Sensory anticipation for autonomous selection of robot landmarks

Abstract

There are many ways to define what constitutes a suitable landmark for mobile robot navigation, and automatically extracting landmarks from an environment as the robot travels is an open research problem. This paper describes an automatic landmark selection algorithm that chooses as landmarks any places where a trained sensory anticipation model makes poor predictions. The model is applied to a route navigation task, and the results are evaluated according to how well landmarks align between different runs on the same route. The quality of landmark matches is compared for several types of sensory anticipation models and also against a non-anticipatory landmark selector. We extend and correct the analysis presented in [6] and also present a more complete picture of the importance of sensory anticipation to the landmark selection process. Finally, we show that the system can navigate reliably in a goal-oriented route-following task, and we compare success rates using only metric distances with using a combination of odometric and landmark category information.