Design of Unsignalized Roundabouts Driving Policy of Autonomous Vehicles Using Deep Reinforcement Learning

Abstract

Driving at an unsignalized roundabout is a complex traffic scenario that requires both traffic safety and efficiency. At the unsignalized roundabout, the driving policy does not simply maintain a safe distance for all vehicles. Instead, it pays more attention to vehicles that potentially have conflicts with the ego-vehicle, while guessing the intentions of other obstacle vehicles. In this paper, a driving policy based on the Soft actor-critic (SAC) algorithm combined with interval prediction and self-attention mechanism is proposed to achieve safe driving of ego-vehicle at unsignalized roundabouts. The objective of this work is to simulate a roundabout scenario and train the proposed algorithm in a low-dimensional environment, and then test and validate the policy in the CARLA simulator to ensure safety while reducing costs. By using a self-attention network and interval prediction algorithms to enable ego-vehicle to focus on more temporal and spatial features, the risk of driving into and out of the roundabout is predicted, and safe and effective driving decisions are made. Simulation results show that our proposed driving policy can provide collision risk avoidance and improve vehicle driving safety, resulting in a 15% reduction in collisions. Finally, the trained model is transferred to the complete vehicle system of CARLA to validate the possibility of real-world deployment of the policy model.