Multi-Agent Cooperative Decision-Making for Connected and Automated Vehicles: A Survey of Methods, Scenarios, and Deployment Challenges

Connected and Automated Vehicles (CAVs) are expected to improve road safety and traffic efficiency, but achieving this remains challenging. This paper presents a methodologically organized survey of cooperative decision-making approaches for CAV systems, tracing their chronological development within each methodological family. The literature is categorized into rule-based and control-theoretic methods; game-theoretic models; single-agent and multi-agent reinforcement learning (MARL) approaches; and large language model (LLM)-, vision language model (VLM)-, multimodal large language model (MLLM)-, and foundation-model-based decision support. The reviewed studies address a range of traffic scenarios, including platooning, lane changes, merging, intersections, traffic signals, and mixed-traffic environments. The review compares these methods based on safety, scalability, communication requirements, interpretability, sim-to-real transfer, and mixed-traffic performance. The selected literature illustrates MARL-based approaches to cooperative CAV decision-making; its composition is not used to infer publication growth or the prevalence of methods across the field. Nevertheless, the reviewed MARL methods face challenges concerning safety, communication, scalability, interpretability, and transfer from simulation to real-world deployment. Overall, no single method addresses all the requirements of cooperative CAV systems. Hybrid architectures combining MARL with complementary methods are therefore a promising direction for future CAV systems.