ def make_policy(
         # TODO: (jadechoghari) - add a check_state(cfg, features) and check_action(cfg, features)
 
     return policy
+
+
+def _get_policy_cls_from_policy_name(name: str) -> type[PreTrainedConfig]:
+    """Get policy class from its registered name using dynamic imports.
+
+    This is used as a helper function to import policies from 3rd party lerobot plugins.
+
+    Args:
+        name: The name of the policy.
+    Returns:
+        The policy class corresponding to the given name.
+    """
+    if name not in PreTrainedConfig.get_known_choices():
+        raise ValueError(
+            f"Unknown policy name '{name}'. Available policies: {PreTrainedConfig.get_known_choices()}"
+        )
+
+    config_cls = PreTrainedConfig.get_choice_class(name)
+    config_cls_name = config_cls.__name__
+
+    model_name = config_cls_name.removesuffix("Config")  # e.g., DiffusionConfig -> Diffusion
+    if model_name == config_cls_name:
+        raise ValueError(
+            f"The config class name '{config_cls_name}' does not follow the expected naming convention."
+            f"Make sure it ends with 'Config'!"
+        )
+    cls_name = model_name + "Policy"  # e.g., DiffusionConfig -> DiffusionPolicy
+    module_path = config_cls.__module__.replace(
+        "configuration_", "modeling_"
+    )  # e.g., configuration_diffusion -> modeling_diffusion
+
+    module = importlib.import_module(module_path)
+    policy_cls = getattr(module, cls_name)
+    return policy_cls
+
+
+def _make_processors_from_policy_config(
+    config: PreTrainedConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[Any, Any]:
+    """Create pre- and post-processors from a policy configuration using dynamic imports.
+
+    This is used as a helper function to import processor factories from 3rd party lerobot plugins.
+
+    Args:
+        config: The policy configuration object.
+        dataset_stats: Dataset statistics for normalization.
+    Returns:
+        A tuple containing the input (pre-processor) and output (post-processor) pipelines.
+    """
+
+    policy_type = config.type
+    function_name = f"make_{policy_type}_pre_post_processors"
+    module_path = config.__class__.__module__.replace(
+        "configuration_", "processor_"
+    )  # e.g., configuration_diffusion -> processor_diffusion
+    logging.debug(
+        f"Instantiating pre/post processors using function '{function_name}' from module '{module_path}'"
+    )
+    module = importlib.import_module(module_path)
+    function = getattr(module, function_name)
+    return function(config, dataset_stats=dataset_stats)

src/lerobot/policies/groot/modeling_groot.py

@@ -49,7 +49,7 @@ class GrootPolicy(PreTrainedPolicy):
     name = "groot"
     config_class = GrootConfig
 
-    def __init__(self, config: GrootConfig):
+    def __init__(self, config: GrootConfig, **kwargs):
         """Initialize Groot policy wrapper."""
         super().__init__(config)
         config.validate_features()

src/lerobot/policies/pi0/configuration_pi0.py

@@ -23,6 +23,8 @@ from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
 from lerobot.utils.constants import OBS_IMAGES
 
+DEFAULT_IMAGE_SIZE = 224
+
 
 @PreTrainedConfig.register_subclass("pi0")
 @dataclass
@@ -51,7 +53,10 @@ class PI0Config(PreTrainedConfig):
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
 
-    image_resolution: tuple[int, int] = (224, 224)  # see openpi `preprocessing_pytorch.py`
+    image_resolution: tuple[int, int] = (
+        DEFAULT_IMAGE_SIZE,
+        DEFAULT_IMAGE_SIZE,
+    )  # see openpi `preprocessing_pytorch.py`
 
     # Add empty images. Used to add empty cameras when no image features are present.
     empty_cameras: int = 0

src/lerobot/policies/pi0/modeling_pi0.py

@@ -41,7 +41,7 @@ else:
     PaliGemmaForConditionalGeneration = None
 
 from lerobot.configs.policies import PreTrainedConfig
-from lerobot.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.policies.pi0.configuration_pi0 import DEFAULT_IMAGE_SIZE, PI0Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
 from lerobot.utils.constants import (
@@ -93,10 +93,11 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
 
 
 def sample_beta(alpha, beta, bsize, device):  # see openpi `sample_beta` (exact copy)
-    alpha_t = torch.as_tensor(alpha, dtype=torch.float32, device=device)
-    beta_t = torch.as_tensor(beta, dtype=torch.float32, device=device)
+    # Beta sampling uses _sample_dirichlet which isn't implemented for MPS, so sample on CPU
+    alpha_t = torch.tensor(alpha, dtype=torch.float32)
+    beta_t = torch.tensor(beta, dtype=torch.float32)
     dist = torch.distributions.Beta(alpha_t, beta_t)
-    return dist.sample((bsize,))
+    return dist.sample((bsize,)).to(device)
 
 
 def make_att_2d_masks(pad_masks, att_masks):  # see openpi `make_att_2d_masks` (exact copy)
@@ -337,6 +338,7 @@ class PaliGemmaWithExpertModel(
         action_expert_config,
         use_adarms=None,
         precision: Literal["bfloat16", "float32"] = "bfloat16",
+        image_size: int = DEFAULT_IMAGE_SIZE,
     ):
         if use_adarms is None:
             use_adarms = [False, False]
@@ -356,6 +358,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
+        vlm_config_hf.vision_config.image_size = image_size
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
@@ -519,11 +522,17 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         paligemma_config = get_gemma_config(config.paligemma_variant)
         action_expert_config = get_gemma_config(config.action_expert_variant)
 
+        if config.image_resolution[0] != config.image_resolution[1]:
+            raise ValueError(
+                f"PaliGemma expects square image resolution, invalid resolution: {config.image_resolution}"
+            )
+
         self.paligemma_with_expert = PaliGemmaWithExpertModel(
             paligemma_config,
             action_expert_config,
             use_adarms=[False, False],
             precision=config.dtype,
+            image_size=config.image_resolution[0],
         )
 
         self.action_in_proj = nn.Linear(config.max_action_dim, action_expert_config.width)
@@ -812,16 +821,13 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         )
 
         dt = -1.0 / num_steps
-        dt = torch.tensor(dt, dtype=torch.float32, device=device)
 
         x_t = noise
-        time = torch.tensor(1.0, dtype=torch.float32, device=device)
-        while time >= -dt / 2:
-            expanded_time = time.expand(bsize)
+        for step in range(num_steps):
+            time = 1.0 + step * dt
+            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
 
-            # Define a closure function to properly capture expanded_time
-            # This avoids the lambda expression (E731) and loop variable binding (B023) issues
-            def denoise_step_partial_call(input_x_t, current_timestep=expanded_time):
+            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
                 return self.denoise_step(
                     state=state,
                     prefix_pad_masks=prefix_pad_masks,
@@ -846,15 +852,11 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             else:
                 v_t = denoise_step_partial_call(x_t)
 
-            # Euler step
-            x_t += dt * v_t
+            x_t = x_t + dt * v_t
 
-            # Record x_t and v_t after Euler step
             if self.rtc_processor is not None and self.rtc_processor.is_debug_enabled():
                 self.rtc_processor.track(time=time, x_t=x_t, v_t=v_t)
 
-            time += dt
-
         return x_t
 
     def denoise_step(
@@ -906,6 +908,7 @@ class PI0Policy(PreTrainedPolicy):
     def __init__(
         self,
         config: PI0Config,
+        **kwargs,
     ):
         """
         Args:
@@ -1234,9 +1237,15 @@ class PI0Policy(PreTrainedPolicy):
 
         return actions
 
-    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
-        """Run the batch through the model and compute the loss for training."""
+    def forward(self, batch: dict[str, Tensor], reduction: str = "mean") -> tuple[Tensor, dict]:
+        """Run the batch through the model and compute the loss for training.
 
+        Args:
+            batch: Training batch containing observations and actions.
+            reduction: How to reduce the loss. Options:
+                - "mean": Return scalar mean loss (default, backward compatible)
+                - "none": Return per-sample losses of shape (batch_size,) for RA-BC weighting
+        """
         # Prepare inputs
         images, img_masks = self._preprocess_images(batch)
         lang_tokens, lang_masks = batch[f"{OBS_LANGUAGE_TOKENS}"], batch[f"{OBS_LANGUAGE_ATTENTION_MASK}"]
@@ -1250,11 +1259,17 @@ class PI0Policy(PreTrainedPolicy):
         original_action_dim = self.config.output_features[ACTION].shape[0]
         losses = losses[:, :, :original_action_dim]
 
-        loss = losses.mean()
-
         loss_dict = {
-            "loss": loss.item(),
             "loss_per_dim": losses.mean(dim=[0, 1]).detach().cpu().numpy().tolist(),
         }
 
-        return loss, loss_dict
+        if reduction == "none":
+            # Return per-sample losses (B,) by averaging over time and action dims
+            per_sample_loss = losses.mean(dim=(1, 2))
+            loss_dict["loss"] = per_sample_loss.mean().item()
+            return per_sample_loss, loss_dict
+        else:
+            # Default: return scalar mean loss
+            loss = losses.mean()
+            loss_dict["loss"] = loss.item()
+            return loss, loss_dict

src/lerobot/policies/pi05/configuration_pi05.py

@@ -22,6 +22,8 @@ from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
 
+DEFAULT_IMAGE_SIZE = 224
+
 
 @PreTrainedConfig.register_subclass("pi05")
 @dataclass
@@ -50,7 +52,10 @@ class PI05Config(PreTrainedConfig):
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
 
-    image_resolution: tuple[int, int] = (224, 224)  # see openpi `preprocessing_pytorch.py`
+    image_resolution: tuple[int, int] = (
+        DEFAULT_IMAGE_SIZE,
+        DEFAULT_IMAGE_SIZE,
+    )  # see openpi `preprocessing_pytorch.py`
 
     # Add empty images. Used to add empty cameras when no image features are present.
     empty_cameras: int = 0

src/lerobot/policies/pi05/modeling_pi05.py

@@ -41,7 +41,7 @@ else:
     PaliGemmaForConditionalGeneration = None
 
 from lerobot.configs.policies import PreTrainedConfig
-from lerobot.policies.pi05.configuration_pi05 import PI05Config
+from lerobot.policies.pi05.configuration_pi05 import DEFAULT_IMAGE_SIZE, PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
 from lerobot.utils.constants import (
@@ -336,6 +336,7 @@ class PaliGemmaWithExpertModel(
         action_expert_config,
         use_adarms=None,
         precision: Literal["bfloat16", "float32"] = "bfloat16",
+        image_size: int = DEFAULT_IMAGE_SIZE,
     ):
         if use_adarms is None:
             use_adarms = [False, False]
@@ -355,6 +356,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
+        vlm_config_hf.vision_config.image_size = image_size
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
@@ -518,11 +520,17 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         paligemma_config = get_gemma_config(config.paligemma_variant)
         action_expert_config = get_gemma_config(config.action_expert_variant)
 
+        if config.image_resolution[0] != config.image_resolution[1]:
+            raise ValueError(
+                f"PaliGemma expects square image resolution, invalid resolution: {config.image_resolution}"
+            )
+
         self.paligemma_with_expert = PaliGemmaWithExpertModel(
             paligemma_config,
             action_expert_config,
             use_adarms=[False, True],
             precision=config.dtype,
+            image_size=config.image_resolution[0],
         )
 
         self.action_in_proj = nn.Linear(config.max_action_dim, action_expert_config.width)
@@ -538,6 +546,8 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         if config.compile_model:
             torch.set_float32_matmul_precision("high")
             self.sample_actions = torch.compile(self.sample_actions, mode=config.compile_mode)
+            # Also compile the main forward pass used during training
+            self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
         msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
 
@@ -785,16 +795,13 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         )
 
         dt = -1.0 / num_steps
-        dt = torch.tensor(dt, dtype=torch.float32, device=device)
 
         x_t = noise
-        time = torch.tensor(1.0, dtype=torch.float32, device=device)
-        while time >= -dt / 2:
-            expanded_time = time.expand(bsize)
+        for step in range(num_steps):
+            time = 1.0 + step * dt
+            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
 
-            # Define a closure function to properly capture expanded_time
-            # This avoids the lambda expression (E731) and loop variable binding (B023) issues
-            def denoise_step_partial_call(input_x_t, current_timestep=expanded_time):
+            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
                 return self.denoise_step(
                     prefix_pad_masks=prefix_pad_masks,
                     past_key_values=past_key_values,
@@ -818,15 +825,11 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             else:
                 v_t = denoise_step_partial_call(x_t)
 
-            # Euler step
-            x_t += dt * v_t
+            x_t = x_t + dt * v_t
 
-            # Record x_t and v_t after Euler step
             if self.rtc_processor is not None and self.rtc_processor.is_debug_enabled():
                 self.rtc_processor.track(time=time, x_t=x_t, v_t=v_t)
 
-            time += dt
-
         return x_t
 
     def denoise_step(
@@ -877,6 +880,7 @@ class PI05Policy(PreTrainedPolicy):
     def __init__(
         self,
         config: PI05Config,
+        **kwargs,
     ):
         """
         Args:
@@ -1206,9 +1210,15 @@ class PI05Policy(PreTrainedPolicy):
 
         return actions
 
-    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
-        """Run the batch through the model and compute the loss for training."""
+    def forward(self, batch: dict[str, Tensor], reduction: str = "mean") -> tuple[Tensor, dict]:
+        """Run the batch through the model and compute the loss for training.
 
+        Args:
+            batch: Training batch containing observations and actions.
+            reduction: How to reduce the loss. Options:
+                - "mean": Return scalar mean loss (default, backward compatible)
+                - "none": Return per-sample losses of shape (batch_size,) for RA-BC weighting
+        """
         # Prepare inputs
         images, img_masks = self._preprocess_images(batch)
         tokens, masks = batch[f"{OBS_LANGUAGE_TOKENS}"], batch[f"{OBS_LANGUAGE_ATTENTION_MASK}"]
@@ -1222,11 +1232,17 @@ class PI05Policy(PreTrainedPolicy):
         original_action_dim = self.config.output_features[ACTION].shape[0]
         losses = losses[:, :, :original_action_dim]
 
-        loss = losses.mean()
-
         loss_dict = {
-            "loss": loss.item(),
             "loss_per_dim": losses.mean(dim=[0, 1]).detach().cpu().numpy().tolist(),
         }
 
-        return loss, loss_dict
+        if reduction == "none":
+            # Return per-sample losses (B,) by averaging over time and action dims
+            per_sample_loss = losses.mean(dim=(1, 2))
+            loss_dict["loss"] = per_sample_loss.mean().item()
+            return per_sample_loss, loss_dict
+        else:
+            # Default: return scalar mean loss
+            loss = losses.mean()
+            loss_dict["loss"] = loss.item()
+            return loss, loss_dict

src/lerobot/policies/pi0fast/README.md

@@ -0,0 +1,49 @@
+# 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)                |
+
+---
+
+## Citation
+
+If you use this work, please cite both **OpenPI** and the π₀.₅ paper:
+
+```bibtex
+@misc{openpi2024,
+  author       = {Physical Intelligence Lab},
+  title        = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
+  year         = {2024},
+  publisher    = {GitHub},
+  howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
+  license      = {Apache-2.0}
+}
+
+@misc{intelligence2025pi05visionlanguageactionmodelopenworld,
+  title        = {π₀.₅: a Vision-Language-Action Model with Open-World Generalization},
+  author       = {Physical Intelligence and Kevin Black and Noah Brown and James Darpinian and Karan Dhabalia and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Manuel Y. Galliker and Dibya Ghosh and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Devin LeBlanc and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Allen Z. Ren and Lucy Xiaoyang Shi and Laura Smith and Jost Tobias Springenberg and Kyle Stachowicz and James Tanner and Quan Vuong and Homer Walke and Anna Walling and Haohuan Wang and Lili Yu and Ury Zhilinsky},
+  year         = {2025},
+  eprint       = {2504.16054},
+  archivePrefix= {arXiv},
+  primaryClass = {cs.LG},
+  url          = {https://arxiv.org/abs/2504.16054},
+}
+```
+
+---
+
+## License
+
+This port follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).

src/lerobot/policies/pi0fast/__init__.py

@@ -0,0 +1,21 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_pi05 import PI05Config
+from .modeling_pi05 import PI05Policy
+from .processor_pi05 import make_pi05_pre_post_processors
+
+__all__ = ["PI05Config", "PI05Policy", "make_pi05_pre_post_processors"]

src/lerobot/policies/pi0fast/configuration_pi05.py

@@ -0,0 +1,164 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.optim.optimizers import AdamWConfig
+from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+
+DEFAULT_IMAGE_SIZE = 224
+
+
+@PreTrainedConfig.register_subclass("pi05")
+@dataclass
+class PI05Config(PreTrainedConfig):
+    paligemma_variant: str = "gemma_2b"
+    action_expert_variant: str = "gemma_300m"
+    dtype: str = "float32"  # Options: "bfloat16", "float32"
+
+    n_obs_steps: int = 1
+    chunk_size: int = 50  # Number of action steps to predict, in openpi called "action_horizon"
+    n_action_steps: int = 50  # Number of action steps to execute
+
+    # Shorter state and action vectors will be padded to these dimensions
+    max_state_dim: int = 32
+    max_action_dim: int = 32
+
+    # Flow matching parameters: see openpi `PI0Pytorch`
+    num_inference_steps: int = 10
+    time_sampling_beta_alpha: float = 1.5
+    time_sampling_beta_beta: float = 1.0
+    time_sampling_scale: float = 0.999
+    time_sampling_offset: float = 0.001
+    min_period: float = 4e-3
+    max_period: float = 4.0
+
+    # Real-Time Chunking (RTC) configuration
+    rtc_config: RTCConfig | None = None
+
+    image_resolution: tuple[int, int] = (
+        DEFAULT_IMAGE_SIZE,
+        DEFAULT_IMAGE_SIZE,
+    )  # see openpi `preprocessing_pytorch.py`
+
+    # Add empty images. Used to add empty cameras when no image features are present.
+    empty_cameras: int = 0
+
+    tokenizer_max_length: int = 200  # see openpi `__post_init__`
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.QUANTILES,  # Pi0.5 uses quantiles for state
+            "ACTION": NormalizationMode.QUANTILES,  # Pi0.5 uses quantiles for action
+        }
+    )
+
+    # Training settings
+    gradient_checkpointing: bool = False  # Enable gradient checkpointing for memory optimization
+    compile_model: bool = False  # Whether to use torch.compile for model optimization
+    compile_mode: str = "max-autotune"  # Torch compile mode
+    device: str | None = None  # Device to use for the model (None = auto-detect)
+
+    # Optimizer settings: see openpi `AdamW`
+    optimizer_lr: float = 2.5e-5  # see openpi `CosineDecaySchedule: peak_lr`
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 0.01
+    optimizer_grad_clip_norm: float = 1.0
+
+    # Scheduler settings: see openpi `CosineDecaySchedule`
+    # Note: These will auto-scale if --steps < scheduler_decay_steps
+    # For example, --steps=3000 will scale warmup to 100 and decay to 3000
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    tokenizer_max_length: int = 200  # see openpi `__post_init__`
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        # Validate configuration
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"n_action_steps ({self.n_action_steps}) cannot be greater than chunk_size ({self.chunk_size})"
+            )
+
+        if self.paligemma_variant not in ["gemma_300m", "gemma_2b"]:
+            raise ValueError(f"Invalid paligemma_variant: {self.paligemma_variant}")
+
+        if self.action_expert_variant not in ["gemma_300m", "gemma_2b"]:
+            raise ValueError(f"Invalid action_expert_variant: {self.action_expert_variant}")
+
+        if self.dtype not in ["bfloat16", "float32"]:
+            raise ValueError(f"Invalid dtype: {self.dtype}")
+
+    def validate_features(self) -> None:
+        """Validate and set up input/output features."""
+        for i in range(self.empty_cameras):
+            key = f"observation.images.empty_camera_{i}"
+            empty_camera = PolicyFeature(
+                type=FeatureType.VISUAL,
+                shape=(3, *self.image_resolution),  # Use configured image resolution
+            )
+            self.input_features[key] = empty_camera
+
+        if "observation.state" not in self.input_features:
+            state_feature = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(self.max_state_dim,),  # Padded to max_state_dim
+            )
+            self.input_features["observation.state"] = state_feature
+
+        if "action" not in self.output_features:
+            action_feature = PolicyFeature(
+                type=FeatureType.ACTION,
+                shape=(self.max_action_dim,),  # Padded to max_action_dim
+            )
+            self.output_features["action"] = action_feature
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.optimizer_grad_clip_norm,
+        )
+
+    def get_scheduler_preset(self):
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    @property
+    def observation_delta_indices(self) -> None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

