Title Feasibility Study Of Fully Autonomous Vehicles Using Decision-theoretic Control Permalink

This project investigated the feasibility of constructing an autonomous vehicle controller based on probabilistic inference and utility maximization. We believed, and still believe, that such methods are required for autonomous operation in unconstrained mixed human/automated traffic. The project has been successful in identifying the major technical issues to be resolved. Contrary to our expectations, several significant theoretical and algorithmic advances were required in order to create an inference system capable of handling vehicle monitoring in a real-time fashion. New methods were also developed for learning probabilistic models from data, and for learning control policies given reward/penalty feedback. Our experience in hand-coding controllers for high-level control suggests that it will be quite easy to attain a fairly high level of performance, equivalent to perhaps safe operation over a 40-mile trip. However, we require safe operation over, say, 100,000 miles or more. We believe that testing and refinement against an incident scenario library, combined with reinforcement learning techniques for controller optimization, will enable us to reach this goal; however, we do not yet have a mechanism for verifying that the goal has been reached, since each test would require about three months of simulator time. Thus, we believe on the basis of our experience in this project that the high level driving problem can be handled satisfactorily, although much remains to be done in terms of real-time operation.