Two Waymo cars hit the same towed truck minutes apart

Two collisions with the same towed pickup

On December 11, 2023, a Waymo robotaxi driving itself through Phoenix came up behind a pickup truck under tow. The pickup was being hauled backwards and angled across a shared center turn lane, so it did not track in line behind the tow truck the way a trailer would. The Waymo's front left struck it. The tow truck did not stop. A few minutes later a second Waymo, driving the same stretch, made contact with the same off-angle pickup, still strung across more than one lane. Both were low-speed, same-direction contacts, and neither hurt anyone.¹

Waymo's own account, published February 13, 2024, called the pickup "improperly towed" and "persistently angled across a center turn lane," and named the cause in plain words: the car "incorrectly predicted the future motion of the towed vehicle."² The company treated the filing as voluntary and pushed a software update to its fleet, finished by mid-January. It reported the collisions to the National Highway Traffic Safety Administration as recall 24E-013, covering 444 vehicles.¹ It was the first recall in Waymo's history, and the wire press carried it: Reuters reported the same count and the same cause, an inaccurate prediction of where the towed truck would move.³

1. Dec 11, 2023

   Two Waymo cars contact the towed pickup¹

   In Phoenix, minutes apart, two driverless Waymos strike the same backwards, angled pickup. No injuries.

2. Dec 20, 2023

   Software update begins deploying

   Waymo starts pushing the corrected release across the affected fleet.

3. Jan 12, 2024

   Update finished on all affected vehicles

   The fix is in the field a month before the paperwork is filed.

4. Feb 13, 2024

   Recall 24E-013 filed with NHTSA, 444 vehicles¹

   The crash was in December. The recall came two months later, in February.

Perception, then prediction

A self-driving car does three jobs in sequence before it commits to a move. It perceives the scene, detecting and tracking the objects around it. It predicts where each moving object will go next. Then it plans its own path around those predictions. In the research literature these are separate modules, built and tuned on their own, with the output of one handed forward to the next.⁴

The order matters, because an error in an early step travels forward: a wrong prediction hands the planner a future that never happens.⁴ At the towed pickup, perception was not the problem. The car detected a vehicle, and the filing does not fault what its sensors returned. The next step broke. The system predicted where a truck would go, and the truck it was watching did not move like a truck. Hauled backwards and angled, it sat where the model expected motion and moved where the model expected stillness, so the predicted path and the real one came apart.¹

Why a rare shape breaks a learned guess

A prediction model learns from examples. It is shown enormous numbers of real scenes and the paths the objects in them actually took. From those it absorbs the regularities: a car in a lane keeps going down the lane, a vehicle under tow tracks in line behind the one pulling it. What it learns is the distribution of situations it was shown. Give it a situation far outside that distribution and its guess is only as good as the nearest thing it saw in training, which may be nothing like the thing in front of it. That failure has a name, and this paper's lesson on distribution shift covers why it is the normal condition of a deployed model rather than a rare edge case.

In autonomous driving it is unavoidable. The "Shifts" benchmark, built partly on real self-driving fleet data, states that distributional shift is ubiquitous in the domain. It frames motion prediction as exactly the task that broke here: predicting the distribution over the possible future states of the agents around the car.⁵ A pickup towed backwards across a turn lane sits far out on the tail of that distribution. Waymo's cars had surely handled unusual vehicles before without trouble, but those clean passes were never proof that this particular configuration was safe. The December collisions showed it was not, and showed it twice, the same way, within minutes.

The pole was a different failure

A recall is the mechanism by which a manufacturer both fixes a defect and puts it on the public record. For an ordinary car it ends with letters to owners and a trip to the dealer. Waymo's ended differently: the company owns every vehicle in its fleet and has never sold one, so there was no owner to notify, and the remedy was a software release already deployed before the report was filed.¹

The second recall, 24E-049, filed on June 10, 2024, is the useful counterpart to the first. On May 21, 2024, a Waymo in Phoenix collided with a wooden utility pole in an alley while pulling over at low speed. No one was aboard and no one was hurt; the car was damaged.⁶ The filing covered 672 vehicles and named two fixes rather than one, a software update and a map update.⁶ TechCrunch, reporting the recall, identified the cars as Jaguar I-Pace robotaxis and put the collision speed at about 8 miles per hour, a figure the filing itself does not give.⁷

The pole was not a repeat of the towed-truck bug. This time the failure sat one step earlier, in perception and mapping. The filing says the system assigned the pole a low damage score, and the map lacked a hard road edge between the pole and the drivable surface. Nothing marked the pole as a fixed object to steer well clear of.⁶ Same class of problem, different broken part. The towed truck defeated the step that guesses motion; the pole defeated the step that judges what an object is and where the road ends. Read side by side, the two recalls are a clean picture of a pipeline. Two different stages, each failing on a configuration the training and the map did not cover.

Where does the weakness live now? NHTSA opened a preliminary evaluation of Waymo's fifth-generation driving system in May 2024, logging 22 incidents at the outset, single-party crashes and apparent traffic-law violations across the same fleet.⁸ The two specific bugs were patched. The general weakness behind them belongs to every deployed self-driving system. Each one carries a residual of road situations too rare to have appeared in its training, and a recall is what happens when one of them surfaces as a crash.