Earlier this month, GM CEO Mary Barra told the crowd at a financial conference that her company was on track to unveil a ridesharing service in 2019 that would be powered by autonomous vehicles. It sounded like pretty big news, but to understand company announcements about the commercial viability of driverless cars, one must first parse the semantics.
Because there is no legal framework in place at either the state or federal level for fully autonomous vehicles to operate on the nation’s roads, there would presumably be human “safety attendants” on board in GM’s ridesharing fleet. “The vehicles can currently run safely at speeds of up to about 30 miles per hour, and the service will be limited to a small geographical area,” Barra said at the DealBook event, according to the New York Times. Now that we’ve examined the details, the whiz-bang rideshare service starts sounding more like a glorified golf cart.
GM may be preparing to launch a semi-autonomous rideshare service which has features that others currently don’t, but the automaker is not entering fundamentally uncharted territory. There are already commercial self-driving shuttles and pilots underway, equipped with human attendants and navigating “geofenced” areas.
May Mobility is running them in Detroit and Columbus, OH; Waymo is doing it in Phoenix; Navya has shuttles in Las Vegas and on the campus of the University of Michigan; Drive.ai is operating a driverless service in Frisco, TX; and Boston’s Optimus Ride this week said it will build the world’s first “fully autonomous fleet service for geofenced deployments” in 2019. And this list hardly represents all the driverless deployments underway across the country and world.
But as autonomous vehicle pilot programs grow, the industry is simultaneously pulling back from some of its early timeline hubris that had both auto and tech execs claiming that they’d have AVs out on the road, at least in fleet deployments, by 2021.
Mobility has so far undergone an interesting hype cycle. First there was mass skepticism that AVs were even possible or that car manufacturers had much of a role in their design. But as the artificial intelligence technology underpinning them got more sophisticated and the projected profits grew, breathless prognosticators predicted that we could hand off many driving duties to robotic chauffeurs within the decade.
Now, industry insiders that we’ve been talking to lately have swung back around to estimating full autonomy—vehicles with no human intervention navigating the open road in all types of weather—is closer to 30 years away than 10. And, as the self-driving Uber fatality earlier this year demonstrated, more caution is well-placed.
Chris Mentzer oversees autonomous vehicle research at the nonprofit Southwest Research Institute (SWRI) in San Antonio, TX. He remembers his organization’s first demonstration of “unmanned ground vehicles,” which took place in Manhattan in 2008.
“At that time, automakers were literally declaring bankruptcy,” he recalls. Back then, much of the SWRI’s autonomous research work was being done for the military, but since then, researchers have slowly shifted to developing more auto industry applications for the technology. Describing the 2,600-person organization’s work, Mentzer says it provides components or expertise for companies that don’t have those capabilities in-house. And that process takes time.
“People [in the mobility industry] are starting to coalesce around the idea of slowing things down,” Mentzer maintains. “The realists—the people working on [AVs] a while—have been thinking along these lines the whole time. The new people are spending a ton of money, but at the end of the day, it’s a really hard technical problem to solve.”
You can make lots of progress early and pull off a good-looking demo, he says, but there might not be much “there” there.
“The thing holding all of us back from Level 5 autonomy are all these edge cases—it’s impossible to know and test everything. It used to be that with machine learning, you could teach it to recognize images of a car, but not if they were hand-drawn. The new technology is better at generalization and knows the hand-drawn picture is a car, but it’s still not enough.”
He says a major challenge is that the neural networks and A.I. involved in AVs is essentially teaching itself—and researchers can’t get a clear picture of what’s happening inside the car’s brain until it’s been tested over tens or hundreds of millions of miles.
The idea of extending the AV development timeline “was coalescing before the Uber accident—there’s been a lot of talk over the past few years,” Mentzer says. “The early service dates are hitting now and we’re closer, but not where we said we’d be. The reality is sinking in of how hard this is.”
Sven Beiker, founder and managing director of Palo Alto, CA-based Silicon Valley Mobility consultancy (and a Stanford professor), agrees that it’s becoming more acceptable to have a conservative approach to mobility, which he partially attributes to the Uber fatality and Tesla’s fatal Autopilot accidents earlier this year.
“It’s become a good story if someone says they want to get this right and make sure the system is safe,” he says. At one time, he adds, they might have been looked at as innovation-killing naysayers, but now they seem prescient.
“People are also realizing more and more how long it takes” to teach a car to think, anticipate threats and hazards, and discern driver and pedestrian intent like a human. “For a tech company to go from 0 to 60 [on autonomous vehicles] in two years is just not credible,” Beiker says.
Beiker is a self-described admirer of Barra, yet he doesn’t expect a GM vehicle will pick him up at his house without a human safety attendant in the next few years. However, he does see a limited AV deployment within a constrained location, with perhaps pre-registered riders, as a “next sensible step.”
Sensibility is exactly what the SWRI’s Mentzer is advocating.
“People are trying to sell hype with no true accountability for the words they put out there,” he laments. “There’s not a lot of risk in making bold predictions. Level 4 autonomy has a lot of caveats you can throw in there—like a highly constrained scenario where they still claim Level 4, but it’s not high-level autonomy. We probably need a better way to describe it.”
Still, Mentzer is bullish overall on the potential societal benefits that AVs could offer. “It’s a fun industry and a good, interesting challenge,” he says. “The problem is hard, but self-driving cars could benefit a lot of people and really change how we live our lives. We’re seeing AVs come, but probably not as widespread or the way we expected.”
