Abstract
In recent years, reinforcement learning (RL) has emerged as a powerful platform for developing autonomous driving systems. Nevertheless, while RL policies perform well under ideal conditions, they frequently exhibit unstable behavior when subjected to noisy observations or imperfect actuation factors that significantly compromise safety in real‐world deployment. This study seeks to address these vulnerabilities by proposing distribution‐level regularization (DLR). Rather than altering the underlying policy architecture, DLR introduces a training constraint that encourages consistent action distributions across both clean and perturbed inputs, facilitating straightforward integration into existing frameworks. The utility of this approach was tested by incorporating DLR into an on‐policy proximal policy optimization (PPO) model across discrete‐ and continuous‐action spaces, and further extending it to an off‐policy soft actor‐critic (SAC) model under continuous control, evaluating its performance across diverse robustness scenarios, including stochastic noise and actuator faults, in highway and intersection environments. The results provide evidence that DLR effectively minimizes collision frequency and extends stable driving duration without sacrificing performance under nominal conditions. These findings suggest that robust autonomous behavior may be achieved more efficiently through output‐level constraints rather than through the development of increasingly complex architectures.