Learning athletic humanoid tennis skills from imperfect human motion data

The paper introduces LATENT, a system that teaches humanoid robots athletic tennis skills using imperfect, fragmented human motion data instead of complete match sequences. By leveraging quasi-realistic motion primitives with correction and composition techniques, the policy enables consistent ball striking and returns under varied conditions. Deployed on the Unitree G1 robot via sim-to-real transfer, the system achieves stable multi-shot rallies with human players.