Merge pr-2593 (wallx support) into openarms_tmp_rebase

lerobot dataset fix
add evaluate_with_rtc script
2026-05-13 15:49:53 +00:00 · 2025-12-17 16:52:11 +01:00 · 2025-12-17 16:46:43 +01:00 · 2025-12-17 16:46:43 +01:00 · 2025-12-17 16:46:43 +01:00 · 2025-12-17 16:46:43 +01:00
124 changed files with 13539 additions and 8138 deletions
@@ -12,83 +12,57 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-name: "🚀 Issue / Bug / Request"
-description: Report a bug, suggest an improvement, or ask a technical question.
+name: "\U0001F41B Bug Report"
+description: Submit a bug report to help us improve LeRobot
 body:
  - type: markdown
    attributes:
      value: |
-        ### 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
+        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).

  - type: textarea
    id: system-info
    attributes:
-      label: Environment & System Info
-      description: |
-        For bugs or technical questions, please run `lerobot-info` and paste the output.
-        (Optional for feature requests).
+      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.
      render: Shell
-      placeholder: lerobot version, OS, python version, etc.
+      placeholder: lerobot version, OS, python version, numpy version, torch version, and lerobot's configuration
+    validations:
+      required: true
+
+  - type: checkboxes
+    id: information-scripts-examples
+    attributes:
+      label: Information
+      description: 'The problem arises when using:'
+      options:
+        - label: "One of the scripts in the examples/ folder of LeRobot"
+        - label: "My own task or dataset (give details below)"

  - type: textarea
-    id: description
+    id: reproduction
    validations:
      required: true
    attributes:
-      label: Description
+      label: Reproduction
      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?
+        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: |
-        A clear and concise description of the issue or suggestion.
+        Steps to reproduce the behavior:
+
+          1.
+          2.
+          3.

  - type: textarea
-    id: context-repro
+    id: expected-behavior
+    validations:
+      required: true
    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: extras
-    attributes:
-      label: Checklist
-      options:
-        - 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: workaround
-    attributes:
-      label: Additional Info / Workarounds
-      description: Anything else we should know? If you have a workaround, please share it!
+      label: Expected behavior
+      description: "A clear and concise description of what you would expect to happen."
@@ -1,54 +1,41 @@
-## Title
+## What this does

-Short, imperative summary (e.g., "fix(robots): handle None in sensor parser"). See [CONTRIBUTING.md](../CONTRIBUTING.md) for PR conventions.
+Explain what this PR does. Feel free to tag your PR with the appropriate label(s).

-## Type / Scope
+Examples:
+| Title | Label |
+|----------------------|-----------------|
+| Fixes #[issue] | (🐛 Bug) |
+| Adds new dataset | (🗃️ Dataset) |
+| Optimizes something | (⚡️ Performance) |

- **Type**: (Bug | Feature | Docs | Performance | Test | CI | Chore)
- **Scope**: (optional — name of module or package affected)
+## How it was tested

-## Summary / Motivation
+Explain/show how you tested your changes.

- One-paragraph description of what changes and why.
- Why this change is needed and any trade-offs or design notes.
+Examples:

-## Related issues
+- 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.

- Fixes / Closes: # (if any)
- Related: # (if any)
+## How to checkout & try? (for the reviewer)

-## What changed
+Provide a simple way for the reviewer to try out your changes.

- Short, concrete bullets of the modifications (files/behaviour).
- Short note if this introduces breaking changes and migration steps.
+Examples:

-## How was this tested
+```bash
+pytest -sx tests/test_stuff.py::test_something
+```

- Tests added: list new tests or test files.
- Manual checks / dataset runs performed.
+```bash
+lerobot-train --some.option=true
+```

-## How to run locally (reviewer)
+## SECTION TO REMOVE BEFORE SUBMITTING YOUR PR

- Run the relevant tests:
+**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.

-  ```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.
+**Note**: Before submitting this PR, please read the [contributor guideline](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md#submitting-a-pull-request-pr).
@@ -1,69 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-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/**'
@@ -85,7 +85,7 @@ jobs:
          python-version: ${{ env.PYTHON_VERSION }}

      - name: Install lerobot with all extras
-        run: uv sync --all-extras --no-extra groot --no-extra wallx # TODO(Steven): Make flash-attn optional
+        run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional

      - name: Run pytest (all extras)
        run: uv run pytest tests -vv --maxfail=10
@@ -1,89 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# 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
-
-            // Domain Specific
-            if (matches(/\b(bug|error|issue|fault|crash|exception)\b/i)) labelsToAdd.add('bug');
-            if (matches(/\b(feature|enhancement|improvement|support|implement|proposal)\b/i)) labelsToAdd.add('enhancement');
-            if (matches(/\b(question|help|how to||clarify|explain|unclear)\b/i)) labelsToAdd.add('question');
-            if (matches(/\b(maintenance|documentation|docs|readme|tutorial|guide|wiki)\b/i)) labelsToAdd.add('documentation');
-            if (matches(/\b(example|script|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|loss|optimizer|backward|gradient|wandb|sac)\b/i)) labelsToAdd.add('training');
-            if (matches(/\b(rerun|plot|video|render|visualiz|gif)/i)) labelsToAdd.add('visualization');
-            if (matches(/\b(camera|realsense|lidar|depth|sensor|imu|microphone|rgbd)\b/i)) labelsToAdd.add('sensors');
-            if (matches(/\b(aloha|koch|so-100|so100|mobile|teleop|manipulator|robots?)\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|p0licy)\b/i)) labelsToAdd.add('policies');
-            if (matches(/\b(processors?|pipeline)\b/i)) labelsToAdd.add('processor');
-            if (matches(/\b(eval|evaluate|evaluation|metrics?|score|benchmark)\b/i)) labelsToAdd.add('evaluation');
-
-            // Infrastructure & Code Quality
-            if (matches(/\b(tests?|pytest|unittest|failing test)\b/i)) labelsToAdd.add('tests');
-            if (matches(/\b(ci|github actions|workflow|gha|actions?|pipeline)\b/i)) {
-              labelsToAdd.add('CI');
-              labelsToAdd.add('github_actions');
-            }
-            if (matches(/\b(perf|latency|throughput|fps|speed|performance)\b/i)) labelsToAdd.add('performance');
-            if (matches(/\b(dependency|requirements|pip|conda|install error|importerror|package not found)\b/i)) labelsToAdd.add('dependencies');
-            if (matches(/\b(python|pyproject|requirements(\.txt)?|pip install|typing error)\b/i)) labelsToAdd.add('python');
-
-            // Documentation & Meta
-            if (matches(/\b(doc|documentation|docs|readme|typo|how to)\b/i)) labelsToAdd.add('documentation');
-            if (matches(/\b(refactor|cleanup|restructure|rename|modernize code)\b/i)) labelsToAdd.add('refactor');
-            if (matches(/\b(release|changelog|version bump|cut a release|tag v)\b/i)) labelsToAdd.add('release');
-            if (matches(/\b(breaking change|major change)\b/i)) labelsToAdd.add('breaking change');
-
-            // 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,
-              });
-            }
@@ -1,39 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# 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
@@ -78,7 +78,7 @@ jobs:
          echo "Dependencies unbound:" && cat pyproject.toml

      - name: Install lerobot with all extras
-        run: uv sync --all-extras --no-extra groot --no-extra wallx # TODO(Steven): Make flash-attn optional
+        run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional

      - name: Run pytest (all extras)
        run: uv run pytest tests -vv
@@ -87,7 +87,7 @@ repos:
  # TODO(Steven): Uncomment when ready to use
  ##### Static Analysis & Typing #####
  - repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.19.1
+    rev: v1.18.2
    hooks:
      - id: mypy
        args: [--config-file=pyproject.toml]
@@ -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 e-mail address,
+Examples of representing our community include using an official email 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.
+[feedback@huggingface.co](mailto: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
@@ -1,5 +1,7 @@
 <p align="center">
-  <img alt="LeRobot, Hugging Face Robotics Library" src="./media/readme/lerobot-logo-thumbnail.png" width="100%">
+  <img alt="LeRobot, Hugging Face Robotics Library" src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/lerobot-logo-thumbnail.png" width="100%">
+  <br/>
+  <br/>
 </p>

 <div align="center">
@@ -10,130 +12,323 @@
 [![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>

-**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.
+<h2 align="center">
+    <p><a href="https://huggingface.co/docs/lerobot/hope_jr">
+        Build Your Own HopeJR Robot!</a></p>
+</h2>

-🤗 A hardware-agnostic, Python-native interface that standardizes control across diverse platforms, from low-cost arms (SO-100) to humanoids.
+<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 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><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>

-🤗 State-of-the-art policies that have been shown to transfer to the real-world ready for training and deployment.
+  <p><a href="https://huggingface.co/docs/lerobot/hope_jr">
+      See the full HopeJR tutorial here.</a></p>
+</div>

-🤗 Comprehensive support for the open-source ecosystem to democratize physical AI.
+<br/>

-## Quick Start
+<h2 align="center">
+    <p><a href="https://huggingface.co/docs/lerobot/so101">
+        Build Your Own SO-101 Robot!</a></p>
+</h2>

-LeRobot can be installed directly from PyPI.
+<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:

 ```bash
 pip install lerobot
-lerobot-info
 ```

-> [!IMPORTANT]
-> For detailed installation guide, please see the [Installation Documentation](https://huggingface.co/docs/lerobot/installation).
+_This installs only the default dependencies._

-## 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:
+**Extra Features:**
+To install additional functionality, use one of the following:

 ```bash
-lerobot-train \
-  --policy=act \
-  --dataset.repo_id=lerobot/aloha_mobile_cabinet
+pip install 'lerobot[all]'          # All available features
+pip install 'lerobot[aloha,pusht]'  # Specific features (Aloha & Pusht)
+pip install 'lerobot[feetech]'      # Feetech motor support
 ```

-| 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) |
+_Replace `[...]` with your desired features._

-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
+**Available Tags:**
+For a full list of optional dependencies, see:
+https://pypi.org/project/lerobot/

-For detailed policy setup guides, see the [Policy Documentation](https://huggingface.co/docs/lerobot/bring_your_own_policies).
+> [!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

-## Inference & Evaluation
+### Weights & Biases

-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.
+To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with

 ```bash
-# 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
+wandb login
 ```

-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)
+(note: you will also need to enable WandB in the configuration. See below.)

-## Resources
+### Visualize datasets

- **[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.
+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).

 ## Citation

-If you use LeRobot in your research, please cite:
+If you want, you can cite this work with:

