Public routing examples often reduce logistics to drawing the shortest line between dots. Real middle-mile networks must move freight through facilities with fixed schedules, limited throughput and synchronized departures.
Google MilleMiglia is an open-source C++ generator for realistic synthetic middle-mile logistics instances. Google Research built it to give researchers and optimization teams test problems that resemble industrial networks without exposing a company’s private operating data.
Middle mile is the network between the network
Last-mile routing moves parcels from a local facility to final recipients. Middle mile moves freight among warehouses, cross-docks, distribution centers and regional hubs. A plan must coordinate vehicles and facilities over time, often before the final destination sequence is known.
- Vehicles may leave only on fixed schedules.
- A distribution center can process only a limited volume in each interval.
- Loads may need synchronized arrivals or transfers.
- Routes interact through shared hubs rather than independent destinations.
- A feasible plan can be more important than the mathematically shortest path.
What the generator puts into an instance
| Component | Purpose | Failure exposed |
|---|---|---|
| Facilities and links | Define the network freight can traverse. | Algorithms that assume every pair is directly reachable. |
| Fixed schedules | Restrict when a movement can occur. | Solutions that ignore missed departures. |
| Throughput limits | Bound what a hub can process. | Plans that create impossible facility congestion. |
| Synchronization constraints | Coordinate connected movements. | Routes that arrive after the next transfer leaves. |
| Demand and scale | Create small examples or industrial-size cases. | Methods that work only on toy networks. |
Synthetic does not mean arbitrary
A useful synthetic generator preserves important statistical and operational relationships while removing confidential customer and network data. MilleMiglia is designed to generate instances with realistic structure rather than disclose one operator’s facilities, costs or demand.
That makes the tool valuable for public comparisons, but generated realism still needs validation. Researchers should state which distributions and constraints were used, why they represent the target setting and where the synthetic network differs from production.
A reproducible experiment begins with the seed
- Choose a named generator configuration and preserve it in version control.
- Record the repository commit, compiler and dependency versions.
- Set and publish the random seed for each generated instance.
- Store the resulting Protocol Buffers artifact, not only a screenshot.
- Run every solver against the identical instance and resource limits.
- Report feasibility, objective value, runtime, memory and the optimality gap when available.
- Repeat across small, medium and industrial-size batches.
Protocol Buffers give the project a structured exchange format. They help keep an exact instance stable across languages and systems, but reproducibility still requires the generator version and configuration that created it.
Where AI fits into the research loop
MilleMiglia is an optimization generator, not a language model. AI systems can still participate by proposing heuristics, predicting demand scenarios, selecting solver parameters or learning policies from generated examples. Every learned method should be compared with transparent optimization baselines under the same constraints.
| AI role | Required control |
|---|---|
| Generate a candidate route | Validate every capacity and schedule constraint independently. |
| Predict travel or processing time | Test calibration under distribution shift. |
| Choose solver parameters | Compare against a fixed tuning budget. |
| Explain a plan | Derive the explanation from the verified solution artifact. |
| Operate a planning agent | Keep cost, feasibility and final dispatch behind approval gates. |
What is available and what remains planned
Google has released the generator and describes support for instance sizes ranging from small examples to industrial-scale problems. The research post says the team is working on a solver and API. Those planned pieces should not be described as generally available until their artifacts appear.
The MustHave.ai AI GitHub repositories reference provides complementary optimization, data and evaluation tools. Teams building automated planning systems can use our AI agent controls guide to keep generated recommendations inside budget and approval boundaries.
The practical verdict
Google MilleMiglia fills a genuine research gap: realistic public middle-mile instances are scarce because operating networks are commercially sensitive. The generator can make solver comparisons more reproducible and useful than toy routing examples. Its next test is whether outside teams can reproduce realistic workloads, agree on evaluation suites and connect the planned solver and API without weakening the privacy boundary.
Primary sources
Checked September 19, 2026. MilleMiglia is a logistics optimization research tool, not a generative AI model. The solver and API were described as work in progress.
