I’ve been working on a large-scale last-mile routing system and wrote a technical report about the architecture.
The main thing I found is that once the problem gets large enough, the bottleneck is not only route optimization. It becomes a systems problem: partitioning, boundary repair, graph reuse, bounded route-level optimization, and orchestration.
The report evaluates the system on the public Amazon Last Mile Routing Research Challenge dataset using a shared external OSRM/Google measurement protocol. It does not claim to reproduce Amazon’s internal routing objective.
Main results:
- 23.3% less measured distance under the external protocol
- 11.1% fewer routes
- 1M stops processed in ~20 minutes on commodity hardware
I’d be interested in feedback from people who have worked on routing, optimization, dispatch systems, logistics, or large-scale scheduling. What would you expect to see in the evaluation to make this more convincing?