This repository presents a comprehensive technical report titled "Rethinking Last-Mile Routing at Scale: Near-Linear Planning on Commodity Hardware". It explores innovative solutions for optimizing last-mile routing challenges, particularly at a scale of millions of stops, by addressing routing as a systems problem rather than a mere optimization problem.

Overview

The study emphasizes a systematic approach that integrates components such as partitioning, boundary repair, graph reuse, bounded route-level optimization, and orchestration to enhance route planning on standard hardware. This architecture aims to make large-scale routing workloads feasible without the necessity for extensive infrastructure.

Key Findings

The results derived from testing against the public Amazon Last Mile Routing Research Challenge dataset using a shared external measurement protocol based on OSRM and Google Maps include:

• A 23.3% reduction in measured distance compared to baseline Amazon routes.
• 11.1% fewer routes generated.
• A 17.59% reduction in mean depot-level distances.
• The ability to process 1,000,000 stops in approximately 20 minutes using commodity hardware.

What This Report Covers

This technical report details how to manage large last-mile routing requests efficiently. The architecture supports:

• Parallel constraint-aware clustering
• Constraint-aware vehicle allocation
• Distributed boundary rebalancing
• Bounded route-level optimization
• Localized graph and distance reuse

Important Clarification

It is important to note that this report does not aim to replicate Amazon's internal routing objectives. Instead, it compares released Amazon routing files with generated routes over the same stops using an independent measurement protocol, focusing solely on structural comparisons under consistent external metrics.

Why This Is Significant

Many logistics systems and routing APIs function effectively with smaller requests but tend to struggle with very large workloads due to the need for manual partitions or significant infrastructure. This report offers an alternative strategy by breaking down the routing problem into manageable, composable stages, allowing for predictable planning for workloads ranging from thousands to one million stops using standard hardware.

Additional Resources

A verifier repository for inspecting and reproducing the external distance comparison protocol can be found here.

Citation

For academic referencing, use the following BibTeX entry:

@misc{vizzolini2026rethinking,
  author       = {Martin Vizzolini},
  title        = {Rethinking Last-Mile Routing at Scale: Near-Linear Planning on Commodity Hardware},
  year         = {2026},
  howpublished = {Technical report},
  url          = {https://github.com/vizzito/last-mile-optimizer-paper/releases/tag/v1.0.