Badminton's useful number is 42 of 46 hardware returns—not a simulated rally
A humanoid research platform coordinates footwork and striking, but motion capture and a restricted hitting area remain important limitations.
Desk report · Published · The Humanoid Race
The September 2026 revision of Humanoid Whole-Body Badminton via an Annealed Reinforcement Learning Curriculum describes a unified learned controller for a 21-degree-of-freedom humanoid. The research tests coordinated footwork and racket striking on physical hardware as well as in simulation.
For the machine-fed hardware evaluation, the researchers report 46 trials for each controller. The target-known controller returned 42 shuttles successfully, or 91.3%; four hit the racket frame. The prediction-free variant returned 33 successfully, or 71.7%, with nine frame hits and four misses. Those results compare controller variants on one platform, not robot manufacturers.
The target-known controller produced outgoing shuttle speeds up to 19.1 metres per second in hardware tests. The paper also reports a 21-hit rally between simulated robots. That simulation result is not evidence of a 21-hit hardware rally: the real-world human–robot demonstrations sustained only a few exchanges, and the authors identify longer rallies as unfinished work.
Both variants rely on an instrumented motion-capture arena. Removing an explicit prediction module does not remove external sensing. The narrow interception region and limited stroke repertoire matter when judging how close the work is to ordinary-court play. These constraints belong alongside the impressive return percentage, not in a separate claim about commercial readiness.