Online Location Planning for AI-Defined Vehicles: Optimizing Joint Tasks of Order Serving and Spatio-Temporal Heterogeneous Model Fine-Tuning

TL;DR

This paper tackles the challenge of jointly optimizing ride-hailing order serving and urban FM fine-tuning using edge-enabled vehicles. It introduces a decentralized GNN-enhanced MARL framework with RankTuner for adaptive LoRA rank, integrating data freshness and AoI into the optimization objective via QoS = $\alpha\,AD I + \beta\,ADU$. The approach leverages a topology graph with RGCon-based state embeddings and trains with MAPPO in an online setting, achieving superior QoS, ADI, and ADU across multiple FM tasks and PoI distributions. The results demonstrate practical benefits for smart cities by balancing passenger service and real-time model improvement, with insights into rank adaptation, distribution effects, and scalability to multiple tasks.

Abstract

Advances in artificial intelligence (AI) including foundation models (FMs), are increasingly transforming human society, with smart city driving the evolution of urban living.Meanwhile, vehicle crowdsensing (VCS) has emerged as a key enabler, leveraging vehicles' mobility and sensor-equipped capabilities. In particular, ride-hailing vehicles can effectively facilitate flexible data collection and contribute towards urban intelligence, despite resource limitations. Therefore, this work explores a promising scenario, where edge-assisted vehicles perform joint tasks of order serving and the emerging foundation model fine-tuning using various urban data. However, integrating the VCS AI task with the conventional order serving task is challenging, due to their inconsistent spatio-temporal characteristics: (i) The distributions of ride orders and data point-of-interests (PoIs) may not coincide in geography, both following a priori unknown patterns; (ii) they have distinct forms of temporal effects, i.e., prolonged waiting makes orders become instantly invalid while data with increased staleness gradually reduces its utility for model fine-tuning.To overcome these obstacles, we propose an online framework based on multi-agent reinforcement learning (MARL) with careful augmentation. A new quality-of-service (QoS) metric is designed to characterize and balance the utility of the two joint tasks, under the effects of varying data volumes and staleness. We also integrate graph neural networks (GNNs) with MARL to enhance state representations, capturing graph-structured, time-varying dependencies among vehicles and across locations. Extensive experiments on our testbed simulator, utilizing various real-world foundation model fine-tuning tasks and the New York City Taxi ride order dataset, demonstrate the advantage of our proposed method.

Figures (11)

• Figure 1: An illustration of the proposed vehicle scheduling framework that integrates order fulfillment and model fine-tuning tasks in a smart city environment: Box 1 shows idle vehicles dispatched to neighboring areas in preparation for future tasks, while Box 2 illustrates vehicles providing services to order customers. Box 3 depicts the process where a vehicle retrieves the latest model and data from the nearest RSU, fine-tunes the model during movement, and uploads the updated model to a nearby RSU. Each RSU, equipped with a server, stores a complete base model, enabling vehicles to perform real-time fine-tuning as they collect data and transfer the refined models to other RSUs.
• Figure 3: The impact of data freshness and data volume on the fine-tuning accuracy of different tasks under different UFMs.
• Figure 4: GNN-based MARL framework. It includes the environment, a GNN embedding module for processing raw state information, an actor-critic module for decision-making, a replay buffer for experience storage, and the RankTuner module for dynamically adjusting LoRA ranks to balance fine-tuning accuracy and efficiency. These components work together to enable agents to make independent, informed decisions while optimizing their actions based on the dynamic environment.
• Figure 5: An example of constructing a topology graph.
• Figure 6: The distribution probability of orders and PoIs in the target area. Darker grid colors indicate higher probabilities of generating orders or PoIs. PoIs are shown as points, with colors representing task types: red for image classification and purple for image segmentation. Point sizes are proportional to data volumes, reflecting task intensity.