Improving autonomous vehicle in-traffic safety using learning-based action governor

Abstract

The Action Governor (AG) is a supervisory scheme augmenting a nominal control system in order to enhance the system's safety and performance. It acts as an action filter, monitoring the action commands generated by the nominal control policy and adjusting the ones that might lead to undesirable system behavior. In this article, we present an approach based on learning to developing an AG for autonomous vehicle (AV) decision policies to improve their safety for operating in mixed-autonomy traffic (i.e., traffic involving both AVs and human-operated vehicles (HVs)). To achieve this, we demonstrate that it is possible to train the AG in a traffic simulator that is capable of representing in-traffic interactions among AVs and HVs. We illustrate the effectiveness of this learning-based AG approach to improving AV in-traffic safety through simulation case studies.