Thirteen wins, twenty-nine matches: what robot table tennis actually measured
DeepMind's 2024 study offers a match-based result, but its robot is a specialist arm and its opponent mix is essential context.
Desk report · Published · The Humanoid Race
Google DeepMind's 7 August 2024 publication reports a table-tennis agent winning 13 of 29 matches against previously unseen human opponents. The authors report this as a 45% overall match-win rate, describing the performance as amateur human level.
The aggregate hides a large difference between player groups. The robot won every match against beginners and 55% against intermediate players, but lost every match against the most advanced players. That does not support a claim that it beats tournament players or that the same win rate applies to any group of opponents.
The hardware is an ABB IRB 1100 robotic arm on two linear gantries, not a walking humanoid. Cameras track the ball, and a motion-capture system tracks the human opponent's paddle. Its performance should not be treated as proof that a humanoid can reproduce those results while balancing and moving around a table.
For readers comparing technical progress, competitive matches and cooperative rallies answer different questions. This historical study is useful because it reports opponents and match outcomes. It is not presented here as the newest system, a world record or a current leaderboard.