Quantitative evaluation of the exploration strategies of an intelligent vehicle

摘要

Methods are described by which a simple mobile robot with a single rotating sonar sensor can systematically explore its environment and generate high-quality maps. Leonard and Durrant-Whyte's technique (1992) of grouping similar adjacent sonar returns is extended to sparse data. A new measure of map quality is defined, based on predicting the usefulness of the map for planning a benchmark set of routes. This quality metric is then used to evaluate the results of exploration experiments. Wall-following is tested as an exploration strategy. The experimental results show that it is a robust strategy but that it can be inefficient unless knowledge from the developing map is used in addition to the immediately-available sensory data. The quality (number of paths successfully identified) peaks at 84% after 1170 seconds. A minimum square error localisation scheme is introduced and added to the wall-following. The peak quality is then 89% after 755 seconds. 'Supervised Wall-Following' is then implemented and tested. A supervisory process monitors the developing map for exception conditions. When an exception arises the supervisor may change the parameters which control the wall-following (e.g. the step size) or to move directly to a new location (e.g. to eliminate repetitive loops). With the supervisor in place, the quality peaks at 91% after 480 seconds.