Self-driving Cars Can't Handle Water? Waymo Recalls and Halts Robotaxi Services in Multiple Cities

< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400">

2026-05-25 16:48:14 Source: QuantumBit

Tesla Learns to "Avoid Police"

Jessica, driving passenger-side

Smart Car Reference | Official Account AI4Auto

It's 2026, and Waymo’s self-driving cars still can’t handle wet roads?

Recently, with frequent heavy rains, Waymo has had another incident. A self-driving car entered a severely flooded road that was impassable, causing the vehicle to get stuck.

Moreover, within just one month, this is already the second time Waymo has suffered an operational accident due to flooding:

In April, Waymo was reported to have driven into a flooded lane. Due to the water level being much higher than what the system had predicted, the vehicle was swept away by the current.

Waymo later admitted that its software had flaws and conducted a massive recall, using geofencing to restrict vehicle movement at specific times and locations.

However, it seems that these temporary measures are only a band-aid solution, and the issue of flooding remains a persistent problem for this autonomous vehicle company.

Waymo Fails in Water, Suspends Operations in Multiple Cities

The incident occurred recently when a strong rainstorm hit Atlanta, USA.

An empty Waymo self-driving car drove into a severely flooded area that was impassable, getting stuck for about an hour before being towed away.

Waymo officially explained that the storm came suddenly, and some roads began to flood even before the local meteorological bureau issued a flash flood warning.

In other words, Waymo’s fleet still relies on official weather alerts to decide whether to avoid deep water areas.

This explanation actually reveals a key dependency in Waymo’s system design: external information input.

Even more concerning is that this is the second time in recent months that Waymo has experienced an operational accident due to flooding.

On April 20th, in San Antonio, USA, another Waymo self-driving car became stranded in extreme weather — fortunately, there were no passengers in the car at the time.

Investigation documents showed that although the system detected road flooding, the vehicle continued moving slowly and was eventually swept into a stream.

This incident quickly drew attention from traffic authorities.

In mid-May, Waymo submitted a voluntary recall request to the National Highway Traffic Safety Administration (NHTSA), involving 3,791 vehicles equipped with fifth and sixth-generation autonomous driving systems.

The reason given for the recall was software defects that could cause vehicles to continue driving slowly into impassable flooded areas after detecting water.

At the same time, the company also acknowledged that they have not yet fully developed a final solution to identify and avoid flooded zones.

Therefore, Waymo can only temporarily push software updates via OTA, using geofencing to limit vehicle movement in specific times and regions.

But now, such temporary fixes seem more like a band-aid approach.

After the new incident, Waymo decided to suspend operations in Atlanta, Austin, Dallas, and Houston, and also halt highway services in San Francisco, Los Angeles, Phoenix, and Miami.

The company explained that this was to update the software and improve performance around construction zones and flooded roads. Operations will resume gradually once the updates are complete.

Temporarily halting paid services that were already running is not common for Waymo. Although this suspension is partial and limited to certain cities and routes, it has a real impact on order volume and user experience.

This may indicate that the flooding issue cannot be resolved quickly through remote patches and requires all vehicles to stop until upgrades are completed.

So why does Waymo keep failing in puddles? From a technical perspective, there are two core reasons:

First, there are challenges in perception.

Water surfaces have unique physical properties. When LiDAR pulses hit the surface, part of the light is absorbed, and part undergoes specular reflection, leading to sparse or distorted point cloud data.

Cameras also struggle in heavy rain, as visibility is limited and surface reflections can cause overexposure or misidentification.

Waymo’s multi-sensor fusion algorithm needs to integrate real-time data on water depth, flow speed, and vehicle passability, placing high demands on the system’s perception capabilities.

Secondly, Waymo’s system overly depends on predefined rules.

Waymo’s autonomous driving system is essentially a rule-based system, trained on massive amounts of driving data.

Engineers first define thousands of rules—such as “drive within speed limits” or “brake if there’s an obstacle ahead”—and then match these rules against real-world scenarios.

This works well in 90% of cases, but the remaining 10%—the so-called long-tail scenarios—are where things often go wrong.

Flooded roads are a typical example. At present, no single rule can cover all possible situations:

How deep is too deep? How fast must the current be to be dangerous? Is the road surface flat or full of potholes? These variables can’t be covered by fixed rules.

Under normal conditions, when relying on pre-set rules and high-definition maps, self-driving cars can avoid known high-risk flooded areas by updating map data or setting operational restrictions.

But during sudden extreme weather events—like the Atlanta incident—the rule-based and externally informed defense mechanisms may fail.

Thus, how to give the system common sense judgment similar to human drivers is the real challenge engineers at Waymo currently face.

