Enhanced deep reinforcement learning model with bird’s eye view design strategy for decision control in vehicle-road collaboration

Abstract

Autonomous driving in complex traffic scenarios is a vital challenge, and deep reinforcement learning (DRL) has been extensively applied to address this issue. The recent advancement of vehicle-to-everything (V2X) technology has provided abundant perceptual information for DRL agents, improving the accuracy and safety of decision control. However, existing research on green wave traffic scenes has difficulty adapting to multi-signal scenarios with single-signal countdown models, which lack complete signal state information. To address this limitation, an enhanced DRL model with bird’s eye view (BEV) design strategy is proposed for vehicle-road collaborative autonomous driving scenarios. The constructed model introduces a state prediction fusion strategy to compensate for state information. Specifically, state information is first predicted by fusing perception results from vehicles and roadside units (RSUs) at different moments. Then, the recommended velocity is derived for green wave passage, called the green wave velocity belt, and incorporate it into the state space as two variables in the state vector. Finally, a relevant reward term in the reward function is designed to guide agent learning strategies. The proposed method is trained on the basis of the parallel DreamerV3 framework. The results show that the proposed approach can effectively integrate multi-source perceptual information, improving training efficiency and control performance, and demonstrating great effectiveness and practical application value.