Spatial learning mobile robots with a spatial semantic hierarchical model

The goal of this dissertation is to develop a spatial exploration and map-learning strategy for a mobile robot to use in unknown, large-scale environments. Traditional approaches aim at building purely metrically accurate maps. Because of sensorimotor errors, it is hard to construct accurately such maps. However, in spite of sensory and computation limitation, humans explore environments, build cognitive maps from exploration, and successfully path-plan, navigate, and place-find. Based on the study of human cognitive maps, we develop a spatial semantic hierarchical model to replace the global absolute coordinate frame used in traditional approaches. The semantic hierarchical model consists of three levels: control level, topological level, and geometrical level. The topological level provides the basic structure of the hierarchy. At the control level, a robot finds places or follows travel edges which can be described by qualitatively definable features. The distinctive features allow development of distinctiveness measures. The robot uses these measures to find, with negative feedback control, the distinctive places by hill-climbing search algorithms, and the travel edges by edge-following algorithms. Distinctive places and travel edges are connected to build a topological model. This model is created prior to the construction of a global geometrical map. Cumulative location error is essentially eliminated while traveling among distinctive places and travel edges by alternating between the hill-climbing search control algorithms and the edge-following control algorithms. On top of the topological model, metrical information is accumulated first locally and then globally. Using a simulation package with a robot instance, NX, we demonstrate the robustness of our method against sensorimotor errors. The control knowledge for distinctive places and travel edges, the topological matching process, and the metrical matching process with local geometry make our approach robust in the face of metrical errors. In addition to robust navigation at the control and topological levels, our framework can incorporate certain metrically-based methods and thus provide the best of both approaches.