On Sensor Evolution in Robotics

Abstract

In recent years, evolutionary algorithms (EAs) have been successfully used in the design of artificial neural networks for a variety of applications. The suitability of EAs for this design task stems from their ability to adaptively search large spaces in near-optimal ways. One direct application of this advance has been in the area of evolutionary robotics, where EAs are typically used for designing behavior controllers for robots and autonomous agents. While such designs have been found to work well in general, their performance is often limited by the number, placement, quality, efficacy, and reliability of the sensors that the robots are endowed with. In this paper we argue that designing the sensory systems of these robots, in addition to the usual practice of designing the controller, can lead to improvements in the performance of the robot. Our results indicate that the evolution of sensors is a useful enterprise, and can lead to efficient and often counterintuitive controller d...