When it comes to self-driving cars, the future was supposed to be now.
In 2020, you’ll be a “permanent backseat driver,” the Guardian predicted in 2015. “10 million self-driving cars will be on the road by 2020,” blared a Business Insider headline from 2016. Those declarations were accompanied by announcements from General Motors, Google’s Waymo, Toyota, and Honda that they’d be making self-driving cars by 2020. Elon Musk forecast that Tesla would do it by 2018 — and then, when that failed, by 2020.
But the year is here — and the self-driving cars aren’t.
Despite extraordinary efforts from many of the leading names in tech and in automaking, fully autonomous cars are still out of reach except in special trial programs. You can buy a car that will automatically brake for you when it anticipates a collision, or one that helps keep you in its lane, or even a Tesla Model S (which — disclosure — my partner and I own) whose Autopilot mostly handles highway driving.
But almost every one of the above predictions has been rolled back as the engineering teams at those companies struggle to make self-driving cars work properly.
What happened? Here are nine questions you might have had about this long-promised technology, and why the future we were promised still hasn’t arrived.
Engineers have been attempting prototypes of self-driving cars for decades. The idea behind it is really simple: Outfit a car with cameras that can track all of the objects around it and have the car react if it’s about to steer into one. Teach in-car computers the rules of the road and set them loose to navigate to their own destination.
This simple description elides a whole lot of complexity. Driving is one of the more complicated activities humans routinely do. Following a list of rules of the road isn’t enough to drive as well as a human does, because we do things like make eye contact with others to confirm who has the right of way; react to weather conditions; and otherwise make judgment calls that are difficult to encode in hard-and-fast rules.
And even the simple parts of driving — like tracking the objects around a car on the road — are actually much trickier than they sound. Take Google’s sister company Waymo, the industry leader in self-driving cars. Waymo’s cars, which are fairly typical of other self-driving cars, use high-resolution cameras and lidar (light detection and ranging), a way of estimating distances to objects by bouncing light and sound off of things.
The car’s computers combine all of this to build a picture of where other cars, cyclists, pedestrians, and obstacles are, and where they’re moving. For this part, lots of training data is needed — that is, the car has to draw on millions of miles of driving data that Waymo has collected to form expectations about how other objects might move. It’s hard to get enough training data on the road, so the cars also train based on simulation data — but engineers have to be sure that their AI systems will generalize correctly from the simulation data to the real world.
That’s far from a complete description of the systems at work when a self-driving car is on the road. But it illustrates an important principle to keep in mind when wondering where our self-driving cars are: Even the “easy” things turn out to hide surprising complexity.
Self-driving cars rely on artificial intelligence to work. And the 2010s were a great decade for AI. We saw big advances in translation; speech generation; computer vision and object recognition; and game-playing. AI used to have a hard time identifying dogs in pictures; now that’s a trivial task.
It’s this progress in AI that drove the optimistic predictions for self-driving cars in the mid-2010s. Researchers anticipated that we could build on the amazing gains they’d seen (and are still seeing) in other arenas.
But when it came to self-driving cars, the limitations of those gains became very apparent. Even with extraordinary amounts of time, money, and effort invested, no team could figure out how to have AI solve a real-world problem: navigating our roads with the high degree of reliability needed.
Much of the problem is the need for lots of training data. The ideal way to train a self-driving car would be to show it billions of hours of footage of real driving, and use that to teach the computer good driving behavior. Modern machine learning systems do really well when they have abundant data, and very poorly when they have only a little bit of it. But collecting data for self-driving cars is expensive. And since some events are rare — witnessing a car accident ahead, say, or encountering debris on the road — it’s possible for the car to be out of its depth because it has encountered a situation so infrequently in its training data.
Carmakers have tried to get around this in lots of ways. They’ve driven more miles. They’ve trained the cars in simulations. They sometimes engineer specific situations so that they can get more training data about those situations for the cars.
And they are getting closer. Waymo cars do roam the streets of Arizona with no one behind the wheel (a small pool of specially screened people can call them up like they would an Uber). If all goes well, they may expand to more cities later this year (more on this below). But it’s a hard problem, and progress has been slow.
Companies continue to invest despite the setbacks because self-driving cars, when they happen, will change a lot for the world — and make their creators lots of money.
Many consumers will want to upgrade. Imagine being able to read or doze off during your morning drive to work or on long car trips. It also seems likely that taxi and ride-sharing companies will offer self-driving cars, rather than paying drivers (in fact, companies like Uber are betting on it). Self-driving cars should also make a huge difference for Americans with disabilities, many of whom can’t get a driver’s license and have trouble getting to work, the store, and doctor’s appointments.
Experts disagree on whether self-driving cars will change anything fundamental about car ownership in America. Some argue that people won’t need to own a car if they can order one on their phone and get a timely robot ride anywhere.
