Overview
- In real-world tests, reported success ranged from a best-case 96.5% within 2.5 meters of a target to roughly 90% forehand and just under 80% backhand return rates, depending on source and metric.
- Researchers collected about five hours of amateur motion-capture fragments for strokes and footwork to build priors for humanlike movement.
- The team trained high-level policies via reinforcement learning in simulation with dynamics randomization and observation noise before transferring to the Unitree G1.
- Across 10,000 evaluated trials, the system sustained multi-shot rallies with humans and outperformed prior methods in success rate, accuracy, and motion naturalness.
- The preprint is posted on arXiv with open-source code available, and authors note current limits such as reliance on motion capture and simplified return tasks, with planned advances like active vision and multi-agent training.