A Learning-Based Stochastic Model Predictive Control Method for Online Trajectory Control of Autonomous Vehicles at an Unsignalized Intersection

This paper proposes an autonomous driving framework for unsignalized intersections. The framework consists of two modules: vehicle state inference and vehicle motion planning. A network based on Graph Convolutional Neural Network (GCN) and Gated Recurrent Unit (GRU) encoderdecoder is established. Relying on the relevant historical states of vehicles within the unsignalized intersection, the interactions among vehicles are fully taken into account, and the future trajectory distributions of all vehicles and the collision risks between each pair of vehicles are predicted together. The motion planning module, based on the prediction results of future states, adopts the Model Predictive Control (MPC) method to obtain the desired acceleration input, improving traffic efficiency while focusing on safety. We have built a simulation platform for an unsignalized intersection that simulates human interactions in the Sumo simulation software, and used it for dataset generation and simulation experiments. The results show that our framework demonstrates high efficiency and safety under different traffic volumes.