Continuous routing when the plan meets reality

The problem

Moving goods through a city is the vehicle routing problem: which vehicle serves which stop, in what order, within which time window, under which capacity and traffic constraint. It is NP-hard. Mature heuristics handle a single morning plan well, but the city does not hold still: a driver calls in sick, a road closure removes routes at noon, same-day orders arrive after cut-off, an afternoon spike pushes past planned capacity. Each event invalidates the morning plan.

The window to respond is set by the work itself. A driver is committed to the stop in front of them; the decision that matters is where they go next, and it has to be ready before they complete that drop, a few minutes in dense urban delivery. The real task is a high-quality re-plan of the remaining work delivered inside that gap, again and again across the fleet through the day.

Where we think we can add value

An operator running thousands of vehicles and tens of thousands of stops across a metropolitan area sits exactly where conventional re-optimisation breaks: it returns a worse plan, or a good plan too late. This is the same class of problem as our mining work: a large, constraint-heavy schedule rebuilt inside a tight decision window when something breaks.

Status

What is measured and what is not. The market figures on this page come from public research and are established context. Our measured results to date are in mine planning. The performance we describe for logistics is a hypothesis we want to test on a real problem with a real operator, not a delivered result. See /mining.

We have solved a problem of this class at scale: a scheduling problem of more than 200,000 variables returned in a few minutes on a single consumer-grade GPU, against legacy tooling that takes hours or days.