It’s 2020. Where are our self-driving cars?
In the age of AI advances, self-driving cars turned out to be harder than people expected.
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 your 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.
1) How exactly do self-driving cars work?
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 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 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.
2) Why is it taking longer than expected to get self-driving cars on the road?
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.