- A robot just beat some elite table tennis players
- Sony AI’s Project Ace is good at competing against unpredictable human players
- Success here could mean it’ll be easier to use AI to train future robots to handle the real world
In competitive table tennis, the ball can travel at speeds of up to 70mph, and it can go anywhere. Sure, there’s some predictability based on the strike, spin, and how the ball hits the table, but there are also infinite possibilities that now, it appears, a robot has mastered.
Sony AI’s Project Ace is the first robot to beat multiple elite-level table tennis players in an International Table Tennis Federation-style arena and under the watchful eyes of licensed referees.
In a new Nature Article, Outplaying elite table tennis players with an autonomous robot, Sony AI scientist describe their work and how they built and used AI to train a robot, “Project Ace”, to not just play table tennis, but do so at a pro-level.
Article continues below
“Ace achieved three victories in five matches against elite players, along with competitive performances in the remaining matches. These results demonstrate the potential of physical AI agents to outperform human experts in interactive, real-time tasks,” wrote the scientists.
Project Ace is a canny combination of “high-speed perception,” a control system based on reinforcement learning (rewarding good behavior), and “high-speed robot hardware.”
No feet, but a wicked backhand
Ace doesn’t look like a human ping pong player you’ve ever seen. Instead, it glides in four directions on a custom track system, while its trunk rotates 360 degrees, and the fully articulated arm and wrist adjust on the fly to both serve and return the ball. You may have seen table-tennis-playing robots before (I recall seeing a lumbering one at CES 2026), but not like this. The speed alone is astonishing.
Still, it’s the AI-based reinforcement learning and training simulation that makes Project Ace special and successful. In training, it was able to game out all sorts of play scenarios. It even practiced against a virtualized version of itself. But it’s the “model-free” reinforcement learning that’s, at least in part, allowing Project Ace to adapt to unpredictable, elite, human competitors.
Equipped with on-board sensors and an array of nine cameras positioned around the robot, Project Ace can see things that most human competitors, even elite players, might miss. The ball spin, for example, is a determinant of where the ball will go next.
As the researchers explain in the project video, perception is one of the key innovations, “So it’s the only system in the world that can measure spin of an unaltered table tennis ball at this speed.”
Perhaps the secret sauce here, though, is a technique called “privileged critic”, which Sony AI developers used within the training simulations. The privileged critic accessed perfect-match information, which is married to live sensor data. It might be said that it’s comparing what should happen with what does. That learning is how the robot prepares for the unexpected.
There’s a moment in the video where you can see this at work. The elite human player hits a ball that catches the net, sending it careening in a different, perhaps unanticipated direction. Project Ace clearly already had a return planned, but it managed to adjust to the ball’s new trajectory and hit a return. It all happens within milliseconds, and one might argue that a human player would’ve failed to make that same, rapid adjustment.
“It totally blew my mind,” said Sony AI Director and Lead Engineer Peter Dürr in a release on Project Ace.
Like Sony AI’s previous project: teaching the AI how to beat expert human-level players in a Gran Turismo simulation, Project Ace is not about beating pro players and leveling up to Olympic-class tablet tennis players. This is about helping robots operate in an unpredictable world.
Most people who watch humanoid robots operate in home environments comment on their speed, or lack thereof. The robots move deliberately for safety and to manage the unexpected. Project Ace, though, proves that robots can be trained and train themselves to manage an unpredictable world, and at speed.
Also, future table tennis competition is not totally off the table. After all, the Sony AI team is constantly working on improving Project Ace’s game. They note, for instance, that the robot has a tendency to move in and hit earlier than human opponents. Sometimes, stepping back and waiting a beat can provide for a more strategic return.
If they solve for that (or maybe Project Ace solves it on its own), why should the Olympics be off the table?
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
https://cdn.mos.cms.futurecdn.net/qxbkcwLtJhJU7VnNx24YLf-1920-80.jpg
Source link
lance.ulanoff@futurenet.com (Lance Ulanoff)
