Researchers teach humanoid robot to master kung fu and dynamic movement

Humanoid robots have long struggled to replicate fast, complex human movements. Most existing systems can only handle slow, smooth motions even with carefully tuned training methods. A new framework called KungfuBot tackles this challenge by combining motion processing, adaptive learning, and physics-informed control to teach robots highly dynamic behaviors like kung fu and dance.

How the system works#

The approach involves two main stages. First, the team developed a motion processing pipeline that takes recorded human movements, filters out noise, corrects inconsistencies, and adjusts them to work within the robot's physical constraints. This ensures the target motions are realistic and achievable before training begins.

Second, they created an adaptive learning mechanism using a two-level optimization problem. Rather than fixing the difficulty level, the system automatically adjusts how precisely the robot must track the human motion based on current performance. When tracking is poor, it relaxes the requirements; when the robot performs well, it tightens them. This creates a natural curriculum that guides training without manual intervention.

Training and deployment#

The team used an asymmetric actor-critic framework—a machine learning approach where separate neural networks learn to generate actions and evaluate performance. This architecture proved effective for handling the complexity of whole-body control across dynamic motions.

When tested, the resulting policies achieved substantially lower tracking errors than competing methods. The research demonstrates stable, expressive behaviors successfully deployed on a Unitree G1 humanoid robot, suggesting the framework is ready for real-world application.