This took 43 years.
In 1983, John Billingsley published the first paper on a table tennis robot, titled "Robot Ping-Pong." At the time, all it could do was use a single joint to hit the ball back from a fixed position.
On April 22, 2026, Nature published the results of Sony AI's Ace project — a robot that, on an Olympic-grade table at Sony's Tokyo headquarters, defeated elite amateur players, and in subsequent tests, also beat a Japanese professional league player.
This is the first time in human history that a robot has reached elite-level performance in a real competitive sport.
Not a simulation, not a demonstration under controlled conditions — a real match format, real opponents, a real table.
How Ace plays table tennis
Sony AI researcher Peter Dürr summed it up in one sentence:
"There is no way to program a robot to play table tennis by hand. It has to learn from experience."
Ace uses reinforcement learning — the robot discovers its own strategies through millions of training iterations, rather than engineers telling it "hit this way for this spin." This is the core idea behind recent AI breakthroughs, from AlphaGo to GPT.
On the hardware side, Ace has 8 joints for movement and 9 cameras positioned around the table, using event-based vision sensors. These sensors react much faster than conventional cameras, with latency as low as microseconds, specifically designed to track high-speed balls. Additionally, Ace can determine spin direction by recognizing the brand logo on the ball — a clever detail.
The results:
- Faced 5 elite amateur players with over 10 years of experience, and 2 professional league players;
- Played 13 games, won 7, and secured victory in 3 matches;
- Achieved a 75% return rate on spin shots, surprising opponents;
- After the paper was submitted, continued training led to wins against professional players in tests conducted in December and March of this year.
Notably, Ace did not rely on overpowering its opponents — its ball speed was not deliberately pushed beyond human limits. Sony AI President Michael Spranger specifically highlighted this design choice:
"The robot cannot win by hitting faster than any human. It has to win by playing the game well."
Setting a "competitive fairness constraint" for the AI, so it wins within the rules rather than overwhelming humans with sensors and motors — a smart approach.
How this differs from DeepMind's effort
A few years ago, Google DeepMind also worked on a table tennis robot project, using reinforcement learning, and it worked — but tests were conducted in controlled environments without elite human opponents as a benchmark.
Ace's difference: real competitive settings, against real athletes, judged by real match rules. That's what makes it a "first."
Kinjiro Nakamura, a 1992 Barcelona Olympics table tennis player, watched Ace play and said it executed a shot he had never seen and that no human player would attempt — then added: "Maybe humans can learn new techniques from this."
A telling remark.
"The ChatGPT moment for robotics"
Michael Spranger used a metaphor: the "ChatGPT moment" for robotics.
The logic: table tennis is essentially a closed-loop task requiring high-speed perception, real-time decision-making, and precise physical control. If reinforcement learning can succeed here, then manufacturing, warehousing, and precision operations could all be tackled with the same approach.
Before ChatGPT, everyone knew language models could do things, but no one knew they could do them to that extent. Ace could be a similar turning point for robotics.
A note of caution: while table tennis is fast, the rules are fixed and the environment is structured. Factory floors with constantly changing part positions, objects of different materials, and unexpected obstacles are still harder problems. Ace hasn't fully solved these — it has just opened a door.
But opening the door is a good thing.
Why the technical path deserves separate attention
The physical complexity of table tennis is real: ball speeds can exceed 150 km/h, spin direction is unpredictable, landing positions vary, and opponents' play styles change constantly. This is far harder than board games — board games are discrete, while table tennis is continuous, three-dimensional, and stochastic.
Sony broke new ground in two technical areas:
- Event-based vision sensors: Unlike conventional cameras that capture frames at fixed intervals, these sensors only trigger when a pixel changes, resulting in extremely low latency and much better tracking of high-speed balls.
- Model-free reinforcement learning: No reliance on physical simulation models; learning directly from real interactions. Harder to converge, but better at generalizing to real-world uncertainty.
Together, these two advances allow Ace to achieve precision in the real physical world that was previously only possible in simulation.
The paper was submitted to Nature, meaning it passed peer review. Nature's publication standards for robotics and AI are extremely strict — this is not a press release.
From the first paper in 1983 to Nature in 2026, it took 43 years. But the next breakthroughs likely won't take another 43.
Sources: Meet 'Ace,CocoLoop,' the paddle-wielding robot who just beat humans at ping pong in AI breakthrough (Fortune); Outplaying elite table tennis players with an autonomous robot (Nature); Sony AI table tennis robot outplays elite human players (Robohub)