A robotic arm swings. The ping-pong ball rockets back across the net at 16 meters per second, spinning at 600 radians per second. An elite player lunges. Misses.
Sony AI’s Project Ace just won another point. This isn’t a gimmick. It’s the first autonomous machine to beat expert human table tennis athletes under official International Table Tennis Federation rules, as detailed in a Nature paper published April 22, 2026. In matches on an Olympic-sized court in Tokyo, Ace took three out of five contests from five elite amateurs—players with over a decade of training and 20 hours weekly practice. Against two T-League professionals, Minami Ando and Kakeru Sone, it grabbed one game out of seven. Post-paper tweaks pushed it further: victories over pros in December 2025 and March 2026, per The Verge.
Peter Dürr, Sony AI director in Zurich and Ace’s lead, calls it unavoidable. “There’s no way to program a robot by hand to play table tennis. You have to learn how to play from experience,” he told AP News. Humans react in 230 milliseconds. Ace clocks 20.2. That edge comes from nine frame-based cameras triangulating the ball in 3D at 200 Hz, millimeter precision, 10-millisecond latency. Three event-based gaze systems—Sony’s IMX636 sensors—clock spin at 400-700 Hz. No markers on the ball. No altered court. Just raw perception fused by a convolutional neural network, error at 24.8 rad/s.
The arm? Custom eight degrees of freedom. Two prismatic joints for reach, six revolute for racket control—position, orientation, impact force. Lightweight Scalmalloy parts, topology-optimized. Actuators sync at 1 millisecond. Racket: Butterfly Dignics 05 rubber on VICTAS blade. Safety first: collision predictions trigger resets. Early tests had humans in helmets. Now? Bare table, refereed by Japan Table Tennis Association umpires.
Brains in the racket. Model-free reinforcement learning, trained entirely in simulation. Ace “played itself” for thousands of hours, mastering physics—drag, Magnus effect, bounces. Asymmetric actor-critic setup: policy sees noisy sensors, critic gets perfect sim data. Zero-shot transfer to reality. No fine-tuning. Rewards target landing zones, spin types. Policies stack: skill for strokes, tactics for placement, strategy for matches. Rally control queries at 31 Hz, spits 32-ms trajectories via convex optimization. Serves from a genetic-algorithm library, one-armed tosses.
Ace doesn’t overpower. It outlasts. Return rate above 75% on 450 rad/s spins, 14 m/s speeds—human elite territory. Rallies averaged 5 shots, longer than human 3.9. It scored 16 aces off elites, four off pros. Wins via consistency, not velocity spikes. Humans clinch with smashes; Ace grinds varied spins. One net-bounce response? 49 ms.
Development kicked off in 2020. Ball juggling first. Then cooperative rallies. Competitive by 2025. Each loss exposed flaws—like initial drag overestimation on fast shots. Stronger humans revealed it. Fix the sim. Retrain. Repeat. “We played stronger and stronger players and we beat stronger and stronger players,” Dürr said in The Guardian. Olympian Kinjiro Nakamura watched a backspin return. “No one else would have been able to do that. I didn’t think it was possible,” he remarked, per AP. But: “It means that there is a possibility that a human could do it too.”
Elite Rui Takenaka adapted. “If I used a serve with complex spin, Ace also returned the ball with complex spin… But when I used a simple serve… that made it easier for me to attack,” he told The Guardian. No eyes to read. No fatigue. Players miss the mind games.
Peter Stone, Sony AI chief scientist, sees beyond the table. “Once AI can operate at an expert human level under these conditions, it opens the door to an entirely new class of real-world applications,” he said in the initial CNET coverage. President Michael Spranger adds: “The robot cannot just win by hitting the ball faster… but by actually playing the game,” via AP. Demos on Sony AI’s blog and YouTube show rallies blurring to the eye.
Limitations persist. Drops above 16 m/s. Early hits limit deception. Pros still edge it overall. No onboard vision yet—external cams rule. But sim-to-real gaps shrink. Code pseudocode on GitHub: SonyResearch/ace_public.
Industry ripples. Table tennis has dogged robotics for decades—OMRON’s bots topped amateurs, never pros. Ace crosses that line. Reuters notes the feat in Tokyo. Bloomberg dubs it agentic AI triumph. Tech University Darmstadt’s Jan Peters praises: “Truly impressive,” though he flags broader manip tasks, per Guardian.
Manufacturing gains precision assembly. Rehab bots match patient pace. Training: endless rallies for athletes. Dual-arm extensions loom. Sony eyes physical agents in uncertainty—war, factories, homes. “Like the Apollo mission,” Stone says on Sony’s blog. Not commercial yet. Tech spillovers? Immediate. Watch the Project Ace film. Rally on.
WebProNews is an iEntry Publication