Others have pointed out that people generally still own a car even in areas with good ride-share coverage and that self-driving cars might not be any different. Polls suggest that most Americans don’t want to be driven to work by a self-driving car — but that might change fast once such cars actually exist. Gallup polling on this question found a small share (9 percent) of Americans who’d get such a car right away, with a larger contingent (38 percent) saying they’d wait a while, and half holding steadfast that they’d never use one.
Over time, our infrastructure will likely change to make it easier for self-driving cars to navigate, and in fact some researchers have argued that we won’t have widespread self-driving cars until we’ve made major changes to our streets to make it easier to communicate information to those cars. That would be expensive and require nationwide coordination, so it seems likely that it would follow the widespread introduction of self-driving cars rather than precede it.
Almost every major car manufacturer has at least tested the waters with self-driving car research. But some are much more serious about it than others.
There are two core statistics useful for evaluating how advanced a self-driving car program is. One is how many miles it has driven. That’s a proxy for how much training data the company has, and how much investment it has poured into getting its cars on the road.
The other is disengagements — moments when a human driver has to take over because the computer couldn’t handle a situation — per mile driven. Most companies don’t share these statistics, but the state of California requires that they be reported, and so California’s statistics are the best peek into how various companies are doing.
On both fronts, Google’s sister company Waymo is the clear leader. Waymo just announced 20 million miles driven overall, most of those not in California. In 2018, Waymo drove 1.2 million miles in California, with 0.09 disengagements every 1,000 miles. Coming in second is General Motors’ Cruise, with about half a million miles and 0.19 disengagements per 1,000 miles. (Cruise argues that, as they test their cars on San Francisco’s difficult streets, these numbers are even more impressive than they look.)
Those two companies are well ahead of everyone else in both miles driven and disengagements in the state of California. While that’s only a limited snapshot of their efforts, most experts consider them the leading programs in general.
March 18, 2018, was the first time a self-driving car ran down a pedestrian. An Uber car with a safety driver behind the wheel hit and killed Elaine Herzberg, a 49-year-old woman walking her bicycle across the street in Tempe, Arizona.
The incident was a reminder that self-driving car technology still had a long way to go. Some people were quick to point out that humans frequently kill other humans while driving, and that even if self-driving cars are much safer than humans, there will be some fatal incidents with self-driving cars. That’s true as far as it goes. But it misses a key point. Human driving produces one fatal accident in every 100 million miles driven. Waymo, the leader in miles driven, just reached 20 million miles driven. They haven’t had a fatal accident yet, but given the number of miles they’ve driven, it’s simply far too soon to prove that they’re as safe as or safer than a human driver.
Uber hasn’t driven nearly as many miles and has had a fatal incident. The company doesn’t release specific figures, but its filings for its IPO last year said that it had driven “millions” of miles. It’s hard to tell exactly without specific numbers, but it’s fair to wonder if Uber’s driving record is much worse than a human’s.
Furthermore, a review of Herzberg’s death suggests that a lot of preventable mistakes were made. The accident report by the National Transportation Safety Board, released December 2019, found that the near-range cameras and the ultrasonic sensors “were not in use at the time of the crash.”
Additionally, the system was having such a problem with false alarms — detecting dangerous situations when none existed — that it had been programmed with “a one-second period during which the ADS [automated driving system] suppresses planned braking while the (1) system verifies the nature of the detected hazard and calculates an alternative path, or (2) vehicle operator takes control of the vehicle,” according to the NTSB report. So even when the car detected the hazard, it didn’t brake — which could have made the collision avoidable or much less deadly — but instead continued with exactly what it was doing for a full second.
The system was designed to assume that pedestrians would never cross except at a crosswalk, so when one did cross without using one, it failed to identify her. Even worse, when the system was unclear on whether an object was a bicycle or not (as it was with Herzberg), it was unable to retain any information about how the object was moving. The system sensed her presence six full seconds before the impact — and yet did nothing (except possibly braking in the last two-tenths of a second) before colliding with her at deadly speed.
Those are avoidable failures.
Hopefully, they won’t happen again. Uber pulled its cars off the road in response, returning to self-driving car trials a year later with a drastically changed program. “We’ve implemented key safety improvements from both safety reviews, shared our learnings with the larger self-driving industry, and accepted the NTSB’s recommendation to implement a Safety Management System, which is underway today,” Nat Beuse, Uber’s self-driving cars head of safety, told Vox in a statement in response to a request for comment. “As we look ahead to the future, we’ll continue to keep safety at the center of every decision we make.”
As I’ve written before, getting good self-driving cars on the road can save hundreds of thousands of lives. But it takes a lot of engineering work to get the cars good enough to be lifesaving.
Some advocates have argued that self-driving cars will be good for the environment. They claim that it might reduce car trips by making car ownership unnecessary and transitioning society to a model where most people don’t own a car and just call for one when they need one.
In addition, others have argued that human drivers drive in a wasteful way — braking hard, accelerating hard, idling the engine, all of which use up fuel — which a computer could avoid.
