Learning athletic humanoid tennis skills from imperfect human motion data
6 hours ago
- #Humanoid Robotics
- #Sim-to-Real Transfer
- #Tennis Skills
- Proposes LATENT, a system for learning athletic humanoid tennis skills from imperfect human motion data.
- Uses motion fragments capturing primitive tennis skills instead of complete sequences, easing data collection.
- Demonstrates that quasi-realistic data can still provide useful priors for humanoid policy learning.
- Learns a policy enabling humanoid robots to strike and return balls under varied conditions with natural motion styles.
- Includes robust sim-to-real transfer designs, successfully deployed on the Unitree G1 humanoid robot.
- Achieves stable multi-shot rallies with human players in real-world tests.