AUTHOR=Peng Xiting , Liang Jinyan , Zhang Xiaoyu , Yang Haibo , Lei Weimin 

TITLE=Adaptive enhancements of autonomous lane keeping via advanced PER-TD3 framework

JOURNAL=Frontiers in Artificial Intelligence

VOLUME=Volume 8 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2025.1688764

DOI=10.3389/frai.2025.1688764

ISSN=2624-8212

ABSTRACT=With the advancement of autonomous driving technology, efficient and safe lane-keeping has become one of the core issues in this field. Currently, Deep Reinforcement Learning (DRL) methods still face challenges such as low training efficiency, slow algorithm convergence, and a tendency to fall into local optima when addressing lane-keeping issues. To address these challenges, a Prioritized Experience Replay (PER) mechanism designed to adapt to the learning process of the Twin Delayed Deep Deterministic Policy Gradient (TD3) is proposed, referred to as PER-TD3, to enhance the learning efficiency and lane-keeping performance of the vehicle in this work. It adjusts the probability of a selected sample by utilizing the difference between the predicted Q value and the true Q value to assign priority to different samples. By prioritizing samples with higher errors, the algorithm can correct biases in decision-making more quickly, especially when the vehicle deviates from its lane. In addition, introducing a probabilistic sampling mechanism helps to enhance the diversity of samples, ensuring high-frequency playback of high-value experiences, and enabling vehicles to learn accurate and stable lane-keeping strategies in a shorter period. Validation experiments on the TORCS platform demonstrate that the proposed framework can effectively solve the problem of unbalanced training, which is common in DRL, enhances training sample quality, accelerates algorithm convergence, and ultimately improves driving performance while ensuring safety.