A Sensor-Fused Deep Reinforcement Learning Framework for Multi-Agent Decision-Making in Urban Driving Environments

Achieving robust and efficient autonomous driving in complex and dynamically changing urban traffic environments faces numerous significant challenges, especially the need to properly handle complex and time-varying interaction behaviors among multiple agents. This study innovatively proposes a sensor-integrated deep reinforcement learning framework (SIDRL), which organically combines multimodal sensor data fusion technology with multi-agent decision-making methods based on policy optimization. The system inputs include data from lidar, cameras and vehicle-to-everything (V2X), which are initially processed through a fusion perception module and subsequently fed into a decision-making network based on proximal policy optimization (PPO) for training and inference. Comprehensive evaluation experiments were conducted on the high-fidelity CARLA 0.9.15 simulation platform, and comparisons were performed with classical deep Q-network (DQN), asynchronous advantage actor-critic (A3C), as well as advanced methods such as soft actor-critic (SAC) and multi-agent proximal policy optimization (MAPPO). The experimental results clearly demonstrate that the proposed method enhances collision avoidance capability by 23.5% and decision-making efficiency by 17.2% under complex urban traffic scenarios. The research outcomes effectively confirm the critical role of multi-sensor fusion within deep reinforcement learning frameworks in improving environmental adaptability and safety for autonomous driving vehicles, providing a valuable new direction for the development of urban autonomous driving technology.