Multi-dimensional reward function design for safe and efficient roundabout navigation in reinforcement learning-based autonomous driving

Navigating roundabouts is one of the most challenging decision-making problems in autonomous driving. Traditional reinforcement learning reward functions often struggle to guide agents in learning safe and efficient navigation policies. To address this challenge, this paper proposes a multi-dimensional reward function design centered on path completion. This method reinforces global navigation awareness through an aggressive path completion reward, which is combined with lane correctness guidance, a heading consistency reward, and layered safety penalties to effectively balance driving efficiency with safety. Comparative experiments in the MetaDrive simulator show that our proposed method increases the roundabout success rate to 73%, a substantial improvement over the baseline model's 30%. Concurrently, it significantly reduces the collision failure rate from 31% to 8%. More importantly, the trained agent exhibits an "observe-wait-decide" behavioral pattern akin to that of an expert human driver, offering an effective solution for autonomous decision-making in complex traffic scenarios.