Improving the robustness of evolved robot programs

Abstract

This dissertation describes an evolution of robot arm control programs for a visual-reaching task. The robot programs are evolved using Genetic Programming. The evolution process takes place in simulation and in the real world. Due to uncertainty in the real world, robot programs evolved in simulation are not working robustly when transferred to the real robot. This is caused by the difference between the simulation model and the real world. Two methods are proposed to improve the robustness of a robot program running on a real robot. First, multiple robot configurations are used during off-line evolution. Second, control programs are evolved on-line, that is the evolution procedure takes place in the real world on a real robot. On-line evolution does not require the model but it is very time consuming because of the mechanical speed of a physical robot. To accelerate on-line evolution, we propose a memoized function to store the effect of arm motions. Compared to a naive off-line evolution, the robustness of the program evolved off-line with multiple robot configurations is improved by 11%. The on-line evolution improves the robustness by 27%