Industries / Transport and vehicles / Trucking and freight / Last-mile delivery
Routing
Routing is planning which stops each driver makes and in what order, and finding the exact spot at each address where the parcel goes.
4
Milestones achieved, from 2004 to 2023.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2004
Fewer left turns
UPS began planning routes to avoid left turns across oncoming traffic, which waste time and fuel at junctions and cause more crashes. UPS says left turns are now under 10% of its trucks' turns, and credits the rule with 10 million gallons of fuel saved; other reports give lower fuel figures.
Credit
UPS
2015
Free solvers for route planning
A solver is a program that searches for the best answer to a planning problem, here the shortest set of routes. Google says it open-sourced its OR-Tools optimization toolbox, which includes a vehicle routing solver, in 2015, after working on it since 2008; other sources date the first public release to 2010. In September 2016 Julien Coupey released VROOM 1.0, an open-source route optimizer that uses free OpenStreetMap road data. Small carriers can now plan routes without paying for software.
Credit
Google (Laurent Perron, Vincent Furnon and colleagues); Julien Coupey
2021
Learning from drivers' own routes
Amazon and the MIT Center for Transportation and Logistics ran a challenge to predict the routes experienced drivers actually choose, which differ from planned routes because drivers know things such as which roads jam. More than 2,000 people in 229 teams took part, and the $100,000 first prize went to William Cook, Stephan Held and Keld Helsgaun. Amazon released 9,184 real routes from five US metro areas, driven in 2018, as a public dataset.
LessonThe shortest route on paper is not the one good drivers drive.
Credit
Amazon; MIT
2023
A large public courier dataset
Cainiao, the logistics arm of Alibaba, and university partners released LaDe, records of about 10.7 million packages handled by 21,000 couriers over six months in several Chinese cities. It is the largest public last-mile dataset by its authors' account, and researchers use it to test route and arrival-time models.
Credit
Cainiao and partners (Wu and colleagues)
Watch for
AI maps of building entrances
Amazon says its Wellspring system, which uses generative AI and photos taken by drivers at delivery, has mapped 2.8 million building entrances, 85,000 mailrooms and lockers, and 202 million parking spots since 2024. It also says new AI routing adjusts routes for darkness, lunchtime parking and bad weather, and gives no figures for the effect.
Achieved whenan outside party measures how far Amazon's stored entrances are from the real doors, or a reduction in delivery time per stop.
Claimed by
Amazon, June 2025 and September 2026
Doorway data for sale
Mapbox says it now offers entrance locations for more than 100 million US addresses, in public preview. It says nearly 70% of failures in the last 100 m of a delivery happen in the final 10 m, from its own customers' data.
Achieved whenan independent test publishes the distance between Mapbox's entrance points and real entrances.
Claimed by
Mapbox, February 2026
Open
01
Where maps put the door
For one city, a public sample of addresses, especially large apartment complexes and offices, with the real delivery entrance located on foot. Then measure how far free and paid geocoders (services that turn an address into a map point) place each address from that entrance.
The briefCredit · unclaimed02
Solvers scored against drivers
A head-to-head test of free route solvers, PyVRP, VROOM, Google OR-Tools and LKH, on the public Amazon routes. Score each on total travel time and on how closely it matches the routes drivers actually took.
The briefCredit · unclaimed
Open 01 · the brief
Where maps put the door
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
For one city, a public sample of addresses, especially large apartment complexes and offices, with the real delivery entrance located on foot. Then measure how far free and paid geocoders (services that turn an address into a map point) place each address from that entrance.
Why it is still open
The accuracy figures found are vendors' own. TomTom claims more than 99% of unit locations within 10 feet in the US, with no published method; Mapbox and Amazon describe their own gains. Amazon's research paper on choosing the doorstep reports gains only against its own older system. No independent audit turned up.
Built with what exists
A phone with GPS; Nominatim, the free OpenStreetMap geocoder; trial keys for commercial geocoders such as Google, Mapbox and TomTom, used within their terms; QGIS, free mapping software; a spreadsheet.
First step
Pick 30 addresses in one neighbourhood, walk to each entrance to record its position, and compare them with Nominatim.
Done looks like
At least 300 addresses, including at least 30 complexes with more than one building, with the true entrance and each geocoder's point published, and the median and 95th-percentile error in meters for each service.
Open 02 · the brief
Solvers scored against drivers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A head-to-head test of free route solvers, PyVRP, VROOM, Google OR-Tools and LKH, on the public Amazon routes. Score each on total travel time and on how closely it matches the routes drivers actually took.
Why it is still open
An AWS sample compares only LKH and OR-Tools, on the driver-match score alone, and found them nearly equal (0.0374 and 0.0372). This research found no comparison that includes PyVRP or VROOM and reports both scores.
Built with what exists
The Amazon Last Mile dataset; PyVRP, VROOM, OR-Tools and LKH; the challenge's scoring code; Python on one laptop or a small cloud machine.
First step
Download the dataset and solve 100 routes with two solvers to check that the scoring code runs.
Done looks like
All 9,184 routes solved by each solver, with a published table of travel time and driver-match score, and the code to repeat it.