Meanwhile, Tesla, which is also pushing forward with Robotaxi operations, has quietly become smarter.

"Experienced Driver" Tesla Learns to Recognize Police Cars

When it comes to mimicking the subtle, unwritten rules of human driving, Tesla FSD is getting better and better.

Recently, some owners noticed that when FSD detects a police car on the median of a highway, it adjusts its behavior accordingly:

The car was originally speeding at 77 mph (speed limit 70 mph).

Upon spotting the police car, the system promptly slows down and changes lanes, seamlessly blending into slower traffic and avoiding drawing attention from officers.

Hmm... That driving style...

~~Looks just like a seasoned human driver.~~

This kind of response to emergency vehicles has been mentioned in past Tesla software updates.

In October 2025, Tesla explicitly added the capability to handle emergency vehicles (such as police cars, fire trucks, and ambulances) pulling over or yielding in its FSD update.

In a recent FSD V14.3.3 update, Tesla introduced a specialized recognition module for emergency vehicles like ambulances and school buses.

According to feedback from some test users, when the system detects such vehicles, it initiates avoidance strategies about 300 meters in advance—

At a city speed of 50 km/h, 300 meters takes roughly 20 seconds, giving the vehicle enough time to change lanes, slow down, or pull over.

Comparing both approaches brings us back to the technological rivalry between Waymo and Tesla, fundamentally reflecting two different engineering philosophies.

Waymo represents a “top-down” approach, believing that autonomous driving is a system engineering problem that can be decomposed, designed, and verified. It aims directly for L4/L5 full autonomy.

Engineers define modules such as perception, localization, prediction, planning, and control. Each has clear functions and performance metrics. Through careful design and rigorous testing, they integrate into a reliable system.

This approach relies heavily on high-definition maps and multi-sensors, essentially scanning roads beforehand, ensuring nothing is missed through multiple sensors, and constraining the vehicle with countless rules.

Its advantage lies in controllability: system behavior is predictable and explainable. Within map-covered areas, vehicle positioning is highly accurate, and safety is guaranteed.

When accidents occur, engineers can trace issues to specific modules for targeted fixes.

However, the cost of a top-down approach is complexity. Real-world driving scenarios are nearly infinite. When faced with unanticipated situations, the system might behave rigidly or even fail.

Tesla, on the other hand, follows a progressive path, starting from L2, collecting data from millions of user vehicles, and iteratively improving through end-to-end neural networks to achieve L4/L5 autonomy.

Its core relies on cameras capturing surrounding environments, processing visual information through neural networks, and making driving decisions.

This method excels in adaptability and scalability. Since it doesn’t rely on predefined rules, it can handle previously unseen scenarios. As data accumulates, system capabilities grow naturally.

Additionally, during training, the system sees countless scenarios involving emergency vehicles and human responses, making its reactions more “human-like.”

However, purely vision-based systems have physical limitations. Cameras perform poorly in heavy rain, fog, or strong backlighting.

More importantly, neural networks are “black boxes”—engineers may struggle to fully understand why a specific decision was made, and it’s hard to guarantee safe operation under all edge cases.

Thus, Tesla chooses OTA updates to allow the system to continuously learn from new driving data.

Both paths have their pros and cons. Interestingly, both are subtly converging toward each other.

Waymo, in its promotion of the sixth-generation Driver, mentions “lightweight, powerful machine learning models,” indicating they’re introducing more learning-based methods and reducing reliance on HD maps rather than sticking to strict rules.

Tesla, meanwhile, is focusing more on safety redundancy. The FSD V14 version includes more cross-validation mechanisms, no longer relying solely on one neural network.

This sounds somewhat like Waymo’s multi-sensor fusion approach—but with pure visual sensors.

Ultimately, the solution for autonomous driving may not be a pure path from either side, but a combination: using maps for prior knowledge, neural networks for real-time judgments, and rules to ensure basic safety.

The party that achieves better cost efficiency may gain the upper hand sooner.

One More Thing

Another recent news about Tesla that has drawn attention is the latest progress on its FSD Supervised Version, now open for use in China and 10 other countries and regions.

Several job postings related to autonomous driving testing released by Tesla China further confirm the authenticity of this news.

But when FSD arrives in China, it can only call itself “Assisted Driving”...

Perhaps Tesla’s driving technology peak will soon be visible on domestic competition stages.

However, a vice president of XPeng said that among automakers, only XPeng might welcome FSD into China — “the reasons are obvious to everyone.”

What do you think?

© All rights reserved. No reproduction or usage in any form without permission. Violators will be prosecuted.