Evolving Obstacle Avoidance Behavior in a Robot Arm

Existing approaches for learning to control a robot arm rely on supervised methods where correct behavior is explicitly given. It is di cult to learn to avoid obstacles using such methods, however, because examples of obstacle avoidance behavior are hard to generate. This paper presents an alternative approach that evolves neural network controllers through genetic algorithms. No input/output examples are necessary, since neuroevolution learns from a single performance measurement over the entire task of grasping an object. The approach is tested in a simulation of the OSCAR-6 robot arm which receives both visual and sensory input. Neural networks evolved to e ectively avoid obstacles at various locations to reach random target locations.