April 20, 2026 was a rainy day in San Antonio. Sections of roadway flooded. A Waymo robotaxi, empty and repositioning itself across the city, approached one of those flooded sections on a 40 mph road.

The car's sensors registered the standing water. The autonomous driving system processed it. The vehicle slowed down.

Then it drove in.

The car floated off the road surface and came to rest in Salado Creek. It had to be recovered from the waterway days later. No passengers were aboard; no one was hurt. The only immediate casualty was the vehicle - and, depending on your perspective, a modest amount of confidence in the state of autonomous vehicle software.

What the Software Was Supposed to Do

Modern automated driving systems are not simple rule-followers. They're complex perception and decision-making pipelines: cameras, LiDAR, radar, and sensor fusion all feeding into models that classify the road environment and plan a path forward. Detecting hazards is considered a core capability. Waymo, which has been developing autonomous vehicles since the Google Self-Driving Car Project started in 2009, has invested heavily in exactly this kind of environmental awareness.

In theory, the system should detect standing water, classify it as an untraversable obstacle, and either stop or reroute. That's not an exotic requirement - it's the kind of behavior you'd expect from a reasonably cautious human driver.

In practice, the system did the first part (detect) and then got confused about the second part (stop). According to Waymo's own voluntary recall filing with NHTSA on April 30, the problem was specific to higher-speed roadways: the ADS "may slow but not stop in response to detecting a potentially untraversable flooded lane." It would see the water. It would reduce speed. It would not stop.

This is a distinction that matters a great deal when there's a creek on the other side of the floodwater.

Recall at Scale

Waymo filed the recall voluntarily with NHTSA ten days after the incident. The recall covered 3,791 vehicles - the entire fleet of robotaxis running the company's 5th and 6th generation Automated Driving Systems. That means every Waymo operating in Phoenix, San Francisco, Los Angeles, Austin, San Antonio, and Atlanta had the same underlying software flaw.

The good news, from Waymo's perspective, is that the fix was a software update rather than a physical repair requiring vehicles to be brought to a service center. Waymo pushed an over-the-air update to address the immediate issue and suspended all San Antonio operations while it developed additional software safeguards and refined its extreme weather protocols.

This is the modern version of a car recall: instead of mailing a notice and waiting for owners to schedule a dealership appointment, a robotaxi company can patch the behavior remotely. Quick, clean, and it sidesteps the awkward question of "who do we even mail this to?"

Still, the recall filing is a public record that will live in the NHTSA database forever. It says, plainly, that Waymo's autonomous vehicles had a known software defect that caused them to drive into floodwater.

How Hard Is "Don't Drive Into Water"?

At first glance, this seems like the kind of failure that should be embarrassing for an organization that has been working on autonomous driving for over fifteen years. "Don't drive into visible standing water on a flooded road" feels like it should be near the top of the requirements list, somewhere around "stop at red lights" and "don't hit pedestrians."

But autonomous driving in adverse weather is genuinely difficult, and understanding why the failure happened the way it did requires some context.

Detecting standing water from a moving vehicle is not as simple as it sounds. Water on a road surface can look like wet pavement, which looks like dry pavement with glare, which looks like a shadow, which looks like a road marking. Cameras can struggle to distinguish between a puddle and a reflection. LiDAR works by measuring how long it takes a laser pulse to bounce back; water surfaces scatter laser pulses unpredictably, making depth estimation unreliable. Radar penetrates water but can't distinguish between shallow and deep.

What happened in San Antonio was not that the sensors failed to detect anything. The car did slow down, which means the system identified something unusual ahead. The failure was in what happened next: the system's decision-making apparently classified the situation as something manageable at reduced speed rather than something requiring a full stop. It didn't know the water was deep enough to float a car off the road. It knew something was there, it was cautious about it, and then it drove through anyway.

This is a challenging edge case. The gap between "something unusual on the road" and "this road ends in a creek" is not trivial for a perception system to bridge reliably.

The Physics Were Not Kind

When the vehicle hit the floodwater at reduced speed and floated off the road surface, it demonstrated something any flood safety expert will tell you: moving water is powerful, and vehicles are buoyant. The rule of thumb for humans is that six inches of moving water can knock someone down, and two feet can sweep away a car. A Waymo is several thousand pounds of vehicle, but it's still built to float.

The car ended up in Salado Creek. Waymo had to recover it. The company was tight-lipped about specific damage or recovery costs.

San Antonio was hit with significant spring flooding around this period, which contributed to the conditions. But "it was raining unusually hard" is not a compelling defense for a commercial autonomous vehicle operator. Flooding is a documented weather phenomenon. San Antonio gets flash floods regularly. An ADS deployed for public commercial service in San Antonio needs a plan for San Antonio weather.

Not the First Edge-Case Stumble

This was not the first time a Waymo vehicle exhibited unexpected behavior during edge-case scenarios - though prior incidents tended to involve different failure modes. In October 2025, a routing error sent a vehicle not qualified for freeway operation onto US 101 near the Golden Gate Bridge in San Francisco. The company has also dealt with incidents where vehicles got confused by emergency response scenes, construction zones, and unusual traffic patterns.

These are, broadly, the known hard cases for autonomous vehicles. Not the highway cruise, not the well-mapped suburban street, but the moments when reality diverges from the training distribution. A flooded road on a rainy day in a city prone to flash flooding is precisely the kind of scenario the system needs to handle correctly, because it's exactly the kind of scenario that will occur in the real world with some regularity.

Waymo's Response and the Broader Industry Implication

Waymo handled the recall competently in a procedural sense. They filed with NHTSA promptly, acknowledged the flaw clearly, deployed an interim software fix, and suspended operations in the affected city while developing a permanent solution. The company stated it takes "community feedback" and safety seriously - understated phrasing for an incident involving a car in a creek, but the response mechanics were reasonable.

For the autonomous vehicle industry, this incident lands at an uncomfortable time. AV companies have spent years arguing that their vehicles are safer than human drivers in the aggregate, which is often true on clear days on mapped roads. The harder question - how the systems handle genuine edge cases in adverse weather conditions - remains a live debate. Regulators, insurers, and the public are watching these incidents closely, and a car floating into a creek is not the kind of footage that builds confidence.

The recall affected vehicles in six major U.S. cities. There are a lot of weather events across Phoenix, San Francisco, Los Angeles, Austin, San Antonio, and Atlanta. This flaw apparently existed undetected until it produced a vehicle in a creek, which suggests the test scenarios for adverse weather decision-making were not comprehensive enough.

Over-the-air software fixes are genuinely one of the advantages of modern autonomous vehicles over traditional cars. But they only work if the flaw gets discovered before it causes serious harm. In this case, nobody was in the vehicle, and the creek was survivable. The same software defect, same conditions, with a passenger aboard, would be a different story.

For now, San Antonio residents can note that the Waymo that floated into Salado Creek was eventually recovered from the water. The software update, presumably, does not float.