Look, we’ve all seen those robot arms do the little demo where they catch a ball or stack cups or whatever. Cute. Science fair stuff. Then Sony AI rolled out Ace and had it play actual table tennis under official rules against ranked professionals. Not “we let it keep the ball going for a while.” Full matches. Umpires. Scoreboards. And it started taking matches off real players.
The one that lit everything up was Miyu Kihara. World No. 26 in women’s singles at the time. On March 25, 2026, Ace went out there and beat her. The numbers floating around from the session are the kind that make you do a double take: sets like 11-7, 11-8, and then more later. This isn’t some YouTuber with a paddle and a dream. This is a Japanese pro who plays the international circuit. Ace took the matches.
A few weeks later, it did the same to Miu Hirano, the two-time Olympic silver medalist. After that one she basically just asked, “It’s really strong. Is there really anyone who can beat this?” That’s not a polite post-match quote. That’s someone who just got the experience of playing something that doesn’t get tired, doesn’t tilt, and doesn’t telegraph anything.
How the hell does it actually do this?
Ace is not a humanoid. It’s a custom eight-joint robotic arm sitting on one side of the table like a very expensive, very fast factory tool that learned how to play. Three joints get the paddle where it needs to be, two orient it, and three handle the power and the spin. Topology-optimized, lighter than earlier versions, with upgraded motors so it can accelerate harder and put more work on the ball.
The vision system is the real cheat code. Nine regular high-speed cameras (Sony’s own sensors) triangulate the ball in 3D at 200 frames a second with roughly three-millimeter accuracy. On top of that, three separate event-based vision systems track the logo on the ball itself so the robot can read spin in real time. The whole perception loop was running around 10 milliseconds; after more work, they got ball detection down closer to 8.5 ms. End-to-end, from seeing the ball to actually swinging, the system is in the 20-millisecond range. Elite humans are around 230 milliseconds. It’s not “a little faster.” It’s operating in a completely different time scale.
Because it trains almost entirely in simulation—thousands of hours of synthetic play, reinforcement learning with a policy that decides what kind of shot to hit (placement and spin) and then a lower-level controller that executes it—it doesn’t play like a human. No body language. No wind-up you can read. It just gets to the spot and hits. Hirano talked about this specifically: when you play a person, you can feel when they’re rattled or when they’re hunting a particular ball. Against Ace, there’s none of that. It resets every point. It doesn’t have a “tell.”
The Sony team also improved the physics model in sim so the spin, bounce, and aerodynamics transfer better to the real table, and they added rewards that encourage the arm to start moving early toward where the ball is going to be instead of waiting and then lunging. Combined with the hardware upgrades, the thing got meaningfully faster and more consistent after the original Nature paper results.
The “is this even fair?” part
Players who faced it have been pretty open. Hirano said she thinks it already has top-10 women’s potential and maybe around the top 50 on the men’s side, because male pros can still generate more pure power and serve quality that the robot hasn’t fully matched yet. She also said it would actually be useful as a training partner at the highest levels, which is wild when you remember that most ball machines are still just glorified launchers.
The other thing people keep coming back to is the psychology. You can’t mind-game a robot. You can’t get it frustrated. You can’t watch its shoulders or its eyes. One of the earlier elite players who both beat and lost to it noted that simple no-spin serves sometimes gave it more trouble than heavy spin, because the system is so good at reading and answering spin that a knuckleball throws the expected pattern off. Humans are constantly adjusting to what the other person seems to want or fear. Ace just solves the current ball.
What this actually is
This is the first time an autonomous robot has gone out and beaten professional-level players in a real, adversarial, high-speed physical sport under official competition rules. Not a demonstration. Not a lab. Matches with umpires. Table tennis has been a robotics challenge for decades precisely because the ball is small, fast, spinning like crazy, and the reaction window is tiny. Ace cleared the elite level first, then, after more training and hardware work, started clearing ranked pros.
It’s still specialized as hell. The cameras are all around the table, the arm is fixed to one side, and it doesn’t have to do anything except play this one game. That doesn’t make the result less real. It just means the next versions will be about packing that same perception and control into something that can move around, or applying the same low-latency vision-plus-RL stack to other tasks where things are moving fast and you can’t afford to be late.
For now, the videos are the proof. A pro is out there doing everything she’s trained her whole life to do, and on the other side of the net a robot arm with no face and no feelings is just calmly putting the ball back in places she doesn’t like. And it’s not missing.
That’s the part that feels different. The robot isn’t cheating the rules. It’s just better at the parts of the game that used to be reserved for the best humans.
A note on comments: Our advertisers enforce a strict zero-tolerance policy on racism, hate speech, and threats of violence. Any comment that crosses that line will be removed and the account permanently banned — no warnings. Debate hard, but keep it civil.