Sony AI builds the first robot to reach expert level in a sport

The harder problem here isn't the robot's win rate against human opponents — it's the latency and sensorimotor loop required to return a spinning ball in under 200 milliseconds, which is a fundamentally different constraint than the turn-based or static-environment tasks most robot benchmarks use. That physical timing pressure is what makes this benchmark harder to dismiss than a lab demo.

MIT Technology Review's 'How robots learn' piece from April 17 traced exactly the gap this story is closing: decades of robotics research produced narrow industrial arms while humanoid and athletic ambitions stalled. Ace is a narrow system too, purpose-built for one sport, but it's the first to clear an expert threshold in a domain where the environment is adversarial and unpredictable in real time. That's a different category than the warehouse or kitchen tasks Physical Intelligence's π0.7 targets. The two efforts are converging on embodied AI from opposite directions: one chasing generality, the other pushing the ceiling on a single physical skill.

Watch whether Sony AI publishes the evaluation methodology and opponent ratings used to define 'expert level.' If the benchmark holds up to independent replication against ranked competitive players, it anchors a credible new bar for physical AI; if the opponent pool turns out to be narrow or self-selected, the claim deflates quickly.