On 22 August, at the National Speed Skating Oval in Beijing, the second World Humanoid Robot Games opened and state television turned its cameras to a real tennis court. A humanoid robot faced incoming balls travelling above 50 km/h, judged the landing point within milliseconds, moved its feet, twisted its torso and swung. The ball clipped the net tape and landed in.
Serve, forehand, backhand, baseline rallies and net volleys held up across the board. In a warm-up doubles round it covered for its human partner and switched tactics. During a high-speed rally it made an extreme save, then fell hard, pushed itself off the floor and kept playing. The robot that completed the first fully autonomous tennis rally in the world was built by Galbot.
Galbot calls the moment AstraTennis. Its answer is an embodied large model named AstraBrain, which fuses three jobs into one model. The brain understands and decides: where the ball lands, how to play, how to coordinate with a partner in doubles. The cerebellum controls motion: whole-body balance at speed, explosive swings, human-like movement. The brain bridge translates the brain’s decisions into the cerebellum’s action commands.
The value of one integrated architecture
Past systems split those three jobs into separate modules, and the interfaces between them were where things broke. AstraBrain makes them one system, the basis for the claim of the world’s first end-to-end full-body, full-hand large model integrating brain, cerebellum and neural control.
On court the payoff is visible. A ball arrives, the brain judges in milliseconds whether it is deep or short, whether to attack or defend, while the cerebellum already executes the slide, the crouch and the backswing. The instant the racket meets the ball, the brain updates its read.
This real-time think-while-playing loop only holds together with an integrated architecture. In June, at CVPR 2026, Galbot showed AstraBrain-WBC 0.5, whose cerebellum pushed zero-shot generalisation to 92.58 per cent and cut inference latency below 1.5 milliseconds. AstraTennis is that same system’s first outing in real adversarial play.
Learning from imperfect human data
Galbot’s founder and CTO argues the harder test is not a single trick but generalisation: a robot that plays a dynamic sport against a human opponent is a proxy for warehouse, factory and home tasks that demand real-time reaction. The company is now training on imperfect, noisy human demonstration data and adding random perturbations so the robot learns to self-correct, turning scripted motion into adaptive behaviour.
More from the original report
Editor’s note: This is an adapted translation of the original LeiFengWang report. It has been trimmed and restructured for readability for an international business audience.
Translated and adapted from LeiFengWang (https://www.leiphone.com/category/ai/DtuJryomWlykalLZ.html).