Real-Time Tennis Without Scripts
This humanoid robot stands about 4 feet tall, offering a compact frame suited for the court. Galbot Robotics demonstrated it rallying shot-for-shot with a human player. Powered by the LATENT system on the Unitree G1, it moves, adjusts, and competes live, far from scripted athletic robots that rely on pre-programmed actions or remote control.
It tracks fast-moving balls, repositions dynamically, and handles changing trajectories, marking a shift toward autonomous reaction in sports.
Training on Minimal Data
Tennis demands capturing human gameplay data, predicting ball paths, precise racket control, and full-body balance. Researchers bypassed full matches by focusing on small movement segments from five players on a 10-by-16-foot court—17 times smaller than standard.
The system learns individual motions first, then combines them into sequences for coordinated play. Simulation training varied mass, friction, and aerodynamics to build adaptability to real-world variables.
Robot Capabilities in Action
- Instant reactions to opponent shots without delay.
- Positioning across the court for optimal returns.
- Adjusting to unpredictable ball paths mid-rally.
- Sustaining extended exchanges with consistency.
- Placing shots strategically away from the player.
Performance and Limits
In tests, it hit 96% forehand success in simulation and maintains rallies in reality. The demo reveals competitive edge, hinting at decision-making. Yet instability appears at times, motions lack athlete fluidity, and high shots challenge it—progress remains evident but incomplete.
Implications Beyond the Court
This demonstrates robots learning intricate human skills sans exhaustive data. The approach suits tasks like surgery, manufacturing, or other sports lacking motion records. Soon, robots might train athletes or feature in exhibitions, blurring human-machine boundaries in competition.
