The Pulse
Google Open-Sources MilleMiglia for Middle-Mile Logistics
Google Research has released MilleMiglia, an open-source instance generator for creating realistic benchmark problems in middle-mile logistics.
AI.info Team ·
Google targets the missing benchmark in freight optimization
Google Research has released MilleMiglia, an open-source instance generator designed to give researchers realistic test problems for middle-mile logistics. The project creates synthetic freight networks that model how shipments move between distribution centers, change vehicles and wait for scheduled connections across regional or continental routes.
Google researchers Aymane Lotfi and Thibaut Cuvelier describe the release in a September 18, 2026 post. The accompanying GitHub repository contains the C++ implementation, documentation and a sample instance.
“MilleMiglia bridges the gap between academic theory and industrial logistics by providing open-source, realistic benchmarks that allow researchers to optimize complex middle-mile networks, ultimately leading to more robust and efficient global supply chains.”
— Aymane Lotfi, Software Engineer, Ads & Commerce, and Thibaut Cuvelier, Software Engineer, Google Research
Middle-mile operations sit between factory or supplier pickup and final delivery. A shipment can pass through several distribution centers and ride multiple trucks before reaching its destination. That structure makes the problem materially different from the vehicle-routing problems commonly used to study first- and last-mile delivery.
Why standard vehicle-routing tests fall short
Traditional vehicle-routing models generally focus on assigning stops to vehicles and sequencing those stops over a limited period, often a single day. Middle-mile networks add scheduled transfers, intermediate storage and handling limits at distribution centers.
Google models the problem as a multi-commodity flow problem on a space-time graph. In that representation, each node identifies a distribution center and a time interval, while arcs represent either vehicle movement or a shipment remaining at a hub for sorting and later dispatch.
The operational constraints are tightly connected. Trucks follow fixed timetables, distribution centers can process only a limited volume during a given hour, and a shipment must arrive in time to catch its next scheduled vehicle. Missing that connection can leave freight at the hub until the next operating cycle.
MilleMiglia turns private freight patterns into synthetic data
Public, high-quality data for middle-mile research is scarce because logistics operators generally treat network structures and demand volumes as commercially sensitive. MilleMiglia addresses that gap by generating artificial instances that preserve important operational patterns without exposing company data.
The generator places distribution centers using gravity models or spatial clustering intended to reflect population and industrial density. It creates shipment origin-destination pairs with realistic volume and weight distributions, then builds structured vehicle rotations connecting major hubs or linking a major hub with nearby smaller centers.
Google says those statistical distributions combine publicly available information from industrial sources with privately disclosed data. The resulting networks are synthetic rather than records of a particular company’s operations.
C++ format supports both exact solvers and machine learning
MilleMiglia is written in C++ and serializes instances with Protocol Buffers. Each generated problem can be stored in a single compact file and consumed by solvers written in different programming languages.
The format puts fixed schedules, distribution-center throughput limits and shipment-transfer prerequisites into the same instance. That differs from the way many vehicle-routing benchmarks split operational features across separate problem variants, such as capacity constraints or delivery time windows.
Google intends the generator to produce several scales of test case. Small instances can support exact algorithms and classroom-style experiments, while larger, continent-wide cases are intended for heuristics and metaheuristics. The same generator can also create large training collections for machine-learning systems.
A starting point for a middle-mile solver suite
Google positions MilleMiglia as an initial step toward a standardized benchmark collection for middle-mile optimization. The company compares the ambition to CVRPLIB, a widely used library of capacitated vehicle-routing problems.
The project is part of a collaboration involving Google, the University of Brescia and ENPC Paris. Google says the partners are also working on a specialized solver and an API for middle-mile operations, with methods designed around the flow structure of these networks rather than adapted from ordinary vehicle-routing software.
The immediate value of MilleMiglia is narrower and more practical: researchers can now compare methods against shared, reproducible test instances instead of relying mainly on proprietary logistics data or simplified academic examples. The public repository includes a sample instance.