Investigations on the Robustness of an Evolved Learning Mechanism for a Robot Arm

This paper reports on a continuing research effort to evolve robot controllers with neural networks and to understand how Nature builds such systems. In an earlier paper [8] we showed that an artificial evolutionary system was able to evolve and grow a neural network, which learned to control a camera to track a light source and showed that the learning mechanisms could perform its task as well as in simulation. In this paper we continue transferring the evolved controller to the real-world and tested the robustness of the learning mechanism as follows: Once the system was able to track a red object with the cameras, we fixed the positions of the cameras and used the same principles to teach a robot arm to move a colored object from an initial location at the periphery of the visual field to the center of it, using visual (e.g., color and movement detection) and proprioceptive (e.g., position of the joints of the arm) feedback. The arm could solve the task not only for two-degrees of freedom, but also for three degrees of freedom. These results suggest that to overcome the "reality gap" between simulation and the real world one should evolve mechanisms enabling the real-world system to explore its own possibilities instead of specifying the system precisely in simulation