California DMV Pulls Waymo From Freeways. When Should They Do That?

The California DMV has ordered Waymo to stop operating and testing its robotaxis on California freeways after 3 vehicles entered an active construction zone, and another vehicle attempted to pass police vehicles which were driving a “break” to close the freeway. This time, the DMV gave the order, and it "will remain in place until Waymo demonstrates to the Department that it has taken effective corrective actions to address the deficiencies that precipitated the restrictions," according to the DMV.
That’s strong language. Back in May, Waymo voluntarily withdrew from freeways after some incidents where the vehicles also drove into construction zones. Waymo reported that this problem was fixed and they resumed those operations. The recent incidents, according to a report, had Waymo vehicles driving in a live construction site with workers present. Waymo points out that there were no “injuries or collisions.” The DMV disagreed and declared that Waymo’s freeway operations “pose an imminent hazard to public safety.” The DMV is concerned this happened in spite of Waymo’s corrective actions after the May incidents.
Waymo has voluntarily curtailed service a number of times, not just over this, but other incidents with weather and flooding, among other things. They also were very slow to even deploy passenger service on freeways to begin with, operating for years only on city streets before expanding.
While unrelated to this, Waymo also suspended all operations in Northern California on the 18th due to abnormal rains. This is also an abundance of caution, and at least two Waymos got stuck, but something they won’t be able to do when people depend on them. The rainfall was record-setting, though Waymo hasn’t got a great track record with floods.
Robotaxis make mistakes all the time, of course, and will continue to make mistakes into the indefinite future. Nobody has perfection as a goal or anticipates it. As robocar deployments grow in size we will see reports of problems more and more. Indeed, counter to intuition, the number of daily mistakes probably will keep increasing even if the rate per mile driven is actually decreasing. This makes the question of how to handle mistakes more difficult.
When to Curtail?
For the DMV, the question of when to use its strong power to pull vehicles from some or all roads is an important one. For a vendor like Waymo, there are also forces pushing to do or not do voluntary shutdowns of service.
Today, Waymo is still a growing service in development. It will be for a few years to come, though it is also scaling while developing. During development, you want to be extra cautious, and you will curtail service if you are in doubt, just to be on the safe side. That’s why Waymo has voluntarily curtailed most of the time, so it’s interesting that the DMV beat them to the punch. While freeway service is a vital component for customers in the suburbs and for trips to the airport, Waymo won’t suffer much for this because it does not really run a business right now, and the service is small enough that even if riders don’t ride because they need the freeway, there are plenty of other riders with trips that work on surface streets to keep them busy.
It should be noted that Tesla’s pilot service does not run on freeways at all, nor does it run in California at all, except if an employee is in the car. Zoox also does not run on freeways as yet.
Once you have a real production service, things change. Once you have customers who depend on you, customers who give up car ownership or stop using other services, you can’t just say, “we’re not running this week on your routes.“ Shutdowns can happen, but they need to be on the frequency of other big shutdowns, such as freeway closures for severe weather or transit shutdowns, and even that’s a lot to tolerate. If people decide they can’t depend on you, they won’t be loyal customers, and won’t replace cars.
In the earliest days of all teams, any serious safety incident would bring a call to “shut down the fleet.“ That period ended some time ago. Now there is a risk equation. If a problem is discovered, how likely is it, and how severe is it? In addition, there is the more subjective question of “How will others, like the public and regulators, react to it?“
The risk calculated will compare to the risk of not curtailing. This is both business risk as outlined above, but also risk to the public. That’s because today, robotaxis are driving more safely than humans. If you shut down robotaxi service, the passengers in the back will switch to other modes. Some will switch to driving themselves or riding in a human driven taxi like Uber. (Some will also avoid trips or use transit.) When they switch to human driven vehicles, they will have more crashes. Waymo released data that they believed that in their first 271 million miles of operation, they had prevented 871 injuries by taking drivers out from behind the wheel. I suspect that number overstates it, but even at half that, it’s an impressive number. A round number suggests that for every million miles of Waymo operation you shut down, there would be 2 more traffic injuries.
