Sony AI's Ace Becomes the First Robot to Beat Elite Table Tennis Players in Match Play, in a Nature Study
Sony AI's Project Ace, published in Nature on April 22, 2026, is the first autonomous robot to win competitive matches against elite and professional table tennis players under official rules.
Overview
Sony AI says it has built the first autonomous robot capable of competing with elite and professional human table tennis players under official rules, and that claim is now backed by a peer-reviewed paper in one of the world’s most-read science journals. The research paper, Outplaying elite table tennis players with an autonomous robot, was published in Nature on April 22, 2026, describing a system the team calls Ace that combines high-speed event-based vision, model-free reinforcement learning, and a custom robot arm to return spin-heavy serves and rally with humans on a regulation-size table.
The milestone, reported by Fortune and covered on video by Euronews, pushes physical AI past a long-standing benchmark: matching human reaction time and decision-making in a fast, unscripted competitive sport. Earlier table tennis robots have rallied with hobbyists and demonstrated impressive shots, but the Sony AI team’s headline result is that Ace won matches against trained athletes refereed by certified umpires, not just demos.
What We Know
The Nature paper and the experimental setup
The Nature study evaluated Ace against five elite players, each with more than ten years of active table tennis experience, and two professional players from the Japanese league, according to the paper. Matches followed International Table Tennis Federation (ITTF) rules and were officiated by licensed umpires, as detailed in the paper and confirmed by Interesting Engineering. The paper reports that Ace won three of five matches against the elite cohort, while losing both matches against the two professionals. Across the elite series, Ace took seven of thirteen games, according to the Nature paper.
Follow-up matches conducted after manuscript submission extended the result. Interesting Engineering reports that Ace continued to face top human players, demonstrating improved rally speed and shot placement.
How the system works
Unlike scripted industrial arms, Ace is built around end-to-end perception and learned control. According to the Nature paper, the robot relies on a multi-camera high-speed tracking rig that follows the ball in 3D, supplemented by event-based vision sensors, and uses a custom robotic arm mounted on a mobile base to swing and serve under ITTF conditions. Interesting Engineering’s coverage describes the system at a higher level as using nine high-speed cameras, event-based sensors, and reinforcement learning, and notes that the robot reacts in milliseconds.
The perception pipeline is decisive. The Nature paper describes a system designed for very low end-to-end latency between observing the opponent’s contact and committing to a stroke, which is what allows a non-humanoid form factor to keep pace with athletes whose physical reach and footwork are far superior. Trained through deep reinforcement learning rather than hand-coded playbooks, Ace adapts mid-rally to the observed flight and spin of each return, as Interesting Engineering also describes.
What the team is saying
“There’s no way to program a robot by hand to play table tennis. You have to learn how to play from experience,” said Peter Dürr, who leads Sony AI’s Zurich lab and the project, in comments quoted by Fortune. Fortune reports that the matches were played under official table tennis rules and that Ace’s strongest opponents included professional players from the Japanese professional league, who beat Ace in the manuscript-stage matches.
Why It Matters
Table tennis is a deliberately demanding test bed for physical AI. The ball is light, fast, and heavily affected by spin; opponents adapt to weaknesses in real time; and there is no opportunity for the cloud round-trips that flatter many lab demonstrations of robot learning. A system that wins under these conditions has solved a tightly coupled problem in perception, control, and policy learning at the same time.
The paper’s own framing is guarded. The authors describe Ace as, in their words, “to our knowledge the first real-world autonomous system competitive with elite human table tennis players.” In other words, this is not yet a robot beating the world’s best, but it is the first peer-reviewed evidence that a learned system can win refereed matches against players at the upper end of the recreational competitive ladder. Robotics researchers have spent years arguing that the next frontier of AI lies outside text and pixels, in physical interaction. A Nature paper showing that reinforcement learning plus modern event-based sensors can deliver competitive play at this level is a concrete data point for that thesis.
The result also lands in a competitive moment for the field. The Machine Herald has tracked an unusually dense run of physical-AI announcements this month, including Physical Intelligence’s release of the steerable robot foundation model 0.7 and Hyundai’s plan to scale Boston Dynamics’ Atlas to 30,000 units per year. Where those efforts focus on general-purpose humanoids and broad foundation models, Ace is a sharply scoped system optimized for a single fast task — a different bet on how physical AI gets demonstrated and de-risked.
What We Don’t Know
Several questions remain open. The Nature paper documents Ace’s record against seven specific players in tightly controlled conditions; how the system would fare in a longer, multi-day tournament against many opponents adapting to its weaknesses is not addressed in the published results. Coverage to date does not detail how often Ace must be retrained or recalibrated between sessions, nor how much of its performance generalizes across paddle types, table conditions, or lighting changes outside Sony AI’s test environment described by Euronews.
Nor is it clear how the techniques transfer beyond the table. Sony AI frames Ace as a step toward broader physical AI applications, but the hardware — a high-rate camera array around a single court, an arm tuned for one motion class — is purpose-built. Whether the same control stack, retrained, would carry into less constrained settings such as kitchens, factories, or warehouses is the open question that will determine how far the result travels beyond a sports demonstration.
For now, what is settled is narrower and still significant: a refereed, peer-reviewed match record showing that a learned robotic system can win against elite humans in a real-world sport, with publication in Nature on April 22, 2026 marking the milestone.