CHEN Jing, ZHAO Cong, MA Yu-cheng, LIU Dong-jie, JI Yu-xiong, DU Yu-chuan
Ride comfort is a crucial factor influencing the acceptance and trustworthiness of autonomous driving technology, which is the foundation of high-quality autonomous driving services. How to continuously improve ride comfort while ensuring safety and efficiency is a key challenge for the application of autonomous vehicles. However, traffic conditions, pavement quality, and acceleration and deceleration of vehicles have an impact on passenger sensation, leading to difficulties in discerning the factors contributing to passenger comfort in complex traffic environments. This further results in speed planning biases and affects vehicle control effectiveness. Therefore, understanding the causal relationship between pavement quality, traffic conditions, speed decisions, and comfort sensation is crucial for enhancing the ride comfort of autonomous driving. In this regard, based on the vehicle-road-cloud integration architecture, this study proposes an intelligent decision-making and control framework for ride comfort improvements in autonomous driving. First, we consider onboard units and edge clouds as intelligent agents. Then, an intelligent speed planning model is constructed for autonomous driving based on deep reinforcement learning, utilizing counterfactual reasoning and expert recommendations to increase training samples in two directions and enhance the understanding of the driving environment and tasks. In experiments, a simulation environment was established using pavement data from Shanghai and traffic data from Next Generation Simulation (NGSIM). Speed control models were trained and tested under different pavement and traffic conditions. The experimental results show that with sufficient training samples, the counterfactual reasoning model can analyze the causal relationships between pavement and traffic conditions, speed planning decisions, and decision rewards. The counterfactual reasoning model analyzes the importance of states and clarifies the focusing points of speed planning at different stages. On the premise of driving safety and efficiency, the intelligent speed planning model trained with counterfactual reasoning and expert recommendation can decrease the longitudinal jerk and annoyance rate by 25.71% and 18.89%, respectively, compared to traditional reinforcement learning models. This results in a significant improvement in autonomous driving comfort, with interpretable driving decision results. The proposed autonomous driving decision-making, control framework, and intelligent speed planning approach can support online ride comfort improvement of autonomous vehicles and promote the development of autonomous driving mobility service.