Learning both a World Model and Navigation Paths in an Unknown Environment

Abstract

This paper presents a new method for mobile robot navigation in an un- known world. The parti-game learning approach is used to construct a world model, and to learn to navigate from a starting position to a known goal region. The navigation architecture then integrates a new approach, based on the application of the Fuzzy ART neural architec- ture, for on-line map building from actual sensor data. This leads to an improved world model, which is then used to introduce a predictive on-line trajectory ltering method resulting in a new and more eectiv e navigation approach. It is assumed that the robot knows its own current world location, is able to perform sensor-based obstacle detection (not avoidance), and straight-line motions. Simulation and real-robot results obtained with a Nomad 200 mobile robot will be presented demonstrating the eectiv eness of the discussed methods. Quanti- tative results will demonstrate (1) the exploration and planning improvements of the new navigation approach, and (2) that the constructed world model (original or improved) is gen- eral purpose in the sense that its usefulness is not restricted to be used on learning a particular path, but is valuable for learning paths with dieren t (Start,Goal) pairs.