A neural vision based controller for a robot footballer

Abstract

Robot football is growing in popularity both as a research topic and as a sporting event. The football setting provides rich interaction possibilities and a ready source of competition in an environment containing both predictable and non-deterministic elements. Successful players must be able to react quickly in real time, exhibit multiple competences and choose between several possibly conflicting goals. Opportunities exist to explore reflexive behaviour, strategic behaviour and even communication and social behaviour in team events. At the same time, artificial neural networks are increasingly being used in robot controllers to explore new biologically-inspired ideas relating to perception, memory and motor control. The research described in this paper attempts to combine these two areas of study to produce a framework for a neurally based and visually guided football-playing controller. A controller architecture is proposed in which a small set of high-level features in the robot's environment are extracted from raw image data by using a feedforward neural network. These feature signals, collectively termed the “feature bus”, are then available for use by other controller modules. The feature bus signals are sufficiently general and high-level to be used with many different controller strategies, and their low dimensionality compared to the raw visual input makes the implementation of learning controllers more feasible.