← Back to feed

Sony AI’s Project Ace appeared on the cover of Nature on April 23, describing the first robot to beat elite and professional human table tennis players under International Table Tennis Federation rules, with licensed umpires overseeing matches. The result closes a gap that has been open since AI surpassed humans at chess, Go, and complex video games: physical sport, where perception, hardware, and millisecond decision-making all have to work simultaneously against an unpredictable opponent.

What the System Actually Does

Ace is built around a 12-sensor vision stack: nine active-pixel sensor cameras track the ball’s 3D position at 200Hz with millimetre accuracy and approximately 10ms latency; three event-based gaze control systems measure angular velocity and spin up to 700Hz, fast enough to capture motion invisible to human eyes. The full system end-to-end latency is 20.2 milliseconds. Elite human players react in roughly 230 milliseconds.

The control system is model-free reinforcement learning, trained entirely in simulation and transferred directly to a custom-built eight-degree-of-freedom arm manufactured with lightweight alloys. No handcrafted trajectory models. No pre-programmed shot libraries. The policy generalises to situations it was not explicitly trained on — including balls bouncing off the net — because it learned from returns, not rules.

Match Results

In the April 2025 evaluation underlying the Nature paper, Ace won 3 of 5 matches against elite players (athletes with more than ten years of competitive training) under full ITTF rules at Sony’s Olympic-sized court in Tokyo. Against two professional league players in the same evaluation, it lost.

That changed. Additional matches in December 2025 and March 2026 — conducted after the Nature manuscript was submitted — produced wins against professional players. Shot speeds and placement accuracy both improved between evaluations, reflecting continued policy gains under real competitive conditions.

The system consistently achieved over 75% return rate against spins up to 450 rad/s, including unusual shots and rare ball trajectories. Sony AI Director Peter Dürr: “This research has shown that an autonomous robot can, in fact, win at a competitive sport, matching or exceeding the reaction time and decision making of humans in a physical space.”

What It Does Not Do

Ace has not beaten world champions. The team is explicit about the gap. Hardware researcher Stone noted that the April 2025 version was not yet generating the spin and force that top human players use, though subsequent evaluations narrowed that. Some professional players are still better than the current system.

The competitive ceiling matters for context. AI beat the world’s best human Go players in 2016 and the world’s best chess players a decade before that. In physical sport, elite-level performance has now been demonstrated. World-champion-level performance has not. The distance between those two thresholds is where the next five years of physical AI research will be contested.

Why It Matters Beyond Ping-Pong

Table tennis is a useful benchmark for physical AI for the same reason it is difficult: high-velocity ball physics, continuous spin variation, real-time opponent adaptation, and a game clock measured in milliseconds. A system that solves those constraints in a sport generalises toward real-world tasks — robotic manipulation, assembly, and human-machine physical interaction — that require the same combination of perception speed, hardware control, and adaptive decision-making. Ace’s sim-to-real transfer with no manual tuning is as significant as the win count.

The full paper is published in Nature, April 23, 2026.