At the second World Humanoid Robot Games in Beijing on 22 August, a humanoid robot stood on a real tennis court and returned balls travelling faster than 50 km/h, judging the landing point in milliseconds, stepping, twisting and swinging. It served, played forehands and backhands, moved its opponent around the baseline, volleyed at the net, and did not fall apart. In a doubles warm-up it covered for its human partner and changed tactics. During one high-speed rally it made a desperate save, crashed to the ground, stood up on its own and kept playing.
This was the first fully autonomous humanoid tennis match, and the robot came from Galbot, the Beijing-based embodied-intelligence company. Its opponent was Zheng Jie, a former Grand Slam doubles champion. Galbot named the moment AstraTennis.
One model for thinking, balance and reflexes
Past robot demos split the job across modules: a perception system, a planner, a controller, with failures at every interface. Galbot’s answer is a single embodied foundation model, AstraBrain, that welds the three together. The brain understands and decides, reading the ball’s landing point and choosing tactics. The cerebellum handles motor control, full-body balance at speed and explosive swings. The pons translates decisions into movement commands.
The architecture’s value shows in real time. As a ball arrives, the brain judges in milliseconds whether it is deep or short and whether to attack or defend, while the cerebellum is already sliding, crouching and preparing the racket. At contact the brain updates: that shot was not tight enough, the next one must hug the line. Such think-while-acting closure is only possible in a unified architecture. In June, at CVPR 2026, Galbot reported its AstraBrain-WBC 0.5 reaching 92.58 per cent zero-shot generalisation with inference latency under 1.5 milliseconds. AstraTennis is the same system’s first outing in real competition.
Learning from imperfect data
Galbot founder and chief technology officer Wang He calls tennis the ultimate examination for humanoid robots, because it forces a machine to solve cerebellum and brain problems at once. Traditional robot training leans on teleoperation, but in high-speed tennis a human cannot react fast enough to collect usable demonstrations, and full motion-capture of real matches is prohibitively expensive.
Galbot went the opposite way. Its AstraBrain Latent algorithm, billed as the first whole-body real-time control scheme for tennis play, learns from imperfect fragments: ordinary people’s forehands, backhands, side shuffles and crossover steps. Perfect data is too expensive and rare; imperfect data already carries the priors of human movement, how weight shifts, how the arm swings, how the feet push. The algorithm combines, corrects and generalises those fragments into a complete tennis skill. An added hidden-space action barrier keeps motions inside a human-like style band, and random perturbations during training teach the robot to self-correct, so it plays live balls rather than rehearsed ones.
Millions of virtual sparring partners
Data alone was not enough. Galbot built a virtual tennis world on top of Galaxy Stars, its self-developed billion-scale embodied dataset. Training runs in two stages: massed virtual rehearsal against opponents of different levels, then light calibration on the real machine. The crucial step is multi-agent competition. Instead of imitation, one step at a time, two or more agents play each other in simulation, each rally feeding improvement. Skills that were never explicitly taught emerge on their own, the property the industry calls skill emergence. Crucially, virtual skills transferred to the real court, which is why AstraTennis happened in front of global cameras rather than in a lab video.
Why the fall mattered most
The most telling image of the match was the fall. In a high-speed exchange the robot missed an extreme ball and hit the ground hard. No technician ran in. No one stopped the game. It stood up quickly and prepared for the next return. The detail proved the performance was not a lab set piece: the robot’s control was robust enough to recover from unexpected real-world events.
Galbot frames the wider lesson as a reusable paradigm for embodied intelligence, one general motor-control base shared across dancing, inspection, rescue and housework, with marginal cost falling as scale grows. The capability is already running elsewhere: pharmacy robots pick targets from tens of thousands of products, and the Galbot S1 dual-arm carries 50 kg of materials on industrial lines. Tennis is simply the newest tile in the mosaic, the first time AI stood up from the screen and stepped onto a real floor.
Editor’s note: This is an adapted translation of the original Leiphone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/ai/DtuJryomWlykalLZ.html.