Self-Driving Car Development Is Stalling. An Incremental Approach May Be the Best Strategy

The tech hype machine went into 'overdrive' about driverless cars early and often. But research suggests the public perception remains uncertain and that squeamishness about the risks is stubbornly high.

Testing and announcing incremental advances may be the safer, smarter and more realistic approach. JL


Drew Harwell reports in The Washington Post:

The road to self-driving cars is perception and capability. Early breakthroughs such as car-mounted cameras to keep the vehicle in its lane spawned breathless coverage. But self-driving engineers are battling much thornier challenges, including how to signal to other drivers, navigate obstacles and handle rain, heat and snow. Earlier tasks were conquered by installing sensors, pouring in data. But the new hurdles demand a much more complicated form of intelligence. A go-slow approach might be the only option the industry has for producing a car people trust. Overpromising, “makes it harder when the technology is ready for us to say, ‘Honestly, we’re ready now. It’s safe.’”

Three former executives at Google, Tesla and Uber who once raced to be the first to develop self-driving cars have adopted a new strategy: Slow down. And shut up.

At their new company Aurora Innovation, which is developing self-driving technology for carmakers including Volkswagen and Hyundai, the rules are simple: No flashy launches, mind-blowing timelines or hyper-choreographed performances on closed tracks. “No demo candy,” said Chris Urmson, a co-founder and former head of Google’s self-driving car team.

Aurora’s long-game technique reflects a new phase for the hyped promise of computer-piloted supercars: a more subdued, more pragmatic way of addressing the tough  realities of the most complicated robotic system ever built.

In the wake of several high-profile crashes that dented public enthusiasm in autonomous cars, Aurora’s executives are urging their own industry to face a reality check, saying lofty promises risk confusing passengers and dooming the technology before it can truly take off.

The “entire industry” needs “to be more truthful about our capabilities,” said Sterling Anderson, an Aurora co-founder and former head of Tesla’s “Autopilot” system. “We’re talking about building trust in the public. You don’t do that by overstating what the system can and can’t do.”

A safe, sellable, self-driving car has become the holy grail of Silicon Valley: Converting even a fraction of the 3 trillion miles driven every year by American cars and trucks could reduce deadly accidents caused by human error and spawn a multibillion-dollar business.

Fifty-seven companies are authorized to test self-driving vehicles on California roads, state records show, and competition for talent, software and technologies has sparked an arms race even among traditional automakers: Ford and GM have each invested roughly $1 billion into their own respective self-driving firms, Argo AI and Cruise.

The first company to market a true self-driving car could gain incredible cachet, winning buyer trust and marketing potential for which its rivals could only dream.

But industry experts suspect most will fail: The timeline is so uncertain. The development costs are so high. And the competition is so fierce, in the United States and around the world. The investment giant NIO Capital estimated in August that the “survival rate” for China’s hundreds of self-driving start-ups would be about 1 percent.

To stand out, companies have pushed for increasingly strange or captivating visions of the future: Mercedes-Benz recently unveiled self-driving concept cars straight out of science fiction, including a windshield-less vehicle that resembles a giant toaster.

Tesla’s Elon Musk said in late September that his electric cars were only a software update away from being able to drive themselves on “autopilot” across the country — another attempt at the same pledge he made in 2016 and missed meeting last year.

 

But that spin has experts worried the industry is setting itself up for a fall after notable setbacks, including a crash in March during which a woman in Arizona was struck and killed by a self-driving Uber. The company has suspended testing indefinitely; a backup driver in the car was looking at her phone at the time of the crash.

“It’s risky if you’re generating so much hype and you can’t live up to it, which fundamentally has been almost every other company,” said Missy Cummings, the director of Duke University’s Humans and Autonomy Lab, who taught Anderson when she was a professor at Massachusetts Institute of Technology. “They’re under a lot of pressure to perform, and that crash caused everybody to stand back and say, ‘We’re pushing an envelope we don’t understand.’ ”

The crashes have contributed to public anxiety over the technology: More than 60 percent of Americans surveyed this summer in two polls by the Brookings Institution and Advocates for Highway and Auto Safety said they were unwilling to set foot in a self-driving car. Believers in the technology’s potential to save lives worry that such early unease could lead drivers to shun the vehicles, delay their adoption or write them off for good.

