Appearance-Based Navigation, Localization, Mapping, and Map Merging for Heterogeneous Teams of Robots

Spatial awareness is a vital component for most autonomous robots operating in unstructured environments. Appearance-based maps are emerging as an important class of spatial representations for robots. Requiring only a camera instead of an expensive sensor like a laser range finder, appearance-based maps provide a suitable world model to human perception and offer a natural way to exchange information between robots and humans. In this dissertation, we embrace this representation and present a framework that provides navigation, localization, mapping, and map merging capabilities to heterogeneous multi-robot systems using exclusively monocular vision. Our first contribution is integrating different ideas from separately proposed solutions into a robust appearance-based localization and mapping framework that does not suffer from the individual issues tied to the original proposed methods. Next, we introduce a novel visual navigation algorithm that steers a robot between two images through the shortest possible path. Thanks to its invariance to changes in the tilt angle and the elevation of the cameras, the images collected by another robot with a totally different morphology and camera placement can be used for navigation. Furthermore, we tackle the problem of merging together two or more appearance-based maps independently built by robots operating in the same environment, and propose an anytime algorithm aiming to quickly identify the more advantageous parts to merge. Noting the lack of any evaluation criteria for appearance-based maps, we introduce our task specific quality metric that measures the utility of a map with respect to three major robotic tasks: localization, mapping, and navigation. Additionally, in order to measure the quality of merged appearance-based maps, and the performance of the merging algorithm, we propose the use of algebraic connectivity, a concept which we borrowed from graph theory. Finally, we introduce a machine learning based WiFi localization technique which we later embrace as the core of our novel heterogeneous map merging algorithm. All algorithms introduced in this dissertation are validated on real robots