Sony built the first robot to beat elite ping-pong players

Sony AI's autonomous robot Ace won three of five matches against elite human ping-pong players under full official ITTF (table tennis federation) rules. Nature cover paper published today; since submission, Ace has beaten professional players in December and March rematches too. The retired professor who invented this entire research field in 1983 is calling Sony's approach "mob-handed." He's partly right. But he's mostly missing what actually happened.

The guy who invented robot ping-pong thinks Sony cheated

Well, "cheated" is too strong. But John Billingsley (the retired University of Southern Queensland professor who co-ran the first robot ping-pong contests back in 1983, back when the pop music charts were still wondering what was up with Duran Duran) took one look at Sony AI's new Nature cover paper and told the AP: "I would not want to belittle the achievement, but they have gone at the task mob-handed (British slang for throwing overwhelming numbers at a problem, like sending 30 people to do one person's job), and used sledgehammer techniques."

Here's what Sony actually threw at the task: Nine synchronized cameras. Three event-based vision sensors (specialized Sony chips that only report the pixels that CHANGE brightness between frames, instead of snapping full frames the way your phone camera does). A custom 8-joint arm. Five years of engineering. A physics simulator so carefully tuned it needed its own separate neural network just to fix the residual errors. All to beat a human holding a $30 paddle.

And it worked. Ace beat three out of five elite amateur ping-pong players (national-tier players with 10+ years of training, practicing 20 hours per week) under full ITTF rules (the International Table Tennis Federation's official pro rulebook, same one used at the Olympics), on an Olympic-size court (~14m × 7m) that Sony built inside its Tokyo headquarters, with two licensed umpires from the Japanese Table Tennis Association calling the match. That was April 2025. Since Sony submitted the paper, Ace has kept playing and kept winning: in December 2025 it beat both elites and one pro in a new round; in March 2026 it beat all three new professional opponents at least once.

The rematch tour did not go how rematch tours are supposed to go.

So: is this a breakthrough or brute force? Both, actually. But the reason it matters is neither the hardware nor the win-loss record. It's the way Ace wins.

First up, the TL;DR

Sony AI published a Nature cover paper today describing Ace, the first autonomous robot competitive with expert humans in a real sport under unmodified rules. April 2025: 3-2 vs elite amateurs, 0-2 vs Japanese pros. December 2025 rematch: beat both new elites AND one of two new pros. March 2026: beat all three new pros at least once.

Here's what happened:

• The matches. Under full ITTF rules on an Olympic-size court inside Sony's Tokyo HQ. Five elite amateurs and two pros (Minami Ando and Kakeru Sone, both in Japan's T.League, the country's top professional table tennis league) in April; four more in December; three more in March. Ace's win rate went up as opponents got tougher.
• How it sees. Nine cameras triangulate the ball (combine angles from multiple cameras to pin down exact 3D position) at 200 Hz (200 snapshots per second). Three event-based vision cameras watch the logo printed on the ball to measure how fast it's spinning. End-to-end latency: 10.2 milliseconds (roughly 30x faster than a human blink; humans react in ~230 ms).
• How it learns. A deep reinforcement learning (RL) policy (an AI trained by trial-and-error through millions of practice games; same family of methods that taught AlphaGo to beat the world Go champion) trained entirely in simulation (a physics-accurate video-game version of ping-pong Sony spent five years building) and deployed to the real robot with zero fine-tuning on the actual table.
• How it wins. Not with power. Ace's winning shots are statistically indistinguishable from its regular returned shots (p=0.88; in stats, p-values measure how likely a difference is to be due to random chance, so p=0.88 means "basically identical"). Human winning shots are dramatically harder than their averages (p<0.001, meaning "basically certain there's a real difference"). Ace wins because it almost never misses, returning 75% of shots up to 450 rad/s of spin (about 72 full rotations per second in flight). Humans have signature shots. Ace has no signature.

Why this matters: Every prior physical-AI milestone was narrower than Ace in at least one dimension. Gran Turismo Sophy (2022) beat elite drivers, but only in simulation. Autonomous drone racing (2023) beat humans in the real world, but on a fixed track with no opponent firing back. Ace is the first system to clear all three bars: real world, physical robot, adversarial opponent (adversarial meaning a human is actively trying to beat it back, not a puzzle you solve once). Sony AI chief scientist Peter Stone compared the research pipeline to the Apollo program: "not about immediate commercialization but the technologies that come out of it."