The good news is that as long as Waymo can just redirect its vehicles to other riders, this is not true, but it becomes true once curtailment just sends vehicles out of service without equally safe competitors to take up that slack, which will happen in future.
Waymo’s incidents look scary, though for now, there were no collisions, though that’s certainly not guaranteed. More details have not been published. What’s harder to understand is the combination of “This clearly is bad and should not be happening” and “We need to curtail operations if this can happen.“ It is not a binary matter.
The DMV has to make this calculation too. While their job includes assuring that Waymo and other companies they give permits to operate safely, their bigger job is to assure that our roads become safer. They may not have done that when they ordered Cruise to shut down all operations. As a result of that order, Cruise was closed down. While the DMV did not order that it should have been foreseen as a probable consequence. Cruise’s incident had two components. The first was a scary and serious but extremely rare event, most of which was not Cruise’s fault. The second was a culture of secrecy at Cruise that led to a very brief cover-up. There was extremely low risk that Cruise would get another pedestrian thrown in front of a car in the coming weeks, or even that they would drag another one. The cover-up was another issue that needed punishment and deterrent, but not this level. The DMV’s actions, though not entirely intentionally, left the robotaxi field without competition for many years. This was not in the public interest or in the interests of road safety. The cover-up was more serious and deserved strong action, but not on the road.
Why This Failure?
As noted, robocars will make mistakes regularly, and will do so forever. The particulars of mistakes are actually much less important than what they say about the pattern going forward. Will this mistake repeat? Are the vendors being diligent in fixing mistakes? Are they well motivated to do that? What is their trend of improvement? As such, I tend not to cover individual mistakes unless they speak to those patterns, or teach us how to adjust the motivations.
That said, the police break incident is surprising. Police breaks (where patrol cars weave back and forth in front of traffic with lights on to command traffic to stop behind them to close lanes or a road) are not uncommon. Waymo has seen them, and that means they would have recreated them in their simulator. Waymo tests every new version of their software with billions of virtual miles of simulator driving, and that will include every variation of a police break they have ever seen or thought of, as well as an extremely wide variety of example emergency vehicles. Waymo has declined to answer questions about why their systems didn’t identify the police break, and why this wasn’t detected in simulation.
The DMV also complained about Waymos moving into construction zones. On the plus side, a robocar in a construction zone is probably not going to hit anything there, though it might drive into wet concrete and cause property damage. The DMV seemed upset that even though Waymo had previously had a problem with freeway construction in May and had fixed it, that the fix somehow didn’t “stick.” I doubt that’s the case. I suspect it’s more likely that similar results (going into construction zones, or passing school buses) can result from multiple causes. Fixing one cause doesn’t fix the others, though it should help because the team will do a lot more simulation as part of the process of doing the first fix.
The DMV is right that Waymo needs to improve their construction zone (and police break) handling, and Waymo seems to agree, since they also stopped freeway operations outside California voluntarily. What they may not be right about (they declined to respond to questions on this) is their analysis of the balance between the risk of vehicle mistakes and the risk of ordering vehicles off the road.
It would be good if we got more data on the causes of these failures, so we could answer the real questions about overall safety and the path of improvement. Unfortunately, both regulators and vendors seem to go very silent in these situations, which is not ideal.
In the future, when mistakes like this happen--and they will happen--it may not be so practical, or safe, to shut down a fleet or a segment of operations in response. The hard reality is that we never remove all humans from the road when we find out something dangerous they will do, though we are known to pull specific car models or close certain streets. If we take the approach of “these did something dangerous, so they should not be on the roads” they will never be on the roads, which is a poor answer to the big challenge of overall road safety improvement.
Originally published at Forbes.com in California DMV Pulls Waymo From Freeways. When Should They Do That?.
Copyright © 2026 by Brad Templeton








Add new comment