KungfuBot

A humanoid whole-body motion-imitation framework that teaches a Unitree G1 highly dynamic moves like kung fu and dance.

KungfuBot was released in June 2025 by China Telecom's AI Research Institute (TeleAI) together with Shanghai Jiao Tong University, East China University of Science and Technology, Harbin Institute of Technology, and ShanghaiTech University, accepted to NeurIPS 2025, with the method itself named PBHC and its code open-sourced. Earlier humanoid motion-tracking work could mostly only imitate slow, gentle motions. PBHC first processes the data: it extracts SMPL human poses from video, uses physical metrics to filter out motions the robot can't achieve and correct foot contact, then retargets the result onto a Unitree G1; it then trains a tracking policy with reinforcement learning, turning the tolerance for tracking error into a bi-level optimization that adapts automatically to the error, effectively an automatic curriculum. A September 2025 follow-up, KungfuBot2, uses an orthogonal mixture of experts to let a single policy master multiple skills, and can stably track motions lasting up to several minutes.

ExampleThe G1 performs a jumping kick, a spinning kick, a horse-stance punch, and tai chi movements, all copied from human video; the project page showcases 15 such highly dynamic skills in total.

Also called

PBHC, KungfuBot2, VMS

Related

Motion Tracking · Motion Retargeting · Unitree G1 · SMPL · Asymmetric Actor-Critic · Learning-Based Whole-Body Control

Sources

KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills (arXiv 2506.12851)
KungfuBot 项目页 (Chinese)
KungfuBot2: Learning Versatile Motion Skills for Humanoid Whole-Body Control (arXiv 2509.16638)

As of

2025-10