EdgeDriver: optimising autonomous driving assistance with multi-LLM framework in cloud-edge computing environments

Autonomous driving technology has recently made significant strides, laying the groundwork for the future of transportation systems. This advancement is increasingly powered by the emergence of Large Language Models (LLMs), which have become a focal point of research due to their remarkable common sense and reasoning ability. However, the deployment of LLMs is challenged by their computationally intensive nature, making direct implementation on mobile devices impractical. In this paper, we introduce EdgeDriver, an autonomous driving assistance framework leveraging multiple Large Language Models (LLMs) within cloud-edge computing environments to optimise computational efficiency and service quality. The framework comprises three modules: Perception, Planning and Evaluation, which process real-world driving scenarios to generate driving suggestions. Experiments were conducted using the BBD-100K data set, evaluating the impact of model parameters and input lengths on service quality. Results indicate that a 14B parameter model with 150-word input texts achieves optimal performance, balancing accuracy and computational efficiency. The study concludes that the quality of driving assistance services depends on matching input lengths and model capabilities to real-world conditions, providing valuable insights for future smart transportation systems.