A Practical Platform for On-Line Genetic Programming for Robotics

Abstract

There is growing interest in on-line evolution forautonomous robots. On-line learning is critical toachieve high levels of autonomy in the face of dynamicenvironments, tasks, and other variable elementsencountered in real world environments. Although anumber of successes have been achieved with on-lineevolution, these successes are largely limited tofairly simple learning paradigms, e.g. training smallneural networks of relatively few weights and insimulated environments. The shortage of more complexlearning paradigms is largely due to the limitations ofaffordable robotic platforms, which tend to be woefullyunderpowered for such applications.In this paper we introduce a simple robotics platformbased on Commodity Off The Shelf (COTS) designprinciples that makes on-line genetic programming forrobotics practical and affordable. We compare therelative strengths and weaknesses of a number ofdifferent build options. As a proof-of-concept wecompare three variations of evolutionary learningmodels for a colour-following problem on a robot basedon one of the designs: a simple neural network learningframework of the type typically seen in currentresearch, a more extensive learning model that couldnot be supported by traditional low-cost researchrobots, and a simple evolutionary algorithm, but usingstandard tree-based genetic programming representation,which is also beyond the scope of traditional low-costresearch robots. Our results show that the morepowerful evolutionary models enabled by more powerfulrobots significantly improves the on-line evolutionaryperformance and thus that there are practical benefitsto the COTS based