Active perception and map learning for robot navigation

This paper describes a simulated on-line mapping system for robot navigation. This system allows the autonomous creation of topological maps enhanced with metrical information provided by internal (odometry) and external (vision and sonars) sensors. Within such maps, the robot's position is represented and calculated probabilistically according to algorithms that are inspired by Hidden Markov Models. The visual system is very simple and does not allow reliable recognition of specific places but, used jointly with odometry, sonar recordings and an active perception system, it allows reliable localization even when the robot starts exploring its environment, and when it is passively translated from one place to another. Advantages and drawbacks of the current system are discussed, together with means to remediate the latter.