Sony's table tennis robot beats top players
Sony's AI researchers have developed a robot that beats experienced table tennis players. They see this as a milestone on the path to physical AI.
Sony's table tennis robot Ace in action.
(Image: Sony)
Sony's AI lab has introduced the robot Ace, which according to Sony is the first physical AI system capable of competing with professional table tennis players. To achieve this, Sony combined proprietary event-based sensors, a specially designed robotic arm, and a reinforcement learning architecture.
The chief scientist at Sony AI, Peter Stone, calls the project a premiere: For the first time, an AI system is capable of playing a physical sport at the level of human experts. He draws a comparison to the chess computer Deep Blue, which defeated world chess champion Garry Kasparov in 1997 defeated world chess champion Garry Kasparov, thus indicating that Sony has achieved a similar breakthrough in the development of physical AI.
Videos by heise
In a study on the robot published in Nature, the researchers write that Ace won three out of five matches against “elite players” in April 2025. By “elite players,” the researchers mean players with more than ten years of active table tennis experience and around 20 hours of training per week. It also won one out of seven matches against professional players. The AI lab has since further improved Ace. In March 2026, it competed against three professional players and beat each of them at least once.
Sensors, hardware, and AI interact
The robot is among the earliest research projects of the AI lab, launched in 2020, and was in development there for five years. According to director and lead engineer Peter Dürr, this required years of extensive iterations in perception, hardware tuning, and physical modeling.
For perception, the system uses nine cameras and three proprietary event-based sensors. The latter react only to changes in the image instead of capturing entire frames, thus capturing movements particularly quickly. They measure the ball's position 200 times per second and its rotation up to 700 times per second, Sony AI writes in a blog post about the robot.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externes Video (TargetVideo GmbH) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (TargetVideo GmbH) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
Another component was the eight-axis robotic arm developed by Sony, designed for precise striking movements with high acceleration. Its full potential is only realized in combination with the motion model. According to Sony, this is based on reinforcement learning and was trained entirely in simulation, where the AI competed against itself for countless hours. In developing the model, the team also incorporated previous research findings on the racing game AI GT Sophy, which beat top human players in Gran Turismo.
Refined through duels with humans
According to Sony AI, human input was central to the robot's development. For example, playing against increasingly stronger human opponents revealed that the physical model overestimated air resistance at very high ball speeds and had to be adjusted accordingly.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externes YouTube-Video (Google Ireland Limited) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (Google Ireland Limited) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
The researchers admit that the robot has not yet reached the level of world-class players. For this, not only the striking technique needs to be improved, but also two higher-level aspects: tactics and strategy. Tactics determine whether a ball is played safely or aggressively, while strategy defines how these decisions play out over the course of a match, the authors write.
(mki)
