Learning Locomotion by Co-Evolution of Morphological and Neural Parameters

Evolutionary Robotics applies biologically in- spired Evolutionary Algorithms to enable robots to optimize performance in tasks. Generally, experiments are conducted using fixed morphologies, requiring the algorithm to discover the best set of neural parameters for the given problem. Despite the simplicity of this approach, there is no guarantee that the imposed morphology is suitable, and sub-optimality or even failure may occur, particularly when evolving robots face randomly varying environmental circumstances. In this work, we investigate the co-evolution of robot morphological and neural parameters in a novel robot locomotion problem, where the environment is non-stationary. Specifically, we design two versions of the problem with increasing complexity and com- pare two conditions that differ in the parameters undergoing evolution. Performance analysis highlights the advantage of co- evolving morphological and neural parameters over using a fixed morphology in both versions of the problem. Furthermore, examining variations in morphological parameters reveals how the Evolutionary Algorithm discovers solutions in which certain values are preferentially increased or decreased depending on the properties of the problem.