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

TL;DR

PBHC introduces a two-stage physics-based approach for humanoid motion imitation of highly dynamic tasks, combining a motion-processing pipeline (SMPL→CoM/CoP filtering, contact-aware correction, IK retargeting) with an adaptive tracking mechanism guided by a bi-level optimization. The RL framework employs an asymmetric actor-critic with reward vectorization and RSI to enable robust sim-to-real transfer, aided by domain randomization. Experiments show superior tracking accuracy and stability over baselines, with real-world deployment on a Unitree G1 demonstrating agile kungfu and dance motions. The work highlights the importance of physics-consistent motion processing and adaptive curriculum in learning expressive, high-dynamic humanoid behaviors.

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. The project page is https://kungfu-bot.github.io.

Figures (12)

• Figure 1: An overview of PBHC that includes three core components: (a) motion extraction from videos and multi-steps motion processing, (b) adaptive motion tracking based on the optimal tracking factor, (c) the RL training framework and sim-to-real deployment.
• Figure 2: Illustration of the effect of tracking factor $\sigma$ on the reward value.
• Figure 3: Closed-loop adjustment of tracking factor in the proposed adaptive mechanism.
• Figure 4: Example of the right hand $y$-position for 'Horse-stance punch'. The adaptive $\sigma$ can progressively improve the tracking precision. $\sigma_\mathrm{pos\_vr}$ is used for tracking the head and hands.
• Figure 5: Example motions in our constructed dataset. Darker opacity indicates later timestamps.