Listen to this article
This is an experimental feature. Give us your feedback. Thank you for your feedback.
What do you think?
An air of unreality hung in the air two weeks ago as Waymo — the company that began life as Google’s driverless car project — lifted the lid on its secretive testing facility in California’s agricultural heartland.
It was not just that the parade of Chrysler minivans had no one in the front seats as they were put through their paces, navigating streets laid across the 91-acre site built to emulate a real town.
Waymo employees cycled past, crossed the street or faked a car breakdown, as the driverless vehicles negotiated mock-ups of junctions and a roundabout. In one case, the stand-ins replayed dropping boxes from the back of a removal van, like extras in The Truman Show, the film in which actor Jim Carrey is caught in a fake world with an endlessly repeating human backdrop.
The exhaustive repetition has played a central part in refining Waymo’s software from the safety of the carefully staged and well-mapped world of the Castle testing site, located about two hours south of San Francisco, where every inch of roadway is known intimately.
This week, however, Waymo — now a sister company of Google, under holding group Alphabet — said it was finally ready for the real thing. It revealed that its driverless minivans, with company engineers consigned to the back seats, have been negotiating the streets of Chandler, a suburb of Phoenix, Arizona. Within months, members of the public who have enrolled in a test programme will get their chance to summon the vehicles themselves.
Waymo employees create tests for a self-driving Chrysler Pacifica © AP
Residents of Chandler say they have already become accustomed to the sight of driverless cars on their roads. “There are dozens of cars they’ve deployed, they’re everywhere,” says Micah Miranda, the city’s economic development director. “They are spectacularly unspectacular.”
Eventually, the minivans are intended to cover the entire Phoenix metropolitan area, a region the size of Greater London but far easier for a robot vehicle to navigate thanks to the desert climate and regular grid of streets.
Leaving the artificial world behind and taking drivers out of the front seats has made Waymo the first company to put truly driverless cars on to public streets. But crossing the gulf to the real world is just the first step in a process likely to take years, as the company slowly extends the areas in which it feels confident to drive. And technical failure could be catastrophic.
Waymo’s step is a milestone for the industry but the announcement has left many asking whether the company’s approach will prove to be a formula for long-term success in autonomous vehicles or if it will be leapfrogged by younger companies.
On the timescale of Silicon Valley, Waymo is the grandfather of the self-driving car sector. At eight years old and backed by more than $1bn of investment, the company is by far the oldest and best funded of its peers.
“They have an advantage of time and resources,” says Raj Rajkumar, a robotics professor at Carnegie Mellon University. “Many other companies are doing the same thing . . . but they have been doing this for longer.”
The key behind Waymo’s leap to fully driverless cars is the millions of miles of testing the company has completed — more than 3.5m miles driven in the real world, and over 3bn miles tested virtually through computer simulations. No other company has come close to matching that scale. Public records in California show that Waymo performed 600,000 miles of testing in the state last year — more than 30 times as much as all the other testers combined.
Now that it is launching its driverless cars into the wild, Waymo must demonstrate its vehicles are as safe as it says they are and faces the possibility of a public backlash if a serious accident occurs.
There is a growing list of traffic accidents involving autonomous cars in the US — including one on Thursday, in which a self-driving shuttle made by Navya, a French start-up, had a minor collision with a truck in Las Vegas — but these have all been blamed on human error in the other vehicle.
Inside a Chrysler Pacifica minivan equipped with Waymo self-driving technology © AP
Over the years the company has earned a reputation for a conservative, safety-orientated approach, and the vehicles have redundant systems for essential functions such as braking, steering and computing. In initial public tests, a Waymo engineer may sit in the back seat with the ability to hit an emergency button to bring the car to a safe stop (passengers can also do this).
In the longer term, Waymo must also prove its approach is capable of scaling beyond the handful of well-tested areas where every pothole and speed bump has been mapped and modelled.
“The key question is going to be about whether the fundamental architecture of Waymo’s system is massively scalable,” says Olaf Sakkers, a partner at Maniv Mobility, a transport-focused venture capital firm. “Or whether it has been built in a way that is difficult to scale.”
The company’s approach relies on detailed three-dimensional maps to help the car understand its location, as do most self-driving systems. Compiling these maps demands vast resources and presents one potential bottleneck for deploying driverless cars at a large scale.
Waymo also builds its entire self-driving system itself, including not only the software but also hardware such as laser sensors. This has enabled it to achieve a level of precision and integration that would be unmatched if it had used off-the-shelf parts. But it could also leave the company at a disadvantage as the industry expands and specialist suppliers emerge, giving rivals the benefit of the scale economies that are likely to develop.
The other question facing the company is whether it can succeed in merging today’s most advanced artificial intelligence with more traditional approaches to software programming to create vehicles that can adapt and learn from all the data they gather — while also remaining entirely predictable and safe.
Google’s autonomous car research began years before deep learning, whereby an artificial neural network is used to emulate the way a human brain functions, had become the de facto approach for many of the newest self-driving start-ups.
The advantage of a deep learning system is that it can, in theory, result in a more sophisticated algorithm and speed up the learning process. But it requires a vast amount of data to train itself.
A more traditional, rules-based approach, in which lessons learnt from exhaustive tests are hand-coded into the software, is cumbersome to build. It also runs up against the problem that it is impossible to predict the almost infinite number of situations a vehicle could experience.
Waymo uses a mix of both, and its system requires extensive repetitive training — sometimes a scenario will be run hundreds of times with tiny modifications, to make sure the vehicle responds correctly in all instances.
But striking the right balance between the two is a challenge.
Jeff Schneider, Uber’s head of machine learning, notes that “one of the criticisms of the deep learning case is that there are so many edge cases, you might need an exponential amount of data to train the system”.
But the rules-based approach can require an exponential number of software engineers.
“I regard it as a little bit of an open question as to which of these will get us to the end fastest,” he says.
Peers play catch-up
Waymo’s competitors, such as Uber and GM’s Cruise, have invested more and more heavily in the kind of testing and training that enabled the Alphabet company to be the first to launch a driverless car over the past year.
Uber’s driverless car programme has had a challenging 12 months. It has been bogged down by a bitter lawsuit, filed by Waymo, that accuses the company of stealing intellectual property related to laser designs. Uber denies the allegations but the company’s self-driving car unit has been hurt by the legal distraction and Uber fired Anthony Levandowski, its top self-driving engineer who it had poached from Waymo, as a result of the legal proceedings.
Despite these challenges, Uber’s significant financial resources and its access to consumers through its ride-hailing app mean that it is still considered a very serious competitor in autonomous vehicles. Waymo also plans to use its autonomous cars to one day offer a consumer transportation service, which will only heighten the rivalry between the two companies.
Among the traditional automakers, GM is viewed as being at the head of the pack in terms of its autonomous research thanks to its $1bn acquisition last year of Cruise, a self-driving start-up.
Auto components maker Delphi has also made a bold move into the market with its $450m purchase last month of nuTonomy, an autonomous vehicle start-up based in Boston.
None of them has said they will go driverless like Waymo has. But they might not be too far behind.
This post originally appeared on Financial Times