Cooperative and Conflict-Aware Alerting Using Graph-Based Reinforcement Learning in Rural Intersections for CAVs

Abstract

Connected and Automated Vehicles (CAVs) are reshaping mobility, yet rural intersections remain a persistent challenge where heterogeneous traffic, limited infrastructure, and frequent occlusions complicate safe navigation. Existing approaches using multi-agent reinforcement learning (MARL) and Graph Neural Networks (GNNs) show promise, but they often struggle with scalability, efficiency, and priority-aware conflict resolution. This paper presents a conflict-aware intersection alerting framework that integrates Relational Graph Convolutional Networks (RGCNs) with a Hybrid Proximal Policy Optimization (HPPO) to address these limitations. Vehicles are modelled as graph nodes enriched with spatial and kinematic features, while multi-relational edges capture proximity, geometric conflicts, and time-to-collision risks. The Proximal Policy Optimization (PPO) agent, guided by pre-trained critic values from extensive Car Learning to Act (CARLA) simulations, generates context-specific alerts that adapt to dynamic traffic states. A safety controller with PPO, complements reinforcement learning by enforcing collision avoidance, deadlock prevention, and prioritized zone release across vehicle types and with their arrival times. Evaluation across diverse unsignalized intersection geometries demonstrates measurable improvements in safety and traffic efficiency. The proposed framework reduced collision rates from 56.75% to 2% in 3-way intersections and from 37.5% to 22.62% in 4-way intersections, while maintaining queue reductions up to 1.28% and average speeds upto 4.92 m/s. By combining relational reasoning with reinforcement learning and safety-aware control, this framework advances intersection management for CAVs in rural contexts, mitigating risks associated with occlusion, heterogeneous traffic, and limited infrastructure, and contributing to safer, more efficient, and sustainable mobility systems.