# Inside Castle, Where Google’s Waymo Tests Its Self-Driving Cars

32

Life is a complex problem. Consider the simple act of moving from point A to point B. Solving for that equation requires synthesizing numerous variables, like speed and obstacles. What if you make it more complicated? Like putting an unpredictable human in charge of a one-ton vehicle traveling at ever-changing speeds. Throw in some jaywalking pedestrians, rule-bending cyclists, and, the most erratic variable of all—other drivers. Drivers who are potentially distracted, potentially sleepy, potentially ragey, potentially drunk. With so much to account for, it’s easy to see how it can all add up to 40,000 annual traffic fatalities in the US.

For years now, people have heralded a technological solution to this complicated problem: self-driving cars. And eight years ago, Google unleashed a pack of engineers to solve for the many, many x’s. That internal project evolved into Waymo, Alphabet’s self-driving car company (Alphabet is Google’s parent company), and now the company is ready to show a little bit of its work. On Monday, the company gave a gaggle of 30 or so reporters a tour of Castle, its carefully-constructed self-driving car testing grounds in Atwater, California, two-hours southeast of San Francisco.

Plopped amid the brown and yellow expanse of California’s Central Valley, Atwater has been home to Waymo’s closed course testing ground since 2013. Castle takes its name from the Air Force base that once occupied this land, where the military trained World War II bomber pilots. The roads here wind their way through the 91-acre territory, intersecting, merging, forming roundabouts, connecting the low-slung, sand-colored buildings Waymo has taken over. Since moving in, the engineers have added even more complicating features: crosswalks, curbs, driveways, traffic lights, and a railroad crossing (no train), among others.

The whole tableau isn’t much to look at—a few trees, some piles of dirt and rocks, the occasional orange traffic barrier, all of it surrounded by eight-foot-high green fences that deny prying eyes—but for the engineers and test drivers who spend their days here, the scenery matters little. They’re here for the network of roads on which they can choreograph just about any driving scenario they like.

Waymo’s shiny clean autonomous minivans run pre-planned encounters with rampaging bikers, unyielding drivers, malicious potholes. The minivan handles a car cutting it off, a dopey driver pulling out of a faux driveway without looking, and a woman next to a moving truck knowing a pile of boxes into the road while another car approaches from the other direction. On more inspired days, Waymo says, its testers leap out of porta-potties into the street and hurl stacks of paper in front of their vehicles.

They run the drills. And then they run them again, and again, and again. This is what Google’s engineers call “structured testing,” creating discrete tasks for its cars to repeat over and over, honing their skills with each pass. The method complements the 10,000 miles Waymo’s cars drive on public roads each week, and the millions more they traverse in computer simulations.

But these tests represent a tiny slice of the work Waymo has done over the past seven years. The company’s self-driving vehicles have clocked 3.5 million miles in 20 US cities, plus billions more in computer simulations. Waymo has been ferrying its own employees around in its cars for years, and earlier this year, the company launched a program for select riders in Phoenix to take free trips in its fleet of self-driving Chrysler Pacifica minivans. (Waymo has 100 minivans, and is working with the automaker to build 500 more.)

Yet, a bigger question remains unanswered: when will Waymo finally launch a commercial service? “We’re really close,” CEO John Krafcik, says. He’s considering a variety of business models—running a driverless ridehailing service, building trucks, selling the technology straight to automakers—but resists efforts to pin him down on which is most likely, or when Waymo will be ready to start any or all of these.

The goal is to give passengers control, even if they don’t get a steering wheel.

The ridehailing service seems a good bet for a debut, since Waymo’s already testing a version of this in Arizona, where its select passengers use a smartphone app to summon a self-driving minivan whenever they feel like going somewhere. Juliet Rothenberg, Waymo’s product manager for the Phoenix early rider program, declines to reveal how many users the service has, how often they ride, or what limitations Waymo has placed on how far the cars can go and when they’re available. She is, however, happy to show us the user interface Waymo has developed to keep them at ease.

The system consists of two elements. The first is a series of four buttons, mounted on the ceiling of the minivan, just behind the driver and passenger seat headrests: Help connects you to an operator, who can talk you through any problems. (Asked if Waymo would consider the sort of remote control feature Nissan is developing, to help stranded vehicles, Krafcik says, “It’s not a crazy idea.”) Lock/Unlock controls the doors. Pull Over Here lets you hop out of the car before reaching your destination or halt to pick up a friend. And last, in white lettering on a blue background, is Start Ride. The goal, Rothenberg says, is to give passengers control, even if they don’t get a steering wheel. (Yet another unanswered question: Will Waymo cars have steering wheels? In the minivans, passengers sit in the second and third rows. The company retired its fleet of steering wheel- and pedal-free “koala” cars in June.)

The second UX element is the pair of screens stuck to the back of the driver and passenger seat headrests, there to clue passengers into what the car “sees” and is planning to do about it. Compared to the versions of this worldview Waymo demonstrated in 2015 and Uber uses in Pittsburgh, this is a pared down mise-en-scène, highlighting only key elements of the driving landscape.

The car’s path is highlighted in green against a dark blue background. Other vehicles are 2-D blue rectangles, emergency vehicles have flashes of red going round their edges. Traffic cones appear as traffic cones, crosswalks show up in white. Cyclists and pedestrians look like board game pieces, moving around on blue and white circles, respectively. The “status layer” adds an ETA, plus written messages to explain what the car’s up to: “yielding to pedestrian,” “looking for a spot to pull over.” The goal here is to keep the human cargo informed—and duly calm. (A few more unanswered questions: How many configurations did Waymo try before getting to this one? How did it settle on these details? Is this the final version or a work in progress?)

So yes, there’s a lot to wonder about. What’s clear, sitting in the third row of a minivan with no human behind the wheel, hustling through Silicon Valley’s version of a Hollywood film set, is that this technology is indeed real, and it’s on its way. It has mastered the math of the ersatz world. Now comes adding up its skills to solve for the real thing.