Multivehicle Cooperative Lane-Change Trajectory Planning: An Imitation Learning Approach

Multivehicle cooperative lane-change (MVCLC) is a challenging problem in multivehicle trajectory planning, as it involves complex coordination and collision avoidance among vehicles in dynamic environments. While optimal control-based methods have demonstrated promising performance, their high computational complexity often limits practical deployment. To address this, we propose a novel imitation learning (IL) framework for MVCLC trajectory planning, which integrates optimal control-based trajectory generation with neural network (NN) training for efficient real-time inference. Specifically, we first formulate an optimal control framework for MVCLC and employ numerical optimization to generate high-quality reference trajectories. These trajectories are then used to construct a diverse dataset for training two NNs, enabling fast trajectory generation via IL. To ensure the safety and feasibility of the generated trajectories, we introduce a safety monitoring module that guarantees collision-free execution. Extensive simulation results demonstrate that the proposed method achieves near-optimal performance with significantly improved computational efficiency—achieving a reduction in planning time by 97.1% compared to traditional methods.