Evolving Robust Robots Using Homeostatic Oscillators

A network of homeostatic relaxation oscillators is evolved to produce non-rhythmic phototactic behaviour in a simulated robot. Neural oscillations take place at a faster timescale than that of performance, and are designed to maintain an average activation value which is independent of sensory or synaptic input. In this way, neural activation cannot correlate directly with any actionrelevant sensory information, but must be continuously modulated by sensorimotor coupling. Evolution nds robust controllers which work by altering their central oscillation patterns. Robot are evolved with a xed set of body parameters, including sensor positions. Radical sensor robustness is shown by inverting the position of the sensors and also by removing either of them in turn { operations that do not alter the success of the strategy. Fast dynamics and long-term homeostasis seem to be required for robustness; slowing down the timescale of oscillations results in less robustness. The need for long-term homeostasis is shown both by modifying the oscillators and by running control experiments using a network of FitzHugh-Nagumo neurons. In none of these cases robustness is obtained. A general hypothesis is proposed according to which removing functional speci city from the lower-level mechanisms is likely to result in robust performance at the global level