Sony Ace: the first autonomous robot to beat professional table tennis players — published in Nature
🔎 Why table tennis just became the new Turing test for robotics
On April 22, 2026, a study was published in Nature. It was not about a new language model or a revolutionary molecule. It documented the first defeat of professional table tennis players at the hands of an autonomous robot.
The system is called Ace. It is developed by Sony AI. And it just solved a problem that robotics has been struggling with for over forty years: making a machine interact with the physical world at the speed and precision required by a racket sport.
The stakes go far beyond ping-pong. Every rally illustrates exactly what separates an LLM sitting in a data center from an agent that must react in 20 milliseconds to a ball traveling at over 100 km/h with unpredictable spin. This is the entire promise of physical AI — embodied artificial intelligence — playing out on a 2.74-meter-long table.
The essentials
- Ace, developed by Sony AI, is the first autonomous robot to beat professional-level players (and a former Olympian) at table tennis, according to a study published in Nature on April 22, 2026.
- The system combines an 8-joint robotic arm, proprietary Sony high-speed sensors, and training via deep reinforcement learning entirely in simulation before transfer to the real world (sim-to-real).
- Ace beats 3 out of 4 elite-level players during official tests, with a reaction time of 20 ms — well below the ~200 ms of a human.
- This breakthrough follows in the lineage of GT Sophy, the racing AI of Gran Turismo developed by Sony AI using the same simulation → reality approach.
Recommended tools
| Tool | Main use | Price (June 2025, check on ace.ai.sony) | Ideal for |
|---|---|---|---|
| Ace Research Project | Research in physical AI and sports robotics | Non-commercial (research project) | Understanding Sony's sim-to-real architecture |
| GT Sophy | Agentic AI for simulated car racing | Integrated into Gran Turismo 7 | Seeing Ace's conceptual lineage |
What Nature really validated — and what it did not validate
Nature is not a popular science magazine. Publishing in this journal implies a strict peer review: methodology evaluated by independent researchers, reproducible results, and an original scientific contribution.
What the study demonstrates unambiguously: Ace beats elite-level players under regulated match conditions. The tests were conducted against four high-level players, including Yamato Kawamata, in December 2025. The robot won 3 of these matches.
What the study does not demonstrate: systematic dominance. As Implicator's analysis points out, there is an important nuance between the initial claims — "beats elite-level players in peer-reviewed tests" — and Sony's subsequent statements claiming that newer versions beat higher-ranked professionals. Science demands precision on these points.
The Associated Press reminds us that robot table tennis competitions have existed since 1983, initiated by John Billingsley. Forty-three years later, Ace is the first system to cross the threshold that justifies a publication in Nature. This is not insignificant.
Why table tennis is such a demanding benchmark
A table tennis ball weighs 2.7 grams. It can travel at over 100 km/h after a serve with intense spin. The human player has only fractions of a second to decide on their shot, according to LA AIMPA's technical analysis.
Unlike chess or even Go, table tennis imposes insurmountable physical constraints on a system that does not master latency. An LLM like GPT-5.5 can reason about a chess position in a few seconds — largely sufficient. In ping-pong, 200 ms of latency means the ball has already crossed the table twice.
It is precisely this tension between computational speed and the speed of the physical world that makes table tennis a historic benchmark for robotics.
Ace's architecture: sensors, arm, and deep reinforcement learning
Ace is not an arm driven by an LLM. It is a hybrid system where each layer is optimized for a specific aspect of the problem.
The 8-joint robotic arm
The hardware is an 8-degree-of-freedom arm, custom-designed for the project. Eight joints is more than a human arm at the shoulder and elbow level — which allows for racket trajectories that no human player could reproduce.
The Outpost AI documents a reaction time of 20 ms for the complete system. To contextualize: the visual reaction time of a trained human is around 150-200 ms. Ace reacts ten times faster.
High-speed perception
Sony has a major competitive advantage here: the company manufactures its own image sensors. The sensors used by Ace are not standard industrial cameras — they are high-speed sensors developed in-house, capable of tracking the ball's trajectory frame by frame at frequencies well above standard video.
