Learning two-dimensional merging behaviour from vehicle trajectories with imitation learning

Merging behaviour is a fundamental yet challenging driving task which has significant impact on traffic flow operations. While numerous efforts have been made on the modelling of decisionmaking of lane-based merging behaviour, little was focusing on the simulation of the complete merging process, which generates two-dimensional merging trajectories allowing for the investigation of merging behaviour's impact on the traffic flow. This study thus aims to develop a twodimensional data-driven simulation model for merging behaviour based on the emerging imitation learning framework for generating realistic vehicle trajectories and traffic characteristics. To account for the specific goal of merging vehicles (reaching the planned destination) and the presence of suboptimal behaviours in real traffic, we incorporate goal-conditioned and confidence-aware mechanisms into adversarial inverse reinforcement learning (is referred to as GC-AIRL) to learn the merging behaviour from real-traffic demonstrations. Using the vehicle trajectories data extracted from the NGSIM dataset, we demonstrate that the proposed model is capable of generating two-dimensional vehicle trajectories with superior efficiency, safety, and comfort performance compared to human drivers. The superiority of the GC-AIRL model is validated by comparing with several bench-marking models, including the basic AIRL model, generative adversarial imitation learning (GAIL) model, a reinforcement learning (RL) based model, and long short-term memory (LSTM) model. Moreover, we examine the transferability of the proposed model in producing merging behaviour at a new site that indicates its applicability in diverse environments. The findings of this study highlight the great potential of the developed two-dimensional merging behaviour model for future application in connected and automated vehicles.