src/lerobot/policies/pi0fast/modeling_pi0fast.py

@@ -0,0 +1,995 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ONLY AN EXAMPLE FILE, NEVER USED, IT IS OLD CODE
+"""
+π0+FAST: Efficient Action Tokenization for Vision-Language-Action Models
+
+[Paper](https://huggingface.co/papers/2501.09747)
+[Jax code](https://github.com/Physical-Intelligence/openpi)
+
+Designed by Physical Intelligence. Ported from Jax by Hugging Face.
+Disclaimer: It is not expected to perform as well as the original implementation.
+
+Example of finetuning the pi0+FAST pretrained model (`pi0_fast_base` in `openpi`):
+```bash
+lerobot-train \
+--policy.path=lerobot/pi0fast_base \
+--dataset.repo_id=danaaubakirova/koch_test
+```
+
+Example of training the pi0+FAST neural network with from scratch:
+```bash
+lerobot-train \
+--policy.type=pi0fast \
+--dataset.repo_id=danaaubakirova/koch_test
+```
+
+Example of using the pi0 pretrained model outside LeRobot training framework:
+```python
+policy = PI0FASTPolicy.from_pretrained("lerobot/pi0fast_base")
+```
+
+"""
+
+from collections import deque
+from functools import partial
+
+import numpy as np
+import torch
+import torch.nn.functional as F  # noqa: N812
+from PIL import Image
+from scipy.fft import idct
+from torch import Tensor, nn
+from transformers import AutoProcessor, AutoTokenizer, PaliGemmaForConditionalGeneration
+from transformers.cache_utils import HybridCache, StaticCache
+from transformers.models.auto import CONFIG_MAPPING
+
+from lerobot.constants import ACTION, OBS_STATE
+from lerobot.policies.normalize import Normalize, Unnormalize
+from lerobot.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
+from lerobot.policies.pretrained import PreTrainedPolicy
+
+PRECISION = {
+    "float16": torch.float16,
+    "float32": torch.float32,
+    "bfloat16": torch.bfloat16,
+}
+
+
+def normalize(x, min_val, max_val):
+    return (x - min_val) / (max_val - min_val)
+
+
+def unnormalize(x, min_val, max_val):
+    return x * (max_val - min_val) + min_val
+
+
+def safe_arcsin(value):
+    # This ensures that the input stays within
+    # [−1,1] to avoid invalid values for arcsin
+    return torch.arcsin(torch.clamp(value, -1.0, 1.0))
+
+
+def aloha_gripper_to_angular(value):
+    # Aloha transforms the gripper positions into a linear space. The following code
+    # reverses this transformation to be consistent with pi0 which is pretrained in
+    # angular space.
+    #
+    # These values are coming from the Aloha code:
+    # PUPPET_GRIPPER_POSITION_OPEN, PUPPET_GRIPPER_POSITION_CLOSED
+    value = unnormalize(value, min_val=0.01844, max_val=0.05800)
+
+    # This is the inverse of the angular to linear transformation inside the Interbotix code.
+    def linear_to_radian(linear_position, arm_length, horn_radius):
+        value = (horn_radius**2 + linear_position**2 - arm_length**2) / (2 * horn_radius * linear_position)
+        return safe_arcsin(value)
+
+    # The constants are taken from the Interbotix code.
+    value = linear_to_radian(value, arm_length=0.036, horn_radius=0.022)
+
+    # Normalize to [0, 1].
+    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
+    return normalize(value, min_val=0.4, max_val=1.5)
+
+
+def aloha_gripper_from_angular(value):
+    # Convert from the gripper position used by pi0 to the gripper position that is used by Aloha.
+    # Note that the units are still angular but the range is different.
+
+    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
+    value = unnormalize(value, min_val=0.4, max_val=1.5)
+
+    # These values are coming from the Aloha code:
+    # PUPPET_GRIPPER_JOINT_OPEN, PUPPET_GRIPPER_JOINT_CLOSE
+    return normalize(value, min_val=-0.6213, max_val=1.4910)
+
+
+def aloha_gripper_from_angular_inv(value):
+    # Directly inverts the gripper_from_angular function.
+    value = unnormalize(value, min_val=-0.6213, max_val=1.4910)
+    return normalize(value, min_val=0.4, max_val=1.5)
+
+
+class PI0FASTPolicy(PreTrainedPolicy):
+    """Wrapper class around PI0FAST tokenizer and model to train and run inference within LeRobot."""
+
+    config_class = PI0FASTConfig
+    name = "pi0fast"
+
+    def __init__(
+        self,
+        config: PI0FASTConfig,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
+        """
+        Args:
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                    the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
+        """
+
+        super().__init__(config)
+        config.validate_features()
+        self.config = config
+
+        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_targets = Normalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+
+        self.language_tokenizer = AutoProcessor.from_pretrained("google/paligemma-3b-pt-224")
+        self.model = PI0FAST(config)
+
+        self.reset()
+
+    def reset(self):
+        """This should be called whenever the environment is reset."""
+        self._action_queue = deque([], maxlen=self.config.n_action_steps)
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        """Override the from_pretrained method to display important disclaimer."""
+        print(
+            "⚠️  DISCLAIMER: The PI0FAST model is ported from JAX by the Hugging Face team. \n"
+            "   It is not expected to perform as well as the original implementation. \n"
+            "   Original implementation: https://github.com/Physical-Intelligence/openpi"
+        )
+        return super().from_pretrained(*args, **kwargs)
+
+    def get_optim_params(self) -> dict:
+        return self.parameters()
+
+    def _pi_aloha_decode_state(self, state):
+        # Flip the joints.
+        for motor_idx in [1, 2, 8, 9]:
+            state[:, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            state[:, motor_idx] = aloha_gripper_to_angular(state[:, motor_idx])
+        return state
+
+    def _pi_aloha_encode_actions(self, actions):
+        # Flip the joints.
+        for motor_idx in [1, 2, 8, 9]:
+            actions[:, :, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            actions[:, :, motor_idx] = aloha_gripper_from_angular(actions[:, :, motor_idx])
+        return actions
+
+    def _pi_aloha_encode_actions_inv(self, actions):
+        # Flip the joints again.
+        for motor_idx in [1, 2, 8, 9]:
+            actions[:, :, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            actions[:, :, motor_idx] = aloha_gripper_from_angular_inv(actions[:, :, motor_idx])
+        return actions
+
+    @torch.no_grad()
+    def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
+        """Predict a chunk of actions given environment observations."""
+        raise NotImplementedError("Currently not implemented for PI0FAST")
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a single action given environment observations.
+
+        This method wraps `select_actions` in order to return one action at a time for execution in the
+        environment. It works by managing the actions in a queue and only calling `select_actions` when the
+        queue is empty.
+        """
+        self.eval()
+
+        if self.config.adapt_to_pi_aloha:
+            batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
+
+        batch = self.normalize_inputs(batch)
+
+        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
+        # querying the policy.
+        if len(self._action_queue) == 0:
+            actions = self.model.generate_actions(batch)
+
+            actions = actions[:, : self.config.n_action_steps]
+
+            original_action_dim = self.config.action_feature.shape[
+                0
+            ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
+            actions = actions[:, :, :original_action_dim]
+
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            if self.config.adapt_to_pi_aloha:
+                actions = self._pi_aloha_encode_actions(actions)
+
+            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
+            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
+            self._action_queue.extend(actions.transpose(0, 1))
+        return self._action_queue.popleft()
+
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        if self.config.adapt_to_pi_aloha:
+            batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
+            batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
+        batch = self.normalize_inputs(batch)
+        batch = self.normalize_targets(batch)
+        loss_dict = self.model.forward(batch)
+        return loss_dict["loss"], loss_dict
+
+
+def block_causal_update_causal_mask(
+    attention_mask,
+    token_type_ids=None,
+    past_key_values=None,
+    cache_position=None,
+    input_tensor=None,
+    attn_implementation: str = "eager",
+    dtype: torch.dtype = "float32",
+):
+    """
+    Update the causal mask during training and generation. It can be customized to different attention masks.
+    """
+    if attn_implementation == "flash_attention_2":
+        if attention_mask is not None and 0.0 in attention_mask:
+            return attention_mask
+        return None
+    using_static_cache = isinstance(past_key_values, StaticCache)
+    min_dtype = torch.finfo(dtype).min
+
+    if input_tensor is None:
+        input_tensor = attention_mask
+
+    inputs_lead_dim, sequence_length = input_tensor.shape[:2]
+
+    if using_static_cache or isinstance(past_key_values, HybridCache):
+        target_length = past_key_values.get_max_cache_shape()
+    else:
+        target_length = (
+            attention_mask.shape[-1]
+            if isinstance(attention_mask, torch.Tensor)
+            else cache_position[0] + sequence_length + 1
+        )
+
+    # Handle precomputed attention masks
+    if attention_mask is not None and attention_mask.dim() == 4:
+        return attention_mask
+
+    # Causal mask initialization
+    causal_mask = torch.full(
+        (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
+    )
+
+    # Standard causal masking (triu ensures tokens can only attend to past)
+    if sequence_length != 1:
+        causal_mask = torch.triu(causal_mask, diagonal=1)
+
+        # Apply block causal mask
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.to(causal_mask.device).bool()
+            cumsum = torch.cumsum(token_type_ids, dim=1)
+            block_causal_mask = cumsum[:, None, :] <= cumsum[:, :, None]
+
+            # Combine causal_mask with block-wise attention mask
+            causal_mask = torch.where(block_causal_mask, 0.0, causal_mask)
+            causal_mask = causal_mask[:, None, :, :]
+        else:
+            # Apply past cache position constraint
+            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
+                -1, 1
+            )
+            causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
+    else:
+        # Apply past cache position constraint
+        causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
+            -1, 1
+        )
+        causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
+
+    if attention_mask is not None:
+        causal_mask = causal_mask.clone()  # Copy to contiguous memory for in-place edits
+        mask_length = attention_mask.shape[-1]
+
+        # Apply padding mask
+        padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
+            causal_mask.device
+        )
+        padding_mask = padding_mask == 0
+        causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+            padding_mask, min_dtype
+        )
+
+    return causal_mask
+
+
+def prepare_inputs_for_generation(
+    # self,
+    input_ids,
+    past_key_values=None,
+    inputs_embeds=None,
+    cache_position=None,
+    position_ids=None,
+    pixel_values=None,
+    attention_mask=None,
+    token_type_ids=None,
+    use_cache=True,
+    num_logits_to_keep=None,
+    labels=None,
+    self=None,
+    **kwargs,
+):
+    # create block causal attention
+    if cache_position[0] > 0 and input_ids.shape[1] > 0:
+        input_tensor = input_ids[:, -1:]
+        new_positions = (
+            torch.ones(
+                (position_ids.shape[0], input_ids.shape[1]),
+                dtype=position_ids.dtype,
+                device=position_ids.device,
+            ).cumsum(-1)
+            + position_ids[:, -1:]
+        )
+        position_ids = torch.cat([position_ids, new_positions], dim=-1)
+    else:
+        input_tensor = inputs_embeds
+    attention_mask = block_causal_update_causal_mask(
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        cache_position=cache_position,
+        input_tensor=input_tensor,
+        token_type_ids=token_type_ids,
+        dtype=self.dtype,
+        attn_implementation=self.config.text_config._attn_implementation,
+    )
+    # Overwritten -- custom `position_ids` and `pixel_values` handling
+    model_inputs = self.language_model.prepare_inputs_for_generation(
+        input_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        cache_position=cache_position,
+        use_cache=use_cache,
+        num_logits_to_keep=num_logits_to_keep,
+        token_type_ids=token_type_ids,
+        **kwargs,
+    )
+
+    # Position_ids in Paligemma are 1-indexed
+    if model_inputs.get("position_ids") is not None:
+        model_inputs["position_ids"] += 1
+    # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
+    # Otherwise we need pixel values to be passed to model. NOTE: use_cache=False needs pixel_values always
+    if cache_position[0] == 0:
+        model_inputs["pixel_values"] = pixel_values
+    is_training = token_type_ids is not None and labels is not None
+    if cache_position[0] == 0 and isinstance(past_key_values, HybridCache):
+        input_tensor = inputs_embeds if inputs_embeds is not None else input_ids
+        causal_mask = self._update_causal_mask(
+            attention_mask, token_type_ids, past_key_values, cache_position, input_tensor, is_training
+        )
+        model_inputs["attention_mask"] = causal_mask
+
+    return model_inputs
+
+
+class PI0FAST(nn.Module):
+    def __init__(self, config: PI0FASTConfig):
+        super().__init__()
+        self.config = config
+
+        # TODO: move tokenizers in Policy
+        fast_tokenizer_path = "physical-intelligence/fast"
+        pi0_paligemma_path = "google/paligemma-3b-pt-224"
+        self.paligemma_tokenizer = AutoTokenizer.from_pretrained(pi0_paligemma_path)
+        self.processor = AutoProcessor.from_pretrained(pi0_paligemma_path)
+        self.fast_tokenizer = AutoProcessor.from_pretrained(fast_tokenizer_path, trust_remote_code=True)
+        self.fast_skip_tokens = self.config.fast_skip_tokens
+        self.max_input_seq_len = self.config.max_input_seq_len
+        self.action_horizon = self.config.chunk_size
+        self.action_dim = self.config.action_feature.shape[
+            0
+        ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
+        precision = config.precision
+        torch_precision = PRECISION.get(precision, torch.float32)
+        self.pad_token_id = (
+            self.paligemma_tokenizer.pad_token_id
+            if hasattr(self.paligemma_tokenizer, "pad_token_id")
+            else self.paligemma_tokenizer.eos_token_id
+        )
+
+        paligemma_config = CONFIG_MAPPING["paligemma"](
+            transformers_version="4.48.1",
+            _vocab_size=257152,
+            bos_token_id=2,
+            eos_token_id=1,
+            hidden_size=2048,
+            image_token_index=257152,
+            model_type="paligemma",
+            pad_token_id=0,
+            projection_dim=2048,
+            text_config={
+                "hidden_activation": "gelu_pytorch_tanh",
+                "hidden_size": 2048,
+                "intermediate_size": 16384,
+                "model_type": "gemma",
+                "num_attention_heads": 8,
+                "num_hidden_layers": 18,
+                "num_image_tokens": 256,
+                "num_key_value_heads": 1,
+                "torch_dtype": precision,
+                "vocab_size": 257152,
+                "_attn_implementation": "eager",
+            },
+            vision_config={
+                "hidden_size": 1152,
+                "intermediate_size": 4304,
+                "model_type": "siglip_vision_model",
+                "num_attention_heads": 16,
+                "num_hidden_layers": 27,
+                "num_image_tokens": 256,
+                "patch_size": 14,
+                "projection_dim": 2048,
+                "projector_hidden_act": "gelu_pytorch_tanh",
+                "torch_dtype": precision,
+                "vision_use_head": False,
+            },
+        )
+        self.pi0_paligemma = PaliGemmaForConditionalGeneration(config=paligemma_config)
+
+        self.pi0_paligemma.prepare_inputs_for_generation = partial(
+            prepare_inputs_for_generation, self=self.pi0_paligemma
+        )
+        # change important stuff in bf16
+        params_to_change_dtype = [
+            "language_model",
+            "vision_tower",
+            "multi_modal",
+        ]
+        for name, param in self.pi0_paligemma.named_parameters():
+            if any(selector in name for selector in params_to_change_dtype):
+                param.data = param.data.to(dtype=torch_precision)
+        self.set_requires_grad()
+        self.image_keys = self.config.image_features.keys()
+        # TODO: Remove this once we bump transformers to >4.52.0 because the attribute will be removed
+        # AttributeError: 'PaliGemmaConfig' object has no attribute 'ignore_index'
+        self.ignore_index = self.pi0_paligemma.config.ignore_index
+        self.padding_side = self.config.padding_side
+
+    def set_requires_grad(self):
+        if self.config.freeze_vision_encoder:
+            self.pi0_paligemma.vision_tower.eval()
+            for params in self.pi0_paligemma.vision_tower.parameters():
+                params.requires_grad = False
+        # To avoid unused params issue with distributed training
+        if self.config.freeze_lm_head:
+            for name, params in self.pi0_paligemma.named_parameters():
+                if "embed_tokens" in name:  # lm heads and embedding layer are tied
+                    params.requires_grad = False
+
+    def embed_tokens(self, tokens: torch.Tensor):
+        return self.pi0_paligemma.language_model.model.embed_tokens(tokens)
+
+    def prepare_inputs_for_generation(self, *args, **kwargs):
+        return self.pi0_paligemma.prepare_inputs_for_generation(*args, **kwargs)
+
+    def prepare_images(self, batch):
+        """Preprocess LeRobot batch into Pi0 inputs"""
+        images = []
+        img_masks = []
+        present_img_keys = [key for key in self.image_keys if key in batch]
+        if len(present_img_keys) == 0:
+            raise ValueError(
+                f"All image features are missing from the batch. At least one expected. (batch: {batch.keys()}) (image_features:{self.config.image_features})"
+            )
+
+        # Preprocess image features present in the batch
+        num_empty_cameras = 0
+        for key in self.image_keys:
+            if key in present_img_keys:
+                img = batch[key]
+
+                if self.config.resize_imgs_with_padding is not None:
+                    img = resize_with_pad(
+                        img,
+                        *self.config.resize_imgs_with_padding,
+                        pad_value=0,
+                        interpolate_like_pi=self.config.interpolate_like_pi,
+                    )
+
+                # Normalize from range [0,1] to [-1,1] as expected by siglip
+                img = img * 2.0 - 1.0
+
+                bsize = img.shape[0]
+                device = img.device
+                mask = torch.ones(bsize, dtype=torch.bool, device=device)
+            else:
+                if num_empty_cameras >= self.config.empty_cameras:
+                    continue
+                img = torch.ones_like(img) * -1
+                bsize = img.shape[0]
+                device = img.device
+                mask = torch.ones(bsize, dtype=torch.bool, device=device)
+                num_empty_cameras += 1
+
+            images.append(img)
+            img_masks.append(mask)
+        return images, img_masks
+
+    def normalize_actions(self, actions: torch.Tensor) -> torch.Tensor:
+        mins = actions.amin(dim=(1, 2), keepdim=True)  # [0]
+        maxs = actions.amax(dim=(1, 2), keepdim=True)  # [0]
+        return 2 * (actions - mins) / (maxs - mins + 1e-8) - 1
+
+    def _act_tokens_to_paligemma_tokens(self, tokens: torch.Tensor) -> torch.Tensor:
+        out = self.paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens - tokens
+        return out
+
+    def fast_tokenizer_wrapper(self, actions_norm):
+        """
+        A wrapper for self.fast_tokenizer that ensures batch processing,
+        conversion to PyTorch tensors, and returns a dictionary without padding.
+        """
+        batch_tokens = self.fast_tokenizer(actions_norm)
+        fast_out = self.processor.tokenizer.pad({"input_ids": batch_tokens}, return_tensors="pt")
+
+        return fast_out
+
+    def create_token_type_ids(self, padded_mask: torch.Tensor, prefix_len: int) -> torch.Tensor:
+        token_type_ids = torch.zeros_like(padded_mask, dtype=torch.bool)
+        # Compute cumulative sum mask
+        cumsum_mask = (padded_mask != 0).cumsum(dim=1)
+        # Suffix block (everything after prefix_len)
+        suffix_mask = cumsum_mask > prefix_len
+        token_type_ids = suffix_mask
+        return token_type_ids
+
+    def create_input_tokens(self, state, lang_text, actions=None):
+        bsize = state.shape[0]
+        device = state.device
+        bins = torch.linspace(-1, 1, 256 + 1, device=device)[:-1]
+        discretized = torch.bucketize(state, bins) - 1
+        discretized = discretized[:, :32]
+
+        prefix_texts = []
+        state_text = []
+        for txt, disc in zip(lang_text, discretized, strict=False):
+            cleaned = txt.lower().strip().replace("_", " ")
+            state_str = " ".join(str(val.item()) for val in disc)
+            prefix_texts.append(f"Task: {cleaned}, State: {state_str};\n")
+            state_text.append(f"State: {state_str};\n")
+
+        prefix_out = self.paligemma_tokenizer(
+            prefix_texts, add_special_tokens=True, return_tensors="pt", padding="longest", truncation=False
+        )
+        prefix_ids = prefix_out["input_ids"].to(device)
+        prefix_mask = prefix_out["attention_mask"].to(device)
+        prefix_lens = prefix_mask.sum(dim=1)[:, None].cpu()
+
+        if actions is not None:
+            actions_norm = self.normalize_actions(actions)
+            actions_pad = F.pad(
+                actions_norm, (0, max(0, self.config.max_action_dim - actions_norm.shape[2])), value=0
+            )[:, :, : self.config.max_action_dim]
+            fast_out = self.fast_tokenizer_wrapper(
+                actions_pad.cpu(),
+            )
+            act_ids = fast_out["input_ids"]
+            act_mask = fast_out["attention_mask"].to(device)
+
+            act_ids = self._act_tokens_to_paligemma_tokens(act_ids).to(device)
+            # Replace action with 0 to pad tokens
+            act_ids = torch.where(
+                act_ids == self.paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens,
+                self.pad_token_id,
+                act_ids,
+            )
+
+            eos_token = torch.tensor(
+                [self.paligemma_tokenizer.eos_token_id], dtype=torch.long, device=device
+            ).expand(bsize, -1)
+            eos_mask = torch.tensor([1], dtype=torch.long, device=device).expand(bsize, -1)
+            bos = self.paligemma_tokenizer("Action: ", add_special_tokens=False, return_tensors="pt")
+            bos_token = bos["input_ids"].expand(act_ids.shape[0], -1).to(device)
+            bos_mask = bos["attention_mask"].expand(act_ids.shape[0], -1).to(device)
+            act_ids = torch.cat([bos_token, act_ids, eos_token], dim=1)
+            act_mask = torch.cat([bos_mask, act_mask, eos_mask], dim=1)
+            act_mask = act_mask.to(device)
+        else:
+            act_ids = torch.empty(bsize, self.pad_token_id, dtype=torch.long, device=device)
+            act_mask = torch.empty(bsize, 0, dtype=torch.long, device=device)
+        final_ids = torch.cat([prefix_ids, act_ids], dim=1)
+
+        final_mask = torch.cat([prefix_mask, act_mask], dim=1)
+        batch_inputs = {"input_ids": final_ids.tolist(), "attention_mask": final_mask.tolist()}
+
+        # Use tokenizer pad function
+        padded_output = self.paligemma_tokenizer.pad(
+            batch_inputs, padding="longest", max_length=180, return_tensors="pt"
+        )
+        padded_mask = padded_output["attention_mask"]
+
+        # define tensor of padding lengths
+        att_mask = (padded_mask != 0).cumsum(dim=1) > prefix_lens
+
+        token_type_ids = self.create_token_type_ids(padded_mask=padded_mask, prefix_len=prefix_lens)
+
+        padded_output["padded_mask"] = padded_output.pop("attention_mask")
+        padded_output["attention_mask"] = att_mask
+        # loss is computed not on prefix, and not on padding
+        padded_output["loss_mask"] = att_mask & padded_output["padded_mask"]
+        padded_output["token_type_ids"] = token_type_ids
+        return padded_output
+
+    def shift_padding_side(
+        self,
+        tokens: torch.Tensor,
+        ar_mask: torch.Tensor,
+        padding_mask: torch.Tensor,
+        loss_mask: torch.Tensor,
+        targets: torch.Tensor,
+        token_type_ids: torch.Tensor,
+        padding_side: str = "right",
+    ) -> tuple[torch.Tensor]:
+        if padding_side not in ["right", "left"]:
+            return tokens, ar_mask, padding_mask, loss_mask, targets, token_type_ids
+
+        new_tokens = torch.empty_like(tokens)
+        new_ar_masks = torch.empty_like(ar_mask)
+        new_padding_mask = torch.empty_like(padding_mask)
+        new_loss_mask = torch.empty_like(loss_mask)
+        new_targets = torch.empty_like(targets)
+        new_token_type_ids = torch.empty_like(token_type_ids)
+        batch_size = tokens.shape[0]
+        for i in range(batch_size):
+            padding_indices = torch.where(padding_mask[i] == 0)[0]
+            non_padding_indices = torch.where(padding_mask[i] == 1)[0]
+            if padding_side == "left":
+                new_indices = torch.cat((padding_indices, non_padding_indices), dim=0)
+            else:
+                new_indices = torch.cat((non_padding_indices, padding_indices), dim=0)
+            new_tokens[i] = tokens[i].index_select(0, new_indices)
+            new_ar_masks[i] = ar_mask[i].index_select(0, new_indices)
+            new_padding_mask[i] = padding_mask[i].index_select(0, new_indices)
+            new_loss_mask[i] = loss_mask[i].index_select(0, new_indices)
+            new_targets[i] = targets[i].index_select(0, new_indices)
+            new_token_type_ids[i] = token_type_ids[i].index_select(0, new_indices)
+
+        return new_tokens, new_ar_masks, new_padding_mask, new_loss_mask, new_targets, new_token_type_ids
+
+    def forward(self, batch: dict[str, Tensor]):
+        device = batch[OBS_STATE].device
+        # TODO: keep like this or move to the policy .forward
+        images, img_masks = self.prepare_images(batch)
+
+        padded_outs = self.create_input_tokens(
+            state=batch[OBS_STATE],
+            lang_text=batch["task"],
+            actions=batch[ACTION],
+        )
+
+        embs, pad_masks, _, targets, loss_mask, token_type_ids = self.embed_inputs(
+            images,
+            img_masks,
+            padded_outs["input_ids"],
+            padded_outs["padded_mask"],
+            padded_outs["attention_mask"],
+            padded_outs["loss_mask"],
+            padded_outs["token_type_ids"],
+            padding_side=self.padding_side,
+        )
+        position_ids = torch.cumsum(pad_masks, dim=1) - 1
+        token_type_ids = token_type_ids.to(dtype=torch.int64)
+        past_seen_tokens = 0
+        cache_position = torch.arange(past_seen_tokens, past_seen_tokens + embs.shape[1], device=embs.device)
+        pad_masks = block_causal_update_causal_mask(
+            attention_mask=pad_masks,
+            past_key_values=None,
+            cache_position=cache_position,
+            input_tensor=embs,
+            token_type_ids=token_type_ids,
+            dtype=self.pi0_paligemma.dtype,
+            attn_implementation=self.pi0_paligemma.config.text_config._attn_implementation,
+        )
+        outputs = self.pi0_paligemma.forward(
+            input_ids=None,
+            token_type_ids=None,
+            attention_mask=pad_masks,
+            position_ids=position_ids,
+            past_key_values=None,
+            inputs_embeds=embs,
+            use_cache=False,
+            labels=None,
+        )
+
+        logits = outputs.logits
+
+        loss_fct = nn.CrossEntropyLoss(reduction="none")
+
+        # Shift left for next-step prediction
+        logits = logits[:, :-1, :]
+        targets = targets[:, 1:].to(device)  # Shift targets
+        loss_mask = loss_mask[:, 1:].to(device)  # Ensure correct shape
+
+        # Compute per-token loss
+        token_loss = loss_fct(logits.reshape(-1, logits.shape[-1]), targets.reshape(-1))
+
+        # Apply loss mask
+        token_loss = token_loss * loss_mask.reshape(-1)
+
+        # Compute final loss
+        loss = token_loss.sum() / torch.clamp(loss_mask.sum(), min=1)
+
+        # Return loss dictionary
+        loss_dict = {"ce_loss": loss.item(), "loss": loss}
+        return loss_dict
+
+    def decode_actions_with_fast(
+        self,
+        tokens: list[list[int]],
+        *,
+        time_horizon: int | None = None,
+        action_dim: int | None = None,
+        relaxed_decoding: bool = True,
+    ) -> np.array:
+        """
+        Adapt original decoding in FAST to always return actions instead of zeros.
+        """
+        self.time_horizon = (
+            time_horizon or self.fast_tokenizer.time_horizon or self.fast_tokenizer.called_time_horizon
+        )
+        self.action_dim = (
+            action_dim or self.fast_tokenizer.action_dim or self.fast_tokenizer.called_action_dim
+        )
+
+        # Cache the time horizon and action dimension for the next call
+        self.called_time_horizon = self.time_horizon
+        self.called_action_dim = self.action_dim
+
+        assert self.time_horizon is not None and self.action_dim is not None, (
+            "Tokenizer not initialized, call encode() once or pass in time_horizon and action_dim."
+        )
+
+        decoded_actions = []
+        for token in tokens:
+            try:
+                decoded_tokens = self.fast_tokenizer.bpe_tokenizer.decode(token)
+                decoded_dct_coeff = np.array(list(map(ord, decoded_tokens))) + self.fast_tokenizer.min_token
+                if relaxed_decoding:
+                    # Expected sequence length
+                    expected_seq_len = self.time_horizon * self.action_dim
+                    diff = expected_seq_len - decoded_dct_coeff.shape[0]
+                    # Apply truncation if too long
+                    if diff < 0:
+                        decoded_dct_coeff = decoded_dct_coeff[:expected_seq_len]  # Truncate on the right
+                    # Apply padding if too short
+                    elif diff > 0:
+                        decoded_dct_coeff = np.pad(
+                            decoded_dct_coeff, (0, diff), mode="constant", constant_values=0
+                        )
+
+                decoded_dct_coeff = decoded_dct_coeff.reshape(-1, self.action_dim)
+                assert decoded_dct_coeff.shape == (
+                    self.time_horizon,
+                    self.action_dim,
+                ), (
+                    f"Decoded DCT coefficients have shape {decoded_dct_coeff.shape}, expected ({self.time_horizon}, {self.action_dim})"
+                )
+            except Exception as e:
+                print(f"Error decoding tokens: {e}")
+                print(f"Tokens: {token}")
+                decoded_dct_coeff = np.zeros((self.time_horizon, self.action_dim))
+            decoded_actions.append(idct(decoded_dct_coeff / self.fast_tokenizer.scale, axis=0, norm="ortho"))
+        return np.stack(decoded_actions)
+
+    def extract_actions(self, tokens: torch.Tensor, action_horizon: int, action_dim: int) -> torch.Tensor:
+        """
+        Extracts actions from predicted output tokens using the FAST model.
+
+        Args:
+            tokens (torch.Tensor): The input tensor of tokenized outputs.
+            action_horizon (int): The number of timesteps for actions.
+            action_dim (int): The dimensionality of each action.
+
+        Returns:
+            torch.Tensor: The extracted actions as a tensor of shape (action_horizon, action_dim).
+        """
+        # Decode predicted output tokens
+        decoded_tokens = self.paligemma_tokenizer.batch_decode(tokens, skip_special_tokens=True)
+        cleaned_tokens = [
+            tokens_sequence.replace("Action:", "").replace(":", "").strip().split("|")[0].strip()
+            for tokens_sequence in decoded_tokens
+        ]
+        raw_action_tokens = [
+            self.processor.tokenizer.encode(sample_tokens, return_tensors="pt", padding=False)
+            for sample_tokens in cleaned_tokens
+        ]  # something like this should be robust #looks good
+        action_tokens = [
+            self._act_tokens_to_paligemma_tokens(raw_action_token) for raw_action_token in raw_action_tokens
+        ]
+        # returns the tensor of decoded actions per sample in a list
+        decoded_actions = [
+            torch.tensor(
+                self.decode_actions_with_fast(
+                    tok.tolist(),
+                    time_horizon=action_horizon,
+                    action_dim=action_dim,
+                    relaxed_decoding=self.config.relaxed_action_decoding,
+                ),
+                device=tokens.device,
+            ).squeeze(0)
+            for tok in action_tokens
+        ]
+
+        return torch.stack(
+            decoded_actions,
+            dim=0,
+        )
+
+    def generate_actions(self, batch: dict[str, Tensor]):
+        # TODO: keep like this or move to the policy .forward
+        images, img_masks = self.prepare_images(batch)
+
+        padded_outs = self.create_input_tokens(state=batch[OBS_STATE], lang_text=batch["task"], actions=None)
+        embs, pad_masks, att_masks2, targets, loss_mask, token_type_ids = self.embed_inputs(
+            images,
+            img_masks,
+            padded_outs["input_ids"],
+            padded_outs["padded_mask"],
+            padded_outs["attention_mask"],
+            padded_outs["loss_mask"],
+            padded_outs["token_type_ids"],
+            padding_side="left",
+        )
+        token_type_ids = token_type_ids.to(dtype=torch.int64)
+        prefix_position_ids = torch.cumsum(pad_masks, dim=1) - 1
+        output_tokens = self.pi0_paligemma.generate(
+            input_ids=None,
+            attention_mask=pad_masks,
+            position_ids=prefix_position_ids,
+            past_key_values=None,
+            inputs_embeds=embs,
+            use_cache=self.config.use_cache,
+            max_new_tokens=self.config.max_decoding_steps,
+            do_sample=False,
+            num_beams=1,
+            token_type_ids=token_type_ids,
+        )
+        actions = self.extract_actions(output_tokens, self.action_horizon, self.action_dim)
+        return actions
+
+    def embed_image(self, image: torch.Tensor):
+        # Handle different transformers versions
+        if hasattr(self.pi0_paligemma, "get_image_features"):
+            return self.pi0_paligemma.get_image_features(image)
+        else:
+            return self.pi0_paligemma.model.get_image_features(image)
+
+    def embed_inputs(
+        self,
+        images,
+        img_masks,
+        tokens,
+        pad_mask,
+        ar_mask,
+        loss_mask,
+        token_type_ids,
+        padding_side: str = "right",
+    ):
+        # TODO: avoid list in python and torch.cat ; prefer pre-allocation with torch.empty
+        # images are a list of same size
+        # vectorizing everything!
+        device = images[0].device
+        image_embedding_dim = images[0].shape[-1]  # TODO should be from self.config
+        all_images = torch.stack(images, dim=1).to(device)
+        b, n, c, h, w = all_images.shape
+        all_images = all_images.view(b * n, c, h, w)
+        embedded = self.embed_image(all_images).to(device)
+        b_n, p, image_embedding_dim = embedded.shape  # Extract current dimensions
+        m = b_n // b  # Compute the number of images per sample dynamically
+
+        # Reshape dynamically
+        embedded = embedded.view(b, m, p, image_embedding_dim)
+        tokens_embs = self.embed_tokens(tokens.to(device))
+
+        img_masks = torch.stack(img_masks, dim=1).unsqueeze(-1).to(device)
+        num_img_emb = embedded.shape[2]
+        img_pad_masks = img_masks.repeat(1, 1, num_img_emb).view(b, -1)
+        img_att_masks = torch.zeros((b, n, num_img_emb), dtype=torch.long, device=device).reshape(b, -1)
+
+        image_target_tokens = (
+            torch.ones((b, n, num_img_emb), dtype=torch.long, device=device) * self.pad_token_id
+        ).reshape(b, -1)
+        image_loss_mask = torch.zeros((b, n, num_img_emb), dtype=torch.long, device=device).reshape(b, -1)
+
+        embedded = embedded.reshape(b, n * num_img_emb, image_embedding_dim)  # Shape: (B, N*P, D)
+
+        embs = torch.cat([embedded, tokens_embs], dim=1).to(device)
+        pad_masks = torch.cat([img_pad_masks, pad_mask.to(device)], dim=1)
+        att_masks = torch.cat([img_att_masks, ar_mask.to(device)], dim=1)
+        loss_masks = torch.cat([image_loss_mask, loss_mask.to(device)], dim=1)
+        targets = torch.cat([image_target_tokens, tokens.to(device)], dim=1)
+        token_type_ids = torch.cat([img_att_masks, token_type_ids.to(device)], dim=1)
+
+        # Shift pad tokens to the left (.generate()) or right (.train())
+        embs, att_masks, pad_masks, loss_masks, targets, token_type_ids = self.shift_padding_side(
+            embs, att_masks, pad_masks, loss_masks, targets, token_type_ids, padding_side=padding_side
+        )
+
+        targets = torch.where(targets == self.pad_token_id, self.ignore_index, targets)
+        return embs, pad_masks, att_masks, targets, loss_masks, token_type_ids
+
+
+def resize_with_pad(img, width, height, pad_value=0, interpolate_like_pi=True):
+    # assume no-op when width height fits already
+    if img.ndim != 4:
+        raise ValueError(f"(b,c,h,w) expected, but {img.shape}")
+
+    cur_height, cur_width = img.shape[2:]
+
+    ratio = max(cur_width / width, cur_height / height)
+    resized_height = int(cur_height / ratio)
+    resized_width = int(cur_width / ratio)
+
+    if interpolate_like_pi:
+        img = (img * 255.0).to(dtype=torch.uint8)
+        img = img.permute(0, 2, 3, 1)
+        original_device = img.device
+        img = img.to(device="cpu").numpy()
+        imgs = []
+        for sub_img in img:
+            sub_img = Image.fromarray(sub_img)
+            resized_img = sub_img.resize((resized_width, resized_height), resample=2)
+            resized_img = torch.from_numpy(np.array(resized_img))
+            imgs.append(resized_img)
+        img = torch.stack(imgs, dim=0)
+        img = img.permute(0, 3, 1, 2)
+        resized_img = img.to(device=original_device, dtype=torch.float32) / 255.0
+    else:
+        resized_img = F.interpolate(
+            img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
+        )
+
+    pad_height = max(0, int(height - resized_height))
+    pad_width = max(0, int(width - resized_width))
+
+    # pad on left and top of image
+    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
+    return padded_img

