Researchers from Tsinghua University, Peking University, Galbot, and Shanghai AI Laboratory have developed LATENT, a system that teaches a Unitree G1 humanoid robot to play competitive tennis using only fragmentary, imperfect human motion data. The work, accepted as a Spotlight paper at the Conference on Robot Learning (CoRL) 2024, addresses one of the central bottlenecks in <a href="/news/2026-03-15-tu-munich-neura-robotics-robogym-humanoid-training-center">physical AI</a>: the difficulty of acquiring high-quality, task-specific motion data for training athletic behaviors on real hardware.
The key insight behind LATENT is that complete, precise motion-capture sequences from real tennis matches are not necessary to learn capable athletic policies. Instead, the system uses short motion fragments capturing primitive tennis skills — partial swings, stances, footwork patterns — as priors. A reinforcement learning pipeline then corrects and composes these fragments into a coherent whole-body control policy, which is subsequently transferred from simulation to the physical Unitree G1 robot through a series of robustness-oriented sim-to-real design choices. The resulting policy enables the robot to track incoming balls across varied conditions, return them to target locations, maintain natural motion style, and demonstrate reactive footwork — achieving stable multi-shot rallies against multiple human opponents.
The practical significance of LATENT lies in what it implies for data requirements in physical AI development. If quasi-realistic motion fragments are sufficient priors for learning dynamic, contact-rich athletic tasks, the field faces a considerably lower data collection burden than previously assumed. For teams building general-purpose humanoids, <a href="/news/2026-03-14-gig-workers-training-humanoid-robots-physical-ai">ground-truth demonstration data is expensive and slow to collect at scale</a> — a lower floor on data quality changes the calculus. Galbot, a robotics company with multiple co-authors on the paper, has a direct interest in translating these findings toward commercial hardware.
Community discussion around the paper noted the robot's movements retain a distinctly mechanical quality — hesitant and sharp rather than fluid. Some observers found that gap instructive about what learned athletic policies can and can't yet replicate. The authors identify improving motion naturalness and extending the framework to more complex sports as the next targets. Video demonstrations of multi-shot rallies and reactive footwork on real hardware are at zzk273.github.io/LATENT.