Adapted BFM-style motion tracking to the 22-DoF Booster K1, achieving over 95% full-clip completion on the BONES-SEED test set after 5 hours of training on a single node with 8 NVIDIA A10 GPUs. Deployed and verified the policy on the real robot.

Reproduced SuperMimic’s differentiable-world-model approach to humanoid motion tracking on native MuJoCo-Warp through mjlab, achieving 100% full-clip completion across 40 LAFAN1 motions within 5 hours of training on a single RTX 5090. Demonstrated few-shot adaptation to highly dynamic MotionCode motions.

Developed command-conditioned AMP policies for omnidirectional locomotion on Booster K1, achieving stable real-robot velocity tracking from −1.5 to +2.0 m/s forward/backward, ±1.5 m/s laterally, and ±1.7 rad/s in yaw.

Trained an AMP-based fall-recovery policy in MJLab for Booster K1, recovering from tested fallen poses to stable standing within 0.5 s and succeeding in all 30 real-robot trials.

Trained reference-free dribbling on Booster K1 using asymmetric PPO, jointly learning locomotion and head control from ball-velocity commands, proprioception, and camera-frame ball observations.

Developed target-conditioned shooting policies on Booster K1 to aim at specified 3D goal-mouth coordinates from positions across a 14 × 9 m field. Real-robot tests achieved targeting errors within 10 cm in 27 of 30 trials; demonstrations include a goal scored at RoboCup.

Built a Unity–MuJoCo platform for autonomous 3v3, 5v5, and 11v11 humanoid soccer, connecting learned locomotion, kicking, and recovery with perception, team decisions, and match management.