src/lerobot/policies/pi0fast/processor_pi05.py

@@ -0,0 +1,171 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from copy import deepcopy
+from dataclasses import dataclass
+from typing import Any
+
+import numpy as np
+import torch
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.policies.pi05.configuration_pi05 import PI05Config
+from lerobot.policies.pi05.modeling_pi05 import pad_vector
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    RenameObservationsProcessorStep,
+    TokenizerProcessorStep,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.processor.core import EnvTransition, TransitionKey
+from lerobot.utils.constants import (
+    OBS_STATE,
+    POLICY_POSTPROCESSOR_DEFAULT_NAME,
+    POLICY_PREPROCESSOR_DEFAULT_NAME,
+)
+
+
+@ProcessorStepRegistry.register(name="pi05_prepare_state_tokenizer_processor_step")
+@dataclass
+class Pi05PrepareStateTokenizerProcessorStep(ProcessorStep):
+    """
+    Processor step to prepare the state and tokenize the language input.
+    """
+
+    max_state_dim: int = 32
+    task_key: str = "task"
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        transition = transition.copy()
+
+        state = transition.get(TransitionKey.OBSERVATION, {}).get(OBS_STATE)
+        if state is None:
+            raise ValueError("State is required for PI05")
+        tasks = transition.get(TransitionKey.COMPLEMENTARY_DATA, {}).get(self.task_key)
+        if tasks is None:
+            raise ValueError("No task found in complementary data")
+
+        # TODO: check if this necessary
+        state = deepcopy(state)
+
+        # Prepare state (pad to max_state_dim)
+        state = pad_vector(state, self.max_state_dim)
+
+        # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
+        # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
+        state_np = state.cpu().numpy()
+        discretized_states = np.digitize(state_np, bins=np.linspace(-1, 1, 256 + 1)[:-1]) - 1
+
+        full_prompts = []
+        for i, task in enumerate(tasks):
+            cleaned_text = task.strip().replace("_", " ").replace("\n", " ")
+            state_str = " ".join(map(str, discretized_states[i]))
+            full_prompt = f"Task: {cleaned_text}, State: {state_str};\nAction: "
+            full_prompts.append(full_prompt)
+
+        transition[TransitionKey.COMPLEMENTARY_DATA][self.task_key] = full_prompts
+        # Normalize state to [-1, 1] range if needed (assuming it's already normalized by normalizer processor step!!)
+        # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
+        return transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        This step does not alter the feature definitions.
+        """
+        return features
+
+
+def make_pi05_pre_post_processors(
+    config: PI05Config,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Constructs pre-processor and post-processor pipelines for the PI0 policy.
+
+    The pre-processing pipeline prepares input data for the model by:
+    1. Renaming features to match pretrained configurations.
+    2. Normalizing input and output features based on dataset statistics.
+    3. Adding a batch dimension.
+    4. Appending a newline character to the task description for tokenizer compatibility.
+    5. Tokenizing the text prompt using the PaliGemma tokenizer.
+    6. Moving all data to the specified device.
+
+    The post-processing pipeline handles the model's output by:
+    1. Moving data to the CPU.
+    2. Unnormalizing the output features to their original scale.
+
+    Args:
+        config: The configuration object for the PI0 policy.
+        dataset_stats: A dictionary of statistics for normalization.
+        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
+        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
+
+    Returns:
+        A tuple containing the configured pre-processor and post-processor pipelines.
+    """
+
+    # Add remaining processors
+    input_steps: list[ProcessorStep] = [
+        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
+        AddBatchDimensionProcessorStep(),
+        # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
+        # because the tokenizer step expects normalized state in [-1, 1] range for discretization
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        Pi05PrepareStateTokenizerProcessorStep(max_state_dim=config.max_state_dim),
+        TokenizerProcessorStep(
+            tokenizer_name="google/paligemma-3b-pt-224",
+            max_length=config.tokenizer_max_length,
+            padding_side="right",
+            padding="max_length",
+        ),
+        DeviceProcessorStep(device=config.device),
+    ]
+
+    output_steps: list[ProcessorStep] = [
+        UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/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}
+}
+```

