ESLAS - a robust layered learning framework

Abstract

With increasing capabilities today robots get more and more complex to program. Not only the low-level skills and different strategies for different subgoals have to be specified, which by itself is not a trivial task even for simple domains. Both, the skill set and the strategies, also have to be compatible with each other. This turns out to be a major hassle as they are designed and implemented under assumptions about the future environment and conditions the robot will be faced with, that usually do not hold in reality. The Evolving Societies of Learning Autonomous Systems architecture (ESLAS) is targeted to this problem. With minimal need for specification, it is able to learn skills and strategies independently in order to accomplish different goals, which the designer can specify by means of an intuitive motivation system. In addition, it is able to handle system and environ- mental changes by learning autonomously at the different levels of abstraction. It is achieving this 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 provides actions to the strategy module. We demonstrate the feasibility of simultaneously learning low- level skills and high-level strategies in a Capture-The-Flag scenario. Thereby, the robot drastically increases its overall autonomy.