We often see these limits only when something goes wrong
On a flooded road in San Antonio, an empty robotaxi made a decision no human would have — and kept driving until a creek made the choice for it. The incident, small in its immediate harm, has grown into a recall of nearly 3,800 Waymo autonomous vehicles, exposing a quiet truth about self-driving technology: its limits are often only visible in the moment they are crossed. As these systems expand into more cities and more weather, the distance between what software understands and what the world presents will continue to test both the industry and those charged with overseeing it.
- A driverless Waymo car drove into a flooded San Antonio road on April 20th and was swept into a creek — no passenger aboard, but the software's blind spot was suddenly impossible to ignore.
- The incident triggered a voluntary recall of 3,800 robotaxis running fifth and sixth-generation systems, marking the most significant safety action in Waymo's history.
- Waymo has already restricted its vehicles from flood-prone zones and suspended San Antonio service entirely, but these are stopgaps while a permanent software fix is still being built.
- Experts warn that operational blind spots in autonomous systems are rarely discovered before something goes wrong — and as robotaxi fleets grow, more such gaps are likely waiting to surface.
- The pressure now falls on regulators and companies alike to move from reactive incident response toward proactive mapping of where these systems cannot safely go.
On April 20th, a Waymo robotaxi with no passenger drove onto a flooded road in San Antonio, Texas. The software kept the car moving forward. Water rose. The vehicle was swept into a creek. That single, quiet failure has since grown into the largest safety action Waymo has ever taken — a voluntary recall of nearly 3,800 autonomous vehicles across the United States.
The recall, filed with the National Highway Traffic Safety Administration, covers all vehicles running Waymo's fifth and sixth-generation driving systems. The vulnerability is both simple and serious: in wet conditions, when floodwater obscures lane markings and blurs road boundaries, the software cannot reliably detect danger. It proceeds anyway.
Waymo, owned by Alphabet, has moved quickly on temporary measures — restricting vehicles from flash flood-prone areas and suspending San Antonio service entirely. But a permanent fix, a software update capable of recognizing and avoiding flooded roads, is still in development. Until it arrives, thousands of cars operate with a narrowed geography and an unresolved gap.
The company has been transparent, filing the recall voluntarily rather than waiting for regulatory pressure. What it has not disclosed is whether the vehicle sustained damage, or how it was recovered from the creek.
Jack Stilgoe, a science and technology policy professor at University College London, offered a broader frame: every autonomous system has operational limits, and those limits tend to stay invisible until something breaks. Rain, flooding, construction zones, unusual markings — these are not rare exceptions. They are the ordinary texture of roads that human drivers navigate without a second thought. The San Antonio incident hurt no one, but it issued a clear warning. Whether the industry and its regulators choose to anticipate the next one, or simply wait to discover it, remains the open question.
On April 20th, a Waymo robotaxi with no one inside drove onto a flooded road in San Antonio, Texas. The vehicle kept going. Water rose around it. The car was swept into a creek. That single incident—one empty car, one bad decision made by software—has now triggered the largest safety action Waymo has taken: a voluntary recall of nearly 3,800 robotaxis across the United States.
The recall notice, filed with the National Highway Traffic Safety Administration this week, covers all vehicles running Waymo's fifth and sixth-generation autonomous driving systems. The problem is straightforward and alarming: the software that guides these cars cannot reliably detect flooded roads and steer around them. In wet conditions, when water obscures lane markings and road boundaries blur into surrounding terrain, the system fails to recognize danger. It simply proceeds.
Waymo, owned by Google's parent company Alphabet, has already begun damage control. The company has restricted its robotaxis from accessing areas where flash flooding is known to occur. Service in San Antonio itself remains suspended. But these are temporary patches. The real fix—a software update that will teach the system to recognize and avoid flooded roads—is still being developed and tested. Until that update rolls out, thousands of cars sit partially sidelined, their operational geography shrinking.
A spokesperson for Waymo acknowledged the incident and the mitigations now in place, framing the recall as a responsible response to an identified vulnerability. The company has not disclosed whether the vehicle was damaged, whether anyone was nearby when it happened, or how long it took to recover the car from the creek.
Jack Stilgoe, a professor of science and technology policy at University College London, sees this recall as a window into a larger problem. Every autonomous vehicle system, he explained to the BBC, has operational limits—conditions under which it simply cannot function safely. The trouble is that these limits often remain invisible until something breaks. "We often see these limits only when something goes wrong," Stilgoe said. As more robotaxis deploy across more cities, he warned, more such problems will surface. The question is whether regulators and companies will discover these gaps through incident reports and recalls, or whether they will map them out in advance.
The broader context matters here. Over the past year, multiple driverless car companies have experienced incidents that raised public concerns about safety. Each incident reveals not a flaw in the technology itself, necessarily, but a gap between what the technology can do and what the real world demands of it. Rain, flooding, snow, construction zones, unusual road markings—these are not edge cases. They are ordinary conditions that human drivers navigate constantly.
Waymo's response has been measured and transparent. The company did not hide the incident or delay reporting it. It filed the recall voluntarily rather than waiting for regulators to force action. But Stilgoe's point cuts deeper: policymakers would rather anticipate these problems than keep discovering them after the fact. As autonomous vehicles scale up and spread into more cities and weather conditions, the gap between what the software can handle and what the roads actually present will only grow wider. The San Antonio incident was not a catastrophe—no one was hurt, the car was recovered—but it was a warning. The question now is whether the industry and regulators will treat it as such.
Citações Notáveis
We often see these limits only when something goes wrong— Jack Stilgoe, professor of science and technology policy at University College London
Policymakers would prefer to know about these things in advance rather than discovering them in hindsight— Jack Stilgoe
A Conversa do Hearth Outra perspectiva sobre a história
Why does a flooded road trip up autonomous vehicles when human drivers navigate them all the time?
Because the systems rely on visual markers—lane lines, road edges, the texture of pavement—to understand where they are and where they can safely go. Water erases those markers. It also creates reflections and distortions that confuse the cameras and sensors. A human driver has intuition, experience, and can read the landscape in ways the software simply cannot.
So this is a known limitation, or did Waymo discover it by accident?
Waymo discovered it by accident. That's the troubling part. The company presumably tested its vehicles in many conditions, but apparently not thoroughly enough in flooded scenarios. Or perhaps they tested in controlled flooding and the real-world version was different.
What does "mitigations already in place" actually mean in practical terms?
It means they've drawn a digital fence around flood-prone areas on their maps and programmed the vehicles to refuse to enter those zones. It's a workaround, not a fix. The cars still can't handle flooding—they just won't go there anymore.
How long will the software fix take?
The filing doesn't say. Waymo has suspended service in San Antonio, which suggests weeks at minimum, possibly longer. They need to develop the fix, test it extensively, and make sure it doesn't break something else.
Is this a sign that autonomous vehicles aren't ready for real cities?
Not necessarily. It's a sign that the technology has real limits that need to be mapped and managed. The question is whether companies and regulators will do that work proactively or keep learning through incidents.