DMD3QN: Dual Model Combined Driving Based on Dueling Double Deep Q-Network

Fast computation, low latency, and high accuracy are required for autonomous vehicles (AVs) to reach the end in uncertain environments safely. Therefore, high dimensional input data from sensors must be processed efficiently to reduce the state space dimensionality and improve operational efficiency. Deep Reinforcement Learning (DRL) can better ensure an efficient exploration and learning process for each intelligence. Based on this, we propose a Dual-Model Dueling Double Deep Q-Network (DMD3QN) autonomous driving system, which uses D3QN to train braking and driving models separately to combine the driving tasks, enabling the vehicle to localize objects and make braking and driving actions accordingly more accurately. In addition, to reduce the state space, we utilize depth cameras and segmentation cameras for object detection, distance estimation, and image processing. We used the CARLA simulator to conduct 20 test episodes on four routes with uncertain environments, achieving success rates of 95%, 95%, 100%, and 90%, which demonstrates the superiority of our proposed system.