Representing and acquiring geographic knowledge (robotics)

This thesis presents a computational model of memory for spatial relations; how knowledge of geography may be represented, retrieved, and acquired. We focus particularly on the problems of representing incomplete and inexact information in a knowledge base and of learning a large scale area from a sequence of small scale views of the area. Our model, called MERCATOR, includes an original representation scheme for two-dimensional space; routines which retrieve information from the knowledge base in useful form; and a routine (the assimilator) which builds up a knowledge base incrementally from a sequence of scene descriptions. All these have been implemented in a running computer program. The representation approximates object boundaries in terms of sequences of straight edges, and it records object positions in terms of the dimensions of edges. Inexactness information is represented by recording bounds on edge dimensions and on the maximal distance from an edge to a boundary. Incomplete information is represented by recording only part of the boundary and interior of an object. This scheme is capable of representing many kinds of situations and of partial states of knowledge. The assimilator consists of two parts. The matcher finds correspondences between a scene description and the knowledge base. The merger adds the new information from the scene description into the knowledge base. The major result of this thesis is that a representation of space should integrate shape and positional information, but should maintain an abstract level of geometric description separate from object descriptions.