Navigation and Planning in an Unknown Environment Using Vision and a Cognitive Map

Abstract

We present a framework for Simultaneous Localization and Map building of an unknown environment based on vision and dead-reckoning systems. An omnidirectional camera gives a panoramic image from which no a priori defined landmarks are extracted. The set of landmarks and their azimuth relative to the north given by a compass defines a particular location without any need of an external environment map. Transitions between two locations are explicitly coded. They are simultaneously used in two layers of our architecture. First to construct, during exploration (latent learning), a graph (our cognitive map) of the environment where the links are reinforced when the path is used. And second, to be associated, on an another layer, with the integrated movement used for going from one place to the other. During the planning phase, the activity of transition coding for the required goal in the cognitive map spreads along the arcs of this graph giving transitions (nodes) an higher value to the ones closer from this goal. We will show that, when planning to reach a goal in this environment is needed, the interactions of these two levels can lead to the selection of multiple transitions corresponding to the most activated ones according to the current place. Those proposed transitions are finally exploited by a dynamical system (neural field) merging these informations. Stable solution of this system gives a unique movement vector to apply. Experimental results underline the interest of such a soft competition of transition information over a strict one to get a more accurate generalization on the movement selection.