Robotaxi expansion tests where the evidence still holds

Waymo’s Madrid preparations (2026-10-20) invite a broader question: what should a robotaxi operator have to demonstrate when entering another city? The appealing answer is experience. A system that has already driven extensively should bring something valuable with it. The difficult part is deciding exactly which conclusions travel with that experience.

A city name is a convenient commercial milestone, but a poor unit of technical comparison. Two service areas could look similar on a map while presenting quite different demands. Conversely, a particular type of junction in another country could be more familiar to a system than an unusual junction in its original city. Expansion deserves to be examined through the evidence that remains applicable and the evidence that must be rebuilt. The following is an analytical framework, not an assessment of undisclosed Madrid test results.

A map leaves out much of the boundary

NHTSA’s 2018 research framework describes an operational design domain through constraints including road type, speed, lighting and weather. NHTSA research.

For assessing expansion, we would distinguish geographical reach from coverage of conditions. Imagine two operators serving the same district. One accepts trips only during a narrow, predictable window; the other proposes a much broader service. Counting their mapped streets equally would conceal the difference in their promises. This is a hypothetical comparison, not a description of either company’s offering.

Coverage also needs a denominator. The share of all roads, the share of likely passenger journeys and the share of journeys supportable throughout the advertised hours answer different questions. A small excluded connection could matter more to customers than many included streets. A boundary may be technically coherent yet commercially awkward if it separates common origins from destinations.

The useful disclosure would explain where a passenger’s journey stops being supported, rather than present a single impressive area figure. Expansion then becomes a change in a defined service promise that can be evaluated.

More observation is not the same as resolving uncertainty

RAND’s 2016 study found that establishing injury and fatality reliability through driving alone could require impractically large mileage. RAND research.

That finding should not be stretched into an argument that road experience is worthless. Our narrower inference is that the purpose of an additional test matters. Repeated success on a familiar route might strengthen one conclusion without addressing the uncertainty that blocks a different route.

Consider a hypothetical temporary street closure coinciding with an unusually crowded pickup area. The analytical question is whether the combination exposes an assumption that earlier tests left untouched. Counting either condition separately would not necessarily answer it. Nor does inventing a dramatic scenario make it relevant: its plausibility and connection to the proposed service still need explanation.

A strong account of progress would therefore describe which uncertainty a test was intended to reduce, what changed after the result, and what remains unresolved. That is more informative than treating every additional kilometre as an interchangeable increment of readiness.

Reusing technology still requires a local argument

Waymo’s 2025 deployment-readiness description combines qualitative and quantitative evidence across methods and considers uncertainty in new geographical environments. Company methodology.

An operator’s published method should be distinguished from independent verification that a particular deployment satisfies it. We would ask which previous conclusions are being reused, why the new conditions do not invalidate them, and which conclusions require fresh evidence. These questions permit substantial reuse without granting automatic portability.

For a simulated test, the question would be whether the features that decide the outcome are represented credibly. For a physical test, it would be whether success supports the proposed operating claim or only the tested arrangement. Agreement between two tests is less reassuring if both depend on the same unexamined assumption. These are proposed ways to interrogate evidence, not findings about any operator’s test quality.

The UK’s trialling code, updated in June 2026, expects a domain-specific safety case and recommends a public abridgment. It is UK guidance, not Madrid’s approval process. UK guidance.

The wider editorial lesson is about legibility. A useful public explanation should let an outsider understand the claim, the limits of its support and the conditions that would prompt reconsideration. It need not disclose every proprietary detail to distinguish a demonstrated capability from a future ambition.

Availability and utilization can point in different directions

The commercial analysis needs equal care. We use service availability to mean whether a requested journey can actually be supplied under the stated conditions. Fleet utilization concerns how much of the fleet’s time or capacity is used, with the definition stated explicitly. Neither is a substitute for safety evidence.

A hypothetical operator could keep its vehicles busy by concentrating on an easy-to-serve pocket while leaving many desired trips unavailable. Another could reserve vehicles to shorten waits and provide broader availability, accepting lower measured utilization. The first number would look better without necessarily describing the more useful service. Equally, wider availability could be costly enough to undermine an otherwise attractive proposition.

Local adaptation costs should therefore be examined alongside the promise being delivered. In a cost model, mapping work, validation, support staffing, facilities and vehicle downtime belong in separate lines. These are categories worth investigating, not estimates of Waymo’s spending. A favourable cost per occupied trip could still omit significant launch work, empty travel or support capacity. A poor early figure could include genuinely reusable investment. The accounting boundary changes the interpretation.

What would make the next city convincing?

The optimistic interpretation is that a common system can reuse enough capability and testing infrastructure to make successive expansions easier. The sceptical interpretation is that local exceptions absorb much of that saving. Both are coherent hypotheses. City counts alone cannot decide between them.

Better evidence would show how the incremental validation burden changes while the service promise remains comparable. Faster preparation for a narrower offering is different from faster preparation for an equally broad one. Likewise, the ability to suspend or restrict service responsibly may support a safety argument while weakening an availability claim. Both effects deserve to remain visible.

This article establishes neither local collision performance nor weather limits, intervention outcomes or per-city economics. Those require location-specific evidence. Its conclusion is more modest: expansion becomes meaningful when the added operating promise is clear, the transferred evidence is justified and the remaining uncertainty is acknowledged. A larger map is useful only when readers can understand what the new boundary actually means.