Humanoids Learning who are Teammates and who are Opponents

Abstract

RoboCup aims to progressively advance the re- search challenges in robotics by presenting a soccer tournament played by robots. One of the main aspects that shall fade is the need for pre-defined color coding; specially on what humanoids wear. However, most colors have rarely been modified over the years of the competition. The field remains green with white lines on it, the colors of the teams are predefined and the ball continues to have no pattern and is only orange. If colors of objects are no longer predetermined, a new parameter is needed for object recognition, and we propose to use shapes. However, computer-vision techniques for shape recognition are much more CPU-intensive than color recognition and perhaps they are unaffordable during the game. Our proposal here consists of identifying objects by their shape and extracting the colors from within these shapes. We identify the objects on the basis of their shape and learn autonomously the colors of the recognized objects. The chosen shape recognition algorithm is the Histogram of Oriented Gradients. This method has been proved to be capable to recognize complex objects and is widely used for pedestrian recognition. The complex objects we recognize are other humanoid robots (in particular Naos), similar in complexity for their anthropomorphic shape but less variable than human beings. This enables full learning of the environment colors in less than 1 minute on board of a Nao.