TL;DR:

• Sony AI’s Ace robot beat three of five elite players at table tennis under official rules, losing both matches against professionals — a robotics milestone published in Nature this week.
• The result extends the pattern of AI beating humans in increasingly complex real-world tasks, but researchers explicitly warned it does not solve the harder robotics problem of general object manipulation.
• For UK businesses evaluating physical-AI vendors, the headline matters less than the implicit forecast from Darmstadt professor Jan Peters: the ChatGPT-scale moment for embodied AI may arrive this decade, not next.

Sony AI has unveiled Ace, a table-tennis robot that won three out of five matches against elite players and one of seven games across matches with two professionals. A paper on the system was published in Nature on Wednesday. Project lead Peter Dürr, director of Sony AI in Zurich, said the robot had continued to improve since the paper was submitted.

How Ace actually works

The design sidesteps several hard robotics problems. Rather than two eyes and two legs, Ace uses an eight-jointed arm on a movable base and multiple fixed cameras that track the ball’s position and spin from different angles. By zooming in on the ball’s logo, the camera system infers spin and axis of rotation in the few milliseconds before the ball reaches the robot’s side. Policies for dealing with spin and shot selection were learned through 3,000 hours of simulated play, while serves drew on expert human technique.

Elite player Rui Takenaka noted that Ace mirrors the incoming ball’s complexity: complex spin returned complex, simple “knuckle” serves returned simply — a weakness that let him win. Former Olympic player Kinjiro Nakamura described one rapid early-intercept backspin shot as something he had not thought physically possible.

The narrower robotics reality

The Nature result sits in a specific category of AI-beats-humans achievements — chess, Go, poker, StarCraft, Breakout — that have steadily expanded into real-world domains. What Ace demonstrates is that fast perception and learned motor policy can hold up against elite human athletes in a constrained-rules sport. What it does not demonstrate is progress on the harder robotics problem of manipulating unfamiliar objects in unstructured environments. Jan Peters, a professor at Darmstadt who has worked on table-tennis robots, said research of this kind will not solve those challenges — but he also predicted that the ChatGPT-equivalent moment for physical-world robotics may be “closer to now than to 2036”.

Looking forward

For UK manufacturers, logistics operators and research groups tracking embodied AI, the useful signal is not the match score but the compression of the research-to-capability timeline. The Alan Turing Institute has been funding comparable multi-camera perception work; Dyson’s robotics lab in Wiltshire has published on related sim-to-real transfer. Peters’ timeline prediction, if even roughly accurate, puts physical-AI procurement decisions firmly inside the 2026–2030 planning window rather than the speculative long term.