This perception is not used to "see" the ball in the human sense. It feeds a trajectory prediction model that estimates where the ball will arrive, with what spin, and at what exact moment. The arm does not follow the ball — it moves to the calculated interception point.
Deep reinforcement learning in simulation
This is the core of the approach, and it is the same philosophy as GT Sophy, the Gran Turismo AI developed by Sony AI.
The principle: the agent is trained in an ultra-precise physics simulator. Millions of virtual rallies, with variations in spin, speed, and position. The agent learns through reinforcement learning — it receives a reward when it returns the ball onto the table, a penalty when it misses.
The challenge is the transfer to the real world (sim-to-real). The simulation is never perfect: friction, micro-deformations of the ball, and arm vibrations do not exist identically. Sony has not published the full details of its transfer method, but WebProNews confirms that the training combined simulation and progressive adaptation in a real environment.
This approach resembles what we see in autonomous AI agents that first learn in virtual environments before acting in the real world — but with infinitely tighter latency constraints.
Measured results: 3 victories out of 4, but with nuances
The published figures are as follows: Ace faced four elite-level players and beat three of them. Among them was a former Olympic participant, which gives particular weight to the result.
BigGo specifies that this is the first time in history that a robot has reached this level of performance in table tennis. Asharq Al-Awsat specifically documents the match against Yamato Kawamata in December 2025, with a photo via Reuters.
Elite vs. professional: the distinction that matters
Where media coverage sometimes lacked rigor is in the distinction between "elite level" and "top-ranked professional." The tests published in Nature concern elite players — an extremely high level, but not that of the world's top 10.
Sony subsequently stated that later versions of Ace beat higher-ranked professionals, but these results were not subject to the same peer-reviewed validation. Implicator analyzes this difference in claims accurately: science and corporate communication do not operate on the same criteria of proof.
Nor does the robot systematically dominate. "Occasionally beats elite-level players" is BigGo's phrasing — which means it also loses. The performance is not overwhelming domination but real, measured competitiveness.
The GT Sophy → Ace lineage: same philosophy, different world
Sony AI is not starting from scratch. The Ace project is explicitly part of the continuity of GT Sophy, the agent that learned to drive virtual race cars in Gran Turismo at the level of the best human players.
The official Ace project page on the Sony AI website clearly establishes this link. GT Sophy had demonstrated that deep reinforcement learning in simulation could produce competitive behaviors in a complex physical domain. But Gran Turismo remains a video game — the physics are simulated, perfect, deterministic.
Ace is the scaling up to the real world. Same learning architecture, same reward principles, but with all the uncertainty of the physical world: the ball does not bounce exactly the same way every time, the ambient air varies slightly, the arm has mechanical tolerances.
This lineage is important for understanding Sony's strategy. The company is not building a ping-pong robot — it is building a technology stack for physical AI, and table tennis is its public demonstrator.
What this means for physical AI
The term "physical AI" is used by WebProNews to describe Ace's breakthrough. The term refers to AI systems that interact directly with the physical world, as opposed to language models that operate in a purely symbolic space.
Beyond sports: manufacturing, rehabilitation, interaction
The immediate applications are not entertainment. WebProNews identifies three concrete areas:
Manufacturing. A arm capable of reacting in 20 ms with millimetric precision to fast-moving objects has direct implications for production lines. Catching a part on a conveyor belt, orienting it, placing it — these tasks require exactly the type of perception-prediction-action that Ace demonstrates.
Rehabilitation. A robot that can adapt its movement in real time to the unpredictable trajectory of an object (or a human limb) opens up perspectives for motor rehabilitation. The patient hits the ball, the robot adapts to their level and their asymmetric movements.
Human-robot interaction. This is perhaps the most profound implication. Until now, collaborative robots (cobots) operated in predictable environments or at reduced speeds for safety reasons. Ace demonstrates that a robot can physically interact with a human at full speed without injuring them — because it understands and predicts the dynamics of the exchange.
The link with autonomous agents
What Ace does in the physical world, models like GPT-5.5 or Claude Opus 4.7 do in the digital world: perceive an environment, plan an action, execute with real-time adaptation. The difference in allowed latency is simply several orders of magnitude.
An agent like Dexter, which conducts deep financial research, can take 30 seconds to analyze a document. Ace has 20 milliseconds. But the fundamental paradigm is the same: perception-decision-action loop.
The historical context: 1983, Billingsley, and forty years of failures
The Associated Press recalls a crucial detail: John Billingsley organized the first robot table tennis competitions in 1983. The idea was already to use this sport as a benchmark for robotics.
Forty-three years later, no robot had managed to beat a professional-level player in a reproducible and scientifically documented manner. Previous attempts reached an average amateur level — enough to impress in a YouTube video, but not to publish in Nature.
This delay illustrates the specific difficulty of the problem. In 43 years, computing has gone from mainframes to smartphones, Go was solved by AlphaGo, LLMs learned to write code. But making a mechanical arm react to a 2.7-gram ball with spin at 100 km/h? That resisted all these advances.
The reason lies in what roboticists call the sim-to-real gap. You can simulate a table tennis match with perfect physics — the problem is not the simulation. The problem is that the real robot is not the simulated robot. And every millimeter of discrepancy accumulates with every exchange until the system diverges completely.
Ace has bridged this gap. That is the scientific contribution.
The limits that media coverage forgets
Everything is not perfect in this result, and rigorous tech journalism must explain them.
A controlled environment
The matches take place under standardized conditions: controlled lighting, uniform background, ITTF-standard table according to WebProNews. This is not a random environment. Moving Ace into a noisy gym with variable lighting would very likely alter its performance.
One player, not a complete match
Ace plays singles, against a human. It does not handle the variables of a tournament: fatigue, psychological pressure, strategic adaptation over multiple matches. Comparing Ace to a complete professional player would be misleading — it is competitive on shot execution, not on the entire sport.
A demonstrator, not a product
Ace is a research project. Sony is not selling a ping-pong robot. The hardware is custom, the sensors are proprietary, the training required considerable computing resources. The distance between this demonstrator and a consumer robot is immense — comparable to the distance between a Formula 1 prototype and a production car.
❌ Common errors
Error 1: Confusing "elite level" and "best player in the world"
What's wrong: Several articles headlined "Ace beats the world's best players." This is false. The tests involve elite-level players, not the global top. The distinction is scientifically significant. The solution: Reread the primary sources — the Nature study and the official Sony AI page — and use exactly their terminology.
Error 2: Presenting Ace as a roboticized LLM
What's wrong: Ace does not use a language model to decide its shots. It is a perception-prediction-action system based on deep reinforcement learning, not an LLM that "thinks" about the match. The solution: Clearly distinguish RL architectures from language architectures. Ace is closer to AlphaGo than to ChatGPT.
Error 3: Ignoring the sim-to-real nuance
What's wrong: Presenting the result as "we trained in simulation and it worked on the first try." The sim-to-real transfer is the central problem the study solves, and it likely required massive iterations. The solution: Always mention that performance in simulation does not imply performance in reality — that is precisely the contribution of the publication.
❓ Frequently asked questions
Can Ace beat the world table tennis champion?
No. The published tests involve elite-level players, not the world's top 10. Sony stated that later versions beat higher-ranked professionals, but these results have not been validated by peer review.
What AI model does Ace use?
Ace does not use an LLM. It relies on deep reinforcement learning trained in simulation, with a perception pipeline based on Sony high-speed sensors. The approach is comparable to GT Sophy, not GPT-5.
Why table tennis and not another sport?
Table tennis combines high speeds (100+ km/h), complex trajectories (spin), a small space, and minimal equipment. It is a compact benchmark that simultaneously tests perception, prediction, and execution — something the 1983 Billingsley competitions had already understood.
Can you buy an Ace robot?
No. Ace is a Sony AI research project, not commercialized. The arm, sensors, and software are internally developed prototypes.
✅ Conclusion
Ace does not beat the world champion, and that is not where its importance lies. It demonstrates for the first time in a Nature-rank journal that a sim-to-real system can compete at the elite level in a very high-speed sport — crossing a threshold that robotics had been trying to reach since 1983. The underlying tech stack (high-speed perception + simulation RL + real-world transfer) is exactly what consumer robotics will need in the next ten years. Ping-pong was just the demonstrator.