Thread by @amir

Just out: The truth about Waymo...
go.theinformation.com/4596b7125469ea…

...in suburban Phoenix starts here, at a T intersection at East Fairview St. and South 56 St., the closest one to Waymo’s vehicle depot in Chandler. Waymo’s vans sometimes have trouble finding a gap in traffic to turn left, frustrating people who say they get stuck behind them.

But even turning right in the intersection is not always simple. A local worker told me that this month, she was turning left to Fairview, expecting to follow a Waymo van already in the middle of his turn—a maneuver she said she’s done countless times with non-Waymo cars.

She said she nearly hit the van, which froze. She hit the brakes, came to a stop in the the intersection, yelled things she said she didn’t want me to print, then steered around the van. (Yeah, I know, my arrows are amazing.)

I was surprised to hear such negative feedback from people who encounter the Waymo vans the most. Some didn’t seem to think there were problems, but most people I talked to complained, even about how the vans drive on Fairview [see photo], which is a really wide and boring road.

That’s the tip of the iceberg re challenges for Waymo, which are really the challenges for the whole industry. The story includes a list. E.g. vehicles get confused & stop due to all kinds of things like ramp meters. Stops cause fender benders as cars hit Waymo from behind.

Other issues: Errors in digital maps that Waymo’s vans rely on. Hardware failures or miscalibrations have been common too, including with lidars and cameras. “Shit like that pops up all the time,” said one person who has examined data on this at Waymo.

I spoke to 5 people w/direct knowledge of Waymo’s recent travails & will get into more of that later. But first, how did we get here? I won’t take you back to 2009 but we can start in late 2015, when Waymo (then a Google unit) finished a demo of driving a blind man in Austin.

Side note: After the demo, Google fired workers who allegedly celebrated at a party with controlled substances. People refer to this as “Cocaine Cowboys.” It was considered a hush-hush embarrassment. Also Google realized Austin wasn’t the easiest place to test cars anyway.

Heading into 2016, lots of Googlers wanted to build a car w/no steering wheel & pedals, but Larry Page said no. Hence the Chryslers. Initially, the plan was to launch a commercial robotaxi service by end of 2017. That didn’t happen obvi. More on this here: go.theinformation.com/yct1DjV6cXQ

Ultimately, Waymo chose Chandler because, well, look at it. It’s at least several times easier than in Silicon Valley, where Waymo has been testing cars a lot longer. Bonus scoop—Apple recently landed in the area too!

Waymo’s been testing there in earnest since early 2017. In late 2017 we wrote about the issues Waymo was starting to encounter, among them the troubles with unprotected lefts, and other things.
go.theinformation.com/a41f8a998f7934…

There’s been a lot of progress, including on highways, which is critical in a place like Chandler. You often use them to get anywhere farther than a few miles, and pretty much everything is many miles apart. Our article discusses other driving features Waymo has improved.

Where does this leave us? There are times when it seems “autonomy is around the corner,” and the vehicle can go for a day without a human driver intervening, said a person familiar with Waymo. Other days reality sets in because “the edge cases are endless.” Photo not from AZ :)

On the other hand, the minivans never text and drive. They’re not in a rush. And they can sense some objects, including other cars or people, before human drivers do. In theory, that means they can do a better job at not hitting them. Thanks, lidars!

But replicating driving is hard because it’s more complex than you and I ever think about. There’s much we do that’s hard for a robot. And if the robot actually follows the law, that makes things harder because human drivers are rule breakers. [Our article gets into this.]

“The trouble with self-driving cars is people”—specifically, predicting their behavior, Gill Pratt, who is head of Toyota Research Institute, told me recently. Today’s AI doesn’t know that a mother holding a child likely won’t cross the street illegally but two teenagers might.

Same goes for anticipating or solving problems with other drivers. As one Waymo-er put it: “There’s just stuff humans can solve by looking at the other driver” in ways that a machine cannot.

Pratt added the bar is high for any company to launch robotaxis if they aren’t flawless. “Human empathy for another human being that makes a mistake [while driving] will not apply in case of a machine,” he said. See: Uber’s software programming disaster: theinformation.com/articles/uber-…

So lower your expectations for truly driverless taxis to get you through the day. We’re still close to the starting line of this long but worthwhile race. If you are interested in the field & what the worldwide leader is going through, check this out: go.theinformation.com/4596b7125469ea…

Post script: good annual reminder to pay attention to @lessin who said long ago that self driving cars "need to be very, very close to perfect before they are valuable at all. There is no 50% credit."

go.theinformation.com/PHToFF8BgOM