Our take: Billingsley's "mob-handed" critique has a point, but it misses the interesting finding. Yes, Ace has nine camera eyes and a custom arm. Yes, the simulator took five years. But the winning trick is not the hardware; it's that Ace plays differently than humans do. Ace wins by refusing to have a signature shot, forcing its opponent into a no-mistakes marathon at speeds humans can barely track. That's a new strategic category the sport had not seen before. Brute force gets you a robot. A new style of play gets you a breakthrough.

The actually weird thing: Ace wins by being boring

Here's the single most surprising sentence in Sony's paper, rewritten in English:

Ace's winning shots are statistically indistinguishable from its regular shots.

In human ping-pong, winning shots are measurably different from average shots. You wind up, you commit, you hit harder than usual when the opening shows up. Welch's t-test (a standard statistical test for whether two groups of measurements are really different or just look different by chance) on the paper's human-player data returns p<0.001. That's stats-speak for "basically certain the signature-shot phenomenon is real." Humans DO in fact beat opponents by hitting harder when it counts.

For Ace, the same test returns p=0.88. Essentially no difference. Ace has no putaway shot. Ace has no signature. Ace has no "this is the one" move.

Ace wins because it almost never misses.

This sounds boring until you watch what happens to a human trying to play against it. Ace returns 75% of incoming shots up to 450 rad/s of spin (about 72 full rotations per second while the ball is in flight). It holds up to 14 m/s ball speeds consistently (the speed of a hard top-spin drive off an elite pro's paddle). It hits the ball sooner after the table bounce than humans do. Its average rally runs 5.0 shots versus 3.9 in human-versus-human matches. It wears you out with rallies, not with blasts.

And this is a DIFFERENT style of play than any human uses. It's the ping-pong equivalent of how a computer plays chess against a human: not by being smarter per move but by never hanging a piece. When the robot forces you to play perfect ping-pong for 11 points in a row, you cannot. Nobody can. That's the whole trick.

The five elite amateurs Ace beat in April 2025 were not worn down by blistering top-spin rippers. They were worn down by getting every single ball back. Then making one mistake. Then making another. The boredom is the weapon.

This is the part that matters for the Billingsley debate. You can build a robot that's "mob-handed" with cameras and custom hardware. Sure. But once that robot demonstrates a style of play humans have never seen before, it is not brute force anymore. It is a new strategic option nobody had considered.

43 years of asterisks

Which brings us to why this paper matters in the first place.

The first robot ping-pong paper landed in 1983, co-authored by Billingsley himself at the University of Southern Queensland. For 43 years, every paper that followed came with asterisks. Researchers swapped human opponents for ball-launchers. They shrunk the court. They deleted serves. They ignored spin entirely (even though spin is roughly half of competitive ping-pong; ignore spin and you're not playing the actual sport). They modified the rackets. They restricted shot types. One corner-cut at a time, the community built up the pieces of the problem without ever playing the actual game.

Google DeepMind came closest in April 2025 with a paper titled "Achieving human level competitive robot table tennis." Real progress, and still with modifications (restricted playing area, simplified setups). Sony's paper cites Google's work directly and politely contrasts: Ace plays with no such constraints. Full ITTF rules. Full court. Free choice of racket (Ace uses a stock Butterfly Dignics 05 rubber on a modified VICTAS ZX-GEAR OUT blade, same gear a pro would pick off the shelf). Human serves. Licensed umpires. Nittaku 3-star competition balls. The only concession: the "golden point" rule (first to 11 wins a game), which Japan's T.League already uses in actual pro matches.

For the first time, "robot beats human at a sport" means what you think it means. No hidden accommodations.

Sony AI president Michael Spranger put the philosophy plainly: "It's very easy to build a superhuman table tennis robot. You build a machine that sucks in the ball and shoots it out much faster than a human can return it. But that's not the goal. The robot cannot just win by hitting the ball faster than any human ever could; it has to win by actually playing the game."

How the machine actually works

Three key ideas. Light on jargon, heavy on why each choice matters.

1. Ace sees the ball with two different kinds of cameras at once.

Nine conventional cameras (Sony IMX273 sensors) triangulate the ball's 3D position 200 times per second, with 3mm average error. Three gaze control systems (each one a specialized rig combining a pan-tilt mirror, a tunable telephoto lens, and an event-based vision sensor, which is Sony's IMX636 chip that reports only pixels that CHANGED since the last frame instead of the whole picture) lock onto the ball and watch the printed logo as it spins. Two neural networks read the spin from how the logo rotates: one fast, one more accurate, both running in parallel. If one gaze system loses the ball mid-rally, the other two cover.

