Reliable safety decision-making for autonomous vehicles: a safety assurance reinforcement learning

Reinforcement learning (RL) can improve autonomous vehicles (AVs) by supporting adaptive decision-making, but it faces challenges in providing reliable safety across diverse scenarios. This research presents safety assurance RL (SARL), which embeds the safety assurance based on Responsibility-Sensitive Safety. The safety assurance includes three patterns to address different risks in the highway environment. The different scenarios are designed to compare the performance between the SARL and baseline algorithms (D3QN, RL-SA). Specifically, in scenarios with different traffic densities, SARL maintains zero collisions and outperforms D3QN and RL-SA in average velocity by 11.47% and 10.43%, respectively. In the emergency scenario, SARL maintains a low collision rate, while leading D3QN and RL-SA in average velocity by 19.55% and 21.42%, respectively. Furthermore, the analysis indicates that SARL can use lane changes to increase vehicle velocity. The SARL provides a design benchmark that helps integrate safety assurance into research employing RL for AV decision-making.