From Homing Behavior to Cognitive Mapping - Integration of Egocentric Pose Relations and Allocentric Landmark Information in a Graph Model

This thesis describes a behavior based approach to the problem of simultaneous localization and mapping ("SLAM"). The complex behavior of exploring an unknown environment is based on a combination of three local navigation strategies: obstacle avoidance, path integration, and scene based homing. :p: In this context the role of metric pose information is discussed. In the proposed system pose information is used to overcome several shortcomings of topological navigation, especially the problem of spatial aliasing. The spatial memory of the agent is modeled as a graph, which is embedded into the three dimensional pose space. In order to achieve global consistency a modified multidimensional scaling algorithm ("MDS") is used. The proposed system differs from recent robotic systems in several ways. First, pose estimates are derived only between known places, i.e. there is no explicit knowledge about the location of single landmarks. Second, all pose relations are derived from odometry. Third, globally consistent position estimates are calculated separately from globally consistent heading estimates.