 ```bibtex
@misc{cadene2024lerobot,
@@ -144,14 +339,6 @@ If you use LeRobot in your research, please cite:
 }
 ```

-## Contribute
+## Star History

-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>
+[![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)
@@ -19,8 +19,6 @@
    title: Train RL in Simulation
  - local: multi_gpu_training
    title: Multi GPU training
-  - local: peft_training
-    title: Training with PEFT (e.g., LoRA)
  title: "Tutorials"
 - sections:
  - local: lerobot-dataset-v3
@@ -44,10 +42,6 @@
  - local: xvla
    title: X-VLA
  title: "Policies"
- sections:
-  - local: sarm
-    title: SARM
-  title: "Reward Models"
 - sections:
  - local: async
    title: Use Async Inference
@@ -0,0 +1,328 @@
+# OpenArms Robot
+
+OpenArms is a 7 DOF robotic arm with a gripper, designed by [Enactic, Inc.](https://www.enactic.com/) It uses Damiao motors controlled via CAN bus communication and MIT control mode for smooth, precise motion.
+
+## Hardware Overview
+
+- **7 DOF per arm** (14 DOF total for dual arm setup)
+- **1 gripper per arm** (2 grippers total)
+- **Damiao motors** with 4 different types:
+  - **DM8009** (DM-J8009P-2EC) for shoulders (J1, J2) - high torque
+  - **DM4340** for shoulder rotation and elbow (J3, J4)
+  - **DM4310** (DM-J4310-2EC V1.1) for wrist (J5, J6, J7) and gripper (J8)
+- **24V power supply** required
+- **CAN interface device**:
+  - **Linux**: Any SocketCAN-compatible adapter
+  - **macOS**: CANable, PEAK PCAN-USB, or Kvaser USBcan
+- Proper CAN wiring (CANH, CANL, 120Ω termination)
+
+
+## Motor Configuration
+
+Each arm has the following motor configuration based on the [OpenArm setup guide](https://docs.openarm.dev/software/setup/):
+
+| Joint | Motor | Motor Type | Sender CAN ID | Receiver ID | Description |
+|-------|-------|------------|---------------|-------------|-------------|
+| J1 | joint_1 | DM8009 | 0x01 | 0x11 | Shoulder pan |
+| J2 | joint_2 | DM8009 | 0x02 | 0x12 | Shoulder lift |
+| J3 | joint_3 | DM4340 | 0x03 | 0x13 | Shoulder rotation |
+| J4 | joint_4 | DM4340 | 0x04 | 0x14 | Elbow flex |
+| J5 | joint_5 | DM4310 | 0x05 | 0x15 | Wrist roll |
+| J6 | joint_6 | DM4310 | 0x06 | 0x16 | Wrist pitch |
+| J7 | joint_7 | DM4310 | 0x07 | 0x17 | Wrist rotation |
+| J8 | gripper | DM4310 | 0x08 | 0x18 | Gripper |
+
+For dual arm setups, the left arm uses IDs 0x09-0x10 for joints 1-8 with the same motor types.
+
+## Quick Start 
+
+```bash
+# Install system dependencies
+sudo apt install can-utils iproute2
+
+# Install LeRobot with OpenArms support
+pip install -e ".[openarms]"
+```
+
+## Setup Guide
+
+### Step 1: Motor ID Configuration
+
+**IMPORTANT**: Before using the robot, motors must be configured with the correct CAN IDs.
+
+Refer to the [OpenArm Motor ID Configuration Guide](https://docs.openarm.dev/software/setup/motor-id) for detailed instructions using the Damiao Debugging Tools on Windows.
+
+Key points:
+- Each motor needs a unique **Sender CAN ID** (0x01-0x08)
+- Each motor needs a unique **Receiver/Master ID** (0x11-0x18)
+- Use the Damiao Debugging Tools to set these IDs
+
+### Step 2: Setup CAN Interface
+
+Configure your CAN interface as described in the [OpenArm CAN Setup Guide](https://docs.openarm.dev/software/setup/can-setup):
+
+#### Linux (SocketCAN)
+
+```bash
+# Find your CAN interface
+ip link show
+
+# Configure can0, 1, 2, 3
+sudo ip link set can0 down
+sudo ip link set can0 type can bitrate 1000000
+sudo ip link set can0 up
+
+sudo ip link set can1 down
+sudo ip link set can1 type can bitrate 1000000
+sudo ip link set can1 up
+
+sudo ip link set can2 down
+sudo ip link set can2 type can bitrate 1000000
+sudo ip link set can2 up
+
+sudo ip link set can3 down
+sudo ip link set can3 type can bitrate 1000000
+sudo ip link set can3 up
+
+# Verify configuration
+ip link show can0
+```
+
+or run:
+
+`examples/openarms/setup_can.sh`
+
+### Testing canbus and motor connection
+
+Please run this script to check if all motors can be found and to find your can-fd speed: `python examples/openarms/debug_can_communication.py`
+
+## Usage
+
+### Basic Setup
+
+
+```python
+from lerobot.robots.openarms import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+
+# Configure for dual arm setup
+config = OpenArmsFollowerConfig(
+    port="can0",
+    can_interface="socketcan",  # Or "auto" for auto-detection
+    id="openarms_dual",
+    is_dual_arm=True,
+)
+
+robot = OpenArmsFollower(config)
+robot.connect()
+```
+
+### Calibration
+
+On first use, you'll need to calibrate the robot:
+
+```python
+robot.calibrate()
+```
+
+The calibration process will:
+1. Disable torque on all motors
+2. Ask you to position arms in **hanging position with grippers closed**
+3. Set this as the zero position
+4. Ask you to move each joint through its full range
+5. Record min/max positions for each joint
+6. Save calibration to file
+
+### Reading Observations
+
+The robot provides comprehensive state information:
+
+```python
+observation = robot.get_observation()
+
+# Observation includes for each motor:
+# - {motor_name}.pos: Position in degrees
+# - {motor_name}.vel: Velocity in degrees/second  
+# - {motor_name}.torque: Motor torque
+# - {camera_name}: Camera images (if configured)
+
+print(f"Right arm joint 1 position: {observation['right_joint_1.pos']:.1f}°")
+print(f"Right arm joint 1 velocity: {observation['right_joint_1.vel']:.1f}°/s")
+print(f"Right arm joint 1 torque: {observation['right_joint_1.torque']:.3f} N·m")
+```
+
+### Sending Actions
+
+```python
+# Send target positions (in degrees)
+action = {
+    "right_joint_1.pos": 45.0,
+    "right_joint_2.pos": -30.0,
+    # ... all joints
+    "right_gripper.pos": 45.0,  # Half-closed
+}
+
+actual_action = robot.send_action(action)
+```
+
+### Gripper Control
+
+```python
+# Open gripper
+robot.open_gripper(arm="right")
+
+# Close gripper  
+robot.close_gripper(arm="right")
+```
+
+## Safety Features
+
+### 1. Maximum Relative Target
+
+Limits how far a joint can move in a single command to prevent sudden movements:
+
+```python
+config = OpenArmsFollowerConfig(
+    port="can0",
+    # Limit all joints to 10 degrees per command
+    max_relative_target=10.0,
+    
+    # Or set per-motor limits
+    max_relative_target={
+        "right_joint_1": 15.0,  # Slower moving joint
+        "right_joint_2": 10.0,
+        "right_gripper": 5.0,   # Very slow gripper
+    }
+)
+```
+
+**How it works**: If current position is 50° and you command 80°, with `max_relative_target=10.0`, the robot will only move to 60° in that step.
+
+### 2. Torque Limits
+
+Control maximum torque output, especially important for grippers and teleoperation:
+
+```python
+config = OpenArmsFollowerConfig(
+    port="can0",
+    # Gripper torque limit (fraction of motor's max torque)
+    gripper_torque_limit=0.5,  # 50% of max torque
+)
+```
+
+Lower torque limits prevent damage when gripping delicate objects.
+
+### 3. MIT Control Gains
+
+Control responsiveness and stability via PID-like gains:
+
+```python
+config = OpenArmsFollowerConfig(
+    port="can0",
+    position_kp=10.0,  # Position gain (higher = more responsive)
+    position_kd=0.5,   # Velocity damping (higher = more damped)
+)
+```
+
+**Guidelines**:
+- **For following (robot)**: Higher gains for responsiveness
+  - `position_kp=10.0`, `position_kd=0.5`
+- **For teleoperation (leader)**: Lower gains or disable torque for manual movement
+  - `manual_control=True` (torque disabled)
+
+### 4. Velocity Limits
+
+Velocity limits are enforced by the Damiao motors based on motor type. For DM4310:
+- Max velocity: 30 rad/s ≈ 1718°/s
+
+The motors will automatically limit velocity to safe values.
+
+## Teleoperation
+
+### Leader Arm Setup
+
+The leader arm is moved manually (torque disabled) to generate commands:
+
+```python
+from lerobot.teleoperators.openarms import OpenArmsLeader
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+
+config = OpenArmsLeaderConfig(
+    port="can1",  # Separate CAN interface for leader
+    id="openarms_leader",
+    manual_control=True,  # Torque disabled for manual movement
+    is_dual_arm=True,
+)
+
+leader = OpenArmsLeader(config)
+leader.connect()
+
+# Read current position as action
+action = leader.get_action()
+# action contains positions for all joints in degrees
+```
+
+### Safety Considerations for Teleoperation
+
+1. **Use separate CAN interfaces** for leader and follower to avoid conflicts
+2. **Enable max_relative_target** on follower to smooth abrupt movements
+3. **Lower torque limits** on follower to prevent damage from tracking errors
+4. **Test with one arm** before enabling dual arm teleoperation
+5. **Have emergency stop** ready (power switch or CAN disable)
+
+```python
+# Recommended follower config for teleoperation
+follower_config = OpenArmsFollowerConfig(
+    port="can0",
+    max_relative_target=5.0,  # Small steps for smooth following
+    gripper_torque_limit=0.3, # Low torque for safety
+    position_kp=5.0,  # Lower gains for gentler following
+    position_kd=0.3,
+)
+```
+
+## Troubleshooting
+
+### Motor Shaking/Unstable
+
+- **Lower control gains**: Reduce `position_kp` and `position_kd`
+- **Check calibration**: Re-run calibration procedure
+- **Verify power**: Insufficient current can cause instability
+- **Check mechanical**: Loose connections, binding, or damaged components
+
+### CAN Bus Errors
+
+```bash
+# Check for errors
+ip -s link show can0
+
+# Reset CAN interface
+sudo ip link set can0 down
+sudo ip link set can0 up
+```
+
+### Control Mode
+
+OpenArms uses **MIT control mode** which allows simultaneous control of:
+- Position (degrees)
+- Velocity (degrees/second)
+- Torque (N·m)
+- Position gain (Kp)
+- Velocity damping (Kd)
+
+### Communication
+
+- **Protocol**: CAN 2.0 at 1 Mbps (or CAN-FD at 5 Mbps)
+- **Frame format**: Standard 11-bit IDs
+- **Update rate**: Typically 50-100 Hz depending on motor count
+- **Latency**: ~10-20ms per motor command
+
+## References
+
+- [OpenArm Official Documentation](https://docs.openarm.dev/)
+- [OpenArm Setup Guide](https://docs.openarm.dev/software/setup/)
+- [Motor ID Configuration](https://docs.openarm.dev/software/setup/motor-id)
+- [CAN Interface Setup](https://docs.openarm.dev/software/setup/can-setup)
+- [Motor Communication Test](https://docs.openarm.dev/software/setup/configure-test)
+- [Damiao Motor Documentation](https://wiki.seeedstudio.com/damiao_series/)
+- [Enactic GitHub](https://github.com/enactic/openarm_can)
@@ -1,62 +0,0 @@
-# Parameter efficient fine-tuning with 🤗 PEFT
-
-[🤗 PEFT](https://github.com/huggingface/peft) (Parameter-Efficient Fine-Tuning) is a library for efficiently adapting
-large pretrained models such as pre-trained policies (e.g., SmolVLA, π₀, ...) to new tasks without training all
-of the model's parameters while yielding comparable performance.
-
-Install the `lerobot[peft]` optional package to enable PEFT support.
-
-To read about all the possible methods of adaption, please refer to the [🤗 PEFT docs](https://huggingface.co/docs/peft/index).
-
-## Training SmolVLA
-
-In this section we'll show you how to train a pre-trained SmolVLA policy with PEFT on the libero dataset.
-For brevity we're only training on the `libero_spatial` subset. We will use `lerobot/smolvla_base` as the model
-to parameter efficiently fine-tune:
-
-```
-lerobot-train \
- --policy.path=lerobot/smolvla_base \
- --policy.repo_id=your_hub_name/my_libero_smolvla \
- --dataset.repo_id=HuggingFaceVLA/libero \
- --policy.output_features=null \
- --policy.input_features=null \
- --policy.optimizer_lr=1e-3 \
- --policy.scheduler_decay_lr=1e-4 \
- --env.type=libero \
- --env.task=libero_spatial \
- --steps=100000 \
- --batch_size=32 \
- --peft.method_type=LORA \
- --peft.r=64
-```
-
-Note the `--peft.method_type` parameter that let's you select which PEFT method to use. Here we use
-[LoRA](https://huggingface.co/docs/peft/main/en/package_reference/lora) (Low-Rank Adapter) which is probably the most
-popular fine-tuning method to date. Low-rank adaption means that we only fine-tune a matrix with comparably low rank
-instead of the full weight matrix. This rank can be specified using the `--peft.r` parameter. The higher the rank
-the closer you get to full fine-tuning
-
-There are more complex methods that have more parameters. These are not yet supported, feel free to raise an issue
-if you want to see a specific PEFT method supported.
-
-By default, PEFT will target the `q_proj` and `v_proj` layers of the LM expert in SmolVLA. It will also target the
-state and action projection matrices as they are most likely task-dependent. If you need to target different layers
-you can use `--peft.target_modules` to specify which layers to target. You can refer to the respective PEFT method's
-documentation to see what inputs are supported, (e.g., [LoRA's target_modules documentation](https://huggingface.co/docs/peft/main/en/package_reference/lora#peft.LoraConfig.target_modules)).
-Usually a list of suffixes or a regex are supported. For example, to target the MLPs of the `lm_expert` instead of
-the `q` and `v` projections, use:
-
-```
--peft.target_modules='(model\.vlm_with_expert\.lm_expert\..*\.(down|gate|up)_proj|.*\.(state_proj|action_in_proj|action_out_proj|action_time_mlp_in|action_time_mlp_out))'
-```
-
-In case you need to fully fine-tune a layer instead of just adapting it, you can supply a list of layer suffixes
-to the `--peft.full_training_modules` parameter:
-
-```
--peft.full_training_modules=["state_proj"]
-```
-
-The learning rate and the scheduled target learning rate can usually be scaled by a factor of 10 compared to the
-learning rate used for full fine-tuning (e.g., 1e-4 normal, so 1e-3 using LoRA).
@@ -1,586 +0,0 @@
-# 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}
-}
-```
@@ -0,0 +1,416 @@
+#!/usr/bin/env python3
+"""
+Comprehensive debug script for OpenArms CAN FD communication.
+Tests all 4 CAN interfaces with CAN FD support.
+"""
+
+import can
+import time
+import sys
+import subprocess
+
+def check_can_interface(port):
+    """Check if CAN interface is UP and configured."""
+    try:
+        result = subprocess.run(['ip', 'link', 'show', port], 
+                              capture_output=True, text=True)
+        if result.returncode != 0:
+            return False, "Interface not found", None
+        
+        output = result.stdout
+        if 'UP' not in output:
+            return False, "Interface is DOWN", None
+        
+        # Check if CAN FD is enabled
+        is_fd = 'fd on' in output.lower() or 'canfd' in output.lower()
+        
+        return True, "Interface is UP", is_fd
+    except FileNotFoundError:
+        return None, "Cannot check (ip command not found)", None
+
+
+def test_motor_on_interface(bus, motor_id, timeout=2.0, use_fd=False):
+    """
+    Test a single motor and return all responses.
+    
+    Returns:
+        list of (arbitration_id, data) tuples for all responses received
+    """
+    # Send enable command
+    enable_msg = can.Message(
+        arbitration_id=motor_id,
+        data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFC],
+        is_extended_id=False,
+        is_fd=use_fd
+    )
+    
+    try:
+        bus.send(enable_msg)
+    except Exception as e:
+        return None, f"Send error: {e}"
+    
+    # Listen for responses
+    responses = []
+    start_time = time.time()
+    
+    while time.time() - start_time < timeout:
+        msg = bus.recv(timeout=0.1)
+        if msg:
+            responses.append((msg.arbitration_id, msg.data, msg.is_fd if hasattr(msg, 'is_fd') else False))
+    
+    # Send disable command
+    disable_msg = can.Message(
+        arbitration_id=motor_id,
+        data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFD],
+        is_extended_id=False,
+        is_fd=use_fd
+    )
+    try:
+        bus.send(disable_msg)
+    except:
+        pass
+    
+    return responses, None
+
+
+def test_interface(port, interface_type="socketcan", use_can_fd=True):
+    """Test all 8 motors on a single CAN interface."""
+    
+    results = {
+        'interface': port,
+        'status': None,
+        'is_fd': use_can_fd,
+        'motors': {}
+    }
+    
+    # Check interface status
+    status_ok, status_msg, interface_has_fd = check_can_interface(port)
+    
+    if interface_has_fd is not None:
+        results['interface_fd_enabled'] = interface_has_fd
+        if use_can_fd and not interface_has_fd:
+            status_msg += " (CAN FD NOT enabled on interface!)"
+        elif interface_has_fd:
+            status_msg += " (CAN FD enabled)"
+    
+    results['status'] = status_msg
+    
+    if status_ok is False:
+        return results
+    
+    # Try to connect
+    try:
+        if use_can_fd:
+            print(f"  Connecting to {port} with CAN FD (1 Mbps / 5 Mbps)...")
+            bus = can.interface.Bus(
+                channel=port,
+                interface=interface_type,
+                bitrate=1000000,
+                data_bitrate=5000000,
+                fd=True
+            )
+        else:
+            print(f"  Connecting to {port} with CAN 2.0 (1 Mbps)...")
+            bus = can.interface.Bus(
+                channel=port,
+                interface=interface_type,
+                bitrate=1000000
+            )
+    except Exception as e:
+        results['status'] = f"Connection failed: {e}"
+        return results
+    
+    try:
+        # Clear any pending messages
+        while bus.recv(timeout=0.01):
+            pass
+        
+        # Test each motor (0x01 to 0x08)
+        for motor_id in range(0x01, 0x09):
+            responses, error = test_motor_on_interface(bus, motor_id, timeout=1.0, use_fd=use_can_fd)
+            
+            if error:
+                results['motors'][motor_id] = {'error': error}
+            elif responses:
+                results['motors'][motor_id] = {
+                    'found': True,
+                    'responses': responses
+                }
+            else:
+                results['motors'][motor_id] = {
+                    'found': False,
+                    'responses': []
+                }
+            
+            time.sleep(0.05)  # Small delay between motors
+        
+    finally:
+        bus.shutdown()
+    
+    return results
+
+
+def print_results(all_results):
+    """Print formatted results for all interfaces."""
+    
+    print("SUMMARY - Motors Found on Each Interface")
+    
+    motor_names = {
+        0x01: "joint_1 (Shoulder pan)",
+        0x02: "joint_2 (Shoulder lift)",
+        0x03: "joint_3 (Shoulder rotation)",
+        0x04: "joint_4 (Elbow flex)",
+        0x05: "joint_5 (Wrist roll)",
+        0x06: "joint_6 (Wrist pitch)",
+        0x07: "joint_7 (Wrist rotation)",
+        0x08: "gripper",
+    }
+    
+    total_found = 0
+    
+    for result in all_results:
+        interface = result['interface']
+        status = result['status']
+        
+        print(f"{interface}: {status}")
+        if result.get('is_fd'):
+            print(f"  Mode: CAN FD")
+        else:
+            print(f"  Mode: CAN 2.0")
+        
+        if 'Connection failed' in status or 'DOWN' in status:
+            print(f"  ⚠ Cannot test {interface}")
+            continue
+        
+        motors_found = 0
+        
+        for motor_id in range(0x01, 0x09):
+            motor_data = result['motors'].get(motor_id, {})
+            motor_name = motor_names.get(motor_id, "Unknown")
+            
+            if motor_data.get('error'):
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✗ {motor_data['error']}")
+            elif motor_data.get('found'):
+                motors_found += 1
+                total_found += 1
+                responses = motor_data['responses']
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✓ FOUND")
+                
+                for resp_id, data, is_fd in responses:
+                    data_hex = data.hex()
+                    fd_flag = " [FD]" if is_fd else " [2.0]"
+                    print(f"    → Response from 0x{resp_id:02X}{fd_flag}: {data_hex}")
+            else:
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✗ No response")
+        
+        print(f"\n  Summary: {motors_found}/8 motors found on {interface}")
+    
+    # Overall summary
+    print("OVERALL SUMMARY")
+    print(f"Total motors found across all interfaces: {total_found}")
+    
+    # Analyze configuration
+    print("DIAGNOSIS")
+    
+    for result in all_results:
+        interface = result['interface']
+        motors_found = sum(1 for m in result['motors'].values() if m.get('found'))
+        
+        if motors_found == 0:
+            print(f"\n⚠ {interface}: NO MOTORS FOUND")
+            print("  Possible issues:")
+            print("    1. CAN FD mode mismatch (interface vs motor configuration)")
+            print("    2. Missing 120Ω termination resistors at BOTH cable ends")
+            print("    3. Motor timeout parameter set incorrectly (should NOT be 0)")
+            print("    4. CANH/CANL wiring issue")
+            print("    5. Cable too long (>40m for CAN FD at 5Mbps)")
+            
+            # Check FD mismatch
+            if result.get('is_fd') and not result.get('interface_fd_enabled'):
+                print("    ⚠️ CRITICAL: Trying CAN FD but interface NOT configured for FD!")
+                print(f"       Fix: sudo ip link set {interface} type can bitrate 1000000 dbitrate 5000000 fd on")
+                
+        elif motors_found < 8:
+            print(f"\n⚠ {interface}: Only {motors_found}/8 motors responding")
+            print("  Check power and connections for missing motors")
+        else:
+            print(f"\n✓ {interface}: All 8 motors responding correctly!")
+    
+    # Check for unexpected response IDs
+    print("RESPONSE ID ANALYSIS")
+    
+    for result in all_results:
+        interface = result['interface']
+        unexpected = []
+        
+        for motor_id, motor_data in result['motors'].items():
+            if motor_data.get('found'):
+                expected_id = motor_id + 0x10
+                actual_ids = [resp[0] for resp in motor_data['responses']]
+                
+                if expected_id not in actual_ids:
+                    unexpected.append((motor_id, actual_ids))
+        
+        if unexpected:
+            print(f"\n⚠ {interface}: Unexpected response IDs detected")
+            for motor_id, actual_ids in unexpected:
+                expected_id = motor_id + 0x10
+                print(f"  Motor 0x{motor_id:02X}: Expected 0x{expected_id:02X}, "
+                      f"got {[f'0x{id:02X}' for id in actual_ids]}")
+            print("  → Motor Master IDs need reconfiguration")
+        else:
+            motors_found = sum(1 for m in result['motors'].values() if m.get('found'))
+            if motors_found > 0:
+                print(f"\n✓ {interface}: All responding motors use correct IDs")
+
+
+def test_communication_speed(interface, motor_id, num_iterations=100):
+    """
+    Test communication speed with a motor.
+    
+    Returns:
+        tuple: (hz, avg_latency_ms) or (None, None) if test failed
+    """
+    try:
+        # Connect to interface
+        bus = can.interface.Bus(
+            channel=interface,
+            interface="socketcan",
+            bitrate=1000000,
+            data_bitrate=5000000,
+            fd=True
+        )
+        
+        # Send refresh commands and measure round-trip time
+        latencies = []
+        successful = 0
+        
+        for _ in range(num_iterations):
+            start = time.perf_counter()
+            
+            # Send enable command (lightweight operation)
+            enable_msg = can.Message(
+                arbitration_id=motor_id,
+                data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFC],
+                is_extended_id=False,
+                is_fd=True
+            )
+            bus.send(enable_msg)
+            
+            # Wait for response
+            msg = bus.recv(timeout=0.1)
+            
+            if msg:
+                latency = (time.perf_counter() - start) * 1000  # Convert to ms
+                latencies.append(latency)
+                successful += 1
+        
+        bus.shutdown()
+        
+        if successful > 0:
+            avg_latency = sum(latencies) / len(latencies)
+            hz = 1000.0 / avg_latency if avg_latency > 0 else 0
+            return hz, avg_latency
+        
+        return None, None
+        
+    except Exception as e:
+        print(f"    Speed test error: {e}")
+        return None, None
+
+
+def main():
+    """Main function to test all CAN interfaces with CAN FD."""
+    
+    print("\nThis will test all 4 CAN interfaces (can0-can3) with CAN FD")
+    print("Testing motors 0x01-0x08 on each interface")
+    print()
+    print("Make sure:")
+    print("  ✓ Motors are powered (24V)")
+    print("  ✓ CAN interfaces configured with FD mode:")
+    print("    ./examples/openarms/setup_can.sh")
+    print("  ✓ Motor 'timeout' parameter NOT set to 0 (use Damiao tools)")
+    print("  ✓ CAN wiring includes 120Ω termination at BOTH ends")
+    print()
+    
+    input("Press ENTER to start testing...")
+    
+    # Test all 4 interfaces with CAN FD
+    all_results = []
+    
+    for i in range(4):
+        interface = f"can{i}"
+        print(f"Testing {interface}...")
+        
+        result = test_interface(interface, use_can_fd=True)
+        all_results.append(result)
+        
+        # Quick status
+        if 'Connection failed' in result['status'] or 'DOWN' in result['status']:
+            print(f"  ⚠ {interface}: {result['status']}")
+        else:
+            motors_found = sum(1 for m in result['motors'].values() if m.get('found'))
+            print(f"  {interface}: {motors_found}/8 motors found")
+        
+        time.sleep(0.2)
+    
+    # Print detailed results
+    print_results(all_results)
+    
+    print("Testing Complete!")
+    
+    all_found = sum(sum(1 for m in r['motors'].values() if m.get('found')) for r in all_results)
+    
+    if all_found == 0:
+        print("\n⚠️ CRITICAL: No motors found on any interface!")
+        print("\nTop issues to check:")
+        print("  1. Motor 'timeout' parameter (use Damiao tools to set > 0)")
+        print("  2. CAN FD not enabled (run ./examples/openarms/setup_can.sh)")
+        print("  3. Missing termination resistors")
+        print("\nTry:")
+        print("  a) Check motor parameters with Damiao Debugging Tools")
+        print("  b) Verify CAN FD is enabled: ip -d link show can0 | grep fd")
+        print("  c) Run setup script: ./examples/openarms/setup_can.sh")
+    else:
+        # Run speed test on interfaces with motors
+        print("COMMUNICATION SPEED TEST")
+        print("\nTesting maximum communication frequency...")
+        
+        for result in all_results:
+            interface = result['interface']
+            
+            # Find first responding motor
+            responding_motor = None
+            for motor_id, motor_data in result['motors'].items():
+                if motor_data.get('found'):
+                    responding_motor = motor_id
+                    break
+            
+            if responding_motor:
+                print(f"\n{interface}: Testing with motor 0x{responding_motor:02X}...")
+                hz, latency = test_communication_speed(interface, responding_motor, num_iterations=100)
+                
+                if hz:
+                    print(f"  ✓ Max frequency: {hz:.1f} Hz")
+                    print(f"  ✓ Avg latency: {latency:.2f} ms")
+                    print(f"  ✓ Commands per second: ~{int(hz)}")
+                else:
+                    print(f"  ✗ Speed test failed")
+            else:
+                print(f"\n{interface}: No motors found, skipping speed test")
+
+        print()
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except KeyboardInterrupt:
+        print("\n\nTesting interrupted by user.")
+        sys.exit(1)
+    except Exception as e:
+        print(f"\nUnexpected error: {e}")
+        import traceback
+        traceback.print_exc()
+        sys.exit(1)
+
@@ -0,0 +1,360 @@
+#!/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.
+
+"""
+OpenArms Policy Evaluation
+
+Evaluates a trained policy on the OpenArms robot by running inference and recording
+the evaluation episodes to a dataset. Supports optional leader arm for manual resets.
+
+Example usage:
+    python examples/openarms/evaluate.py
+"""
+
+import time
+from pathlib import Path
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.processor import make_default_processors
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.scripts.lerobot_record import record_loop
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import init_rerun
+
+
+HF_MODEL_ID = "lerobot-data-collection/three-folds-pi0"  # TODO: Replace with your trained model
+HF_EVAL_DATASET_ID = "lerobot-data-collection/three-folds-pi0_eval7"  # TODO: Replace with your eval dataset name
+TASK_DESCRIPTION = "three-folds-dataset"  # TODO: Replace with your task, this should match!!
+
+NUM_EPISODES = 1
+FPS = 30
+EPISODE_TIME_SEC = 300
+RESET_TIME_SEC = 60
+
+# Robot CAN interfaces
+FOLLOWER_LEFT_PORT = "can0"
+FOLLOWER_RIGHT_PORT = "can1"
+
+# If enabled, you can manually reset the environment between evaluation episodes
+USE_LEADER_FOR_RESETS = True  # Set to False if you don't want to use leader
+LEADER_LEFT_PORT = "can2"
+LEADER_RIGHT_PORT = "can3"
+
+# Camera configuration
+CAMERA_CONFIG = {
+    "left_wrist": OpenCVCameraConfig(index_or_path="/dev/video5", width=640, height=480, fps=FPS),
+    "right_wrist": OpenCVCameraConfig(index_or_path="/dev/video1", width=640, height=480, fps=FPS),
+    "base": OpenCVCameraConfig(index_or_path="/dev/video3", width=640, height=480, fps=FPS),
+}
+
+def main():
+    """Main evaluation function."""
+    print("OpenArms Policy Evaluation")
+    print(f"\nModel: {HF_MODEL_ID}")
+    print(f"Evaluation Dataset: {HF_EVAL_DATASET_ID}")
+    print(f"Task: {TASK_DESCRIPTION}")
+    print(f"Episodes: {NUM_EPISODES}")
+    print(f"Episode Duration: {EPISODE_TIME_SEC}s")
+    print(f"Reset Duration: {RESET_TIME_SEC}s")
+    print(f"Use Leader for Resets: {USE_LEADER_FOR_RESETS}")
+    
+    follower_config = OpenArmsFollowerConfig(
+        port_left=FOLLOWER_LEFT_PORT,
+        port_right=FOLLOWER_RIGHT_PORT,
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+        cameras=CAMERA_CONFIG,
+    )
+    
+    follower = OpenArmsFollower(follower_config)
+    follower.connect(calibrate=False)
+    
+    if not follower.is_connected:
+        raise RuntimeError("Follower robot failed to connect!")
+
+    
+    leader = None
+    if USE_LEADER_FOR_RESETS:
+        leader_config = OpenArmsLeaderConfig(
+            port_left=LEADER_LEFT_PORT,
+            port_right=LEADER_RIGHT_PORT,
+            can_interface="socketcan",
+            id="openarms_leader",
+            manual_control=False,  # Enable torque control for gravity compensation
+        )
+        
+        leader = OpenArmsLeader(leader_config)
+        leader.connect(calibrate=False)
+        
+        if not leader.is_connected:
+            raise RuntimeError("Leader robot failed to connect!")
+        
+        # Enable gravity compensation
+        if leader.pin_robot is not None:
+            leader.bus_right.enable_torque()
+            leader.bus_left.enable_torque()
+            time.sleep(0.1)
+            print(f"Leader connected with gravity compensation ({LEADER_LEFT_PORT}, {LEADER_RIGHT_PORT})")
+        else:
+            print(f"Leader connected but gravity compensation unavailable (no URDF)")
+
+    # Build default processors for action and observation
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+    
+    # Build dataset features from robot features and processors
+    # For actions, only include positions (no velocity or torque)
+    action_features_hw = {}
+    for key, value in follower.action_features.items():
+        if key.endswith(".pos"):
+            action_features_hw[key] = value
+    
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_action_processor,
+            initial_features=create_initial_features(action=action_features_hw),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_observation_processor,
+            initial_features=create_initial_features(observation=follower.observation_features),
+            use_videos=True,
+        ),
+    )
+    
+    # Check if dataset already exists
+    dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / HF_EVAL_DATASET_ID
+    if dataset_path.exists():
+        print(f"Evaluation dataset already exists at: {dataset_path}")
+        print("This will append new episodes to the existing dataset.")
+        choice = input("  Continue? (y/n): ").strip().lower()
+        if choice != 'y':
+            print("  Aborting evaluation.")
+            follower.disconnect()
+            if leader:
+                leader.disconnect()
+            return
+    
+    # Create dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_EVAL_DATASET_ID,
+        fps=FPS,
+        features=dataset_features,
+        robot_type=follower.name,
+        use_videos=True,
+        image_writer_processes=0,
+        image_writer_threads=12, 
+    )
+    
+    # Load policy config from pretrained model and create policy using factory
+    policy_config = PreTrainedConfig.from_pretrained(HF_MODEL_ID)
+    policy_config.pretrained_path = HF_MODEL_ID
+    policy = make_policy(policy_config, ds_meta=dataset.meta)
+    
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=policy.config,
+        pretrained_path=HF_MODEL_ID,
+        dataset_stats=dataset.meta.stats,
+        preprocessor_overrides={
+            "device_processor": {"device": str(policy.config.device)}
+        },
+    )
+
+    print(f"\nRunning evaluation...")
+    # Initialize keyboard listener and visualization
+    listener, events = init_keyboard_listener()
+    init_rerun(session_name="openarms_evaluation")
+    episode_idx = 0
+    
+    try:
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Evaluating episode {episode_idx + 1} of {NUM_EPISODES}")
+            print(f"\nRunning inference for episode {episode_idx + 1}...")
+            
+            # Run inference with policy
+            record_loop(
+                robot=follower,
+                events=events,
+                fps=FPS,
+                teleop_action_processor=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
+                policy=policy,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
+                dataset=dataset,
+                control_time_s=EPISODE_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+            )
+            
+            # Handle re-recording
+            if events["rerecord_episode"]:
+                log_say("Re-recording episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
+            
+            # Save episode
+            if dataset.episode_buffer is not None and dataset.episode_buffer.get("size", 0) > 0:
+                print(f"Saving episode {episode_idx + 1} ({dataset.episode_buffer['size']} frames)...")
+                dataset.save_episode()
+                episode_idx += 1
+            
+            # Reset environment between episodes (if not last episode)
+            if not events["stop_recording"] and episode_idx < NUM_EPISODES:
+                if USE_LEADER_FOR_RESETS and leader:
+                    log_say("Reset the environment using leader arms")
+                    print(f"\nManual reset period ({RESET_TIME_SEC}s)...")
+                    
+                    # Use leader for manual reset with gravity compensation
+                    import numpy as np
+                    
+                    dt = 1 / FPS
+                    reset_start_time = time.perf_counter()
+                    
+                    while time.perf_counter() - reset_start_time < RESET_TIME_SEC:
+                        if events["exit_early"] or events["stop_recording"]:
+                            break
+                        
+                        loop_start = time.perf_counter()
+                        
+                        # Get leader state
+                        leader_action = leader.get_action()
+                        
+                        # Extract positions and velocities
+                        leader_positions_deg = {}
+                        leader_velocities_deg_per_sec = {}
+                        
+                        for motor in leader.bus_right.motors:
+                            pos_key = f"right_{motor}.pos"
+                            vel_key = f"right_{motor}.vel"
+                            if pos_key in leader_action:
+                                leader_positions_deg[f"right_{motor}"] = leader_action[pos_key]
+                            if vel_key in leader_action:
+                                leader_velocities_deg_per_sec[f"right_{motor}"] = leader_action[vel_key]
+                        
+                        for motor in leader.bus_left.motors:
+                            pos_key = f"left_{motor}.pos"
+                            vel_key = f"left_{motor}.vel"
+                            if pos_key in leader_action:
+                                leader_positions_deg[f"left_{motor}"] = leader_action[pos_key]
+                            if vel_key in leader_action:
+                                leader_velocities_deg_per_sec[f"left_{motor}"] = leader_action[vel_key]
+                        
+                        # Calculate gravity and friction torques
+                        leader_positions_rad = {k: np.deg2rad(v) for k, v in leader_positions_deg.items()}
+                        leader_gravity_torques_nm = leader._gravity_from_q(leader_positions_rad)
+                        
