Object detection for robot coordination in robotics soccer

Abstract

In June 2018, iCog Labs held its second annual robosoccer competition which featured groups of humanoid robots playing soccer against each other. The authors were members of a team called upon to represent Nigeria with the University of Lagos at the competition which took place in Ethiopia. The work here presents a review of the approach taken to address the problem of automating robot coordination in real-world soccer applications. The design methodology relies on the Robot Operating System (ROS) as the platform upon which an asynchronous communication network between each robot and a central server is built. On the network, each robot is a node that consists of sub nodes for object detection and motion control. For object detection the work makes use of the you only look once (YOLO)v2 deep learning algorithm, and a simple decision-making algorithm for controlling vcv the robot based on the objects detected is devised. To quantify the object detection results, the common objects in context (COCO) evaluation metric is used. The results indicate an average recall and precision of 84% across different IOU. For qualitative results on the robot coordination in the ball’s direction, a reference to the open-source implementation of the work has been provided.