Intent-Driven Task-Oriented Resource Collaboration for Cloud-Edge-End Vehicular Networks

Vehicular networks should be with great potentials to coordinate traffic scheduling and ensure road safety, where cloud-edge-end vehicular networks have been as promising paradigms to improve vehicle scheduling efficiency according to user task intents. However, as user demands become increasingly complex and diverse, task intent decomposition is crucial for handling the diversity of task intents and the complexity of their content. Furthermore, the dynamic changes in task intents and vehicular network node resources lead to uncertainty in the task execution, which significantly reduces vehicle scheduling efficiency. In this paper, we present an intent-driven task-oriented resource collaboration framework, referred to as ITRC, which consists of the intent, policy, and infrastructure layers from a top to bottom perspective. Specifically, we propose a coarseto-fine complex intent decomposition algorithm assisted by a generative pre-trained transformer. To mitigate hallucination during the decomposition, a hybrid hallucination suppression mechanism is introduced based on similarity analysis and natural language inference. Furthermore, the task-resource uncertainty is solved using a maximum likelihood estimation approach, and an efficient optimal allocation algorithm is proposed for dynamic task scheduling. Extensive simulations demonstrate that the proposed algorithm consistently outperforms compared baseline methods across multiple metrics. In terms of quality of service, it reduces end-to-end latency by 60% to 89%; in terms of resource utilization, it achieves improvements ranging from 5% to 110%. For task distribution balance, the algorithm significantly reduces the variance of resource allocation across nodes to mitigate load imbalance, with the standard deviation decreasing by 38% to 87%. In terms of execution efficiency, the proposed algorithm maintains the shortest running time across all task scales and avoids the exponential time complexity typically observed in traditional optimization algorithms, demonstrating strong scalability and real-time performance.