But as self-driving cars have inched closer to reality, most of these claimed benefits have started to look less likely.
There’s not much evidence that computers are dramatically more fuel-efficient drivers than humans. There’s one small study suggesting adaptive cruise control improves efficiency a little (5-7 percent), but there’s little else beyond that. Furthermore, researchers have examined the effects of more fuel-efficient cars on miles traveled and found that, under many circumstances, people drive more when cars get more fuel-efficient — so self-driving cars having higher fuel efficiency might not mean that they produce lower emissions.
One study attempting to estimate the effects of self-driving cars on car use behavior simulated a family having a self-driving car by paying for them to have a free chauffeur for a week, and telling them to treat the free chauffeur service the way they’d treat having a car that could drive itself.
The result? They went on a lot more car trips.
It’s still possible that some big transition to a lower-driving world will happen. A study of one week of driving habits isn’t enough to settle the question for sure. The researchers who conducted that study are preparing future studies and it’s possible that those comparisons will turn up more encouraging results.
The above few sections might inspire some pessimism, but there’s plenty of reason to be excited about self-driving cars. They will likely make life easier for elderly and disabled people who cannot safely drive. They might provide better, safer, and cheaper options for people currently forced to own a car to get anywhere. Additional research and development will make them safer — and once the kinks are worked out, there is a likelihood that self-driving cars will be safer than human-driven ones.
In a sense, we’re in an awkward transition moment where we want self-driving cars but aren’t yet in a situation where they’ll be an uncomplicated positive.
Research and development is proceeding anyway, mostly because self-driving cars will probably be a goldmine for the first company to get them on the road. They’ll likely be able to establish themselves in ride-sharing, taxi, and trucking markets while competitors are still struggling to catch up, and then they’ll benefit from the additional miles traveled to further improve their cars.
It’s not uncommon for a technology to be dangerous and barely worth it when it is first invented, only to eventually be refined into a valuable part of modern life. The first airplanes were dangerous and commercially useless, but we improved things dramatically from there.
There is no federal law regarding self-driving cars. Much of the action on policy has largely taken place at the state level. And the laws surrounding self-driving cars vary enormously by state, with 29 states having passed legislation.
Self-driving car development mostly happens in the states that have been friendliest to it — especially California and Arizona — and it’s easy to envision some states banning self-driving cars long after they become commonplace in other states, especially if the safety case for them isn’t a slam-dunk.
When self-driving cars were first proposed, I heard a lot of worries that regulators would unnecessarily delay their implementation. By 2016, it was obvious that hadn’t come to pass. Indeed, in some cases regulators may have been too permissive — for example, in light of Uber pulling its cars and instituting new safety procedures, it seems that the vehicle that killed Elaine Herzberg probably shouldn’t have been on the road at all.
Policy might also shape whether self-driving cars are good or bad for the environment. With high taxes on gasoline, for example, the social costs of carbon emissions could be reflected in the price of using self-driving cars — and the money can be spent on climate adaptation and clean energy. But right now, our transportation policy doesn’t do much of anything about the social costs of driving, and that’s a problem that will only get worse if self-driving cars put more people on the road.
In some senses, we’ve been “close” on self-driving cars for years now. Waymo is doing test runs with no one behind the wheel in Arizona, which they’ve been doing since 2017. Cruise delayed the 2019 launch of their autonomous taxi service, but they think it might happen in 2020. Earlier this year they unveiled a car with no steering wheel … and no timetable for when it’ll be available for sale. Tesla’s periodic software updates make its Autopilot highway self-driving work better, but it remains well short of full self-driving.
There are certainly skeptics. Recently, the CEO of Volkswagen said that fully self-driving cars might “never happen.”
That might be an overly harsh forecast, considering the progress that’s been made. But it is exasperatingly difficult to get a good estimate of how long until self-driving cars happen for real for the typical American, both because no one knows for sure and because companies have incentives to publicize optimistic estimates. The companies boast about their progress but don’t publish their mishaps. Timelines slip, and the change in plans is often publicly acknowledged only long after it’s become obvious that the deadline can’t be met.
At the same time, companies hesitate to actually put their cars on the road when there’s any chance they aren’t ready. They are well aware that killing someone, as Uber did, is not only horrible but also probably spells doom for their business. So there’s ample incentive to say optimistic things and not actually launch.
It’s not hard to imagine their arrival later this year, at least in sufficiently limited contexts; it’s also not hard to imagine that the deadlines will be pushed out again for another three or four years.
Self-driving cars are on their way. They’re closer than they were a year ago. When they’ll actually get here is anyone’s guess.
Sign up for the Future Perfect newsletter and we’ll send you a roundup of ideas and solutions for tackling the world’s biggest challenges — and how to get better at doing good.
Future Perfect is funded in part by individual contributions, grants, and sponsorships. Learn more here.