Humanoids can now rally with you on the tennis court
Chinese researchers just taught a Unitree G1 humanoid robot to play tennis using only five hours of imperfect motion-capture data. Through a framework called LATENT, the robot can track footwork, return balls traveling over 15 m/s with 90% accuracy, and sustain multi-shot rallies against human players.
Step onto a tennis court today, and you might actually find yourself in a rally against a two-legged robot adjusting its footwork.
A team of researchers from Tsinghua University, Peking University, Galbot, Shanghai Qi Zhi Institute, and Shanghai AI Laboratory just introduced LATENT, a system deployed on the Unitree G1 humanoid robot that can sustain real-time, multi-shot tennis rallies with human opponents.
Five hours of messy data is all it took
Teaching a humanoid to play high-speed sports has traditionally been an absolute nightmare because pristine, match-grade human motion data basically does not exist for robots.
Instead of chasing perfect full-game sequences, the team trained LATENT on just five hours of raw motion-capture data. These were simply short, imperfect clips of basic human swings and movement fragments.
The core insight is that even messy fragments carry the essential athletic priors of how humans swing and shift weight. By combining these primitives, correcting the policy in simulation, and deploying a robust sim-to-real transfer pipeline, the robot learned to string together fluid, reactive motions on its own.
Sustaining real rallies at 15 m/s
This isn't a staged tech demo where balls are gently tossed to a fixed target.
Deployed on the Unitree G1, the system successfully returns balls traveling at speeds over 15 meters per second (about 34 mph) with roughly 90% accuracy. The robot tracks footwork dynamically, positions itself, and aims returns back to target zones, keeping continuous rallies alive against real humans.
The team has already open-sourced the approach on GitHub, positioning LATENT as a reusable blueprint for teaching humanoids athletic skills from noisy real-world data.
Why it matters
What makes this genuinely impressive is the intersection of low-cost hardware and data efficiency.
The Unitree G1 retails for around $16,000, which is remarkably cheap by humanoid robotics standards. Seeing a budget platform perform fluid, unchoreographed athletic movements without thousands of hours of hand-tuned data breaks the assumption that agile robotics requires million-dollar setups.
In our view, LATENT's real triumph is treating imperfect human data as a feature rather than a bug. By extracting primitive motor building blocks from just five hours of casual clips, it provides a practical shortcut around robotics' massive data collection bottleneck.
If this framework scales to other dynamic tasks like ping-pong, badminton, or high-speed industrial handling, five hours of rough motion capture might just become the new baseline for teaching robots how to move.
Sources (1)
- zzk273.github.io(opens in a new tab)Originalzzk273.github.io