Iterative Learning-Based Cooperative Motion Planning and Decision-Making for Connected and Autonomous Vehicles Coordination at On-Ramps

This paper proposes an iterative learning-based cooperative motion planning and decision-making approach to achieve time-optimal coordination control of connected and autonomous vehicles (CAVs) at on-ramps. In this work, a decentralized learning-based iterative optimization method (DLIO) is first developed to offline train the vehicle merging trajectories in space-time. To guarantee safety and convergence properties at each iteration, the collision-free terminal constraint set and approximated merging time cost are designed using the historical vehicle states as a dataset. For adapting the trained trajectories into dynamic traffic flow at on-ramps online, we systematically analyze the arrival time of inter-vehicle potential actions and model a decision tree to express all possible vehicle cluster passing sequences. Then, a heuristic Monte-Carlo tree search (HMCTS) algorithm with a modified searching principle is presented to derive a minimum-time passing sequence. Also, a customized two-stage velocity planning method is used to regulate the vehicle flow following the optimal sequence and trained condition. The proposed approach is verified on the SUMO and compared with three baselines under different on-ramps traffic demands. Results show that our approach enables the conflicting vehicles to merge into the mainline without queuing, rendering both robust and high efficiency multi-vehicle coordination.