The On-Demand Delivery Problem: Assignment of Orders to Warehouses and Couriers

The surge in customers’ preference for online shopping has spurred the growth of on-demand delivery services, exemplified by companies like Getir and Flink. These companies promise near-instantaneous deliveries, typically within a few minutes. To fulfill this promise, multiple micro-warehouses and a courier fleet using e-bikes are employed. To address this problem, the current practice of logistics companies is to statically define spatial areas as polygons for each micro-warehouse and assign all customers within this polygon to the respective warehouse. However, such a static assignment neglects real-time information that might be used to achieve a better workload balance of orders between warehouses. In this work, we suggest a dynamic assignment of orders to warehouses and couriers based on the current workload and previously assigned orders to the warehouses. The problem is formalized as a sequential decision problem, as customers arrive dynamically over time. The goal is to minimize total delays. Because of the time commitment and the unpredictability of customer orders, it is not possible to plan in advance and a strategy is needed to make decisions immediately. We develop a decision policy to solve the considered problem and apply it to problem instances on a simplified grid as well as to instances derived from real-world data of Chicago. Our method is benchmarked to current practices from the industry, showing that a dynamic assignment can substantially reduce delays.