Enhancing autonomous driving safety in real lane-changing scenarios under friction variability: A friction-adaptive...

How autonomous vehicles (AVs) can learn safety-aware policies in real-world lane-changing scenarios under friction variability remains underexplored. To address this, a friction-adaptive shield reinforcement learning (FA-SRL) framework is proposed in this study. FA-SRL incorporates a friction-adaptive shield that actively intervenes to address the impact of variable friction on vehicle dynamics through a generalized time-to-collision (GTTC)-based risk assessment and graded risk rewards. Simulation scenarios based on real-world datasets with varying traffic densities are constructed to rigorously train and evaluate FA-SRL variants against their baselines (PPO, SAC, TD3) and naturalistic driving behavior. Training results demonstrate that FA-SRL significantly enhances final rewards and convergence speed, optimizing by 11.31 % and 34.76 %, respectively, compared to baselines. Evaluations indicate substantial safety improvements with FA-SRL variants; particularly, FA-SRL (SAC) achieves the optimal and most balanced performance with success rates of 100 % in low-density and 97.33 % in high-density traffic and statistical safety improvements compared to human driving under varying traffic densities. Specifically, the safe time-step ratio in low-density and high-density scenarios is improved by 1.28 % and 2.97 %, respectively. Moreover, the efficiency performance of learned policies significantly outperforms human driving behavior, with an average reduction of 13.63 % in task completion time. However, the conservative friction-adaptive shield in FA-SRL slightly compromises efficiency relative to baselines. These findings validate the effectiveness and robustness of the proposed FA-SRL under inevitable friction variability, highlight its potential for deployment in safety-critical autonomous driving requiring a careful balance between safety and efficiency, and further suggest its broader value as a principled framework for embedding physics-based safety constraints into policy learning.