Ace the Ping-Pong Robot Can Whup Your Ass

Ping-pong is a deliberately punishing benchmark for robotics: the ball moves faster than most manipulation tasks, tolerates almost no latency between perception and actuation, and requires continuous adaptation rather than a single planned motion. The story doesn't surface whether Ace operates under controlled lighting and spin conditions or handles the full variance a human opponent introduces.

The MIT Technology Review piece from mid-April traced the long gap between humanoid ambitions and the narrow industrial systems roboticists actually shipped. Ace fits squarely in that gap: it is highly capable within a bounded physical domain but is not the general-purpose system Physical Intelligence was pitching when it announced pi0.7 around the same time, a model framed around performing tasks it was never explicitly taught. Ace and pi0.7 represent two distinct bets on how robotic competence gets built, one through deep specialization in a high-speed task, the other through broad transferability across tasks.

The meaningful test is whether Ace's perception-to-actuation pipeline holds up against opponents who vary spin and placement deliberately to exploit system latency. If the team publishes latency and error-rate numbers against ranked amateur players rather than controlled rally conditions, that would clarify how much of the performance is genuine adaptability versus a well-tuned narrow loop.