End-to-end latency: 10.2 ms for position. Human reaction time: ~230 ms. Ace sees the ball 22 times faster than a human does. That number is measured, not guessed.

2. Ace was trained entirely in a video-game version of itself.

All of Ace's ping-pong skill comes from a deep reinforcement learning (RL) policy trained in a physics simulator Sony spent five years building. The algorithm is Soft Actor-Critic (a specific well-known RL method; think of it as the "brand name" choice in the RL world), plus an asymmetric actor-critic twist borrowed from Sony's earlier Gran Turismo Sophy racing work. That twist in plain English: during training, the "coach" part of the network sees perfect information about the ball, while the "player" part sees only noisy sensor readings. Picture coaching a player by showing them the game from both their own seat AND the press box at the same time. Coach sees everything; player is learning to perform under fog-of-war.

Sony did not train one giant policy. It trained a library of policies: one that specializes in topspin returns, one for backspin returns, one for aiming at specific corners, one for playing aggressive, one for playing safe. During a match, a policy sampler (a small selector network) picks the right specialist policy based on the incoming shot. This is closer to how humans actually think about shot selection than a single giant network would be.

The final trick: domain randomization. The simulator deliberately varies lots of conditions (ball weight, air density, camera calibration errors, racket friction coefficients) so that when Ace shows up at the real table, real-world variation looks like just another draw from the training distribution. No real-world fine-tuning needed. This is the thing that makes "trained only in simulation" a real claim rather than marketing.

3. The arm is a human-peer, not a superhuman.

Ace has 8 degrees of freedom (8 independent ways it can move; for comparison, a human shoulder+arm has about 7). Two joints slide along tracks for gross positioning; six rotate for precision. End-effector velocity (the business end, where the racket lives): 20 m/s, about the speed of a pro's drive shot. Workspace: 3.6m × 3.6m, matching the area pros actually use for most shots.

Sony deliberately built a human-peer arm, not a superhuman one. The links were designed using topology optimization (a structural design method that removes material from anywhere it is not needed; picture 3D printing with fancy subtraction) and additively manufactured in Scalmalloy (a high-strength aerospace aluminum alloy). Every actuator syncs at 1 ms intervals and shares a clock with the cameras, so every sensor reading and every joint command lives on the same timeline. The motion planner guarantees provably collision-free trajectories (mathematically impossible for the arm to hit the table or itself), which is table stakes (sorry) for a machine swinging a racket at 20 m/s next to a human being.

The rematch tour nobody expected

April 2025: Ace 3-2 vs elite amateurs, 0-2 vs Japanese pros Minami Ando and Kakeru Sone. The paper goes into peer review. Good story. First robot to beat any tier of serious human ping-pong players under real rules.

Then Sony kept experimenting.

December 2025: Four new opponents (two elite amateurs, two pros). Ace beat both elites and one of the two pros. First time any robot had beaten a professional table tennis player in a full match under unmodified rules.

March 2026: Three new pros. Ace beat all three at least once, with measurably faster shot speeds and more aggressive placement than April. Sony says Ace played closer to the table edge (a technical shift pros make to shorten reaction windows on their opponents) and pushed rally pace harder in every rematch round.

The arc of Ace's first 11 months looks like this: Lose to pros. Beat one pro. Beat every new pro at least once. It keeps improving between test sessions even though the core policy is not being updated mid-match. Partly Sony iterates on the simulator between rounds (better physics, more training). Partly the specialist policies that generalize across opponents get stronger as more human shots get incorporated into the training data.

The surprise is not that Ace won in April. The surprise is how fast the curve is bending.

What this means for physical AI

Peter Stone, Sony AI's chief scientist, compared Ace to Apollo: "Like the Apollo mission, it's not about immediate commercialization but the technologies that come out of it."

Translation: do not judge Ace by what ships next quarter. Judge it by what the pipeline unlocks over the next decade.

Here's the case for taking Stone seriously. Every prior physical-AI milestone was narrower than Ace in at least one dimension:

• Gran Turismo Sophy beat elite drivers in 2022, but only in a racing game simulator. No real car. No real human in the seat next to it.
• Champion-level autonomous drone racing happened in the real world in 2023 (AI-piloted drones beating human champions on first-person-view obstacle courses), but on a fixed pre-mapped track with no moving opponent.
• Robot soccer leagues exist; nothing in them beats humans.
• Autonomous driving works in many conditions, but is not beating experts at anything. It is trying to match a safe human baseline, not outperform experts.

