PingpongDuo: two humanoids play a doubles table tennis match and string together more than ten shots in turn

The system coordinates the movements and strikes of two robots sharing the same space, achieving a 13.2 per cent higher success rate in simulation than the best existing method.

Playing doubles in table tennis imposes a constraint that singles ignores: the two partners must take turns hitting in a cramped space, quickly getting out of the way without disturbing the other. Researchers present PingpongDuo, a hierarchical framework that transposes this exercise to humanoid robots.

A team planner and individual skills

At the collective level, a planner jointly chooses the positions of the two robots and, thanks to a learned estimate of the risk of collision, rules out dangerous propositions, while preserving the target of the player who must strike as a priority. At the individual level, a footwork skill, inspired by movements recorded in human doubles players, is combined with a striking skill; a residual correction on the arm then refines the shot, trained in a simulator based on real measurements of the ball's trajectory and contact.

In simulation, the rate of successful returns increases by 13.2% compared to the state of the art, with a positioning error of about 2 cm during release. On real robots, a humanoid completes a 2.15 m release and return cycle in 2.4 seconds, with a peak of 2.23 m/s, and two machines hold more than ten consecutive strikes alternately.