Analysis of a neural network lateral controller for an autonomous road vehicle

Lateral control of a simulated vehicle in a simulated highway driving environment is explored. Three modules are used: a driving simulator, a visual preprocessor, and a neural network. The driving simulator, called RoadWay, is a three-dimensional computer graphics environment which supports interactive highway design and driving capabilities. The visual preprocessor, RoadVision, receives images from RoadWay, which represent forward-looking views from the cockpit of the simulated vehicle, and encodes these images using a family of oriented two-dimensional Gabor filters. Two Adaptive Resonance Theory neural network architectures, ART2 and ARTMAP, constituting the RoadBrain module, are employed to learn mappings between the visual encodings and emergent image categories, and then to associate these image categories with appropriate steering decisions. Once trained, the networks control the trajectory of the vehicle by accessing a steering decision for implementation by RoadWay at each timestep in response to a visual encoding of an image generated by RoadWay at the previous timestep. The paper presents the development of the three system modules, the creation of training sets, and computational results. Neural network performances are gauged by a number of procedures. Excellent results are achieved for straight roads and curved roads under a variety of initial conditions on the vehicle. Sponsored by the James S. McDonnell Foundation.