iPrism: Characterize and Mitigate Risk by Quantifying Change in Escape Routes

Abstract

This paper tackles the significant challenge of ensuring safety of autonomous vehicles (AVs, also called ego actors) in dynamic, real-world scenarios, where interaction with other actors can cause accidents. To address this challenge, we introduce a new risk metric -- Safety-Threat Indicator (STI). This metric measures the changes in available escape routes for the ego actor. To effectively minimize this risk, quantified by STI, and to avert accidents, we present a reinforcement learning (RL)-based Safety-hazard Mitigation Controller (SMC). The SMC is designed to learn the optimal mitigation policies and provide actions that help avoid accidents. Our evaluation of the RL-based SMC on more than 4000 NHTSA-based safety-critical scenarios shows that (i) SMC significantly enhances safety, reducing accident occurrences by 37% to 98% compared to a baseline Learning-by-Cheating (LBC) agent, and (ii) achieving up to 72.7% accident prevention when compared to the state-of-the-art safety hazard mitigation agents.