Risk‐Aware Federated Hierarchical Reinforcement Learning for cooperative CAV Safety in Intelligent Transportation System

ABSTRACT Autonomous vehicles still face serious safety challenges because sensor readings are uncertain, traffic participants act unpredictably and road conditions shift quickly. This combination can lead to delayed or unsafe collision avoidance decisions while traditional rule based and model based controllers often struggle to deliver reliable performance. To address this problem, this paper presents a five phase deep reinforcement learning framework for safe and cooperative autonomous collision avoidance using the Udacity Self Driving Car dataset. The framework combines uncertainty aware multimodal perception through a retentive vision transformer based fusion of camera LiDAR and radar data with a hierarchical deep reinforcement learning policy regularized by Lyapunov safety constraints. It also incorporates vehicle to everything communication and federated multi‐agent learning to support cooperative decision making without centralized sharing of raw data. To improve robustness and readiness for deployment, the training pipeline includes curriculum learning domain randomization explainable validation and compliance oriented auditing. Experimental results show that the framework reduces collision rate by 37%, improves reaction time by 24%, lowers travel time by 13.4% and increases mean episodic reward by 22.5% while maintaining a safety score of 0.98, a comfort index of 0.91 and an explainability confidence of 0.87. Comparing with the state‐of‐the‐art baselines and ablated alternatives evaluated, the proposed framework delivers consistently better safety efficiency and control stability under dynamic driving conditions in this study. These results have shown that the framework provides a practical and reproducible foundation for reliable real time and cooperative autonomous driving in future intelligent transportation systems.