+                        leader_velocities_rad_per_sec = {k: np.deg2rad(v) for k, v in leader_velocities_deg_per_sec.items()}
+                        leader_friction_torques_nm = leader._friction_from_velocity(
+                            leader_velocities_rad_per_sec,
+                            friction_scale=1.0
+                        )
+                        
+                        # Combine torques
+                        leader_total_torques_nm = {}
+                        for motor_name in leader_gravity_torques_nm:
+                            gravity = leader_gravity_torques_nm.get(motor_name, 0.0)
+                            friction = leader_friction_torques_nm.get(motor_name, 0.0)
+                            leader_total_torques_nm[motor_name] = gravity + friction
+                        
+                        # Apply compensation
+                        for motor in leader.bus_right.motors:
+                            full_name = f"right_{motor}"
+                            position = leader_positions_deg.get(full_name, 0.0)
+                            torque = leader_total_torques_nm.get(full_name, 0.0)
+                            kd = leader.get_damping_kd(motor)
+                            
+                            leader.bus_right._mit_control(
+                                motor=motor, kp=0.0, kd=kd,
+                                position_degrees=position,
+                                velocity_deg_per_sec=0.0,
+                                torque=torque,
+                            )
+                        
+                        for motor in leader.bus_left.motors:
+                            full_name = f"left_{motor}"
+                            position = leader_positions_deg.get(full_name, 0.0)
+                            torque = leader_total_torques_nm.get(full_name, 0.0)
+                            kd = leader.get_damping_kd(motor)
+                            
+                            leader.bus_left._mit_control(
+                                motor=motor, kp=0.0, kd=kd,
+                                position_degrees=position,
+                                velocity_deg_per_sec=0.0,
+                                torque=torque,
+                            )
+                        
+                        # Send leader positions to follower
+                        follower_action = {}
+                        for joint in leader_positions_deg.keys():
+                            pos_key = f"{joint}.pos"
+                            if pos_key in leader_action:
+                                follower_action[pos_key] = leader_action[pos_key]
+                        
+                        if follower_action:
+                            follower.send_action(follower_action)
+                        
+                        # Maintain loop rate
+                        loop_duration = time.perf_counter() - loop_start
+                        sleep_time = dt - loop_duration
+                        if sleep_time > 0:
+                            time.sleep(sleep_time)
+                    
+                    print("Reset complete")
+                else:
+                    log_say("Waiting for manual reset")
+                    print(f"Manually reset the environment and press ENTER to continue")
+                    input("Press ENTER when ready...")
+        
+        print(f"Evaluation complete! {episode_idx} episodes recorded")
+        log_say("Evaluation complete", blocking=True)
+    
+    except KeyboardInterrupt:
+        print("\n\nEvaluation interrupted by user")
+    
+    finally:
+        if leader:
+            leader.bus_right.disable_torque()
+            leader.bus_left.disable_torque()
+            time.sleep(0.1)
+            leader.disconnect()
+
+        follower.disconnect()
+        
+        if listener is not None:
+            listener.stop()
+        
+        dataset.finalize()
+        print("\nUploading to Hugging Face Hub...")
+        dataset.push_to_hub(private=True)
+
+
+if __name__ == "__main__":
+    main()
+
@@ -0,0 +1,653 @@
+#!/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.
+
+"""
+OpenArms Policy Evaluation with Real-Time Chunking (RTC)
+
+Evaluates a trained policy on the OpenArms robot using RTC for smooth, continuous motion.
+RTC enables large flow-matching policies (Pi0, Pi0.5, SmolVLA) to produce reactive motion
+despite high inference latency by asynchronously generating action chunks.
+
+Features:
+- Thread-based asynchronous action generation and execution
+- RTC for smooth transitions between action chunks
+- Dataset recording for evaluation episodes
+
+Example usage:
+    python examples/openarms/evaluate_with_rtc.py
+
+    # With custom RTC parameters
+    python examples/openarms/evaluate_with_rtc.py \
+        --rtc.execution_horizon=12 \
+        --rtc.max_guidance_weight=10.0
+"""
+
+import logging
+import math
+import sys
+import time
+import traceback
+from dataclasses import dataclass, field
+from pathlib import Path
+from threading import Event, Lock, Thread
+
+import torch
+from torch import Tensor
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.processor import make_default_processors
+from lerobot.rl.process import ProcessSignalHandler
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import init_logging, log_say
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+# ============================================================================
+# Default Configuration Constants
+# ============================================================================
+
+DEFAULT_HF_MODEL_ID = "lerobot-data-collection/three-folds-pi0"
+DEFAULT_HF_EVAL_DATASET_ID = "lerobot-data-collection/three-folds-pi0_eval_rtc"
+DEFAULT_TASK_DESCRIPTION = "three-folds-dataset"
+
+DEFAULT_NUM_EPISODES = 1
+DEFAULT_FPS = 30
+DEFAULT_EPISODE_TIME_SEC = 300
+DEFAULT_RESET_TIME_SEC = 60
+
+DEFAULT_FOLLOWER_LEFT_PORT = "can0"
+DEFAULT_FOLLOWER_RIGHT_PORT = "can1"
+
+DEFAULT_CAMERA_CONFIG = {
+    "left_wrist": OpenCVCameraConfig(index_or_path="/dev/video5", width=640, height=480, fps=DEFAULT_FPS),
+    "right_wrist": OpenCVCameraConfig(index_or_path="/dev/video1", width=640, height=480, fps=DEFAULT_FPS),
+    "base": OpenCVCameraConfig(index_or_path="/dev/video3", width=640, height=480, fps=DEFAULT_FPS),
+}
+
+
+# ============================================================================
+# Thread-Safe Robot Wrapper
+# ============================================================================
+
+
+class RobotWrapper:
+    """Thread-safe wrapper for robot operations."""
+
+    def __init__(self, robot: OpenArmsFollower):
+        self.robot = robot
+        self.lock = Lock()
+
+    def get_observation(self) -> dict[str, Tensor]:
+        with self.lock:
+            return self.robot.get_observation()
+
+    def send_action(self, action: dict) -> None:
+        with self.lock:
+            self.robot.send_action(action)
+
+    @property
+    def observation_features(self) -> dict:
+        with self.lock:
+            return self.robot.observation_features
+
+    @property
+    def action_features(self) -> dict:
+        with self.lock:
+            return self.robot.action_features
+
+    @property
+    def name(self) -> str:
+        return self.robot.name
+
+
+# ============================================================================
+# Configuration
+# ============================================================================
+
+
+@dataclass
+class OpenArmsRTCEvalConfig(HubMixin):
+    """Configuration for OpenArms evaluation with RTC."""
+
+    policy: PreTrainedConfig | None = None
+
+    rtc: RTCConfig = field(
+        default_factory=lambda: RTCConfig(
+            enabled=True,
+            execution_horizon=10,
+            max_guidance_weight=10.0,
+            prefix_attention_schedule=RTCAttentionSchedule.EXP,
+        )
+    )
+
+    model_id: str = DEFAULT_HF_MODEL_ID
+    eval_dataset_id: str = DEFAULT_HF_EVAL_DATASET_ID
+    task: str = DEFAULT_TASK_DESCRIPTION
+
+    num_episodes: int = DEFAULT_NUM_EPISODES
+    fps: float = DEFAULT_FPS
+    episode_time_sec: float = DEFAULT_EPISODE_TIME_SEC
+    reset_time_sec: float = DEFAULT_RESET_TIME_SEC
+
+    follower_left_port: str = DEFAULT_FOLLOWER_LEFT_PORT
+    follower_right_port: str = DEFAULT_FOLLOWER_RIGHT_PORT
+
+    device: str = "cuda"
+
+    # Should be higher than inference_delay + execution_horizon
+    action_queue_size_to_get_new_actions: int = 30
+
+    record_dataset: bool = True
+    push_to_hub: bool = True
+
+    use_torch_compile: bool = False
+    torch_compile_backend: str = "inductor"
+    torch_compile_mode: str = "default"
+    torch_compile_disable_cudagraphs: bool = True
+
+    def __post_init__(self):
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+            self.model_id = policy_path
+        elif self.model_id:
+            self.policy = PreTrainedConfig.from_pretrained(self.model_id)
+            self.policy.pretrained_path = self.model_id
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        return ["policy"]
+
+
+# ============================================================================
+# Action Generation Thread
+# ============================================================================
+
+
+def get_actions_thread(
+    policy,
+    robot: RobotWrapper,
+    robot_observation_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: OpenArmsRTCEvalConfig,
+    episode_active: Event,
+):
+    """Thread function to asynchronously generate action chunks from the policy."""
+    try:
+        logger.info("[GET_ACTIONS] Starting action generation thread")
+
+        latency_tracker = LatencyTracker()
+        time_per_chunk = 1.0 / cfg.fps
+
+        hw_features = hw_to_dataset_features(robot.observation_features, "observation")
+        policy_device = policy.config.device
+
+        logger.info(f"[GET_ACTIONS] Loading preprocessor/postprocessor from {cfg.policy.pretrained_path}")
+
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=None,
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.device},
+            },
+        )
+
+        logger.info("[GET_ACTIONS] Preprocessor/postprocessor loaded successfully")
+
+        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
+        if not cfg.rtc.enabled:
+            get_actions_threshold = 0
+
+        while not shutdown_event.is_set():
+            if not episode_active.is_set():
+                time.sleep(0.01)
+                continue
+
+            if action_queue.qsize() <= get_actions_threshold:
+                current_time = time.perf_counter()
+                action_index_before_inference = action_queue.get_action_index()
+                prev_actions = action_queue.get_left_over()
+
+                inference_latency = latency_tracker.max()
+                inference_delay = math.ceil(inference_latency / time_per_chunk) if inference_latency else 0
+
+                obs = robot.get_observation()
+                obs_processed = robot_observation_processor(obs)
+
+                obs_with_policy_features = build_dataset_frame(
+                    hw_features, obs_processed, prefix="observation"
+                )
+
+                for name in obs_with_policy_features:
+                    obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
+                    if "image" in name:
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].type(torch.float32) / 255
+                        )
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].permute(2, 0, 1).contiguous()
+                        )
+                    obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0)
+                    obs_with_policy_features[name] = obs_with_policy_features[name].to(policy_device)
+
+                obs_with_policy_features["task"] = [cfg.task]
+                obs_with_policy_features["robot_type"] = robot.name
+
+                preprocessed_obs = preprocessor(obs_with_policy_features)
+
+                actions = policy.predict_action_chunk(
+                    preprocessed_obs,
+                    inference_delay=inference_delay,
+                    prev_chunk_left_over=prev_actions,
+                )
+
+                original_actions = actions.squeeze(0).clone()
+                postprocessed_actions = postprocessor(actions).squeeze(0)
+
+                new_latency = time.perf_counter() - current_time
+                new_delay = math.ceil(new_latency / time_per_chunk)
+                latency_tracker.add(new_latency)
+
+                if cfg.action_queue_size_to_get_new_actions < cfg.rtc.execution_horizon + new_delay:
+                    logger.warning(
+                        "[GET_ACTIONS] action_queue_size_to_get_new_actions too small. "
+                        "Should be higher than inference delay + execution horizon."
+                    )
+
+                action_queue.merge(
+                    original_actions, postprocessed_actions, new_delay, action_index_before_inference
+                )
+
+                logger.debug(
+                    f"[GET_ACTIONS] Generated chunk, latency={new_latency:.3f}s, "
+                    f"delay={new_delay}, queue_size={action_queue.qsize()}"
+                )
+            else:
+                time.sleep(0.01)
+
+        logger.info("[GET_ACTIONS] Action generation thread shutting down")
+    except Exception as e:
+        logger.error(f"[GET_ACTIONS] Fatal exception: {e}")
+        logger.error(traceback.format_exc())
+        shutdown_event.set()
+        sys.exit(1)
+
+
+# ============================================================================
+# Action Execution Thread
+# ============================================================================
+
+
+def actor_thread(
+    robot: RobotWrapper,
+    robot_action_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: OpenArmsRTCEvalConfig,
+    episode_active: Event,
+    dataset: LeRobotDataset | None,
+    dataset_lock: Lock,
+    teleop_action_processor,
+    robot_observation_processor,
+):
+    """Thread function to execute actions on the robot."""
+    try:
+        logger.info("[ACTOR] Starting actor thread")
+
+        action_count = 0
+        action_interval = 1.0 / cfg.fps
+        action_keys = [k for k in robot.action_features.keys() if k.endswith(".pos")]
+
+        while not shutdown_event.is_set():
+            if not episode_active.is_set():
+                time.sleep(0.01)
+                continue
+
+            start_time = time.perf_counter()
+            action = action_queue.get()
+
+            if action is not None:
+                action = action.cpu()
+
+                action_dict = {}
+                for i, key in enumerate(action_keys):
+                    if i < len(action):
+                        action_dict[key] = action[i].item()
+
+                action_processed = robot_action_processor((action_dict, None))
+                robot.send_action(action_processed)
+
+                if cfg.record_dataset and dataset is not None:
+                    with dataset_lock:
+                        obs = robot.get_observation()
+                        obs_processed = robot_observation_processor(obs)
+                        action_for_dataset = teleop_action_processor((action_dict, None))
+
+                        frame = {}
+                        for key, value in obs_processed.items():
+                            frame[f"observation.{key}"] = value
+                        for key, value in action_for_dataset.items():
+                            frame[f"action.{key}"] = value
+                        frame["task"] = cfg.task
+
+                        dataset.add_frame(frame)
+
+                action_count += 1
+
+            dt_s = time.perf_counter() - start_time
+            sleep_time = max(0, action_interval - dt_s - 0.001)
+            if sleep_time > 0:
+                time.sleep(sleep_time)
+
+        logger.info(f"[ACTOR] Actor thread shutting down. Total actions executed: {action_count}")
+    except Exception as e:
+        logger.error(f"[ACTOR] Fatal exception: {e}")
+        logger.error(traceback.format_exc())
+        shutdown_event.set()
+        sys.exit(1)
+
+
+# ============================================================================
+# Main Evaluation Function
+# ============================================================================
+
+
+def _apply_torch_compile(policy, cfg: OpenArmsRTCEvalConfig):
+    """Apply torch.compile to the policy's predict_action_chunk method."""
+    if policy.name in ["pi05", "pi0"]:
+        return policy
+
+    try:
+        if not hasattr(torch, "compile"):
+            logger.warning(
+                f"torch.compile not available. Requires PyTorch 2.0+. "
+                f"Current version: {torch.__version__}. Skipping compilation."
+            )
+            return policy
+
+        logger.info("Applying torch.compile to predict_action_chunk...")
+
+        compile_kwargs = {
+            "backend": cfg.torch_compile_backend,
+            "mode": cfg.torch_compile_mode,
+        }
+
+        if cfg.torch_compile_disable_cudagraphs:
+            compile_kwargs["options"] = {"triton.cudagraphs": False}
+
+        original_method = policy.predict_action_chunk
+        compiled_method = torch.compile(original_method, **compile_kwargs)
+        policy.predict_action_chunk = compiled_method
+        logger.info("Successfully compiled predict_action_chunk")
+
+    except Exception as e:
+        logger.error(f"Failed to apply torch.compile: {e}")
+        logger.warning("Continuing without torch.compile")
+
+    return policy
+
+
+@parser.wrap()
+def main(cfg: OpenArmsRTCEvalConfig):
+    """Main evaluation function with RTC."""
+    init_logging()
+
+    print("=" * 60)
+    print("OpenArms Policy Evaluation with RTC")
+    print("=" * 60)
+    print(f"\nModel: {cfg.model_id}")
+    print(f"Evaluation Dataset: {cfg.eval_dataset_id}")
+    print(f"Task: {cfg.task}")
+    print(f"Episodes: {cfg.num_episodes}")
+    print(f"Episode Duration: {cfg.episode_time_sec}s")
+    print(f"RTC Enabled: {cfg.rtc.enabled}")
+    print(f"RTC Execution Horizon: {cfg.rtc.execution_horizon}")
+    print(f"RTC Max Guidance Weight: {cfg.rtc.max_guidance_weight}")
+    print(f"Device: {cfg.device}")
+    print("=" * 60)
+
+    signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
+    shutdown_event = signal_handler.shutdown_event
+    episode_active = Event()
+
+    # Initialize Robot
+    follower_config = OpenArmsFollowerConfig(
+        port_left=cfg.follower_left_port,
+        port_right=cfg.follower_right_port,
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+        cameras=DEFAULT_CAMERA_CONFIG,
+    )
+
+    follower = OpenArmsFollower(follower_config)
+    follower.connect(calibrate=False)
+
+    if not follower.is_connected:
+        raise RuntimeError("Follower robot failed to connect!")
+
+    robot = RobotWrapper(follower)
+    logger.info("Follower robot connected")
+
+    # Build Processors and Dataset Features
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+
+    action_features_hw = {}
+    for key, value in follower.action_features.items():
+        if key.endswith(".pos"):
+            action_features_hw[key] = value
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_action_processor,
+            initial_features=create_initial_features(action=action_features_hw),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_observation_processor,
+            initial_features=create_initial_features(observation=follower.observation_features),
+            use_videos=True,
+        ),
+    )
+
+    # Create or Load Dataset
+    dataset = None
+    dataset_lock = Lock()
+
+    if cfg.record_dataset:
+        dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / cfg.eval_dataset_id
+        if dataset_path.exists():
+            logger.info(f"Evaluation dataset exists at: {dataset_path}")
+            logger.info("New episodes will be appended.")
+            choice = input("Continue? (y/n): ").strip().lower()
+            if choice != "y":
+                logger.info("Aborting evaluation.")
+                follower.disconnect()
+                return
+
+        dataset = LeRobotDataset.create(
+            repo_id=cfg.eval_dataset_id,
+            fps=int(cfg.fps),
+            features=dataset_features,
+            robot_type=follower.name,
+            use_videos=True,
+            image_writer_processes=0,
+            image_writer_threads=12,
+        )
+        logger.info(f"Dataset created: {cfg.eval_dataset_id}")
+
+    # Load Policy
+    logger.info(f"Loading policy from: {cfg.model_id}")
+
+    policy_class = get_policy_class(cfg.policy.type)
+    config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
+
+    if cfg.policy.type in ["pi05", "pi0"]:
+        config.compile_model = cfg.use_torch_compile
+
+    policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
+
+    policy.config.rtc_config = cfg.rtc
+    policy.init_rtc_processor()
+
+    assert policy.name in ["smolvla", "pi05", "pi0"], "Only smolvla, pi05, and pi0 are supported for RTC"
+
+    policy = policy.to(cfg.device)
+    policy.eval()
+
+    if cfg.use_torch_compile:
+        policy = _apply_torch_compile(policy, cfg)
+
+    logger.info(f"Policy loaded: {policy.name}")
+
+    # Create Action Queue and Start Threads
+    action_queue = ActionQueue(cfg.rtc)
+
+    get_actions_t = Thread(
+        target=get_actions_thread,
+        args=(
+            policy,
+            robot,
+            robot_observation_processor,
+            action_queue,
+            shutdown_event,
+            cfg,
+            episode_active,
+        ),
+        daemon=True,
+        name="GetActions",
+    )
+    get_actions_t.start()
+    logger.info("Started action generation thread")
+
+    actor_t = Thread(
+        target=actor_thread,
+        args=(
+            robot,
+            robot_action_processor,
+            action_queue,
+            shutdown_event,
+            cfg,
+            episode_active,
+            dataset,
+            dataset_lock,
+            teleop_action_processor,
+            robot_observation_processor,
+        ),
+        daemon=True,
+        name="Actor",
+    )
+    actor_t.start()
+    logger.info("Started actor thread")
+
+    # Run Evaluation Episodes
+    episode_idx = 0
+
+    try:
+        while episode_idx < cfg.num_episodes and not shutdown_event.is_set():
+            log_say(f"Evaluating episode {episode_idx + 1} of {cfg.num_episodes}")
+            logger.info(f"\n{'='*40}")
+            logger.info(f"Episode {episode_idx + 1} / {cfg.num_episodes}")
+            logger.info(f"{'='*40}")
+
+            action_queue = ActionQueue(cfg.rtc)
+            episode_active.set()
+            episode_start_time = time.time()
+
+            while (time.time() - episode_start_time) < cfg.episode_time_sec:
+                if shutdown_event.is_set():
+                    break
+
+                elapsed = time.time() - episode_start_time
+                if int(elapsed) % 10 == 0 and int(elapsed) > 0:
+                    logger.info(
+                        f"[MAIN] Episode progress: {elapsed:.0f}/{cfg.episode_time_sec}s, "
+                        f"queue_size={action_queue.qsize()}"
+                    )
+
+                time.sleep(0.5)
+
+            episode_active.clear()
+            logger.info(f"Episode {episode_idx + 1} completed")
+
+            if cfg.record_dataset and dataset is not None:
+                with dataset_lock:
+                    if dataset.episode_buffer is not None and dataset.episode_buffer.get("size", 0) > 0:
+                        logger.info(
+                            f"Saving episode {episode_idx + 1} "
+                            f"({dataset.episode_buffer['size']} frames)"
+                        )
+                        dataset.save_episode()
+
+            episode_idx += 1
+
+            # Manual reset between episodes
+            if not shutdown_event.is_set() and episode_idx < cfg.num_episodes:
+                log_say("Waiting for manual reset")
+                logger.info("Manually reset the environment and press ENTER to continue")
+                input("Press ENTER when ready...")
+
+        logger.info(f"Evaluation complete! {episode_idx} episodes recorded")
+        log_say("Evaluation complete", blocking=True)
+
+    except KeyboardInterrupt:
+        logger.info("\n\nEvaluation interrupted by user")
+
+    finally:
+        shutdown_event.set()
+        episode_active.clear()
+
+        if get_actions_t.is_alive():
+            logger.info("Waiting for action generation thread to finish...")
+            get_actions_t.join(timeout=5.0)
+
+        if actor_t.is_alive():
+            logger.info("Waiting for actor thread to finish...")
+            actor_t.join(timeout=5.0)
+
+        follower.disconnect()
+        logger.info("Follower disconnected")
+
+        if cfg.record_dataset and dataset is not None:
+            dataset.finalize()
+            if cfg.push_to_hub:
+                logger.info("Uploading to Hugging Face Hub...")
+                dataset.push_to_hub(private=True)
+
+        logger.info("Cleanup completed")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,216 @@
+import time
+import numpy as np
+
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+
+
+# Friction model parameters from OpenArms config/follower.yaml
+# τ_fric(ω) = Fo + Fv·ω + Fc·tanh(k·ω)
+# For 8 motors: [joint_1, joint_2, joint_3, joint_4, joint_5, joint_6, joint_7, gripper]
+FRICTION_PARAMS = {
+    "Fc": [0.306, 0.306, 0.40, 0.166, 0.050, 0.093, 0.172, 0.0512],  # Coulomb friction [Nm]
+    "k": [28.417, 28.417, 29.065, 130.038, 151.771, 242.287, 7.888, 4.000],  # tanh steepness
+    "Fv": [0.063, 0.0630, 0.604, 0.813, 0.029, 0.072, 0.084, 0.084],  # Viscous friction [Nm·s/rad]
+    "Fo": [0.088, 0.088, 0.008, -0.058, 0.005, 0.009, -0.059, -0.050],  # Offset torque [Nm]
+}
+
+# Constants from OpenArms C++ implementation
+AMP_TMP = 1.0
+COEF_TMP = 0.1
+
+FRICTION_SCALE = 1.0  # OpenArms C++ uses 0.3 factor in unilateral mode
+DAMPING_KD = [0.5, 0.5, 0.5, 0.5, 0.1, 0.1, 0.1, 0.1]  # Damping gains for stability
+
+def compute_friction_torque(velocity_rad_per_sec: float, motor_index: int) -> float:
+    """
+    Compute friction torque for a single motor using the tanh friction model.
+    
+    Args:
+        velocity_rad_per_sec: Angular velocity in rad/s
+        motor_index: Index of the motor (0-7)
+        
+    Returns:
+        Friction torque in N·m (scaled for stability)
+    """
+    
+    Fc = FRICTION_PARAMS["Fc"][motor_index]
+    k = FRICTION_PARAMS["k"][motor_index]
+    Fv = FRICTION_PARAMS["Fv"][motor_index]
+    Fo = FRICTION_PARAMS["Fo"][motor_index]
+    
+    # Friction model: τ_fric = amp * Fc * tanh(coef * k * ω) + Fv * ω + Fo
+    friction_torque = (
+        AMP_TMP * Fc * np.tanh(COEF_TMP * k * velocity_rad_per_sec) +
+        Fv * velocity_rad_per_sec +
+        Fo
+    )
+    
+    # Scale down friction compensation for stability at lower control rates
+    # (OpenArms C++ uses 0.3 factor in unilateral mode)!!
+    friction_torque *= FRICTION_SCALE
+    
+    return friction_torque
+
+
+def main() -> None:
+    config = OpenArmsFollowerConfig(
+        port_left="can0",
+        port_right="can1",
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=5.0,
+    )
+    
+    print("Initializing robot...")
+    follower = OpenArmsFollower(config)
+    follower.connect(calibrate=True)
+    
+    print(f"Applying friction compensation")
+    print("  1. Support the arm before starting")
+    print("  2. The arm will be held in place by friction compensation")
+    print("  3. You should be able to move it with gentle force")
+    print("\nPress ENTER when ready to start...")
+    input()
+    
+    print(f"✓ Motors enabled")
+    print("\nStarting friction compensation loop...")
+    print("Press Ctrl+C to stop\n")
+    
+    loop_times = []
+    last_print_time = time.perf_counter()
+    
+    # Motor name to index mapping
+    motor_name_to_index = {
+        "joint_1": 0,
+        "joint_2": 1,
+        "joint_3": 2,
+        "joint_4": 3,
+        "joint_5": 4,
+        "joint_6": 5,
+        "joint_7": 6,
+        "gripper": 7,
+    }
+    
+    try:
+        while True:
+            loop_start = time.perf_counter()
+            
+            # Get current joint positions and velocities from robot
+            obs = follower.get_observation()
+            
+            # Extract velocities in degrees per second
+            velocities_deg_per_sec = {}
+            positions_deg = {}
+            
+            for motor in follower.bus_right.motors:
+                vel_key = f"right_{motor}.vel"
+                pos_key = f"right_{motor}.pos"
+                if vel_key in obs:
+                    velocities_deg_per_sec[f"right_{motor}"] = obs[vel_key]
+                if pos_key in obs:
+                    positions_deg[f"right_{motor}"] = obs[pos_key]
+            
+            for motor in follower.bus_left.motors:
+                vel_key = f"left_{motor}.vel"
+                pos_key = f"left_{motor}.pos"
+                if vel_key in obs:
+                    velocities_deg_per_sec[f"left_{motor}"] = obs[vel_key]
+                if pos_key in obs:
+                    positions_deg[f"left_{motor}"] = obs[pos_key]
+            
+            # Convert velocities to rad/s and compute friction torques
+            friction_torques_nm = {}
+            for motor_full_name, velocity_deg_per_sec in velocities_deg_per_sec.items():
+                # Extract motor name without arm prefix
+                if motor_full_name.startswith("right_"):
+                    motor_name = motor_full_name.removeprefix("right_")
+                elif motor_full_name.startswith("left_"):
+                    motor_name = motor_full_name.removeprefix("left_")
+                else:
+                    continue
+                
+                # Get motor index for friction parameters
+                motor_index = motor_name_to_index.get(motor_name, 0)
+                
+                # Convert velocity to rad/s
+                velocity_rad_per_sec = np.deg2rad(velocity_deg_per_sec)
+                
+                # Compute friction torque
+                friction_torque = compute_friction_torque(velocity_rad_per_sec, motor_index)
+                friction_torques_nm[motor_full_name] = friction_torque
+            
+            # Apply friction compensation to right arm (all joints INCLUDING gripper)
+            for motor in follower.bus_right.motors:
+                full_name = f"right_{motor}"
+                position = positions_deg.get(full_name, 0.0)
+                torque = friction_torques_nm.get(full_name, 0.0)
+                
+                # Get motor index for damping gain
+                motor_index = motor_name_to_index.get(motor, 0)
+                kd = DAMPING_KD[motor_index]
+                
+                # Send MIT control command with friction compensation + damping
+                follower.bus_right._mit_control(
+                    motor=motor,
+                    kp=0.0,  # No position control
+                    kd=kd,   # Add damping for stability
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque
+                )
+
+            # Apply friction compensation to left arm (all joints INCLUDING gripper)
+            for motor in follower.bus_left.motors:
+                full_name = f"left_{motor}"
+                position = positions_deg.get(full_name, 0.0)
+                torque = friction_torques_nm.get(full_name, 0.0)
+                
+                # Get motor index for damping gain
+                motor_index = motor_name_to_index.get(motor, 0)
+                kd = DAMPING_KD[motor_index]
+                
+                # Send MIT control command with friction compensation + damping
+                follower.bus_left._mit_control(
+                    motor=motor,
+                    kp=0.0,  # No position control
+                    kd=kd,   # Add damping for stability
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque
+                )
+            
+            # Measure loop time
+            loop_end = time.perf_counter()
+            loop_time = loop_end - loop_start
+            loop_times.append(loop_time)
+            
+            # Print status every 2 seconds
+            if loop_end - last_print_time >= 2.0:
+                if loop_times:
+                    avg_time = sum(loop_times) / len(loop_times)
+                    current_hz = 1.0 / avg_time if avg_time > 0 else 0
+                    
+                    print(f"{current_hz:.1f} Hz")
+
+                    loop_times = []
+                    last_print_time = loop_end
+            
+            time.sleep(0.001)
+            
+    except KeyboardInterrupt:
+        print("\n\nStopping friction compensation...")
+    
+    finally:
+        print("\nDisabling all motors and disconnecting...")
+        follower.bus_right.disable_torque()
+        follower.bus_left.disable_torque()
+        time.sleep(0.1)
+        follower.disconnect()
+        print("✓ Safe shutdown complete")
+
+
+if __name__ == "__main__":
+    main()
+
@@ -0,0 +1,142 @@
+import time
+import numpy as np
+import pinocchio as pin
+from os.path import join, dirname, exists, expanduser
+
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+
+
+def main() -> None:
+    config = OpenArmsFollowerConfig(
+        port_left="can0",
+        port_right="can1",
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=5.0,
+    )
+    
+    
+    print("Initializing robot...")
+    follower = OpenArmsFollower(config)
+    follower.connect(calibrate=True)
+    
+    # Load URDF for Pinocchio dynamics
+    urdf_path = "/home/croissant/Documents/openarm_description/openarm_bimanual_pybullet.urdf"
+    
+    pin_robot = pin.RobotWrapper.BuildFromURDF(urdf_path, dirname(urdf_path))
+    pin_robot.data = pin_robot.model.createData()
+    print(f"✓ Loaded Pinocchio model with {pin_robot.nq} DoFs")
+    
+    follower.pin_robot = pin_robot
+    
+    print(f"Applying gravity compensation")
+    print("  1. Support the arm before starting")
+    print("  2. The arm will be held in place by gravity compensation")
+    print("  3. You should be able to move it with gentle force")
+    print("\nPress ENTER when ready to start...")
+    input()
+    
+    print(f"✓ Motors enabled")
+    print("\nStarting gravity compensation loop...")
+    print("Press Ctrl+C to stop\n")
+    
+    loop_times = []
+    last_print_time = time.perf_counter()
+    
+    try:
+        while True:
+            loop_start = time.perf_counter()
+            
+            # Get current joint positions from robot
+            obs = follower.get_observation()
+            
+            # Extract positions in degrees
+            positions_deg = {}
+            for motor in follower.bus_right.motors:
+                key = f"right_{motor}.pos"
+                if key in obs:
+                    positions_deg[f"right_{motor}"] = obs[key]
+            
+            for motor in follower.bus_left.motors:
+                key = f"left_{motor}.pos"
+                if key in obs:
+                    positions_deg[f"left_{motor}"] = obs[key]
+            
+            # Convert to radians and calculate gravity torques
+            # Use the built-in method from OpenArmsFollower
+            positions_rad = {k: np.deg2rad(v) for k, v in positions_deg.items()}
+            torques_nm = follower._gravity_from_q(positions_rad)
+            
+            # Apply gravity compensation to right arm (all joints except gripper)
+            for motor in follower.bus_right.motors:
+                if motor == "gripper":
+                    continue  # Skip gripper
+                    
+                full_name = f"right_{motor}"
+                position = positions_deg.get(full_name, 0.0)
+                torque = torques_nm.get(full_name, 0.0)
+                
+                # Send MIT control command with gravity compensation torque
+                follower.bus_right._mit_control(
+                    motor=motor,
+                    kp=0.0,  # No position control
+                    kd=0.0,  # No velocity damping
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque
+                )
+
+            # Apply gravity compensation to left arm (all joints except gripper)
+            for motor in follower.bus_left.motors:
+                if motor == "gripper":
+                    continue  # Skip gripper
+                    
+                full_name = f"left_{motor}"
+                position = positions_deg.get(full_name, 0.0)
+                torque = torques_nm.get(full_name, 0.0)
+                
+                # Send MIT control command with gravity compensation torque
+                follower.bus_left._mit_control(
+                    motor=motor,
+                    kp=0.0,  # No position control
+                    kd=0.0,  # No velocity damping
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque
+                )
+            
+            # Measure loop time
+            loop_end = time.perf_counter()
+            loop_time = loop_end - loop_start
+            loop_times.append(loop_time)
+            
+            # Print status every 2 seconds
+            if loop_end - last_print_time >= 2.0:
+                if loop_times:
+                    avg_time = sum(loop_times) / len(loop_times)
+                    current_hz = 1.0 / avg_time if avg_time > 0 else 0
+                    
+                    print(f"{current_hz:.1f} Hz ({avg_time*1000:.1f} ms)")
+
+                    loop_times = []
+                    last_print_time = loop_end
+            
+            time.sleep(0.005)
+            
+    except KeyboardInterrupt:
+        print("\n\nStopping gravity compensation...")
+    
+    finally:
+        print("\nDisabling all motors and disconnecting...")
