KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills

Weiji Xie^{* 1,2}, Jinrui Han^{* 1,2}, Jiakun Zheng^{* 1,3}, Huanyu Li^1,4, Xinzhe Liu^1,5, Jiyuan Shi¹, Weinan Zhang², Chenjia Bai^†1, Xuelong Li¹
* Equal Contribution  † Corresponding Author
¹Institute of Artificial Intelligence (TeleAI), China Telecom   ²Shanghai Jiao Tong University   ³East China University of Science and Technology   ⁴Harbin Institute of Technology   ⁵ShanghaiTech University

KungfuBot2: Learning Versatile Motion Skills for Humanoid Whole-Body Control

Jinrui Han ^1,2, Weiji Xie ^1,2, Jiakun Zheng ^1,3, Jiyuan Shi¹, Weinan Zhang², Ting Xiao², Chenjia Bai^†1,
† Corresponding Author
¹Institute of Artificial Intelligence (TeleAI), China Telecom   ²Shanghai Jiao Tong University   ³East China University of Science and Technology  

Demo

News

• [2025-10] Release support for general motion tracking.
• [2025-09] KungfuBot is accepted by NeurIPS 2025!
• [2025-06] We release the code and paper for PBHC.

Contents

• Demo
• News
• Contents
• About
• Usage
• Folder Structure
• Citation
• License
• Acknowledgements
• Contact

About

This is the official implementation of the paper KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills, supporting general motion tracking of the paper KungfuBot2: Learning Versatile Motion Skills for Humanoid Whole-Body Control.

Our paper introduces a physics-based control framework that enables humanoid robots to learn and reproduce challenging motions through multi-stage motion processing and adaptive policy training.

This repository includes:

• Motion processing pipeline
  • Collect human motion from various sources (video, LAFAN, AMASS, etc.) to a unified SMPL format (motion_source/)
  • Filter, correct and retarget human motion to the robot (smpl_retarget/)
  • Visualize and analyze the processed motions (smpl_vis/, robot_motion_process/)
• RL-based motion imitation framework (humanoidverse/)
  • Train the policy in IsaacGym
  • Deploy trained policies in MuJoCo for sim2sim verification. The framework is designed for easy extension--custom policies and real-world deployment modules can be plugged in with minimal effort
• Example data (example/)
  • Sample motion data in our experiments (example/motion_data/, you can visualize the motion data with tools in robot_motion_process/)
  • A pretrained policy checkpoint (example/pretrained_hors_stance_pose/)

Usage

• Refer to INSTALL.md for environment setup and installation instructions.

• Each module folder (e.g., humanoidverse, smpl_retarget) contains a dedicated README.md explaining its purpose and usage.

• How to let your robot perform a new motion?

  • Collect the motion data from the source and process the motion data to the SMPL format (motion_source/).
  • Retarget the motion data to the robot (smpl_retarget/, choose Mink or PHC pipeline as you like).
  • Visualize the processed motion to check whether the motion quality is satisfiable (smpl_vis/, robot_motion_process/).
  • Train a policy for the processed motion in IsaacGym (humanoidverse/).
  • Deploy the policy in MuJoCo or real-world robot (humanoidverse/).

Folder Structure

• description: provide description file for SMPL and G1 robot.
• motion_source: docs for getting SMPL format data.
• smpl_retarget: tools for SMPL to G1 robot retargeting.
• smpl_vis: tools for visualizing SMPL format data.
• robot_motion_process: tools for processing robot format motion. Including visualization, interpolation, and trajectory analysis.
• humanoidverse: training RL policy
• example: example motion and ckpt for using PBHC

Citation

If you find our work helpful, please cite:

@article{xie2025kungfubot,
  title={KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills},
  author={Xie, Weiji and Han, Jinrui and Zheng, Jiakun and Li, Huanyu and Liu, Xinzhe and Shi, Jiyuan and Zhang, Weinan and Bai, Chenjia and Li, Xuelong},
  journal={Advances in Neural Information Processing Systems},
  year={2025}
}

@article{han2025kungfubot2,
  title={KungfuBot2: Learning Versatile Motion Skills for Humanoid Whole-Body Control},
  author={Han, Jinrui and Xie, Weiji and Zheng, Jiakun and Shi, Jiyuan and Zhang, Weinan and Xiao, Ting and Bai, Chenjia},
  journal={arXiv preprint arXiv:2509.16638},
  year={2025}
}

License

This codebase is under CC BY-NC 4.0 license. You may not use the material for commercial purposes, e.g., to make demos to advertise your commercial products.

Acknowledgements

• ASAP: We use ASAP library to build our RL codebase.
• Beyondmimic: We incorporate Beyondmimic features into policy training.
• RSL_RL: We use rsl_rl library for the PPO implementation.
• Unitree: We use Unitree G1 as our testbed robot.
• Maskedmimic: We use the retargeting pipeline in Maskedmimic, which based on Mink.
• PHC: We incorporate the retargeting pipeline from PHC into our implementation.
• GVHMR: We use GVHMR to extract motions from videos.
• IPMAN: We filter motions based on IPMAN codebase.

Contact

Feel free to open an issue or discussion if you encounter any problems or have questions about this project.

For collaborations, feedback, or further inquiries, please reach out to:

• Weiji Xie: [email protected] or Weixin shisoul
• Jinrui Han: [email protected] or Weixin Bw_rooneY
• Chenjia Bai (Corresponding Author): [email protected]

We welcome contributions and are happy to support the community in building upon this work!