DRL-GAT-SA: Deep reinforcement learning for autonomous driving planning based on graph attention networks and simplex architecture

Abstract

Self-driving cars need to make decisions while sharing the road with human drivers whose behavior is uncertain. However, the presence of uncertainty leads to a trade-off between two conflicting goals: safety and efficiency. In this work, we propose DRL-GAT-SA to achieve safe and efficient autonomous driving, which is a new runtime assurance approach. First, we construct a graph attention reinforcement learning controller (GARL) based on the safety field model, which combines graph attention network with deep reinforcement learning. We learn the interaction between vehicles through the representation of dynamic relationships between vehicles and make decisions. Then DRL-GAT-SA provides safety assurance for the vehicle when the GARL produces unsafe control, while also retraining the GARL to improve its efficiency. DRL-GAT-SA can restore GARL control of the system under conditions that confirm safety without undue sacrifice of efficiency. In two different driving environments, DRL-GAT-SA consistently shows excellent performance.