USING EMERGENT MODULARITY TO DEVELOP CONTROL-SYSTEMS FOR MOBILE ROBOTS

A new way of building control systems, known as behavior-based robotic s, has recently been proposed to overcome the difficulties of the trad itional artificial intelligence approach to robotics. This new approac h is based on she idea of providing the robot with a range of simple b ehaviors and letting the environment determine which behavior should h ave control at any given time. We will present a set of experiments in which neural networks with different architectures have been trained so control a mobile robot designed to keep an arena clear by picking u p trash objects and releasing them outside the arena. Controller weigh ts are selected using a form of genetic algorithm and do not change du ring the lifetime (i.e., no learning occurs). We will compare, in simu lation and on a real robot, five different network architectures and w ill show that a network that allows for fine-grained modularity achiev es significantly better performance. By comparing the functionality of each network module and ifs interaction with a description of the sim ple behavior components, we will show that it is not possible to find simple correlations; rather module switching and interaction are corre lated with low-level sensorimotor mappings. This implies that the engi neering-oriented approach to behavior-based robotics might have seriou s limitations because it is difficult to know in advance the appropria te mappings between behavior components and sensorimotor activity for complex tasks.