A tennis robot points to the harder question: how messy can robot training data be?
Manual Codex image generation📷 AI-generated / Tech&Space
- ★Researchers collected five hours of fragmented tennis motion from amateur players
- ★LATENT was deployed on a Unitree G1 humanoid after simulation training
- ★The system still depends on motion capture and needs active vision for more realistic deployment
LATENT looks like a story about a robot playing tennis, but the more useful reading is stricter: researchers showed that a humanoid can learn a dynamic sports skill from imperfect human fragments. TechXplore describes a team in China that, working with the Chinese AI robotics company Galbot, deployed the system on a Unitree G1 humanoid robot.
The data was not perfect professional motion. The team collected five hours of primitive tennis fragments from amateur players: forehand, backhand, lateral shuffle, and crossover step. From that, they built a latent action space, then a policy that uses reinforcement learning and simulation to correct and compose those fragments into real robot control.
The result is concrete enough to matter. Across 10,000 real-world evaluation trials, the robot's best setup reached a 96.5% success rate, defined as returning an incoming ball within 2.5 meters of the target location. That is not a poetic claim that it "plays tennis"; it is a measurable control task.
The spotlight still has to move away from the racket. The biggest signal is not that the robot can look athletic, but that incomplete demonstrations can contain enough structure for skill learning. If that pattern works beyond tennis, it could reduce the data collection cost for manipulation, industrial tasks, and assistive robots.
The real result is not the robot forehand; it is evidence that dynamic skills can be composed from messy, fragmentary demonstrations.
Manual Codex image generation📷 AI-generated / Tech&Space
The limitation is explicit in the source: the current system relies on motion capture for real-world deployment. The researchers say active vision is a future direction because it could reduce dependence on external infrastructure. Without that, LATENT is not yet an autonomous player on an ordinary tennis court.
The task formulation is another constraint. The work focuses on returning randomly initialized incoming balls to target locations, which is not the same as a full match against an unpredictable opponent. The robot can sustain multi-shot rallies, but that does not mean it understands strategy, serves, spin, material fatigue, or safety around people.
Hardware is not a footnote either. The Unitree G1 is a commercially available humanoid, but athletic motion creates loads that differ from walking on a flat floor. Every stroke is a test of joints, actuators, battery, and balance control. Lab success has to survive a maintenance interval, not just a demo video.
That is why LATENT is best read as a physical skill-learning method, not an announcement of robot tennis academies. Tennis is a visible and difficult demonstration language. The real value will be in tasks where humans cannot provide perfect demonstrations, but can provide fragments good enough for a robot to learn the motion.