src/lerobot/policies/sarm/compute_rabc_weights.py

@@ -0,0 +1,870 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Compute SARM progress values for RA-BC (Reward-Aware Behavior Cloning) weighting.
+
+This script processes all frames in a dataset with SARM to compute progress values [0, 1].
+The results are saved as a parquet file that can be loaded during training for RA-BC weighting.
+
+Uses multi-output extraction: each SARM query returns progress for 9 frames, so we only
+need ~num_frames/30 queries instead of one per frame (~30x speedup).
+
+Usage:
+    # Full RA-BC computation with visualizations
+    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
+        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
+        --reward-model-path pepijn223/sarm_single_uni4
+
+    # Faster computation with stride (compute every 5 frames, interpolate the rest)
+    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
+        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
+        --stride 5
+
+    # Visualize predictions only (no RA-BC computation)
+    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
+        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
+        --visualize-only \\
+        --num-visualizations 5
+
+The output is saved to the dataset's local cache directory as 'sarm_progress.parquet'.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import matplotlib.gridspec as gridspec
+import matplotlib.pyplot as plt
+import numpy as np
+import pyarrow as pa
+import pyarrow.parquet as pq
+import torch
+from tqdm import tqdm
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.policies.sarm.modeling_sarm import SARMRewardModel
+from lerobot.policies.sarm.processor_sarm import make_sarm_pre_post_processors
+from lerobot.policies.sarm.sarm_utils import normalize_stage_tau
+
+
+def get_reward_model_path_from_parquet(parquet_path: Path) -> str | None:
+    """Read reward_model_path from parquet metadata if available."""
+    if not parquet_path.exists():
+        return None
+    try:
+        metadata = pq.read_metadata(parquet_path).schema.to_arrow_schema().metadata
+        if metadata and b"reward_model_path" in metadata:
+            return metadata[b"reward_model_path"].decode()
+    except Exception:  # nosec B110
+        return None
+    return None
+
+
+def load_sarm_resources(
+    dataset_repo_id: str,
+    reward_model_path: str,
+    device: str = "cuda",
+) -> tuple[LeRobotDataset, SARMRewardModel, any]:
+    """
+    Load SARM model, dataset, and preprocessor.
+
+    Returns:
+        Tuple of (dataset, reward_model, preprocessor)
+    """
+    logging.info(f"Loading model: {reward_model_path}")
+    reward_model = SARMRewardModel.from_pretrained(reward_model_path)
+    reward_model.config.device = device
+    reward_model.to(device).eval()
+
+    image_key = reward_model.config.image_key
+    state_key = reward_model.config.state_key
+    delta_indices = reward_model.config.observation_delta_indices
+
+    logging.info(f"Loading dataset: {dataset_repo_id}")
+    temp_dataset = LeRobotDataset(dataset_repo_id, download_videos=True)
+    fps = temp_dataset.fps
+
+    delta_timestamps = {
+        image_key: [idx / fps for idx in delta_indices],
+        state_key: [idx / fps for idx in delta_indices],
+    }
+    dataset = LeRobotDataset(dataset_repo_id, delta_timestamps=delta_timestamps)
+    logging.info(f"Dataset: {dataset.num_episodes} episodes, {dataset.num_frames} frames")
+
+    preprocess, _ = make_sarm_pre_post_processors(
+        config=reward_model.config,
+        dataset_stats=dataset.meta.stats,
+        dataset_meta=dataset.meta,
+    )
+
+    return dataset, reward_model, preprocess
+
+
+def to_numpy_image(img) -> np.ndarray:
+    """Convert image tensor to numpy uint8 (H, W, C)."""
+    if isinstance(img, torch.Tensor):
+        img = img.cpu().numpy()
+    if img.ndim == 4:
+        # Take center frame for bidirectional sampling
+        img = img[img.shape[0] // 2]
+    if img.shape[0] in [1, 3]:
+        img = np.transpose(img, (1, 2, 0))
+    if img.dtype != np.uint8:
+        # Handle normalized images (may have negative values or values > 1)
+        img = img.astype(np.float32)
+        img = (img - img.min()) / (img.max() - img.min() + 1e-8)  # Normalize to [0, 1]
+        img = (img * 255).astype(np.uint8)
+    return img
+
+
+def visualize_episode(
+    frames, progress_preds, stage_preds, title, output_path, stage_labels, gt_progress=None, gt_stages=None
+):
+    """Create visualization with progress plot, stage probabilities, and sample frames.
+
+    Same as sarm_inference_visualization.py
+    """
+    num_stages = stage_preds.shape[1]
+    colors = plt.cm.tab10(np.linspace(0, 1, num_stages))
+    frame_indices = np.arange(len(progress_preds))
+
+    fig = plt.figure(figsize=(14, 12))
+    gs = gridspec.GridSpec(3, 1, height_ratios=[2, 1, 1], hspace=0.3)
+    ax_progress, ax_stages, ax_frames = fig.add_subplot(gs[0]), fig.add_subplot(gs[1]), fig.add_subplot(gs[2])
+
+    # Progress plot
+    ax_progress.plot(frame_indices, progress_preds, linewidth=2, color="#2E86AB", label="Predicted")
+    ax_progress.fill_between(frame_indices, 0, progress_preds, alpha=0.3, color="#2E86AB")
+    if gt_progress is not None:
+        ax_progress.plot(
+            frame_indices, gt_progress, linewidth=2, color="#28A745", linestyle="--", label="Ground Truth"
+        )
+    ax_progress.axhline(y=1.0, color="gray", linestyle="--", alpha=0.5)
+    ax_progress.set_ylabel("Progress")
+    ax_progress.set_title(f'Task: "{title}"', fontweight="bold")
+    ax_progress.set_ylim(-0.05, 1.1)
+    ax_progress.legend(loc="upper left")
+    ax_progress.grid(True, alpha=0.3)
+
+    # Stage predictions
+    ax_stages.stackplot(
+        frame_indices,
+        *[stage_preds[:, i] for i in range(num_stages)],
+        colors=colors,
+        alpha=0.8,
+        labels=stage_labels,
+    )
+    if gt_stages is not None:
+        for change_idx in np.where(np.diff(gt_stages) != 0)[0] + 1:
+            ax_stages.axvline(x=change_idx, color="black", linestyle="-", alpha=0.7, linewidth=1.5)
+    ax_stages.set_xlabel("Frame")
+    ax_stages.set_ylabel("Stage Probability")
+    ax_stages.set_ylim(0, 1)
+    ax_stages.legend(loc="upper left", ncol=min(num_stages, 5), fontsize=8)
+    ax_stages.grid(True, alpha=0.3)
+
+    # Sample frames
+    ax_frames.axis("off")
+    num_sample = 8
+    sample_indices = np.linspace(0, len(frames) - 1, num_sample, dtype=int)
+    h, w = frames[0].shape[:2]
+    combined = np.zeros((h, w * num_sample, 3), dtype=np.uint8)
+    for i, idx in enumerate(sample_indices):
+        frame = frames[idx]
+        if frame.shape[-1] == 1:
+            frame = np.repeat(frame, 3, axis=-1)
+        combined[:, i * w : (i + 1) * w] = frame
+        stage_name = stage_labels[np.argmax(stage_preds[idx])][:12]
+        ax_frames.text(
+            i * w + w / 2,
+            -10,
+            f"Frame {idx}\n{progress_preds[idx]:.2f}\n{stage_name}",
+            ha="center",
+            va="top",
+            fontsize=7,
+        )
+    ax_frames.imshow(combined)
+    ax_frames.set_title("Sample Frames", pad=20)
+
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    plt.savefig(output_path, dpi=150, bbox_inches="tight")
+    plt.close()
+    print(f"Saved: {output_path}")
+
+
+def visualize_sarm_predictions(
+    dataset: LeRobotDataset,
+    reward_model: SARMRewardModel,
+    preprocess,
+    episode_indices: list[int],
+    head_mode: str,
+    output_dir: Path,
+    num_display_frames: int = 5,
+    stride: int = 1,
+):
+    """
+    Visualize SARM predictions for multiple episodes.
+
+    Computes predictions for every frame by default. With stride > 1, computes predictions
+    every N frames and interpolates (progress + stage probabilities) for visualization.
+
+    Args:
+        dataset: LeRobotDataset with delta_timestamps configured
+        reward_model: Loaded SARM model
+        preprocess: Preprocessor from make_sarm_pre_post_processors
+        episode_indices: List of episode indices to visualize
+        head_mode: "sparse", "dense", or "both"
+        output_dir: Directory to save visualizations
+        num_display_frames: Number of frames to display in thumbnail strip (default: 5)
+        stride: Compute predictions every N frames, interpolate the rest (default: 1)
+    """
+    output_dir = Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    image_key = reward_model.config.image_key
+    state_key = reward_model.config.state_key
+    dual_mode = reward_model.config.uses_dual_heads
+    device = reward_model.device
+
+    # Center frame index for bidirectional sampling
+    target_idx = reward_model.config.n_obs_steps // 2
+
+    # Determine which heads to visualize
+    schemes_to_viz = []
+    if head_mode in ("sparse", "both") or not dual_mode:
+        schemes_to_viz.append("sparse")
+    if head_mode in ("dense", "both") and dual_mode:
+        schemes_to_viz.append("dense")
+
+    # Set preprocessor to eval mode to disable augmentations
+    if hasattr(preprocess, "eval"):
+        preprocess.eval()
+    for step in preprocess.steps:
+        if hasattr(step, "eval"):
+            step.eval()
+
+    for episode_idx in episode_indices:
+        ep = dataset.meta.episodes[episode_idx]
+        ep_start = ep["dataset_from_index"]
+        ep_end = ep["dataset_to_index"]
+        task = dataset[ep_start].get("task", "perform the task")
+        num_frames = ep_end - ep_start
+
+        # Select frames for display thumbnails (evenly sampled from begin to end)
+        display_indices = set(
+            [
+                ep_start + int(i * (num_frames - 1) / (num_display_frames - 1))
+                for i in range(num_display_frames)
+            ]
+            if num_frames >= num_display_frames
+            else list(range(ep_start, ep_end))
+        )
+        viz_frames = {}
+
+        # Load display frames up-front (stride mode might skip them otherwise).
+        for frame_idx in display_indices:
+            sample = dataset[frame_idx]
+            viz_frames[frame_idx] = to_numpy_image(sample[image_key])
+
+        # Initialize storage for each scheme
+        scheme_data = {}
+        for scheme in schemes_to_viz:
+            num_stages = getattr(reward_model.config, f"num_{scheme}_stages")
+            scheme_data[scheme] = {
+                "viz_progress": np.full(num_frames, np.nan),
+                "viz_stages": np.full((num_frames, num_stages), np.nan),
+                "viz_gt_progress": np.full(num_frames, np.nan),
+                "viz_gt_stages": np.full(num_frames, np.nan),
+                "target_key": f"{scheme}_targets",
+                "num_stages": num_stages,
+                "temporal_props": getattr(reward_model.config, f"{scheme}_temporal_proportions"),
+                "subtask_names": getattr(reward_model.config, f"{scheme}_subtask_names"),
+            }
+
+        if stride > 1:
+            logging.info(f"Visualization stride={stride}: inferring every {stride} frames and interpolating")
+
+        # Process frames one at a time to avoid memory buildup
+        frame_indices = list(range(ep_start, ep_end, stride))
+        if (ep_end - 1) not in frame_indices:
+            frame_indices.append(ep_end - 1)
+        frame_indices = sorted(set(frame_indices))
+
+        for frame_idx in tqdm(frame_indices, desc=f"Episode {episode_idx}", leave=False):
+            local_idx = frame_idx - ep_start
+            sample = dataset[frame_idx]
+
+            batch = {
+                image_key: sample[image_key],
+                "task": task,
+                "index": frame_idx,
+                "episode_index": episode_idx,
+            }
+            if state_key in sample:
+                batch[state_key] = sample[state_key]
+
+            with torch.no_grad():
+                processed = preprocess(batch)
+                video_features = processed["video_features"].to(device)
+                text_features = processed["text_features"].to(device)
+                state_features = processed.get("state_features")
+                if state_features is not None:
+                    state_features = state_features.to(device)
+                lengths = processed.get("lengths")
+
+                for scheme in schemes_to_viz:
+                    sd = scheme_data[scheme]
+
+                    # Ground truth
+                    # In stride visualization mode, ground-truth plots can be misleading
+                    # (only sparse points are available), so we skip GT.
+                    if stride == 1 and sd["target_key"] in processed:
+                        gt_target = processed[sd["target_key"]][0, target_idx].cpu().item()
+                        sd["viz_gt_stages"][local_idx] = int(gt_target)
+                        sd["viz_gt_progress"][local_idx] = normalize_stage_tau(
+                            gt_target,
+                            num_stages=sd["num_stages"],
+                            temporal_proportions=sd["temporal_props"],
+                            subtask_names=sd["subtask_names"],
+                        )
+
+                    # Predictions
+                    reward, stage_probs = reward_model.calculate_rewards(
+                        text_embeddings=text_features,
+                        video_embeddings=video_features,
+                        state_features=state_features,
+                        lengths=lengths,
+                        return_all_frames=True,
+                        return_stages=True,
+                        head_mode=scheme,
+                    )
+
+                    # Handle both tensor and numpy outputs
+                    if isinstance(reward, torch.Tensor):
+                        reward = reward.cpu().numpy()
+                        stage_probs = stage_probs.cpu().numpy()
+
+                    if reward.ndim == 2:
+                        sd["viz_progress"][local_idx] = reward[0, target_idx]
+                        sd["viz_stages"][local_idx] = stage_probs[0, target_idx, :]
+                    else:
+                        sd["viz_progress"][local_idx] = reward[target_idx]
+                        sd["viz_stages"][local_idx] = stage_probs[target_idx, :]
+
+                # Clear GPU memory after each frame
+                del processed, video_features, text_features
+                if state_features is not None:
+                    del state_features
+
+            torch.cuda.empty_cache()
+
+        # Interpolate predictions back to per-frame arrays for smooth visualization.
+        if stride > 1:
+            all_local = np.arange(num_frames)
+            for scheme in schemes_to_viz:
+                sd = scheme_data[scheme]
+
+                valid = np.isfinite(sd["viz_progress"])
+                valid_idx = np.where(valid)[0]
+                if valid_idx.size >= 1:
+                    sd["viz_progress"] = interpolate_progress(
+                        valid_idx, sd["viz_progress"][valid_idx], all_local
+                    )
+
+                    stage_interp = np.zeros_like(sd["viz_stages"], dtype=np.float32)
+                    for s in range(sd["num_stages"]):
+                        stage_interp[:, s] = interpolate_progress(
+                            valid_idx, sd["viz_stages"][valid_idx, s], all_local
+                        )
+
+                    stage_interp = np.clip(stage_interp, 0.0, 1.0)
+                    row_sums = stage_interp.sum(axis=1, keepdims=True)
+                    nz = row_sums.squeeze(-1) > 0
+                    stage_interp[nz] = stage_interp[nz] / row_sums[nz]
+                    sd["viz_stages"] = stage_interp
+                else:
+                    # No valid points: keep NaNs/zeros; visualization will be empty.
+                    sd["viz_stages"] = np.nan_to_num(sd["viz_stages"], nan=0.0)
+
+        # Generate visualization for each head
+        ordered_viz_frames = [viz_frames[idx] for idx in sorted(display_indices)]
+        for scheme in schemes_to_viz:
+            sd = scheme_data[scheme]
+            stage_labels = sd["subtask_names"] or [f"Stage {i + 1}" for i in range(sd["num_stages"])]
+            viz_path = output_dir / f"sarm_prediction_ep{episode_idx}_{scheme}.png"
+
+            visualize_episode(
+                frames=np.array(ordered_viz_frames),
+                progress_preds=sd["viz_progress"],
+                stage_preds=sd["viz_stages"],
+                title=f"{task} (Episode {episode_idx})",
+                output_path=viz_path,
+                stage_labels=stage_labels,
+                gt_progress=sd["viz_gt_progress"] if not np.all(np.isnan(sd["viz_gt_progress"])) else None,
+                gt_stages=sd["viz_gt_stages"] if not np.all(np.isnan(sd["viz_gt_stages"])) else None,
+            )
+
+        # Clear memory between episodes
+        torch.cuda.empty_cache()
+
+    logging.info(f"Visualizations saved to: {output_dir.absolute()}")
+
+
+def generate_all_frame_indices(ep_start: int, ep_end: int, frame_gap: int = 30) -> list[int]:
+    """Generate all frame indices, ordered by offset for cache-friendly access.
+
+    Orders frames as: [0, 30, 60...], [1, 31, 61...], ..., [29, 59, 89...]
+    This groups frames that share similar temporal windows together.
+    """
+    num_frames = ep_end - ep_start
+    indices = []
+    for offset in range(frame_gap):
+        for frame_rel in range(offset, num_frames, frame_gap):
+            indices.append(ep_start + frame_rel)
+    return indices
+
+
+def interpolate_progress(
+    computed_indices: np.ndarray,
+    computed_values: np.ndarray,
+    all_indices: np.ndarray,
+) -> np.ndarray:
+    """Linearly interpolate values to fill in gaps (robust to NaNs / edge cases)."""
+    computed_indices = np.asarray(computed_indices)
+    computed_values = np.asarray(computed_values)
+    all_indices = np.asarray(all_indices)
+
+    mask = np.isfinite(computed_values)
+    if mask.sum() == 0:
+        return np.full(all_indices.shape, np.nan, dtype=np.float32)
+    if mask.sum() == 1:
+        return np.full(all_indices.shape, float(computed_values[mask][0]), dtype=np.float32)
+
+    out = np.interp(all_indices, computed_indices[mask], computed_values[mask])
+    return out.astype(np.float32)
+
+
+def compute_sarm_progress(
+    dataset_repo_id: str,
+    reward_model_path: str,
+    output_path: str | None = None,
+    head_mode: str = "sparse",
+    device: str = "cuda",
+    num_visualizations: int = 5,
+    output_dir: str = "./sarm_viz",
+    stride: int = 1,
+):
+    """
+    Compute SARM progress predictions for all frames in a dataset.
+
+    Args:
+        dataset_repo_id: HuggingFace dataset repo ID or local path
+        reward_model_path: Path to pretrained SARM model
+        output_path: Path to save results. If None, saves to dataset's cache directory
+        head_mode: SARM head to use ("sparse", "dense", or "both")
+        device: Device to use for inference
+        num_visualizations: Number of episodes to visualize (0 to skip)
+        output_dir: Directory to save visualizations
+        stride: Compute progress every N frames, interpolate the rest (default: 1 = every frame)
+    """
+    dataset, reward_model, preprocess = load_sarm_resources(dataset_repo_id, reward_model_path, device)
+
+    # Set preprocessor to eval mode to disable augmentations
+    if hasattr(preprocess, "eval"):
+        preprocess.eval()
+    for step in preprocess.steps:
+        if hasattr(step, "eval"):
+            step.eval()
+
+    image_key = reward_model.config.image_key
+    state_key = reward_model.config.state_key
+    frame_gap = reward_model.config.frame_gap
+    num_episodes = dataset.num_episodes
+    total_frames = dataset.num_frames
+    logging.info(f"Processing {total_frames} frames across {num_episodes} episodes")
+
+    # Determine which heads to compute
+    dual_mode = reward_model.config.uses_dual_heads
+    compute_sparse = head_mode in ("sparse", "both") or not dual_mode
+    compute_dense = head_mode in ("dense", "both") and dual_mode
+
+    # Storage arrays
+    all_indices = []
+    all_episode_indices = []
+    all_frame_indices = []
+    all_progress_sparse = [] if compute_sparse else None
+    all_progress_dense = [] if compute_dense else None
+
+    if stride > 1:
+        logging.info(f"Using stride={stride}: computing every {stride} frames, interpolating the rest")
+
+    # Process all episodes
+    for episode_idx in tqdm(range(num_episodes), desc="Episodes"):
+        ep = dataset.meta.episodes[episode_idx]
+        ep_start = ep["dataset_from_index"]
+        ep_end = ep["dataset_to_index"]
+
+        # Get task description
+        task = dataset[ep_start].get("task", "perform the task")
+
+        # Generate frames to compute (with stride applied)
+        all_ep_indices = generate_all_frame_indices(ep_start, ep_end, frame_gap)
+        if stride > 1:
+            # Only compute every stride-th frame (relative to episode start)
+            compute_indices = [idx for idx in all_ep_indices if (idx - ep_start) % stride == 0]
+            # Always include last frame for better interpolation at episode end
+            last_frame = ep_end - 1
+            if last_frame not in compute_indices:
+                compute_indices.append(last_frame)
+            compute_indices = sorted(set(compute_indices))
+        else:
+            compute_indices = all_ep_indices
+
+        center_idx = reward_model.config.n_obs_steps // 2  # Center of bidirectional window
+
+        # Dictionary to collect results
+        frame_results = {}
+
+        for query_idx in tqdm(compute_indices, desc=f"  Ep {episode_idx}", leave=False):
+            try:
+                sample = dataset[query_idx]
+
+                batch = {
+                    image_key: sample[image_key],
+                    "task": task,
+                    "index": query_idx,
+                    "episode_index": episode_idx,
+                }
+                if state_key in sample:
+                    batch[state_key] = sample[state_key]
+
+                with torch.no_grad():
+                    processed = preprocess(batch)
+                    video_features = processed["video_features"].to(device)
+                    text_features = processed["text_features"].to(device)
+                    state_features = processed.get("state_features")
+                    if state_features is not None:
+                        state_features = state_features.to(device)
+                    lengths = processed.get("lengths")
+
+                    sparse_val = np.nan
+                    dense_val = np.nan
+
+                    # Compute sparse prediction for center frame
+                    if compute_sparse:
+                        sparse_progress = reward_model.calculate_rewards(
+                            text_embeddings=text_features,
+                            video_embeddings=video_features,
+                            state_features=state_features,
+                            lengths=lengths,
+                            return_all_frames=True,
+                            head_mode="sparse",
+                        )
+                        sparse_val = float(
+                            sparse_progress[0, center_idx]
+                            if sparse_progress.ndim == 2
+                            else sparse_progress[center_idx]
+                        )
+
+                    # Compute dense prediction for center frame
+                    if compute_dense:
+                        dense_progress = reward_model.calculate_rewards(
+                            text_embeddings=text_features,
+                            video_embeddings=video_features,
+                            state_features=state_features,
+                            lengths=lengths,
+                            return_all_frames=True,
+                            head_mode="dense",
+                        )
+                        dense_val = float(
+                            dense_progress[0, center_idx]
+                            if dense_progress.ndim == 2
+                            else dense_progress[center_idx]
+                        )
+
+                    frame_results[query_idx] = (sparse_val, dense_val)
+
+            except Exception as e:
+                logging.warning(f"Failed to process frame {query_idx}: {e}")
+
+        # Interpolate to get values for all frames
+        computed_indices = np.array(sorted(frame_results.keys()))
+        computed_sparse = (
+            np.array([frame_results[i][0] for i in computed_indices]) if compute_sparse else None
+        )
+        computed_dense = np.array([frame_results[i][1] for i in computed_indices]) if compute_dense else None
+
+        # All frame indices for this episode
+        all_frame_idx_array = np.arange(ep_start, ep_end)
+
+        if stride > 1 and len(computed_indices) > 1:
+            # Interpolate progress values
+            if compute_sparse:
+                interp_sparse = interpolate_progress(computed_indices, computed_sparse, all_frame_idx_array)
+            if compute_dense:
+                interp_dense = interpolate_progress(computed_indices, computed_dense, all_frame_idx_array)
+        else:
+            # No interpolation needed
+            interp_sparse = computed_sparse if compute_sparse else None
+            interp_dense = computed_dense if compute_dense else None
+
+        # Store results for all frames
+        for i, frame_idx in enumerate(all_frame_idx_array):
+            local_idx = frame_idx - ep_start
+            all_indices.append(frame_idx)
+            all_episode_indices.append(episode_idx)
+            all_frame_indices.append(local_idx)
+            if compute_sparse:
+                if stride > 1 and len(computed_indices) > 1:
+                    all_progress_sparse.append(float(interp_sparse[i]))
+                elif frame_idx in frame_results:
+                    all_progress_sparse.append(frame_results[frame_idx][0])
+                else:
+                    all_progress_sparse.append(np.nan)
+            if compute_dense:
+                if stride > 1 and len(computed_indices) > 1:
+                    all_progress_dense.append(float(interp_dense[i]))
+                elif frame_idx in frame_results:
+                    all_progress_dense.append(frame_results[frame_idx][1])
+                else:
+                    all_progress_dense.append(np.nan)
+
+    # Create output table
+    table_data = {
+        "index": np.array(all_indices, dtype=np.int64),
+        "episode_index": np.array(all_episode_indices, dtype=np.int64),
+        "frame_index": np.array(all_frame_indices, dtype=np.int64),
+    }
+    if compute_sparse:
+        table_data["progress_sparse"] = np.array(all_progress_sparse, dtype=np.float32)
+    if compute_dense:
+        table_data["progress_dense"] = np.array(all_progress_dense, dtype=np.float32)
+
+    # Sort by index
+    df = pa.table(table_data).to_pandas()
+    df = df.sort_values("index").reset_index(drop=True)
+    final_table = pa.Table.from_pandas(df, preserve_index=False)
+
+    # Add metadata with reward model path
+    metadata = {b"reward_model_path": reward_model_path.encode()}
+    final_table = final_table.replace_schema_metadata(metadata)
+
+    # Determine output path
+    output_path = Path(dataset.root) / "sarm_progress.parquet" if output_path is None else Path(output_path)
+
+    # Save
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    pq.write_table(final_table, output_path)
+    logging.info(f"Saved {len(final_table)} frame progress values to {output_path}")
+
+    # Print statistics
+    if "progress_sparse" in df.columns:
+        valid = df["progress_sparse"].dropna()
+        logging.info(
+            f"Sparse progress: mean={valid.mean():.4f}, std={valid.std():.4f}, "
+            f"min={valid.min():.4f}, max={valid.max():.4f}"
+        )
+
+    if "progress_dense" in df.columns:
+        valid = df["progress_dense"].dropna()
+        logging.info(
+            f"Dense progress: mean={valid.mean():.4f}, std={valid.std():.4f}, "
+            f"min={valid.min():.4f}, max={valid.max():.4f}"
+        )
+
+    # Visualize episodes after processing
+    if num_visualizations > 0:
+        viz_episodes = list(range(min(num_visualizations, num_episodes)))
+        logging.info(f"Generating {len(viz_episodes)} visualizations...")
+        visualize_sarm_predictions(
+            dataset=dataset,
+            reward_model=reward_model,
+            preprocess=preprocess,
+            episode_indices=viz_episodes,
+            head_mode=head_mode,
+            output_dir=Path(output_dir),
+            stride=stride,
+        )
+
+    return output_path
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Compute SARM progress values for RA-BC weighting or visualize SARM predictions",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+Examples:
+    # Full RA-BC computation with visualizations
+    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
+        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
+        --reward-model-path pepijn223/sarm_single_uni4
+
+    # Visualize predictions only (no RA-BC computation)
+    python src/lerobot/policies/sarm/compute_rabc_weights.py \\
+        --dataset-repo-id lerobot/aloha_sim_insertion_human \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
+        --visualize-only \\
+        --num-visualizations 10
+        """,
+    )
+    parser.add_argument(
+        "--dataset-repo-id",
+        type=str,
+        required=True,
+        help="HuggingFace dataset repo ID or local path",
+    )
+    parser.add_argument(
+        "--reward-model-path",
+        type=str,
+        default=None,
+        help="Path to pretrained SARM model (reads from existing parquet metadata if not provided)",
+    )
+    parser.add_argument(
+        "--output-path",
+        type=str,
+        default=None,
+        help="Output path for parquet. If not set, saves to dataset's cache directory",
+    )
+    parser.add_argument(
+        "--head-mode",
+        type=str,
+        default="sparse",
+        choices=["sparse", "dense", "both"],
+        help="SARM head to use (default: sparse)",
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+        help="Device to use (default: cuda)",
+    )
+    # Visualization options
+    parser.add_argument(
+        "--visualize-only",
+        action="store_true",
+        help="Only visualize SARM predictions (no RA-BC computation)",
+    )
+    parser.add_argument(
+        "--num-visualizations",
+        type=int,
+        default=5,
+        help="Number of episodes to visualize (default: 5, set to 0 to skip)",
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default="./sarm_viz",
+        help="Output directory for visualizations (default: ./sarm_viz)",
+    )
+    parser.add_argument(
+        "--push-to-hub",
+        action="store_true",
+        help="Upload progress file to the dataset repo on HuggingFace Hub",
+        default=True,
+    )
+    parser.add_argument(
+        "--stride",
+        type=int,
+        default=1,
+        help="Compute progress every N frames, interpolate the rest (default: 1 = every frame)",
+    )
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
+
+    # Try to get reward_model_path from parquet metadata if not provided
+    reward_model_path = args.reward_model_path
+    if reward_model_path is None:
+        # Load dataset to find parquet path
+        temp_dataset = LeRobotDataset(args.dataset_repo_id, download_videos=False)
+        parquet_path = Path(temp_dataset.root) / "sarm_progress.parquet"
+        reward_model_path = get_reward_model_path_from_parquet(parquet_path)
+        if reward_model_path:
+            logging.info(f"Using reward model from parquet metadata: {reward_model_path}")
+        else:
+            raise ValueError(
+                "--reward-model-path is required (no existing parquet with model metadata found)"
+            )
+
+    # Handle visualize-only mode
+    if args.visualize_only:
+        dataset, reward_model, preprocess = load_sarm_resources(
+            args.dataset_repo_id, reward_model_path, args.device
+        )
+        logging.info(f"Visualization-only mode: visualizing {args.num_visualizations} episodes")
+        viz_episodes = list(range(min(args.num_visualizations, dataset.num_episodes)))
+        visualize_sarm_predictions(
+            dataset=dataset,
+            reward_model=reward_model,
+            preprocess=preprocess,
+            episode_indices=viz_episodes,
+            head_mode=args.head_mode,
+            output_dir=Path(args.output_dir),
+            stride=args.stride,
+        )
+        print(f"\nVisualizations saved to: {Path(args.output_dir).absolute()}")
+        return
+
+    # Full RABC computation (compute_sarm_progress loads model/dataset itself)
+    output_path = compute_sarm_progress(
+        dataset_repo_id=args.dataset_repo_id,
+        reward_model_path=reward_model_path,
+        output_path=args.output_path,
+        head_mode=args.head_mode,
+        device=args.device,
+        num_visualizations=args.num_visualizations,
+        output_dir=args.output_dir,
+        stride=args.stride,
+    )
+
+    print(f"\nSARM progress values saved to: {output_path}")
+
+    # Upload to Hub if requested
+    if args.push_to_hub:
+        from huggingface_hub import HfApi
+
+        api = HfApi()
+        hub_path = "sarm_progress.parquet"
+
+        print(f"\nUploading to Hub: {args.dataset_repo_id}/{hub_path}")
+        api.upload_file(
+            path_or_fileobj=str(output_path),
+            path_in_repo=hub_path,
+            repo_id=args.dataset_repo_id,
+            repo_type="dataset",
+        )
+        print(
+            f"Successfully uploaded to: https://huggingface.co/datasets/{args.dataset_repo_id}/blob/main/{hub_path}"
+        )
+
+        print("\nTo use in training, add to your config:")
+        print("  use_rabc: true")
+        print(f"  rabc_progress_path: hf://datasets/{args.dataset_repo_id}/{hub_path}")
+        print("  rabc_head_mode: sparse  # or dense")
+    else:
+        print("\nTo use in training, add to your config:")
+        print("  use_rabc: true")
+        print(f"  rabc_progress_path: {output_path}")
+        print("  rabc_head_mode: sparse  # or dense")
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/policies/sarm/configuration_sarm.py

