Autonomous Decision-Making in Roundabouts Using Deep Reinforcement Learning and Heuristics Rules

Haikuo Zhang, Weizhen Han, Haiyong Shi, Haiyang Feng, Jin Hu

Abstract: Roundabouts frequently present challenging traffic scenarios, often leading to accidents, making it crucial to design a safe and efficient automatic driving decision-making system for effective navigation. While deep reinforcement learning (DRL) has shown promise in this field, it faces challenges such as poor sample efficiency and safety concerns. To address these issues, this paper proposes a hybrid approach that combines DRL with heuristic rules. We introduce an action filter module that pre-screens the action space to select valid actions based on the strategy, thereby improving sample efficiency. Additionally, we propose a novel reward function that encourages vehicles to switch to the outer lane before exiting the roundabout. This function helps mitigate safety concerns and traffic congestion under varying vehicle densities, enhancing the robustness of the algorithm. The proposed method is tested in a two-lane roundabout scenario. Experimental results demonstrate that the action filter module significantly improves training speed and traffic safety, while the novel reward function enhances the adaptability and robustness of the model.