Learning Localization Skills of Soccer Robot Using by Simulated Omni-Vision Camera

Abstract

In the RoboCup, each robot must perform localization, strategy selection, etc. within its limited resources. For this reason, the problem is that the programming cost is high for performing optimization according to the characteristics of the opponent, role, and environment so that resources can be utilized to the limit. Therefore, in order to reduce the cost of this tuning, we will consider a transfer learning method using a simulator. Since a soccer field can be regarded as a semi-structured environment, we study how to learn these features using CNN. Also, we think a method for generating test images used for CNN learning, in particular, a method for generating omnidirectional images using the simulator. We performed a simple computer experiment for localization using omnidirectional images and CNN, and then examined learning data suitable for tuning.