Edge-Aided Collaborative Inference for Visual Language Navigation-Based Intelligent Vehicles

In Visual Language Navigation (VLN)–based autonomous driving, the limited field of view of an individual vehicle hinders the acquisition of comprehensive environmental information, thereby affecting navigation reliability. Collaborative perception alleviates this issue by enabling information-sensing vehicles (ISVs) to collect visual observations along the trip, which information-requesting vehicles (IRVs) can then exploit to enhance scene understanding and navigation decisions. To further reduce the substantial reasoning burden that exceeds the computational capacity of individual vehicles, edge devices (EDs) equipped with large language models (LLMs) can provide efficient VLN inference services. While such collaboration significantly enriches scene awareness, it also introduces several challenges. First, ISVs and EDs possess heterogeneous resources and behave as self-interested entities that incur energy, communication, and computation costs, necessitating appropriate incentive mechanisms to sustain their participation. Second, an ISV may be simultaneously covered by multiple EDs, and selecting the most suitable ED depends on the dynamic characteristics of wireless links, vehicle states, and edge resources. Furthermore, the computation cost of large-model inference at the edge is difficult to model accurately due to the complexity of VLN pipelines. To address these challenges, this paper proposes an edge-aided collaborative inference framework that coordinates ISVs, IRVs, and EDs within a hierarchical incentive structure to enable stable and efficient cooperation. A vehicle–edge matching algorithm is developed to determine optimal associations, and an empirical validation framework is introduced to accurately estimate the costs of large-model–based VLN inference at the edge. Simulation results demonstrate that the proposed framework converges rapidly and substantially improves the utilities of ISVs, EDs, and IRVs, as well as the social welfare.