A new table tennis robot is capable of competing in the human professional league and beating even experienced elite players. With high-speed cameras, an AI-based learning system and a highly mobile robotic arm, it returns even the most difficult balls. The development shows the potential of robots in complex, interactive, real-time tasks in the physical world. Although this robot specializes in table tennis, the underlying technology could also be used in other areas.
Computers have long been superior to humans in many domains – be it chess, Go or complex video games. However, if the playing field moves to the real world, robots and artificial intelligence quickly reach their limits, especially when it comes to quick perception and reactions. Table tennis is considered a special challenge for robots because it is important to predict the ball’s trajectory, including complex spins, within milliseconds and to react immediately. Mastering these tasks with a robotic system has been a popular subject of study projects and doctorates for years. However, most previous models could only compete against amateur players and return a few friendly balls.
AI support
Now a team led by Peter Dürr from Sony AI in Zurich has developed a table tennis robot called Ace that can compete with human professional players. “Ace addresses the challenges of real-time physical interaction through a new high-speed perception system with event-based image sensors and a new control system based on model-free reinforcement learning, as well as state-of-the-art high-speed robotic hardware,” report the researchers.
The robot registers the position of the ball as well as its speed and rotation at any time using twelve cameras positioned around the playing field. This information is evaluated in real time by an AI system, which then calculates an appropriate reaction and sends appropriate commands to the robot arm. This in turn is equipped with eight joints that allow it great freedom of movement. Ace was trained in a virtual environment, where he learned to design his shots so that the ball landed on the opponent’s side of the table tennis table in exactly the desired way.
Test against professionals
To test their robot’s abilities under real conditions, the researchers had it compete according to official competition rules against five elite Japanese table tennis players who had been actively playing their sport for at least ten years and trained an average of 20 hours per week. And indeed: Ace won three out of five matches against the elite players. In games against the two table tennis professionals Minami Ando and Kakeru Sone, who play successfully in the Japanese professional league, Ace did well, but was unable to achieve a victory. Nevertheless, Ace demonstrated a sophisticated range of skills in all games, including skillful handling of spins and quick reactions to unusual shots, such as when the ball bounced off the net and unexpectedly changed its trajectory.
“These results highlight the potential of physical AI agents to perform complex, interactive tasks in real time, suggesting broader applications in areas requiring fast and precise human-robot interaction,” Dürr and his colleagues conclude. “Similar techniques could also be applicable in other areas with fast, real-time control and human interaction, including, for example, manufacturing and service robotics.” However, Jan Peters from the Technical University of Darmstadt, who was not involved in the study, is skeptical about its transferability: “The approach presented was developed specifically for table tennis. Therefore, it is unlikely that practical tasks will benefit from it,” he says.
Ace could possibly become a training partner for real table tennis players. Table tennis expert Kinjiro Nakamura, who competed at the 1992 Olympic Games, even thinks that people could learn new techniques from robots in the future – even if they don’t have an arm with eight joints or eyes around the playing field. Nakamura commented on a punch from Ace with the words “No one else would have done that. I wouldn’t have thought it possible. But the fact that it was possible means that a human could do it too.”
Source: Peter Dürr (Sony AI, Zurich, Switzerland) et al., Nature, doi: 10.1038/s41586-026-10338-5