Large Models for Cooperative Control of Connected and Autonomous Vehicles

Integrating large models (LMs) into future vehicles and transportation systems marks a significant advancement in mobility and transportation technology. Incorporating artificial intelligence and machine learning, these LMs are poised to revolutionize various aspects of transportation. This paper proposes LMs-based approaches for cooperative control and coordination of connected and autonomous vehicle (CAV) fleets. Specifically, algorithms based on the alternating direction method of multipliers (ADMM) are developed for distributed optimization of CAV trajectories. The synchronous ADMM and asynchronous ADMM algorithms enable parallelized coordination of large-scale CAV systems. Simultaneously, we propose a distributed training scheme where each CAV trains its cost and dynamics networks on simulators local to each vehicle. A central coordinator interacts with the vehicles to tune the coupling networks. Then, we introduce an innovative car-following model named the integrated velocity and acceleration fusion model that integrates state information from multiple lead and following vehicles to determine the optimal acceleration for the subject CAV. While we utilize graph sample and aggregate –based neural network and the gated recurrent unit and propose a model for recognizing driving intentions and predicting the trajectories of surrounding vehicles based on these theories. Simulation results demonstrate enhanced traffic efficiency, safety, robustness, and scalability using LMs for cooperative control of CAV.

I. Introduction

Connected and autonomous vehicles (CAVs), endowed with advanced sensing, communication, and control technologies, offer transformative potential for the future of transportation [1]. The primary advantages of CAVs include improved traffic flow, enhanced road safety, reduced congestion, lower emissions, and superior mobility services [2], [3], [4]. However, several complex challenges must be addressed to harness these benefits on a large scale fully. These challenges encompass perception, planning, control, and coordination within expansive transportation networks that include CAVs and human-driven vehicles (HDVs), pedestrians, and various infrastructure components [5], [6]. Effectively integrating CAVs into this dynamic mix demands sophisticated solutions that can handle the complexities of real-world traffic environments and the diverse behaviors of different road users. Achieving this integration is crucial for creating more intelligent, responsive, sustainable transportation systems.