Risk Assessment of Autonomous Vehicles across Diverse Driving Contexts

Traffic crashes are a leading cause of death in the US, claiming more than 36,000 lives each year. Autonomous vehicles (AVs) promise a future without crashes. At the same time, permitting deployment of AVs without an adequate safety assessment might actually increase rather than reduce crashes. Extensive on-road testing is needed to ensure that AVs bring about the intended safety benefits. However, testing AVs across all possible driving contexts is impractical. Moreover, since crashes are rare events, AV safety assessment requires comprehensive approaches that account for diversity in driving contexts without testing in all possible scenarios. In this paper, we present a risk assessment framework that leverages human driving data, in addition to on-road testing data, to provide insights into the safety capabilities of AVs. Using our framework, we show that crash risk varies substantially across diverse maneuvers and environments. For instance, we estimate the crash probability for a left turn to be about 30 times that for a lane change in our region of study. Finally, we discuss specific use cases of our risk assessment framework for AV companies and regulators.