Autonomy Roulette: Why Getting a Robotaxi Is an Operations Problem

A rider recently recorded a trip on a robotaxi in Las Vegas. They took it twice, once out and once back, through a ride-hail app.

We did not take the ride. We are working from their account, and we have kept to what they reported. But it is a useful account, because it separates two things that are easy to run together: how well the car drives, and how easy the service is to use.

The driving was good. Getting the car took about ten tries.

The rider said the service had made a large leap in quality since earlier this year. The driving was smooth, and pedestrian crossings were handled cleanly and reliably.

Then the other half. The rider reported needing roughly ten attempts to be paired with a robotaxi on each leg of the trip. They called it autonomy roulette, and the name fits.

Why getting one is a matter of odds

The reason is in how the service is offered. Riders who request one of several standard ride types may be matched with a robotaxi at no extra cost. They are told when it happens, and they can accept it or switch to a regular ride.¹

So the rider does not simply request a robotaxi. The app offers one when a robotaxi is near and available. Riders can raise their chances by opting in under Ride Preferences in the app’s settings.¹ The rider did not say whether they had.

When the service launched on March 13, 2026, it covered designated rideshare zones at two resorts, curbside spots downtown and the Town Square shopping district.¹

A small service area and, by the rider’s count, odds of about one in ten. None of that says anything about the driving.

It also says something about how this kind of service grows. A ride-hail service with human drivers can bring in more cars when demand rises. A robotaxi service has the vehicles it has. That is our reading, not a claim from the sources, but it is why availability, more than driving skill, will shape what riders think of these services for a while.

Smooth driving does not fix a scarce service

If a rider needs ten tries, the experience they remember is the ten tries.

The ride is the part the technology controls most directly. The service around it is not only the technology. It is where the vehicles are, how many are ready, how fast they are turned around, and whether one is parked close to the person who just opened the app.

This is the same point we made in our post about trains and stations. The vehicle should be waiting where the person is, not on its way back from somewhere else. A robotaxi that spends its idle time driving to a distant depot is a robotaxi that is not near the next rider.

Staging matters in numbers, too. In a San Francisco simulation of 1,700 shared vehicles, banning curb staging raised daily fleet travel by nearly 60%, because vehicles that cannot wait nearby have to keep moving.² Every extra mile is a vehicle that is not available when a rider opens the app.

The service area is a list of properties that said yes

Look at where the Las Vegas service started: rideshare zones at resorts, a downtown curb, a shopping center.¹ Those are places, and each is controlled by someone.

A robotaxi can only pick up where it is allowed to, and it can only wait where someone has agreed to host it. The map of a service is a map of those agreements. That is why the property side is not a footnote. It decides how far the service can reach.

What changes when the supervisor leaves

The rider also noted that a safety supervisor was present, and that it affects the experience. At launch the vehicles had an operator behind the wheel, and the companies said they expected a fully driverless service to begin by the end of this year.¹

That change matters for operations. A person in the car does small jobs without anyone noticing: reporting a smell, a warning light, a mess. Take the person out, and those jobs have to be done somewhere else, on a schedule, by someone, with a record.

What we are watching

We are watching three things in Las Vegas and in every market that follows.

First, whether matching gets easier as the fleet and the zones grow. That is the number riders will feel before any other.

Second, what happens at the point the supervisor leaves the vehicle. The cleaning, checking and reporting that a person did in passing will need a home.

Third, where the vehicles wait. Every ride that starts quickly begins with a vehicle that was parked nearby, charged, clean and checked. Someone had to agree to that space, and someone had to record that it was done.

That is the work we are building XoomPark OS around: find suitable sites near demand, agree each site’s terms, and keep the record of what happened there.

The driving is the part you can see. The operations are the part that decides whether you get the ride.