Evolving robot arm controllers for continued adaptation

Abstract

The practical objective was to develop a robust adaptive controller for a visually guided robot arm. The controller had to be able to adapt to drift errors or to the repositioning or replacement of equipment. A GA was employed to evolve artificial neural network controllers that continued to adapt throughout their life cycle. The genotype consisted of genes for encoding the network architecture, learning parameters, method of data creation, and the size of the robot arm movement. However, the GA did not evolve the weight values of the networks. These were adapted in the performance of the task using backpropagation. Evolution and testing of individuals was carried out both in simulation and in the real world with successful results. One of the best evolved controllers was tested on a real visually guided arm and learned to pick-up 98% of target objects.