AI in Last-Mile Delivery and Route Optimization

Routing delivery vehicles is a textbook case of the vehicle routing problem (VRP), a generalization of the traveling salesman problem. Given a depot, a fleet and a set of stops, the goal is routes that minimize total cost while meeting constraints. These include vehicle capacity, driver shift length, customer time windows, and sometimes pairs of pickups and deliveries. The problem is NP-hard: exact solutions become impractical as the number of stops grows, so real systems rely on heuristics (rules of thumb that find good, not guaranteed-best, answers). Classic methods include the Clarke-Wright savings algorithm from 1964 and local-search moves such as 2-opt, which untangles route segments that cross. Modern solvers, including open-source tools like Google's OR-Tools, combine these with metaheuristics that explore many candidate solutions. UPS's ORION system, rolled out in the 2010s, is a well-known example. It optimizes the order of each driver's stops. UPS is also known for a long-standing preference for routes that avoid left turns across traffic where practical, trading slightly longer distances for less idling and fewer risky turns. Machine learning adds prediction on top of optimization. Models estimate travel times by time of day, service time at each stop (finding parking, climbing stairs, getting a signature), and the chance a delivery fails. In 2021 Amazon and MIT ran a research challenge asking teams to learn the routes experienced drivers actually prefer, because the mathematically shortest route isn't always the one that works on the street. Food and grocery apps face a harder, dynamic version. Orders arrive all the time, so dispatch decides in real time which courier gets which order and whether to bundle orders, balancing speed against keeping couriers busy. A common misconception is that routing is just shortest-path navigation. Navigation finds a road path between two points. Route optimization decides the order and assignment of hundreds of stops.