Avoiding local optima with interactive evolutionary robotics

Abstract

The main bottleneck in evolutionary robotics has traditionally been the time required to evolve robot controllers. However with the continued acceleration in computational resources, the main bottleneck is now the time required for an investigator to create a robot simulator, a neural network, evolutionary algorithm and fitness function that together produce the desired behavior. Here we introduce a software framework that allows a user to conduct evolutionary robotics experiments without having to write any software themselves: the user defines the robot morphology, task environment and fitness function interactively; a neural network is constructed based on the robot's morphology; and an evolutionary algorithm optimizes desired behavior. We here show that this approach allows users to overcome one of the main limitations of evolutionary algorithms---recognizing and then preventing entrapment in local optima---in a continuous, code free manner.