Developmental Analysis in Evolutionary Robotics

Abstract

This paper presents a developmental analysis of robot controllers created using evolutionary robotics (ER) methods. ER uses artificial evolution to automatically design and synthesize intelligent robot controllers. An aggregate fitness function that injects relatively little a priori task knowledge into the evolving controllers was used. We analyze the course of development of robot controllers evolving to perform a competitive goal-locating task. To sample the course of evolution, controllers were taken from progressively more advanced generations, and were then tested in a novel environment. Developments and changes in the controllers' abilities and competencies were identified and correlated with overall controller fitness. As evolution progressed, it was found that robots evolved more complex high-level behaviors that were not explicitly selected for by the fitness function.