Learning Complex Robot Behaviours by Evolutionary Computing with Task Decomposition

Abstract

Building robots can be a tough job because the designer has to predict the interactions between the robot and the environment as well as to deal with them. One solution to cope the difficulties in designing robots is to adopt learning methods. Evolution-based approaches are a special kind of machine learning method and during the last few years some researchers have shown the advantages of using this kind of approach to automate the design of robots. However, the tasks achieved so far are fairly simple. In this work, we analyse the difficulties of applying evolutionary approaches to learn complex behaviours for mobile robots. And, instead of evolving the controller as a whole, we propose to take the control architecture of a behavior-based system and to learn the separate behaviours and the arbitration by the use of an evolutionary approach. By using the technique of task decomposition, the job of defining fitness functions becomes more straightforward and the tasks become easier to achieve. To assess the performance of the developed approach, we have evolved a control system to achieve an application task of box-pushing as an example. Experimental results show the promise and efficiency of the presented approach.