Evolving morphological robustness for collective robotics

Abstract

This study evaluates a Neuro-Evolution (NE) method for controller evolution in simulated robot teams, where the goal is to evaluate the morphological robustness of evolved controllers. Artificial Neural Network (ANN) controllers are evolved for a specific sensory configuration (morphology) and then evaluated on a set of different morphologies. The morphological robustness of evolved controllers is evaluated according to team task performance given a collective construction task of increasing complexity. The overall objective was to ascertain an appropriate method for evolving ANN controllers that are readily transferable to robot teams with varied morphologies. Such controller transfer is necessary if task specifications change and different sensory configurations are required, or if robots are damaged and some sensors become disabled. In both cases it is ideal if teams continue to exhibit consistent behavior and a similar task performance. Results indicate that an indirect (developmental) encoding NE method consistently evolves controllers that fully function when transferred to teams with varied morphologies. That is, where comparable or higher task performances were yielded compared to controllers evolved specifically for the varied morphology.