Sony AI's Ace becomes the first robot to beat elite table tennis players under official rules, Nature paper shows
Sony AI's autonomous table tennis system Ace won three of five matches against elite players and lost both matches against two Japanese T.League professionals in April 2025 trials, according to a paper published in Nature on April 22, 2026. The robot combines event-based vision sensors from Sony Semiconductor Solutions, an eight-axis arm and reinforcement learning policies trained in simulation, and reaches a perception latency of 10.2 milliseconds.
Sony's research arm Sony AI has published in Nature what it describes as the first real-world autonomous system competitive with elite human table tennis players. The robot, called Ace, won three of five matches against elite players under official competition rules and lost both of its matches against professionals, according to the paper, which appeared on April 22, 2026 in volume 652 of the journal.
The numbers are specific. Against five elite players, three women and two men with more than ten years of intensive training and about 20 hours of practice a week, Ace won 7 of 13 games played. Against two professionals from Japan's T.League, Minami Ando and Kakeru Sone, it won 1 of 7 games. All matches took place in April 2025, and every opponent faced the robot for the first time, the authors write.
Serving was a strength. Using 15 different serve types, Ace scored 16 direct points from its serve against the elite group, which together scored 8. Against the professionals it used 13 serve types and scored 4 such points, against 7 by its opponents. Sony AI's project page, which covers matches against 12 players in April and December 2025, reports a first victory over a professional player, a result that is not part of the peer-reviewed paper.
The system has three main parts. For perception, nine synchronised conventional cameras triangulate the ball in three dimensions at 200 Hz, and three gaze-controlled units fitted with event-based vision sensors track the printed logo on the ball to estimate spin. The average latency for locating the ball is 10.2 milliseconds. Sony AI says the event sensors are IMX636 parts developed with the French company Prophesee, and the conventional cameras use Sony Pregius IMX273 sensors.
The arm is custom-built with eight degrees of freedom, two prismatic joints for lateral travel and six revolute joints, and its structure is topology-optimised in Scalmalloy. Actuators are synchronised at 1 millisecond, and the robot plays with a Butterfly Dignics 05 rubber, the paper says. It returns balls consistently at up to 14 meters per second, keeps a return rate above 75% for spin up to 450 radians per second, and reaches a maximum linear velocity of 16.4 meters per second.
Control comes from deep reinforcement learning. The team trained policies with the Soft Actor-Critic algorithm entirely in simulation, using an asymmetric design in which the policy sees noisy inputs while the critic receives ground truth. The simulator models a velocity-dependent Magnus effect, variable bounce on the table and a neural-network correction for ball and racket contact. Policies choose actions at 31.25 Hz, which are turned into collision-free trajectories executed at 1 kHz, and a serve library was generated with a genetic algorithm.
The paper lists more than 50 authors, with Peter Dürr as first author and Michael Spranger as last author, and gives Sony AI in Tokyo as the main affiliation. Sony AI names Sony Semiconductor Solutions as its sensor supplier and the table tennis brand Victas as athletic partner for recruiting players and supporting data collection.
Table tennis is a demanding test because the ball arrives near the limit of human reaction time and every shot is adversarial. Industrial robots from makers such as FANUC, Yaskawa Electric and ABB are fast, but they repeat known motions in fenced cells. Ace shows that a learned policy can react to unpredictable, fast physical input within tens of milliseconds, which is the capability that collaborative robots, logistics pickers and humanoids still lack.
The limits are stated in the results. Ace lost to both professionals in the paper's trials, the system is a fixed installation with twelve cameras around one table, and the authors give no figures for compute, training time or cost. Whether the approach transfers to tasks without a tightly modelled ball and table is not tested.
- Elite players
- 5
- 5
- 3
- 13
- 7
- 16
- 8
- Professional players (T.League)
- 2
- 2
- 0
- 7
- 1
- 4
- 7
All figures as disclosed in the Nature paper (DOI 10.1038/s41586-026-10338-5). Matches played equals players because each opponent played one match against Ace. The December 2025 results cited on Sony AI's project page are not included.
As of Oct 1, 2026
ROBOTNESS analysis
Ace matters less as a sports milestone than as proof that event-based vision combined with simulation-trained reinforcement learning can close the latency gap that keeps robots out of fast, contact-rich work.
The evidence is in the stack. A 10.2 millisecond perception loop, a policy that runs at 31.25 Hz on top of a 1 kHz trajectory layer, and returns at up to 14 meters per second against human opponents are all measured under competition rules, not in staged demonstrations. Sony owns the sensor, the learning research and the integration, which few robot companies can say.
The strongest counter-argument is that table tennis is a closed world. The ball's physics can be modelled precisely, the workspace is fixed and the cameras are placed around it. Factory and warehouse tasks involve deformable objects, clutter and occlusion, where the same simulation fidelity is much harder to reach.
Bull case. Event-based sensors become standard on robots that must react quickly, and Sony Semiconductor Solutions turns Ace into a reference design for that market. The training methods carry over to high-speed sorting, catching and human handover tasks.
Bear case. Ace remains a research showcase like earlier game-playing systems. Sensor customers keep buying conventional cameras because event-based vision requires new software, and Sony does not build a robotics business around the result.
Signals to watch:
- Whether Sony Semiconductor Solutions announces event-based sensor products or reference kits aimed at robot makers during its fiscal year ending March 2027.
- Any peer-reviewed report of the December 2025 matches, including the win over a professional player cited on Sony AI's project page.
- Follow-up papers from Sony AI that apply the same perception and control stack to manipulation tasks outside sport, for example at CoRL 2026 or ICRA 2027.
- 8 degrees of freedom (2 prismatic, 6 revolute), 1 ms actuator sync
- Consistent up to 14 m/s; above 75% return rate up to 450 rad/s spin
- 9 conventional cameras at 200 Hz plus 3 event-based vision units
- Soft Actor-Critic reinforcement learning trained in simulation
- Sony AI, Tokyo
- Nature, April 22, 2026, vol. 652, issue 8111, pages 886 to 891
- April 2025, under official competition rules
- Won 3 of 5 matches, 7 of 13 games
- 10.2 ms average
- Lost 2 of 2 matches, won 1 of 7 games (Minami Ando, Kakeru Sone, T.League)
Why it matters
Most robot learning results are judged on success rates in slow, staged tasks. Ace was judged on wins and losses against people who were trying to beat it, at speeds where a few milliseconds decide the outcome. That makes it one of the clearest public demonstrations that a learned controller can handle fast, adversarial physical interaction, a requirement for any robot that works next to people.
It also matters for Sony's positioning. The company combines image sensors, AI research and system integration in one group, and the paper documents how those pieces fit together in a working system. That is a capability story for Sony Semiconductor Solutions as much as for Sony AI.
Rival analysis
No other company has published peer-reviewed results of a robot winning matches against elite table tennis players under official rules, according to the paper's own claim of a first. The relevant competition is therefore in the enabling technologies. Prophesee, which co-developed the IMX636 sensor with Sony, is the best-known independent event-based vision company. In robot control, Google DeepMind, Physical Intelligence and Skild AI are building general models, but none has shown this level of reactive speed in public results.
Industrial robot makers such as FANUC, Yaskawa Electric and ABB dominate fast, repeatable motion but rely on programmed paths. The gap Ace illustrates is reaction to unpredictable input, which is where learning-based control and low-latency sensing intersect.
Valuation context
Sony does not disclose spending on Sony AI or on the Ace project, so the work cannot be tied to a valuation figure. For context on how investors price robot intelligence, Skild AI raised $1.4 billion at a $14 billion valuation in January 2026 and Figure AI raised $1 billion at $39 billion in September 2025, both according to their own announcements.
For Sony, the more direct link is the sensor business. A credible robotics use case for event-based vision supports demand for that product line.
Supply-chain implications
Ace relies on two Sony sensor families, the IMX636 event-based sensor developed with Prophesee and the Pregius IMX273 sensor. The arm is custom hardware with topology-optimised Scalmalloy parts and 1 millisecond synchronised actuators, and the racket rubber comes from Butterfly. Sony AI names Victas as its athletic partner.
The design points to a supply chain in which sensor makers move closer to robot integration. Event-based vision needs dedicated processing software, which gives the sensor owner a reason to supply full perception modules rather than bare chips.
Signals to watch
Sensor product announcements from Sony Semiconductor Solutions that target robot makers would show commercial intent. A second signal is a peer-reviewed account of the December 2025 matches cited on the project page. A third is whether Sony AI applies the same stack to manipulation, logistics or human handover tasks.
Useful dates include CoRL 2026 in the autumn, Sony's fiscal year results in May 2027 and ICRA 2027.
Analyst view
Thesis: Ace is the strongest public evidence so far that low-latency event-based perception plus simulation-trained reinforcement learning can make robots competitive in fast physical interaction with people. Confidence: medium.
The result is peer reviewed, measured in matches under official rules and documented with exact latency and speed figures, which supports the thesis. Confidence is not high because the environment is unusually well modelled, the installation is fixed and large, and the robot still lost to both professionals in the paper's trials.
Questions you should be asking
How much of Ace's performance depends on the precise physics of a table tennis ball, and what happens with objects that cannot be simulated as accurately? What compute and training time did the system need, given that the paper does not report them?
Will Sony commercialise any part of the stack, either as sensor modules or as a perception reference design? And will the December 2025 professional win be documented to the same standard as the April 2025 trials?