Some in the industry argue such fears are misplaced, given that tens of thousands of Americans already die in car crashes every year. But Urmson says it’s reasonable for people to have concerns over a new and unproven technology when the marketing and the real-world capabilities appear so vastly mismatched.

“The status quo itself is broken. It's not like we have safe transportation today on the roads,” he said. “But it’s important that we be as straightforward as possible about what you should expect, as someone using the technology

and what the limitations are. We can’t oversell.”

Early breakthroughs in self-driving technology, such as using car-mounted cameras and computers to keep the vehicle in its lane, spawned breathless media coverage and spurred industry marketers to suggest true autonomy was close at hand. Some of those features are now sold on car lots as a step closer to real self-driving, known as “driver-assistance” systems.

But self-driving engineers and designers are battling much thornier challenges, including how to dodge pedestrians, signal to other drivers, navigate obstacles and handle rain, heat and snow. Earlier tasks were conquered by installing sensors, pouring in data or teaching the computers useful tricks. But the new hurdles demand a much more complicated form of intelligence — a mix of road awareness, physical sturdiness and human psychology, among other qualities — that could take years to research and understand.

The road to self-driving cars is largely broken down into levels of perception and capability, from the simpler driver-assist systems of Level 1 to the fully self-driving models of Level 5. Most on-the-road technologies now are around Level 2, demanding lots of human interaction and oversight. And experts worry the next phase of mostly incremental improvements needed to get to true autonomy could leave many disappointed or disinterested. (In artificial intelligence, this cycle of hype and disillusionment has repeated so often it has a name: “AI winter.”)

Companies such as Tesla have pushed to roll out self-driving features piece by piece, gradually shifting the car’s most basic functions from human to machine. But Aurora’s executives argue that such approaches give conflicting messages, suggesting drivers should both let the car handle itself but also be ready to avert disaster at a second’s notice.

“Drivers cannot be inattentively attentive,” they wrote online in September. “We will not release a system that misleads drivers about its capabilities and thus raises the already too-high risks of driving.”

Aurora has quickly become one of the industry’s most promising challengers, with a founding trio of veteran automation experts, including Urmson, Anderson and Drew Bagnell, formerly of Uber’s self-driving division.

The company has raised $90 million in funding since launching in 2016 and now employs more than 150 in Palo Alto, San Francisco and Pittsburgh, where the company runs daily road tests using unmarked cars such as the Lincoln MKZ with spinning laser sensors on the roof.

But few have seen how the cars actually drive, and the company has worked intensely to keep its technology under wraps: A rare public sighting came in June, when one of its covert test cars was rear-ended in Palo Alto while slowing down to dodge a squirrel.

Aurora’s model of signing deals to build the self-driving technology for automakers — including Volkswagen, Hyundai and the Chinese start-up Byton — has helped it avoid the pressure of having to win over mass-market car buyers. Aurora has sought to shop its technology of sensors and artificial-intelligence software to multiple automakers, helping it guard against the risks of partnering with a single automotive giant or the expensive slog of building cars.

The company’s major automotive partners say they’re content to let the team work at the pace the tech demands. Johann Jungwirth, the Volkswagen Group chief leading self-driving efforts for the world’s largest automaker, said he’s “completely relaxed” about other companies winning the technology’s early attention.

With Aurora’s help, the automaker is building a “double-digit number” of self-driving e-Golf electric cars, which they’re testing in the Bay Area this year. Jungwirth and his team moved this summer from Germany to Silicon Valley to expand the rollout on what he called a “need-to-know basis approach.”

“If the competition wants to disclose and show us where they are, it might even help us, to see for ourselves without having to disclose ourselves,” Jungwirth said.

Unlike companies such as Ford, which says it will have a fully autonomous vehicle in commercial operation by 2021, Aurora offers no hint of a set timeline. And it is not the only self-driving team pushing for more managed expectations. John Krafcik, chief of Waymo, the self-driving division of Google’s parent Alphabet that is experimenting with giving people rides in robo-vans around Phoenix, said in July that widely available self-driving cars will take “longer than you think.”

Cummings, the Duke director, said that go-slow approach might be the only option the industry has for producing a self-driving car that people actually trust. Overpromising, she said, “makes it harder when the technology is actually ready to go for us to say, ‘Honestly, seriously, we’re ready now. It’s safe.’”