@@ -0,0 +1,248 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Qianzhong Chen, Justin Yu, Mac Schwager, Pieter Abbeel, Yide Shentu, Philipp Wu
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation.
+Paper: https://arxiv.org/abs/2509.25358
+"""
+
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.optim.optimizers import AdamWConfig
+from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+
+
+@PreTrainedConfig.register_subclass("sarm")
+@dataclass
+class SARMConfig(PreTrainedConfig):
+    """Configuration class for SARM (Stage-Aware Reward Modeling).
+
+    Supports three annotation modes:
+
+    1. single_stage (default): No annotations needed. Uses the episode's task description
+       as a single stage covering the entire episode.
+
+    2. dense_only: Uses dense (fine-grained) annotations from VLM, with an auto-generated
+       single sparse "task" stage covering the full episode. The dense head learns detailed
+       subtask progression while sparse provides overall task completion.
+
+    3. dual: Full dual-head mode with both sparse (high-level) and dense (fine-grained)
+       annotations from VLM. Both heads are trained on their respective annotations.
+
+    The annotation_mode determines how sparse_temporal_proportions and dense_temporal_proportions
+    are loaded/generated during model initialization.
+    """
+
+    annotation_mode: str = "single_stage"  # "single_stage", "dense_only", or "dual"
+    n_obs_steps: int = 8  # Number of observation history steps
+    frame_gap: int = 30  # Frame gap between frames (at 30 fps = 1 second)
+    max_rewind_steps: int = 4  # Maximum rewind steps for temporal augmentation
+
+    # Total frames = 1 + n_obs_steps + max_rewind_steps (computed in property)
+    # During training with rewind: [obs_frames] + [rewind_frames]
+    # During inference: [obs_frames] only
+
+    # Architecture params
+    image_dim: int = 512
+    text_dim: int = 512
+    hidden_dim: int = 768
+    num_heads: int = 12
+    num_layers: int = 8
+    max_state_dim: int = 32
+    drop_n_last_frames: int = 1
+    batch_size: int = 64
+    clip_batch_size: int = 64
+    dropout: float = 0.1
+    stage_loss_weight: float = 1.0  # Weight for stage classification loss when using subtask annotations
+
+    rewind_probability: float = 0.8
+    language_perturbation_probability: float = 0.2
+
+    # Sparse annotations (high-level stages)
+    num_sparse_stages: int = 1
+    sparse_subtask_names: list | None = None
+    sparse_temporal_proportions: list | None = None
+
+    # Dense annotations (fine-grained stages)
+    num_dense_stages: int | None = None
+    dense_subtask_names: list | None = None
+    dense_temporal_proportions: list | None = None
+
+    pretrained_model_path: str | None = None
+    device: str | None = None
+    image_key: str = "observation.images.top"  # Key for image used from the dataset
+    state_key: str = "observation.state"
+
+    # Populated by the processor (video_features, state_features, text_features)
+    input_features: dict = field(default_factory=lambda: {})
+
+    # Output features (updated in __post_init__)
+    output_features: dict = field(
+        default_factory=lambda: {
+            "stage": PolicyFeature(shape=(9, 5), type=FeatureType.REWARD),
+            "progress": PolicyFeature(shape=(9, 1), type=FeatureType.REWARD),
+        }
+    )
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,
+            "LANGUAGE": NormalizationMode.IDENTITY,
+            "REWARD": NormalizationMode.IDENTITY,
+        }
+    )
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        if self.annotation_mode not in ["single_stage", "dense_only", "dual"]:
+            raise ValueError(
+                f"annotation_mode must be 'single_stage', 'dense_only', or 'dual', got {self.annotation_mode}"
+            )
+
+        if self.annotation_mode == "single_stage":
+            # Use task description as stage name, full episode as one stage
+            self.num_sparse_stages = 1
+            self.sparse_subtask_names = ["task"]
+            self.sparse_temporal_proportions = [1.0]
+            self.num_dense_stages = None
+            self.dense_subtask_names = None
+            self.dense_temporal_proportions = None
+
+        elif self.annotation_mode == "dense_only":
+            self.num_sparse_stages = 1
+            self.sparse_subtask_names = ["task"]
+            self.sparse_temporal_proportions = [1.0]
+
+        self.input_features = {}
+        self.output_features = {}
+
+        if self.image_key:
+            self.input_features[self.image_key] = PolicyFeature(shape=(480, 640, 3), type=FeatureType.VISUAL)
+
+        self.input_features[self.state_key] = PolicyFeature(
+            shape=(self.max_state_dim,),
+            type=FeatureType.STATE,
+        )
+
+        # Update output features based on annotation_mode
+        if self.annotation_mode in ["dense_only", "dual"]:
+            self.output_features["sparse_stage"] = PolicyFeature(
+                shape=(self.num_frames, self.num_sparse_stages), type=FeatureType.REWARD
+            )
+            self.output_features["sparse_progress"] = PolicyFeature(
+                shape=(self.num_frames, 1), type=FeatureType.REWARD
+            )
+            dense_stages = self.num_dense_stages or self.num_sparse_stages
+            self.output_features["dense_stage"] = PolicyFeature(
+                shape=(self.num_frames, dense_stages), type=FeatureType.REWARD
+            )
+            self.output_features["dense_progress"] = PolicyFeature(
+                shape=(self.num_frames, 1), type=FeatureType.REWARD
+            )
+        else:
+            self.output_features["sparse_stage"] = PolicyFeature(
+                shape=(self.num_frames, self.num_sparse_stages), type=FeatureType.REWARD
+            )
+            self.output_features["sparse_progress"] = PolicyFeature(
+                shape=(self.num_frames, 1), type=FeatureType.REWARD
+            )
+
+        if self.max_rewind_steps >= self.n_obs_steps:
+            raise ValueError(
+                f"max_rewind_steps ({self.max_rewind_steps}) must be less than n_obs_steps ({self.n_obs_steps})"
+            )
+        if self.num_sparse_stages < 1:
+            raise ValueError(f"num_sparse_stages must be at least 1, got {self.num_sparse_stages}")
+        if (
+            self.annotation_mode in ["dense_only", "dual"]
+            and self.num_dense_stages is not None
+            and self.num_dense_stages < 2
+        ):
+            raise ValueError(f"num_dense_stages must be at least 2, got {self.num_dense_stages}")
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        """Get default optimizer configuration for SARM training."""
+        return AdamWConfig(
+            lr=5e-5,
+            weight_decay=1e-3,
+            betas=(0.9, 0.999),
+            eps=1e-8,
+        )
+
+    def get_scheduler_preset(self) -> CosineDecayWithWarmupSchedulerConfig:
+        """Get default learning rate scheduler configuration."""
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=5e-5,
+            decay_lr=5e-6,
+            num_warmup_steps=500,
+            num_decay_steps=50000,
+        )
+
+    def validate_features(self) -> None:
+        pass
+
+    @property
+    def uses_dual_heads(self) -> bool:
+        """Whether the model uses dual heads (dense_only or dual annotation modes)."""
+        return self.annotation_mode in ["dense_only", "dual"]
+
+    @property
+    def num_frames(self) -> int:
+        """Total number of frames in sequence.
+
+        For training: 1 + n_obs_steps + max_rewind_steps
+        The sequence is: [obs_frames (n_obs_steps + 1)] + [rewind_frames (max_rewind_steps)]
+        """
+        return 1 + self.n_obs_steps + self.max_rewind_steps
+
+    @property
+    def max_length(self) -> int:
+        return self.num_frames
+
+    @property
+    def observation_delta_indices(self) -> list[int]:
+        """Bidirectional frame sampling centered on target frame.
+
+        Example with n_obs_steps=8, gap=30:
+        Before: [-120, -90, -60, -30]  (4 frames)
+        Current: [0]                   (1 frame)
+        After:  [30, 60, 90, 120]      (4 frames)
+        Total: 9 frames
+        """
+        half_steps = self.n_obs_steps // 2
+
+        past_deltas = [-self.frame_gap * i for i in range(half_steps, 0, -1)]
+        future_deltas = [self.frame_gap * i for i in range(1, half_steps + 1)]
+        obs_deltas = past_deltas + [0] + future_deltas
+
+        # Rewind placeholders
+        rewind_deltas = [-self.frame_gap * (i + 1) for i in range(self.max_rewind_steps)]
+
+        return obs_deltas + rewind_deltas
+
+    @property
+    def action_delta_indices(self) -> None:
+        """SARM is a reward model, not an action policy."""
+        return None
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

