Unguided robot navigation using continuous action space

Abstract

In this paper, we propose a new method for robot vision-based navigation. This method is distinct from other methods, in the sense that it works even when the actions are not pre-determined and when their space is continuous. An if-then rule is introduced in the form of (from-position ^ action → to-position) for path and action planning. These if-then rules are learned incrementally in an unsupervised manner. When planning the path, the sequence of path fragments is retrieved from these rules. An experiment is performed with a realistic robot simulator called Webots; the results show that our method succeed in reaching our goals. © 2010 Springer-Verlag Berlin Heidelberg.