Learning to Evaluate Potential Safety Risk of Ego-Vehicle via Inverse Reinforcement Learning

The flourishing development of autonomous driving technology has brought forth a pivotal question, which is how to enhance the safety of autonomous vehicles. Ensuring vehicle safety requires comprehensive attention not only to the surrounding environment but also an in-depth exploration and monitoring of crucial factors intrinsic to the vehicle itself. One such key factor is the slip ratio, a determinant of the traction force acting on the vehicle, which importantly impacts the safety of the vehicle. The precise quantification of slip ratio’s impact on vehicle safety holds crucial significance for mitigating potential hazards and enhancing overall safety. This paper introduces a novel approach, which utilizes deep inverse reinforcement learning (DIRL) to extract the evaluation of slip ratio’s influence on vehicle safety from expert demonstration, by considering the single step action case to get rid of the dependence on vehicle model. The proposed method provides a general and direct safety assessment approach, imparting novel insights for practical applications in this field. The effectiveness of the proposed method is empirically verified by experimental results, which demonstrate the accuracy of the proposed method in assessing vehicle safety.