Research on an Integrated Decision-Making Mechanism for Logistics Last-Mile Sorting and Delivery Based on Multimodal Large Models

Abstract

With China's annual express delivery business volume exceeding 120 billion parcels, the logistics last mile faces systematic challenges such as low sorting efficiency, static delivery route planning, and insufficient multimodal data fusion. This study focuses on small and medium-sized express delivery stations, proposing an intelligent decision-making mechanism based on multimodal large models to achieve deep synergy and dynamic optimization between sorting and delivery processes. The research constructs a multimodal fusion architecture integrating visual perception, textual semantics, and spatiotemporal data. An improved YOLOv8 model combined with a Dual-Branch Routing Attention (DBRA) mechanism is employed to enhance waybill recognition accuracy in complex scenarios to 99.5%. A Spatio-Temporal Graph Convolutional Network (STGCN) is designed for dynamic route planning, which integrates multi-source real-time information such as traffic, orders, and courier status through causal inference, improving delivery efficiency by over 30%. Pilot implementations at multiple stations in Jinhua City demonstrate that the system significantly reduces sorting error rates and delivery overtime rates, forming an intelligent closed-loop of "perception-decision-collaboration."