Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data
Summary
LATENT presents a system for learning athletic humanoid tennis skills from imperfect human motion data, using motion fragments as priors to train a policy that can strike and rally balls with a Unitree G1 humanoid. The work emphasizes robust sim-to-real transfer, and provides demonstrations of multi-shot rallies, reactive footwork, and self-play in simulation.