src/lerobot/policies/sarm/modeling_sarm.py

@@ -0,0 +1,793 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Qianzhong Chen, Justin Yu, Mac Schwager, Pieter Abbeel, Yide Shentu, Philipp Wu
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation.
+
+Paper: https://arxiv.org/abs/2509.25358
+
+- StageTransformer: Predicts stage classification (sparse/dense)
+- SubtaskTransformer: Predicts within-stage progress (tau) conditioned on stage
+"""
+
+import json
+import logging
+import random
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor
+
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.sarm.configuration_sarm import SARMConfig
+from lerobot.policies.sarm.sarm_utils import (
+    normalize_stage_tau,
+    pad_state_to_max_dim,
+)
+
+
+class StageTransformer(nn.Module):
+    """
+    Stage classification transformer for SARM.
+
+    Predicts which stage/subtask the current frame belongs to.
+    Supports both sparse (high-level) and dense (fine-grained) annotation schemes.
+
+    Input streams: [vis_proj, lang_proj, state_proj] concatenated -> (B, N+2, T, D)
+    Output: stage logits (B, T, num_classes)
+    """
+
+    def __init__(
+        self,
+        d_model: int = 512,
+        vis_emb_dim: int = 512,
+        text_emb_dim: int = 512,
+        state_dim: int = 32,
+        n_layers: int = 6,
+        n_heads: int = 8,
+        dropout: float = 0.1,
+        num_cameras: int = 1,
+        num_classes_sparse: int = 4,
+        num_classes_dense: int = 8,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.num_cameras = num_cameras
+
+        # Projections
+        self.lang_proj = nn.Linear(text_emb_dim, d_model)
+        self.visual_proj = nn.Linear(vis_emb_dim, d_model)
+        self.state_proj = nn.Linear(state_dim, d_model)
+
+        # Encoder
+        enc_layer = nn.TransformerEncoderLayer(d_model, n_heads, 4 * d_model, dropout, batch_first=True)
+        self.transformer = nn.TransformerEncoder(enc_layer, n_layers)
+
+        # Positional bias on first visual frame
+        self.first_pos = nn.Parameter(torch.zeros(1, d_model))
+
+        # Shared fusion MLP
+        # Fuses (num_cameras + 2) streams: cameras + lang + state
+        fused_in = d_model * (num_cameras + 2)
+        self.fusion_backbone = nn.Sequential(
+            nn.LayerNorm(fused_in),
+            nn.Linear(fused_in, d_model),
+            nn.ReLU(),
+        )
+
+        # Scheme-specific heads
+        self.heads = nn.ModuleDict(
+            {
+                "sparse": nn.Linear(d_model, num_classes_sparse),
+                "dense": nn.Linear(d_model, num_classes_dense),
+            }
+        )
+
+    def _prep_lang(self, lang_emb: torch.Tensor, B: int, T: int, D: int) -> torch.Tensor:  # noqa: N803
+        """
+        Prepare language embeddings for fusion.
+
+        Accepts lang_emb of shape:
+          - (B, text_emb_dim) -> broadcast across time
+          - (B, T, text_emb_dim) -> per-timestep (dense annotation mode)
+
+        Returns: (B, 1, T, D)
+        """
+        if lang_emb.dim() == 3:
+            # (B, T, E) -> (B, T, D) -> (B, 1, T, D)
+            lang_proj = self.lang_proj(lang_emb).unsqueeze(1)
+        else:
+            # (B, E) -> (B, 1, 1, D) -> expand to (B, 1, T, D)
+            lang_proj = self.lang_proj(lang_emb).unsqueeze(1).unsqueeze(2).expand(B, 1, T, D)
+        return lang_proj
+
+    def forward(
+        self,
+        img_seq: torch.Tensor,  # (B, N, T, vis_emb_dim)
+        lang_emb: torch.Tensor,  # (B, E) or (B, T, E)
+        state: torch.Tensor,  # (B, T, state_dim)
+        lengths: torch.Tensor,  # (B,) - valid sequence lengths
+        scheme: str = "sparse",  # "sparse" or "dense"
+    ) -> torch.Tensor:
+        """
+        Forward pass for stage classification.
+
+        Args:
+            img_seq: Image embeddings (B, N, T, vis_emb_dim) where N=num_cameras
+            lang_emb: Language embeddings (B, E) or (B, T, E) for dense
+            state: State features (B, T, state_dim)
+            lengths: Valid sequence lengths (B,) for masking
+            scheme: "sparse" or "dense" for head selection
+
+        Returns:
+            Stage logits (B, T, num_classes)
+        """
+        assert scheme in self.heads, f"Unknown scheme '{scheme}'. Use one of {list(self.heads.keys())}."
+
+        B, N, T, _ = img_seq.shape  # noqa: N806
+        D = self.d_model  # noqa: N806
+        device = img_seq.device
+
+        # Project inputs
+        vis_proj = self.visual_proj(img_seq)  # (B, N, T, D)
+        state_proj = self.state_proj(state).unsqueeze(1)  # (B, 1, T, D)
+        lang_proj = self._prep_lang(lang_emb, B, T, D)  # (B, 1, T, D)
+
+        # Concatenate streams
+        # cameras + lang + state -> (B, N+2, T, D)
+        x = torch.cat([vis_proj, lang_proj, state_proj], dim=1)
+
+        # Add positional bias to first visual frame
+        x[:, :N, 0, :] = x[:, :N, 0, :] + self.first_pos
+
+        # Flatten to tokens for Transformer
+        x_tokens = x.view(B, (N + 2) * T, D)
+        L = x_tokens.size(1)  # noqa: N806
+
+        # Create padding mask
+        base_mask = torch.arange(T, device=device).expand(B, T) >= lengths.unsqueeze(1)  # (B, T)
+        mask = base_mask.unsqueeze(1).expand(B, N + 2, T).reshape(B, (N + 2) * T)
+
+        # Create causal mask
+        causal_mask = torch.triu(torch.ones(L, L, device=device, dtype=torch.bool), diagonal=1)
+
+        # Encode
+        h = self.transformer(x_tokens, mask=causal_mask, src_key_padding_mask=mask, is_causal=True)
+
+        # Reshape and fuse
+        h = h.view(B, N + 2, T, D).permute(0, 2, 1, 3).reshape(B, T, (N + 2) * D)
+        fused = self.fusion_backbone(h)  # (B, T, D)
+
+        # Scheme-specific logits
+        logits = self.heads[scheme](fused)  # (B, T, num_classes)
+        return logits
+
+
+class SubtaskTransformer(nn.Module):
+    """
+    Subtask progress regression transformer for SARM.
+
+    Predicts within-stage normalized progress (tau) conditioned on stage prior.
+    The stage prior is a one-hot encoding passed from StageTransformer predictions.
+
+    Input streams: [vis_proj, lang_proj, state_proj, stage_emb] -> (B, N+3, T, D)
+    Output: tau predictions (B, T) in [0, 1]
+    """
+
+    def __init__(
+        self,
+        d_model: int = 512,
+        vis_emb_dim: int = 512,
+        text_emb_dim: int = 512,
+        state_dim: int = 32,
+        n_layers: int = 6,
+        n_heads: int = 8,
+        dropout: float = 0.1,
+        num_cameras: int = 1,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.num_cameras = num_cameras
+
+        # Projections
+        self.lang_proj = nn.Linear(text_emb_dim, d_model)
+        self.visual_proj = nn.Linear(vis_emb_dim, d_model)
+        self.state_proj = nn.Linear(state_dim, d_model)
+
+        # Encoder
+        enc = nn.TransformerEncoderLayer(d_model, n_heads, 4 * d_model, dropout, batch_first=True)
+        self.transformer = nn.TransformerEncoder(enc, n_layers)
+
+        # Learned bias on first visual frame
+        self.first_pos = nn.Parameter(torch.zeros(1, d_model))
+
+        # Shared fusion backbone
+        # Fuses (num_cameras + 3) streams: cameras + lang + state + stage_emb
+        fused_in = d_model * (num_cameras + 3)
+        self.fusion_backbone = nn.Sequential(
+            nn.LayerNorm(fused_in),
+            nn.Linear(fused_in, d_model),
+            nn.ReLU(),
+        )
+
+        # Scheme-specific regression heads
+        self.heads = nn.ModuleDict(
+            {
+                "sparse": nn.Linear(d_model, 1),
+                "dense": nn.Linear(d_model, 1),
+            }
+        )
+
+    def _prep_lang(self, lang_emb: torch.Tensor, B: int, T: int, D: int) -> torch.Tensor:  # noqa: N803
+        """
+        Prepare language embeddings for fusion.
+        """
+        if lang_emb.dim() == 3:
+            # (B, T, E) -> (B, T, D) -> (B, 1, T, D)
+            return self.lang_proj(lang_emb).unsqueeze(1)
+        else:
+            # (B, E) -> (B, 1, 1, D) -> (B, 1, T, D)
+            return self.lang_proj(lang_emb).unsqueeze(1).unsqueeze(2).expand(B, 1, T, D)
+
+    def _stage_to_dmodel(self, stage_prior: torch.Tensor) -> torch.Tensor:
+        """
+        Deterministic projection of one-hot stage to d_model by pad/truncate.
+
+        Args:
+            stage_prior: One-hot stage embedding (B, 1, T, C)
+
+        Returns:
+            Projected stage embedding (B, 1, T, d_model)
+        """
+        B, one, T, C = stage_prior.shape  # noqa: N806
+        D = self.d_model  # noqa: N806
+        if D == C:
+            return stage_prior
+        elif D > C:
+            pad = torch.zeros(B, one, T, D - C, device=stage_prior.device, dtype=stage_prior.dtype)
+            return torch.cat([stage_prior, pad], dim=-1)
+        else:
+            return stage_prior[..., :D]
+
+    def forward(
+        self,
+        img_seq: torch.Tensor,  # (B, N, T, vis_emb_dim)
+        lang_emb: torch.Tensor,  # (B, E) or (B, T, E)
+        state: torch.Tensor,  # (B, T, state_dim)
+        lengths: torch.Tensor,  # (B,) - valid sequence lengths
+        stage_prior: torch.Tensor,  # (B, 1, T, C) one-hot from gen_stage_emb
+        scheme: str = "sparse",  # "sparse" or "dense"
+    ) -> torch.Tensor:
+        """
+        Forward pass for subtask progress regression.
+
+        Args:
+            img_seq: Image embeddings (B, N, T, vis_emb_dim)
+            lang_emb: Language embeddings (B, E) or (B, T, E)
+            state: State features (B, T, state_dim)
+            lengths: Valid sequence lengths (B,) for masking
+            stage_prior: One-hot stage prior (B, 1, T, num_classes)
+            scheme: "sparse" or "dense" for head selection
+
+        Returns:
+            Tau predictions (B, T) in [0, 1] via sigmoid
+        """
+        assert scheme in self.heads, f"Unknown scheme '{scheme}'. Use one of {list(self.heads.keys())}."
+
+        B, N, T, _ = img_seq.shape  # noqa: N806
+        D = self.d_model  # noqa: N806
+        device = img_seq.device
+
+        # Project inputs
+        vis_proj = self.visual_proj(img_seq)  # (B, N, T, D)
+        state_proj = self.state_proj(state).unsqueeze(1)  # (B, 1, T, D)
+        lang_proj = self._prep_lang(lang_emb, B, T, D)  # (B, 1, T, D)
+        stage_emb = self._stage_to_dmodel(stage_prior)  # (B, 1, T, D)
+
+        # Concatenate all streams
+        # cameras + lang + state + stage_emb -> (B, N+3, T, D)
+        x = torch.cat([vis_proj, lang_proj, state_proj, stage_emb], dim=1)
+
+        # Add positional bias to first visual frame
+        x[:, :N, 0, :] = x[:, :N, 0, :] + self.first_pos
+
+        # Flatten to tokens
+        x_tokens = x.view(B, (N + 3) * T, D)
+        L = x_tokens.size(1)  # noqa: N806
+
+        # Create padding mask
+        base_mask = torch.arange(T, device=device).expand(B, T) >= lengths.unsqueeze(1)
+        mask = base_mask.unsqueeze(1).expand(B, N + 3, T).reshape(B, (N + 3) * T)
+
+        # Create causal mask
+        causal_mask = torch.triu(torch.ones(L, L, device=device, dtype=torch.bool), diagonal=1)
+
+        # Encode
+        h = self.transformer(x_tokens, mask=causal_mask, src_key_padding_mask=mask, is_causal=True)
+
+        # Reshape and fuse
+        h = h.view(B, N + 3, T, D)
+        h_flat = h.permute(0, 2, 1, 3).reshape(B, T, (N + 3) * D)
+        fused = self.fusion_backbone(h_flat)  # (B, T, D)
+
+        # Scheme-specific regression head -> sigmoid
+        r = torch.sigmoid(self.heads[scheme](fused)).squeeze(-1)  # (B, T)
+        return r
+
+
+def gen_stage_emb(num_classes: int, targets: torch.Tensor) -> torch.Tensor:
+    """
+    Generate one-hot stage embeddings from targets.
+
+    Args:
+        num_classes: Number of stage classes
+        targets: Target values (B, T) where integer part is stage index
+
+    Returns:
+        One-hot stage embedding (B, 1, T, num_classes)
+    """
+    # Integer part of float targets -> [0, C-1]
+    idx = targets.long().clamp(min=0, max=num_classes - 1)  # (B, T)
+    C = num_classes  # noqa: N806
+    # Identity-lookup one-hot
+    stage_onehot = torch.eye(C, device=targets.device)[idx]  # (B, T, C)
+    stage_onehot = stage_onehot.unsqueeze(1)  # (B, 1, T, C)
+    return stage_onehot
+
+
+class SARMRewardModel(PreTrainedPolicy):
+    """
+    SARM Reward Model for stage-aware task completion rewards.
+
+    Uses two separate transformer models:
+    - StageTransformer: Classifies which stage/subtask
+    - SubtaskTransformer: Predicts within-stage progress (tau)
+
+    Training uses 75%/25% GT/predicted stage conditioning (teacher forcing).
+    """
+
+    name = "sarm"
+    config_class = SARMConfig
+
+    def __init__(self, config: SARMConfig, dataset_stats: dict | None = None, dataset_meta=None):
+        super().__init__(config, dataset_stats)
+        config.validate_features()
+        self.config = config
+        self.dataset_stats = dataset_stats
+        self.device = torch.device(
+            config.device if config.device else "cuda" if torch.cuda.is_available() else "cpu"
+        )
+
+        # Load temporal proportions based on annotation_mode
+        if config.annotation_mode == "single_stage":
+            logging.info(f"Using single_stage mode: sparse_subtask_names={config.sparse_subtask_names}")
+        elif dataset_meta is not None:
+            self._load_temporal_proportions(dataset_meta)
+
+        # Create two separate models
+        self.stage_model = StageTransformer(
+            d_model=config.hidden_dim,
+            vis_emb_dim=config.image_dim,
+            text_emb_dim=config.text_dim,
+            state_dim=config.max_state_dim,
+            n_layers=config.num_layers,
+            n_heads=config.num_heads,
+            dropout=config.dropout,
+            num_cameras=1,  # Single camera for now
+            num_classes_sparse=config.num_sparse_stages,
+            num_classes_dense=config.num_dense_stages or config.num_sparse_stages,
+        )
+
+        self.subtask_model = SubtaskTransformer(
+            d_model=config.hidden_dim,
+            vis_emb_dim=config.image_dim,
+            text_emb_dim=config.text_dim,
+            state_dim=config.max_state_dim,
+            n_layers=config.num_layers,
+            n_heads=config.num_heads,
+            dropout=config.dropout,
+            num_cameras=1,
+        )
+
+        self.stage_model.to(self.device)
+        self.subtask_model.to(self.device)
+
+        # GT/predicted stage ratio for teacher forcing
+        self.gt_stage_ratio = 0.75
+
+        if config.uses_dual_heads:
+            logging.info(
+                f"SARM initialized with dual heads: {config.num_sparse_stages} sparse stages, "
+                f"{config.num_dense_stages} dense stages"
+            )
+        else:
+            logging.info(f"SARM initialized with sparse head only: {config.num_sparse_stages} stages")
+
+        logging.info(f"SARM initialized on {self.device}")
+
+    def _load_proportions_from_json(self, path, annotation_type: str) -> tuple[list[str], list[float]]:
+        """Load temporal proportions from a JSON file (preserving order)."""
+        if not path.exists():
+            raise ValueError(
+                f"{annotation_type.capitalize()} temporal proportions not found at {path}. "
+                f"Run the subtask annotation tool with --{annotation_type}-subtasks to generate annotations."
+            )
+        with open(path) as f:
+            proportions_dict = json.load(f)
+        names = list(proportions_dict.keys())
+        logging.info(f"Loaded {len(names)} {annotation_type} subtasks: {names}")
+        logging.info(f"{annotation_type.capitalize()} temporal proportions: {proportions_dict}")
+        return names, [proportions_dict[name] for name in names]
+
+    def _load_temporal_proportions(self, dataset_meta) -> None:
+        """Load temporal proportions based on annotation_mode."""
+        meta_path = dataset_meta.root / "meta"
+
+        if self.config.annotation_mode == "dual":
+            names, props = self._load_proportions_from_json(
+                meta_path / "temporal_proportions_sparse.json", "sparse"
+            )
+            (
+                self.config.num_sparse_stages,
+                self.config.sparse_subtask_names,
+                self.config.sparse_temporal_proportions,
+            ) = len(names), names, props
+
+        if self.config.annotation_mode in ["dense_only", "dual"]:
+            names, props = self._load_proportions_from_json(
+                meta_path / "temporal_proportions_dense.json", "dense"
+            )
+            (
+                self.config.num_dense_stages,
+                self.config.dense_subtask_names,
+                self.config.dense_temporal_proportions,
+            ) = len(names), names, props
+            if self.config.annotation_mode == "dense_only":
+                logging.info(f"Using auto-generated sparse 'task' stage: {self.config.sparse_subtask_names}")
+
+    def to(self, device):
+        """Override to method to ensure all components move together."""
+        super().to(device)
+        self.device = device if isinstance(device, torch.device) else torch.device(device)
+        self.stage_model.to(device)
+        self.subtask_model.to(device)
+        return self
+
+    @torch.no_grad()
+    def calculate_rewards(
+        self,
+        text_embeddings: np.ndarray | torch.Tensor,
+        video_embeddings: np.ndarray | torch.Tensor,
+        state_features: np.ndarray | torch.Tensor | None = None,
+        lengths: np.ndarray | torch.Tensor | None = None,
+        return_all_frames: bool = False,
+        return_stages: bool = False,
+        return_confidence: bool = False,
+        head_mode: str | None = "sparse",
+        frame_index: int | None = None,
+    ) -> np.ndarray | tuple:
+        """
+        Calculate rewards for given text, video, and state representations.
+
+        This is the canonical method for SARM reward computation, used for:
+        - Inference/visualization
+        - RA-BC weight computation
+
+        Args:
+            text_embeddings: Encoded text representations (batch_size, 512)
+            video_embeddings: Encoded video representations (batch_size, num_frames, 512)
+            state_features: Joint state features (batch_size, num_frames, state_dim)
+            lengths: Valid sequence lengths (batch_size,)
+            return_all_frames: If True, return rewards for all frames
+            return_stages: If True, also return stage predictions
+            return_confidence: If True, also return stage confidence
+            head_mode: Which head to use ("sparse" or "dense")
+            frame_index: Index of the target frame to extract (default: n_obs_steps).
+
+        Returns:
+            Rewards and optionally stage probs/confidence.
+        """
+        if isinstance(text_embeddings, np.ndarray):
+            text_embeddings = torch.tensor(text_embeddings, dtype=torch.float32)
+        if isinstance(video_embeddings, np.ndarray):
+            video_embeddings = torch.tensor(video_embeddings, dtype=torch.float32)
+        if state_features is not None and isinstance(state_features, np.ndarray):
+            state_features = torch.tensor(state_features, dtype=torch.float32)
+
+        # Handle single sample case
+        if text_embeddings.dim() == 1:
+            text_embeddings = text_embeddings.unsqueeze(0)
+            video_embeddings = video_embeddings.unsqueeze(0)
+            if state_features is not None:
+                state_features = state_features.unsqueeze(0)
+            single_sample = True
+        else:
+            single_sample = False
+
+        batch_size = video_embeddings.shape[0]
+        seq_len = video_embeddings.shape[1]
+
+        scheme = head_mode
+
+        # Default lengths if not provided
+        if lengths is None:
+            lengths = torch.full((batch_size,), seq_len, dtype=torch.int32)
+        elif isinstance(lengths, np.ndarray):
+            lengths = torch.tensor(lengths, dtype=torch.int32)
+
+        # Reshape video to (B, N, T, D) for multi-camera format
+        # Currently single camera: (B, T, D) -> (B, 1, T, D)
+        img_seq = video_embeddings.unsqueeze(1).to(self.device)
+        lang_emb = text_embeddings.to(self.device)
+        state = (
+            state_features.to(self.device)
+            if state_features is not None
+            else torch.zeros(batch_size, seq_len, self.config.max_state_dim, device=self.device)
+        )
+        lens = lengths.to(self.device)
+
+        # Pad state to max_state_dim
+        state = pad_state_to_max_dim(state, self.config.max_state_dim)
+
+        # Get num_classes for this scheme
+        num_classes = self.config.num_sparse_stages if scheme == "sparse" else self.config.num_dense_stages
+
+        # Run stage model
+        stage_logits = self.stage_model(img_seq, lang_emb, state, lens, scheme=scheme)
+        stage_probs = F.softmax(stage_logits, dim=-1)  # (B, T, num_classes)
+        stage_idx = stage_probs.argmax(dim=-1)  # (B, T)
+        stage_conf = stage_probs.gather(-1, stage_idx.unsqueeze(-1)).squeeze(-1)  # (B, T)
+
+        # Create one-hot stage prior
+        stage_onehot = F.one_hot(stage_idx, num_classes=num_classes).float()  # (B, T, C)
+        stage_emb = stage_onehot.unsqueeze(1)  # (B, 1, T, C)
+
+        # Run subtask model
+        tau_pred = self.subtask_model(img_seq, lang_emb, state, lens, stage_emb, scheme=scheme)
+
+        # Compute final reward: stage + tau
+        raw_reward = stage_idx.float() + tau_pred  # (B, T)
+
+        # Normalize to [0, 1] using temporal proportions for proper weighting
+        if scheme == "sparse":
+            normalized_reward = normalize_stage_tau(
+                raw_reward,
+                num_stages=num_classes,
+                temporal_proportions=self.config.sparse_temporal_proportions,
+                subtask_names=self.config.sparse_subtask_names,
+            )
+        else:
+            normalized_reward = normalize_stage_tau(
+                raw_reward,
+                num_stages=num_classes,
+                temporal_proportions=self.config.dense_temporal_proportions,
+                subtask_names=self.config.dense_subtask_names,
+            )
+
+        # Default frame index is n_obs_steps (last observation frame)
+        if frame_index is None:
+            frame_index = self.config.n_obs_steps
+
+        # Prepare outputs (batch mode or no smoothing)
+        if return_all_frames:
+            rewards = normalized_reward.cpu().numpy()
+        else:
+            rewards = normalized_reward[:, frame_index].cpu().numpy()
+
+        if single_sample:
+            rewards = rewards[0] if not return_all_frames else rewards[0]
+
+        outputs = [rewards]
+        if return_stages:
+            probs = stage_probs.cpu().numpy()
+            if single_sample:
+                probs = probs[0]
+            outputs.append(probs)
+        if return_confidence:
+            conf = stage_conf.cpu().numpy()
+            if single_sample:
+                conf = conf[0]
+            outputs.append(conf)
+
+        return outputs[0] if len(outputs) == 1 else tuple(outputs)
+
+    def train(self, mode: bool = True):
+        """Set training mode for both models."""
+        super().train(mode)
+        self.stage_model.train(mode)
+        self.subtask_model.train(mode)
+        return self
+
+    def eval(self):
+        """Set evaluation mode for both models."""
+        return self.train(False)
+
+    def parameters(self):
+        """Override to return trainable parameters from both models."""
+        from itertools import chain
+
+        return chain(self.stage_model.parameters(), self.subtask_model.parameters())
+
+    def get_optim_params(self):
+        """Override to return optimizer parameters from both models."""
+        return self.parameters()
+
+    def reset(self):
+        """Required by PreTrainedPolicy but not used for reward models."""
+        pass
+
+    def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
+        """Required by PreTrainedPolicy but not used for reward models."""
+        raise NotImplementedError("SARM model does not predict action chunks")
+
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Required by PreTrainedPolicy but not used for SARM."""
+        raise NotImplementedError("SARM model does not select actions")
+
+    def _train_step(
+        self,
+        img_emb: torch.Tensor,  # (B, N, T, D)
+        lang_emb: torch.Tensor,  # (B, E) or (B, T, E)
+        state: torch.Tensor,  # (B, T, state_dim)
+        lengths: torch.Tensor,  # (B,)
+        targets: torch.Tensor,  # (B, T) - format: stage.tau
+        scheme: str,
+    ) -> dict[str, torch.Tensor]:
+        """
+        Single training step for one annotation scheme.
+
+        Implements 75%/25% GT/predicted stage conditioning.
+
+        Args:
+            img_emb: Image embeddings (B, N, T, D)
+            lang_emb: Language embeddings
+            state: State features
+            lengths: Valid sequence lengths
+            targets: Target values where floor=stage, remainder=tau
+            scheme: "sparse" or "dense"
+
+        Returns:
+            Dict with stage_loss, subtask_loss, total_loss
+        """
+        num_classes = self.config.num_sparse_stages if scheme == "sparse" else self.config.num_dense_stages
+
+        # Ground truth: stage (integer) and tau (fractional)
+        # Clamp stage indices to valid range [0, num_classes-1] to handle edge cases
+        # where targets may exceed expected range (e.g., frames between subtasks)
+        gt_stage = torch.floor(targets).long().clamp(0, num_classes - 1)  # (B, T)
+        gt_tau = torch.remainder(targets, 1.0)  # (B, T)
+
+        # Run stage model
+        stage_pred = self.stage_model(img_emb, lang_emb, state, lengths, scheme=scheme)
+
+        # 75%/25% GT/predicted stage conditioning
+        if random.random() < self.gt_stage_ratio:
+            # Mode 1: Use ground truth stage -> one-hot
+            stage_emb = gen_stage_emb(num_classes, targets)  # (B, 1, T, C)
+        else:
+            # Mode 2: Use predicted stage argmax -> one-hot
+            stage_idx = stage_pred.argmax(dim=-1)  # (B, T)
+            stage_onehot = F.one_hot(stage_idx, num_classes=num_classes).float()  # (B, T, C)
+            stage_emb = stage_onehot.unsqueeze(1)  # (B, 1, T, C)
+
+        # Run subtask model with stage prior
+        tau_pred = self.subtask_model(img_emb, lang_emb, state, lengths, stage_emb, scheme=scheme)
+
+        # Compute losses
+        stage_loss = F.cross_entropy(stage_pred.view(-1, num_classes), gt_stage.view(-1), reduction="mean")
+        subtask_loss = F.mse_loss(tau_pred, gt_tau, reduction="mean")
+
+        return {
+            "stage_loss": stage_loss,
+            "subtask_loss": subtask_loss,
+            "total_loss": stage_loss + subtask_loss,
+        }
+
+    def forward(self, batch):
+        """
+        Forward pass for SARM reward model training.
+
+        Uses stage+tau target format where:
+        - Integer part = stage index
+        - Fractional part = within-stage progress (tau)
+
+        Training uses 75%/25% GT/predicted stage conditioning.
+
+        Args:
+            batch: Dictionary with 'observation' containing:
+                - 'video_features': (B, T, 512) pre-encoded video features
+                - 'text_features': (B, 512) or (B, T, 512) text features
+                - 'state_features': (B, T, state_dim) joint state features
+                - 'lengths': (B,) valid sequence lengths
+                - 'sparse_targets': (B, T) sparse targets (stage.tau format)
+                - 'dense_targets': (B, T) dense targets (optional, for dual mode)
+
+        Returns:
+            Tuple of (total_loss, output_dict with loss components)
+        """
+        observation = batch.get("observation", batch)
+
+        # Extract features
+        video_features = observation["video_features"].to(self.device)
+        text_features = observation["text_features"].to(self.device)
+        state_features = observation.get("state_features")
+        if state_features is not None:
+            state_features = state_features.to(self.device)
+
+        batch_size = video_features.shape[0]
+        seq_len = video_features.shape[1]
+
+        # Get lengths (default to full sequence)
+        lengths = observation.get("lengths")
+        if lengths is None:
+            lengths = torch.full((batch_size,), seq_len, dtype=torch.int32, device=self.device)
+        else:
+            lengths = lengths.to(self.device)
+
+        # Reshape video to (B, N, T, D) - single camera
+        img_emb = video_features.unsqueeze(1)
+
+        # Pad state to max_state_dim
+        if state_features is None:
+            state_features = torch.zeros(batch_size, seq_len, self.config.max_state_dim, device=self.device)
+        else:
+            state_features = pad_state_to_max_dim(state_features, self.config.max_state_dim)
+
+        output_dict = {}
+        total_loss = torch.tensor(0.0, device=self.device)
+
+        # Sparse training (always)
+        sparse_targets = observation.get("sparse_targets")
+        if sparse_targets is None:
+            # Try legacy format
+            sparse_targets = observation.get("targets")
+        if sparse_targets is None:
+            raise ValueError("sparse_targets (or targets) is required for SARM training")
+        sparse_targets = sparse_targets.to(self.device)
+
+        sparse_result = self._train_step(
+            img_emb, text_features, state_features, lengths, sparse_targets, scheme="sparse"
+        )
+        output_dict["sparse_stage_loss"] = sparse_result["stage_loss"].item()
+        output_dict["sparse_subtask_loss"] = sparse_result["subtask_loss"].item()
+        total_loss = total_loss + sparse_result["total_loss"]
+
+        # Dense training (if dual mode)
+        if self.config.uses_dual_heads:
+            dense_targets = observation.get("dense_targets")
+            if dense_targets is not None:
+                dense_targets = dense_targets.to(self.device)
+                dense_result = self._train_step(
+                    img_emb, text_features, state_features, lengths, dense_targets, scheme="dense"
+                )
+                output_dict["dense_stage_loss"] = dense_result["stage_loss"].item()
+                output_dict["dense_subtask_loss"] = dense_result["subtask_loss"].item()
+                total_loss = total_loss + dense_result["total_loss"]
+
+        output_dict["total_loss"] = total_loss.item()
+        return total_loss, output_dict
+
+
+def compute_stage_loss(stage_logits: torch.Tensor, target_stages: torch.Tensor) -> torch.Tensor:
+    """Compute cross-entropy loss for stage classification."""
+    _, _, num_stages = stage_logits.shape
+    stage_logits_flat = stage_logits.reshape(-1, num_stages)
+    # Clamp target stage indices to valid range [0, num_stages-1]
+    target_stages_flat = target_stages.reshape(-1).clamp(0, num_stages - 1)
+    return F.cross_entropy(stage_logits_flat, target_stages_flat)

