TH-MOS Gait & Soccer Skills
Learning and deploying locomotion, shooting, and fall recovery on Booster K1.
Role: Gait Control Group Lead · Decision-Making Algorithms Lead, TH-MOS Robot Soccer Team, Tsinghua University.
I develop learning-based control for humanoid soccer, spanning command-conditioned adversarial motion priors (AMP), target-conditioned shooting, and fast fall recovery. My work includes policy training, simulation validation, and deployment on Booster K1.
On the field
A goal scored during an international RoboCup match shows the shooting skill in a competitive setting.
Recovery and shooting development
The recovery video shows physical robot testing; the shooting video shows the policy in simulation.
Methods and hardware evaluation
- Omnidirectional locomotion: command-conditioned AMP for velocity tracking. Hardware tracking ranges reach −1.5 to +2.0 m/s forward/backward, −1.5 to +1.5 m/s laterally, and −1.7 to +1.7 rad/s in yaw.
- Target-conditioned shooting: policies aim at specified 3D coordinates within the goal mouth from positions across a 14 × 9 m field. Separate hardware tests achieved targeting errors within 10 cm in 27 of 30 trials.
- Fall recovery: an AMP-based policy trained in MJLab completed the transition from tested fallen poses to stable standing within 0.5 s, succeeding in all 30 hardware trials.
Current work explores onboard-perception-guided curved shots for obstacle avoidance.