Evaluating robustness in a two layer simulated robot architecture

Abstract

Many two layer robot architectures have been proposed and implemented. While justification for the design can be well argued, how does one know it is really a good idea? In this paper, one describes a two layer architecture (reinforcement learning in the bottom layer and POMDP planning at the top) for a simulated robot and summarize a set of three experiments in which one evaluated the design. To address the many difficulties of evaluating robot architectures, one advocates an experimental approach in which design criteria are elucidated and then form the basis for the evaluation experiments. In our case, one tests the implementation for its reliability and generalization (our design criteria) by comparing our architecture to one in which a key component is substituted; in these experiments, one demonstrates significant performance gains on the design criteria for our architecture.