+        follower.bus_right.disable_torque()
+        follower.bus_left.disable_torque()
+        time.sleep(0.1)
+        follower.disconnect()
+        print("✓ Safe shutdown complete")
+
+
+if __name__ == "__main__":
+    main()
+
@@ -0,0 +1,395 @@
+"""
+OpenArms Dataset Recording with Gravity + Friction Compensation
+
+Records a dataset using OpenArms follower robot with leader teleoperator.
+Leader arms have gravity and friction compensation for weightless, easy movement.
+Includes 3 cameras: left wrist, right wrist, and base camera.
+
+Uses the same compensation approach as teleop_with_compensation.py
+"""
+
+import shutil
+import time
+from pathlib import Path
+
+import numpy as np
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+
+# Recording parameters
+NUM_EPISODES = 1
+FPS = 30
+EPISODE_TIME_SEC = 600
+RESET_TIME_SEC = 120
+TASK_DESCRIPTION = "OpenArms task description"
+
+# Friction compensation scale factor (1.0 = full, 0.3 = 30% for stability)
+FRICTION_SCALE = 1.0
+
+def record_loop_with_compensation(
+    robot,
+    leader,
+    events,
+    fps,
+    dataset,
+    dataset_features,
+    control_time_s,
+    single_task,
+    display_data=True,
+):
+    """
+    Custom record loop that applies gravity + friction compensation to leader.
+    Based on record_loop but with integrated compensation.
+    """
+    dt = 1 / fps
+    episode_start_time = time.perf_counter()
+    
+    # All joints (both arms)
+    all_joints = []
+    for motor in leader.bus_right.motors:
+        all_joints.append(f"right_{motor}")
+    for motor in leader.bus_left.motors:
+        all_joints.append(f"left_{motor}")
+    
+    while True:
+        loop_start = time.perf_counter()
+        elapsed = loop_start - episode_start_time
+        
+        # Check if we should exit
+        if elapsed >= control_time_s or events["exit_early"] or events["stop_recording"]:
+            break
+        
+        # Get leader state
+        leader_action = leader.get_action()
+        
+        # Extract positions and velocities in degrees
+        leader_positions_deg = {}
+        leader_velocities_deg_per_sec = {}
+        
+        for motor in leader.bus_right.motors:
+            pos_key = f"right_{motor}.pos"
+            vel_key = f"right_{motor}.vel"
+            if pos_key in leader_action:
+                leader_positions_deg[f"right_{motor}"] = leader_action[pos_key]
+            if vel_key in leader_action:
+                leader_velocities_deg_per_sec[f"right_{motor}"] = leader_action[vel_key]
+        
+        for motor in leader.bus_left.motors:
+            pos_key = f"left_{motor}.pos"
+            vel_key = f"left_{motor}.vel"
+            if pos_key in leader_action:
+                leader_positions_deg[f"left_{motor}"] = leader_action[pos_key]
+            if vel_key in leader_action:
+                leader_velocities_deg_per_sec[f"left_{motor}"] = leader_action[vel_key]
+        
+        # Calculate gravity torques for leader using built-in method
+        leader_positions_rad = {k: np.deg2rad(v) for k, v in leader_positions_deg.items()}
+        leader_gravity_torques_nm = leader._gravity_from_q(leader_positions_rad)
+        
+        # Calculate friction torques for leader using built-in method
+        leader_velocities_rad_per_sec = {k: np.deg2rad(v) for k, v in leader_velocities_deg_per_sec.items()}
+        leader_friction_torques_nm = leader._friction_from_velocity(
+            leader_velocities_rad_per_sec,
+            friction_scale=FRICTION_SCALE
+        )
+        
+        # Combine gravity + friction torques
+        leader_total_torques_nm = {}
+        for motor_name in leader_gravity_torques_nm:
+            gravity = leader_gravity_torques_nm.get(motor_name, 0.0)
+            friction = leader_friction_torques_nm.get(motor_name, 0.0)
+            leader_total_torques_nm[motor_name] = gravity + friction
+        
+        # Apply gravity + friction compensation to leader RIGHT arm (all joints including gripper)
+        for motor in leader.bus_right.motors:
+            full_name = f"right_{motor}"
+            position = leader_positions_deg.get(full_name, 0.0)
+            torque = leader_total_torques_nm.get(full_name, 0.0)
+            
+            # Get damping gain for stability
+            kd = leader.get_damping_kd(motor)
+            
+            leader.bus_right._mit_control(
+                motor=motor,
+                kp=0.0,
+                kd=kd,  # Add damping for stability
+                position_degrees=position,
+                velocity_deg_per_sec=0.0,
+                torque=torque,
+            )
+        
+        # Apply gravity + friction compensation to leader LEFT arm (all joints including gripper)
+        for motor in leader.bus_left.motors:
+            full_name = f"left_{motor}"
+            position = leader_positions_deg.get(full_name, 0.0)
+            torque = leader_total_torques_nm.get(full_name, 0.0)
+            
+            # Get damping gain for stability
+            kd = leader.get_damping_kd(motor)
+            
+            leader.bus_left._mit_control(
+                motor=motor,
+                kp=0.0,
+                kd=kd,  # Add damping for stability
+                position_degrees=position,
+                velocity_deg_per_sec=0.0,
+                torque=torque,
+            )
+        
+        # Send leader positions to follower (both arms)
+        follower_action = {}
+        for joint in all_joints:
+            pos_key = f"{joint}.pos"
+            if pos_key in leader_action:
+                follower_action[pos_key] = leader_action[pos_key]
+        
+        # Send action to robot
+        if follower_action:
+            robot.send_action(follower_action)
+        
+        # Get observation from robot (includes camera images)
+        observation = robot.get_observation()
+        
+        # Add to dataset if we have a dataset
+        if dataset is not None:
+            # Build properly formatted observation frame
+            obs_frame = build_dataset_frame(dataset_features, observation, prefix="observation")
+            
+            # Build properly formatted action frame (keep .pos suffix - it matches the feature names)
+            action_frame = build_dataset_frame(dataset_features, follower_action, prefix="action")
+            
+            # Combine into single frame
+            frame = {**obs_frame, **action_frame}
+            
+            # Add metadata (task is required, timestamp will be auto-calculated by add_frame)
+            frame["task"] = single_task
+            
+            dataset.add_frame(frame)
+        
+        # Display data if requested
+        if display_data:
+            log_rerun_data(observation=observation, action=follower_action)
+        
+        # Maintain loop rate
+        loop_duration = time.perf_counter() - loop_start
+        sleep_time = dt - loop_duration
+        if sleep_time > 0:
+            time.sleep(sleep_time)
+
+
+def main():
+    """Main recording loop with gravity compensation."""
+    
+    print("=" * 70)
+    print("OpenArms Dataset Recording with Compensation")
+    print("=" * 70)
+    
+    # Create camera configurations (3 cameras: left wrist, right wrist, base)
+    # Using actual device paths found by lerobot-find-cameras opencv
+    camera_config = {
+        "left_wrist": OpenCVCameraConfig(index_or_path="/dev/video0", width=640, height=480, fps=FPS),
+        "right_wrist": OpenCVCameraConfig(index_or_path="/dev/video1", width=640, height=480, fps=FPS),
+        "base": OpenCVCameraConfig(index_or_path="/dev/video7", width=640, height=480, fps=FPS),
+    }
+    
+    # Configure follower robot with cameras
+    follower_config = OpenArmsFollowerConfig(
+        port_left="can2",
+        port_right="can3",
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+        cameras=camera_config,
+    )
+    
+    # Configure leader teleoperator (no cameras needed)
+    leader_config = OpenArmsLeaderConfig(
+        port_left="can0",
+        port_right="can1",
+        can_interface="socketcan",
+        id="openarms_leader",
+        manual_control=False,  # Enable torque control for gravity compensation
+    )
+    
+    # Initialize robot and teleoperator
+    print("\nInitializing devices...")
+    follower = OpenArmsFollower(follower_config)
+    leader = OpenArmsLeader(leader_config)
+    
+    # Connect devices
+    print("Connecting and calibrating...")
+    follower.connect(calibrate=True)
+    leader.connect(calibrate=True)
+    
+    # Verify URDF is loaded for gravity compensation
+    if leader.pin_robot is None:
+        raise RuntimeError("URDF model not loaded on leader. Gravity compensation not available.")
+    
+    # Configure the dataset features
+    # For actions, we only want to record positions (not velocity or torque)
+    action_features_hw = {}
+    for key, value in follower.action_features.items():
+        if key.endswith(".pos"):
+            action_features_hw[key] = value
+    
+    action_features = hw_to_dataset_features(action_features_hw, "action")
+    obs_features = hw_to_dataset_features(follower.observation_features, "observation")
+    dataset_features = {**action_features, **obs_features}
+    
+    # Create the dataset
+    print("\nCreating dataset...")
+    repo_id = "<hf_username>/<dataset_repo_id>"  # TODO: Replace with your Hugging Face repo
+    
+    # Check if dataset already exists and prompt user
+    dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / repo_id
+    while dataset_path.exists():
+        print(f"\nDataset already exists at: {dataset_path}")
+        print("\nOptions:")
+        print("  1. Overwrite existing dataset")
+        print("  2. Use a different name")
+        print("  3. Abort")
+        
+        choice = input("\nEnter your choice (1/2/3): ").strip()
+        
+        if choice == '1':
+            print(f"Removing existing dataset...")
+            shutil.rmtree(dataset_path)
+            print("✓ Existing dataset removed")
+            break
+        elif choice == '2':
+            print("\nCurrent repo_id:", repo_id)
+            new_repo_id = input("Enter new repo_id (format: <username>/<dataset_name>): ").strip()
+            if new_repo_id and '/' in new_repo_id:
+                repo_id = new_repo_id
+                dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / repo_id
+                print(f"✓ Using new repo_id: {repo_id}")
+                # Loop will continue if this new path also exists
+            else:
+                print("Invalid repo_id format. Please use format: <username>/<dataset_name>")
+        elif choice == '3':
+            print("Aborting. Please remove the existing dataset manually or restart with a different repo_id.")
+            follower.disconnect()
+            leader.disconnect()
+            return
+        else:
+            print("Invalid choice. Please enter 1, 2, or 3.")
+    
+    dataset = LeRobotDataset.create(
+        repo_id=repo_id,
+        fps=FPS,
+        features=dataset_features,
+        robot_type=follower.name,
+        use_videos=True,
+        image_writer_threads=4,
+    )
+    
+    # Initialize keyboard listener and visualization
+    _, events = init_keyboard_listener()
+    init_rerun(session_name="openarms_recording")
+    
+    # Enable motors on both leader arms for gravity compensation
+    leader.bus_right.enable_torque()
+    leader.bus_left.enable_torque()
+    time.sleep(0.1)
+    
+    print("\n" + "=" * 70)
+    print(f"Recording {NUM_EPISODES} episodes")
+    print(f"Task: {TASK_DESCRIPTION}")
+    print("=" * 70)
+    print("\nLeader BOTH arms: Gravity + Friction comp | Follower BOTH arms: Teleop")
+    print("\nKeyboard controls:")
+    print("  - Press 'q' to stop recording")
+    print("  - Press 'r' to re-record current episode")
+    print("=" * 70)
+    
+    episode_idx = 0
+    
+    try:
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+            
+            # Record episode with compensation active
+            record_loop_with_compensation(
+                robot=follower,
+                leader=leader,
+                events=events,
+                fps=FPS,
+                dataset=dataset,
+                dataset_features=dataset_features,
+                control_time_s=EPISODE_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+            )
+            
+            # Reset the environment if not stopping or re-recording
+            if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+                log_say("Reset the environment")
+                record_loop_with_compensation(
+                    robot=follower,
+                    leader=leader,
+                    events=events,
+                    fps=FPS,
+                    dataset=None,  # Don't save reset period
+                    dataset_features=dataset_features,
+                    control_time_s=RESET_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                )
+            
+            # Handle re-recording
+            if events["rerecord_episode"]:
+                log_say("Re-recording episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
+            
+            # Only save episode if frames were recorded
+            if dataset.episode_buffer is not None and dataset.episode_buffer["size"] > 0:
+                dataset.save_episode()
+                episode_idx += 1
+            else:
+                log_say("No frames recorded, skipping episode save")
+                # Clear the empty buffer
+                dataset.episode_buffer = None
+    
+    except KeyboardInterrupt:
+        print("\n\nStopping recording...")
+    
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        try:
+            leader.bus_right.disable_torque()
+            leader.bus_left.disable_torque()
+            time.sleep(0.1)
+            leader.disconnect()
+            follower.disconnect()
+            print("✓ Shutdown complete")
+        except Exception as e:
+            print(f"Shutdown error: {e}")
+        
+        # Upload dataset
+        print("\nUploading dataset to Hugging Face Hub...")
+        try:
+            dataset.push_to_hub()
+            print("✓ Dataset uploaded successfully")
+        except Exception as e:
+            print(f"Warning: Failed to upload dataset: {e}")
+            print("You can manually upload later using: dataset.push_to_hub()")
+        
+        print("✓ Recording complete!")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,166 @@
+#!/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.
+
+"""
+OpenArms Dataset Replay Example
+
+Replays position actions from a recorded dataset on an OpenArms follower robot.
+Only position commands (ending with .pos) are replayed, not velocity or torque.
+
+Example usage:
+    python examples/openarms/replay.py
+"""
+
+import time
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.utils.constants import ACTION
+from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.utils import log_say
+
+# Configuration
+EPISODE_IDX = 0
+DATASET_REPO_ID = "lerobot-data-collection/replay-this-2025-11-02-17-58"  # TODO: Replace with your dataset
+DATASET_ROOT = None  # Use default cache location, or specify custom path
+
+# Robot configuration - adjust these to match your setup
+ROBOT_CONFIG = OpenArmsFollowerConfig(
+    port_left="can2",      # CAN interface for left arm
+    port_right="can3",     # CAN interface for right arm
+    can_interface="socketcan", 
+    id="openarms_follower",
+    disable_torque_on_disconnect=True,
+    max_relative_target=10.0,  # Safety limit: max degrees to move per step
+)
+
+
+def main():
+    """Main replay function."""
+    print("=" * 70)
+    print("OpenArms Dataset Replay")
+    print("=" * 70)
+    print(f"\nDataset: {DATASET_REPO_ID}")
+    print(f"Episode: {EPISODE_IDX}")
+    print(f"Robot: {ROBOT_CONFIG.id}")
+    print(f"  Left arm: {ROBOT_CONFIG.port_left}")
+    print(f"  Right arm: {ROBOT_CONFIG.port_right}")
+    print("\n" + "=" * 70)
+    
+    # Initialize the robot
+    print("\n[1/3] Initializing robot...")
+    robot = OpenArmsFollower(ROBOT_CONFIG)
+    
+    # Load the dataset
+    print(f"\n[2/3] Loading dataset '{DATASET_REPO_ID}'...")
+    dataset = LeRobotDataset(
+        DATASET_REPO_ID,
+        root=DATASET_ROOT,
+        episodes=[EPISODE_IDX]
+    )
+    
+    # Filter dataset to only include frames from the specified episode
+    # (required for dataset V3.0 where episodes are chunked)
+    episode_frames = dataset.hf_dataset.filter(
+        lambda x: x["episode_index"] == EPISODE_IDX
+    )
+    
+    if len(episode_frames) == 0:
+        raise ValueError(
+            f"No frames found for episode {EPISODE_IDX} in dataset {DATASET_REPO_ID}"
+        )
+    
+    print(f"  Found {len(episode_frames)} frames in episode {EPISODE_IDX}")
+    
+    # Extract action features from dataset
+    action_features = dataset.features.get(ACTION, {})
+    action_names = action_features.get("names", [])
+    
+    # Filter to only position actions (ending with .pos)
+    position_action_names = [name for name in action_names if name.endswith(".pos")]
+    
+    if not position_action_names:
+        raise ValueError(
+            f"No position actions found in dataset. Action names: {action_names}"
+        )
+    
+    print(f"  Found {len(position_action_names)} position actions to replay")
+    print(f"  Actions: {', '.join(position_action_names[:5])}{'...' if len(position_action_names) > 5 else ''}")
+    
+    # Select only action columns from dataset
+    actions = episode_frames.select_columns(ACTION)
+    
+    # Connect to the robot
+    print(f"\n[3/3] Connecting to robot...")
+    robot.connect(calibrate=False)  # Skip calibration for replay
+    
+    if not robot.is_connected:
+        raise RuntimeError("Robot failed to connect!")
+    
+    print("\n" + "=" * 70)
+    print("Ready to replay!")
+    print("=" * 70)
+    print("\nThe robot will replay the recorded positions.")
+    print("Press Ctrl+C to stop at any time.\n")
+    
+    input("Press ENTER to start replaying...")
+    
+    # Replay loop
+    log_say(f"Replaying episode {EPISODE_IDX}", blocking=True)
+    
+    try:
+        for idx in range(len(episode_frames)):
+            loop_start = time.perf_counter()
+            
+            # Extract action array from dataset
+            action_array = actions[idx][ACTION]
+            
+            # Build action dictionary, but only include position actions
+            action = {}
+            for i, name in enumerate(action_names):
+                # Only include position actions (ending with .pos)
+                if name.endswith(".pos"):
+                    action[name] = float(action_array[i])
+            
+            # Send action to robot
+            robot.send_action(action)
+            
+            # Maintain replay rate (use dataset fps)
+            loop_duration = time.perf_counter() - loop_start
+            dt_s = 1.0 / dataset.fps - loop_duration
+            busy_wait(dt_s)
+            
+            # Progress indicator every 100 frames
+            if (idx + 1) % 100 == 0:
+                progress = (idx + 1) / len(episode_frames) * 100
+                print(f"Progress: {idx + 1}/{len(episode_frames)} frames ({progress:.1f}%)")
+        
+        print(f"\n✓ Successfully replayed {len(episode_frames)} frames")
+        log_say("Replay complete", blocking=True)
+        
+    except KeyboardInterrupt:
+        print("\n\nReplay interrupted by user")
+    finally:
+        # Disconnect robot
+        print("\nDisconnecting robot...")
+        robot.disconnect()
+        print("✓ Replay complete!")
+
+
+if __name__ == "__main__":
+    main()
+
@@ -0,0 +1,73 @@
+#!/bin/bash
+# Setup all OpenArms CAN interfaces with CAN FD
+
+set -e
+
+echo "=========================================="
+echo "OpenArms CAN FD Interface Setup"
+echo "=========================================="
+echo ""
+echo "Mode: CAN FD"
+echo "  - Nominal bitrate: 1 Mbps"
+echo "  - Data bitrate: 5 Mbps"
+echo ""
+echo "Configuring interfaces can0, can1, can2, can3..."
+echo ""
+
+# Configure each CAN interface with CAN FD
+for i in 0 1 2 3; do
+    interface="can$i"
+    
+    # Check if interface exists
+    if ! ip link show "$interface" &> /dev/null; then
+        echo "⚠ $interface: Not found, skipping"
+        continue
+    fi
+    
+    # Bring down interface
+    sudo ip link set "$interface" down 2>/dev/null
+    
+    # Configure CAN FD mode
+    sudo ip link set "$interface" type can \
+        bitrate 1000000 \
+        dbitrate 5000000 \
+        fd on
+    
+    # Bring up interface
+    sudo ip link set "$interface" up
+    
+    # Verify configuration
+    if ip link show "$interface" | grep -q "UP"; then
+        echo "✓ $interface: Configured and UP"
+    else
+        echo "✗ $interface: Failed to bring UP"
+    fi
+done
+
+echo ""
+echo "=========================================="
+echo "Verification"
+echo "=========================================="
+echo ""
+
+# Show detailed status for each interface
+for i in 0 1 2 3; do
+    interface="can$i"
+    if ip link show "$interface" &> /dev/null; then
+        echo "$interface:"
+        # Show key parameters
+        ip -d link show "$interface" | grep -E "can|state|bitrate|dbitrate" | head -3
+        echo ""
+    fi
+done
+
+echo "=========================================="
+echo "Setup Complete!"
+echo "=========================================="
+echo ""
+echo "All interfaces configured for CAN FD mode"
+echo ""
+echo "Next steps:"
+echo "  1. Test motors: python debug_can_communication.py"
+echo "  2. Run teleoperation: python examples/openarms/teleop.py"
+echo ""
@@ -0,0 +1,148 @@
+"""
+OpenArms Teleoperation Example - Full Dual Arms
+
+This script demonstrates teleoperation of OpenArms follower robot using an OpenArms leader arm.
+It first calibrates both devices, then enters a teleoperation loop for both arms.
+"""
+
+import time
+
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+
+
+follower_config = OpenArmsFollowerConfig(
+    port_left="can2",   # CAN interface for follower left arm
+    port_right="can3",  # CAN interface for follower right arm
+    can_interface="socketcan",  # Linux SocketCAN
+    id="openarms_follower",
+    disable_torque_on_disconnect=True,
+    max_relative_target=5.0,  # Safety limit
+)
+
+
+leader_config = OpenArmsLeaderConfig(
+    port_left="can0",   # CAN interface for leader left arm
+    port_right="can1",  # CAN interface for leader right arm
+    can_interface="socketcan",  # Linux SocketCAN
+    id="openarms_leader",
+    manual_control=True,  # Enable manual control (torque disabled)
+)
+
+print("=" * 60)
+print("OpenArms Teleoperation - Full Dual Arms")
+print("=" * 60)
+
+# Initialize devices
+print("\n[1/4] Initializing devices...")
+follower = OpenArmsFollower(follower_config)
+leader = OpenArmsLeader(leader_config)
+
+# Connect and calibrate follower
+print("\n[2/4] Connecting and calibrating follower robot...")
+print("Note: If you have existing calibration, just press ENTER to use it.")
+follower.connect(calibrate=True)
+
+# Connect and calibrate leader
+print("\n[3/4] Connecting and calibrating leader arm...")
+print("Note: The leader arm will have torque disabled for manual control.")
+leader.connect(calibrate=True)
+
+# Wait for user to be ready
+print("\n[4/4] Ready for teleoperation!")
+print("\nBoth arms will be controlled (16 motors total):")
+print("  RIGHT ARM: joints 1-7 + gripper")
+print("  LEFT ARM: joints 1-7 + gripper")
+
+print("\nPress ENTER to start teleoperation...")
+input()
+
+print("\nTeleoperation started! Move both leader arms.")
+print("Press Ctrl+C to stop.\n")
+
+# All joints for both arms (16 motors total)
+all_joints = [
+    # Right arm
+    "right_joint_1",
+    "right_joint_2",
+    "right_joint_3",
+    "right_joint_4",
+    "right_joint_5",
+    "right_joint_6",
+    "right_joint_7",
+    "right_gripper",
+    # Left arm
+    "left_joint_1",
+    "left_joint_2",
+    "left_joint_3",
+    "left_joint_4",
+    "left_joint_5",
+    "left_joint_6",
+    "left_joint_7",
+    "left_gripper",
+]
+
+# Performance monitoring
+loop_times = []
+start_time = time.perf_counter()
+last_print_time = start_time
+
+try:
+    while True:
+        loop_start = time.perf_counter()
+        
+        # Get action from leader
+        leader_action = leader.get_action()
+        
+        # Filter to only position data for all joints (both arms)
+        joint_action = {}
+        for joint in all_joints:
+            pos_key = f"{joint}.pos"
+            if pos_key in leader_action:
+                joint_action[pos_key] = leader_action[pos_key]
+        
+        # Send action to follower (both arms)
+        if joint_action:
+            follower.send_action(joint_action)
+        
+        # Measure loop time
+        loop_end = time.perf_counter()
+        loop_time = loop_end - loop_start
+        loop_times.append(loop_time)
+        
+        # Print stats every 2 seconds
+        if loop_end - last_print_time >= 2.0:
+            if loop_times:
+                avg_time = sum(loop_times) / len(loop_times)
+                current_hz = 1.0 / avg_time if avg_time > 0 else 0
+                min_time = min(loop_times)
+                max_time = max(loop_times)
+                max_hz = 1.0 / min_time if min_time > 0 else 0
+                min_hz = 1.0 / max_time if max_time > 0 else 0
+                
+                print(f"[Hz Stats] Avg: {current_hz:.1f} Hz | "
+                      f"Range: {min_hz:.1f}-{max_hz:.1f} Hz | "
+                      f"Avg loop time: {avg_time*1000:.1f} ms")
+                
+                # Reset for next measurement window
+                loop_times = []
+                last_print_time = loop_end
+            
+except KeyboardInterrupt:
+    print("\n\nStopping teleoperation...")
+finally:
+    # Disconnect devices
+    print("Disconnecting devices...")
+    try:
+        follower.disconnect()
+    except Exception as e:
+        print(f"Error disconnecting follower: {e}")
+    
+    try:
+        leader.disconnect()
+    except Exception as e:
+        print(f"Error disconnecting leader: {e}")
+    
+    print("Done!")
@@ -0,0 +1,197 @@
+"""
+OpenArms Mini Teleoperation Example
+
+This script demonstrates teleoperation of an OpenArms follower robot using 
+an OpenArms Mini leader (Feetech-based) with dual arms (16 motors total).
+
+The OpenArms Mini has:
+- Right arm: 8 motors (joint_1 to joint_7 + gripper)
+- Left arm: 8 motors (joint_1 to joint_7 + gripper)
+
+Note on gripper normalization:
+- OpenArms Mini gripper: 0-100 scale (0=closed, 100=open)
+- OpenArms follower gripper: degrees (0=closed, -65=open)
+- This script automatically converts between the two ranges
+"""
+
+import time
+import os
+import sys
+
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.teleoperators.openarms_mini.openarms_mini import OpenArmsMini
+from lerobot.teleoperators.openarms_mini.config_openarms_mini import OpenArmsMiniConfig
+from lerobot.utils.robot_utils import busy_wait
+
+# Target control frequency
+TARGET_FPS = 30
+
+# Configure the OpenArms follower (Damiao motors on CAN bus)
+follower_config = OpenArmsFollowerConfig(
+    port_left="can0",   # CAN interface for follower left arm
+    port_right="can1",  # CAN interface for follower right arm
+    can_interface="socketcan",  # Linux SocketCAN
+    id="openarms_follower",
+    disable_torque_on_disconnect=True,
+    max_relative_target=10.0,  # Safety limit (degrees per step)
+)
+
+# Configure the OpenArms Mini leader (Feetech motors on serial)
+leader_config = OpenArmsMiniConfig(
+    port_right="/dev/ttyACM0",  # Serial port for right arm
+    port_left="/dev/ttyACM1",   # Serial port for left arm
+    id="openarms_mini",
+    use_degrees=True,
+)
+
+print("OpenArms Mini → OpenArms Follower Teleoperation")
+
+# Initialize devices
+follower = OpenArmsFollower(follower_config)
+leader = OpenArmsMini(leader_config)
+
+# Connect and calibrate follower
+print("Note: If you have existing calibration, just press ENTER to use it.")
+follower.connect(calibrate=True)
+
+# Connect and calibrate leader
+print("Note: The leader arms will have torque disabled for manual control.")
+leader.connect(calibrate=True)
+
+print("\nPress ENTER to start teleoperation...")
+input()
+
+print("Press Ctrl+C to stop.\n")
+
+# All joints for both arms (16 motors total)
+all_joints = [
+    # Right arm
+    "right_joint_1",
+    "right_joint_2",
+    "right_joint_3",
+    "right_joint_4",
+    "right_joint_5",
+    "right_joint_6",
+    "right_joint_7",
+    "right_gripper",
+    # Left arm
+    "left_joint_1",
+    "left_joint_2",
+    "left_joint_3",
+    "left_joint_4",
+    "left_joint_5",
+    "left_joint_6",
+    "left_joint_7",
+    "left_gripper",
+]
+
+# Performance monitoring
+loop_times = []
+avg_loop_time = 0.0
+min_loop_time = float('inf')
+max_loop_time = 0.0
+stats_update_interval = 1.0  # Update stats every 1 second
+last_stats_update = time.perf_counter()
+
+
+SWAPPED_JOINTS = {
+    "right_joint_6": "right_joint_7",
+    "right_joint_7": "right_joint_6",
+    "left_joint_6": "left_joint_7",
+    "left_joint_7": "left_joint_6",
+}
+
+try:
+    while True:
+        loop_start = time.perf_counter()
+
+        # Get actions and observations
+        leader_action = leader.get_action()
+        follower_obs = follower.get_observation()
+
+        joint_action = {}
+        for joint in all_joints:
+            leader_key = f"{joint}.pos"
+
+            # Determine which follower joint this leader joint controls
+            follower_joint = SWAPPED_JOINTS.get(joint, joint)
+            follower_key = f"{follower_joint}.pos"
+
+            # Get leader position (default 0 if missing)
+            pos = leader_action.get(leader_key, 0.0)
+
+            # Convert gripper values: Mini uses 0-100, OpenArms uses 0 to -65 degrees
+            if "gripper" in joint:
+                # Map 0-100 (Mini) to 0 to -65 (OpenArms)
+                # 0 (closed) -> 0°, 100 (open) -> -65°
+                pos = (pos / 100.0) * -65.0
+
+            # Store in action dict for follower
+            joint_action[follower_key] = pos
+
+        follower.send_action(joint_action)
+
+        # Loop timing
+        loop_end = time.perf_counter()
+        loop_time = loop_end - loop_start
+        loop_times.append(loop_time)
+
+        # Update stats periodically
+        current_time = time.perf_counter()
+        if current_time - last_stats_update >= stats_update_interval:
+            if loop_times:
+                avg_loop_time = sum(loop_times) / len(loop_times)
+                min_loop_time = min(loop_times)
+                max_loop_time = max(loop_times)
+                loop_times = []
+                last_stats_update = current_time
+
+        # Display everything
+        sys.stdout.write("\033[H\033[J")  # Clear screen
+        
+        # Show timing stats at the top
+        if avg_loop_time > 0:
+            avg_hz = 1.0 / avg_loop_time
+            min_hz = 1.0 / max_loop_time if max_loop_time > 0 else 0
+            max_hz = 1.0 / min_loop_time if min_loop_time > 0 and min_loop_time < float('inf') else 0
+            print(f"[Performance] Target: {TARGET_FPS} Hz | Avg: {avg_hz:.1f} Hz | Range: {min_hz:.1f}-{max_hz:.1f} Hz | Loop: {avg_loop_time*1000:.1f} ms\n")
+        else:
+            print(f"[Performance] Target: {TARGET_FPS} Hz | Measuring...\n")
+
+        # Show joint positions
+        print(f"{'Joint':<20} {'Leader':>15} {'Follower':>15}")
+        print(f"{'':20} {'(0-100/deg)':>15} {'(deg)':>15}")
+        print("-" * 52)
+
+        for joint in all_joints:
+            leader_key = f"{joint}.pos"
+            follower_joint = SWAPPED_JOINTS.get(joint, joint)
+            follower_key = f"{follower_joint}.pos"
+
+            leader_pos = leader_action.get(leader_key, 0.0)
+            follower_pos = follower_obs.get(follower_key, 0.0)
+
+            print(f"{joint:<20} {leader_pos:>15.2f} {follower_pos:>15.2f}")
+
+        # Smart sleep to maintain target FPS
+        dt_s = time.perf_counter() - loop_start
+        busy_wait(max(0, 1.0 / TARGET_FPS - dt_s))
+            
+except KeyboardInterrupt:
+    print("\n\nStopping teleoperation...")
+finally:
+    # Disconnect devices
+    print("Disconnecting devices...")
+    try:
+        follower.disconnect()
+    except Exception as e:
+        print(f"Error disconnecting follower: {e}")
+    
+    try:
+        leader.disconnect()
+    except Exception as e:
+        print(f"Error disconnecting leader: {e}")
+    
+    print("Done!")
+
@@ -0,0 +1,202 @@
+"""
+OpenArms Teleoperation with Gravity + Friction Compensation
+
+Leader arms (both LEFT and RIGHT): Gravity + Friction compensation (weightless, easy to move)
+Follower arms (both LEFT and RIGHT): Mirror leader movements
+
+Uses the URDF file from the lerobot repository.
+"""
+
+import time
+
+import numpy as np
+
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+
+# Friction compensation scale factor (1.0 = full, 0.3 = 30% for stability)
+FRICTION_SCALE = 1.0
+
+def main():
+    """Main teleoperation loop with gravity compensation"""
+
+    print("=" * 70)
+    print("OpenArms Teleoperation with Gravity Compensation")
+    print("=" * 70)
+
+    # Configuration
+    follower_config = OpenArmsFollowerConfig(
+        port_left="can2",
+        port_right="can3",
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+    )
+
+    leader_config = OpenArmsLeaderConfig(
+        port_left="can0",
+        port_right="can1",
+        can_interface="socketcan",
+        id="openarms_leader",
+        manual_control=False,  # Enable torque control for gravity compensation
+    )
+
+    # Initialize and connect
+    print("\nInitializing devices...")
+    follower = OpenArmsFollower(follower_config)
+    leader = OpenArmsLeader(leader_config)
+
+    follower.connect()
+    leader.connect()
+
+    # URDF is automatically loaded in the leader constructor
+    if leader.pin_robot is None:
+        raise RuntimeError("URDF model not loaded on leader. Gravity compensation not available.")
+
+    print("\nLeader BOTH arms: Gravity + Friction comp | Follower BOTH arms: Teleop")
+    print("Press ENTER to start...")
+    input()
+
+    # Enable motors on both leader arms for gravity compensation
+    leader.bus_right.enable_torque()
+    leader.bus_left.enable_torque()
+    time.sleep(0.1)
+
+    print("Press Ctrl+C to stop\n")
+
+    # Main control loop
+    loop_times = []
+    last_print_time = time.perf_counter()
+
+    # All joints (both arms)
+    all_joints = []
+    for motor in leader.bus_right.motors:
+        all_joints.append(f"right_{motor}")
+    for motor in leader.bus_left.motors:
+        all_joints.append(f"left_{motor}")
+
+    try:
+        while True:
+            loop_start = time.perf_counter()
+
+            # Get leader state
+            leader_action = leader.get_action()
+
+            # Extract positions and velocities in degrees
+            leader_positions_deg = {}
+            leader_velocities_deg_per_sec = {}
+            
+            for motor in leader.bus_right.motors:
+                pos_key = f"right_{motor}.pos"
+                vel_key = f"right_{motor}.vel"
+                if pos_key in leader_action:
+                    leader_positions_deg[f"right_{motor}"] = leader_action[pos_key]
+                if vel_key in leader_action:
+                    leader_velocities_deg_per_sec[f"right_{motor}"] = leader_action[vel_key]
+
+            for motor in leader.bus_left.motors:
+                pos_key = f"left_{motor}.pos"
+                vel_key = f"left_{motor}.vel"
+                if pos_key in leader_action:
+                    leader_positions_deg[f"left_{motor}"] = leader_action[pos_key]
+                if vel_key in leader_action:
+                    leader_velocities_deg_per_sec[f"left_{motor}"] = leader_action[vel_key]
+
+            # Calculate gravity torques for leader using built-in method
+            leader_positions_rad = {k: np.deg2rad(v) for k, v in leader_positions_deg.items()}
+            leader_gravity_torques_nm = leader._gravity_from_q(leader_positions_rad)
+            
+            # Calculate friction torques for leader using built-in method
+            leader_velocities_rad_per_sec = {k: np.deg2rad(v) for k, v in leader_velocities_deg_per_sec.items()}
+            leader_friction_torques_nm = leader._friction_from_velocity(
+                leader_velocities_rad_per_sec,
+                friction_scale=FRICTION_SCALE
+            )
+            
+            # Combine gravity + friction torques 
+            leader_total_torques_nm = {}
+            for motor_name in leader_gravity_torques_nm:
+                gravity = leader_gravity_torques_nm.get(motor_name, 0.0)
+                friction = leader_friction_torques_nm.get(motor_name, 0.0)
+                leader_total_torques_nm[motor_name] = gravity + friction
+
+            # Apply gravity + friction compensation to leader RIGHT arm (all joints including gripper)
+            for motor in leader.bus_right.motors:
+                full_name = f"right_{motor}"
+                position = leader_positions_deg.get(full_name, 0.0)
+                torque = leader_total_torques_nm.get(full_name, 0.0)
+                
+                # Get damping gain for stability
+                kd = leader.get_damping_kd(motor)
+
+                leader.bus_right._mit_control(
+                    motor=motor,
+                    kp=0.0,
+                    kd=kd,  # Add damping for stability
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque,
+                )
+
+            # Apply gravity + friction compensation to leader LEFT arm (all joints including gripper)
+            for motor in leader.bus_left.motors:
+                full_name = f"left_{motor}"
+                position = leader_positions_deg.get(full_name, 0.0)
+                torque = leader_total_torques_nm.get(full_name, 0.0)
+                
+                # Get damping gain for stability
+                kd = leader.get_damping_kd(motor)
+
+                leader.bus_left._mit_control(                    
+                    motor=motor,
+                    kp=0.0,
+                    kd=kd,  # Add damping for stability
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque,
+                )
+
+            # Send leader positions to follower (both arms)
+            follower_action = {}
+            for joint in all_joints:
+                pos_key = f"{joint}.pos"
+                if pos_key in leader_action:
+                    follower_action[pos_key] = leader_action[pos_key]
+
+            if follower_action:
+                follower.send_action(follower_action)
+
+            # Performance monitoring
+            loop_end = time.perf_counter()
+            loop_time = loop_end - loop_start
+            loop_times.append(loop_time)
+
+            if loop_end - last_print_time >= 2.0:
+                if loop_times:
+                    avg_time = sum(loop_times) / len(loop_times)
+                    current_hz = 1.0 / avg_time if avg_time > 0 else 0
+
+                    print(f"{current_hz:.1f} Hz ({avg_time*1000:.1f} ms)")
+
+                    loop_times = []
+                    last_print_time = loop_end
+
+    except KeyboardInterrupt:
+        print("\n\nStopping...")
+    finally:
+        try:
+            leader.bus_right.disable_torque()
+            leader.bus_left.disable_torque()
+            time.sleep(0.1)
+            leader.disconnect()
+            follower.disconnect()
+            print("✓ Shutdown complete")
+        except Exception as e:
+            print(f"Shutdown error: {e}")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,745 @@
+body {
+  margin: 0;
+  padding: 0;
+  font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen, Ubuntu, sans-serif;
+  background: #f5f5f5;
+}
+
+main {
+  min-height: 100vh;
+  padding: 2rem;
+}
+
+header {
+  text-align: center;
+  margin-bottom: 2rem;
+}
+
+h1 {
+  font-size: 2rem;
+  font-weight: 600;
+  color: #333;
+  margin: 0;
+}
+
+h2 {
+  font-size: 1.25rem;
+  font-weight: 600;
+  color: #333;
+  margin: 0 0 1rem 0;
+}
+
+h3 {
+  font-size: 0.875rem;
+  font-weight: 600;
+  color: #666;
+  margin: 0 0 0.5rem 0;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+}
+
+.container {
+  max-width: 1920px;
+  margin: 0 auto;
+  display: grid;
+  grid-template-columns: minmax(500px, 600px) 1fr;
+  gap: 2rem;
+  align-items: start;
+}
+
+/* Left column container */
+.left-column {
+  display: flex;
+  flex-direction: column;
+  gap: 1.5rem;
+}
+
+/* Right column container */
+.right-column {
+  display: flex;
+  flex-direction: column;
+  gap: 1.5rem;
+}
+
+/* Responsive: Stack on smaller screens */
+@media (max-width: 1200px) {
+  .container {
+    grid-template-columns: 1fr;
+  }
+}
+
+.panel {
+  background: white;
+  border-radius: 8px;
+  padding: 1.5rem;
+  box-shadow: 0 1px 3px rgba(0,0,0,0.1);
+}
+
+.config-panel {
+  border: 2px solid #e5e7eb;
+}
+
+.config-header {
+  display: flex;
+  justify-content: space-between;
+  align-items: center;
+  cursor: pointer;
+  user-select: none;
+  padding: 0.5rem 0;
+}
+
+.config-header:hover {
+  opacity: 0.7;
+}
+
+.toggle-icon {
+  font-size: 1rem;
+  color: #6b7280;
+  transition: transform 0.2s;
+}
+
+.config-content {
+  margin-top: 1rem;
+  padding-top: 1rem;
+  border-top: 1px solid #e5e7eb;
+}
+
+.robot-setup {
+  margin-bottom: 0.5rem;
+}
+
+.robot-status {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  padding: 1rem;
+  border-radius: 6px;
+  font-weight: 500;
+  gap: 1rem;
+}
+
+.robot-status.ready {
+  background: linear-gradient(135deg, #d1fae5 0%, #a7f3d0 100%);
+  color: #065f46;
+  border: 1px solid #10b981;
+}
+
+.robot-status.not-ready {
+  background: linear-gradient(135deg, #fef3c7 0%, #fde68a 100%);
+  color: #92400e;
+  border: 1px solid #f59e0b;
+}
+
+.btn-setup {
+  background: #10b981;
+  color: white;
+  border: none;
+  padding: 0.5rem 1rem;
+  border-radius: 4px;
+  font-size: 0.875rem;
+  font-weight: 500;
+  cursor: pointer;
+  transition: background 0.2s;
+}
+
+.btn-setup:hover:not(:disabled) {
+  background: #059669;
+}
+
+.btn-setup:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+}
+
+.btn-zero {
+  background: #8b5cf6;
+  color: white;
+  border: none;
+  padding: 0.5rem 1rem;
+  border-radius: 4px;
+  font-size: 0.875rem;
+  font-weight: 500;
+  cursor: pointer;
+  transition: background 0.2s;
+}
+
+.btn-zero:hover:not(:disabled) {
+  background: #7c3aed;
+}
+
+.btn-zero:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+}
+
+.zero-position-section {
+  margin-top: 1rem;
+  padding-top: 1rem;
+  border-top: 1px solid #e5e7eb;
+}
+
+.btn-zero-large {
+  width: 100%;
+  background: #8b5cf6;
+  color: white;
+  border: none;
+  padding: 0.875rem 1.5rem;
+  border-radius: 8px;
+  font-size: 1rem;
+  font-weight: 600;
+  cursor: pointer;
+  transition: all 0.2s;
+  box-shadow: 0 2px 4px rgba(139, 92, 246, 0.2);
+}
+
+.btn-zero-large:hover:not(:disabled) {
+  background: #7c3aed;
+  box-shadow: 0 4px 8px rgba(139, 92, 246, 0.3);
+  transform: translateY(-1px);
+}
+
+.btn-zero-large:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+  box-shadow: none;
+  transform: none;
+}
+
+.delete-episode-section {
+  margin-top: 1rem;
+  padding-top: 1rem;
+  border-top: 1px solid #e5e7eb;
+}
+
+.btn-delete {
+  width: 100%;
+  background: #ef4444;
+  color: white;
+  border: none;
+  padding: 0.875rem 1.5rem;
+  border-radius: 8px;
+  font-size: 1rem;
+  font-weight: 600;
+  cursor: pointer;
+  transition: all 0.2s;
+  box-shadow: 0 2px 4px rgba(239, 68, 68, 0.2);
+}
+
+.btn-delete:hover:not(:disabled) {
+  background: #dc2626;
+  box-shadow: 0 4px 8px rgba(239, 68, 68, 0.3);
+  transform: translateY(-1px);
+}
+
+.btn-delete:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+  box-shadow: none;
+  transform: none;
+}
+
+.delete-info {
+  margin-top: 0.5rem;
+  font-size: 0.875rem;
+  color: #666;
+  text-align: center;
+  font-style: italic;
+}
+
+.btn-disconnect {
+  background: #ef4444;
+  color: white;
+  border: none;
+  padding: 0.5rem 1rem;
+  border-radius: 4px;
+  font-size: 0.875rem;
+  font-weight: 500;
+  cursor: pointer;
+  transition: background 0.2s;
+}
+
+.btn-disconnect:hover {
+  background: #dc2626;
+}
+
+.btn-refresh {
+  background: #3b82f6;
+  color: white;
+  border: none;
+  padding: 0.4rem 0.8rem;
+  border-radius: 4px;
+  font-size: 0.75rem;
+  font-weight: 500;
+  cursor: pointer;
+  transition: background 0.2s;
+}
+
+.btn-refresh:hover:not(:disabled) {
+  background: #2563eb;
+}
+
+.btn-refresh:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+}
+
+.control-panel {
+  border: 2px solid #10b981;
+}
+
+.status-banner {
+  display: flex;
+  align-items: center;
+  gap: 1rem;
+  padding: 1rem 1.5rem;
+  border-radius: 6px;
+  margin-bottom: 1.5rem;
+  font-weight: 500;
+  font-size: 0.95rem;
+}
+
+.status-banner.initializing {
+  background: linear-gradient(135deg, #dbeafe 0%, #bfdbfe 100%);
+  color: #1e40af;
+  border-left: 4px solid #3b82f6;
+}
+
+.status-banner.encoding {
+  background: linear-gradient(135deg, #fef3c7 0%, #fde68a 100%);
+  color: #92400e;
+  border-left: 4px solid #f59e0b;
+}
+
+.status-banner.uploading {
+  background: linear-gradient(135deg, #e0e7ff 0%, #c7d2fe 100%);
+  color: #3730a3;
+  border-left: 4px solid #6366f1;
+}
+
+.status-banner.success {
+  background: linear-gradient(135deg, #d1fae5 0%, #a7f3d0 100%);
+  color: #065f46;
+  border-left: 4px solid #10b981;
+}
+
+.status-banner.warning {
+  background: linear-gradient(135deg, #fee2e2 0%, #fecaca 100%);
+  color: #991b1b;
+  border-left: 4px solid #ef4444;
+}
+
+.spinner {
+  width: 20px;
+  height: 20px;
+  border: 3px solid rgba(0, 0, 0, 0.1);
+  border-top-color: currentColor;
+  border-radius: 50%;
+  animation: spin 0.8s linear infinite;
+}
+
+@keyframes spin {
+  to { transform: rotate(360deg); }
+}
+
+.control-horizontal {
+  display: flex;
+  flex-direction: column;
+  gap: 1.5rem;
+}
+
+.control-left {
+  display: flex;
+  flex-direction: column;
+  gap: 1rem;
+}
+
+.control-right {
+  display: flex;
+  align-items: center;
+  justify-content: center;
+}
+
+.input-group {
+  display: flex;
+  gap: 0.5rem;
+  margin-bottom: 0;
+}
+
+input[type="text"] {
+  flex: 1;
+  padding: 0.75rem;
+  border: 1px solid #ddd;
+  border-radius: 4px;
+  font-size: 1rem;
+}
+
+input[type="text"]:disabled {
+  background: #f5f5f5;
+  cursor: not-allowed;
+}
+
+input[type="text"]:focus {
+  outline: none;
+  border-color: #10b981;
+}
+
+button {
+  padding: 0.75rem 1.5rem;
+  border: none;
+  border-radius: 4px;
+  font-size: 1rem;
+  font-weight: 500;
+  cursor: pointer;
+  transition: all 0.2s;
+}
+
+.btn-set-task {
+  background: #3b82f6;
+  color: white;
+  min-width: 120px;
+}
+
+.btn-set-task:hover:not(:disabled) {
+  background: #2563eb;
+}
+
+.btn-set-task:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+}
+
+.btn-start {
+  background: #10b981;
+  color: white;
+}
+
+.btn-start:hover:not(:disabled) {
+  background: #059669;
+}
+
+.btn-start:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+}
+
+.btn-stop {
+  background: #ef4444;
+  color: white;
+}
+
+.btn-stop:hover {
+  background: #dc2626;
+}
+
+.btn-reset {
+  padding: 0.5rem 1rem;
+  background: #6b7280;
+  color: white;
+  font-size: 0.875rem;
+}
+
+.btn-reset:hover {
+  background: #4b5563;
+}
+
+.status {
+  display: flex;
+  align-items: center;
+  gap: 0.75rem;
+  padding: 1rem;
+  border-radius: 4px;
+  margin-bottom: 1rem;
+}
+
+.status.recording {
+  background: #fee2e2;
+  color: #991b1b;
+}
+
+.status.recording.recording-active {
+  display: flex;
+  flex-direction: column;
+  gap: 1rem;
+  background: #dc2626;
+  color: white;
+  padding: 1.5rem;
+  border: 4px solid #991b1b;
+  box-shadow: 0 4px 12px rgba(220, 38, 38, 0.4);
+  font-weight: 700;
+  font-size: 1rem;
+}
+
+.status.recording.recording-active .indicator {
+  width: 20px;
+  height: 20px;
+  background: #fef2f2;
+  animation: pulse-strong 1s ease-in-out infinite;
+}
+
+@keyframes pulse-strong {
+  0%, 100% { 
+    opacity: 1;
+    transform: scale(1);
+  }
+  50% { 
+    opacity: 0.7;
+    transform: scale(1.1);
+  }
+}
+
+.status.recording.recording-active .time-display {
+  display: flex;
+  flex-direction: column;
+  gap: 0.5rem;
+  font-size: 1.5rem;
+  font-weight: 700;
+  color: white;
+}
+
+.fps-display {
+  font-size: 1rem;
+  font-weight: 500;
+  opacity: 0.95;
+}
+
+.fps-warning {
+  color: #fef2f2;
+  animation: pulse-warning 1s ease-in-out infinite;
+}
+
+@keyframes pulse-warning {
+  0%, 100% { opacity: 1; }
+  50% { opacity: 0.5; }
+}
+
+.status.recording.recording-active .btn-stop {
+  align-self: stretch;
+}
+
+.ramp-up-countdown {
+  display: flex;
+  justify-content: center;
+  margin-bottom: 1rem;
+}
+
+.countdown-box {
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  justify-content: center;
+  padding: 2rem 3rem;
+  background: linear-gradient(135deg, #fef3c7 0%, #fde68a 100%);
+  border: 4px solid #f59e0b;
+  border-radius: 16px;
+  box-shadow: 0 6px 20px rgba(245, 158, 11, 0.4);
+  min-width: 280px;
+  animation: pulse-warm 1.5s ease-in-out infinite;
+}
+
+@keyframes pulse-warm {
+  0%, 100% {
+    box-shadow: 0 6px 20px rgba(245, 158, 11, 0.4);
+  }
+  50% {
+    box-shadow: 0 6px 25px rgba(245, 158, 11, 0.6);
+  }
+}
+
+.countdown-label {
+  font-size: 1rem;
+  color: #92400e;
+  text-transform: uppercase;
+  letter-spacing: 1.5px;
+  font-weight: 800;
+  margin-bottom: 1rem;
+  text-align: center;
+}
+
+.countdown-value {
+  font-size: 4.5rem;
+  font-weight: 900;
+  color: #d97706;
+  font-family: 'Courier New', monospace;
+  line-height: 1;
+  text-shadow: 2px 2px 6px rgba(0, 0, 0, 0.15);
+  margin-bottom: 0.5rem;
+}
+
+.countdown-subtitle {
+  font-size: 0.875rem;
+  color: #78350f;
+  font-weight: 600;
+  font-style: italic;
+  text-align: center;
+  margin-top: 0.5rem;
+}
+
+.status.idle {
+  background: #f3f4f6;
+  color: #374151;
+}
+
+.indicator {
+  width: 12px;
+  height: 12px;
+  border-radius: 50%;
+  background: #ef4444;
+  animation: pulse 1.5s ease-in-out infinite;
+}
+
+@keyframes pulse {
+  0%, 100% { opacity: 1; }
+  50% { opacity: 0.5; }
+}
+
+.counter {
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  gap: 0.75rem;
+  padding: 1.5rem;
+  background: linear-gradient(135deg, #f9fafb 0%, #f3f4f6 100%);
+  border-radius: 8px;
+  border: 2px solid #e5e7eb;
+  min-width: 200px;
+}
+
+.counter-label {
+  font-size: 0.75rem;
+  color: #6b7280;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+  font-weight: 600;
+}
+
+.counter-value {
+  font-size: 3rem;
+  font-weight: 700;
+  color: #10b981;
+  line-height: 1;
+}
+
+.time-display {
+  font-size: 1.5rem;
+  font-weight: 600;
+  font-family: 'Courier New', monospace;
+}
+
+.error-box {
+  padding: 1rem;
+  background: #fee2e2;
+  color: #991b1b;
+  border-radius: 4px;
+  border-left: 4px solid #ef4444;
+  font-size: 0.875rem;
+}
+
+.config-section {
+  margin-bottom: 1.5rem;
+}
+
+.config-section:last-child {
+  margin-bottom: 0;
+}
+
+.config-grid {
+  display: grid;
+  grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
+  gap: 1rem;
+}
+
+label {
+  display: flex;
+  flex-direction: column;
+  gap: 0.5rem;
+  font-size: 0.875rem;
+  color: #374151;
+  font-weight: 500;
+}
+
+select {
+  padding: 0.5rem;
+  border: 1px solid #ddd;
+  border-radius: 4px;
+  font-size: 0.875rem;
+  background: white;
+}
+
+select:disabled {
+  background: #f5f5f5;
+  cursor: not-allowed;
+}
+
+/* Camera Layout */
+.camera-layout {
+  display: flex;
+  flex-direction: column;
+  gap: 1.5rem;
+}
+
+.camera-base {
+  width: 100%;
+}
+
+.camera-wrist-container {
+  display: grid;
+  grid-template-columns: repeat(2, 1fr);
+  gap: 1.5rem;
+}
+
+.camera-wrist {
+  width: 100%;
+}
+
+.camera {
+  border: 1px solid #e5e7eb;
+  border-radius: 4px;
+  overflow: hidden;
+}
+
+.camera h3 {
+  padding: 0.75rem;
+  background: #f9fafb;
+  border-bottom: 1px solid #e5e7eb;
+  margin: 0;
+}
+
+.camera img {
+  width: 100%;
+  height: auto;
+  display: block;
+  background: #000;
+  min-height: 300px;
+  object-fit: cover;
+}
+
+.camera-placeholder {
+  text-align: center;
+  padding: 4rem 2rem;
+  background: #f9fafb;
+  border-radius: 4px;
+  border: 2px dashed #d1d5db;
+}
+
+.camera-placeholder p {
+  margin: 0.5rem 0;
+  font-size: 1rem;
+  color: #6b7280;
+}
+
+.camera-placeholder p:first-child {
+  font-size: 1.25rem;
+  font-weight: 500;
+  color: #374151;
+}
+
+.hint {
+  margin-top: 0.5rem;
+  font-size: 0.75rem;
+  color: #6b7280;
+  display: flex;
+  align-items: center;
+  gap: 0.5rem;
+  flex-wrap: wrap;
+}
+
@@ -0,0 +1,857 @@
+import { useState, useEffect, useCallback, useRef } from 'react';
+import './App.css';
+
+const API_BASE = 'http://localhost:8000/api';
+
+function App() {
+  // State
+  const [task, setTask] = useState('');
+  const [isRecording, setIsRecording] = useState(false);
+  const [isInitializing, setIsInitializing] = useState(false);
+  const [isEncoding, setIsEncoding] = useState(false);
+  const [isUploading, setIsUploading] = useState(false);
+  const [robotsReady, setRobotsReady] = useState(false);
+  const [elapsedTime, setElapsedTime] = useState(0);
+  const [currentFps, setCurrentFps] = useState(0);
+  const [loopFps, setLoopFps] = useState(0);
+  const [episodeCount, setEpisodeCount] = useState(0);
+  const [error, setError] = useState(null);
+  const [statusMessage, setStatusMessage] = useState('Ready');
+  const [uploadStatus, setUploadStatus] = useState(null);
+  const [rampUpRemaining, setRampUpRemaining] = useState(0);
+  const [movingToZero, setMovingToZero] = useState(false);
+  const [configExpanded, setConfigExpanded] = useState(false);
+  const [latestRepoId, setLatestRepoId] = useState(null);
+
+  // Configuration
+  const [config, setConfig] = useState({
+    leader_type: 'openarms',  // 'openarms' or 'openarms_mini'
+    leader_left: 'can0',
+    leader_right: 'can1',
+    follower_left: 'can2',
+    follower_right: 'can3',
+    left_wrist: '/dev/video0',
+    right_wrist: '/dev/video1',
+    base: '/dev/video4'
+  });
+
+  // Available options
+  const [availableCameras, setAvailableCameras] = useState([]);
+  const [availableUsbPorts, setAvailableUsbPorts] = useState([]);
+  const canInterfaces = ['can0', 'can1', 'can2', 'can3'];
+
+  const statusIntervalRef = useRef(null);
+  const hasInitializedRef = useRef(false);
+
+  const loadConfig = () => {
+    try {
+      const saved = localStorage.getItem('openarms_config');
+      if (saved) {
+        const loadedConfig = JSON.parse(saved);
+        setConfig(prev => ({ ...prev, ...loadedConfig }));
+      }
+    } catch (e) {
+      console.error('Load config error:', e);
+    }
+  };
+
+  const saveConfig = (newConfig) => {
+    try {
+      localStorage.setItem('openarms_config', JSON.stringify(newConfig || config));
+    } catch (e) {
+      console.error('Save config error:', e);
+    }
+  };
+
+  // Fetch status periodically
+  const fetchStatus = async () => {
+    try {
+      const response = await fetch(`${API_BASE}/status`);
+      const data = await response.json();
+
+      setIsRecording(data.is_recording);
+      setIsInitializing(data.is_initializing);
+      setIsEncoding(data.is_encoding);
+      setIsUploading(data.is_uploading);
+      setRobotsReady(data.robots_ready);
+      setElapsedTime(data.elapsed_time);
+      setCurrentFps(data.current_fps || 0);
+      setLoopFps(data.loop_fps || 0);
+      setEpisodeCount(data.episode_count);
+      setError(data.error);
+      setStatusMessage(data.status_message || 'Ready');
+      setUploadStatus(data.upload_status);
+      setRampUpRemaining(data.ramp_up_remaining || 0);
+      setMovingToZero(data.moving_to_zero || false);
+      
+      // Track the latest repo_id from the backend
+      if (data.latest_repo_id) {
+        setLatestRepoId(data.latest_repo_id);
+      }
+
+      if (data.config) {
+        // Only merge server config if we don't have a saved config (first load)
+        if (!localStorage.getItem('openarms_config')) {
+          setConfig(prev => {
+            const merged = { ...data.config, ...prev };
+            localStorage.setItem('openarms_config', JSON.stringify(merged));
+            return merged;
+          });
+        }
+      }
+    } catch (e) {
+      console.error('Failed to fetch status:', e);
+    }
+  };
+
+  const setupRobots = async () => {
+    // Show warning to verify camera positions
+    const confirmed = window.confirm(
+      '⚠️ IMPORTANT: Before connecting robots, please verify:\n\n' +
+      '📹 Check that cameras are correctly positioned:\n' +
+      '   • LEFT wrist camera is actually on the LEFT arm\n' +
+      '   • RIGHT wrist camera is actually on the RIGHT arm\n' +
+      '   • BASE camera is actually the BASE/overhead camera\n\n' +
+      'Incorrect camera positioning will result in invalid training data!\n\n' +
+      'Click OK to continue with robot setup, or Cancel to review configuration.'
+    );
+    
+    if (!confirmed) {
+      return; // User cancelled, don't proceed
+    }
+    
+    setError(null);
+    try {
+      const response = await fetch(`${API_BASE}/robots/setup`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify(config)
+      });
+
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to setup robots');
+      }
+
+      await response.json();
+      saveConfig(config);
+    } catch (e) {
+      setError(`Robot setup failed: ${e.message}`);
+    }
+  };
+
+  // Disconnect robots
+  const disconnectRobots = async () => {
+    try {
+      await fetch(`${API_BASE}/robots/disconnect`, { method: 'POST' });
+      setRobotsReady(false);
+    } catch (e) {
+      console.error('Failed to disconnect robots:', e);
+    }
+  };
+
+  // Discover cameras
+  const discoverCameras = async () => {
+    try {
+      const response = await fetch(`${API_BASE}/cameras/discover`);
+      const data = await response.json();
+      const cameras = data.cameras || [];
+      setAvailableCameras(cameras);
+
+      // Get list of valid camera IDs
+      const validCameraIds = cameras.map(cam => String(cam.id));
+
+      // Auto-fix config if current values are invalid or not set
+      const updated = { ...config };
+      let changed = false;
+
+      // Auto-fix invalid camera config
+      if (!config.left_wrist || !validCameraIds.includes(config.left_wrist)) {
+        if (cameras.length >= 1) {
+          updated.left_wrist = String(cameras[0].id);
+          changed = true;
+        }
+      }
+
+      if (!config.right_wrist || !validCameraIds.includes(config.right_wrist)) {
+        if (cameras.length >= 2) {
+          updated.right_wrist = String(cameras[1].id);
+          changed = true;
+        }
+      }
+
+      if (!config.base || !validCameraIds.includes(config.base)) {
+        if (cameras.length >= 3) {
+          updated.base = String(cameras[2].id);
+          changed = true;
+        }
+      }
+
+      if (changed) {
+        setConfig(updated);
+        saveConfig(updated);
+      }
+
+      if (cameras.length === 0) {
+        setError('No cameras detected! Please connect cameras and refresh.');
+      }
+    } catch (e) {
+      console.error('Failed to discover cameras:', e);
+      setError(`Camera discovery failed: ${e.message}`);
+    }
+  };
+
+  // Discover USB ports
+  const discoverUsbPorts = async () => {
+    try {
+      const response = await fetch(`${API_BASE}/usb/discover`);
+      const data = await response.json();
+      const ports = data.ports || [];
+      setAvailableUsbPorts(ports);
+
+      // Auto-fix config if OpenArms Mini is selected and ports are invalid
+      if (config.leader_type === 'openarms_mini') {
+        const updated = { ...config };
+        let changed = false;
+
+        if (ports.length >= 1 && !ports.includes(config.leader_left)) {
+          updated.leader_left = ports[0];
+          changed = true;
+        }
+
+        if (ports.length >= 2 && !ports.includes(config.leader_right)) {
+          updated.leader_right = ports[1];
+          changed = true;
+        }
+
+        if (changed) {
+          setConfig(updated);
+          saveConfig(updated);
+        }
+      }
+
+      if (ports.length === 0) {
+        console.warn('No USB ports detected for OpenArms Mini');
+      }
+    } catch (e) {
+      console.error('Failed to discover USB ports:', e);
+    }
+  };
+
+  // Set task only (for pedal use)
+  const setTaskOnly = async () => {
+    if (!task.trim()) {
+      setError('Please enter a task description');
+      return;
+    }
+
+    setError(null);
+
+    try {
+      const response = await fetch(`${API_BASE}/recording/set-task`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ task, ...config })
+      });
+
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to set task');
+      }
+
+      const result = await response.json();
+      setStatusMessage(result.message || `Task set: ${task}`);
+      saveConfig(config);
+      
+      // Clear success message after 3 seconds
+      setTimeout(() => {
+        if (!isRecording && !isInitializing) {
+          setStatusMessage('Ready');
+        }
+      }, 3000);
+    } catch (e) {
+      setError(e.message);
+    }
+  };
+
+  // Start recording
+  const startRecording = async () => {
+    if (!task.trim()) {
+      setError('Please enter a task description');
+      return;
+    }
+
+    setError(null);
+
+    try {
+      const response = await fetch(`${API_BASE}/recording/start`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ task, ...config })
+      });
+
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to start recording');
+      }
+
+      await response.json();
+      saveConfig(config);
+    } catch (e) {
+      setError(e.message);
+    }
+  };
+
+  // Stop recording
+  const stopRecording = async () => {
+    try {
+      const response = await fetch(`${API_BASE}/recording/stop`, {
+        method: 'POST'
+      });
+
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to stop recording');
+      }
+
+      const data = await response.json();
+      setError(null);
+      // Update latest repo_id after recording
+      if (data.dataset_name) {
+        setLatestRepoId(`lerobot-data-collection/${data.dataset_name}`);
+      }
+    } catch (e) {
+      setError(e.message);
+    }
+  };
+
+  const deleteLatestEpisode = async () => {
+    if (!latestRepoId) {
+      setError('No episode to delete');
+      return;
+    }
+    
+    const confirmed = window.confirm(
+      `WARNING: This will permanently delete the repository:\n\n${latestRepoId}\n\nThis action cannot be undone. Continue?`
+    );
+    
+    if (!confirmed) {
+      return;
+    }
+    
+    try {
+      const response = await fetch(`${API_BASE}/recording/delete-latest`, { method: 'POST' });
+      
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to delete episode');
+      }
+
+      const data = await response.json();
+      setLatestRepoId(null);
+      setEpisodeCount(Math.max(0, episodeCount - 1));
+      setStatusMessage(`Deleted: ${data.deleted_repo}`);
+      
+      setTimeout(() => {
+        if (!isRecording && !isInitializing) {
+          setStatusMessage('Ready');
+        }
+      }, 3000);
+    } catch (e) {
+      setError(`Delete failed: ${e.message}`);
+    }
+  };
+
+  // Reset counter
+  const resetCounter = async () => {
+    try {
+      await fetch(`${API_BASE}/counter/reset`, { method: 'POST' });
+      setEpisodeCount(0);
+    } catch (e) {
+      console.error('Failed to reset counter:', e);
+    }
+  };
+
+  // Move robot to zero position
+  const moveToZero = async () => {
+    setError(null);
+    try {
+      const response = await fetch(`${API_BASE}/robots/move-to-zero`, { method: 'POST' });
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to move to zero position');
+      }
+      await response.json();
+    } catch (e) {
+      setError(`Move to zero failed: ${e.message}`);
+    }
+  };
+
+  // Format time as MM:SS
+  const formatTime = (seconds) => {
+    const mins = Math.floor(seconds / 60);
+    const secs = Math.floor(seconds % 60);
+    return `${mins.toString().padStart(2, '0')}:${secs.toString().padStart(2, '0')}`;
+  };
+
+  // Update config and save
+  const updateConfig = (key, value) => {
+    const updated = { ...config, [key]: value };
+    setConfig(updated);
+    saveConfig(updated);
+  };
+
+  // Initialize on mount only
+  useEffect(() => {
+    // Prevent double-initialization in development
+    if (hasInitializedRef.current) {
+      return;
+    }
+    hasInitializedRef.current = true;
+
+    loadConfig();
+    discoverCameras();
+    discoverUsbPorts();
+    fetchStatus();
+    statusIntervalRef.current = setInterval(fetchStatus, 1000);
+
+    return () => {
+      if (statusIntervalRef.current) {
+        clearInterval(statusIntervalRef.current);
+      }
+    };
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, []); // Run only once on mount
+
+  // Discover USB ports when leader type changes to Mini
+  useEffect(() => {
+    if (config.leader_type === 'openarms_mini') {
+      discoverUsbPorts();
+    }
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [config.leader_type]);
+
+  return (
+    <main>
+      <header>
+        <h1>OpenArms Recording</h1>
+      </header>
+
+      <div className="container">
+        {/* Left Column: Configuration and Recording Control */}
+        <div className="left-column">
+          {/* Configuration Panel */}
+          <section className="panel config-panel">
+          <div
+            className="config-header"
+            onClick={() => setConfigExpanded(!configExpanded)}
+            role="button"
+            tabIndex={0}
+            onKeyDown={(e) => e.key === 'Enter' && setConfigExpanded(!configExpanded)}
+          >
+            <h2>⚙️ Configuration</h2>
+            <span className="toggle-icon">{configExpanded ? '▼' : '▶'}</span>
+          </div>
+
+          {configExpanded && (
+            <div className="config-content">
+              {/* Robot Setup */}
+              <div className="config-section">
+                <h3>🤖 Robot Setup</h3>
+                <div className="robot-setup">
+                  {robotsReady ? (
+                    <div className="robot-status ready">
+                      <span>✅ Robots Ready - Recording will start instantly</span>
+                      <button onClick={disconnectRobots} className="btn-disconnect">
+                        Disconnect Robots
+                      </button>
+                    </div>
+                  ) : (
+                    <div className="robot-status not-ready">
+                      <span>⚠️ Robots not initialized - Recording will take ~10 seconds</span>
+                      <button
+                        onClick={setupRobots}
+                        disabled={isRecording || isInitializing}
+                        className="btn-setup"
+                      >
+                        🚀 Setup Robots
+                      </button>
+                    </div>
+                  )}
+                </div>
+              </div>
+
+              {/* Leader Type Selection */}
+              <div className="config-section">
+                <h3>🎮 Leader Type</h3>
+                <div className="config-grid">
+                  <label style={{gridColumn: '1 / -1'}}>
+                    Leader Arm Type
+                    <select
+                      value={config.leader_type}
+                      onChange={(e) => updateConfig('leader_type', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      <option value="openarms">OpenArms (CAN Bus - Damiao Motors)</option>
+                      <option value="openarms_mini">OpenArms Mini (USB - Feetech Motors)</option>
+                    </select>
+                  </label>
+                </div>
+              </div>
+
+              {/* Leader Interfaces (CAN or USB based on type) */}
+              <div className="config-section">
+                <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'center', marginBottom: '0.5rem' }}>
+                  <h3>
+                    {config.leader_type === 'openarms_mini' 
+                      ? `Leader Ports (USB/Serial) ${availableUsbPorts.length > 0 ? `(${availableUsbPorts.length} detected)` : ''}` 
+                      : 'Leader Interfaces (CAN)'}
+                  </h3>
+                  {config.leader_type === 'openarms_mini' && (
+                    <button
+                      onClick={discoverUsbPorts}
+                      className="btn-refresh"
+                      disabled={isRecording || robotsReady}
+                    >
+                      🔄 Refresh
+                    </button>
+                  )}
+                </div>
+                
+                <div className="config-grid">
+                  <label>
+                    Leader Left
+                    <select
+                      value={config.leader_left}
+                      onChange={(e) => updateConfig('leader_left', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {config.leader_type === 'openarms_mini' ? (
+                        availableUsbPorts.length > 0 ? (
+                          availableUsbPorts.map((port) => (
+                            <option key={port} value={port}>{port}</option>
+                          ))
+                        ) : (
+                          <option value="">No USB ports detected</option>
+                        )
+                      ) : (
+                        canInterfaces.map((iface) => (
+                          <option key={iface} value={iface}>{iface}</option>
+                        ))
+                      )}
+                    </select>
+                  </label>
+
+                  <label>
+                    Leader Right
+                    <select
+                      value={config.leader_right}
+                      onChange={(e) => updateConfig('leader_right', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {config.leader_type === 'openarms_mini' ? (
+                        availableUsbPorts.length > 0 ? (
+                          availableUsbPorts.map((port) => (
+                            <option key={port} value={port}>{port}</option>
+                          ))
+                        ) : (
+                          <option value="">No USB ports detected</option>
+                        )
+                      ) : (
+                        canInterfaces.map((iface) => (
+                          <option key={iface} value={iface}>{iface}</option>
+                        ))
+                      )}
+                    </select>
+                  </label>
+                </div>
+              </div>
+
+              {/* Follower CAN Interfaces */}
+              <div className="config-section">
+                <h3>Follower Interfaces (CAN)</h3>
+                
+                <div className="config-grid">
+                  <label>
+                    Follower Left
+                    <select
+                      value={config.follower_left}
+                      onChange={(e) => updateConfig('follower_left', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {canInterfaces.map((iface) => (
+                        <option key={iface} value={iface}>{iface}</option>
+                      ))}
+                    </select>
+                  </label>
+
+                  <label>
+                    Follower Right
+                    <select
+                      value={config.follower_right}
+                      onChange={(e) => updateConfig('follower_right', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {canInterfaces.map((iface) => (
+                        <option key={iface} value={iface}>{iface}</option>
+                      ))}
+                    </select>
+                  </label>
+                </div>
+              </div>
+
+              {/* Camera Configuration */}
+              <div className="config-section">
+                <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'center', marginBottom: '0.5rem' }}>
+                  <h3>Cameras {availableCameras.length > 0 && `(${availableCameras.length} detected)`}</h3>
+                  <button
+                    onClick={discoverCameras}
+                    className="btn-refresh"
+                    disabled={isRecording || robotsReady}
+                  >
+                    🔄 Refresh
+                  </button>
+                </div>
+                <div className="config-grid">
+                  <label>
+                    Left Wrist
+                    <select
+                      value={config.left_wrist}
+                      onChange={(e) => updateConfig('left_wrist', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {availableCameras.map((cam) => (
+                        <option key={cam.id} value={String(cam.id)}>
+                          {cam.name || `Camera @ ${cam.id}`}
+                        </option>
+                      ))}
+                    </select>
+                  </label>
+
+                  <label>
+                    Right Wrist
+                    <select
+                      value={config.right_wrist}
+                      onChange={(e) => updateConfig('right_wrist', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {availableCameras.map((cam) => (
+                        <option key={cam.id} value={String(cam.id)}>
+                          {cam.name || `Camera @ ${cam.id}`}
+                        </option>
+                      ))}
+                    </select>
+                  </label>
+
+                  <label>
+                    Base Camera
+                    <select
+                      value={config.base}
+                      onChange={(e) => updateConfig('base', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {availableCameras.map((cam) => (
+                        <option key={cam.id} value={String(cam.id)}>
+                          {cam.name || `Camera @ ${cam.id}`}
+                        </option>
+                      ))}
+                    </select>
+                  </label>
+                </div>
+              </div>
+            </div>
+          )}
+        </section>
+
+        {/* Control Panel */}
+        <section className="panel control-panel">
+          <h2>🎬 Recording Control</h2>
+
+          {/* Status Banner - Always show important statuses */}
+          {isInitializing && (
+            <div className="status-banner initializing">
+              <div className="spinner"></div>
+              <span>{statusMessage}</span>
+            </div>
+          )}
+
+          {isEncoding && (
+            <div className="status-banner encoding">
+              <div className="spinner"></div>
+              <span>📹 {statusMessage}</span>
+            </div>
+          )}
+
+          {isUploading && (
+            <div className="status-banner uploading">
+              <div className="spinner"></div>
+              <span>☁️ {statusMessage}</span>
+            </div>
+          )}
+
+          {uploadStatus && !isRecording && !isEncoding && !isUploading && (
+            <div className={`status-banner ${uploadStatus.startsWith('✓') ? 'success' : 'warning'}`}>
+              <span>{uploadStatus}</span>
+            </div>
+          )}
+
+          <div className="control-horizontal">
+            {/* Task Input and Status */}
+            <div className="control-left">
+              <div className="input-group">
+                <input
+                  type="text"
+                  value={task}
+                  onChange={(e) => setTask(e.target.value)}
+                  placeholder="Task description (e.g., 'pick and place')"
+                  disabled={isRecording || isInitializing || isEncoding || isUploading}
+                  onKeyPress={(e) => {
+                    if (e.key === 'Enter' && robotsReady) {
+                      setTaskOnly();
+                    }
+                  }}
+                />
+                <button
+                  onClick={setTaskOnly}
+                  disabled={isRecording || isInitializing || isEncoding || isUploading || !robotsReady}
+                  className="btn-set-task"
+                  title={!robotsReady ? 'Please setup robots first' : 'Store task for pedal use (Enter key)'}
+                >
+                  💾 Set Task
+                </button>
+                <button
+                  onClick={startRecording}
+                  disabled={isRecording || isInitializing || isEncoding || isUploading || !robotsReady}
+                  className="btn-start"
+                  title={!robotsReady ? 'Please setup robots first' : ''}
+                >
+                  {isInitializing
+                    ? '⏳ Initializing...'
+                    : isRecording
+                    ? '⏺ Recording...'
+                    : robotsReady
+                    ? '⏺ Start Recording'
+                    : '⏺ Setup Robots First'}
+                </button>
+              </div>
+
+              {/* Ramp-up Countdown */}
+              {isRecording && rampUpRemaining > 0 && (
+                <div className="ramp-up-countdown">
+                  <div className="countdown-box">
+                    <div className="countdown-label">⚡ WARMING UP - PID RAMP-UP</div>
+                    <div className="countdown-value">{rampUpRemaining.toFixed(1)}s</div>
+                    <div className="countdown-subtitle">Recording will start automatically...</div>
+                  </div>
+                </div>
+              )}
+
+              {/* Recording Status - Only show after ramp-up */}
+              {isRecording && rampUpRemaining <= 0 && (
+                <div className="status recording recording-active">
+                  <div className="indicator"></div>
+                  <div className="time-display">
+                    <span>{formatTime(elapsedTime)}</span>
+                    <span className="fps-display">
+                      Loop: {loopFps.toFixed(1)} Hz
+                      {loopFps > 0 && loopFps < 29 && <span className="fps-warning"> ⚠️</span>}
+                    </span>
+                    <span className="fps-display">Recording: {currentFps.toFixed(1)} FPS</span>
+                  </div>
+                  <button onClick={stopRecording} className="btn-stop">
+                    ⏹ Stop
+                  </button>
+                </div>
+              )}
+            </div>
+
+            {/* Episode Counter */}
+            <div className="control-right">
+              <div className="counter">
+                <div className="counter-label">Episodes Recorded</div>
+                <div className="counter-value">{episodeCount}</div>
+                <button onClick={resetCounter} className="btn-reset">
+                  Reset
+                </button>
+              </div>
+            </div>
+          </div>
+
+          {/* Delete Latest Episode Button */}
+          {!isRecording && !isInitializing && latestRepoId && (
+            <div className="delete-episode-section">
+              <button 
+                onClick={deleteLatestEpisode}
+                className="btn-delete"
+                title="Delete the latest recorded episode from HuggingFace Hub"
+              >
+                Delete Latest Episode
+              </button>
+              <div className="delete-info">Will delete: {latestRepoId}</div>
+            </div>
+          )}
+
+          {/* Move to Zero Button */}
+          {robotsReady && !isRecording && !isInitializing && (
+            <div className="zero-position-section">
+              <button 
+                onClick={moveToZero} 
+                disabled={movingToZero}
+                className="btn-zero-large"
+                title="Move both leader and follower robots to zero position (2s)"
+              >
+                {movingToZero ? '⏳ Moving to Zero Position...' : '🎯 Move to Zero Position (Leader + Follower)'}
+              </button>
+            </div>
+          )}
+
+          {/* Error Display */}
+          {error && (
+            <div className="error-box">
+              ⚠️ {error}
+            </div>
+          )}
+        </section>
+        </div>
+
+        {/* Right Column: Camera Feeds */}
+        <div className="right-column">
+          <section className="panel cameras">
+          <h2>📹 Camera Views</h2>
+          {robotsReady || isRecording || isInitializing ? (
+            <div className="camera-layout">
+              {/* Base camera - full width */}
+              <div className="camera camera-base">
+                <h3>Base Camera</h3>
+                <img src={`${API_BASE}/camera/stream/base`} alt="Base Camera" />
+              </div>
+
+              {/* Wrist cameras - side by side */}
+              <div className="camera-wrist-container">
+                <div className="camera camera-wrist">
+                  <h3>Left Wrist</h3>
+                  <img src={`${API_BASE}/camera/stream/left_wrist`} alt="Left Wrist Camera" />
+                </div>
+
+                <div className="camera camera-wrist">
+                  <h3>Right Wrist</h3>
+                  <img src={`${API_BASE}/camera/stream/right_wrist`} alt="Right Wrist Camera" />
+                </div>
+              </div>
+            </div>
+          ) : (
+            <div className="camera-placeholder">
+              <p>📷 Camera feeds will appear when robots are set up</p>
+              <p className="hint">Click "Setup Robots" above to preview camera feeds</p>
+            </div>
+          )}
+        </section>
+        </div>
+
+      </div>
+    </main>
+  );
+}
+
+export default App;
+
@@ -0,0 +1,41 @@
+# OpenArms Web Recording Interface
+
+A web interface for recording OpenArms datasets.
+
+## Installation
+
+```bash
+cd examples/openarms_web_interface
+npm install
+```
+
+## Usage
+
+**Start everything with one command:**
+
+```bash
+./launch.sh
+```
+
+This will:
+- Start the FastAPI backend on port 8000
+- Start the React frontend on port 5173
+- Show live logs from both services
+
+Then open your browser to: **http://localhost:5173**
+
+**Stop with:** `Ctrl+C`
+
+---
+
+## Workflow
+
+1. **Configure CAN interfaces** and **camera paths** in the dropdowns
+2. Click **"Setup Robots"** to initialize (once at start)
+3. Enter a **task description**
+4. Click **"Start Recording"** to begin an episode
+5. Click **"Stop Recording"** when done
+6. Dataset is automatically encoded and uploaded to HuggingFace Hub as **private**
+7. Repeat steps 3-6 for more episodes (no need to re-setup robots!)
+
+---
@@ -0,0 +1,12 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="UTF-8" />
+    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+    <title>OpenArms Recording Interface</title>
+  </head>
+  <body>
+    <div id="root"></div>
+    <script type="module" src="/main.jsx"></script>
+  </body>
+</html>
@@ -0,0 +1,142 @@
+#!/bin/bash
+
+# OpenArms Web Interface Launcher
+# Starts Rerun viewer, FastAPI backend, and React frontend
+
+set -e
+
+# Colors for output
+GREEN='\033[0;32m'
+BLUE='\033[0;34m'
+YELLOW='\033[1;33m'
+RED='\033[0;31m'
+NC='\033[0m' # No Color
+
+# Get script directory
+SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
+cd "$SCRIPT_DIR"
+
+echo -e "${BLUE}╔════════════════════════════════════════╗${NC}"
+echo -e "${BLUE}║   OpenArms Web Recording Interface    ║${NC}"
+echo -e "${BLUE}╚════════════════════════════════════════╝${NC}"
+echo ""
+
+# Function to cleanup on exit
+cleanup() {
+    echo ""
+    echo -e "${YELLOW}Shutting down services...${NC}"
+    
+    # Kill all child processes
+    pkill -P $$ 2>/dev/null || true
+    
+    # Kill specific services by port
+    lsof -ti:8000 | xargs kill -9 2>/dev/null || true  # Backend
+    lsof -ti:5173 | xargs kill -9 2>/dev/null || true  # Frontend
+    lsof -ti:9876 | xargs kill -9 2>/dev/null || true  # Rerun (if spawned)
+    
+    echo -e "${GREEN}✓ Services stopped${NC}"
+    exit 0
+}
+
+# Register cleanup on script exit
+trap cleanup EXIT INT TERM
+
+# Check if required commands exist
+command -v rerun >/dev/null 2>&1 || { 
+    echo -e "${RED}✗ Error: 'rerun' not found. Please install: pip install rerun-sdk${NC}"
+    exit 1
+}
+
+command -v python >/dev/null 2>&1 || { 
+    echo -e "${RED}✗ Error: 'python' not found${NC}"
+    exit 1
+}
+
+command -v npm >/dev/null 2>&1 || { 
+    echo -e "${RED}✗ Error: 'npm' not found${NC}"
+    exit 1
+}
+
+# Check if node_modules exists
+if [ ! -d "node_modules" ]; then
+    echo -e "${YELLOW}⚠ node_modules not found. Running npm install...${NC}"
+    npm install
+    echo -e "${GREEN}✓ Dependencies installed${NC}"
+    echo ""
+fi
+
+echo -e "${GREEN}Starting services...${NC}"
+echo ""
+
+# 1. Start FastAPI backend (Rerun will start when recording begins)
+echo -e "${BLUE}[1/2]${NC} Starting FastAPI backend on port 8000..."
+cd "$SCRIPT_DIR"
+
+# Use Python from current environment (if lerobot env is active, it will use that)
+# Otherwise, check if we need to use conda run
+if [[ "$CONDA_DEFAULT_ENV" == "lerobot" ]]; then
+    # Already in lerobot environment
+    echo -e "${GREEN}✓ Using active lerobot environment${NC}"
+    PYTHON_CMD="python"
+elif command -v conda >/dev/null 2>&1 && conda env list | grep -q "^lerobot "; then
+    # lerobot env exists but not active - use conda run
+    echo -e "${YELLOW}Using conda run with lerobot environment...${NC}"
+    PYTHON_CMD="conda run -n lerobot --no-capture-output python"
+else
+    # Fall back to system python
+    echo -e "${YELLOW}⚠ Warning: lerobot environment not found, using system python${NC}"
+    PYTHON_CMD="python"
+fi
+
+$PYTHON_CMD web_record_server.py > /tmp/openarms_backend.log 2>&1 &
+BACKEND_PID=$!
+sleep 3
+
+if ps -p $BACKEND_PID > /dev/null; then
+    echo -e "${GREEN}✓ Backend started${NC} (PID: $BACKEND_PID)"
+    echo -e "      URL: ${BLUE}http://localhost:8000${NC}"
+else
+    echo -e "${RED}✗ Failed to start backend${NC}"
+    echo -e "${YELLOW}Check logs: tail -f /tmp/openarms_backend.log${NC}"
+    exit 1
+fi
+echo ""
+
+# 2. Start React frontend
+echo -e "${BLUE}[2/2]${NC} Starting React frontend on port 5173..."
+cd "$SCRIPT_DIR"
+npm run dev > /tmp/openarms_frontend.log 2>&1 &
+FRONTEND_PID=$!
+sleep 3
+
+if ps -p $FRONTEND_PID > /dev/null; then
+    echo -e "${GREEN}✓ Frontend started${NC} (PID: $FRONTEND_PID)"
+    echo -e "      URL: ${BLUE}http://localhost:5173${NC}"
+else
+    echo -e "${RED}✗ Failed to start frontend${NC}"
+    echo -e "${YELLOW}Check logs: tail -f /tmp/openarms_frontend.log${NC}"
+    exit 1
+fi
+echo ""
+
+# Display status
+echo -e "${GREEN}╔════════════════════════════════════════╗${NC}"
+echo -e "${GREEN}║     All services running! 🚀           ║${NC}"
+echo -e "${GREEN}╚════════════════════════════════════════╝${NC}"
+echo ""
+echo -e "🔧 ${BLUE}Backend:${NC}  http://localhost:8000"
+echo -e "🌐 ${BLUE}Frontend:${NC} http://localhost:5173"
+echo -e "📊 ${BLUE}Rerun:${NC}    Will spawn automatically when recording starts"
+echo ""
+echo -e "${YELLOW}Open your browser to:${NC} ${BLUE}http://localhost:5173${NC}"
+echo ""
+echo -e "${YELLOW}Logs:${NC}"
+echo -e "  • Backend:  tail -f /tmp/openarms_backend.log"
+echo -e "  • Frontend: tail -f /tmp/openarms_frontend.log"
+echo ""
+echo -e "${RED}Press Ctrl+C to stop all services${NC}"
+echo ""
+
+# Keep script running and wait for any service to exit
+wait
+
@@ -0,0 +1,7 @@
+import { createRoot } from 'react-dom/client'
+import App from './App.jsx'
+
+createRoot(document.getElementById('root')).render(
+  <App />
+)
+
@@ -0,0 +1,21 @@
+{
+  "name": "openarms-web-interface",
+  "private": true,
+  "version": "0.0.0",
+  "type": "module",
+  "scripts": {
+    "dev": "vite",
+    "build": "vite build",
+    "preview": "vite preview"
+  },
+  "dependencies": {
+    "react": "^18.3.1",
+    "react-dom": "^18.3.1"
+  },
+  "devDependencies": {
+    "@types/react": "^18.3.12",
+    "@types/react-dom": "^18.3.1",
+    "@vitejs/plugin-react": "^4.3.4",
+    "vite": "^6.0.1"
+  }
+}
@@ -0,0 +1,17 @@
+import { defineConfig } from 'vite'
+import react from '@vitejs/plugin-react'
+
+// https://vite.dev/config/
+export default defineConfig({
+  plugins: [react()],
+  server: {
+    port: 5173,
+    strictPort: false,
+    host: true,
+    open: false
+  },
+  build: {
+    outDir: 'dist',
+    sourcemap: true
+  }
+})
@@ -455,18 +455,7 @@ def demo_cli(cfg: RTCDemoConfig):
    if cfg.policy.type == "pi05" or cfg.policy.type == "pi0":
        config.compile_model = cfg.use_torch_compile

-    if config.use_peft:
-        from peft import PeftConfig, PeftModel
-
-        peft_pretrained_path = cfg.policy.pretrained_path
-        peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
-
-        policy = policy_class.from_pretrained(
-            pretrained_name_or_path=peft_config.base_model_name_or_path, config=config
-        )
-        policy = PeftModel.from_pretrained(policy, peft_pretrained_path, config=peft_config)
-    else:
-        policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
+    policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)

    # Turn on RTC
    policy.config.rtc_config = cfg.rtc
@@ -0,0 +1,10 @@
+from huggingface_hub import HfApi, list_datasets
+
+api = HfApi()
+datasets = list_datasets(author="lerobot-data-collection")
+print('"[', end="")
+i=0
+for dataset in datasets:
+    if "three-folds-dataset" in dataset.id:
+        print("'" + dataset.id + "',", end="")
+print(']"',)
@@ -96,14 +96,16 @@ dependencies = [
 # Common
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.10.0"]
-transformers-dep = ["transformers>=4.57.1,<5.0.0"]
+transformers-dep = ["transformers>=4.53.0,<5.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"] # TODO: Bumb dependency (compatible with wandb)

 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
+damiao = ["python-can>=4.2.0,<5.0.0"]

 # Robots
+openarms = ["lerobot[damiao]"]
 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"]
@@ -121,11 +123,16 @@ phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]

 # Policies
 wallx = [
+    "torch==2.6.0",
+    "torchvision==0.21.0",
+    "torchaudio==2.6.0",
    "transformers==4.49.0",
+    "accelerate==1.10.1",
    "peft==0.17.1",
    "scipy==1.15.3",
    "torchdiffeq==0.2.5",
-    "qwen_vl_utils==0.0.11"
+    "qwen_vl_utils==0.0.11",
+    "flash-attn==2.7.4.post1"
 ]
 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"]
@@ -140,16 +147,14 @@ 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"]
-peft = ["lerobot[transformers-dep]", "peft>=0.18.0"]

 # 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", "mypy>=1.19.1"]
+dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1"]
 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"]

@@ -162,13 +167,14 @@ metaworld = ["metaworld==3.0.0"]
 # All
 all = [
    "lerobot[dynamixel]",
+    "lerobot[openarms]",
    "lerobot[gamepad]",
    "lerobot[hopejr]",
    "lerobot[lekiwi]",
    "lerobot[reachy2]",
    "lerobot[kinematics]",
    "lerobot[intelrealsense]",
-    # "lerobot[wallx]",
+    "lerobot[wallx]",
    "lerobot[pi]",
    "lerobot[smolvla]",
    # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
@@ -183,8 +189,6 @@ all = [
    "lerobot[phone]",
    "lerobot[libero]",
    "lerobot[metaworld]",
-    "lerobot[sarm]",
-    "lerobot[peft]",
 ]

 [project.scripts]
@@ -239,7 +243,6 @@ 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
@@ -331,9 +334,9 @@ 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.*"
@@ -383,40 +386,3 @@ 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 = "hilserl" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "libero" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "all" },
-    ],
-]
@@ -67,31 +67,3 @@ class EvalConfig:
                f"to increase the number of episodes to match the batch size (e.g. `eval.n_episodes={self.batch_size}`), "
                f"or lower the batch size (e.g. `eval.batch_size={self.n_episodes}`)."
            )
-
-
-@dataclass
-class PeftConfig:
-    # PEFT offers many fine-tuning methods, layer adapters being the most common and currently also the most
-    # effective methods so we'll focus on those in this high-level config interface.
-
-    # Either a string (module name suffix or 'all-linear'), a list of module name suffixes or a regular expression
-    # describing module names to target with the configured PEFT method. Some policies have a default value for this
-    # so that you don't *have* to choose which layers to adapt but it might still be worthwhile depending on your case.
-    target_modules: list[str] | str | None = None
-
-    # Names/suffixes of modules to fully fine-tune and store alongside adapter weights. Useful for layers that are
-    # not part of a pre-trained model (e.g., action state projections). Depending on the policy this defaults to layers
-    # that are newly created in pre-trained policies. If you're fine-tuning an already trained policy you might want
-    # to set this to `[]`. Corresponds to PEFT's `modules_to_save`.
-    full_training_modules: list[str] | None = None
-
-    # The PEFT (adapter) method to apply to the policy. Needs to be a valid PEFT type.
-    method_type: str = "LORA"
-
-    # Adapter initialization method. Look at the specific PEFT adapter documentation for defaults.
-    init_type: str | None = None
-
-    # We expect that all PEFT adapters are in some way doing rank-decomposition therefore this parameter specifies
-    # the rank used for the adapter. In general a higher rank means more trainable parameters and closer to full
-    # fine-tuning.
-    r: int = 16
@@ -55,18 +55,14 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno

    n_obs_steps: int = 1

-    # `input_features` can be set to None/null in order to infer those values from the dataset.
-    input_features: dict[str, PolicyFeature] | None = field(default_factory=dict)
-    output_features: dict[str, PolicyFeature] | None = field(default_factory=dict)
+    input_features: dict[str, PolicyFeature] = field(default_factory=dict)
+    output_features: dict[str, PolicyFeature] = field(default_factory=dict)

    device: str | None = None  # e.g. "cuda", "cuda:0", "cpu", or "mps"
    # `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
    # automatic gradient scaling is used.
    use_amp: bool = False

-    # Whether the policy employed PEFT for training.
-    use_peft: bool = False
-
    push_to_hub: bool = True  # type: ignore[assignment] # TODO: use a different name to avoid override
    repo_id: str | None = None

@@ -129,8 +125,6 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno

    @property
    def robot_state_feature(self) -> PolicyFeature | None:
-        if self.input_features is None:
-            return None
        for ft_name, ft in self.input_features.items():
            if ft.type is FeatureType.STATE and ft_name == OBS_STATE:
                return ft
@@ -138,8 +132,6 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno

    @property
    def env_state_feature(self) -> PolicyFeature | None:
-        if self.input_features is None:
-            return None
        for _, ft in self.input_features.items():
            if ft.type is FeatureType.ENV:
                return ft
@@ -147,14 +139,10 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno

    @property
    def image_features(self) -> dict[str, PolicyFeature]:
-        if self.input_features is None:
-            return {}
        return {key: ft for key, ft in self.input_features.items() if ft.type is FeatureType.VISUAL}

    @property
    def action_feature(self) -> PolicyFeature | None:
-        if self.output_features is None:
-            return None
        for ft_name, ft in self.output_features.items():
            if ft.type is FeatureType.ACTION and ft_name == ACTION:
                return ft
@@ -24,7 +24,7 @@ from huggingface_hub.errors import HfHubHTTPError

 from lerobot import envs
 from lerobot.configs import parser
-from lerobot.configs.default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
+from lerobot.configs.default import DatasetConfig, EvalConfig, WandBConfig
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.optim import OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
@@ -65,18 +65,9 @@ class TrainPipelineConfig(HubMixin):
    scheduler: LRSchedulerConfig | None = None
    eval: EvalConfig = field(default_factory=EvalConfig)
    wandb: WandBConfig = field(default_factory=WandBConfig)
-    peft: PeftConfig | None = 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"
-
+    checkpoint_path: Path | None = field(init=False, default=None)
    # 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.
@@ -140,14 +131,6 @@ 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`"""
@@ -23,11 +23,13 @@ from pathlib import Path

 import datasets
 import numpy as np
+import os
 import packaging.version
 import pandas as pd
 import PIL.Image
 import pyarrow as pa
 import pyarrow.parquet as pq
+from concurrent.futures import ProcessPoolExecutor
 import torch
 import torch.utils
 from huggingface_hub import HfApi, snapshot_download
@@ -1199,6 +1201,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
        use_batched_encoding = self.batch_encoding_size > 1

        if has_video_keys and not use_batched_encoding:
+            video_paths = self._encode_multiple_temporary_episode_videos(self.meta.video_keys, episode_index)
+            for (video_key, video_path) in zip(self.meta.video_keys, video_paths):
+                ep_metadata.update(self._save_episode_video(video_key, episode_index, video_path))
            num_cameras = len(self.meta.video_keys)
            if parallel_encoding and num_cameras > 1:
                # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
@@ -1528,6 +1533,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
        """
        return _encode_video_worker(video_key, episode_index, self.root, self.fps)

+    def _encode_multiple_temporary_episode_videos(self, video_keys, episode_index):
+        temp_paths = []
+        img_dirs = []
+        for video_key in video_keys:
+            temp_paths.append(Path(tempfile.mkdtemp(dir=self.root)) / f"{video_key}_{episode_index:03d}.mp4")
+            img_dirs.append(self._get_image_file_dir(episode_index, video_key))
+        fps = [self.fps]*len(video_keys)
+
+        with ProcessPoolExecutor(max_workers=len(video_keys)) as executor:
+            executor.map(encode_video_frames,img_dirs,temp_paths,fps)
+
+        for img_dir in img_dirs:
+            shutil.rmtree(img_dir)
+
+        return temp_paths
+
    @classmethod
    def create(
        cls,
@@ -310,7 +310,7 @@ def encode_video_frames(
    crf: int | None = 30,
    fast_decode: int = 0,
    log_level: int | None = av.logging.ERROR,
-    overwrite: bool = False,
+    overwrite: bool = True,
    preset: int | None = None,
 ) -> None:
    """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
@@ -355,6 +355,9 @@ def encode_video_frames(
    if crf is not None:
        video_options["crf"] = str(crf)

+    #TEMPORARY FIX
+    video_options["preset"] = "12"
+
    if fast_decode:
        key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
        value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
@@ -14,4 +14,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-from .motors_bus import Motor, MotorCalibration, MotorNormMode, MotorsBus
+from .motors_bus import (
+    Motor,
+    MotorCalibration,
+    MotorNormMode,
+    MotorsBus,  # Backward compatibility (alias for SerialMotorsBus)
+    MotorsBusBase,
+    SerialMotorsBus,
+)
@@ -1,3 +1,5 @@
+#!/usr/bin/env python
+
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -11,3 +13,6 @@
 # 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 .damiao import DamiaoMotorsBus
+from .tables import *
@@ -0,0 +1,905 @@
+# 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.
+
+# TODO(pepijn): add license of: https://github.com/cmjang/DM_Control_Python?tab=MIT-1-ov-file#readme
+
+import logging
+import time
+from contextlib import contextmanager
+from copy import deepcopy
+from functools import cached_property
+from typing import Dict, List, Optional, Tuple, Union
+
+import can
+import numpy as np
+
+from lerobot.motors import Motor, MotorCalibration, MotorNormMode, MotorsBusBase
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.utils import enter_pressed, move_cursor_up
+
+from .tables import (
+    AVAILABLE_BAUDRATES,
+    CAN_CMD_DISABLE,
+    CAN_CMD_ENABLE,
+    CAN_CMD_REFRESH,
+    CAN_CMD_SET_ZERO,
+    CAN_PARAM_ID,
+    DEFAULT_BAUDRATE,
+    DEFAULT_TIMEOUT_MS,
+    MODEL_RESOLUTION,
+    MOTOR_LIMIT_PARAMS,
+    NORMALIZED_DATA,
+    MotorType,
+)
+
+logger = logging.getLogger(__name__)
+
+NameOrID = Union[str, int]
+Value = Union[int, float]
+
+
+class DamiaoMotorsBus(MotorsBusBase):
+    """
+    The Damiao implementation for a MotorsBus using CAN bus communication.
+    
+    This class uses python-can for CAN bus communication with Damiao motors.
+    For more info, see:
+    - python-can documentation: https://python-can.readthedocs.io/en/stable/
+    - Seedstudio documentation: https://wiki.seeedstudio.com/damiao_series/
+    - DM_Control_Python repo: https://github.com/cmjang/DM_Control_Python
+    """
+
+    # CAN-specific settings
+    available_baudrates = deepcopy(AVAILABLE_BAUDRATES)
+    default_baudrate = DEFAULT_BAUDRATE
+    default_timeout = DEFAULT_TIMEOUT_MS
+    
+    # Motor configuration
+    model_resolution_table = deepcopy(MODEL_RESOLUTION)
+    normalized_data = deepcopy(NORMALIZED_DATA)
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, Motor],
+        calibration: dict[str, MotorCalibration] | None = None,
+        can_interface: str = "auto",
+        use_can_fd: bool = True,
+        bitrate: int = 1000000,
+        data_bitrate: int | None = 5000000,
+    ):
+        """
+        Initialize the Damiao motors bus.
+
+        Args:
+            port: CAN interface name (e.g., "can0" for Linux, "/dev/cu.usbmodem*" for macOS)
+            motors: Dictionary mapping motor names to Motor objects
+            calibration: Optional calibration data
+            can_interface: CAN interface type - "auto" (default), "socketcan" (Linux), or "slcan" (macOS/serial)
+            use_can_fd: Whether to use CAN FD mode (default: True for OpenArms)
+            bitrate: Nominal bitrate in bps (default: 1000000 = 1 Mbps)
+            data_bitrate: Data bitrate for CAN FD in bps (default: 5000000 = 5 Mbps), ignored if use_can_fd is False
+        """
+        super().__init__(port, motors, calibration)
+        self.port = port
+        self.can_interface = can_interface
+        self.use_can_fd = use_can_fd
+        self.bitrate = bitrate
+        self.data_bitrate = data_bitrate
+        self.canbus = None
+        self._is_connected = False
+        
+        # Map motor names to CAN IDs
+        self._motor_can_ids = {}
+        self._recv_id_to_motor = {}
+        
+        # Store motor types and recv IDs
+        self._motor_types = {}
+        for name, motor in self.motors.items():
+            if hasattr(motor, "motor_type"):
+                self._motor_types[name] = motor.motor_type
+            else:
+                # Default to DM4310 if not specified
+                self._motor_types[name] = MotorType.DM4310
+            
+            # Map recv_id to motor name for filtering responses
+            if hasattr(motor, "recv_id"):
+                self._recv_id_to_motor[motor.recv_id] = name
+
+    @property
+    def is_connected(self) -> bool:
+        """Check if the CAN bus is connected."""
+        return self._is_connected and self.canbus is not None
+
+    def connect(self, handshake: bool = True) -> None:
+        """
+        Open the CAN bus and initialize communication.
+
+        Args:
+            handshake: If True, ping all motors to verify they're present
+        """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(
+                f"{self.__class__.__name__}('{self.port}') is already connected."
+            )
+
+        try:
+            # Auto-detect interface type based on port name
+            if self.can_interface == "auto":
+                if self.port.startswith("/dev/"):
+                    # Serial device (macOS/Windows)
+                    self.can_interface = "slcan"
+                    logger.info(f"Auto-detected slcan interface for port {self.port}")
+                else:
+                    # Network interface (Linux)
+                    self.can_interface = "socketcan"
+                    logger.info(f"Auto-detected socketcan interface for port {self.port}")
+            
+            # Connect to CAN bus
+            if self.can_interface == "socketcan":
+                # Linux SocketCAN with CAN FD support
+                if self.use_can_fd and self.data_bitrate is not None:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface="socketcan",
+                        bitrate=self.bitrate,
+                        data_bitrate=self.data_bitrate,
+                        fd=True
+                    )
+                    logger.info(f"Connected to {self.port} with CAN FD (bitrate={self.bitrate}, data_bitrate={self.data_bitrate})")
+                else:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface="socketcan",
+                        bitrate=self.bitrate
+                    )
+                    logger.info(f"Connected to {self.port} with CAN 2.0 (bitrate={self.bitrate})")
+            elif self.can_interface == "slcan":
+                # Serial Line CAN (macOS, Windows, or USB adapters)
+                # Note: SLCAN typically doesn't support CAN FD
+                self.canbus = can.interface.Bus(
+                    channel=self.port,
+                    interface="slcan",
+                    bitrate=self.bitrate
+                )
+                logger.info(f"Connected to {self.port} with SLCAN (bitrate={self.bitrate})")
+            else:
+                # Generic interface (vector, pcan, etc.)
+                if self.use_can_fd and self.data_bitrate is not None:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface=self.can_interface,
+                        bitrate=self.bitrate,
+                        data_bitrate=self.data_bitrate,
+                        fd=True
+                    )
+                else:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface=self.can_interface,
+                        bitrate=self.bitrate
+                    )
+            
+            self._is_connected = True
+            
+            if handshake:
+                self._handshake()
+                
+            logger.debug(f"{self.__class__.__name__} connected via {self.can_interface}.")
+        except Exception as e:
+            self._is_connected = False
+            raise ConnectionError(f"Failed to connect to CAN bus: {e}")
+
+    def _handshake(self) -> None:
+        """Verify all motors are present by refreshing their status."""
+        for motor_name in self.motors:
+            self._refresh_motor(motor_name)
+            time.sleep(0.01)  # Small delay between motors
+
+    def disconnect(self, disable_torque: bool = True) -> None:
+        """
+        Close the CAN bus connection.
+
+        Args:
+            disable_torque: If True, disable torque on all motors before disconnecting
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(
+                f"{self.__class__.__name__}('{self.port}') is not connected."
+            )
+
+        if disable_torque:
+            try:
+                self.disable_torque()
+            except Exception as e:
+                logger.warning(f"Failed to disable torque during disconnect: {e}")
+
+        if self.canbus:
+            self.canbus.shutdown()
+            self.canbus = None
+        self._is_connected = False
+        logger.debug(f"{self.__class__.__name__} disconnected.")
+
+    def configure_motors(self) -> None:
+        """Configure all motors with default settings."""
+        # Damiao motors don't require much configuration in MIT mode
+        # Just ensure they're enabled
+        for motor in self.motors:
+            self._enable_motor(motor)
+            time.sleep(0.01)
+
+    def _enable_motor(self, motor: NameOrID) -> None:
+        """Enable a single motor."""
+        motor_id = self._get_motor_id(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [0xFF] * 7 + [CAN_CMD_ENABLE]
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        self._recv_motor_response(expected_recv_id=recv_id)
+
+    def _disable_motor(self, motor: NameOrID) -> None:
+        """Disable a single motor."""
+        motor_id = self._get_motor_id(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [0xFF] * 7 + [CAN_CMD_DISABLE]
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        self._recv_motor_response(expected_recv_id=recv_id)
+
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Enable torque on selected motors."""
+        motors = self._get_motors_list(motors)
+        for motor in motors:
+            for _ in range(num_retry + 1):
+                try:
+                    self._enable_motor(motor)
+                    break
+                except Exception as e:
+                    if _ == num_retry:
+                        raise e
+                    time.sleep(0.01)
+
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Disable torque on selected motors."""
+        motors = self._get_motors_list(motors)
+        for motor in motors:
+            for _ in range(num_retry + 1):
+                try:
+                    self._disable_motor(motor)
+                    break
+                except Exception as e:
+                    if _ == num_retry:
+                        raise e
+                    time.sleep(0.01)
+
+    @contextmanager
+    def torque_disabled(self, motors: str | list[str] | None = None):
+        """
+        Context manager that guarantees torque is re-enabled.
+        
+        This helper is useful to temporarily disable torque when configuring motors.
+        
+        Examples:
+            >>> with bus.torque_disabled():
+            ...     # Safe operations here with torque disabled
+            ...     pass
+        """
+        self.disable_torque(motors)
+        try:
+            yield
+        finally:
+            self.enable_torque(motors)
+
+    def set_zero_position(self, motors: str | list[str] | None = None) -> None:
+        """Set current position as zero for selected motors."""
+        motors = self._get_motors_list(motors)
+        for motor in motors:
+            motor_id = self._get_motor_id(motor)
+            recv_id = self._get_motor_recv_id(motor)
+            data = [0xFF] * 7 + [CAN_CMD_SET_ZERO]
+            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+            self.canbus.send(msg)
+            self._recv_motor_response(expected_recv_id=recv_id)
+            time.sleep(0.01)
+
+    def _refresh_motor(self, motor: NameOrID) -> Optional[can.Message]:
+        """Refresh motor status and return the response."""
+        motor_id = self._get_motor_id(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
+        msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        return self._recv_motor_response(expected_recv_id=recv_id)
+
+    def _recv_motor_response(self, expected_recv_id: Optional[int] = None, timeout: float = 0.001) -> Optional[can.Message]:
+        """
+        Receive a response from a motor.
+        
+        Args:
+            expected_recv_id: If provided, only return messages from this CAN ID
+            timeout: Timeout in seconds (default: 1ms for high-speed operation)
+        
+        Returns:
+            CAN message if received, None otherwise
+        """
+        try:
+            start_time = time.time()
+            messages_seen = []
+            while time.time() - start_time < timeout:
+                msg = self.canbus.recv(timeout=0.0001)  # 100us timeout for fast polling
+                if msg:
+                    messages_seen.append(f"0x{msg.arbitration_id:02X}")
+                    # If no filter specified, return any message
+                    if expected_recv_id is None:
+                        return msg
+                    # Otherwise, only return if it matches the expected recv_id
+                    if msg.arbitration_id == expected_recv_id:
+                        return msg
+                    else:
+                        logger.debug(f"Ignoring message from CAN ID 0x{msg.arbitration_id:02X}, expected 0x{expected_recv_id:02X}")
+            
+            # Only log warnings if we're in debug mode to reduce overhead
+            if logger.isEnabledFor(logging.DEBUG):
+                if messages_seen:
+                    logger.debug(f"Received {len(messages_seen)} message(s) from IDs {set(messages_seen)}, but expected 0x{expected_recv_id:02X}")
+                else:
+                    logger.debug(f"No CAN messages received (expected from 0x{expected_recv_id:02X})")
+        except Exception as e:
+            logger.debug(f"Failed to receive CAN message: {e}")
+        return None
+    
+    def _recv_all_responses(self, expected_recv_ids: list[int], timeout: float = 0.002) -> dict[int, can.Message]:
+        """
+        Efficiently receive responses from multiple motors at once.
+        Uses the OpenArms pattern: collect all available messages within timeout.
+        
+        Args:
+            expected_recv_ids: List of CAN IDs we expect responses from
+            timeout: Total timeout in seconds (default: 2ms)
+        
+        Returns:
+            Dictionary mapping recv_id to CAN message
+        """
+        responses = {}
+        expected_set = set(expected_recv_ids)
+        start_time = time.time()
+        
+        try:
+            while len(responses) < len(expected_recv_ids) and (time.time() - start_time) < timeout:
+                msg = self.canbus.recv(timeout=0.0002)  # 200us poll timeout (increased from 100us for better reliability)
+                if msg and msg.arbitration_id in expected_set:
+                    responses[msg.arbitration_id] = msg
+                    if len(responses) == len(expected_recv_ids):
+                        break  # Got all responses, exit early
+        except Exception as e:
+            logger.debug(f"Error receiving responses: {e}")
+        
+        return responses
+
+    def _mit_control(
+        self,
+        motor: NameOrID,
+        kp: float,
+        kd: float,
+        position_degrees: float,
+        velocity_deg_per_sec: float,
+        torque: float,
+    ) -> None:
+        """
+        Send MIT control command to a motor.
+
+        Args:
+            motor: Motor name or ID
+            kp: Position gain
+            kd: Velocity gain
+            position_degrees: Target position (degrees)
+            velocity_deg_per_sec: Target velocity (degrees/s)
+            torque: Target torque (N·m)
+        """
+        motor_id = self._get_motor_id(motor)
+        motor_name = self._get_motor_name(motor)
+        motor_type = self._motor_types.get(motor_name, MotorType.DM4310)
+        
+        # Convert degrees to radians for motor control
+        position_rad = np.radians(position_degrees)
+        velocity_rad_per_sec = np.radians(velocity_deg_per_sec)
+        
+        # Get motor limits
+        pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+        
+        # Encode parameters
+        kp_uint = self._float_to_uint(kp, 0, 500, 12)
+        kd_uint = self._float_to_uint(kd, 0, 5, 12)
+        q_uint = self._float_to_uint(position_rad, -pmax, pmax, 16)
+        dq_uint = self._float_to_uint(velocity_rad_per_sec, -vmax, vmax, 12)
+        tau_uint = self._float_to_uint(torque, -tmax, tmax, 12)
+        
+        # Pack data
+        data = [0] * 8
+        data[0] = (q_uint >> 8) & 0xFF
+        data[1] = q_uint & 0xFF
+        data[2] = dq_uint >> 4
+        data[3] = ((dq_uint & 0xF) << 4) | ((kp_uint >> 8) & 0xF)
+        data[4] = kp_uint & 0xFF
+        data[5] = kd_uint >> 4
+        data[6] = ((kd_uint & 0xF) << 4) | ((tau_uint >> 8) & 0xF)
+        data[7] = tau_uint & 0xFF
+        
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        recv_id = self._get_motor_recv_id(motor)
+        self._recv_motor_response(expected_recv_id=recv_id)
+    
+    def _mit_control_batch(
+        self,
+        commands: Dict[NameOrID, Tuple[float, float, float, float, float]],
+    ) -> None:
+        """
+        Send MIT control commands to multiple motors in batch (optimized).
+        Sends all commands first, then collects responses. Much faster than sequential.
+        
+        Args:
+            commands: Dict mapping motor name/ID to (kp, kd, position_deg, velocity_deg/s, torque)
+                     Example: {'joint_1': (10.0, 0.5, 45.0, 0.0, 0.0), ...}
+        """
+        if not commands:
+            return
+        
+        expected_recv_ids = []
+        
+        # Step 1: Send all MIT control commands (no waiting)
+        for motor, (kp, kd, position_degrees, velocity_deg_per_sec, torque) in commands.items():
+            motor_id = self._get_motor_id(motor)
+            motor_name = self._get_motor_name(motor)
+            motor_type = self._motor_types.get(motor_name, MotorType.DM4310)
+            
+            # Convert degrees to radians
+            position_rad = np.radians(position_degrees)
+            velocity_rad_per_sec = np.radians(velocity_deg_per_sec)
+            
+            # Get motor limits
+            pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+            
+            # Encode parameters
+            kp_uint = self._float_to_uint(kp, 0, 500, 12)
+            kd_uint = self._float_to_uint(kd, 0, 5, 12)
+            q_uint = self._float_to_uint(position_rad, -pmax, pmax, 16)
+            dq_uint = self._float_to_uint(velocity_rad_per_sec, -vmax, vmax, 12)
+            tau_uint = self._float_to_uint(torque, -tmax, tmax, 12)
+            
+            # Pack data
+            data = [0] * 8
+            data[0] = (q_uint >> 8) & 0xFF
+            data[1] = q_uint & 0xFF
+            data[2] = dq_uint >> 4
+            data[3] = ((dq_uint & 0xF) << 4) | ((kp_uint >> 8) & 0xF)
+            data[4] = kp_uint & 0xFF
+            data[5] = kd_uint >> 4
+            data[6] = ((kd_uint & 0xF) << 4) | ((tau_uint >> 8) & 0xF)
+            data[7] = tau_uint & 0xFF
+            
+            # Send command
+            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+            self.canbus.send(msg)
+            
+            # Track expected response
+            recv_id = self._get_motor_recv_id(motor)
+            expected_recv_ids.append(recv_id)
+        
+        # Step 2: Collect all responses at once
+        self._recv_all_responses(expected_recv_ids, timeout=0.002)
+
+    def _float_to_uint(self, x: float, x_min: float, x_max: float, bits: int) -> int:
+        """Convert float to unsigned integer for CAN transmission."""
+        x = max(x_min, min(x_max, x))  # Clamp to range
+        span = x_max - x_min
+        data_norm = (x - x_min) / span
+        return int(data_norm * ((1 << bits) - 1))
+
+    def _uint_to_float(self, x: int, x_min: float, x_max: float, bits: int) -> float:
+        """Convert unsigned integer from CAN to float."""
+        span = x_max - x_min
+        data_norm = float(x) / ((1 << bits) - 1)
+        return data_norm * span + x_min
+
+    def _decode_motor_state(self, data: bytes, motor_type: MotorType) -> Tuple[float, float, float, int, int]:
+        """
+        Decode motor state from CAN data.
+
+        Returns:
+            Tuple of (position_degrees, velocity_deg_per_sec, torque, temp_mos, temp_rotor)
+        """
+        if len(data) < 8:
+            raise ValueError("Invalid motor state data")
+        
+        # Extract encoded values
+        q_uint = (data[1] << 8) | data[2]
+        dq_uint = (data[3] << 4) | (data[4] >> 4)
+        tau_uint = ((data[4] & 0x0F) << 8) | data[5]
+        t_mos = data[6]
+        t_rotor = data[7]
+        
+        # Get motor limits
+        pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+        
+        # Decode to physical values (radians)
+        position_rad = self._uint_to_float(q_uint, -pmax, pmax, 16)
+        velocity_rad_per_sec = self._uint_to_float(dq_uint, -vmax, vmax, 12)
+        torque = self._uint_to_float(tau_uint, -tmax, tmax, 12)
+        
+        # Convert to degrees
+        position_degrees = np.degrees(position_rad)
+        velocity_deg_per_sec = np.degrees(velocity_rad_per_sec)
+        
+        return position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor
+
+    def read(
+        self,
+        data_name: str,
+        motor: str,
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> Value:
+        """Read a value from a single motor. Positions are always in degrees."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        
+        # Refresh motor to get latest state
+        msg = self._refresh_motor(motor)
+        if msg is None:
+            motor_id = self._get_motor_id(motor)
+            recv_id = self._get_motor_recv_id(motor)
+            raise ConnectionError(
+                f"No response from motor '{motor}' (send ID: 0x{motor_id:02X}, recv ID: 0x{recv_id:02X}). "
+                f"Check that: 1) Motor is powered (24V), 2) CAN wiring is correct, "
+                f"3) Motor IDs are configured correctly using Damiao Debugging Tools"
+            )
+        
+        motor_type = self._motor_types.get(motor, MotorType.DM4310)
+        position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
+        
+        # Return requested data (already in degrees for position/velocity)
+        if data_name == "Present_Position":
+            value = position_degrees
+        elif data_name == "Present_Velocity":
+            value = velocity_deg_per_sec
+        elif data_name == "Present_Torque":
+            value = torque
+        elif data_name == "Temperature_MOS":
+            value = t_mos
+        elif data_name == "Temperature_Rotor":
+            value = t_rotor
+        else:
+            raise ValueError(f"Unknown data_name: {data_name}")
+        
+        # For Damiao, positions are always in degrees, no normalization needed
+        # We keep the normalize parameter for compatibility but don't use it
+        return value
+
+    def write(
+        self,
+        data_name: str,
+        motor: str,
+        value: Value,
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> None:
+        """Write a value to a single motor. Positions are always in degrees."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        
+        # Value is expected to be in degrees for positions
+        if data_name == "Goal_Position":
+            # Use MIT control with position in degrees
+            self._mit_control(motor, 10.0, 0.5, value, 0, 0)
+        else:
+            raise ValueError(f"Writing {data_name} not supported in MIT mode")
+
+    def sync_read(
+        self,
+        data_name: str,
+        motors: str | list[str] | None = None,
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> Dict[str, Value]:
+        """
+        Read the same value from multiple motors simultaneously.
+        Uses batched operations: sends all refresh commands, then collects all responses.
+        This is MUCH faster than sequential reads (OpenArms pattern).
+        """
+        motors = self._get_motors_list(motors)
+        result = {}
+
+        # Step 1: Send refresh commands to ALL motors first (no waiting)
+        for motor in motors:
+            motor_id = self._get_motor_id(motor)
+            data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
+            msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False)
+            self.canbus.send(msg)
+        
+        # Step 2: Collect all responses at once (batch receive)
+        expected_recv_ids = [self._get_motor_recv_id(motor) for motor in motors]
+        responses = self._recv_all_responses(expected_recv_ids, timeout=0.01)  # 10ms total timeout
+        
+        # Step 3: Parse responses
+        for motor in motors:
+            try:
+                recv_id = self._get_motor_recv_id(motor)
+                msg = responses.get(recv_id)
+                
+                if msg is None:
+                    logger.warning(f"No response from motor '{motor}' (recv ID: 0x{recv_id:02X})")
+                    result[motor] = 0.0
+                    continue
+                
+                motor_type = self._motor_types.get(motor, MotorType.DM4310)
+                position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
+                
+                # Return requested data
+                if data_name == "Present_Position":
+                    value = position_degrees
+                elif data_name == "Present_Velocity":
+                    value = velocity_deg_per_sec
+                elif data_name == "Present_Torque":
+                    value = torque
+                elif data_name == "Temperature_MOS":
+                    value = t_mos
+                elif data_name == "Temperature_Rotor":
+                    value = t_rotor
+                else:
+                    raise ValueError(f"Unknown data_name: {data_name}")
+                
+                result[motor] = value
+                
+            except Exception as e:
+                logger.warning(f"Failed to read {data_name} from {motor}: {e}")
+                result[motor] = 0.0
+        
+        return result
+    
+    def sync_read_all_states(
+        self,
+        motors: str | list[str] | None = None,
+        *,
+        num_retry: int = 0,
+    ) -> Dict[str, Dict[str, Value]]:
+        """
+        Read ALL motor states (position, velocity, torque) from multiple motors in ONE refresh cycle.
+        This is 3x faster than calling sync_read() three times separately.
+        
+        Returns:
+            Dictionary mapping motor names to state dicts with keys: 'position', 'velocity', 'torque'
+            Example: {'joint_1': {'position': 45.2, 'velocity': 1.3, 'torque': 0.5}, ...}
+        """
+        motors = self._get_motors_list(motors)
+        result = {}
+
+        # Step 1: Send refresh commands to ALL motors first (with small delays to reduce bus congestion)
+        for motor in motors:
+            motor_id = self._get_motor_id(motor)
+            data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
+            msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False)
+            self.canbus.send(msg)
+            time.sleep(0.0001)  # 100us delay between commands to reduce bus congestion
+        
+        # Step 2: Collect all responses at once (batch receive)
+        expected_recv_ids = [self._get_motor_recv_id(motor) for motor in motors]
+        responses = self._recv_all_responses(expected_recv_ids, timeout=0.015)  # 15ms timeout (increased for reliability)
+        
+        # Step 3: Parse responses and extract ALL state values
+        for motor in motors:
+            try:
+                recv_id = self._get_motor_recv_id(motor)
+                msg = responses.get(recv_id)
+                
+                if msg is None:
+                    logger.warning(f"No response from motor '{motor}' (recv ID: 0x{recv_id:02X})")
+                    result[motor] = {"position": 0.0, "velocity": 0.0, "torque": 0.0}
+                    continue
+                
+                motor_type = self._motor_types.get(motor, MotorType.DM4310)
+                position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
+                
+                # Return all state values in one dict
+                result[motor] = {
+                    "position": position_degrees,
+                    "velocity": velocity_deg_per_sec,
+                    "torque": torque,
+                    "temp_mos": t_mos,
+                    "temp_rotor": t_rotor,
+                }
+                
+            except Exception as e:
+                logger.warning(f"Failed to read state from {motor}: {e}")
+                result[motor] = {"position": 0.0, "velocity": 0.0, "torque": 0.0}
+        
+        return result
+
+    def sync_write(
+        self,
+        data_name: str,
+        values: Dict[str, Value],
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> None:
+        """
+        Write different values to multiple motors simultaneously. Positions are always in degrees.
+        Uses batched operations: sends all commands first, then collects responses (OpenArms pattern).
+        """
+        if data_name == "Goal_Position":
+            # Step 1: Send all MIT control commands first (no waiting)
+            for motor, value_degrees in values.items():
+                motor_id = self._get_motor_id(motor)
+                motor_name = self._get_motor_name(motor)
+                motor_type = self._motor_types.get(motor_name, MotorType.DM4310)
+                
+                # Convert degrees to radians
+                position_rad = np.radians(value_degrees)
+                
+                # Default gains for position control
+                kp, kd = 10.0, 0.5
+                
+                # Get motor limits and encode parameters
+                pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+                kp_uint = self._float_to_uint(kp, 0, 500, 12)
+                kd_uint = self._float_to_uint(kd, 0, 5, 12)
+                q_uint = self._float_to_uint(position_rad, -pmax, pmax, 16)
+                dq_uint = self._float_to_uint(0, -vmax, vmax, 12)
+                tau_uint = self._float_to_uint(0, -tmax, tmax, 12)
+                
+                # Pack data
+                data = [0] * 8
+                data[0] = (q_uint >> 8) & 0xFF
+                data[1] = q_uint & 0xFF
+                data[2] = dq_uint >> 4
+                data[3] = ((dq_uint & 0xF) << 4) | ((kp_uint >> 8) & 0xF)
+                data[4] = kp_uint & 0xFF
+                data[5] = kd_uint >> 4
+                data[6] = ((kd_uint & 0xF) << 4) | ((tau_uint >> 8) & 0xF)
+                data[7] = tau_uint & 0xFF
+                
+                msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+                self.canbus.send(msg)
+                time.sleep(0.0001)  # 100us delay between commands to reduce bus congestion
+            
+            # Step 2: Collect all responses at once
+            expected_recv_ids = [self._get_motor_recv_id(motor) for motor in values.keys()]
+            self._recv_all_responses(expected_recv_ids, timeout=0.015)  # 15ms timeout (increased for reliability)
+        else:
+            # Fall back to individual writes for other data types
+            for motor, value in values.items():
+                self.write(data_name, motor, value, normalize=normalize, num_retry=num_retry)
+
+    def read_calibration(self) -> dict[str, MotorCalibration]:
+        """Read calibration data from motors."""
+        # Damiao motors don't store calibration internally
+        # Return existing calibration or empty dict
+        return self.calibration if self.calibration else {}
+
+    def write_calibration(self, calibration_dict: dict[str, MotorCalibration], cache: bool = True) -> None:
+        """Write calibration data to motors."""
+        # Damiao motors don't store calibration internally
+        # Just cache it in memory
+        if cache:
+            self.calibration = calibration_dict
+
+    def record_ranges_of_motion(
+        self, motors: NameOrID | list[NameOrID] | None = None, display_values: bool = True
+    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
+        """
+        Interactively record the min/max values of each motor in degrees.
+
+        Move the joints by hand (with torque disabled) while the method streams live positions.
+        Press Enter to finish.
+        """
+        if motors is None:
+            motors = list(self.motors.keys())
+        elif isinstance(motors, (str, int)):
+            motors = [motors]
+
+        # Disable torque for manual movement
+        self.disable_torque(motors)
+        time.sleep(0.1)
+
+        # Get initial positions (already in degrees)
+        start_positions = self.sync_read("Present_Position", motors, normalize=False)
+        mins = start_positions.copy()
+        maxes = start_positions.copy()
+
+        print("\nMove joints through their full range of motion. Press ENTER when done.")
+        user_pressed_enter = False
+        
+        while not user_pressed_enter:
+            positions = self.sync_read("Present_Position", motors, normalize=False)
+            
+            for motor in motors:
+                if motor in positions:
+                    mins[motor] = min(positions[motor], mins.get(motor, positions[motor]))
+                    maxes[motor] = max(positions[motor], maxes.get(motor, positions[motor]))
+
+            if display_values:
+                print("\n" + "=" * 50)
+                print(f"{'MOTOR':<20} | {'MIN (deg)':>12} | {'POS (deg)':>12} | {'MAX (deg)':>12}")
+                print("-" * 50)
+                for motor in motors:
+                    if motor in positions:
+                        print(f"{motor:<20} | {mins[motor]:>12.1f} | {positions[motor]:>12.1f} | {maxes[motor]:>12.1f}")
+
+            if enter_pressed():
+                user_pressed_enter = True
+
+            if display_values and not user_pressed_enter:
+                # Move cursor up to overwrite the previous output
+                move_cursor_up(len(motors) + 4)
+
+            time.sleep(0.05)
+
+        # Re-enable torque
+        self.enable_torque(motors)
+        
+        # Validate ranges
+        for motor in motors:
+            if motor in mins and motor in maxes:
+                if abs(maxes[motor] - mins[motor]) < 5.0:  # At least 5 degrees of range
+                    raise ValueError(f"Motor {motor} has insufficient range of motion (< 5 degrees)")
+
+        return mins, maxes
+
+    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
+        """Convert motor specification to list of motor names."""
+        if motors is None:
+            return list(self.motors.keys())
+        elif isinstance(motors, str):
+            return [motors]
+        elif isinstance(motors, list):
+            return motors
+        else:
+            raise TypeError(f"Invalid motors type: {type(motors)}")
+
+    def _get_motor_id(self, motor: NameOrID) -> int:
+        """Get CAN ID for a motor."""
+        if isinstance(motor, str):
+            if motor in self.motors:
+                return self.motors[motor].id
+            else:
+                raise ValueError(f"Unknown motor: {motor}")
+        else:
+            return motor
+
+    def _get_motor_name(self, motor: NameOrID) -> str:
+        """Get motor name from name or ID."""
+        if isinstance(motor, str):
+            return motor
+        else:
+            for name, m in self.motors.items():
+                if m.id == motor:
+                    return name
+            raise ValueError(f"Unknown motor ID: {motor}")
+    
+    def _get_motor_recv_id(self, motor: NameOrID) -> Optional[int]:
+        """Get motor recv_id from name or ID."""
+        motor_name = self._get_motor_name(motor)
+        motor_obj = self.motors.get(motor_name)
+        if motor_obj and hasattr(motor_obj, "recv_id"):
+            return motor_obj.recv_id
+        return None
+
+    @cached_property
+    def is_calibrated(self) -> bool:
+        """Check if motors are calibrated."""
+        return bool(self.calibration)
@@ -0,0 +1,209 @@
+# 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.
+
+"""Configuration tables for Damiao motors."""
+
+from enum import IntEnum
+from typing import Dict, List, Tuple
+
+# Motor type definitions
+class MotorType(IntEnum):
+    DM3507 = 0
+    DM4310 = 1
+    DM4310_48V = 2
+    DM4340 = 3
+    DM4340_48V = 4
+    DM6006 = 5
+    DM8006 = 6
+    DM8009 = 7
+    DM10010L = 8
+    DM10010 = 9
+    DMH3510 = 10
+    DMH6215 = 11
+    DMG6220 = 12
+
+# Control modes
+class ControlMode(IntEnum):
+    MIT = 1
+    POS_VEL = 2
+    VEL = 3
+    TORQUE_POS = 4
+
+# Motor variable IDs (RID)
+class MotorVariable(IntEnum):
+    UV_VALUE = 0
+    KT_VALUE = 1
+    OT_VALUE = 2
+    OC_VALUE = 3
+    ACC = 4
+    DEC = 5
+    MAX_SPD = 6
+    MST_ID = 7
+    ESC_ID = 8
+    TIMEOUT = 9
+    CTRL_MODE = 10
+    DAMP = 11
+    INERTIA = 12
+    HW_VER = 13
+    SW_VER = 14
+    SN = 15
+    NPP = 16
+    RS = 17
+    LS = 18
+    FLUX = 19
+    GR = 20
+    PMAX = 21
+    VMAX = 22
+    TMAX = 23
+    I_BW = 24
+    KP_ASR = 25
+    KI_ASR = 26
+    KP_APR = 27
+    KI_APR = 28
+    OV_VALUE = 29
+    GREF = 30
+    DETA = 31
+    V_BW = 32
+    IQ_C1 = 33
+    VL_C1 = 34
+    CAN_BR = 35
+    SUB_VER = 36
+    U_OFF = 50
+    V_OFF = 51
+    K1 = 52
+    K2 = 53
+    M_OFF = 54
+    DIR = 55
+    P_M = 80
+    XOUT = 81
+
+# Motor limit parameters [PMAX, VMAX, TMAX] 
+# PMAX: Maximum position (rad)
+# VMAX: Maximum velocity (rad/s)
+# TMAX: Maximum torque (N·m)
+MOTOR_LIMIT_PARAMS = {
+    MotorType.DM3507: (12.5, 30, 10),
+    MotorType.DM4310: (12.5, 30, 10),
+    MotorType.DM4310_48V: (12.5, 50, 10),
+    MotorType.DM4340: (12.5, 8, 28),
+    MotorType.DM4340_48V: (12.5, 10, 28),
+    MotorType.DM6006: (12.5, 45, 20),
+    MotorType.DM8006: (12.5, 45, 40),
+    MotorType.DM8009: (12.5, 45, 54),
+    MotorType.DM10010L: (12.5, 25, 200),
+    MotorType.DM10010: (12.5, 20, 200),
+    MotorType.DMH3510: (12.5, 280, 1),
+    MotorType.DMH6215: (12.5, 45, 10),
+    MotorType.DMG6220: (12.5, 45, 10),
+}
+
+# Motor model names
+MODEL_NAMES = {
+    MotorType.DM3507: "dm3507",
+    MotorType.DM4310: "dm4310",
+    MotorType.DM4310_48V: "dm4310_48v",
+    MotorType.DM4340: "dm4340",
+    MotorType.DM4340_48V: "dm4340_48v",
+    MotorType.DM6006: "dm6006",
+    MotorType.DM8006: "dm8006",
+    MotorType.DM8009: "dm8009",
+    MotorType.DM10010L: "dm10010l",
+    MotorType.DM10010: "dm10010",
+    MotorType.DMH3510: "dmh3510",
+    MotorType.DMH6215: "dmh6215",
+    MotorType.DMG6220: "dmg6220",
+}
+
+# Motor resolution table (encoder counts per revolution)
+MODEL_RESOLUTION = {
+    "dm3507": 65536,
+    "dm4310": 65536,
+    "dm4310_48v": 65536,
+    "dm4340": 65536,
+    "dm4340_48v": 65536,
+    "dm6006": 65536,
+    "dm8006": 65536,
+    "dm8009": 65536,
+    "dm10010l": 65536,
+    "dm10010": 65536,
+    "dmh3510": 65536,
+    "dmh6215": 65536,
+    "dmg6220": 65536,
+}
+
+# CAN baudrates supported by Damiao motors
+AVAILABLE_BAUDRATES = [
+    125000,   # 0: 125 kbps
+    200000,   # 1: 200 kbps
+    250000,   # 2: 250 kbps
+    500000,   # 3: 500 kbps
+    1000000,  # 4: 1 mbps (default for OpenArms)
+    2000000,  # 5: 2 mbps
+    2500000,  # 6: 2.5 mbps
+    3200000,  # 7: 3.2 mbps
+    4000000,  # 8: 4 mbps
+    5000000,  # 9: 5 mbps
+]
+DEFAULT_BAUDRATE = 1000000  # 1 Mbps is standard for OpenArms
+
+# Default timeout in milliseconds
+DEFAULT_TIMEOUT_MS = 1000
+
+# Data that should be normalized
+NORMALIZED_DATA = ["Present_Position", "Goal_Position"]
+
+# OpenArms specific configurations
+# Based on: https://docs.openarm.dev/software/setup/configure-test
+# OpenArms has 7 DOF per arm (14 total for dual arm)
+OPENARMS_ARM_MOTOR_IDS = {
+    "joint_1": {"send": 0x01, "recv": 0x11},  # J1 - Shoulder pan
+    "joint_2": {"send": 0x02, "recv": 0x12},  # J2 - Shoulder lift
+    "joint_3": {"send": 0x03, "recv": 0x13},  # J3 - Elbow flex
+    "joint_4": {"send": 0x04, "recv": 0x14},  # J4 - Wrist flex
+    "joint_5": {"send": 0x05, "recv": 0x15},  # J5 - Wrist roll
+    "joint_6": {"send": 0x06, "recv": 0x16},  # J6 - Wrist pitch
+    "joint_7": {"send": 0x07, "recv": 0x17},  # J7 - Wrist rotation
+}
+
+OPENARMS_GRIPPER_MOTOR_IDS = {
+    "gripper": {"send": 0x08, "recv": 0x18},  # J8 - Gripper
+}
+
+# Default motor types for OpenArms
+OPENARMS_DEFAULT_MOTOR_TYPES = {
+    "joint_1": MotorType.DM8009,   # Shoulder pan - high torque
+    "joint_2": MotorType.DM8009,   # Shoulder lift - high torque
+    "joint_3": MotorType.DM4340,   # Shoulder rotation
+    "joint_4": MotorType.DM4340,   # Elbow flex
+    "joint_5": MotorType.DM4310,   # Wrist roll
+    "joint_6": MotorType.DM4310,   # Wrist pitch
+    "joint_7": MotorType.DM4310,   # Wrist rotation
+    "gripper": MotorType.DM4310,   # Gripper
+}
+
+# MIT control parameter ranges
+MIT_KP_RANGE = (0.0, 500.0)
+MIT_KD_RANGE = (0.0, 5.0)
+
+# CAN frame command IDs
+CAN_CMD_ENABLE = 0xFC
+CAN_CMD_DISABLE = 0xFD
+CAN_CMD_SET_ZERO = 0xFE
+CAN_CMD_REFRESH = 0xCC
+CAN_CMD_QUERY_PARAM = 0x33
+CAN_CMD_WRITE_PARAM = 0x55
+CAN_CMD_SAVE_PARAM = 0xAA
+
+# CAN ID for parameter operations
+CAN_PARAM_ID = 0x7FF
@@ -24,7 +24,7 @@ from enum import Enum

 from lerobot.motors.encoding_utils import decode_twos_complement, encode_twos_complement

-from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
+from ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
 from .tables import (
    AVAILABLE_BAUDRATES,
    MODEL_BAUDRATE_TABLE,
@@ -100,7 +100,7 @@ def _split_into_byte_chunks(value: int, length: int) -> list[int]:
    return data


-class DynamixelMotorsBus(MotorsBus):
+class DynamixelMotorsBus(SerialMotorsBus):
    """
    The Dynamixel implementation for a MotorsBus. It relies on the python dynamixel sdk to communicate with
    the motors. For more info, see the Dynamixel SDK Documentation:
@@ -19,7 +19,7 @@ from pprint import pformat

 from lerobot.motors.encoding_utils import decode_sign_magnitude, encode_sign_magnitude

-from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
+from ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
 from .tables import (
    FIRMWARE_MAJOR_VERSION,
    FIRMWARE_MINOR_VERSION,
@@ -96,7 +96,7 @@ def patch_setPacketTimeout(self, packet_length):  # noqa: N802
    self.packet_timeout = (self.tx_time_per_byte * packet_length) + (self.tx_time_per_byte * 3.0) + 50


-class FeetechMotorsBus(MotorsBus):
+class FeetechMotorsBus(SerialMotorsBus):
    """
    The FeetechMotorsBus class allows to efficiently read and write to the attached motors. It relies on the
    python feetech sdk to communicate with the motors, which is itself based on the dynamixel sdk.
@@ -165,7 +165,7 @@ class FeetechMotorsBus(MotorsBus):

    def _handshake(self) -> None:
        self._assert_motors_exist()
-        self._assert_same_firmware()
+        #self._assert_same_firmware()

    def _find_single_motor(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
        if self.protocol_version == 0:
@@ -19,6 +19,8 @@
 # TODO(aliberts): Add block noqa when feature below is available
 # https://github.com/astral-sh/ruff/issues/3711

+from __future__ import annotations
+
 import abc
 import logging
 from contextlib import contextmanager
@@ -41,6 +43,92 @@ Value: TypeAlias = int | float
 logger = logging.getLogger(__name__)


+class MotorsBusBase(abc.ABC):
+    """
+    Base class for all motor bus implementations.
+    
+    This is a minimal interface that all motor buses must implement, regardless of their
+    communication protocol (serial, CAN, etc.).
+    """
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, Motor],
+        calibration: dict[str, MotorCalibration] | None = None,
+    ):
+        self.port = port
+        self.motors = motors
+        self.calibration = calibration if calibration else {}
+
+    @abc.abstractmethod
+    def connect(self, handshake: bool = True) -> None:
+        """Establish connection to the motors."""
+        pass
+
+    @abc.abstractmethod
+    def disconnect(self, disable_torque: bool = True) -> None:
+        """Disconnect from the motors."""
+        pass
+
+    @property
+    @abc.abstractmethod
+    def is_connected(self) -> bool:
+        """Check if connected to the motors."""
+        pass
+
+    @abc.abstractmethod
+    def read(self, data_name: str, motor: str, *, normalize: bool = True, num_retry: int = 0) -> Value:
+        """Read a value from a single motor."""
+        pass
+
+    @abc.abstractmethod
+    def write(
+        self, data_name: str, motor: str, value: Value, *, normalize: bool = True, num_retry: int = 0
+    ) -> None:
+        """Write a value to a single motor."""
+        pass
+
+    @abc.abstractmethod
+    def sync_read(
+        self, data_name: str, motors: str | list[str] | None = None, *, normalize: bool = True
+    ) -> dict[str, Value]:
+        """Read a value from multiple motors."""
+        pass
+
+    @abc.abstractmethod
+    def sync_write(
+        self,
+        data_name: str,
+        values: Value | dict[str, Value],
+        motors: str | list[str] | None = None,
+        *,
+        normalize: bool = True,
+    ) -> None:
+        """Write values to multiple motors."""
+        pass
+
+    @abc.abstractmethod
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Enable torque on selected motors."""
+        pass
+
+    @abc.abstractmethod
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Disable torque on selected motors."""
+        pass
+
+    @abc.abstractmethod
+    def read_calibration(self) -> dict[str, MotorCalibration]:
+        """Read calibration parameters from the motors."""
+        pass
+
+    @abc.abstractmethod
+    def write_calibration(self, calibration_dict: dict[str, MotorCalibration], cache: bool = True) -> None:
+        """Write calibration parameters to the motors."""
+        pass
+
+
 def get_ctrl_table(model_ctrl_table: dict[str, dict], model: str) -> dict[str, tuple[int, int]]:
    ctrl_table = model_ctrl_table.get(model)
    if ctrl_table is None:
@@ -203,15 +291,15 @@ class GroupSyncWrite(Protocol):
    def txPacket(self): ...


-class MotorsBus(abc.ABC):
+class SerialMotorsBus(MotorsBusBase):
    """
-    A MotorsBus allows to efficiently read and write to the attached motors.
+    A SerialMotorsBus allows to efficiently read and write to motors connected via serial communication.
    It represents several motors daisy-chained together and connected through a serial port.
-    There are currently two implementations of this abstract class:
+    There are currently two implementations of this class:
        - DynamixelMotorsBus
        - FeetechMotorsBus

-    Note: This class may evolve in the future should we add support for other types of bus.
+    This class is specifically for serial-based motor protocols (Dynamixel, Feetech, etc.).

    A MotorsBus subclass instance requires a port (e.g. `FeetechMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
    To find the port, you can run our utility script:
@@ -1212,3 +1300,7 @@ class MotorsBus(abc.ABC):
        for id_, value in ids_values.items():
            data = self._serialize_data(value, length)
            self.sync_writer.addParam(id_, data)
+
+
+# Backward compatibility alias
+MotorsBus = SerialMotorsBus
@@ -35,8 +35,6 @@ 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
@@ -14,7 +14,6 @@
 # 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
@@ -30,17 +29,6 @@ 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):
@@ -57,24 +45,13 @@ class OptimizerConfig(draccus.ChoiceRegistry, abc.ABC):
        return "adam"

    @abc.abstractmethod
-    def build(self, params: OptimizerParams) -> torch.optim.Optimizer | dict[str, torch.optim.Optimizer]:
+    def build(self) -> 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.
        """
@@ -90,7 +67,7 @@ class AdamConfig(OptimizerConfig):
    weight_decay: float = 0.0
    grad_clip_norm: float = 10.0

-    def build(self, params: OptimizerParams) -> torch.optim.Optimizer:
+    def build(self, params: dict) -> torch.optim.Optimizer:
        kwargs = asdict(self)
        kwargs.pop("grad_clip_norm")
        return torch.optim.Adam(params, **kwargs)
@@ -105,7 +82,7 @@ class AdamWConfig(OptimizerConfig):
    weight_decay: float = 1e-2
    grad_clip_norm: float = 10.0

-    def build(self, params: OptimizerParams) -> torch.optim.Optimizer:
+    def build(self, params: dict) -> torch.optim.Optimizer:
        kwargs = asdict(self)
        kwargs.pop("grad_clip_norm")
        return torch.optim.AdamW(params, **kwargs)
@@ -121,7 +98,7 @@ class SGDConfig(OptimizerConfig):
    weight_decay: float = 0.0
    grad_clip_norm: float = 10.0

-    def build(self, params: OptimizerParams) -> torch.optim.Optimizer:
+    def build(self, params: dict) -> torch.optim.Optimizer:
        kwargs = asdict(self)
        kwargs.pop("grad_clip_norm")
        return torch.optim.SGD(params, **kwargs)
@@ -162,19 +139,21 @@ class XVLAAdamWConfig(OptimizerConfig):
    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:
+    def build(self, params: dict) -> torch.optim.Optimizer:
        """
        Build AdamW optimizer with differential learning rates.

+        Expects `named_parameters()` as input (dict of name -> param).
+        Applies:
+        - lr * 0.1 for all VLM-related parameters
+        - lr * soft_prompt_lr_scale for soft-prompt parameters (with optional warmup)
+        - full lr for all other parameters
+
        Args:
-            params: Must be a dict[str, Parameter] from dict(model.named_parameters())
-                or equivalent.
+            params: Dictionary of parameter names to parameters (from named_parameters())

        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."

@@ -195,7 +174,7 @@ class XVLAAdamWConfig(OptimizerConfig):
            # 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]] = [
+        param_groups = [
            {
                "params": vlm_group,
                "lr": self.lr * 0.1,
@@ -245,25 +224,19 @@ class MultiAdamConfig(OptimizerConfig):
    grad_clip_norm: float = 10.0
    optimizer_groups: dict[str, dict[str, Any]] = field(default_factory=dict)

-    def build(self, params: OptimizerParams) -> dict[str, torch.optim.Optimizer]:
+    def build(self, params_dict: dict[str, list]) -> dict[str, torch.optim.Optimizer]:
        """Build multiple Adam optimizers.

        Args:
-            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).
+            params_dict: Dictionary mapping parameter group names to lists of parameters
+                         The keys should match the keys in optimizer_groups

        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, group_params in params.items():
+        for name, params in params_dict.items():
            # Get group-specific hyperparameters or use defaults
            group_config = self.optimizer_groups.get(name, {})

@@ -275,7 +248,7 @@ class MultiAdamConfig(OptimizerConfig):
                "weight_decay": group_config.get("weight_decay", self.weight_decay),
            }

-            optimizers[name] = torch.optim.Adam(group_params, **optimizer_kwargs)
+            optimizers[name] = torch.optim.Adam(params, **optimizer_kwargs)

        return optimizers

@@ -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 | None
+    num_warmup_steps: int

    @property
    def type(self) -> str:
@@ -21,8 +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
+from .wall_x.configuration_wall_x import WallXConfig as WallXConfig

 __all__ = [
    "ACTConfig",
@@ -30,7 +30,6 @@ __all__ = [
    "PI0Config",
    "PI05Config",
    "SmolVLAConfig",
-    "SARMConfig",
    "TDMPCConfig",
    "VQBeTConfig",
    "GrootConfig",
@@ -50,7 +50,6 @@ class ACTPolicy(PreTrainedPolicy):
    def __init__(
        self,
        config: ACTConfig,
-        **kwargs,
    ):
        """
        Args:
@@ -56,7 +56,6 @@ class DiffusionPolicy(PreTrainedPolicy):
    def __init__(
        self,
        config: DiffusionConfig,
-        **kwargs,
    ):
        """
        Args:
@@ -37,13 +37,12 @@ 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.policies.wall_x.configuration_wall_x import WallXConfig
 from lerobot.processor import PolicyAction, PolicyProcessorPipeline
 from lerobot.processor.converters import (
    batch_to_transition,
@@ -107,10 +106,6 @@ 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

@@ -349,14 +344,6 @@ 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

@@ -374,7 +361,7 @@ def make_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

@@ -464,47 +451,11 @@ def make_policy(
        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 not cfg.pretrained_path and cfg.use_peft:
-        raise ValueError(
-            "Instantiating a policy with `use_peft=True` without a checkpoint is not supported since that requires "
-            "the PEFT config parameters to be set. For training with PEFT, see `lerobot_train.py` on how to do that."
-        )
-
-    if cfg.pretrained_path and not cfg.use_peft:
+    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).
        kwargs["pretrained_name_or_path"] = cfg.pretrained_path
        policy = policy_cls.from_pretrained(**kwargs)
-    elif cfg.pretrained_path and cfg.use_peft:
-        # Load a pretrained PEFT model on top of the policy. The pretrained path points to the folder/repo
-        # of the adapter and the adapter's config contains the path to the base policy. So we need the
-        # adapter config first, then load the correct policy and then apply PEFT.
-        from peft import PeftConfig, PeftModel
-
-        logging.info("Loading policy's PEFT adapter.")
-
-        peft_pretrained_path = cfg.pretrained_path
-        peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
-
-        kwargs["pretrained_name_or_path"] = peft_config.base_model_name_or_path
-        if not kwargs["pretrained_name_or_path"]:
-            # This means that there's a bug or we trained a policy from scratch using PEFT.
-            # It is more likely that this is a bug so we'll raise an error.
-            raise ValueError(
-                "No pretrained model name found in adapter config. Can't instantiate the pre-trained policy on which "
-                "the adapter was trained."
-            )
-
-        policy = policy_cls.from_pretrained(**kwargs)
-        policy = PeftModel.from_pretrained(policy, peft_pretrained_path, config=peft_config)
-
    else:
        # Make a fresh policy.
        policy = policy_cls(**kwargs)
@@ -49,7 +49,7 @@ class GrootPolicy(PreTrainedPolicy):
    name = "groot"
    config_class = GrootConfig

-    def __init__(self, config: GrootConfig, **kwargs):
+    def __init__(self, config: GrootConfig):
        """Initialize Groot policy wrapper."""
        super().__init__(config)
        config.validate_features()
@@ -907,7 +907,6 @@ class PI0Policy(PreTrainedPolicy):
    def __init__(
        self,
        config: PI0Config,
-        **kwargs,
    ):
        """
        Args:
@@ -1236,15 +1235,9 @@ class PI0Policy(PreTrainedPolicy):

        return actions

-    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.
+    def forward(self, batch: dict[str, Tensor]) -> 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}"]
@@ -1258,17 +1251,11 @@ 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(),
        }

-        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
+        return loss, loss_dict
@@ -880,7 +880,6 @@ class PI05Policy(PreTrainedPolicy):
    def __init__(
        self,
        config: PI05Config,
-        **kwargs,
    ):
        """
        Args:
@@ -1210,15 +1209,9 @@ class PI05Policy(PreTrainedPolicy):

        return actions

-    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.
+    def forward(self, batch: dict[str, Tensor]) -> 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}"]
@@ -1232,17 +1225,11 @@ 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(),
        }

-        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
+        return loss, loss_dict
@@ -206,7 +206,6 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
    def push_model_to_hub(
        self,
        cfg: TrainPipelineConfig,
-        peft_model=None,
    ):
        api = HfApi()
        repo_id = api.create_repo(
@@ -217,14 +216,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
        with TemporaryDirectory(ignore_cleanup_errors=True) as tmp:
            saved_path = Path(tmp) / repo_id

-            if peft_model is not None:
-                # Since PEFT just forwards calls to `push_model_to_hub`, `self` is not the PeftModel wrapper
-                # but the actual policy which is why we need the PEFT model passed to us to save the adapter.
-                # That also means that we need to store the policy config ourselves since PEFT can't.
-                peft_model.save_pretrained(saved_path)
-                self.config.save_pretrained(saved_path)
-            else:
-                self.save_pretrained(saved_path)  # Calls _save_pretrained and stores model tensors
+            self.save_pretrained(saved_path)  # Calls _save_pretrained and stores model tensors

            card = self.generate_model_card(
                cfg.dataset.repo_id, self.config.type, self.config.license, self.config.tags
@@ -1,14 +0,0 @@
-## Paper
-
-https://arxiv.org/abs/2509.25358
-
-## Citation
-
-```bibtex
-@article{chen2025sarm,
-  title={SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation},
-  author={Chen, Qianzhong and Yu, Justin and Schwager, Mac and Abbeel, Pieter and Shentu, Yide and Wu, Philipp},
-  journal={arXiv preprint arXiv:2509.25358},
-  year={2025}
-}
-```
@@ -1,870 +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.
-
-"""
-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()
@@ -1,248 +0,0 @@
-#!/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
@@ -1,793 +0,0 @@
-#!/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)
@@ -1,518 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""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,
-        ),
-    )
@@ -1,295 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-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])
@@ -231,7 +231,6 @@ class SmolVLAPolicy(PreTrainedPolicy):
    def __init__(
        self,
        config: SmolVLAConfig,
-        **kwargs,
    ):
        """
        Args:
@@ -353,19 +352,8 @@ 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, 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
-        """
+    def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> dict[str, Tensor]:
+        """Do a full training forward pass to compute the loss"""
        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])
@@ -389,16 +377,11 @@ class SmolVLAPolicy(PreTrainedPolicy):
        losses = losses[:, :, : self.config.max_action_dim]
        loss_dict["losses_after_rm_padding"] = losses.clone()

-        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
+        # For backward pass
+        loss = losses.mean()
+        # For backward pass
+        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
@@ -65,7 +65,6 @@ class TDMPCPolicy(PreTrainedPolicy):
    def __init__(
        self,
        config: TDMPCConfig,
-        **kwargs,
    ):
        """
        Args:
@@ -231,20 +231,11 @@ 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}
-
-    # 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):
+    if not provided_visuals.issubset(expected_visuals):
        raise_feature_mismatch_error(provided_visuals, expected_visuals)
@@ -47,7 +47,6 @@ class VQBeTPolicy(PreTrainedPolicy):
    def __init__(
        self,
        config: VQBeTConfig | None = None,
-        **kwargs,
    ):
        """
        Args:
@@ -6,13 +6,12 @@ This repository contains the Hugging Face port of **WALL-OSS**, a Vision-Languag

 ## 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      |
-
+| 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
@@ -33,3 +32,4 @@ If you use this work, please cite:
 ## License

 This port follows the **Apache 2.0 License**.
+
@@ -15,5 +15,7 @@
 # limitations under the License.

 from .configuration_wall_x import WallXConfig
+from .modeling_wall_x import WallXPolicy
+from .processor_wall_x import make_wall_x_pre_post_processors

 __all__ = ["WallXConfig", "WallXPolicy", "make_wall_x_pre_post_processors"]
@@ -49,19 +49,17 @@ class WallXConfig(PreTrainedConfig):
        }
    )

-    # ==================== Action Prediction ====================
+    # ==================== Action Prediction ====================    
    # Pretrained model paths
    pretrained_name_or_path: str = "x-square-robot/wall-oss-flow"

-    # Tokenizer settings
-    action_tokenizer_path: str | None = "physical-intelligence/fast"
-
    # Action prediction mode: "diffusion" or "fast"
    prediction_mode: str = "diffusion"

-    # Attention Implementation, options: "eager", "flash_attention_2", "sdpa"
-    # NOTE: flash-attn==2.7.4.post1 is required for flash_attention_2 implementation
-    attn_implementation: str = "eager"
+    # Tokenizer settings
+    use_fast_tokenizer: bool = False  # True: train FAST, False: train Flow
+    action_tokenizer_path: str | None = None  # Path to action tokenizer (for FAST mode)
+

    # ==================== Optimizer Presets ====================
    optimizer_lr: float = 2e-5
@@ -85,16 +83,15 @@ class WallXConfig(PreTrainedConfig):
            )

        if self.prediction_mode not in ["diffusion", "fast"]:
-            raise ValueError(f"prediction_mode must be 'diffusion' or 'fast', got {self.prediction_mode}")
+            raise ValueError(
+                f"prediction_mode must be 'diffusion' or 'fast', got {self.prediction_mode}"
+            )

-        # Assign use_fast_tokenizer based on prediction_mode
-        if self.prediction_mode == "fast":
-            self.use_fast_tokenizer = True
-        elif self.prediction_mode == "diffusion":
-            self.use_fast_tokenizer = False
-            self.action_tokenizer_path = None  # disable action tokenizer for diffusion mode
+        # Sync prediction_mode with use_fast_tokenizer
+        if self.use_fast_tokenizer:
+            self.prediction_mode = "fast"
        else:
-            raise ValueError(f"prediction_mode must be 'diffusion' or 'fast', got {self.prediction_mode}")
+            self.prediction_mode = "diffusion"

    def validate_features(self) -> None:
        """Validate and set up input/output features."""
@@ -18,6 +18,8 @@
 Wall-X Constants and Configuration Data.
 """

+from lerobot.utils.constants import OBS_STATE, OBS_IMAGES, ACTION
+
 CAMERA_NAME_MAPPING = {
    "face_view": "front view",
    "left_wrist_view": "left wrist view",
@@ -38,4 +40,4 @@ PRIORITY_ORDER = None
 GENERATE_SUBTASK_RATIO = 0.0
 MODEL_TYPE = "qwen2_5"

-TOKENIZER_MAX_LENGTH = 768
+TOKENIZER_MAX_LENGTH = 768
@@ -33,8 +33,6 @@ from lerobot.processor import (
 )
 from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
 from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
-
-
 def make_wall_x_pre_post_processors(
    config: WallXConfig,
    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
@@ -77,7 +75,9 @@ def make_wall_x_pre_post_processors(

    output_steps = [
        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+            features=config.output_features,
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats
        ),
        DeviceProcessorStep(device="cpu"),
    ]
@@ -123,7 +123,9 @@ class WallXTaskProcessor(ComplementaryDataProcessorStep):
                new_complementary_data["task"] = f"{task}."
        elif isinstance(task, list) and all(isinstance(t, str) for t in task):
            # List of strings: format each
-            new_complementary_data["task"] = [t if t.endswith(".") else f"{t}." for t in task]
+            new_complementary_data["task"] = [
+                t if t.endswith(".") else f"{t}." for t in task
+            ]

        return new_complementary_data

@@ -209,7 +209,6 @@ class Qwen2_5_VLConfig(PretrainedConfig):
        self.use_sliding_window = use_sliding_window
        self.sliding_window = sliding_window
        self.max_window_layers = max_window_layers
-        self.layer_types = ["dense"] * num_hidden_layers

        # for backward compatibility
        if num_key_value_heads is None:
@@ -232,6 +231,11 @@ class Qwen2_5_VLConfig(PretrainedConfig):
        self.attention_moe = attention_moe
        self.mlp_moe = mlp_moe

+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        # and change type from 'mrope' to 'default' because `mrope` does defeault RoPE calculations
+        # one can set it to "linear"/"dynamic" etc. to have scaled RoPE
+        # TODO: @raushan update config in the hub
        if self.rope_scaling is not None and "type" in self.rope_scaling:
            if self.rope_scaling["type"] == "mrope":
                self.rope_scaling["type"] = "default"
@@ -240,9 +244,5 @@ class Qwen2_5_VLConfig(PretrainedConfig):

        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

-    @property
-    def text_config(self):
-        return self
-

 __all__ = ["Qwen2_5_VLConfig"]
@@ -21,11 +21,12 @@ Contains data processing utilities, text formatting functions, and helper classe
 for the Wall-X cross-embodiment robotic control model.
 """

+import json
 import random
 import re
 from collections import OrderedDict
 from dataclasses import dataclass, field
-from typing import Any
+from typing import Any, Dict, List, Optional, Tuple, Union

 import torch
 from transformers import BatchFeature
@@ -46,11 +47,11 @@ class X2RDataProcessingConfig:
    """

    # Action prediction configuration
-    predict_action_keys: list[str] = field(default_factory=list)
-    obs_action_keys: list[str] = field(default_factory=list)
+    predict_action_keys: List[str] = field(default_factory=list)
+    obs_action_keys: List[str] = field(default_factory=list)

    # Image resolution settings for different views
-    resolution: dict[str, int] = field(
+    resolution: Dict[str, int] = field(
        default_factory=lambda: {
            "face_view": -1,
            "left_wrist_view": 128,
@@ -63,7 +64,7 @@ class X2RDataProcessingConfig:
    split_seed: int = 42

    # Instruction handling
-    priority_order: dict[str, float] | None = None
+    priority_order: Optional[Dict[str, float]] = None

    # Vision model parameters
    model_type: str = "qwen2_5"
@@ -77,9 +78,11 @@ class X2RDataProcessingConfig:
        """Post-initialization validation and setup."""
        # Validate train/test split
        if not 0 < self.train_test_split < 1:
-            raise ValueError(f"train_test_split must be between 0 and 1, got {self.train_test_split}")
+            raise ValueError(
+                f"train_test_split must be between 0 and 1, got {self.train_test_split}"
+            )

-    def as_dict(self) -> dict:
+    def as_dict(self) -> Dict:
        """Convert configuration to dictionary format.

        Returns:
@@ -103,15 +106,14 @@ class X2RDataProcessingConfig:
                raise ValueError(f"Unknown configuration parameter: {key}")
        return self

-
 def preprocesser_call(
    processor,
-    images: list | Any | None = None,
-    text: str | list[str] | None = None,
-    videos: list | Any | None = None,
-    padding: bool | str = False,
-    truncation: bool | None = None,
-    max_length: int | None = None,
+    images: Optional[Union[List, Any]] = None,
+    text: Optional[Union[str, List[str]]] = None,
+    videos: Optional[Union[List, Any]] = None,
+    padding: Union[bool, str] = False,
+    truncation: Optional[bool] = None,
+    max_length: Optional[int] = None,
    return_tensors: str = "pt",
 ) -> BatchFeature:
    """Unified preprocessing function for Wall-X model handling text, image and video inputs.
@@ -144,7 +146,9 @@ def preprocesser_call(
    """
    # Process image inputs
    if images is not None and len(images) > 0:
-        image_inputs = processor.image_processor(images=images, videos=None, return_tensors=return_tensors)
+        image_inputs = processor.image_processor(
+            images=images, videos=None, return_tensors=return_tensors
+        )
        image_grid_thw = image_inputs["image_grid_thw"]
    else:
        image_inputs = {}
@@ -152,7 +156,9 @@ def preprocesser_call(

    # Process video inputs
    if videos is not None:
-        videos_inputs = processor.image_processor(images=None, videos=videos, return_tensors=return_tensors)
+        videos_inputs = processor.image_processor(
+            images=None, videos=videos, return_tensors=return_tensors
+        )
        video_grid_thw = videos_inputs["video_grid_thw"]
    else:
        videos_inputs = {}
@@ -178,7 +184,9 @@ def preprocesser_call(
                    break
                # Replace image placeholder with actual token count
                token_count = image_grid_thw[index].prod() // merge_length
-                text[i] = text[i].replace("<|image_pad|>", "<|placeholder|>" * token_count, 1)
+                text[i] = text[i].replace(
+                    "<|image_pad|>", "<|placeholder|>" * token_count, 1
+                )
                index += 1
            text[i] = text[i].replace("<|placeholder|>", "<|image_pad|>")

@@ -190,7 +198,9 @@ def preprocesser_call(
            while "<|video_pad|>" in text[i]:
                # Replace video placeholder with actual token count
                token_count = video_grid_thw[index].prod() // merge_length
-                text[i] = text[i].replace("<|video_pad|>", "<|placeholder|>" * token_count, 1)
+                text[i] = text[i].replace(
+                    "<|video_pad|>", "<|placeholder|>" * token_count, 1
+                )
                index += 1
            text[i] = text[i].replace("<|placeholder|>", "<|video_pad|>")

@@ -212,7 +222,9 @@ def preprocesser_call(
    labels = torch.full_like(text_inputs.input_ids, -100)
    assistant_marker = "<|im_start|>assistant\n"
    im_end_token_id = processor.tokenizer.convert_tokens_to_ids("<|im_end|>")
-    assistant_tokens = processor.tokenizer("<|im_start|>assistant\n", add_special_tokens=False).input_ids
+    assistant_tokens = processor.tokenizer(
+        "<|im_start|>assistant\n", add_special_tokens=False
+    ).input_ids

    for i in range(len(text)):
        assistant_regions = []
@@ -238,7 +250,9 @@ def preprocesser_call(
            # From second part onwards, each part starts with assistant response
            for k in range(current_pos + 1, len(text_inputs.input_ids[i])):
                if text_inputs.input_ids[i][k] == im_end_token_id:
-                    assistant_regions.append((current_pos + len(assistant_tokens), k + 2))
+                    assistant_regions.append(
+                        (current_pos + len(assistant_tokens), k + 2)
+                    )
                    break
            current_pos += len(part_tokens) + 3

@@ -331,7 +345,7 @@ def process_grounding_points(
                coords = [new_x1, new_y1, new_x2, new_y2]

            # Return processed point tag
-            return f"<point>[{', '.join(map(str, coords))}]</point>"
+            return f'<point>[{", ".join(map(str, coords))}]</point>'

        except (ValueError, TypeError):
            # Return original content if processing fails
@@ -343,10 +357,10 @@ def process_grounding_points(


 def get_frame_instruction(
-    instruction_info: dict[str, Any],
-    frame_idx: int | None = None,
-    truncate_keys: list[str] | None = None,
-) -> tuple[dict[str, Any], int | None]:
+    instruction_info: Dict[str, Any],
+    frame_idx: Optional[int] = None,
+    truncate_keys: Optional[List[str]] = None,
+) -> Tuple[Dict[str, Any], Optional[int]]:
    """Extract frame-specific instruction from instruction dictionary.

    Args:
@@ -375,7 +389,11 @@ def get_frame_instruction(
                start_frame, end_frame = map(int, frame_range.split(" "))
                if start_frame <= frame_idx < end_frame or (start_frame == frame_idx):
                    instruction_for_frame[key] = frame_instruction
-                    if truncate_keys is not None and split_end is None and key in truncate_keys:
+                    if (
+                        truncate_keys is not None
+                        and split_end is None
+                        and key in truncate_keys
+                    ):
                        split_end = end_frame + 1
                    break
        else:
@@ -385,7 +403,7 @@ def get_frame_instruction(


 def get_task_instruction(
-    frame_instruction_info: dict[str, Any], priority_order: OrderedDict | None = None
+    frame_instruction_info: Dict[str, Any], priority_order: Optional[OrderedDict] = None
 ) -> str:
    """Construct task instruction from available instruction fields using priority sampling.

@@ -433,13 +451,13 @@ def get_task_instruction(


 def get_wallx_normal_text(
-    instruction_info: dict[str, Any],
+    instruction_info: Dict[str, Any],
    action_chunk_size: int,
    frame_idx: int,
-    priority_order: OrderedDict | None = None,
-    img_keys: list[str] | None = None,
+    priority_order: Optional[OrderedDict] = None,
+    img_keys: Optional[List[str]] = None,
    generate_subtask_ratio: float = 0.0,
-) -> tuple[str, bool]:
+) -> Tuple[str, bool]:
    """Construct complete multimodal prompt text for Wall-X model.

    Formats input using special tokens including:
@@ -471,7 +489,9 @@ def get_wallx_normal_text(
    action_fast_symbol = "<|action_fast|>"

    # System prologue
-    prologue = f"{role_start_symbol}system\nYou are a helpful assistant.{role_end_symbol}\n"
+    prologue = (
+        f"{role_start_symbol}system\nYou are a helpful assistant.{role_end_symbol}\n"
+    )

    # User request with observation
    user_request = f"{role_start_symbol}user\nObservation:"
@@ -482,7 +502,9 @@ def get_wallx_normal_text(
    user_request += "\nInstruction:"

    # Get frame-specific instruction
-    frame_instruction_info, _ = get_frame_instruction(instruction_info, frame_idx=frame_idx)
+    frame_instruction_info, _ = get_frame_instruction(
+        instruction_info, frame_idx=frame_idx
+    )

    generate_subtask = False
    priority_keys = ["subtask_generation", "distribute"]
@@ -503,11 +525,15 @@ def get_wallx_normal_text(
                output_instruction = frame_instruction_info[key]
                break

-        assistant_output = f"{role_start_symbol}assistant\n{output_instruction}\n{role_end_symbol}"
+        assistant_output = (
+            f"{role_start_symbol}assistant\n{output_instruction}\n{role_end_symbol}"
+        )
        generate_subtask = True
    else:
        # Generate actions
-        instruction = get_task_instruction(frame_instruction_info, priority_order=priority_order)
+        instruction = get_task_instruction(
+            frame_instruction_info, priority_order=priority_order
+        )
        text_prompt = f"\nPredict the next action in robot action.\nProprioception: {propri_symbol}\n"
        user_message = f"{user_request} {instruction}{text_prompt}{role_end_symbol}\n"
        assistant_output = f"{role_start_symbol}assistant\n{action_fast_symbol}{role_end_symbol}\n{action_symbol * action_chunk_size}"
@@ -515,8 +541,7 @@ def get_wallx_normal_text(
    complete_text = prologue + user_message + assistant_output
    return complete_text, generate_subtask

-
-def img_key_mapping(img_keys: list[str]) -> list[str]:
+def img_key_mapping(img_keys: List[str]) -> List[str]:
    """Map image keys to camera names.

    Args:
@@ -531,15 +556,16 @@ def img_key_mapping(img_keys: list[str]) -> list[str]:
        if key in CAMERA_NAME_MAPPING:
            key = CAMERA_NAME_MAPPING[key]
        else:
-            if "view" in key:
-                key = key.replace("_", " ")
+            if 'view' in key:
+                key = key.replace('_', ' ')
            else:
                key = key + " view"
        processed_img_keys.append(key)
    return processed_img_keys

-
-def get_action_tokens(normalized_actions: torch.Tensor | list, action_tokenizer) -> list[list[str]]:
+def get_action_tokens(
+    normalized_actions: Union[torch.Tensor, List], action_tokenizer
+) -> List[List[str]]:
    """Convert normalized actions to action token strings.

    Args:
@@ -565,9 +591,8 @@ def get_action_tokens(normalized_actions: torch.Tensor | list, action_tokenizer)


 def pad_action_token_strs(
-    actions_token_lists: list[list[str]],
-    pad_token: str = "<|endoftext|>",  # nosec B107
-) -> list[str]:
+    actions_token_lists: List[List[str]], pad_token: str = "<|endoftext|>"
+) -> List[str]:
    """Pad action token lists to same length and join as strings.

    Args:
@@ -581,32 +606,39 @@ def pad_action_token_strs(
    padded_action_strs = []

    for tokens in actions_token_lists:
-        padded_tokens = tokens + ["<|im_end|>\n"] + [pad_token] * (max_len - len(tokens))
+        padded_tokens = (
+            tokens + ["<|im_end|>\n"] + [pad_token] * (max_len - len(tokens))
+        )
        padded_action_strs.append("".join(padded_tokens))

    return padded_action_strs


 def replace_action_token(
-    text: list[str],
-    norm_action: torch.Tensor | None,
+    text: List[str],
+    norm_action: Optional[torch.Tensor],
    action_tokenizer,
-    dof_masks: torch.Tensor | None = None,
-) -> list[str]:
+    dof_masks: Optional[torch.Tensor] = None,
+) -> List[str]:
    """Replace action placeholders in text with actual action tokens.

    Args:
        text: List of text strings with action placeholders
        norm_action: Normalized action tensors
        action_tokenizer: Tokenizer for converting actions to tokens
+        dataset_names: Names of datasets for each sample
        dof_masks: Masks for degrees of freedom

    Returns:
        List of text strings with action tokens replaced
    """
+    # Filter out multimodal dataset names
    if action_tokenizer is not None and norm_action is not None:
        # Extract actions based on chunk sizes and DOF masks
-        norm_action = [action[:32, dof_masks[i, 0].bool()] for i, action in enumerate(norm_action)]
+        norm_action = [
+            action[: 32, dof_masks[i, 0].bool()]
+            for i, action in enumerate(norm_action)
+        ]

        # Convert to action tokens and pad
        actions_fast_tokens = get_action_tokens(norm_action, action_tokenizer)
@@ -629,3 +661,4 @@ def replace_action_token(
        text = [t.replace("<|action_fast|><|im_end|>\n", "") for t in text]

    return text
+
@@ -273,7 +273,7 @@ class XVLAPolicy(PreTrainedPolicy):
    config_class = XVLAConfig
    name = "xvla"

-    def __init__(self, config: XVLAConfig, **kwargs):
+    def __init__(self, config: XVLAConfig):
        super().__init__(config)
        config.validate_features()
        florence_config = config.get_florence_config()
@@ -170,9 +170,8 @@ def _extract_complementary_data(batch: dict[str, Any]) -> dict[str, Any]:
    task_key = {"task": batch["task"]} if "task" in batch else {}
    index_key = {"index": batch["index"]} if "index" in batch else {}
    task_index_key = {"task_index": batch["task_index"]} if "task_index" in batch else {}
-    episode_index_key = {"episode_index": batch["episode_index"]} if "episode_index" in batch else {}

-    return {**pad_keys, **task_key, **index_key, **task_index_key, **episode_index_key}
+    return {**pad_keys, **task_key, **index_key, **task_index_key}


 def create_transition(
@@ -112,32 +112,7 @@ class WandBLogger:
        artifact_name = f"{self._group}-{step_id}"
        artifact_name = get_safe_wandb_artifact_name(artifact_name)
        artifact = self._wandb.Artifact(artifact_name, type="model")
-        pretrained_model_dir = checkpoint_dir / PRETRAINED_MODEL_DIR
-
-        # Check if this is a PEFT model (has adapter files instead of model.safetensors)
-        adapter_model_file = pretrained_model_dir / "adapter_model.safetensors"
-        standard_model_file = pretrained_model_dir / SAFETENSORS_SINGLE_FILE
-
-        if adapter_model_file.exists():
-            # PEFT model: add adapter files and configs
-            artifact.add_file(adapter_model_file)
-            adapter_config_file = pretrained_model_dir / "adapter_config.json"
-            if adapter_config_file.exists():
-                artifact.add_file(adapter_config_file)
-            # Also add the policy config which is needed for loading
-            config_file = pretrained_model_dir / "config.json"
-            if config_file.exists():
-                artifact.add_file(config_file)
-        elif standard_model_file.exists():
-            # Standard model: add the single safetensors file
-            artifact.add_file(standard_model_file)
-        else:
-            logging.warning(
-                f"No {SAFETENSORS_SINGLE_FILE} or adapter_model.safetensors found in {pretrained_model_dir}. "
-                "Skipping model artifact upload to WandB."
-            )
-            return
-
+        artifact.add_file(checkpoint_dir / PRETRAINED_MODEL_DIR / SAFETENSORS_SINGLE_FILE)
        self._wandb.log_artifact(artifact)

    def log_dict(
@@ -0,0 +1,20 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .config_openarms_follower import OpenArmsFollowerConfig
+from .openarms_follower import OpenArmsFollower
+
+__all__ = ["OpenArmsFollower", "OpenArmsFollowerConfig"]
@@ -0,0 +1,118 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+from typing import Dict, Optional
+
+from lerobot.cameras import CameraConfig
+from lerobot.motors.damiao.tables import MotorType
+
+from ..config import RobotConfig
+
+
+@RobotConfig.register_subclass("openarms_follower")
+@dataclass
+class OpenArmsFollowerConfig(RobotConfig):
+    """Configuration for the OpenArms follower robot with Damiao motors."""
+    
+    # CAN interfaces - one per arm
+    # Right arm CAN interface (e.g., "can0")
+    # Left arm CAN interface (e.g., "can1")
+    # Linux: "can0", "can1", etc.
+    # macOS: "/dev/cu.usbmodem*" (serial device)
+    port_right: str = "can0"    # CAN interface for right arm
+    port_left: str = "can1"     # CAN interface for left arm
+    
+    # CAN interface type: "socketcan" (Linux), "slcan" (macOS/serial), or "auto" (auto-detect)
+    can_interface: str = "socketcan"
+    
+    # CAN FD settings (OpenArms uses CAN FD by default)
+    use_can_fd: bool = True
+    can_bitrate: int = 1000000      # Nominal bitrate (1 Mbps)
+    can_data_bitrate: int = 5000000  # Data bitrate for CAN FD (5 Mbps)
+    
+    # Whether to disable torque when disconnecting
+    disable_torque_on_disconnect: bool = True
+    
+    # Safety limit for relative target positions
+    # Set to a positive scalar for all motors, or a dict mapping motor names to limits
+    max_relative_target: Optional[float | Dict[str, float]] = None
+    
+    # Camera configurations
+    cameras: Dict[str, CameraConfig] = field(default_factory=dict)
+    
+    # Motor configuration for OpenArms (7 DOF per arm)
+    # Maps motor names to (send_can_id, recv_can_id, motor_type)
+    # Based on: https://docs.openarm.dev/software/setup/configure-test
+    # OpenArms uses 4 types of motors:
+    # - DM8009 (DM-J8009P-2EC) for shoulders (high torque)
+    # - DM4340P and DM4340 for shoulder rotation and elbow
+    # - DM4310 (DM-J4310-2EC V1.1) for wrist and gripper
+    motor_config: Dict[str, tuple[int, int, str]] = field(default_factory=lambda: {
+        "joint_1": (0x01, 0x11, "dm8009"),   # J1 - Shoulder pan (DM8009)
+        "joint_2": (0x02, 0x12, "dm8009"),   # J2 - Shoulder lift (DM8009)
+        "joint_3": (0x03, 0x13, "dm4340"),   # J3 - Shoulder rotation (DM4340)
+        "joint_4": (0x04, 0x14, "dm4340"),   # J4 - Elbow flex (DM4340)
+        "joint_5": (0x05, 0x15, "dm4310"),   # J5 - Wrist roll (DM4310)
+        "joint_6": (0x06, 0x16, "dm4310"),   # J6 - Wrist pitch (DM4310)
+        "joint_7": (0x07, 0x17, "dm4310"),   # J7 - Wrist rotation (DM4310)
+        "gripper": (0x08, 0x18, "dm4310"),   # J8 - Gripper (DM4310)
+    })
+    
+    # MIT control parameters for position control (used in send_action)
+    # List of 8 values: [joint_1, joint_2, joint_3, joint_4, joint_5, joint_6, joint_7, gripper]
+    position_kp: list[float] = field(default_factory=lambda: [240.0, 240.0, 240.0, 240.0, 24.0, 31.0, 25.0, 25.0])
+    position_kd: list[float] = field(default_factory=lambda: [3.0, 3.0, 3.0, 3.0, 0.2, 0.2, 0.2, 0.2])
+    
+    # Damping gains for stability when applying torque compensation (gravity/friction)
+    # Used when kp=0 and only torque is applied
+    damping_kd: list[float] = field(default_factory=lambda: [0.5, 0.5, 0.5, 0.5, 0.1, 0.1, 0.1, 0.1])
+    
+    # Friction model parameters: τ_fric(ω) = Fo + Fv·ω + Fc·tanh(k·ω)
+    # From OpenArms config/follower.yaml
+    friction_fc: list[float] = field(default_factory=lambda: [0.306, 0.306, 0.40, 0.166, 0.050, 0.093, 0.172, 0.0512])  # Coulomb friction [Nm]
+    friction_k: list[float] = field(default_factory=lambda: [28.417, 28.417, 29.065, 130.038, 151.771, 242.287, 7.888, 4.000])  # tanh steepness
+    friction_fv: list[float] = field(default_factory=lambda: [0.063, 0.0630, 0.604, 0.813, 0.029, 0.072, 0.084, 0.084])  # Viscous friction [Nm·s/rad]
+    friction_fo: list[float] = field(default_factory=lambda: [0.088, 0.088, 0.008, -0.058, 0.005, 0.009, -0.059, -0.050])  # Offset torque [Nm]
+    
+    # Calibration parameters
+    calibration_mode: str = "manual"  # "manual" or "auto"
+    zero_position_on_connect: bool = False  # Set zero position on connect
+    
+    # Joint limits for position clipping (degrees)
+    # Format: [min, max] for each joint
+    # These limits clip commands in send_action to prevent mechanical damage
+    joint_limits_right: Dict[str, tuple[float, float]] = field(default_factory=lambda: {
+        "joint_1": (-75.0, 75.0),
+        "joint_2": (-9.0, 90.0),
+        "joint_3": (-85.0, 85.0),
+        "joint_4": (0.0, 135.0),
+        "joint_5": (-85.0, 85.0),
+        "joint_6": (-40.0, 40.0),
+        "joint_7": (-80.0, 80.0),
+        "gripper": (-65.0, 0.0),
+    })
+    
+    joint_limits_left: Dict[str, tuple[float, float]] = field(default_factory=lambda: {
+        "joint_1": (-75.0, 75.0),
+        "joint_2": (-90.0, 9.0),
+        "joint_3": (-85.0, 85.0),
+        "joint_4": (0.0, 135.0),
+        "joint_5": (-85.0, 85.0),
+        "joint_6": (-40.0, 40.0),
+        "joint_7": (-80.0, 80.0),
+        "gripper": (-65.0, 0.0),
+    })
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Michel Aractingi	b6050e6242	Merge pr-2593 (wallx support) into openarms_tmp_rebase	2025-12-17 16:52:11 +01:00
Michel Aractingi	08d2ed8015	lerobot dataset fix	2025-12-17 16:46:43 +01:00
Michel Aractingi	4bcd14b8de	add evaluate_with_rtc script	2025-12-17 16:46:43 +01:00
Michel Aractingi	c34935090d	integrate delete button openarm UI (#2535 ) * add visualize_dataset call from `lerobot_dataset_viz` in web record server * add delete button * fixes * remove viz * unused import	2025-12-17 16:46:43 +01:00
CarolinePascal	9cfd56587e	fix(num processes)	2025-12-17 16:46:43 +01:00
Caroline Pascal	ff8584a025	fix(os version) Signed-off-by: Caroline Pascal <caroline8.pascal@gmail.com>	2025-12-17 16:46:43 +01:00
Caroline Pascal	6bc1e5186a	fix(import os) Signed-off-by: Caroline Pascal <caroline8.pascal@gmail.com>	2025-12-17 16:46:43 +01:00
Caroline Pascal	69dc8165ae	fix(max workers) Signed-off-by: Caroline Pascal <caroline8.pascal@gmail.com>	2025-12-17 16:46:42 +01:00
CarolinePascal	021bca2ad9	feat(multi-processes): adding support for multiprocess encoding	2025-12-17 16:46:42 +01:00
CarolinePascal	4e0ee0d643	feat(preset): adding encoding preset	2025-12-17 16:46:42 +01:00
croissant	0a8aa85871	ruse video datasets	2025-12-17 16:46:42 +01:00
croissant	76ddd8b948	use image datasets and change ui	2025-12-17 16:46:42 +01:00
croissant	bf08733068	frontend set correct port openarms mini	2025-12-17 16:46:42 +01:00
croissant	e38f56c071	add default mini arms	2025-12-17 16:46:41 +01:00
croissant	19fe69dac0	add improv openarm mini	2025-12-17 16:46:41 +01:00
pepijn kooijmans	14319ee608	add openarms mini	2025-12-17 16:46:41 +01:00
croissant	9b04fd25b6	cam res	2025-12-17 16:46:41 +01:00
Pepijn	40e98ba690	fix calibration of gripper and add max clip positions for openarm for safety	2025-12-17 16:46:41 +01:00
pepijn kooijmans	894d65d58a	add openarms to setup motors	2025-12-17 16:46:41 +01:00
Pepijn	f58d508df2	cleanuo	2025-12-17 16:46:40 +01:00
Pepijn	e22b909e7c	Add mini openarms to test	2025-12-17 16:46:40 +01:00
croissant	09f1673cbf	add longer timeout	2025-12-17 16:46:40 +01:00
croissant	4744f99990	add timing debugging, foot pedal and eval script	2025-12-17 16:46:40 +01:00
croissant	01c1735739	add disable torque	2025-12-17 16:46:40 +01:00
croissant	6808a42455	add pid ramp	2025-12-17 16:46:40 +01:00
croissant	fff719cb4f	add web interface example	2025-12-17 16:46:39 +01:00
croissant	e2c00f6ed8	speedup	2025-12-17 16:46:39 +01:00
croissant	0f90db23c5	add full bimanual gravity comp	2025-12-17 16:46:39 +01:00
Michel Aractingi	96b192f2ae	Add gravity compensation to the openarms teleoperation (#2352 ) * adding first attempt at gcompensation to open arms * add teleop with gravity compensation script	2025-12-17 16:46:39 +01:00
Pepijn	ecdc34a699	faster canbus	2025-12-17 16:46:39 +01:00
croissant	fa6a2fb9b7	pos teleop	2025-12-17 16:46:39 +01:00
Pepijn	b011643dc9	add tests and debug	2025-12-17 16:46:38 +01:00
Pepijn	30c10c1c6e	Add damiao motors and open arm robot	2025-12-17 16:46:38 +01:00
Pepijn	56e2360072	add damiao	2025-12-17 16:46:38 +01:00
Pepijn	2ef1de78b6	Merge branch 'main' into main	2025-12-09 09:41:02 +01:00
Geoffrey19	2852b968b9	add wallx dependencies	2025-12-07 15:30:14 +08:00
Geoffrey19	56d20caa1e	fixed dtype bugs	2025-12-07 15:30:13 +08:00
Geoffrey19	b4a7586b27	reduce to least config and params & pass lerobot basic test	2025-12-07 15:30:13 +08:00
Geoffrey19	78995621fa	update the policy methods	2025-12-07 15:30:12 +08:00
Geoffrey19	5be8b6de6b	incorporate wallx model into lerobot	2025-12-07 15:30:11 +08:00
Geoffrey19	b185fa0f87	fix bugs in flow	2025-12-07 15:30:10 +08:00
vincentchen	73a6f20e58	support wallx	2025-12-07 15:29:23 +08:00