A Chinese humanoid robot has demonstrated real-time tennis rallying, marking a step forward in embodied artificial intelligence.<\/p>\n
The system, developed by Galbot and Tsinghua University in Beijing\u2019s Haidian district, can track incoming balls, predict trajectories, reposition itself, and return shots using autonomous learning and full-body coordination.<\/p>\n
According to the team, with a forehand return success rate of 90.9 percent, the robot highlights growing capability in dynamic, adversarial environments\u2014one of the most challenging benchmarks for humanoid robotics, according to local media reports.<\/p>\n
In January 2026, UBTech Robotics\u2019 Walker S2 demonstrated<\/a> real-world tennis capabilities, combining perception, balance, and precision to deliver powerful, accurate strokes during human-robot rally tests.<\/p>\n Efficient skill acquisition<\/p>\n The tennis robot is powered by a training framework called LATENT, developed by a team from Tsinghua University, working with Chinese AI robotics firm Galbot, that enables humanoid robots to acquire complex sports skills from imperfect human motion data. <\/p>\n The system<\/a> focuses on decomposing tennis into fundamental motion primitives\u2014such as forehand and backhand strokes, lateral shuffles, and crossover steps\u2014allowing robots to learn in a structured and scalable manner.<\/p>\n<\/p>\n Rather than relying on high-precision motion capture or detailed kinematic modeling, LATENT uses \u201cquasi-realistic\u201d inputs derived from amateur players. Approximately 5 hours of motion data were collected using a compact capture setup, yielding noisy yet meaningful representations of human movement. This data is mapped into a latent action space, where the robot can interpret, refine, and recombine motion elements into coherent actions.<\/p>\n The framework integrates reinforcement learning with large-scale simulation, enabling the system to adapt to varying ball trajectories and gameplay conditions while preserving fluid, human-like motion. The trained policy was deployed on the Unitree G1 humanoid robot, where it demonstrated reliable ball striking and controlled returns.<\/p>\n The team claims that by reducing dependence on high-quality datasets and enabling learning from imperfect inputs, LATENT addresses a key bottleneck in robotics: replicating fast, dynamic, and precise human behaviors in unstructured environments.<\/p>\n High accuracy returns<\/p>\n