A Computational Model of Driving for Autonomous Vehicles

Abstract

Driving models are needed by many researchers to improve traffic safety and to advance autonomous vehicle design. However, existing models of driving maneuver selection are generally too abstract and do not describe the computation needed to select actions after observing objects. In this paper we present a dynamic task analysis and use it to develop a computational model of driving in traffic. This model has been implemented in a driving program called Ulysses as part of our research program in robot vehicle development. Ulysses encodes legal, safe and practical driving rules as constraints on acceleration and lane selection. The application of constraints depends on particular objects in the world; thus, when constraints are evaluated, they show exactly where the driver needs to look at that moment. We explain the specific knowledge in Ulysses with illustrations from a series of driving scenarios of increasing complexity. We also briefly discuss the computer perception system that Ulysses needs. Finally, we describe how Ulysses drives a robot in a simulated environment provided by our new traffic simulator called PHAROS, which is similar in spirit to previous simulators (such as NETSIM) but far more detailed. Our new driving model is a key component for developing autonomous vehicles and intelligent driver aids that operate in traffic, and provides a new tool for traffic research in general.