A humanoid can return a tennis ball. That is not the same as winning a match.
LATENT's Unitree G1 reports 90.90% forehand and 77.78% backhand return success under externally tracked research conditions.
Desk report · Published · The Humanoid Race
The LATENT research paper, submitted on 13 March 2026, describes a learned tennis controller deployed on a 29-degree-of-freedom Unitree G1. The authors train from fragments of human strokes and footwork rather than complete recorded tennis matches, then transfer the policy from simulation to physical hardware.
In the paper's real-world evaluation, a successful return means the ball lands within the opponent's court boundaries. Table 5 reports 90.90% forehand and 77.78% backhand success for the complete method. These are stroke-specific return rates, not the probability of beating a human opponent. The evaluation describes 20 human–robot rally matches; individual stroke sample counts are not separately stated.
The setup is instrumented. An optical motion-capture system tracks the robot and a ball covered in reflective tape, while a racket attached to the right wrist replaces the robot's hand. This demonstrates useful whole-body coordination, but not autonomous visual perception on an ordinary tennis court.
The authors explicitly distinguish the incoming-ball return task from a real two-player match. For companies comparing controllers, the useful questions are how tracking is supplied, what counts as a successful return and whether a result was measured on hardware. For investors and early adopters, the result is evidence of a particular athletic skill, not proof of a commercially ready tennis opponent.