With 672 driverless cars recalled after a Phoenix taxi pressed its bumper into a utility pole at walking pace, a second software flaw pulled in 1,212 more

The alley was quiet, the vehicle moving slowly, and the wooden pole had stood there for years without incident.

Then a driverless taxi in Phoenix pressed its bumper right into it.

Nobody was riding inside.

Nobody was hurt.

And the pole that barely moved ended up touching off one of the more revealing sequences in the short history of autonomous cars.

What the software could not see

A driverless vehicle navigates by fusing what several sensor systems tell it at once: a spinning laser that builds a three dimensional map of the surroundings, cameras reading lane markings and signs, radar bouncing signals off nearby objects, and a precisely pre mapped version of every street the car is allowed to drive. When all four agree, the car moves with confidence.

The problem is that a narrow, vertical object can slip between those readings in ways that wider obstacles almost never do. A wooden utility pole is about 7 to 12 inches across at the base and rises straight up with no horizontal edge to catch on. Radar tends to pass through it or scatter. Lidar, the spinning laser, can miss a very thin target if it sits between two sweep angles.

Cameras can see the pole, but if the pre mapped version placed it a few inches to one side, the planning software may decide there is a gap large enough to pass through. All three of those slip points landed together in that Phoenix alley, and the bumper found the pole before any alert reached the braking system. The gap that mattered was not in the hardware. It was in the code that decides what a pole shaped object actually means for the car’s path.

The Phoenix alley, at walking pace

On May 21, 2024, a fully driverless vehicle was making a low speed pullover maneuver in an alley in Phoenix, heading to pick up a passenger, when its front end contacted a wooden utility pole. The vehicle was moving at roughly walking pace. No passengers were aboard, no other vehicle was involved, and the pole stayed upright.

The incident was reported to federal safety regulators, as every autonomous vehicle collision must be. Investigators reviewed the sensor logs and the software decisions made in the seconds before contact. They found that the software had assigned a low damage score to the pole, and that the maps did not properly account for the road edge in the alleyway, leaving the vehicle unable to judge the narrow object’s true hazard during a slow maneuver.

Because the flaw was not unique to the one car in the alley, every vehicle carrying the same software version became part of the answer.

672 vehicles, then 1,212 more

No injuries occurred in the Phoenix incident, but it led to a voluntary recall of 672 driverless cars, the entire fleet at the time. A software update improved the vehicles’ ability to detect pole shaped objects, along with mapping improvements to better define road boundaries. The fix traveled invisibly, pushed over the air to every car overnight.

Then, roughly eleven months later, a second recall covered 1,212 autonomous vehicles for a software issue that could lead to minor collisions with roadside barriers such as gates and chains. The mechanism was closely related: another category of narrow, low profile obstacle that sat at the edge of what the sensor suite reliably resolved. A gate arm, like a utility pole, is thin enough to slip between lidar sweeps and close enough to the ground to confuse the mapping layer.

Together the two actions covered nearly 1,900 vehicles. Neither recall involved a human injury, and both were resolved entirely through software, with no car returned to a factory floor.

Over the air software fixes represent a category of remedy that simply did not exist for conventional vehicle recalls. As the lead researcher on autonomous vehicle safety at one federally funded transportation institute put it, a mechanical defect cannot be patched remotely, but a software flaw can reach every car in a fleet before sunrise.

The part of the road the map has never visited

What the Phoenix pole made plain is that the hardest territory for a driverless car is not the highway at speed but the slow, unpredictable geometry of alleys, driveways and parking structures. In the same period, federal safety regulators opened investigations into multiple autonomous vehicle operators over incidents where vehicles were filmed driving erratically, including a 2025 Austin pilot in which robotaxis were recorded driving down the wrong side of the road and braking suddenly.

The alley in Phoenix added a different problem: a place where the map and reality disagreed by just a few inches, and a few inches was enough. Alleys rarely get the mapping attention that main roads do. Pre mapped surveys are refreshed by a dedicated vehicle making deliberate passes, but a residential alley may be re surveyed infrequently, and a pole can be replaced, shifted or added between visits.

So the car may be operating from a map that is accurate in every large detail and wrong about a single object standing 8 inches from where the software expects it. That mismatch is nearly invisible until a bumper finds it. The broader question of what happens when a software patch falls short of a mechanical fix has surfaced in very different vehicles too.

What a pole and a gate have in common

The two recalls share a geometry: a thin, mostly vertical structure whose base sits close to the ground and whose width is narrow enough to slip through sensor resolution at low speed. Utility poles, gate arms and chains are all members of this family, and they appear most often exactly where driverless cars move most slowly, near buildings, in parking structures, at property edges.

Slow speed is the condition that saves people from injury when a miss occurs, but it is also the condition where sensors have the least time to build a clean picture before a decision must be made. No injuries were reported across either recall, and that record is genuinely significant.

Software patches can travel to 1,212 vehicles in a single night, something a mechanical recall simply cannot do. The open question is whether every member of the thin object family has now been found, or whether an alley somewhere holds a variation the updated code has not yet met. For a fiberglass hull or a steel bumper, the answer comes from a physical test in a controlled yard; for software, it comes from the next alley the car has never driven before.