Evolvable Robot Hardware

Abstract

The theory and practice of evolutionary robotics is well established (Nolfi and Floreano, 2000). However, the overwhelming majority of work in evolutionary robotics has, to date, been concerned with evolving a robot’s control system. The robot’s control system is normally expressed entirely in software: it may be a program, or weights in an artificial neural network (ANN). Then the instantiation of a robot from its genome typically requires only that the control program coded by the genome is uploaded into the fixed robot hardware, ready for fitness evaluation. This is not to belittle the difficult challenges of evolutionary robotics, but to point out that conventional practice has, with a very small number of notable exceptions, avoided the altogether greater challenge of evolving a robot’s physical body.