Risk-Aware Reinforcement Learning for Cooperative Autonomous Vehicle Coordination with Adaptive Risk Sensitivity and Multi-Agent Optimization

Ensuring safe and efficient coordination of autonomous vehicles (AVs) in intelligent transportation systems is particularly difficult under dense, uncertain, and rapidly changing traffic conditions. Many existing reinforcement learning (RL) methods show good performance in simplified environments but fail to fully account for heterogeneous risk exposure and non-stationary, multi-agent interactions. To address this gap, this paper introduces a Risk-Aware Reinforcement Learning (RARL) framework that couples adaptive risk sensitivity with Bayesian risk estimation in a cooperative multi-agent setting. Within RARL, reward signals are dynamically reshaped using real-time probabilistic risk measures, allowing AV agents to jointly balance safety and traffic efficiency across signalised intersections, multi-lane highways, and roundabout scenarios. The proposed approach is evaluated using SUMO, CARLA, NGSIM, and INTERACTION benchmarks. Compared with strong multi-agent RL baselines such as Bi-AC, MACPO, and MAPPO-L, RARL achieves up to 30% fewer collisions, about 25% higher throughput, roughly 30% improvement in scenario-recognition accuracy, and around 20% faster training convergence. These empirical results show that explicit and adaptive risk modelling significantly enhances policy robustness, scalability, and cooperative behaviour in heterogeneous traffic. By tightly integrating risk-aware decision making with multi-agent coordination, RARL provides a scalable and practically deployable paradigm for next-generation autonomous driving, improving safety, reliability, and real-time adaptability.