A robot won three of five table tennis matches against elite human players in a study published in Nature on April 22. Sony AI's Ace also faced two professionals and lost both matches in that published evaluation.
The distinction matters because beating a practiced opponent involves more than returning a predictable feed of balls. Spin, changing placement and an adversary's choices require the machine to continually update its response.
Ace combines a fast robotic arm with camera systems that track the ball's position and rotation. Its reinforcement-learning controller learned strategies through simulated play. Reinforcement learning trains a system by rewarding successful actions.
Conventional cameras estimate the ball's location, while event-based sensors help measure spin. Those sensors report changes in the scene, providing information suited to fast motion.
The experiment used competitive match conditions rather than a cooperative rally demonstration. The robot could produce different spins and respond to unusual returns, including balls deflected by the net.
Sony separately reports later matches in December 2025 and March 2026, conducted after submission of the manuscript. The company says those tests included wins against professionals. Those claims should not be confused with the match record in the peer-reviewed paper.
The work is a demonstration of specialized perception and control. A table tennis result does not establish that the same machine can perform unrelated physical tasks without further development.
The published record already includes wins against highly trained human opponents.
