Humanoid Badminton

Humanoid Badminton is a research project from Tsinghua University, accepted at CoRL 2026. It presents a three-stage hierarchical reinforcement learning framework that lets a humanoid robot (shown in videos as a Unitree G1) return shuttlecocks in real-world rallies against human players.

The method first uses task-randomized motion augmentation to expand sparse annotated hitting events from limited human motion data into executable, target-conditioned stroke variations that form a continuous latent skill space. A high-level planner then outputs latent skill codes online based on the observed shuttle state. A context-conditioned adversarial regularizer encourages more natural skill usage without reducing return performance.

The authors report real-world forehand, backhand and jump-return skills and sustained multi-skill human-robot rallies, and claim this is the first real-world humanoid racket-sport system to show multi-skill rallies including jump returns. The project page offers demonstration videos only. No code, weights or datasets were found in the provided material, so it is a research result rather than a product users can access.