Orientation system in Robots: Merging Allothetic and Idiothetic Estimations

For the last decade, we have developed a bioinspired control architecture for the autonomous navigation of mobile robots. The robot is able to learn to reproduce a homing or a route following behavior by interacting with a human teacher. However, the system strongly relies on the estimation of the orientation provided by a magnetic compass. We propose in this paper a model of visual compass in order to avoid the use of a magnetic compass. Each online learned visual landmark is associated with the shift between its position in the visual field and a direction of reference. The projection and the integration of these data on a one dimensional neural field allow to build a visual compass which accurately reconstructs the local reference in the neighborhood of the locations of learning. We also investigate how this visual compass can be used to calibrate an orientation system, which could be maintained by the odometry. Finally, the global system is validated in an experiment of route learning.