Evolving sufficient robot controllers

Abstract

Different methods exist for reducing the time consumption in evolutionary robotics experiments. One is to use simulations, while another is to evolve controllers that are no more complex than task fulfilment requires. Behaviors such as exploration and homing, that seemingly demand a complex control system, only require a perceptron that connects a robot's sensors to its motors. This is shown by evolving such neurocontrollers for the Khepera robot. An exploitation of the robot's perception of the environment's geometrical shape allows the robot to encode time, even though explicitly it is not presented with the time and there are no recurrent connections in the neurocontroller.