Limits and Potential of an Experience Metric Space in a Mobile Domestic Robot

This paper discusses the concept of an Interaction History Architecture, its use in training a simple task on a mobile domestic robot and some negative results which point to limitations of such an approach. We begin with a brief history its use which motivated the current research. It is based upon Shannon Information Theory and has previously been used in both humanoid and non-humanoid robots. These studies hinted at the ability of using the Interaction History Architecture to classify actions on a broader scale. The experiment outlined is an early test-bed for the use of this on a domestic robot as well as introducing negative rewards to the system. We then present the results from this and discuss and explain some of the difficulties and limitations that were uncovered in this type of approach.