Humanoid Robot Plays Tennis on Live TV: Galaxy General's 'AstraTennis Moment'

According to a report by Leiphone, the robot moved with human-like fluidity: its feet positioned first, torso rotated, arms swung, and it struck the ball with precision. It could serve, hit forehands and backhands, control court placement, volley, and adapt strategies in both singles and doubles. During a high-speed exchange, it fell heavily after a desperate save, then pushed itself up and continued playing without assistance.

The robot's opponent was Zheng Jie, a former Grand Slam champion, making the performance more striking. The capabilities come from an integrated system called AstraBrain, described as the world's first end-to-end large model that unifies brain, cerebellum, and neural control. The brain handles understanding and decision-making, the cerebellum manages motor control, and the pons translates decisions into movement commands. This architecture allows millisecond-level closed-loop adjustment, enabling the robot to "think while acting."

To develop tennis skills, Galaxy General adopted an unconventional approach. Instead of relying on expensive, perfect teleoperation data, they collected fragmented, imperfect motion clips from ordinary people—swings, steps, and directional movements. An algorithm called AstraBrain Latent extracts latent patterns from this imperfect data to compose complete tennis strokes. A "latent space motion barrier" keeps movements human-like and responsive to the incoming ball.

The company also built a virtual tennis world based on its hundred-billion-level embodied dataset Galaxy Xingfang. The robot trained by playing massive simulated matches against virtual opponents of varying levels, using multi-agent self-play. This process led to skill emergence, where untaught abilities spontaneously developed, and the skills transferred successfully to the physical court.

Wang He, founder and CTO of Galaxy General, said tennis is the "ultimate exam" for humanoid robots, because it proves a robot has simultaneously solved both the "cerebellum" (motor control) and the "brain" (decision-making) problems.

The most telling moment, as noted in the report, was the robot's fall and recovery. This demonstrated that the system could operate outside a controlled laboratory and handle unexpected real-world events. Beyond tennis, the same technological base is already used in smart pharmacies for picking items from thousands of products, and in industrial production lines where the Galbot S1 robot arm carries up to 50 kilograms of materials. Tennis is only the latest demonstration of a reusable paradigm for embodied intelligence, suggesting that a general motor-control base can be shared across different tasks, lowering the marginal cost of learning new skills.