Dribbling
Whole-body humanoid dribbling with active vision on Booster K1.
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.
Whole-body reinforcement learning
Reimplemented and adapted Joga-style dribbling to K1 with asymmetric PPO, jointly controlling all 22 joints, including head yaw and pitch. The actor uses three frames of proprioception, ball-velocity commands, and camera-frame ball observations; the critic receives additional physical state during training.
Active vision and command curriculum
Modeled camera field of view, ball visibility, and observation latency. A curriculum increases the frequency of command changes, with rewards for ball-velocity tracking, close ball control, and balance.
Simulation and deployment
Built GPU-parallel training in Isaac Lab with Newton/MuJoCo-Warp, physical randomization, and ONNX export with consistent observations, joint ordering, and actuator limits. The video shows continuous dribbling on Booster K1, four-view simulation, and a first-person demonstration.