Peter Dürr could barely track the ball as it blurred across the net. On one side: Taira Mayuka, one of the world’s top table-tennis players. On the other: a robotic arm called Ace.
Mayuka launched a twisting smash that should have ended the rally. Ace answered with a return no one saw coming.
Table tennis is a brutal sport — players read spin exceeding 160 rotations per second and react in fractions of a moment. What made Ace’s performance remarkable, though, wasn’t raw speed. It was something far less expected about how the robot chose to play.
A robot arm versus the world’s best
Sony AI’s Ace isn’t a humanoid or a full machine — it’s a single robotic arm. Yet in competitive matches against seven top-tier human players — five elite and two professional — it beat three of them outright. That result places Ace in a lineage stretching back to Deep Blue’s 1997 defeat of chess champion Garry Kasparov, through AI victories in Jeopardy, Go, and StarCraft II. Those were all virtual contests. Ace brought AI’s winning streak into the physical world.
Independent researchers Carlos Ribeiro and Esther Colombini, from the Aeronautics Institute of Technology and the University of Campinas respectively, called the achievement an “important milestone” for robotics. Ace arrived the same week a humanoid robot broke the half-marathon world record in Beijing — a coincidence that illustrated just how quickly autonomous machines are moving from controlled labs into messy, unpredictable environments.
Eyes, brain, and a custom-built arm
Ace’s performance starts with perception. High-speed cameras positioned around the court track the ball’s position roughly 200 times per second, while a separate event-based image sensor reads spin — together giving the robot enough information to anticipate where the ball is heading before most humans could consciously register it.
That sensory data feeds into several AI algorithms working in concert. One, borrowed from image processing, uses an attention mechanism to focus on the most relevant parts of each frame, speeding up analysis. Another is a deep reinforcement learning model trained across thousands of simulated hours — learning through trial and error, gradually building a library of effective shots without ever being explicitly shown how to play.
Then there’s the arm itself. Off-the-shelf hardware wasn’t fast enough, so the team built a custom six-jointed, lightweight arm capable of whipping a racket at over 20 meters per second. It reacts roughly 11 times faster than a human — not because the AI thinks faster, but because the physical system was engineered to close the gap between decision and motion.
Winning through invention, not just speed
Here’s where Ace defies the obvious assumption. A robot beating humans at table tennis sounds like a story about mechanical advantage — faster reflexes, tireless repetition. That’s not what happened.
Ace won by being creative. It varied its spin, improvised returns, and placed the ball with consistent precision, introducing shots its opponents hadn’t prepared for rather than relying on predictable power. Olympic player Kinjiro Nakamura, watching from the sideline, was visibly struck. “No one else would have been able to do that,” he said. “I didn’t think it was possible.”
The robot’s limits are equally telling. Ace dominated the five elite players but lost to both professionals — a clear ceiling the team openly acknowledges. The study notes that Ace has continued improving in the months since those results were written up, which suggests the gap with professional-level play is narrowing rather than fixed.
Sports as a proving ground for the real world
Nobody at Sony AI is trying to build the world’s best table-tennis player. The ambition runs considerably larger.
“We wanted to prove that AI doesn’t just exist in virtual spaces,” said Michael Spranger, president of Sony AI. The physical experience, he argued, is the point — demonstrating that AI is ready to operate in environments where timing, friction, and unpredictability can’t be abstracted away.
Colombini, whose research background includes soccer-playing robots, frames it more directly. Sports are a proxy — a legible, observable environment where researchers can watch robots develop agility and improvisation in real time. Those same capabilities are what autonomous machines will need to work safely alongside people in factories, hospitals, and homes. The table-tennis court isn’t the destination. It’s a stress test.
What comes next
Ace’s story isn’t finished. With the robot still improving post-study, and humanoid machines already competing in endurance athletics, the pace of development in physical AI is accelerating in ways that were difficult to anticipate even a few years ago.
The more interesting question may not be how fast these robots get, but how inventive. Nakamura’s reaction — that a human watching Ace might now attempt shots they’d never previously considered — hints at something unexpected: robots may end up expanding what human athletes believe is possible, not just competing with them. That dynamic, if it holds, would be the real development worth watching.