Adaptive double soft actor-critic architecture for intelligent vehicle platooning

Handling ralely happened safety-critical events remains a central challenge in developing reliable, learning-driven automated driving technologies. Existing reinforcement learning (RL) models, especially those trained on normal driving data, often lack robustness when encountering uncovered scenarios, such as sudden full braking by the lead vehicle, due to the underrepresentation of these events in training data. This study addresses this gap by introducing a Double Soft Actor-Critic (Double SAC) framework tailored for car-following control in both routine and emergency situations, with distinct reward functions designed for each scenario. A unique stochastic training environment is proposed to incorporate simulated emergency braking scenarios, enhancing model resilience in diverse and challenging conditions. Simulation results demonstrate that the Double SAC algorithm outperforms traditional methods in key metrics such as safety, efficiency, and passenger comfort, especially in emergency braking situations. Additionally, with an increase in CAV penetration rates, the stability of platoon control is greatly improved, traffic oscillations are reduced, and the system exhibits enhanced cooperative control capabilities.