Deep learning applied to humanoid soccer robotics: playing without using any color information

Nicolás Cruz, Francisco Leiva, Javier Ruiz-del-Solar

Abstract: The goal of this paper is to describe a vision system for humanoid robot soccer players that does not use any color information, and whose object detectors are based on the use of convolutional neural networks. The main features of this system are the following: (i) real-time operation in computationally constrained humanoid robots, and (ii) the ability to detect the ball, the pose of the robot players, as well as the goals, lines and other key field features robustly. The proposed vision system is validated in the RoboCup Standard Platform League, where humanoid NAO robots are used. Tests are carried out under realistic and highly demanding game conditions, where very high performance is obtained: a robot detection accuracy of 94.90%, a ball detection accuracy of 97.10%, and a correct determination of the robot orientation 99.88% of the times when the observed robot is static, and 95.52% when the robot is moving.