Crashing Castle: An autonomous ride in Waymo's playground

What once was an Air Force base is now the testing site for one of the most advanced autonomous cars on the planet. Join us for a ride.

Tim Stevens Former editor at large for CNET Cars
Tim Stevens got his start writing professionally while still in school in the mid '90s, and since then has covered topics ranging from business process management to video game development to automotive technology.
Tim Stevens
8 min read

Decommissioned military bases tend to look like tired, mildly depressing, curiously unpopulated small towns, places where a community once thrived but now nothing much is going on. And that's basically exactly what they are: systems of buildings and streets still functional but now unnecessary. Some bases get rebooted into halfhearted industrial complexes, others are spun into municipal facilities, but one became something special: a pretend town for Waymo and its driverless cars.

Waymo is the former Google Self-Driving Car Project, now a separate corporate entity beneath Alphabet. Waymo set up shop at the former Castle Air Force Base back in 2013, when autonomous cars still seemed like some crazy technology for a crazy future. Today, we're achingly close to that technology being ready, but before then it needs more testing, and as it turns out that post-apocalyptic city vibe given off by decommissioned military bases makes the perfect scene for autonomous testing.

Let's take a tour, and a ride, together.

Waymo's autonomous Pacifica cruising through Castle

See all photos


The former Castle Air Force Base is huge, with a runway over a mile long that was once part of the Strategic Air Command. Long-range B-52 bombers flew from here to support combat operations in southeast Asia, though historically the base was largely used for crew training, including that of KC-135 mid-air refueling craft. It was officially closed in 1995.

Waymo's licensing of the location dates back four years, the company leasing 91 acres of vacant housing and office space, plus a matrix of empty streets and bare lots in between. Since then, Waymo has revised and expanded the facility, repaving some sections, intentionally braking up others. Though there's a railway just a few hundred yards away, Waymo built a fake crossing on site to test that setup, also deploying various devices like traffic lights and Botts' dots.

It's here that the company does the majority of its structured testing scenarios, preplanned and almost real-world challenges to its systems. There are 20,000 such test scenarios that have played out thus far, all with the goal of creating the world's most advanced autonomous car.

Today's version of the future of autonomy.


Minivan chic

If you're creating the most advanced, smartest, safest car on the planet, you might be tempted to make it look radically futuristic, to give it some sort of strong visual cue of intent. With that in mind, a minivan probably wouldn't be your top pick for silhouette. But that's exactly where you'll find Waymo's latest and greatest hardware and software, inside (and outside) of a Chrysler Pacifica Hybrid. We're quite fond of the Pacifica ourselves here at Roadshow, but beyond being a great car, a number of factors actually make it a great autonomous car.

Not the least of them is a pair of generous sliding doors, one on either side, that open automatically to admit you into a comfortable seating area fronted by touchscreens that have now been taken over by Waymo's user interface experts. The user experience of an autonomous car may feel secondary to the technology that makes it autonomous, but this is an overlooked part of the equation that Google is, for the first time, really showing off.

You might think it would be as simple as stepping into the car and telling it where you want to go, but that's just the beginning. In the real world, passengers demand a little more information from their driver, everything from indications of why they left the highway an exit early to why they aren't moving through the intersection when it looks clear.

"Predictability and communication are key to trust," Juliet Rothenberg said at Castle. Rothenberg is Waymo's product manager for user experience, and a lot of that predictability comes from the communication. For Waymo's van, that communication is all handled through the pair of rear-seat entertainment displays in the Pacifica, which instead of playing movies or games now show what the car can see -- a hint of it, anyway.

Instead of a full representation of everything detected by the car's LIDAR sensors, the car presents a reduced graphical representation to prevent information overload. Blue boxes represent cars, while pedestrians get ghostly outlines. The system even puts blue platforms beneath the wheels of cyclists, to show it can tell the difference. "The goal of this interface is to give riders a sense of what the car can see, and also what the car intends to do," Ryan Powell said, Waymo's head of UX design.

There are hundreds of subtle cues built into the system, cues that you'd likely never notice, all intended to give the same sort of confidence you would get when getting in the car with a confident driver you trust. But of course to really build confidence you need to execute reliably, and Waymo's working on that, too.

Iteration through simulation

Waymo has a pretty simple and interesting tagline for its business: "Building the world's most experienced driver." That phrase doesn't mention technology nor process, merely the end result: a driver so overwhelmingly competent you'd have to be a technophobe of epic proportions to not feel safe.

To reinforce that idea, the company has a flurry of numbers to throw at you. Waymo's autonomous cars have covered a whopping 3.5 million autonomous miles, a massive amount when you figure the average American driver covers less than 14,000 miles per year. Let's optimistically say that someone gets 70 years of active driving in a lifetime, and that's still less than 1 million miles -- a third of what Waymo's amassed thus far.

