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 Li1,4, Xinzhe Liu1,5, Jiyuan Shi1, Weinan Zhang2, Chenjia Bai† 1, Xuelong Li1
1Institute of Artificial Intelligence (TeleAI), China Telecom 2Shanghai Jiao Tong University 3East China University of Science and Technology 4Harbin Institute of Technology 5ShanghaiTech University *Equal contributions †Corresponding Author
Paper arXiv Code

Video

Abstract

Humanoid robots are promising to acquire various skills by imitating human behaviors. However, existing algorithms are only capable of tracking smooth, low-speed human motions, even with delicate reward and curriculum design. This paper presents a physics-based humanoid control framework, aiming to master highly-dynamic human behaviors such as Kungfu and dancing through multi-steps motion processing and adaptive motion tracking. For motion processing, we design a pipeline to extract, filter out, correct, and retarget motions, while ensuring compliance with physical constraints to the maximum extent. For motion imitation, we formulate a bi-level optimization problem to dynamically adjust the tracking accuracy tolerance based on the current tracking error, creating an adaptive curriculum mechanism. We further construct an asymmetric actor-critic framework for policy training. In experiments, we train whole-body control policies to imitate a set of highly-dynamic motions. Our method achieves significantly lower tracking errors than existing approaches and is successfully deployed on the Unitree G1 robot, demonstrating stable and expressive behaviors.

Method

图片1

Real World Experiments

Jump kick
Roundhouse kick
Side kick
Front kick
Back Kick
360-degree spin
Dance
Bruce Lee's pose
Fighting combo
Horse stance punch
Hooks punch
Jabs punch
Horse stance pose
Stretch leg
Tai Chi

Experiment Results

Main Results

Main results comparing different methods across difficulty levels. PBHC consistently outperforms deployable baselines and approaches oracle-level performance. Results are reported as mean ±one standard deviation. Bold indicates methods within one standard deviation of the best result, excluding MaskedMimic.

图片2


Ablation on Adaptive Motion Tracking

Ablation study comparing the adaptive motion tracking mechanism with fixed tracking factor variants. The adaptive mechanism consistently achieves near-optimal performance across all motions, whereas fixed variants exhibit varying performance depending on motions.

图片2


Sim-to-Real Result

Comparison of tracking performance of Tai Chi between real-world and simulation. The robot root is fixed to the origin since it’s inaccessible in real-world.

图片2

BibTeX

@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={arXiv preprint arXiv:2506.12851},
        year={2025}
      }