Blurred vision: Simulation-reality transfer of a visually guided robot

Abstract

This paper investigates the evolution of robot controllers utilising only visual environment input data, capable of performing a hard task, playing football, in the real world. The techniques of minimal simulation, where the robot controller is forced to ignore certain features through making those features unreliable, are used to construct a noisy simulated environment for a robot with an onboard vision system. Robot control structures evolved in this simulation are then transferred on to a real robot, where the behaviours shown in simulation are displayed in the real world. In the experiment presented, finding a tennis ball and pushing it towards a goal, good controllers capable of performing the same behaviours in simulation and in the real world, are evolved only once sufficient unreliability is incorporated into the simulation. The success in evolving in simulation a robot controller incorporating only distal visual environment input data and displaying the same behaviours in both simulation and the real world, goes some way to addressing the argument that evolution is suitable only for toy problems.