Evolving a nervous system of spiking neurons for a behaving robot

Abstract

We describe the artificial evolution of 'nervous systems' for the ARBIB robot, which control the way it interacts with its environment. The present work differs from earlier attempts to evolve robot controllers by use of a biologically-inspired heterogeneous network of spiking neurons. An obstacle-avoidance task is defined so as to provide an appropriate fitness function for the evolutionary process, and ten separate runs were undertaken in a simulated environment. Evolved nervous systems from 7 of the 10 runs led to an 'emergent' wall-following behaviour. Two of the seven, both having just 7 neurons, are examined and described. They are considerably simpler than our earlier, manually-designed solutions which had some 30-50 neurons, although the latter were additionally capable of ontogenetic (during lifetime) learning. One of the two example nervous system promotes photo-taxis as well as wall-following. These two evolved systems are tested on a real Khepera robot, and behave entirely as expected from the simulations.