It has been 10 years since artificial intelligence (AI) surpassed the world's top players in the intellectual game of Go, and now it has defeated elite athletes in a physical sport. A robot powered by AI has repeatedly beaten table tennis players in matches played under official rules. Table tennis robots existed before, but this is the first time one has beaten elite players.
Sony AI in Zurich, Switzerland, said on the 23rd in the international journal Nature as a cover paper that "the autonomous AI robot Ace faced five elite table tennis players under the rules of Japan's professional table tennis league and won three times." The players Ace faced were not professionals, but all had at least 10 years of experience and trained 20 hours a week.
According to the paper, Ace lost both matches against two professional players. It managed to take only one set. But Sony AI said that after submitting the paper to Nature, it further improved Ace's capabilities and defeated professional players as well. It means Ace is now at the position AlphaGo was 10 years ago.
◇ Detecting pingpong balls with cameras and sensors
Ace is a robot arm with eight joints. It looks like an industrial robot arm that assembles cars. With the help of cameras surrounding the table tennis court and a network of acceleration and vision sensors, the robot determines the ball's position in 3D and detects acceleration and spin.
Sony AI said Ace achieved three major advances in autonomous robotics. First, it uses an "event-based sensor." That means the robot focuses on the parts of the camera-captured image that show changes in motion or brightness in specific areas—namely, those important for tracking the pingpong ball's trajectory.
Second, the robot's table tennis skills were built using "reinforcement learning," the method used to develop AlphaGo, the Go AI. Peter Dürr of Sony AI in Zurich, Switzerland, who led Ace's development, said, "AI learns through simulated game experience rather than being taught how to play table tennis." In other words, it is like giving a dog praise or food as a reward when it performs a specific action, rather than repeatedly explaining the action.
Lastly, the researchers said they applied high-speed robot hardware so Ace could play with human-level agility. Dürr said it takes an athlete about 0.23 seconds to react, whereas Ace's reaction latency is just 0.020 seconds.
The players who faced Ace tried to hide their serve motions, but said they were surprised that Ace immediately sensed the spin and responded. In particular, Ace even returned balls that clipped the net and bounced out. Even the developers had not anticipated that skill. Dürr said it was a capability that emerged on its own from the AI.
Experts said they expect AI will also help maximize human abilities. Nakamura Kinjiro, a Japanese table tennis player who competed in the 1992 Barcelona Olympics, said in the paper, "A particular shot Ace made was previously thought to be impossible," adding, "But the fact that it was possible means humans can do it too."
◇ Beyond games, AI surpasses humans in physical sports
Ace's success shows that the moment is approaching when AI can surpass humans in physical sports beyond intellectual games. Earlier, in 1997, IBM's AI supercomputer Deep Blue defeated world chess champion Garry Kasparov, and Google DeepMind's Go AI AlphaGo beat 9-dan Lee Sedol 4-1 in 2016.
Dürr said, "Intellectual games such as chess and Go have long served as a testbed for demonstrating AI capabilities," adding, "While previous AI milestones happened online, Ace shows important progress in that it held its own against professional table tennis champions in the real world."
Sony AI said it has continued to improve Ace's capabilities over the past year. The researchers said Ace defeated a professional player for the first time in December last year and notched wins in Mar. this year against Kihara Miyu, a female professional player ranked around 25th in the world, and male professional players Ryuzaki Tonin and Igarashi Fumiya. Dürr said, "Surpassing the world champion will be possible in the future." The researchers said Ace will also be implemented in a humanoid form.
Peter Stone, chief scientist at Sony AI, said, "This achievement goes far beyond table tennis," adding, "If AI can demonstrate expert-level capabilities in complex, rapidly changing real-world environments that demand precision and speed, it will open the door to entirely new types of real-world applications that were previously out of reach."
Esther Colombini, a professor at the Institute of Computing at the State University of Campinas in Brazil, also wrote in a companion commentary in Nature the same day, "One might expect a machine to rely mainly on power, but Ace achieved those results not with shots faster than those of humans, but with a variety of spins and consistency in returning the ball," adding, "Ace is an important milestone that shows the potential of next-generation AI that interacts with the physical world."
References
Nature (2026), DOI: https://doi.org/10.1038/s41586-026-10338-5