Safety-driven Path Selection Using Reinforcement-Learning in Autonomous Driving


 Accepted by RSAE 2025

Abstract—Autonomous driving systems rely heavily on high-quality data in the path selection process to ensure safety and time efficiency during driving. However, data reliability and dynamic situational changes pose significant challenges for path decision-making. This paper proposes a safety-driven path selection method based on peer-to-peer (P2P) reporting and reinforcement learning (RL) to improve data credibility in path selection. The main idea is as follows. First, key parameters from several traffic accident datasets are identified, and a machine learning algorithm is used to evaluate the relative importance of these parameters on accident risks. Vehicles generate reports when passing through each road segment. An RL model is designed to evaluate and validate the road segment reports within a specific time window and dynamically update the trust level of the reporting vehicle. This trust computation enables the system to gradually identify and exclude malicious or faulty vehicles from contributing unreliable data. Finally, combined with the Kshortest path algorithm and lower-risk road segment selection, the vehicle can select the optimal path in real time to balance safety and driving efficiency. Experimental results show that the proposed method improves the safety and time efficiency in path selection, especially in low error rate environments, where the RL model can effectively distinguish between real and false events and ensure the reliability of decision-making.

Keywords—Autonomous driving, road conditions, data quality, reinforcement learning, P2P network.


Leave a Reply

Your email address will not be published. Required fields are marked *