The Uber self-driving program died quietly. A

Uber’s self-driving ambitions are officially dead. The company sold its autonomous vehicle division to Aurora Innovation in a deal that values the unit at around $4 billion, which sounds like a lot until you remember that Uber poured over $2.5 billion into it.

The story ends not with a crash but with a quiet transaction.

But the story really ended in Tempe, Arizona, in March 2018.

The gap

An Uber test vehicle hit and killed a pedestrian named Elaine Herzberg as she crossed a road at night with her bicycle. The NTSB investigation found that the car’s system had detected her 6 seconds before impact but classified and reclassified her multiple times, cycling between “vehicle,” “bicycle,” and “other.” It never resolved the classification in time to brake.

The safety driver was watching a video on her phone.

A woman died because a machine couldn’t decide what it was looking at, and the human backup was distracted.

I’ve been thinking about this for four years now. The technology worked, in a narrow sense. The car detected an object. The LIDAR returned the points. The system saw something was there. It just couldn’t understand what it was fast enough to act on that understanding.

And that gap, between detection and comprehension, between “the sensor sees it” and “the system understands what to do about it,” is the gap that kills people. It’s also the gap that separates “a computer can drive” from “society trusts a computer to drive.”

Those are different problems. The first is engineering. The second is culture, law, psychology, and grief. The first might take a decade. The second might take a generation.

What Uber got wrong

I think Uber treated self-driving as a feature. A thing to add to the ride-hailing app. A competitive advantage, not a moral obligation.

Waymo treated it differently. Years of mapping. Thousands of test miles. A geographic approach that starts with a small area and expands slowly. Waymo is cautious in a way that isn’t exciting but might be appropriate when the downside is killing someone.

Uber moved fast. In self-driving, moving fast has a different cost than in software. When your app crashes, a user is annoyed. When your car crashes, a person is dead.

The legacy

Aurora now has Uber’s talent, data, and patents. They’re working on autonomous trucks, which is a smarter starting point. Highways are more predictable than city streets. Fewer pedestrians. Fewer cyclists. Fewer situations where a machine needs to classify an ambiguous shape at night in a fraction of a second.

But I keep coming back to Elaine Herzberg. The first pedestrian killed by an autonomous vehicle. Her death didn’t stop the industry. It slowed it. Made it more careful, maybe. But slowed isn’t the same as stopped, and careful isn’t the same as safe.

The gap between “can drive” and “trusted to drive” is still there. It’s closing. Slowly. Measured in trust built and trust broken.

Uber broke trust. The technology will recover. The question is whether the memory will last long enough to keep the industry honest.

Related thinking: