| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen3.5-4B |
| tags: |
| - robotics |
| - vision-language-action |
| - mr0 |
| - pytorch |
| - mmdit |
| --- |
| |
| # Meituan-Robotics-0 |
|
|
| [Technical Report](https://arxiv.org/abs/2611.08577) |
|
|
| MR0 continuous-action pretraining model with a Qwen3.5-4B vision-language backbone and an MMDiT action head. Adapt this pretrained VLA to downstream robot embodiments and tasks. |
|
|
| [Code](https://github.com/meituan/Meituan-Robotics-0) · [Getting started](https://github.com/meituan/Meituan-Robotics-0/blob/main/GETTING_STARTED.md) · [File checksums](SHA256SUMS) |
|
|
| | Property | Value | |
| |---|---| |
| | Framework | QwenMMDiT | |
| | Action dimensions | 20 | |
| | Action horizon | 16 | |
| | Weights | `checkpoints/model.pt` | |
| | Configuration | `config.full.yaml` | |
| | Normalization statistics | `dataset_statistics.json` | |
|
|
| ## Download and load |
|
|
| Install the MR0 code at commit [`88c66793`](https://github.com/meituan/Meituan-Robotics-0/commit/88c667933dfc941a71c9b142eb84b716ac4f0707) or later, following the Getting Started guide. This loader reads the complete VLA checkpoint without a separate Qwen weight file. Then download this repository: |
|
|
| ```python |
| import os |
| from pathlib import Path |
| from huggingface_hub import snapshot_download |
| |
| root = Path(snapshot_download( |
| repo_id="MeituanRobotics/Meituan-Robotics-0", |
| local_dir="Meituan-Robotics-0", |
| )).resolve() |
| os.environ["MR0_RELEASE_DIR"] = str(root) |
| |
| from mr0.model.framework.base_framework import baseframework |
| |
| model = baseframework.from_pretrained(root / "checkpoints/model.pt") |
| model = model.eval().to("cuda") |
| ``` |
|
|
| `checkpoints/model.pt` contains the complete Qwen3.5 backbone and MMDiT action head. `base_vlm/Qwen3.5-4B-Action/` contains only the matching tokenizer, processor, and architecture configuration; it is not a second model. The release also includes the VLA configuration and normalization statistics. Choose the normalization key in `dataset_statistics.json` that matches your robot embodiment. |
|
|
| The checkpoint bytes are unchanged. `release.json` records the model provenance and `SHA256SUMS` records file identities. For downstream training, adapt the dataset settings and set both `framework.qwenvl.checkpoint_path` and `trainer.pretrained_checkpoint` to the downloaded complete checkpoint. Set `framework.qwenvl.base_vlm` to its metadata directory. |
|
|
| ## File verification |
|
|
| [SHA256SUMS](SHA256SUMS) is the single SHA-256 checksum manifest for this release. It covers every repository file except the manifest itself, including weights, configuration, tokenizer/processor metadata, and documentation. |
|
|
| After downloading the complete repository, run from its root directory: |
|
|
| ```bash |
| sha256sum -c SHA256SUMS |
| ``` |
|
|
| ## Acknowledgements |
|
|
| We thank [Qwen](https://github.com/QwenLM), [StarVLA](https://github.com/starVLA/starVLA), and the [RoboTwin](https://github.com/RoboTwin-Platform/RoboTwin) community. |
|
|
