The driverless mile is now a business

The DARPA Grand Challenge in 2004 ended with every vehicle failing, the best one after seven miles. A year later, five cars finished a 132-mile desert course. That gap, from total failure to routine success in twelve months, taught the field a lesson it then had to unlearn: the last few percent of driving takes far longer than the first ninety.

What took two decades

Driving on a closed desert course is a navigation problem. Driving in a city is a prediction problem. The car has to guess what a cyclist will do, whether the pedestrian at the kerb has seen it, whether the double-parked van will move, and what the human driver behind it will tolerate. None of that is solved by better sensing alone.

Three developments got the industry over the line. Machine learning replaced hand-written rules for prediction and planning. Sensor costs fell by more than an order of magnitude, particularly for lidar. And operators learned to launch narrowly, mapping a single service area intensely rather than attempting general driving.

The sensor argument, briefly

One camp uses cameras, radar and lidar, and builds a detailed prior map of the service area. The argument is redundancy: lidar measures distance directly and works in the dark, so the system does not depend on inferring depth from images. The cost is hardware and mapping effort.

The other camp uses cameras only, and argues that humans drive with two eyes, that a vision system trained on enough data will exceed them, and that anything requiring a prior map cannot scale to every road. The cost is that everything depends on one sensing modality being right.

The honest summary in 2026 is that the lidar camp has more driverless miles on public roads and the camera camp has a cheaper bill of materials and a far larger fleet collecting data. Both claims can be true, and the question of which approach wins is a question about cost curves, not about physics.

The economics that actually decide it

A robotaxi replaces a wage with capital and support. Capital is the vehicle plus the sensor suite plus depreciation. Support is remote assistance staff, cleaning, charging, depots, maintenance and insurance. The first is falling fast. The second is the reason profitability is slower than it looks.

Watch the ratio of remote operators to vehicles. A service with one operator per fifty cars has a different cost structure from one at one per three, and companies do not usually publish the number. Watch utilisation too: a taxi earns only while occupied, and a fleet sized for the Friday evening peak sits idle on Tuesday morning.

Insurance is the quiet variable. Driverless fleets have so far reported lower claim rates per mile than human drivers in the same areas, which matters more to a balance sheet than any demo. A single severe incident with clear fault, however, can pause a service for months, and a pause costs revenue on an asset that still depreciates.

What it does to cities

The optimistic story is fewer parked cars, fewer collisions and mobility for people who cannot drive. The pessimistic story is more empty miles, because a car with no driver has no reason to park and every reason to circle. Both happen, and which dominates is a matter of road pricing and kerb regulation, not technology.

Driver employment is the other question. The transition is slower than headlines suggest, because the vehicles are concentrated in a few dense, well-mapped, warm, sunny cities and the job of driving includes loading, helping and dealing with exceptions.

What to watch

Watch for the first operator to publish cost per vehicle mile honestly. Watch for service in a city with real winter. And watch the insurance filings, which contain more truth about safety than any press release.