Increasing the Autonomy of Mobile Robots by On-line Learning Simultaneously at Different Levels of Abstraction

We present a framework that is able to handle system and environmental changes by learning autonomously at different levels of abstraction. It is able to do so in continuous and noisy environments by 1) an active strategy learning module that uses reinforcement learning and 2) a dynamically adapting skill module that proactively explores the robot's own action capabilities and thereby providing actions to the strategy module. We present results that show the feasibility of simultaneously learning low-level skills and high-level strategies in order to reach a goal while reacting to disturbances like hardware damages. Thereby, the robot drastically increases its overall autonomy.

1 Introduction

For a robot to be called a truly autonomous system it needs to provide means at all the different abstraction levels of behavior to autonomously monitor and control the according behaviour. That means that one central module in the system which provides alternative plans for reaction at unforeseen events no more suffices. Instead, all the different levels of behavior have to be autonomous on their own - and this leads to problems of coordination and control of the different learning processes. Typically, current research is thus sticking to one learning means for the whole task a robot has to learn. This usually has one of the following consequences: Either the low-level behavior to learn is only applicable to blocks-worlds domains with simple actions like LEFT or UP, or the convergence of the whole task is too slow to be used in real-world domains.