AI can make takeaway deliveries more efficient

Wolt, Just Eat and Uber Eats. We are ordering takeaway like never before.

In 2025, delivery accounted for 22 per cent of global consumer spending on restaurants and other food-service businesses – up from just nine per cent in 2019, according to market research company Euromonitor International.

The growing number of orders presents delivery platforms with a dilemma: Should a courier leave as soon as the pizza is ready? Or should the order be held for a few minutes so that it can be delivered together with another order?

Combining several orders allows couriers to travel fewer kilometres and make deliveries more efficiently. But if the platform waits too long, the customer risks both a delay and a cold pizza.

Researchers have now developed an AI-based model that can help platforms find the right balance. The model combines more orders on the same trip without increasing the time for which they are held, while maintaining an on-time delivery rate of more than 95 per cent.

The model was tested on more than 650,000 orders from Meituan, one of China’s largest on-demand food delivery platforms. The data cover 22 geographical service areas, ranging from busy urban districts to suburban and lower-demand areas.

The study is the first paper with an SDU affiliation published in Transportation Science, the flagship journal of the Transportation Science and Logistics Society of INFORMS.

More orders delivered on the same trip

During busy periods, the researchers’ model required up to six per cent fewer couriers per order than the platform’s existing approach. At the same time, more than 95 per cent of orders were delivered on time.

The model was also better at identifying orders that could be delivered together. Depending on the time of day, it grouped between 13 and 24 per cent of orders, compared with between 7 and 17 per cent under the platform’s existing practice.

Importantly, these gains were achieved without holding orders for longer.

– The key point is that the model does not simply try to make deliveries as fast or as inexpensive as possible. It weighs several competing objectives and finds a balance between efficient routes, punctual deliveries and the available courier capacity, says Yihua Wang, Assistant Professor at the Department of Technology and Innovation at the University of Southern Denmark.

AI learns from real-world decisions

When an order comes in, a delivery platform does not know which new orders will arrive a few minutes later. It must therefore make its decisions under considerable uncertainty.

The researchers combine two methods to address this challenge. First, the model uses inverse optimisation to learn from the platform’s previous decisions and determine how different operational objectives are weighted. It then uses deep reinforcement learning to identify which decisions produce the best results over time.

The model assesses which orders should be dispatched immediately, which can be held briefly, and how the dispatched orders can best be grouped.

The results show that in areas and periods with relatively few new orders, holding an order slightly longer can increase the possibility of combining it with another delivery. Orders travelling longer distances, by contrast, are less suitable for holding because they are less likely to find a compatible order with which they can be grouped.

Ready for testing in a live system

According to Yihua Wang, customers would primarily benefit from more reliable and timely deliveries. Couriers could deliver more orders on each trip and travel a shorter distance per order, while platforms could reduce their overall logistics costs and make better use of their available courier capacity.

The technology required to use the model already exists, but the approach has not yet been deployed as part of a platform’s live delivery system.

– The next step would be to work closely with a delivery platform to integrate the model into its system and test it under real operating conditions. However, the solution has been designed specifically for the type of high-frequency decisions that these platforms already make on a daily basis. While the approach has so far been tested in food delivery, it could also be explored with logistics companies operating large networks of distribution centres, including those in the Danish-German border region. This could provide a strong foundation for future industry collaboration and broader practical applications, says Yihua Wang.

She contributed to developing the modelling and computational framework, implementing and testing the methods using real-world data, and analysing the operational implications of the results.

The research paper was written by Yihua Wang, Long He, Zhengling Qi and Stefan Minner.

Read the scientific paper