The easiest road on the network is the one that beat Waymo

On 19 May 2026, Waymo quietly stopped taking passengers onto freeways. On 8 June it filed a voluntary software recall with the American regulator covering nearly 4,000 vehicles. On 29 July it started again, Phoenix first. The cause, in the filing's own terms, was not velocity, not weather, not a pedestrian and not another driver. Thirteen times — six in Phoenix in April, seven in San Francisco in May — a Waymo had driven into a stretch of freeway that was closed for construction. It “did not recognize and drove past ramp closure signs”.

Nobody was hurt. No collision was reported. And yet this is, on the evidence of the last year, the most instructive failure in autonomous driving, because of where it happened.

Every regulator in the world that has licensed a car to drive itself while the owner reads email has licensed it on a motorway. Every robotaxi company that has taken a human out of the driver's seat has done it off the motorway, and gone there last, if at all. Two industries, the same word, and completely opposite conclusions about which road is hard.

Both cannot be right, so we did what we did with the export gap: took each explanation on offer and looked for the case that breaks it.

The two answers, stated plainly

Start with the cars you can buy. The most permissive consumer system on sale in Europe is BMW's Motorway Assistant, approved under the new UN Regulation 171 and launched with the Neue Klasse iX3. It will drive hands-off at up to 130 km/h in nine countries — Germany, Austria, Italy, France, Belgium, the Netherlands, Luxembourg, Switzerland and Liechtenstein — and it will change lanes when the driver glances in the wing mirror. The single condition on all of it is the road: motorways, and nothing else.

Mercedes-Benz goes one step further in law and several steps back in scope. Drive Pilot is genuine Level 3 — the manufacturer, not the driver, is responsible while it is engaged — approved in Germany to 95 km/h. Also motorway only. And, in the ordinary case, it wants a vehicle in front of it to follow. The most legally advanced self-driving car on sale in Europe needs a human being to be driving. In the other car.

Then Lotus Hyper Pilot, delivered by over-the-air update to European Eletre and Emeya owners on 18 August 2026: lane centring, adaptive speed, supervised overtaking — Germany only, approved sections of the A3, A5, A8 and A99, hands on the wheel.

Now the robotaxis. Waymo carried passengers on surface streets in Phoenix from 2020 and did not put a paying rider on a freeway until November 2025 — more than five years of fully driverless commercial service later. Zoox built a purpose-designed vehicle with no steering wheel and launched it in Las Vegas on city streets. Baidu's Apollo Go scaled to hundreds of thousands of rides a week in Chinese cities, on city roads.

For the robotaxis the motorway was the graduation exercise. For the car makers it is the only classroom they are allowed into. One of these groups has badly misjudged the road.

Explanation one: the motorway is easy because nothing unexpected is on it

This is the standard account, and it is the reason the regulators picked it. A motorway has no pedestrians, no cyclists, no oncoming traffic, no traffic lights, no unprotected turns, no cars emerging from driveways. Traffic is separated, one-directional and legible. It is, in the language of the trade, a narrow operational design domain.

The recall falsifies it. Not one of the thirteen Waymo incidents involved a pedestrian, a cyclist, a junction or an oncoming vehicle. They involved signs. A contractor had closed a ramp, put out the prescribed signage, and the most experienced driverless system on earth drove past it. Waymo's own account of why is the part worth reading twice: the vehicles were “prioritizing the avoidance of other freeway hazards and/or failing to recognize the construction zone”.

Read that as an engineer rather than a publicist. It is not primarily a seeing problem. It is a ranking problem: a system attending correctly to the hazards it was built to attend to, and therefore not attending to the one that had been added to the road that morning by a third party with a truck full of cones.

Explanation two: it is a speed problem

Falsified outright, by the cars. BMW does 130 km/h hands-off on public roads in nine European countries today, with an international type approval. Speed is a known quantity: it costs you stopping distance and reaction time, both of which are arithmetic, and arithmetic is the one thing a computer is definitely better at than a person.

What speed actually takes away is something else, and it is not on the spec sheet. On a surface street, the safe fallback is to stop. A robotaxi that becomes uncertain at 25 km/h can halt in a lane, put its hazards on, wait for a remote operator, and the worst outcome is a queue and an angry taxi driver — which is precisely what happened in San Francisco during the December 2025 blackout, when the cars stopped at dead intersections and waited. Annoying; not dangerous.

On a motorway that fallback is the hazard. A stationary vehicle in a live lane at 110 km/h is one of the most dangerous objects in road transport. The motorway does not remove difficulty; it removes the option to defer. Every uncertainty has to be resolved at speed, in traffic, now.

Which explains the shape of every consumer approval on the market. The regulators did not license the motorway because it is easy. They licensed it because it is the only road where a narrow, well-specified competence is useful and a qualified human being is still sitting there to take the decision the machine cannot defer. Level 3 on the autobahn is not a driverless car. It is a hand-off protocol with a car attached.

Explanation three: it is a perception problem, and better sensors fix it

Partly true, and the part that is true is the least interesting part. Yes, a ramp closure sign is an object to be detected, and detection improves.

