Waymo self-driving cars repeatedly drive into flooded areas, Robotaxi suspends operations in four cities.

Waymo has initiated recalls for 3,791 of its robotaxis and suspended service in four cities, including Atlanta, due to vehicles being unable to recognize flooded road sections. This follows a previous incident in San Antonio in April where a vehicle drove into a stream, and on May 21, Atlanta experienced the same issue.
(Background: Waymo blocked ambulances rushing to shooting scenes; emergency responders: the situation is worsening)
(Additional context: Tesla's robotaxi pilot in Texas, autonomous taxis challenging Waymo and Uber's positions)

Table of Contents

Toggle

• From San Antonio to Atlanta: The same mistake replayed within two months
• Boundary cases that software updates can't prevent: Structural gaps in AI training data
• NHTSA intervention and series investigations: Waymo faces more than just technical issues

Autonomous vehicle technology in Silicon Valley labs handles complex decisions like identifying pedestrians within milliseconds and predicting behaviors at intersections. But after a heavy rain in Texas, Waymo's robotaxi overturned over a more basic problem.

From San Antonio to Atlanta: The same mistake replayed within two months

The incident began on April 20. A Waymo robotaxi in San Antonio, Texas, drove into a flooded area and was ultimately swept into a stream. This accident drew the attention of the NHTSA (National Highway Traffic Safety Administration) and forced Waymo to confront structural issues in its fleet's software.

On May 12, Waymo issued a recall for 3,791 of its robotaxis, but this was not a traditional recall where owners bring vehicles back to the factory for repairs. Instead, it was an over-the-air (OTA) update—remote software push via the internet—allowing vehicles to be updated without visiting a service center.

Waymo also admitted that this update was only a temporary measure: it limited vehicle operation in "high flood risk areas and specific times," but a "final solution" had not yet been completed.

Before long, on May 21, Atlanta experienced another incident. A driverless Waymo robotaxi entered a flooded street, got stuck for nearly an hour, and had to be towed out. The rainfall that day was so heavy that street flooding occurred, yet the National Weather Service (NWS) had not even issued a flood warning.

This indicates that the "high-risk time restrictions" set by the OTA update, which were supposed to prevent vehicles from operating during dangerous conditions, completely failed during this rapidly developing regional storm.

Following the incident, Waymo announced service suspensions in Atlanta, San Antonio, Dallas, and Houston.

Boundary cases that software updates can't prevent: Structural gaps in AI training data

To understand why Waymo repeatedly fails on the same issue, we must return to the essence of autonomous vehicle technology.

Robotaxi perception systems rely heavily on large labeled datasets for training: which surfaces are drivable, where obstacles are, and the lane boundaries. In the distribution of training data, 99% of city streets are "dry, normal pavement." Flooded areas are edge cases in machine learning language—rare scenarios not sufficiently covered by training data.

The problem is that when a robotaxi's sensors (cameras + LiDAR) scan a flooded area, the reflection characteristics of water are very similar to wet asphalt. Without explicit training to recognize "this scene is dangerous," the model isn't making a "mistake" in judgment; it was simply not designed to brake in such scenarios.

Waymo's OTA update attempts to address this by using "geofencing + time restrictions": prohibiting or limiting vehicle operation in known flood-prone areas during certain times. The logic itself is sound, but it relies on a premise: that alerts come faster than real-world events. The Atlanta case on May 21 proves that a storm can develop faster than the NWS warning or the geofence trigger.

Waymo's commercial operations cover 11 U.S. cities, including San Francisco, Los Angeles, Phoenix, Austin, and Miami. The suspension in four cities means over a third of its operational footprint is paused. This is not just a regional issue but a systemic risk shared across the fleet due to reliance on a single software logic.

NHTSA intervention and series investigations: Waymo faces more than just technical problems

Flooding incidents are not the only regulatory pressures Waymo faces.

Beyond the flooded area recall, the NHTSA and NTSB (National Transportation Safety Board) are jointly investigating another ongoing issue: Waymo robotaxis repeatedly illegally overtaking stopped school buses. U.S. state regulations require vehicles to stop and wait beside school buses with extended stop arms, but Waymo vehicles in Austin have repeatedly violated this, and even after corrective measures, the problem persists. On May 15, NHTSA issued a second data request.

At the same time, on January 23 of this year, a Waymo robotaxi in Santa Monica, Los Angeles, collided at about 10 km/h with a child, causing minor injuries. The incident occurred near a school.

NHTSA stated that "actions will be taken as necessary" regarding the flooding incidents. The implication is that regulators are reserving the option for further intervention—including stricter recall mandates, suspending commercial licenses in specific cities, or delaying expansion plans into new cities.

The promise of autonomous vehicles has never been "better than human drivers in all scenarios," but rather "statistically safer and more reliable." However, statistical superiority cannot hide the failure of boundary cases.

Every time a robotaxi drives into water, blocks an ambulance, or overtakes a school bus, it reminds us of one fundamental truth: large-scale commercial deployment inherently exposes the gaps in training data of the real world.