Humanoid robot plays tennis with 96.5% accuracy using new AI framework

Researchers in China, in collaboration with AI robotics firm Galbot, have developed a novel system called LATENT that enables humanoid robots to play tennis with a 96.5% accuracy rate. Unlike previous methods relying on perfect motion capture, LATENT uses imperfect, amateur human motion data broken down into basic movement fragments such as strokes and footwork. This approach allows the robot to refine, combine, and respond to tennis movements naturally through reinforcement learning and large-scale simulations. The system was successfully implemented on the Unitree G1 humanoid robot, which demonstrated consistent ball striking, targeted returns, and the ability to sustain multi-shot rallies against human players. The LATENT framework addresses longstanding challenges in robotic sports training, particularly replicating fast, dynamic, and precise human athletic behavior with limited or imperfect data. Testing over 10,000 trials showed the robot outperformed earlier methods in accuracy, success rate, and motion naturalness, though it is not yet at the level of professional players. Current limitations include