Running across the reality gap: Octopod locomotion evolved in a minimal simulation

Abstract

This paper describes experiments in which neural network control architectures were evolved in minimal simulation for an octopod robot. The robot is around 30cm long and has 4 infra red sensors that point ahead and to the side, various bumpers and whiskers, and ten ambient light sensors positioned strategically around the body. Each of the robot's eight legs is controlled by two servo motors, one for movement in the horizontal plane, and one for movement in the vertical plane, which means that the robots motors have a total of sixteen degrees of freedom. The aim of the experiments was to evolve neural network control architectures that would allow the robot to wander around its environment avoiding objects using its infra-red sensors and backing away from objects that it hits with its bumpers. This is a hard behaviour to evolve when one considers that in order to achieve any sort of coherent movement the controller has to control not just one or two motors in a coordinated fashion but sixteen. Moreover it is an extremely difficult set-up to simulate using traditional techniques since the physical outcome of sixteen motor movements is rarely predictable in all but the simplest cases. The evolution of this behaviour in a minimal simulation, with perfect transference to reality, therefore, provides essential evidence that complex motor behaviours can be evolved in simulations built according to the theory and methodology of minimal simulations.