Zoox recalls 105 robotaxis after one entered smoke at an active fire scene

Zoox is fixing the fault with a software update

According to documents filed with the US National Highway Traffic Safety Administration (NHTSA), the recall covers 105 Zoox-owned vehicles. None needs to visit a service centre, because the defect can be remedied with a software update.

Zoox began installing the software version containing the fault in its robotaxis on 23 April 2026. The corrected version was deployed to all 105 affected vehicles on 15 July, one day before the company submitted the formal recall notice.

This recall therefore involves no replacement of a mechanical component. Zoox is changing the decision-making logic of its automated driving system so that the vehicle responds more cautiously when poor visibility coincides with an active emergency scene.

The robotaxi drove into dense smoke

The problem emerged on 20 June, when an unoccupied Zoox vehicle approached the scene of a fire. Dense smoke had reduced visibility, and emergency crews had not yet cordoned off the area with traffic cones.

The robotaxi entered the smoke, then began an avoidance manoeuvre, braked sharply and came to a stop. A remote-assistance operator directed the system to reverse, which the vehicle then did autonomously. Only afterwards did emergency crews use cones to close two of the three traffic lanes.

No one was injured. Zoox identified only one such incident in its analysis, but treated all 105 vehicles running the same software version as potentially affected.

The remote-assistance operator was not driving the vehicle directly. The operator issued a manoeuvre instruction, while the robotaxi remained in control of its steering, braking and propulsion. This kind of support can help a vehicle extricate itself from a difficult situation, but it is no substitute for a human driver at the scene with a complete view of the emergency response.

The weakness may have been in how the scene was interpreted

Zoox’s robotaxi uses visible-light cameras, infrared cameras, lidar and radar to monitor its surroundings. It has no steering wheel or pedals, and its symmetrical design allows it to change direction without turning around.

The official filing does not identify a specific sensor or algorithm as the source of the error. There is therefore no basis for claiming that the robotaxi failed to detect the smoke. A more plausible explanation is that the automated driving system did not interpret the combined effects of reduced visibility and an active emergency response cautiously enough.

Visible-light cameras become less effective in smoke as image contrast deteriorates. Airborne particles can also degrade a lidar point cloud. Radar generally copes better with smoke, but on its own it does not provide enough detail for the system to recognise the wider scene as an emergency response.

The software must therefore fuse data from several sensors into a reliable model of the surroundings. It needs to assess not only whether an obstacle lies ahead, but also whether firefighters may be moving through the area, hoses may be stretched across the road or an emergency vehicle may be approaching.

Emergency scenes remain a wider challenge for robotaxis

The Zoox incident is not an isolated case within the industry. In July 2026, NHTSA warned automated-vehicle developers about incidents in which driverless vehicles entered fire or police response zones, obstructed emergency vehicles or failed to respond correctly to flashing lights, flares, smoke and traffic cones.

The agency does not regard these situations as rare edge cases. They occur often enough in urban traffic that developers are expected to account for them when designing system behaviour.

Following ordinary road signs and lane markings is not enough. In an emergency, a robotaxi must be able to deviate from its planned path, yield to emergency vehicles and stop somewhere that does not obstruct rescue work.

Zoox has updated its software before

This is not Zoox’s first software-related recall. In late 2025, the company updated 332 automated driving systems because some robotaxis could drift across the centre line into oncoming traffic near junctions, or stop with part of the vehicle protruding into the opposing lane.

The two faults are not directly related, but they illustrate the iterative nature of robotaxi development. A company gathers data from real-world driving, identifies a rare pattern of behaviour and changes the system’s decision-making in the next software release.

In a conventional car, a recall often involves a mechanical component, wiring or an airbag. In a robotaxi, the defect may lie in a decision rule that works correctly in ordinary traffic but selects the wrong manoeuvre in an unusual situation.