MotionTrans-Pi0

This repository contains the Pi0-VLA code for MotionTrans: Human VR Data Enable Motion-Level Learning for Robotic Manipulation Policies, which is a modification of the original Pi0 codebase from Physics Intelligence.

For more details of the whole project, please refer to the main MotionTrans repository.

Installation

Please follow the installation instructions in the original Pi0 repository.

Data Preparation

Please follow the data preparation instructions in the original MotionTrans repository. All data should be processed as .zarr files for Pi0-VLA training.

Training and Evaluation

bash scripts_exp/train_cotrain.sh
bash scripts_exp/eval.sh

Inference

First change the policy path in Line 89 of scripts\serve_policy.py, and then run:

bash scripts_exp/serve_policy.sh

This will launch a web server in local. You can then open the client following the instructions in MotionTrans repository to use Pi0-VLA to control the real robot.

Acknowledgment

We thanks Ruiqian Nai and Fanqi Lin for their great help on the development of this MotionTrans-Pi0-VLA codebase. This repository is based on the code from OneTwoVLA, OpenPi andUMI. We sincerely appreciate their contribution to the open-source community, which have significantly supported this project. We also sincerely thank our AI-collaborators ChatGPT, Kimi and Github Copilot !!

Citation

If you find this repository useful, please kindly acknowledge our work :

@article{yuan2025motiontrans,
  title={MotionTrans: Human VR Data Enable Motion-Level Learning for Robotic Manipulation Policies},
  author={Yuan, Chengbo and Zhou, Rui and Liu, Mengzhen and Hu, Yingdong and Wang, Shengjie and Yi, Li and Wen, Chuan and Zhang, Shanghang and Gao, Yang},
  journal={arXiv preprint arXiv:2509.17759},
  year={2025}
}