But this site has spent the year documenting an industry where sensing has stopped being the constraint. The Xiaomi YU7 carries lidar, 4D radar, eleven cameras, twelve ultrasonic sensors and 700 TOPS on every trim, and drives hands-on. The Zeekr 9X puts five lidar and 1,400 TOPS on its top Chinese trim. Caterpillar, which has run trucks with nobody in the cab for eleven years, will not even tell you what sensors it uses. The correlation between hardware and permission is, across our fifty-five reviews, approximately zero.

And the specific failure was not a failure to see. Waymo's fix, deployed before the July restart, is described as “enhanced scene recognition and routing” — recognition and routing. The second word is the admission: the car also needed to be told where not to go.

Explanation four: it is a mapping problem

This one survives, and it is the same explanation that survived our export-gap examination, which should give it weight.

A construction zone is a map that changed this morning and was published to nobody. There is no feed. The authority that closed the ramp told the public with a sign, because the public is assumed to be a person who can read. The map in the vehicle — whether it is Waymo's centimetre-accurate prior or the navigation database that an Xpeng uses to know a motorway interchange is coming — describes the road as it was surveyed. The gap between those two states of the world is where the thirteen incidents live.

That is why Xpeng signing up to the Google Maps Automotive SDK in July 2026, in order to run its navigation-assisted driving outside China, mattered more than it looked. And it is why the rate at which roads are approved, resurveyed and re-validated is now the number that matters more than any TOPS figure. Nobody publishes it. Lotus does not tell you how quickly a fifth autobahn joins the list. Waymo does not publish its freeway mileage separately. Mercedes does not say which stretches of German motorway are in Drive Pilot's geofence and which are out.

Explanation five: put a human back in, and the problem goes away

The industry's working answer, and it is weaker than it sounds.

The consumer Level 3 cars have a human in them, alert and legally required to be ready, and they still carve the motorway down to a sliver: one country, four named roads, hands on, or 95 km/h behind a lead vehicle. If a supervising human made the motorway tractable, the approvals would be broader than they are. They are narrow because the approval reflects what has been validated, and validation is per-road.

Nor does the human-in-the-loop answer survive contact with what the trucks do. Aurora runs genuinely driverless Class 8 tractors on interstate freight in the Sun Belt — ten routes, roughly 440,000 driverless miles by June 2026, a stated target of 200 trucks by the end of this year. Motorway autonomy without a driver is being done. But look at where the sector's actual driver-out milestones have landed: Kodiak went fully driverless hauling sand on private lease roads in the Permian Basin; Gatik built a business on fixed middle-mile routes where one customer owns both ends; Einride put the human in a remote seat and the cabless vehicle inside industrial sites.

Three companies, three different ways of saying the same thing: the road was the problem, so we found a road we controlled.

Where the paradox is already resolved

There is one fleet that has been driving with nobody aboard, commercially, at speed, in mixed traffic with staffed machines, for eleven years, and has moved 13 billion tonnes over more than 455 million kilometres doing it. It is not a car company. Caterpillar's Cat Command for hauling put autonomous trucks into two Western Australian iron ore mines in 2013, and last week extended the system to its third and fourth Virginia quarries.

Ask what Caterpillar has that Waymo does not, and the answer is not sensors, compute, talent or capital. It is this: when the road changes, the owner of the road is the one who changes it. Caterpillar's own description of MineStar includes the “instant alteration or redesign of mine maps to meet changing operational needs”. A mine's equivalent of a closed ramp is a closure the mine itself drew, on the map the trucks are already reading, before the cones went out.

That is the whole of it. The unsolved problem in autonomous driving is not perception, and it is not speed. It is that on a public road, the map is authored by thousands of uncoordinated parties who publish to human eyes and owe the machine nothing.

What follows, and what to watch

Three things follow from taking the paradox seriously.

• “Motorway only” is not a stepping stone to everywhere. It is a different problem with a different failure mode, and the consumer approvals are narrow for a structural reason rather than a timid one. A system that does 130 km/h hands-off on the A5 is not most of the way to a robotaxi.
• The next real milestone is administrative, not technical. A live, machine-readable feed of roadworks, lane closures and temporary restrictions — authored by highway authorities, not scraped — would do more for motorway autonomy than another generation of lidar. No jurisdiction has one. It is unglamorous, cheap by comparison, and nobody's product roadmap.
• Watch the road count, not the chip. How many autobahn sections Lotus adds. Whether Mercedes publishes its geofence. Whether Waymo separates freeway miles in its safety reporting. Whether Aurora's eleventh route takes as long as its tenth. Those numbers describe the pace of validation, which is the real rate limit, and all four companies currently decline to give them.

Waymo's freeway service is running again, across its whole service area, construction zones included. That is a genuine engineering achievement delivered in ten weeks and it deserves saying. But the fix was “enhanced scene recognition and routing” for the closures the system now knows to expect. The next contractor with a truck full of cones has not been consulted.