ATHLETE teaches a humanoid tennis from generic movements and videos

The system combines reusable locomotion with gestures from tennis videos. Real trials report 90% of successful forehands and exchanges with humans.

ATHLETE avoids relying solely on expensive tennis-specific motion captures. It combines LAFAN1 locomotion data with video-extracted strikes, and then constructs body references tailored to ball trajectories.

A teacher-student phase and reinforcement refinement produce a reactive policy based on observations of the robot and the ball. On a real humanoid, the authors report 90% success in forehand and 86.67% in backhand, as well as exchanges of several shots. These rates describe the test conditions presented in the study.

Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.