Ace clears all three bars at once: real world (not simulation), physical robot (not a puzzle-solver), adversarial opponent (a human actively trying to beat it back). The techniques that got it there span several domains: event-based vision for low-latency sensing, simulation training with domain randomization, asymmetric actor-critic RL, and convex optimization (a well-understood math technique for finding the best smooth path between two points) layered on top for safety. All of it general purpose. Not ping-pong specific.

Sony AI president Michael Spranger called the past year "a kind of ChatGPT moment for robotics": a burst of new AI-driven approaches teaching machines about real-world environments fast enough that the catalog of things robots can suddenly do surprises researchers every month. Backflips. Kung-fu demos at Chinese New Year festivals. And now competitive ping-pong against pros. High-speed manufacturing, emergency response, service robotics; they are all versions of the same core problem Ace solved. Perceive, decide, act in the 10-millisecond window a human cannot.

But Billingsley was right about one thing

Four honest caveats. The first belongs to the field's grandfather.

The mob-handed critique. Ace has nine cameras and three event-based gaze systems. A human has two eyes. Ace has a custom 8-joint arm tuned to the exact physics of ping-pong. A human has a skeleton. Whether "AI beats humans" is a fair framing when the AI has nine eyes and a custom skeleton is a legitimate question. What Sony actually demonstrated is that specialized hardware plus reinforcement learning in a carefully-tuned simulator can solve a narrow interactive task. Whether that pipeline generalizes to domains where you can't spend five years building a bespoke simulator is the open question.

The April 2025 elites are very good but not world-best. Ten-plus years of training, national championships, 20 hours a week. Not Olympic medalists. The April pros beat Ace. The Nature paper does not overclaim on this point, but headlines will.

"Trained only in simulation" is doing a lot of work. Technically true. Also: Sony's simulator is a five-year hand-tuned artifact. It uses velocity-dependent Magnus coefficient models (the Magnus force is the aerodynamic effect that makes a spinning ball curve in flight; it is why a curveball curves and why ping-pong spin matters, and Sony needed a more sophisticated version than the standard textbook model to handle 450 rad/s pro-tier spins). Ball-racket contact models that required their own residual-correction neural network (a small AI layer that patches the errors the physics model couldn't cover) to fix 4% of systematic error. Explicitly modeled sensor noise and dropout. This is NOT zero-shot transfer from a generic robotics simulator. It is a carefully crafted sim-to-real pipeline for one specific sport. Each new domain will need its own tuned simulator, not a plug-and-play physics prior.

Sample sizes are still tiny. Sony evaluated against 12 humans total across three rounds. The statistical significance of "Ace beats pros" is limited. The more informative test is what happens when the same pros face Ace five or six times and start studying Ace tape. That hasn't happened yet.

None of this undercuts the achievement. All of it tempers the claim.

Where this goes next

Sony's roadmap points two ways.

Opponent modeling. Ace currently tracks the ball; it has no model of the human across the table. The next generation will model opponent tendencies (is their backhand weak? Do they over-commit on the first shot?). This is where human pros actually make their living, and it is the last competitive edge humans hold over the robot.

Online learning. Ace's current policy is fixed during a match. Future versions will update in real time from match experience. The counter-adaptation loop between a self-improving robot and a human studying its habits is the most interesting dynamic to watch over the next few years.

Meanwhile, Google DeepMind is still working on the same problem. Expect a response paper from Mountain View within 12 months. Expect Sony's follow-up either as an Ace extension or as a new domain entirely. Spranger called this the "ChatGPT moment for robotics," and the whole point of a ChatGPT moment is that the next thing arrives faster than anyone expected.

The tidy summary, via ScienceAlert: "First Pong, now ping-pong." A 1972 Atari video game with two lines and a dot has become, in 54 years, a real robot hitting a real ball against a real Olympian in a real sport, on the cover of Nature. In another 54 years the robot will almost certainly beat the Olympian every time. The interesting question is what happens in the decade between.

The bar just moved. Next time someone publishes a "robot beats humans at a sport" paper, the standard will be "full rules, no modifications, adversarial opponent, plus a year of rematches." Sony set it.