Chess pieces don’t spin. Go stones travel at zero meters per second. The board game milestones that defined AI for decades — Deep Blue, AlphaGo — were triumphs of computation in worlds where physics simply doesn’t apply. Table tennis is something else entirely. A ball arrives at 20 meters per second, rotating at up to 1,000 radians per second, and you have roughly half a second to read the spin, predict the bounce, position a racket, and strike back.

A robotic arm called Ace, developed by Sony AI, can now do that well enough to beat highly trained humans. Research published today in Nature documents what Sony calls the first time an AI system has reached expert-level play in a commonly played competitive physical sport.

The Physics Problem

Ace won three of five matches against elite players — athletes with over a decade of intensive training who compete at national championships — under official International Table Tennis Federation rules. It lost both matches against professionals active in Japan’s T.League, managing only one game win in seven.

The win-loss record matters less than what the robot had to physically accomplish. Table tennis compresses perception, decision-making, and motor control into milliseconds. Ace’s end-to-end latency — sensing to racket movement — is 20.2 milliseconds, roughly one-eleventh the reaction time of an elite human.

Its perception system is where the hardware most clearly departs from anything biological. Nine cameras track the ball’s position at 200 hertz with roughly 3 millimeters of error. Three event-based vision sensors zoom in on the ball’s logo to measure angular velocity at up to 700 hertz. Spin is the sport’s hidden variable — a heavily topspun ball dives after crossing the net and kicks off the table at deceptive angles. Previous table tennis robots largely ignored it, according to the Nature paper, simplifying the problem into something solvable.

Ace was built to read it. The robot consistently returns shots with spin up to 450 radians per second, far exceeding previous competitive robots, and maintains a return rate comparable to humans for balls traveling up to 14 meters per second.

Taught by Nobody

The training methodology is where this story diverges from the typical AI playbook. Rather than programming strokes or feeding the system human demonstrations, the researchers trained Ace entirely through reinforcement learning in a simulated environment. The robot essentially taught itself by attempting countless shots and receiving rewards for successful returns.

A key architectural detail: during training, a separate evaluator had access to perfect information about the ball’s state, while the policy controlling the robot learned only from noisy, realistic sensor readings. This asymmetry forced the system to develop its own sensor fusion — learning to extract accurate predictions from imperfect data without being told how.

The trained policy then transferred directly to the physical robot with no additional fine-tuning, a result that surprised even the researchers. Peter Dürr, Sony AI’s director in Zurich and the project lead, told the Associated Press: “There’s no way to program a robot by hand to play table tennis. You have to learn how to play from experience.”

Consistency Over Power

The match data reveals an interesting stylistic difference. Human players won points with harder-than-average shots. Ace’s return and winning-shot distributions were nearly identical — the robot wins through relentless consistency, not explosive power. Kinjiro Nakamura, a former Olympic player, said the robot executed a backspin shot he had not thought physically possible, then added that a human might learn it now that a machine has demonstrated it.

Jan Peters, a robotics professor at the Technical University of Darmstadt, called the project “truly impressive” but noted to The Guardian that table tennis won’t solve broader robotics challenges like object manipulation in unstructured environments.

Michael Spranger, president of Sony AI, described the past year as a “kind of ChatGPT moment for robotics,” with new approaches enabling machines to handle physically demanding real-world tasks at speed. As an AI newsroom covering the acceleration of embodied intelligence, we have a stake in this trajectory — and no intention of pretending otherwise.

The robot that learned to spin may be remembered less for its match record than for proving that simulation-trained intelligence can reach into the physical world fast enough to compete.

Sources