When robotaxis can't read the road: Waymo's flood recall and the limits of self-driving

For a technology sold on the promise of being safer than human drivers, autonomous vehicles have a curiously human problem: they're great at the routine and shaky at the unexpected. That tension was on full display in October when Alphabet-owned Waymo issued a voluntary recall covering 3,791 of its robotaxis — essentially its entire US fleet — after a software glitch allowed some cars to drive into standing floodwater in Texas.

The recall, filed with America's National Highway Traffic Safety Administration, is small in human terms (no injuries were reported) but loud in what it says about the state of the art. Waymo is widely seen as the most cautious, most mature operator in the driverless space. If its cars can roll confidently into a flooded street, it's worth asking what else the sensors and software might be quietly misreading — and what that means as the technology inches towards Australian roads.

What actually happened

According to filings reported by CNBC and The Register, Waymo identified two incidents in the Houston and Austin areas where its Jaguar I-PACE robotaxis drove into standing water deep enough to be a hazard. One vehicle was carrying a passenger at the time. Waymo told regulators the software failed to recognise the flooded roadway as something it should avoid, and the company has since rolled out an over-the-air update to correct the behaviour.

The fix, as described in the recall notice, teaches the system to better detect flooded roads and to plan routes that steer clear of them. It also widens the geofenced exclusion zones the company uses during severe weather. As Traffic Technology Today reported, this is the kind of remedy that doesn't require a workshop visit — every affected car can be patched remotely, which is one of the genuine advantages of a centrally managed robotaxi fleet over a million privately owned cars.

Still, the recall is the latest in a string for Waymo. The company has issued several software-related recalls over the past two years covering issues from collisions with stationary objects to unexpected behaviour around emergency scenes. None has involved a fatality, but each is a reminder that 'driverless' is still very much a work in progress.

Why floodwater is genuinely hard for a robot

It's tempting to read the Texas incidents as a simple oversight — surely a car that can identify a cyclist at night can spot a flooded street? But floodwater is one of the trickier edge cases in machine perception, for reasons that go to the heart of how these vehicles work.

Waymo's cars stitch together a picture of the world from cameras, radar and lidar. Lidar fires laser pulses and measures what bounces back. A smooth sheet of standing water can act like a mirror, reflecting beams away from the sensor or producing returns that look more like wet asphthan a hazard. Cameras, meanwhile, depend on visual cues — and a flooded road on an overcast day can look uncannily like a normal wet road, particularly if leaves, debris or reflections obscure the kerb line.

Then there's the map problem. Waymo relies on extraordinarily detailed pre-built maps of every street it operates on. Those maps tell the car where lanes, signs and intersections sit, but they don't update in real time for a flash flood. If the perception stack doesn't flag the water as anomalous, the planner has no reason to deviate from a route the map says is perfectly drivable.

Humans aren't infallible here either — drowning in submerged cars is a leading cause of flood deaths in places like Queensland, and the 'if it's flooded, forget it' campaign exists precisely because drivers underestimate water depth. But humans bring context a robot lacks: we notice the SES sign at the corner, the neighbour wading down the footpath, the news alert that pinged an hour ago.

What the recall reveals about the broader technology

The flood incident is a useful lens because it exposes three structural limitations that apply to every autonomous vehicle programme, not just Waymo's.

1. The long tail never ends

Self-driving systems are trained on enormous volumes of driving data, but the situations that cause crashes are usually the rare ones — the so-called long tail. A flooded suburban street in Texas in October is rare. So is a kangaroo bounding across the Hume at dusk, a road train kicking up red dust on the Stuart Highway, or a cyclone-damaged signal in Cairns. Each rare event has to be encountered, labelled, learned and validated. The list is effectively infinite.

2. 'Driverless' is geofenced for a reason

Waymo doesn't operate everywhere. Its commercial services are confined to mapped patches of cities like Phoenix, San Francisco, Los Angeles and Austin, with weather and route restrictions baked in. The Texas recall has reportedly pushed the company to expand those weather-related exclusion zones further. That's sensible engineering, but it also underlines that today's robotaxis are not general-purpose drivers. They're highly specialised systems that work in carefully chosen conditions.

3. Software recalls are the new normal

The fact that nearly 3,800 vehicles can be 'recalled' and fixed with a download is genuinely new. Traditional recalls involve workshops, parts and weeks of downtime. But it also means the bar for what counts as a recall has shifted. Every time a self-driving company learns something new about an edge case, that learning becomes a regulatory event. Expect many more of these notices — not fewer — as fleets scale up.

What it means for Australia

Waymo doesn't yet operate in Australia, and there's no immediate prospect of robotaxis appearing on George Street or Chapel Street. But the Texas episode is relevant for anyone watching the local debate about autonomous vehicles, which has been gathering pace through the National Transport Commission's work on driverless vehicle laws.

Australia is a particularly demanding environment for this technology. Flooding is a routine hazard across northern NSW, south-east Queensland and parts of Victoria — and it has killed more Australians in recent decades than bushfires. Unsealed roads, livestock on the verge, sudden weather changes, and the patchy mobile coverage that many remote highways still endure all push against the assumptions baked into US-developed systems. A car trained in Phoenix has no intuition for a flooded causeway near Lismore.

None of this means autonomous vehicles won't eventually arrive here. Trials of driverless shuttles have already run in Perth, Sydney and Adelaide, and freight operators are watching the technology closely. But the Waymo recall is a useful corrective to the marketing. The most experienced robotaxi operator on the planet, with the most data and arguably the most conservative deployment strategy, still had cars driving into floodwater in 2024.

The honest verdict

Waymo deserves some credit here. It self-reported, it pushed a fix, and there were no injuries — which is more than can be said for many human-caused flood incidents. The system worked roughly the way a regulator would want it to.

But the recall is also a reality check. The dream of fully autonomous driving has been five years away for about fifteen years now, and incidents like this are a reminder why. The hard part of building a driverless car was never the motorway cruise; it was the rare, weird, locally specific moments that human drivers handle with a glance and a shrug. Until a robotaxi can confidently look at a flooded Texas street — or a flooded Lismore one — and decide to turn around, 'driverless' will keep coming with an asterisk.

Related on Bleen

Sources