Ace Reaches Professional Play

Sony AI says Ace, its autonomous table tennis robot, has reached a new threshold: between February and April 2026, the system won matches against seven ranked professional players under official competition rules. The post extends the Nature paper in which Ace could beat elite players but had not yet outplayed professionals. The verified fact is narrow and important: after additional hardware and software iterations, Sony AI says Ace can now repeatedly defeat professional opponents, including Miyuu Kihara, then world No. 26 in women’s singles, and Miu Hirano, a two-time Olympic silver medalist.

This is best read as a robotics milestone, not as a sports novelty. Table tennis forces a robot to see, predict and act within tens of milliseconds. The ball changes speed, spin and trajectory after the bounce, while the opponent adapts shot by shot. For Ace, that means a full technical chain: fast perception, ball-state estimation, shot selection, arm motion and racket control. Sony AI says the matches followed International Table Tennis Federation rules and used official umpires, making the benchmark more demanding than a staged laboratory demonstration.

The reported gains came from several layers working together. On control, Ace’s serves were optimized with Bayesian optimization, a method that builds a probabilistic model of which motions are most likely to work and then searches efficiently around them. On learning, shots from previous matches and internal assessments against professionals were added to training, exposing the system to professional-level ball distributions. On hardware, Sony AI says it reduced the weight of some base components by more than 4 kg and upgraded motors to improve acceleration. On perception, ball-detection latency fell from about 10 ms to 8.5 ms. Those numbers matter because, in high-speed robotics, a smarter model cannot always compensate for a slow sensor or a heavy arm.

The practical meaning goes beyond sport. Ace is not a product ready for sale, but it is a testbed for physical AI: machines trained in simulation, transferred into the real world, and then improved as engineers discover the gaps caused by drag, vibration, sensing error and human unpredictability. Sony AI is also clear about the remaining limits. Ace still does not match the extreme spin values generated by top human players, and its tactical and strategic layers can improve. Still, the signal is useful for industrial robotics, assistive systems and rehabilitation: progress will come from tightly integrated sensing, actuation, control and learning, not from AI models alone.