src/lerobot/policies/sarm/processor_sarm.py

@@ -0,0 +1,518 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""SARM Processor for encoding images/text and generating stage+tau targets."""
+
+import random
+from typing import Any
+
+import numpy as np
+import pandas as pd
+import torch
+from faker import Faker
+from PIL import Image
+from transformers import CLIPModel, CLIPProcessor
+
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.policies.sarm.configuration_sarm import SARMConfig
+from lerobot.policies.sarm.sarm_utils import (
+    apply_rewind_augmentation,
+    compute_absolute_indices,
+    find_stage_and_tau,
+    pad_state_to_max_dim,
+)
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    RenameObservationsProcessorStep,
+)
+from lerobot.processor.converters import (
+    from_tensor_to_numpy,
+    policy_action_to_transition,
+    transition_to_policy_action,
+)
+from lerobot.processor.core import EnvTransition, TransitionKey
+from lerobot.processor.pipeline import PipelineFeatureType
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+
+class SARMEncodingProcessorStep(ProcessorStep):
+    """ProcessorStep that encodes images and text with CLIP and generates stage and progress labels for SARM."""
+
+    def __init__(
+        self,
+        config: SARMConfig,
+        image_key: str | None = None,
+        dataset_meta=None,
+        dataset_stats: dict | None = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.image_key = image_key or config.image_key
+        self.dataset_meta = dataset_meta
+        self.dataset_stats = dataset_stats
+        self.annotation_mode = config.annotation_mode
+
+        # Helper to create temporal proportions dict
+        def make_props_dict(names, props):
+            return dict(zip(names, props, strict=True)) if names and props else None
+
+        # Sparse annotations (always needed)
+        self.sparse_temporal_proportions = make_props_dict(
+            config.sparse_subtask_names, config.sparse_temporal_proportions
+        )
+        self.sparse_subtask_names = config.sparse_subtask_names
+
+        # Dense annotations (only for dual mode)
+        self.dense_subtask_names = config.dense_subtask_names if config.uses_dual_heads else None
+        self.dense_temporal_proportions = (
+            make_props_dict(config.dense_subtask_names, config.dense_temporal_proportions)
+            if config.uses_dual_heads
+            else None
+        )
+
+        self.device = torch.device(
+            self.config.device if self.config.device else "cuda" if torch.cuda.is_available() else "cpu"
+        )
+
+        self.clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        self.clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32", use_fast=True)
+        self.clip_model.to(self.device)
+        self.clip_model.eval()
+
+        self.verbs = ["move", "grasp", "rotate", "push", "pull", "slide", "lift", "place"]
+        self.fake = Faker()
+
+    def _find_episode_for_frame(self, frame_idx: int) -> int:
+        """Find the episode index for a given frame index."""
+        for ep_idx in range(len(self.dataset_meta.episodes)):
+            ep_start = self.dataset_meta.episodes[ep_idx]["dataset_from_index"]
+            ep_end = self.dataset_meta.episodes[ep_idx]["dataset_to_index"]
+            if ep_start <= frame_idx < ep_end:
+                return ep_idx
+        return 0
+
+    def _get_episode_indices(self, frame_indices: np.ndarray, episode_index) -> np.ndarray:
+        """Get episode indices for each frame index."""
+        if episode_index is None:
+            return np.array([self._find_episode_for_frame(int(f)) for f in frame_indices])
+
+        episode_indices = np.atleast_1d(np.asarray(from_tensor_to_numpy(episode_index)))
+
+        # If single episode but multiple frames, compute episode for each frame
+        if len(episode_indices) == 1 and len(frame_indices) > 1:
+            return np.array([self._find_episode_for_frame(int(f)) for f in frame_indices])
+
+        return episode_indices
+
+    def _generate_perturbed_task(self) -> str:
+        """Generate a random perturbed task string for language perturbation."""
+        num_words = random.randint(1, 5)
+        verb = random.choice(self.verbs)
+        phrase = " ".join([verb] + self.fake.words(nb=num_words))
+        return phrase
+
+    def _get_annotation_config(self, annotation_type: str) -> tuple[list[str], dict[str, float] | None]:
+        """Get global subtask names and temporal proportions for an annotation type."""
+        if annotation_type == "dense":
+            return self.dense_subtask_names, self.dense_temporal_proportions
+        return self.sparse_subtask_names, self.sparse_temporal_proportions
+
+    def _load_episode_annotations(
+        self,
+        ep_idx: int,
+        episodes_df: pd.DataFrame | None,
+        annotation_type: str,
+        global_names: list[str],
+    ) -> tuple[list | None, list | None, list | None]:
+        """Load subtask annotations for an episode from DataFrame."""
+        # Single-stage mode: (linear progress 0→1)
+        if episodes_df is None or len(global_names) == 1:
+            return None, None, None
+
+        # Resolve column name with fallback
+        def col(suffix):
+            prefixed = f"{annotation_type}_{suffix}"
+            return prefixed if prefixed in episodes_df.columns else suffix
+
+        col_names = col("subtask_names")
+        if col_names not in episodes_df.columns or ep_idx >= len(episodes_df):
+            return None, None, None
+
+        subtask_names = episodes_df.loc[ep_idx, col_names]
+        if subtask_names is None or (isinstance(subtask_names, float) and pd.isna(subtask_names)):
+            return None, None, None
+
+        return (
+            subtask_names,
+            episodes_df.loc[ep_idx, col("subtask_start_frames")],
+            episodes_df.loc[ep_idx, col("subtask_end_frames")],
+        )
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """
+        Encode images, text, and normalize states in the transition.
+
+        Implements SARM training data preparation:
+        - Applies language perturbation (20% probability)
+        - Applies rewind augmentation (80% probability)
+        - Generates stage+tau targets for all frames
+        - Outputs lengths tensor for valid sequence masking
+        """
+        new_transition = transition.copy() if hasattr(transition, "copy") else dict(transition)
+        observation = new_transition.get(TransitionKey.OBSERVATION)
+        comp_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+
+        frame_index = comp_data.get("index")
+        episode_index = comp_data.get("episode_index")
+
+        if frame_index is None:
+            raise ValueError("Frame index ('index') not found in COMPLEMENTARY_DATA")
+        if episode_index is None:
+            raise ValueError("Episode index ('episode_index') not found in COMPLEMENTARY_DATA")
+
+        frame_indices = np.atleast_1d(np.asarray(from_tensor_to_numpy(frame_index)))
+        episode_indices = self._get_episode_indices(frame_indices, episode_index)
+
+        image = observation.get(self.image_key)
+        if isinstance(image, torch.Tensor):
+            image = image.cpu().numpy()
+
+        # If 4D (T, C, H, W) from delta_timestamps, add batch dim
+        # If 3D (C, H, W) single frame, add batch and time dims
+        if image.ndim == 4:
+            image = image[np.newaxis, ...]  # (T, C, H, W) -> (1, T, C, H, W)
+        elif image.ndim == 3:
+            image = image[np.newaxis, np.newaxis, ...]  # (C, H, W) -> (1, 1, C, H, W)
+
+        batch_size = image.shape[0]
+        total_frames = image.shape[1]  # Should be 13: 9 obs + 4 rewind placeholders
+        n_obs_steps = self.config.n_obs_steps
+        max_rewind_steps = self.config.max_rewind_steps
+        n_obs_frames = 1 + n_obs_steps  # 9 observation frames (including current)
+
+        # Rewind augmentation
+        rewind_steps = torch.zeros(batch_size, dtype=torch.int32)
+        apply_rewind = self.training and random.random() < self.config.rewind_probability
+
+        if apply_rewind and self.dataset_meta is not None:
+            for b_idx, (ep_idx, frame_idx) in enumerate(
+                zip(episode_indices.tolist(), frame_indices.tolist(), strict=True)
+            ):
+                ep_idx, frame_idx = int(ep_idx), int(frame_idx)
+                ep_start = self.dataset_meta.episodes[ep_idx]["dataset_from_index"]
+
+                rewind_step, _ = apply_rewind_augmentation(
+                    frame_idx, ep_start, n_obs_steps, max_rewind_steps, frame_gap=self.config.frame_gap
+                )
+                rewind_steps[b_idx] = rewind_step
+
+        # Compute valid lengths: n_obs_frames + rewind_steps
+        lengths = n_obs_frames + rewind_steps  # (B,)
+
+        # Apply rewind masking to images
+        # For frames beyond valid length, we mask with zeros (or copy last valid frame)
+        for b_idx in range(batch_size):
+            valid_len = lengths[b_idx].item()
+            if valid_len < total_frames:
+                image[b_idx, valid_len:] = 0  # Zero out frames beyond valid length
+
+        # Encode images with CLIP
+        video_features = self._encode_images_batch(image)
+        observation["video_features"] = video_features
+
+        state_key = self.config.state_key
+        state_data = observation.get(state_key)
+
+        if isinstance(state_data, torch.Tensor):
+            state_tensor = state_data.float()
+        else:
+            state_tensor = torch.tensor(state_data, dtype=torch.float32)
+
+        if state_tensor.ndim == 2:
+            state_tensor = state_tensor.unsqueeze(0)  # (T, D) -> (1, T, D)
+        elif state_tensor.ndim == 1:
+            state_tensor = state_tensor.unsqueeze(0).unsqueeze(0)  # (D,) -> (1, 1, D)
+
+        # Apply same rewind masking to state
+        for b_idx in range(batch_size):
+            valid_len = lengths[b_idx].item()
+            if valid_len < state_tensor.shape[1]:
+                state_tensor[b_idx, valid_len:] = 0  # Zero out frames beyond valid length
+
+        observation["state_features"] = pad_state_to_max_dim(state_tensor, self.config.max_state_dim)
+
+        task = comp_data.get("task")
+        if isinstance(task, list):
+            task = task[0] if task else ""
+
+        # Apply language perturbation during training (20% probability)
+        # When perturbed, targets will be zeroed to train model to output low values for irrelevant text
+        apply_perturbation = self.training and random.random() < self.config.language_perturbation_probability
+        if apply_perturbation:
+            task = self._generate_perturbed_task()
+
+        # Encode text with CLIP
+        observation["text_features"] = self._encode_text_clip(task, batch_size)
+
+        # Store lengths for model
+        observation["lengths"] = lengths
+
+        # When language is perturbed, targets are zero so perturbed samples don't contribute to progress loss
+        if self.dataset_meta is not None:
+            episodes_df = None
+            if self.sparse_subtask_names != ["task"]:
+                episodes_df = self.dataset_meta.episodes.to_pandas()
+
+            # Generate sparse targets
+            if self.sparse_temporal_proportions is not None:
+                if apply_perturbation:
+                    # Zero targets when language is perturbed
+                    sparse_targets = torch.zeros(batch_size, total_frames, dtype=torch.float32)
+                else:
+                    sparse_targets = self._compute_batch_targets(
+                        frame_indices, episode_indices, lengths, rewind_steps, episodes_df, "sparse"
+                    )
+                observation["sparse_targets"] = sparse_targets
+
+            # Generate dense targets (for dual mode)
+            if self.config.uses_dual_heads and self.dense_temporal_proportions is not None:
+                if apply_perturbation:
+                    # Zero targets when language is perturbed
+                    dense_targets = torch.zeros(batch_size, total_frames, dtype=torch.float32)
+                else:
+                    dense_targets = self._compute_batch_targets(
+                        frame_indices, episode_indices, lengths, rewind_steps, episodes_df, "dense"
+                    )
+                observation["dense_targets"] = dense_targets
+
+        new_transition[TransitionKey.OBSERVATION] = observation
+        return new_transition
+
+    def _compute_batch_targets(
+        self,
+        frame_indices: np.ndarray,
+        episode_indices: np.ndarray,
+        lengths: torch.Tensor,
+        rewind_steps: torch.Tensor,
+        episodes_df: pd.DataFrame | None,
+        annotation_type: str,
+    ) -> torch.Tensor:
+        """Compute stage+tau targets for a batch of samples."""
+        batch_size = len(frame_indices)
+        n_obs_steps = self.config.n_obs_steps
+        max_rewind_steps = self.config.max_rewind_steps
+        total_frames = 1 + n_obs_steps + max_rewind_steps
+        frame_gap = self.config.frame_gap
+
+        global_names, temporal_props = self._get_annotation_config(annotation_type)
+        targets = torch.zeros(batch_size, total_frames, dtype=torch.float32)
+
+        for b_idx in range(batch_size):
+            ep_idx = int(episode_indices[b_idx])
+            frame_idx = int(frame_indices[b_idx])
+
+            ep_start = self.dataset_meta.episodes[ep_idx]["dataset_from_index"]
+            ep_end = self.dataset_meta.episodes[ep_idx]["dataset_to_index"]
+            ep_length = ep_end - ep_start
+
+            subtask_names, subtask_start_frames, subtask_end_frames = self._load_episode_annotations(
+                ep_idx, episodes_df, annotation_type, global_names
+            )
+
+            # Compute observation frame indices
+            obs_indices, _ = compute_absolute_indices(
+                frame_idx, ep_start, ep_end, n_obs_steps, frame_gap=frame_gap
+            )
+            obs_indices = obs_indices.tolist()
+
+            # Compute targets for observation frames
+            for t_idx, abs_idx in enumerate(obs_indices):
+                rel_frame = abs_idx - ep_start
+                targets[b_idx, t_idx] = find_stage_and_tau(
+                    rel_frame,
+                    ep_length,
+                    subtask_names,
+                    subtask_start_frames,
+                    subtask_end_frames,
+                    global_names,
+                    temporal_props,
+                    return_combined=True,
+                )
+
+            # Compute targets for rewind frames (if any)
+            rewind_step = rewind_steps[b_idx].item()
+            if rewind_step > 0:
+                _, rewind_indices = apply_rewind_augmentation(
+                    frame_idx,
+                    ep_start,
+                    n_obs_steps,
+                    max_rewind_steps,
+                    frame_gap=frame_gap,
+                    rewind_step=rewind_step,
+                )
+
+                for r_idx, abs_idx in enumerate(rewind_indices[:rewind_step]):
+                    rel_frame = max(0, abs_idx - ep_start)
+                    targets[b_idx, n_obs_steps + 1 + r_idx] = find_stage_and_tau(
+                        rel_frame,
+                        ep_length,
+                        subtask_names,
+                        subtask_start_frames,
+                        subtask_end_frames,
+                        global_names,
+                        temporal_props,
+                        return_combined=True,
+                    )
+
+        return targets
+
+    @property
+    def training(self) -> bool:
+        return getattr(self, "_training_mode", True)
+
+    def train(self, mode: bool = True):
+        """Set training mode for augmentation decisions."""
+        self._training_mode = mode
+        return self
+
+    def eval(self):
+        """Set evaluation mode (disable augmentations)."""
+        return self.train(False)
+
+    @torch.no_grad()
+    def _encode_images_batch(self, images: np.ndarray) -> torch.Tensor:
+        """Encode a batch of images using CLIP.
+
+        Args:
+            images: Batched images with shape: (B, T, C, H, W)
+
+        Returns:
+            Encoded feature vectors with shape (B, T, 512)
+        """
+
+        batch_size, seq_length = images.shape[0], images.shape[1]
+        images = images.reshape(batch_size * seq_length, *images.shape[2:])
+
+        num_frames = images.shape[0]
+        images_list = []
+        for i in range(num_frames):
+            img = images[i]
+            if img.shape[0] in [1, 3]:  # Channel first (C, H, W)
+                img = img.transpose(1, 2, 0)
+
+            # Handle single channel
+            if img.shape[-1] == 1:
+                img = np.repeat(img, 3, axis=-1)
+
+            if img.dtype != np.uint8:
+                img = (img * 255).astype(np.uint8) if img.max() <= 1.0 else img.astype(np.uint8)
+
+            images_list.append(Image.fromarray(img))
+
+        all_embeddings = []
+        for i in range(0, num_frames, self.config.clip_batch_size):
+            batch_imgs = images_list[i : i + self.config.clip_batch_size]
+
+            inputs = self.clip_processor(images=batch_imgs, return_tensors="pt")
+            inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+            # Get image embeddings
+            embeddings = self.clip_model.get_image_features(**inputs).detach().cpu()
+
+            # Handle single frame case
+            if embeddings.dim() == 1:
+                embeddings = embeddings.unsqueeze(0)
+
+            all_embeddings.append(embeddings)
+
+        all_embeddings = torch.cat(all_embeddings)  # (B*T, 512)
+        all_embeddings = all_embeddings.reshape(batch_size, seq_length, -1)  # (B, T, 512)
+
+        return all_embeddings
+
+    @torch.no_grad()
+    def _encode_text_clip(self, text: str, batch_size: int) -> torch.Tensor:
+        """Encode text using CLIP text encoder (per SARM paper A.4).
+
+        Args:
+            text: Task description text to encode
+            batch_size: Batch size to replicate for
+
+        Returns:
+            Encoded text features with shape (B, 512)
+        """
+        inputs = self.clip_processor.tokenizer([text], return_tensors="pt", padding=True, truncation=True)
+        inputs = {k: v.to(self.device) for k, v in inputs.items()}
+
+        text_embedding = self.clip_model.get_text_features(**inputs).detach().cpu()
+        text_embedding = text_embedding.expand(batch_size, -1)
+
+        return text_embedding
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """Add encoded features to the observation features."""
+        features[PipelineFeatureType.OBSERVATION]["video_features"] = PolicyFeature(
+            type=FeatureType.VISUAL, shape=(self.config.num_frames, self.config.image_dim)
+        )
+        features[PipelineFeatureType.OBSERVATION]["text_features"] = PolicyFeature(
+            type=FeatureType.LANGUAGE, shape=(self.config.text_dim,)
+        )
+        features[PipelineFeatureType.OBSERVATION]["state_features"] = PolicyFeature(
+            type=FeatureType.STATE, shape=(self.config.num_frames, self.config.max_state_dim)
+        )
+        return features
+
+
+def make_sarm_pre_post_processors(
+    config: SARMConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    dataset_meta=None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Create pre-processor and post-processor pipelines for SARM."""
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=[
+                AddBatchDimensionProcessorStep(),
+                RenameObservationsProcessorStep(rename_map={}),
+                NormalizerProcessorStep(
+                    features={**config.input_features, **config.output_features},
+                    norm_map=config.normalization_mapping,
+                    stats=dataset_stats,
+                ),
+                SARMEncodingProcessorStep(
+                    config=config, dataset_meta=dataset_meta, dataset_stats=dataset_stats
+                ),
+                DeviceProcessorStep(device=config.device),
+            ],
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=[DeviceProcessorStep(device="cpu")],
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/sarm/sarm_utils.py

