Robustness of Robot Programs Generated by Genetic Programming

Abstract

The goal of this research is to study whether robot programs which are generated by means of genetic programming (GP) are robust. The simulated robot task is a box moving problem. In order to test the robustness of generated programs, we conduct two experiments. In the first experiment, the initial conditions of the robot are fluxed. This is to prevent the overgeneralization of the specific situations. In the second one, the robot's sensors and actuators are assumed to be noisy. As a result of experiments, we have observed that the robot behaves robustly on the both experiments. GP has generated programs which evaluate almost effective commands for each sensor state. We suppose that the robot behaves robustly due to the redundancy of the programs, and we discuss how this redundancy is realized in GP framework for our robot programs.