Outplaying Elite Table Tennis Players with an Autonomous Robot

TL;DR

• First robot to beat professionally ranked players head-to-head under official rules. Earlier milestones were either not physical (GT Sophy in Gran Turismo) or not head-to-head adversarial (drone racing).
• Three contributions had to land together: a high-speed perception stack (9 cameras for ball XYZ plus 3 event-based cameras tracking spin), reinforcement learning for control, and custom hardware redesigned for speed.
• RL operates only at the lowest "skill" layer — given an incoming ball, how to swing to hit a target location and spin. Tactics (where to aim) and strategy (how to play the point or match) sit above it.
• The action space was the hard engineering problem: the reachable set for a 6-DoF arm on a 2-DoF base is deeply counterintuitive, so they map it onto a hypercube where every point is a feasible action, then hand execution to an MPC controller.
• The hardware was optimized by subtraction: strip every gram not needed for structural stability, 5 kg lighter than the original design — the only way to hit at 20 m/s every 0.8 seconds.
• Honest about the ceiling: ball speed now slightly exceeds professional humans, but spin does not; beating a world champion is an open question; doing this on a humanoid is "still very far."

Key Points

Framing: the research question behind two decades of work (~00:14–00:16)

Stone splits his time between UT Austin and Sony AI. In two weeks he becomes the head of UT Austin's new School of Computing (merging three departments); he founded Texas Robotics, and CoRL is hosted there this November. Side notes: the robotics facility lives in the old women's gymnasium, the Good Systems ethical-AI initiative, a public AI-literacy course (Essentials of AI for Life and Society), and a $10,000 online master's.

One question unifies his career: to what degree can autonomous intelligent agents learn in the presence of teammates and/or adversaries in real-time dynamic domains? That thread runs through RoboCup soccer (the 2050 goal of a humanoid team beating the World Cup champions), RoboCup@Home service robots, GT Sophy — the Nature cover four years ago, the first AI agent to beat the world's best humans at a real-time control task — and social navigation on Austin's hike-and-bike trails with Joydeep Biswas.

He also plugged Oopsie Data, announced days earlier: a community-sourced dataset of robot failures. Most robotics datasets only capture successes; this one collects real failed and suboptimal rollouts, and he asked the audience to contribute.

Why table tennis (~00:19–00:20)

Games have been AI's benchmark for decades, but the classic milestones are turn-taking (chess, Go, poker). GT Sophy was the first real-time control task where AI beat the best humans; the first real-world case was drone racing from Davide Scaramuzza's lab (also a Nature cover), but that was not head-to-head competitive.

Table tennis is both one-on-one adversarial and physically real-time, and has motivated robotics research since the mid-1980s. Stone showed a Google video — the next speaker, Vincent Vanhoucke, was involved — of a robot rallying well in a cooperative setting. But no robot had beaten expert players under Olympic rules.

The work was led by Peter Dürr at Sony AI ("he deserves the lion's share of the credit"), with a large team, and made the cover of Nature on April 23.

The match setup (~00:20–00:22)

Regulation player area. The robot serves its own ball — often skipped in prior systems — using a cup that tosses the ball up. The single concession to it being a robot: for safety, the human cannot cross the net. Otherwise, effectively Olympic rules. The opponent shown is a two-time Olympic silver medalist who peaked at world No. 5 and was ranked around No. 11 at the time of the match.

System: perception, RL, hardware (~00:21–00:26)

Perception. Nine cameras around the rig recover the ball's XYZ position; three event-based gaze-control cameras track spin. There is no artificial marker on the ball — spin is read off the printed logo. A wrist-mounted camera view, heavily slowed down, shows the tracking: "most robot videos you see sped up a lot; this one you have to slow down."

Control. The core problem is how to move the arm as the ball arrives, given a desired target and spin. Stone calls that a skill, distinct from tactics (where to place the ball) and strategy (how to play the point or the match); the RL lives almost entirely at the skill level.

Training runs mostly in a physics-informed simulator whose noise model, ball distribution model, and physics model are all built from real-world data. The learner is an asymmetric actor-critic: true state to the critic, noisy state to the actor, and the actor is what deploys.

The subtle piece is the action space. With six revolute joints and two base joints in the XY plane, the set of reachable end-effector poses is awkward; they found a mapping onto a hypercube in which every point is a feasible action ("feasible action for optimal control"). A model predictive controller then drives the end effector to the target the RL agent picks.

Hardware. The robot is custom-built, with an optimization pass stripping out all mass not needed for structural stability — 5 kg lighter than the original design — because professional-level play means hitting the ball at 20 m/s every 0.8 seconds.

Serves. An evolutionary process produced a library of diverse serves (backspin, topspin, varied placement) so the robot isn't predictable; it picks one to start the point, then the standard RL controller takes over for the rally.

A striking clip: the robot commits to a forehand, the ball clips the net and changes trajectory, and it reverses its swing mid-motion and still returns the ball.

Results and open questions (~00:26–00:28)

The Nature paper covers results through December 2025, when the robot was winning some and losing some against university-level players (red = losses, green = won at least one game, T = professional). A blog released about a month before the talk covers this spring: wins over a woman formerly ranked No. 5 in the world and a man formerly ranked No. 99 — genuine professionals.

By April 2026 the robot hits slightly faster than professional humans but still imparts less spin. Male professionals hit harder than the female professionals; the robot beat them too.

Summary: for the first time, AI agents reach expert-level performance in tasks requiring fast real-time decision-making in physical space. Three open questions: can it beat a world champion (unknown); can pros use the robot to improve their own game (yes — some professionals have already picked up shot ideas they wouldn't otherwise have tried); and could this be done on a humanoid ("we're still very far from that").

He closed by pointing to the 3:10 PM main-stage talk from Sony AI's Michael Spranger, covering Ace and GT Sophy.

Quotes

"The first robot that can beat a professional athlete at their own sport." (~00:14)

His own one-line framing of the result.

"The only concession that we make to this being a robot is that we don't allow the player to cross the net for safety reasons, but otherwise it's effectively regulation Olympic rules." (~00:22)

The credibility of the whole result rests on this line: the robot serves, on a regulation court, under regulation rules.

"For the first time, we've shown that AI agents can reach expert level performance in tasks that require fast real-time decision-making in physical space." (~00:28)

The claim is not "a robot plays table tennis" — it's expert-level real-time physical decision-making as a general capability.