@@ -0,0 +1,295 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import random
+
+import numpy as np
+import torch
+import torch.nn.functional as F  # noqa: N812
+
+
+def find_stage_and_tau(
+    current_frame: int,
+    episode_length: int,
+    subtask_names: list | None,
+    subtask_start_frames: list | None,
+    subtask_end_frames: list | None,
+    global_subtask_names: list,
+    temporal_proportions: dict,
+    return_combined: bool = False,
+) -> tuple[int, float] | float:
+    """Find stage and within-stage progress (tau) for a frame.
+
+    Args:
+        current_frame: Frame index relative to episode start
+        episode_length: Total frames in episode
+        subtask_names: Subtask names for this episode (None for single_stage)
+        subtask_start_frames: Subtask start frames
+        subtask_end_frames: Subtask end frames
+        global_subtask_names: Global list of all subtask names
+        temporal_proportions: Dict of temporal proportions
+        return_combined: If True, return stage+tau as float; else (stage_idx, tau) tuple
+
+    Returns:
+        Float (stage.tau) if return_combined, else (stage_idx, tau) tuple
+    """
+    stage_idx, tau = 0, 0.0
+    num_stages = len(global_subtask_names)
+
+    # Single-stage mode: linear progress from 0 to 1
+    if num_stages == 1:
+        tau = min(1.0, max(0.0, current_frame / max(episode_length - 1, 1)))
+    elif subtask_names is None:
+        pass  # stage_idx=0, tau=0.0
+    elif current_frame < subtask_start_frames[0]:
+        pass  # Before first subtask: stage_idx=0, tau=0.0
+    elif current_frame > subtask_end_frames[-1]:
+        stage_idx, tau = num_stages - 1, 0.999  # After last subtask
+    else:
+        # Find which subtask this frame belongs to
+        found = False
+        for name, start, end in zip(subtask_names, subtask_start_frames, subtask_end_frames, strict=True):
+            if start <= current_frame <= end:
+                stage_idx = global_subtask_names.index(name) if name in global_subtask_names else 0
+                tau = compute_tau(current_frame, start, end)
+                found = True
+                break
+        # Frame between subtasks - use previous subtask's end state
+        if not found:
+            for j in range(len(subtask_names) - 1):
+                if subtask_end_frames[j] < current_frame < subtask_start_frames[j + 1]:
+                    name = subtask_names[j]
+                    stage_idx = global_subtask_names.index(name) if name in global_subtask_names else j
+                    tau = 1.0
+                    break
+
+    if return_combined:
+        # Clamp to avoid overflow at end
+        if stage_idx >= num_stages - 1 and tau >= 1.0:
+            return num_stages - 1 + 0.999
+        return stage_idx + tau
+    return stage_idx, tau
+
+
+def compute_absolute_indices(
+    frame_idx: int,
+    ep_start: int,
+    ep_end: int,
+    n_obs_steps: int,
+    frame_gap: int = 30,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Compute absolute frame indices with clamping for bidirectional observation sequence.
+
+    Bidirectional sampling centered on target frame:
+    - Before: [-frame_gap * half_steps, ..., -frame_gap] (half_steps frames)
+    - Current: [0] (1 frame)
+    - After: [frame_gap, ..., frame_gap * half_steps] (half_steps frames)
+    - Total: n_obs_steps + 1 frames
+
+    Out-of-bounds frames are clamped (duplicated from boundary).
+
+    Args:
+        frame_idx: Target frame index (center frame of sequence)
+        ep_start: Episode start index
+        ep_end: Episode end index (exclusive)
+        n_obs_steps: Number of observation steps (must be even for symmetric sampling)
+        frame_gap: Gap between observation frames
+
+    Returns:
+        Tuple of (indices, out_of_bounds_flags)
+    """
+    half_steps = n_obs_steps // 2
+
+    # Bidirectional deltas: past + current + future
+    past_deltas = [-frame_gap * i for i in range(half_steps, 0, -1)]
+    future_deltas = [frame_gap * i for i in range(1, half_steps + 1)]
+    delta_indices = past_deltas + [0] + future_deltas
+
+    frames = []
+    out_of_bounds = []
+
+    for delta in delta_indices:
+        target_idx = frame_idx + delta
+        # Clamp to episode bounds (duplicate boundary frames for out-of-bounds)
+        clamped_idx = max(ep_start, min(ep_end - 1, target_idx))
+        frames.append(clamped_idx)
+        # Flag as out-of-bounds if clamping occurred
+        out_of_bounds.append(1 if target_idx != clamped_idx else 0)
+
+    return torch.tensor(frames), torch.tensor(out_of_bounds)
+
+
+def apply_rewind_augmentation(
+    frame_idx: int,
+    ep_start: int,
+    n_obs_steps: int,
+    max_rewind_steps: int,
+    frame_gap: int = 30,
+    rewind_step: int | None = None,
+) -> tuple[int, list[int]]:
+    """
+    Generate rewind frame indices for temporal augmentation.
+
+    Rewind simulates going backwards through previously seen frames,
+    starting from before the earliest observation frame (for bidirectional sampling).
+    Appends reversed frames after the observation sequence.
+
+    Args:
+        frame_idx: Target frame index (center of bidirectional observation window)
+        ep_start: Episode start index
+        n_obs_steps: Number of observation steps
+        max_rewind_steps: Maximum rewind steps
+        frame_gap: Gap between frames
+        rewind_step: If provided, use this exact rewind step (for deterministic behavior).
+                     If None, sample randomly.
+
+    Returns:
+        Tuple of (rewind_step, rewind_indices)
+    """
+    # For bidirectional sampling, earliest obs frame is at frame_idx - half_steps * frame_gap
+    half_steps = n_obs_steps // 2
+    earliest_obs_frame = frame_idx - half_steps * frame_gap
+
+    # Required history: frames before earliest observation frame
+    if earliest_obs_frame <= ep_start:
+        return 0, []  # No history before observation window
+
+    # Max valid rewind steps based on available history before earliest obs frame
+    available_history = earliest_obs_frame - ep_start
+    max_valid_step = available_history // frame_gap
+    max_rewind = min(max_rewind_steps, max(0, max_valid_step))
+
+    if max_rewind <= 0:
+        return 0, []
+
+    # Sample rewind steps if not provided
+    rewind_step = random.randint(1, max_rewind) if rewind_step is None else min(rewind_step, max_rewind)
+
+    if rewind_step == 0:
+        return 0, []
+
+    # Generate rewind indices going backwards from earliest obs frame
+    # rewind_indices[0] is closest to obs window, rewind_indices[-1] is furthest back
+    rewind_indices = []
+    for i in range(1, rewind_step + 1):
+        idx = earliest_obs_frame - i * frame_gap
+        idx = max(ep_start, idx)  # Clamp to episode start
+        rewind_indices.append(idx)
+
+    return rewind_step, rewind_indices
+
+
+def compute_tau(current_frame: int | float, subtask_start: int | float, subtask_end: int | float) -> float:
+    """Compute τ_t = (t - s_k) / (e_k - s_k) ∈ [0, 1]. Returns 1.0 for zero-duration subtasks."""
+    duration = subtask_end - subtask_start
+    if duration <= 0:
+        return 1.0
+    return float(np.clip((current_frame - subtask_start) / duration, 0.0, 1.0))
+
+
+def pad_state_to_max_dim(state: torch.Tensor, max_state_dim: int) -> torch.Tensor:
+    """Pad the state tensor's last dimension to max_state_dim with zeros."""
+    current_dim = state.shape[-1]
+    if current_dim >= max_state_dim:
+        return state[..., :max_state_dim]  # Truncate if larger
+
+    # Pad with zeros on the right
+    padding = (0, max_state_dim - current_dim)  # (left, right) for last dim
+    return F.pad(state, padding, mode="constant", value=0)
+
+
+def temporal_proportions_to_breakpoints(
+    temporal_proportions: dict[str, float] | list[float] | None,
+    subtask_names: list[str] | None = None,
+) -> list[float] | None:
+    """Convert temporal proportions to cumulative breakpoints for normalization."""
+    if temporal_proportions is None:
+        return None
+
+    if isinstance(temporal_proportions, dict):
+        if subtask_names is not None:
+            proportions = [temporal_proportions.get(name, 0.0) for name in subtask_names]
+        else:
+            proportions = list(temporal_proportions.values())
+    else:
+        proportions = list(temporal_proportions)
+
+    total = sum(proportions)
+    if total > 0 and abs(total - 1.0) > 1e-6:
+        proportions = [p / total for p in proportions]
+
+    breakpoints = [0.0]
+    cumsum = 0.0
+    for prop in proportions:
+        cumsum += prop
+        breakpoints.append(cumsum)
+    breakpoints[-1] = 1.0
+
+    return breakpoints
+
+
+def normalize_stage_tau(
+    x: float | torch.Tensor,
+    num_stages: int | None = None,
+    breakpoints: list[float] | None = None,
+    temporal_proportions: dict[str, float] | list[float] | None = None,
+    subtask_names: list[str] | None = None,
+) -> float | torch.Tensor:
+    """
+    Normalize stage+tau reward to [0, 1] with custom breakpoints.
+
+    Maps stage index + within-stage tau to normalized progress [0, 1].
+    The breakpoints are designed to give appropriate weight to each stage
+    based on their importance in the task (using temporal proportions).
+
+    Priority: breakpoints > temporal_proportions > linear fallback
+
+    Args:
+        x: Raw reward value (stage index + tau) where stage ∈ [0, num_stages-1] and tau ∈ [0, 1)
+        num_stages: Number of stages (required if breakpoints/proportions not provided)
+        breakpoints: Optional custom breakpoints list of length num_stages + 1.
+        temporal_proportions: Optional temporal proportions dict/list to compute breakpoints.
+        subtask_names: Optional ordered list of subtask names (for dict proportions)
+
+    Returns:
+        Normalized progress value ∈ [0, 1]
+    """
+    if breakpoints is not None:
+        num_stages = len(breakpoints) - 1
+    elif temporal_proportions is not None:
+        breakpoints = temporal_proportions_to_breakpoints(temporal_proportions, subtask_names)
+        num_stages = len(breakpoints) - 1
+    elif num_stages is not None:
+        breakpoints = [i / num_stages for i in range(num_stages + 1)]
+    else:
+        raise ValueError("Either num_stages, breakpoints, or temporal_proportions must be provided")
+
+    if isinstance(x, torch.Tensor):
+        result = torch.zeros_like(x)
+        for i in range(num_stages):
+            mask = (x >= i) & (x < i + 1)
+            tau_in_stage = x - i
+            result[mask] = breakpoints[i] + tau_in_stage[mask] * (breakpoints[i + 1] - breakpoints[i])
+        result[x >= num_stages] = 1.0
+        return result.clamp(0.0, 1.0)
+    else:
+        if x < 0:
+            return 0.0
+        if x >= num_stages:
+            return 1.0
+        stage = int(x)
+        tau = x - stage
+        return breakpoints[stage] + tau * (breakpoints[stage + 1] - breakpoints[stage])

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -231,6 +231,7 @@ class SmolVLAPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: SmolVLAConfig,
+        **kwargs,
     ):
         """
         Args:
@@ -352,8 +353,19 @@ class SmolVLAPolicy(PreTrainedPolicy):
     def _rtc_enabled(self) -> bool:
         return self.config.rtc_config is not None and self.config.rtc_config.enabled
 
-    def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> dict[str, Tensor]:
-        """Do a full training forward pass to compute the loss"""
+    def forward(
+        self, batch: dict[str, Tensor], noise=None, time=None, reduction: str = "mean"
+    ) -> dict[str, Tensor]:
+        """Do a full training forward pass to compute the loss.
+
+        Args:
+            batch: Training batch containing observations and actions.
+            noise: Optional noise tensor for flow matching.
+            time: Optional time tensor for flow matching.
+            reduction: How to reduce the loss. Options:
+                - "mean": Return scalar mean loss (default, backward compatible)
+                - "none": Return per-sample losses of shape (batch_size,) for RA-BC weighting
+        """
         if self.config.adapt_to_pi_aloha:
             batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
             batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
@@ -377,11 +389,16 @@ class SmolVLAPolicy(PreTrainedPolicy):
         losses = losses[:, :, : self.config.max_action_dim]
         loss_dict["losses_after_rm_padding"] = losses.clone()
 
-        # For backward pass
-        loss = losses.mean()
-        # For backward pass
-        loss_dict["loss"] = loss.item()
-        return loss, loss_dict
+        if reduction == "none":
+            # Return per-sample losses (B,) by averaging over time and action dims
+            per_sample_loss = losses.mean(dim=(1, 2))
+            loss_dict["loss"] = per_sample_loss.mean().item()
+            return per_sample_loss, loss_dict
+        else:
+            # Default: return scalar mean loss
+            loss = losses.mean()
+            loss_dict["loss"] = loss.item()
+            return loss, loss_dict
 
     def prepare_images(self, batch):
         """Apply SmolVLA preprocessing to the images, like resizing to 224x224 and padding to keep aspect ratio, and
@@ -527,6 +544,7 @@ class VLAFlowMatching(nn.Module):
             num_vlm_layers=self.config.num_vlm_layers,
             self_attn_every_n_layers=self.config.self_attn_every_n_layers,
             expert_width_multiplier=self.config.expert_width_multiplier,
+            device=self.config.device if self.config.device is not None else "auto",
         )
         self.state_proj = nn.Linear(
             self.config.max_state_dim, self.vlm_with_expert.config.text_config.hidden_size
@@ -783,18 +801,15 @@ class VLAFlowMatching(nn.Module):
             use_cache=self.config.use_cache,
             fill_kv_cache=True,
         )
-        dt = -1.0 / self.config.num_steps
-        dt = torch.tensor(dt, dtype=torch.float32, device=device)
+        num_steps = self.config.num_steps
+        dt = -1.0 / num_steps
 
         x_t = noise
-        time = torch.tensor(1.0, dtype=torch.float32, device=device)
+        for step in range(num_steps):
+            time = 1.0 + step * dt
+            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
 
-        while time >= -dt / 2:
-            expanded_time = time.expand(bsize)
-
-            # Define a closure function to properly capture expanded_time
-            # This avoids the lambda expression (E731) and loop variable binding (B023) issues
-            def denoise_step_partial_call(input_x_t, current_timestep=expanded_time):
+            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
                 return self.denoise_step(
                     x_t=input_x_t,
                     prefix_pad_masks=prefix_pad_masks,
@@ -818,15 +833,11 @@ class VLAFlowMatching(nn.Module):
             else:
                 v_t = denoise_step_partial_call(x_t)
 
-            # Euler step
-            x_t += dt * v_t
+            x_t = x_t + dt * v_t
 
-            # Record x_t and v_t after Euler step (other params are recorded in rtc_processor.denoise_step)
             if self.rtc_processor is not None and self.rtc_processor.is_debug_enabled():
                 self.rtc_processor.track(time=time, x_t=x_t, v_t=v_t)
 
-            time += dt
-
         return x_t
 
     def denoise_step(

src/lerobot/policies/tdmpc/modeling_tdmpc.py

@@ -65,6 +65,7 @@ class TDMPCPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: TDMPCConfig,
+        **kwargs,
     ):
         """
         Args:

src/lerobot/policies/utils.py

@@ -231,11 +231,20 @@ def validate_visual_features_consistency(
     """
     Validates visual feature consistency between a policy config and provided dataset/environment features.
 
+    Validation passes if EITHER:
+    - Policy's expected visuals are a subset of dataset (policy uses some cameras, dataset has more)
+    - Dataset's provided visuals are a subset of policy (policy declares extras for flexibility)
+
     Args:
         cfg (PreTrainedConfig): The model or policy configuration containing input_features and type.
         features (Dict[str, PolicyFeature]): A mapping of feature names to PolicyFeature objects.
     """
     expected_visuals = {k for k, v in cfg.input_features.items() if v.type == FeatureType.VISUAL}
     provided_visuals = {k for k, v in features.items() if v.type == FeatureType.VISUAL}
-    if not provided_visuals.issubset(expected_visuals):
+
+    # Accept if either direction is a subset
+    policy_subset_of_dataset = expected_visuals.issubset(provided_visuals)
+    dataset_subset_of_policy = provided_visuals.issubset(expected_visuals)
+
+    if not (policy_subset_of_dataset or dataset_subset_of_policy):
         raise_feature_mismatch_error(provided_visuals, expected_visuals)

src/lerobot/policies/vqbet/modeling_vqbet.py

@@ -47,6 +47,7 @@ class VQBeTPolicy(PreTrainedPolicy):
     def __init__(
         self,
         config: VQBeTConfig | None = None,
+        **kwargs,
     ):
         """
         Args: