CRUISE-RL: A Risk-Belief Driven Multi-agent Reinforcement Learning Framework for Cooperative Autonomous Driving with Adaptive Safety–efficiency Trade-off

We present CRUISE-RL, a Risk-Belief Driven Multi-Agent Reinforcement Learning framework for cooperative autonomous driving.Safe coordination of multiple autonomous vehicles in dense, mixed traffic is difficult for conventional RL, which typically relies on fixed risk surrogates and static rewards.CRUISE-RL introduces an explicit Bayesian risk-belief state for each vehicle, updated online from noisy observations.This belief is fed into the policy and value networks and used to construct a risk-adaptive cooperative objective that continuously rebalances safety and efficiency across agents.In addition, CRUISE-RL employs risk-modulated exploration, suppressing exploratory actions under high estimated risk while allowing more aggressive learning when conditions are benign, and incorporates potential-based safety shaping linked to risk-belief to guide learning toward safer behaviours.We evaluate CRUISE-RL in coordinated driving scenarios derived from SUMO and CARLA.Compared with strong multi-agent baselines such as Bi-AC, MACPO and MAPPO-L, CRUISE-RL achieves up to 30% fewer collisions, 25% higher throughput, 30% better scenario recognition accuracy and 20% faster convergence across intersections, highways and roundabouts, with results reported as averages over multiple random seeds (mean ± standard deviation).These results show that embedding probabilistic risk-belief directly into the multi-agent learning loop markedly improves stability, scalability and real-time adaptability for next-generation autonomous driving systems.