Visually Augmented POMDP for Indoor Robot Navigation

This paper presents a new approach to robustly track a robot’s location in indoor environments using a partially observable Markov model. This model is constructed from topological representation of the environment, plus actuator an sensor characteristics. The system takes into account various sources of uncertainty to maintain a probability distribution over all possible locations of the robot. A novel feature of our approach is the integration of visual information to augment the robustness of the system. We show the first results of this approach in localizing an actual mobile robot navigating corridors. Key-Words: Robot Localization, Partially Observable Markov Decision Processes, Topological Maps, Natural Landmarks and Artificial Vision.