RideSmart: Pre-trained Large Models for Delivery Route Planning
Zhao Li, Yusheng Jiao, Yuduo Shi, Donghui Ding, Jiacheng Wang, Jiarun Zhang, Haitao Xu · 2025
Millions of e-bike delivery riders in China navigate complex urban environments daily, delivering a wide range of goods to hundreds of millions of customers. Their reputations-and, consequently, their earnings-are largely determined by their ability to ensure fast and timely deliveries. Despite the critical importance of efficiency, most riders rely solely on their accumulated experience to optimize their delivery routes. Notably, there is a lack of an intelligent route planning system that leverages the vast delivery knowledge embedded in the collective experience of these riders to significantly enhance delivery efficiency. In this paper, we introduce RideSmart, the first spatiotemporal pre-trained large model specifically designed for efficient delivery route optimization. Built on an extensive dataset of 1.5 million trajectory records collected from delivery riders, RideSmart utilizes pre-trained models to capture valuable rider expertise, thereby planning highly optimized routes that enable riders to complete tasks with maximum efficiency, leading to higher earnings. Experimental results show that, given the same input of origins and destinations, RideSmart consistently generates routes that are significantly more time-efficient and distance-conserving compared to those planned by riders based on their individual experience. The system achieves a reachability rate of 90.5%, with 82.4% of its generated routes being as optimal as or more efficient than human-experienced routes. By effectively utilizing the collective experience of millions of riders, RideSmart delivers substantial benefits to every individual rider. A video demonstration of RideSmart is available here: https://youtu.be/JD0N2kQTGcc.