Sony's new robot can beat pro ping-pong players

Table tennis has long been robotics’ ultimate nightmare: a 40 mm plastic ball hurtling at over 20 meters per second, spinning at more than 9,000 revolutions per minute, leaving mere milliseconds to read physics and strike.

Sony AI took the cover of Nature to reveal Ace, an autonomous table-tennis robot that just cracked that exact loop. Playing under official ITTF rules, Ace not only racked up wins against elite athletes but, for the first time, defeated professional players on a full-size court.

Sub-millisecond eyes on a spinning ball

Human pros anticipate spin by reading the opponent’s shoulder, paddle angle, and subtle body twitches. Ace has to read the raw physics directly out of thin air.

To pull this off, Sony built a hybrid vision suite packed with 12 high-speed sensors. Nine Sony Pregius IMX273 active pixel cameras run at 200 Hz to triangulate the ball in 3D. Meanwhile, three IMX636 event-based vision sensors track micro-changes in brightness at sub-millisecond precision, locking onto the logo printed on the ball to gauge spin exceeding 160 rotations per second.

This setup drops overall perception latency down to just 10.2 ms. End-to-end, Ace calculates and swings in roughly 20 ms—about ten times faster than human biological limits.

3,000 hours in sim, zero human demos

Seeing the ball is only half the battle. Swinging an eight-joint industrial arm with pinpoint precision in live rallies requires flawless motor control.

Sony AI leveraged the reinforcement learning architecture pioneered in its GT Sophy racing project. Ace trained entirely inside a physics-accurate simulation, battling copies of itself for roughly 3,000 hours without human demonstrations before transferring directly to physical hardware.

The custom arm features two prismatic and six revolute joints, complete with a vacuum cup on the end effector for autonomous one-armed serves. Its low-level control loop ticks at an astonishing 1 kHz, allowing the paddle trajectory to adjust mid-swing to return balls at speeds up to 19.6 m/s.

During match trials, Ace baffled elite players and veteran pros alike. One 1992 Olympian was reportedly left flat-footed courtside by an unreturnable backspin slice. Human opponents noted that Ace is uniquely terrifying because it has no facial expressions, no tells, and zero competitive anxiety.

Why it matters

We think Ace represents a monumental milestone for physical AI that goes far beyond ping-pong.

For athletes and coaches, a room-scale rig with industrial arms and dozens of cameras is obviously not hitting consumer living rooms anytime soon. But for elite training centers, it provides something that has never existed: a sparring partner with superhuman consistency that can quantify spin and trajectory down to the millimeter.

For the robotics industry, Sony’s approach is a masterclass in Sim2Real transfer. Most humanoid and industrial robots today operate in forgiving, low-speed environments where hesitation costs nothing. Table tennis permits zero lag. Proving that an agent trained strictly in virtual simulation can master millisecond-level physical dynamics in the real world validates a crucial playbook.

The limitation is obvious: Ace relies on a tightly controlled indoor arena with high-speed sensor rigs mounted around the table. Translating this agility onto mobile platforms operating in chaotic outdoor environments remains a massive leap. Still, the core perceptual-motor loop is solved—and high-speed automated manufacturing, drone interception, and emergency collision avoidance will likely borrow heavily from Ace's playbook.