Adaptive method runtime safety assurance for autonomous driving in dynamic scenarios

Abstract

An adaptive method runtime safety assurance for autonomous driving in dynamic scenarios was proposed to address the issue that existing safety control approaches lack the ability to adjust dynamically to actual driving environments, resulting in reduced traffic efficiency. A runtime safety assurance autonomous control model (RTA-AutoSafe) was presented for autonomous driving. In this model, a deep Q-network algorithm based on an adaptive dual-buffer prioritized experience replay mechanism was designed for the performance controller to enhance the adaptability of decision-making strategies in dynamic traffic environments and to optimize traffic efficiency. An adaptive responsibility-sensitive safety (ARSS) model was designed based on vehicle dynamics to improve the adaptability of safety assessments. Based on this ARSS model, a dynamic bidirectional switching logic and safety controller integrating real-time vehicle feedback and traffic adaptation was constructed to achieve real-time safety assurance and dynamic coordination between dual controllers. Simulation results show that, compared with other safety control methods, the proposed method reduces the limitation of safety redundancy control on traffic efficiency in dynamic environments and achieves dynamic compatibility between real-time safety response and efficient operational strategies.

Keywords

• autonomous driving
• deep reinforcement learning
• runtime assurance
• responsibility sensitive safety model
• prioritized experience replay