On any given day, Alphabet’s Waymo runs its driverless minivans through dozens of staged situations on an old Air Force base in Atwater, Calif., more than 100 miles outside the company’s Mountain View headquarters.
The company is preparing its cars for the real world. So, a man might jump into the lane, a cyclist might pop a wheelie, a human driver might cut the robot car off. The car figures out how to react and then staffers run the test again. And again. Until the system learns something new.
There’s not always a safety driver behind the wheel waiting to take over. But that’s only when there’s little chance for unpredictability, such as jay-walking pedestrians.
The structured tests run on the site, called Castle, where a nondescript fence defines a 91-acre area of roadways and intersections where the company has, over time, set up a total of more than 20,000 different scenarios. Today, Alphabet’s Waymo is predominantly running its Chrysler Pacifica vans for their tests.
But the tech isn’t ready to do all of what it has learned outside the walls of Castle, especially without a driver.
In a rare moment for the company, Alphabet showed a completely driverless test to more than two dozen invited reporters, including from Recode, the New York Times, Bloomberg and Vice News. It was the first time the secretive Alphabet showed off its self-driving vans in any extensive showcase.
But if the test illustrated anything, it’s that these driverless systems have a lot more to learn.
Alphabet knows this. It will only make driverless cars a reality outside the confines of Castle after the company is confident the system can navigate these one-off situations, and even then only in small, geo-fenced areas.
Waymo CEO John Krafcik wouldn’t give a timeline for when consumers will start to see cars without drivers, but said the company is exploring several opportunities like trucking, ride-sharing, ride-hailing and working with cities. The company wouldn’t specify which cities it’s currently in talks with.
“We’re getting to the point now where … we can say we’re getting close,” Krafcik said. “[I’m not going to give] a specific date, but we’ll do it when we’re convinced that we’re ready.”
Making sure the cars are ready to navigate even a single situation and its many possible variables without a driver in the fake world of Castle, much less in the real world, is a long process.
Waymo wouldn’t go into too many details of what the process of preparing this demo entailed.
What we know is that when the car learns how to perform a single task, the company’s operations team and engineers throw in a new variable on the barren roads of Castle or in simulation, the virtual world the company created to teach its cars. The company’s cars “drive” 10 million miles a day in simulation.
This process continues until the system picks up something new. This is also why it’s important that Waymo has been racking up real-world miles as well. Outside the walls of Castle, Waymo has 100 minivans driving 10,000 miles on real roads every day. That’s supplemented by those 10 million simulated miles and the 20,000 structured tests run on Castle.
Still, it’s not an entirely revolutionary feat for the car to drive on its own when it’s on a predetermined route with staged scenarios. But it is one of the first times the company has been willing to demonstrate that to the press.
“We wanted to show [people] just how far we’ve come,” Krafcik told Recode when asked why the company, which was secretive about its progress early on, was being vocal about it now.
But there are still many unanswered questions, such as how the company will know or decide when it’s ready to launch a fully driverless service in a specific area and whether the company intends to license its tech to other players and automakers.
By comparison, when I rode in an Nvidia-powered car without a driver around a small test track at the Consumer Electronics Show earlier this year, there were no variables and the car still had to be remotely disengaged out of precaution.
And two years ago, when I rode around the Alameda Navy Base in a Mercedes concept autonomous vehicle without a safety driver to take over, the track was completely obstacle-free. In fact, reporters were specifically told not to get too close to the car as it pulled out of its parking spot.
Waymo’s test was incredibly smooth and uneventful. The car responded quickly and efficiently to the choreographed scenarios along the predetemined route. There were no sudden stops or jerky movements.
But the real world isn’t a test track with choreographed scenarios for a car to navigate.
On a bright Monday afternoon, Waymo took us for a spin in its driverless Chrysler Pacifica minivans through a loop around a small part of Castle. Everyone, including the safety driver, sat in the second or third row of the van.
After we pressed a physical start button on a console above our seats, the car began driving itself.
But it drove itself through a previously specified route along which the company choreographed different situations the car might encounter in the real world.
The test center, usually bustling with activity, was quiet except for the multiple Waymo operations staffers positioned on a variety of corners waiting to perform their assigned task.
One man, purportedly stranded because his car broke down, blocked a road. A man and a woman crossed an intersection in front of the car. A cyclist rode quickly alongside us.
The car turned down a different road, stopped for the pedestrians, and drove safely alongside the cyclist.
The route we took was not unlike some of the structured tests the company demonstrated for us — packed with situations the car has already been taught to navigate. But we couldn’t throw in our own variables.
Alongside the start button was a “pull over” button — a feature the company said it wasn’t testing at the moment when we asked if we could press it.
We did push the “help” button on the console, however, which connected us with a customer support agent.
The LED consoles showed the streets we were driving on and a version of what the car was seeing as the car looped through sometimes unregulated intersections. It communicated via text on the console when the car was about to pull over to come to a stop and when it was yielding to pedestrians.
In a way, the consumer-facing console and the overall user experience Waymo demonstrated was the most revolutionary part of the ride.
Consumer experience, particularly for a company like Google and Alphabet, will be increasingly important as real humans get in driverless cars. It’s what will set one self-driving network apart from another, and it’s what will help garner the trust of consumers in a driverless system in the first place.
It was easy to feel comfortable as the console guided us through what the car was doing and seeing, making it natural to take my attention away from the wheel as it turned on its own.