A humanoid robot has demonstrated its ability to play soccer in real time, scoring a goal against a human goalkeeper.

The robot uses a vision-driven controller based on reinforcement learning that combines visual perception with movement patterns to track the ball and perform multidirectional kicks while adapting to changing conditions.

Researchers developed the system to help the robot follow rapidly moving objects, anticipate movements and adjust its posture for accurate kicks. Unlike systems that separate perception and control, the new framework allows continuous adaptation and reactive behavior.

The control policy was trained in simulation to make the robot kick the ball toward a goal, with additional inputs used to refine its movement patterns.

The system was then applied to the Booster T1 humanoid robot. It adapted its gait, including shuffling and moving sideways, to reach the ball and play interactive matches against a human goalkeeper.

The robot was also able to deliver fast and powerful shots from different positions on the field.

The technology helped the team win the Adult-Size Humanoid League at RoboCup 2025 and the 2025 World Humanoid Robot Games.

“Beyond the specific domain of soccer, our results suggest an approach for advancing real-world humanoid capabilities through the integrated learning of perception and action, bringing robots closer to autonomous and adaptive embodied systems,” Yushi Wang and colleagues write.

Podcast: Play in new window | Download