Sony AI's table tennis robot, Ace, won 3 of 5 matches against high-level amateur players under official rules, according to a study published in Nature on April 25, 2026. Against two professionals from the Japanese league, Minami Ando and Kakeru Sone, the robot fared worse, taking 1 of 7 matches. The work is being framed as a demonstration that an autonomous system can compete with humans in a physical, real-time sport, not just a simulated game board.
The headline number is control, not power. Ace returned 75% of balls in play, winning points by keeping rallies alive and forcing errors rather than overpowering opponents. That is a different profile from the brute-force playstyle robotics demos usually advertise.
Ace pairs three systems: a perception stack that reads the ball's spin and trajectory mid-flight, a decision-making model that picks shots in real time, and an 8-jointed robotic arm fast enough to position the racket precisely on each return. Sony AI says the combination is what let Ace stay in points against opponents who train daily.
Key facts
- 01Sony AI's Ace robot won 3 of 5 matches against high-level amateur players under official table tennis rules.
- 02Against two Japanese league pros, Minami Ando and Kakeru Sone, Ace won 1 of 7 matches.
- 03Ace successfully returned 75% of balls, gaining points through control rather than power.
- 04The robot uses an 8-jointed arm paired with a perception system that reads ball spin and trajectory.
- 05Results were published in Nature on April 25, 2026.
Peter Dürr, director of Sony AI and the project lead on Ace, said the work shows "an autonomous robot can actually win in a sports competition, equaling or exceeding the reaction time and decisionmaking ability of humans in a physical space." He added that table tennis "is a game of enormous complexity that requires split-second decisions as well as speed and power."
“Ace returned 75% of balls in play, winning points through control rather than power, and took three of five matches against high-level amateurs under official rules.”— Jaeden Schafer
The contrast with prior AI milestones is the point. Systems have already cleared chess, Go, and StarCraft II, all played in software where the environment is fully observable and the action space is discrete. Table tennis introduces aerodynamics, spin, latency, and a moving racket that has to land in the right place at the right millisecond. The 1-of-7 result against Ando and Sone is a reminder that the gap to elite human play is still real.
Peter Stone, chief scientific officer for artificial intelligence at Sony, pitched the result as a generalization claim. "This breakthrough is much more important than table tennis," he said. "It represents a pivotal moment in AI research, demonstrating for the first time that an AI system can perceive, reason, and act effectively in complex and rapidly changing real-world environments that require accuracy and speed."
Stone went further, arguing that once AI matches human-expert performance in physical tasks, "it will pave the way for a whole new class of real-world applications that were previously unattainable." The unstated targets are the obvious ones: warehouse manipulation, surgical assistance, household robotics, anything where reaction time and fine motor control gate deployment.
The 75% return rate is the most useful single statistic in the paper, because it isolates what Ace is actually good at. The robot is not yet generating winners against pros. It is denying them, dragging rallies out, and converting opponent mistakes. That is a coherent strategy for a system whose perception is reliable but whose offensive shot selection still trails human intuition.
Caveats are worth naming. Five matches against amateurs and seven against two pros is a small sample, and Sony AI has not detailed how opponents were briefed, how serves were handled, or how Ace performs across a wider range of playstyles and spins. The Nature paper is the formal record, but the broader question — can Ace beat a top-100 player, not a Japanese league pro on a given afternoon — remains open.
There is also the gap between a purpose-built table tennis rig and a general-purpose robot. Ace's 8-jointed arm and perception stack are tuned for one sport on one table. Transferring the underlying perception-and-control loop to messier domains, like a kitchen or a factory floor, is a separate research program.
Still, the direction matters. The frontier in AI has spent two years arguing about token costs, context windows, and agent reliability inside a browser. Sony AI just put a robot across a net from a professional and won a game. For a field that often confuses benchmarks with capability, a 75% return rate against humans hitting back is a useful, physical data point.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