And that's only scratching the surface. In addition to cranking out mile after mile on the roads, Waymo's cars are doing it in parallel in the virtual world to the tune of 25,000 cars driving 10 million simulated miles every day. That's 10 lifetimes worth of vehicular learning racked up by Waymo's systems every single day, and thanks to the (mildly disconcerting) hive mind nature of distributed systems like this, the learnings of any of these virtual vehicular agents is shared to all the rest.

This is exactly why we're doomed when the robot uprising occurs, but in the halcyon years between now and then you can see why these are, indeed, the most experienced drivers on the road. And so I was quite confident when my turn came to slip into the back of a Waymo Pacificas and go for a spin.

Around the block without a driver

My route in one of Waymo's autonomous Pacifica vans was preprogrammed, so all I needed to do was sit down, buckle up and tap the blue "START RIDE" button mounted on the ceiling. There are three other buttons up there, all labeled in Braille, incidentally. The first says "HELP" and immediately connects you with Waymo's call center in case you need to talk to an honest-to-gosh human. The next is the lock/unlock button, should your smart yet non-sentient van route you through a bad part of town. Then there's the "PULL OVER" button, whose functionality is self-explanatory.

Anyway, I pushed the blue button and away we went at a pace that was hardly breathtaking. However, it was, I have to admit, impressive. I've ridden in a few autonomous cars before and they tend to kind of just ease their way to a start, creeping up to speed at a snail's pace. Waymo's van took off like a normal driver would and proceeded to make short work of a tight U-turn before entering onto the street. There was nobody behind the wheel, not even a hands-off engineer.

Along my short tour of Castle, Waymo had set up a number of predefined surprise events courtesy of "fauxes" -- faux pedestrians and cyclists and the like meant to challenge the car's systems. In one case, a cyclist rode by us on the right, darting through a four-way stop. The car saw them coming and waited for them to pass before continuing.

Waymo Castle Test

Waymo exclusively hires the world's worst movers.


In another case, a car was broken down in an oncoming lane, opposite the Waymo van which wanted to turn left. As a human being, I could tell that car wasn't going anywhere, but to a combination of LIDAR and imaging sensors the car surely looked like it was just waiting for its turn to go through the intersection. The Waymo van stopped with its blinker on, paused for a moment and then, when it was clear the other car wasn't moving, proceeded through the intersection.

Obviously these were scripted events, and it's always hard to tell just how much of the autonomous car's reactions were similarly preplanned, but I have to say its behavior didn't feel contrived. The van really felt like it was truly processing these situations as they arose, dealing with them appropriately and moving on with the task of getting me to where I was going -- which, as it happened, was exactly where I started.

It wasn't all perfect, the van did lurch and slow down a few times, adjusting its line through turns here and there. It also spent a bit more time at intersections than a human probably would and overall felt just a smidge hesitant. There is room for improvement, but I never felt anything less than totally relaxed and perfectly safe.

Next steps

So it works, or at least seems to. That's great. The question, of course, is what comes next. Neither Waymo nor Google build cars, nor indeed does any other Alphabet company, making it easy to ask the question of "Why?" The leasing of 91 acres of land in California is surely not cheap, even a disused military base. And then of course there's the cost of retrofitting all those Pacificas (100 so far, with 500 more in the works) and the never-inconsequential factor of salaries for all those engineers building all those simulations.

It's an expensive game, and while Waymo CEO John Krafcik declined to single out the silver bullet that will turn all this into an immediately profitable endeavor, he did iterate through a number of possible revenue sources. The first two? Trucking, on the commercial scale, and ride sharing, a pair of opportunities that "make a lot of sense for a company like Waymo," he said. Trucking in particular was a $726 billion industry last year in the US alone, an industry not only ripe for reinvention but one that is struggling to find new drivers.

On the ride-sharing front, there's the partnership with Lyft that could, if all goes well, bear some seriously shiny golden eggs. Autonomy is the missing piece that stands poised to make all these ride-sharing startups like Lyft and Uber instantly profitable -- or at least closer to profitable.

And then there are potential partnerships with OEMs, Waymo licensing its technology to be built into (and onto) cars made by other companies. With so many manufacturers investing billions of their own dollars to make autonomy work it's hard to imagine many wanting to outsource the solution. But, for smaller marques that run the risk of being left behind when systems like this hit the road for real, something like this could be their salvation.

Still, before the question of revenue is answered there are still the not-inconsiderable issues of legality and liability of autonomous cars, topics that still lack resolutions. And even then there will be the notion of consumer confidence. Study after study shows that people overwhelmingly don't want and don't trust cars that drive themselves. But I'm here to say, skeptical humans, that one quick lap around the block will change your mind.