A Practical Platform for On-Line Genetic Programming for Robotics

Abstract

There is growing interest in on-line evolution forautonomous robots. On-line learningis critical to achieve high levels of autonomy in the face of dynamic environments, tasks, and other variable elements encountered in real world environments. Although a number of successes have been achieved with on-line evolution, these successes are largely limited to fairly simple learning paradigms, e.g. training small neural networks of relatively few weights and in simulated environments. The shortage of more complex learning paradigms is largely due to the limitations of affordable robotic platforms, which tend to be woefully underpowered for such applications.

In this paper we introduce a simple robotics platform based on Commodity Off The Shelf (COTS) designprinciples that makes on-line genetic programming for robotics practical and affordable. We compare the relative strengths and weaknesses of a number of different build options. As a proof-of-concept we compare three variations of evolutionary learning models for a color-following problem on a robot based on one of the designs: a simple neural network learning framework of the type typically seen in current research, a more extensive learning model that could not be supported by traditional low-cost research robots, and a simple evolutionary algorithm, but using standard tree-based genetic programming representation, which is also beyond the scope of traditional low-cost research robots. Our results show that the more powerful evolutionary models enabled by more powerful robots significantly improves the on-line evolutionary performance and thus that there are practical benefits to the COTS based