Evolution of Robot Behavior and Its Robustness

Abstract

Evolution of an autonomous robot “Khepera” navigating in a course is studied, with a behavior model of automata and Genetic Algorithms. A set of infra-red ray sensor signals of the robot are inputed to an automaton which drives the two wheels. Any a priori maps or rules are not given to the robot. The transition table of the automaton is sliced into the chromosome and is evolved with Genetic Algorithms. It was demonstrated that a sensori-action model of this type can make the robot navigate successfully in a twisted course with obstacles and narrowness. The acquired behaviors, however, were found to be sensitive to the changes in starting conditions, where rugged fitness landscapes were observed. The fractal mechanism inherent to the wall-following navigation is proposed to explain the ruggedness. Averaging the ruggedness, the robustness was defined by the normalized average fitness. Several simulation experiments were carried out, where it was noticed that the robustness is lost after the major evolution is achieved. Several methods were tested, among which two methods; the fitness averaging over different starting conditions and the changing starting conditions during the evolution, were found to be effective in maintaining the robustness.