Reliably mapping a robot's environment using fast vision and local, but not global, metric data

摘要

A robot operating in an initially unknown environment must, in general, acquire and maintain information about that environment in order to perform tasks which involve repeated returns to particular locations, or reliable traversal of particular routes. Many practical applications of mobile robots in industrial and office environments require these capabilities. A basic prerequisite for this mapping functionality is the ability to reliably determine where the obstacles in the environment actually are. This thesis describes a system which provides reliable mapping functionality in a mobile robot, obtaining reliable information from cheap and readily available sensors without requiring great computational resources. The basic sensor used to detect obstacles in the environment is a camera. Using assumptions about the environment which are quite reliable in common office environments, the system is able to reliably detect and estimate the range to visible obstacles at very little computational expense. The system uses this data to build up a graph-structured representation of the environment. While the system uses local metric information, it does not assign a single set of global coordinates to any of the obstacles it detects, for two reasons. Firstly, assignment of global coordinates consistent with all observations may be impossible. For instance, if there are ramps in the environment, the measured distances between obstacles will not be consistent with any embedding into a plane. Secondly, global coordinates are not actually required for any task the robot must actually perform. However, purely local metric data is reliable enough to be of substantial use, and there are sometimes few alternatives to using it. Finally, the thesis describes parallels between the internal operations of the navigation system and what is believed to be a spatial memory system operating in the brains (specifically, the hippocampi) of rodents, and discusses how more a brain-like system might have certain advantages over the one described here. (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)