Detecting and diagnosing mistakes in inexact vision-based navigation

摘要

Navigation is the process by which a mobile entity finds its way through an environment to achieve a goal. Traditional approaches in mobile robotics assume that navigation can be accomplished using precise measurements of motion to maintain position estimates, complete, precise, and accurate a priori knowledge of the world, and complete, precise, and accurate perceptual information. However, natural environments are complex and dynamic, and cannot be known completely in advance. The consequences of physical actions cannot be predicted precisely or accurately. Perceptual information is incomplete, uncertain, inaccurate, imprecise, ambiguous, and contains categorical errors. As a result, navigational mistakes are inevitable. The problem addressed by this dissertation is how to detect and diagnose mistakes in navigation through large-scale space using vision. Mistakes are perceptual, cognitive, or motoric events that cause one to stray significantly from the intended route. Detecting a mistake involves realizing that one has done something wrong. Diagnosing the mistake consists of determining when the mistake occurred and what it was. The solution described in this thesis with mistakes explicitly by using specialized mechanisms and strategies for detecting and diagnosing mistakes. It uses detailed symbolic representations that incorporate both high-level and low-level visual information. It uses methods for comparing visual information at a symbolic level that make use of a priori knowledge. A simulation system called M scUCKLE implements this solution and tests its effectiveness. The thesis describes a variety of experiments run using M scUCKLE that demonstrate the ability of the approach to detect and diagnose diverse perceptual and motor mistakes. This approach for detecting and diagnosing mistakes delivers correct or nearly correct results most of the time. The research described in this thesis makes several contributions. It provides a thorough analysis of navigational mistakes and their causes. It develops a formalism for symbolic visual information and a scheme for comparing this information. It describes a solution to the problem of detecting and diagnosing navigational mistakes. This solution enables mobile robots to navigate more robustly with less precise, more ambiguous instructions.