We speed, tailgate, blow red lights and drift between lanes while juggling cellphones and spilling coffee. Yet the deeply flawed human driver is still superior to a computerized one.
Self-driving cars piloted by software are clearly on the way, and once perfected, they could dramatically improve auto safety and transform other elements of transportation. But they’re further off than a lot of breathless headlines suggest. “We don’t have any autonomous car yet where the car, on average, is better than a person,” Gill Pratt, CEO of the Toyota Research Institute, said recently at a conference held at MIT. “People are really good at driving safely. These are the standards artificial intelligence has to beat.”
News snippets on self-driving cars make it sound as if we’ll be chauffeured by robots in just a few years. Nearly every automaker is developing autonomous vehicle technology, and Google (GOOGL) publishes regular updates on its own self-driving car program. Upstart electric-car maker Tesla (TSLA) touts an “autopilot” feature it’s developing, and some enthusiasts are waiting for Apple (APPL) to jump into the fray and revolutionize automobiles.
The demise of the driver has been greatly exaggerated, however. “The Google car has been overhyped,” MIT engineering professor John Leonard said at the conference, “and the technology is often misunderstood.” Autonomous technology developed by Google and others is very good, for example, at handling the rote and predictable tasks of driving, such as staying in a lane and speeding up or slowing down in a flow of traffic. A few such autonomous systems are already rolling into dealerships, such as the Cadillac “supercruise” feature that should be available next year.
The trickiest challenge for self-driving cars -- which may still take decades to hone -- is getting computers to process unpredictable situations as quickly and effectively as a human being can. At the MIT conference, sponsored by the New England Motor Press Association, Leonard showed some dashcam video footage of his own car navigating Boston traffic, which highlighted powers of the human brain computers can’t yet match. In one scene, the car approaches an intersection where the traffic light is red but a police officer is waving cars through. A couple blocks later, the opposite occurs: the traffic light is green but a police officer is signaling cars to stop.
Most people know intuitively that the police officer trumps the traffic light, and they’ll do what the cop directs them to do. That’s very hard to build into a computer algorithm, which would basically follow the general rule of doing what the traffic light says. Police presence can be programmed into the algo, but it’s still hard for software to tell if somebody waving their arms at an intersection is a law-enforcement officer or a loon.
Drivers make these types of judgment calls all the time – usually accurately – without even noticing, drawing on years of knowledge stored in our brains. “When a human driver sees a mother and child in a stroller on the sidewalk, our belief in whether they’ll cross road is completely different than if we see a teenage male,” says Pratt. We know the mother is likely to move carefully and deliberately, whereas the teenager could dart into the road with no warning. Computer sensors can easily detect a person on the sidewalk, but software isn’t yet smart enough to guess what they might do based on what kind of people they are.
Judgments about the surrounding environment often determine how we drive. To make a left turn from a dead stop, across a lane of oncoming traffic, into a flow of traffic going the other way, for instance, sometimes requires edging the nose of your car out, or signaling other drivers to let you in. Those are instances in which a bit of aggressive driving may be needed to get where you’re going. A self-driving car programmed only to pull out when there are safe distances in every direction could sit for hours before finding an opening, enraging every car behind and choking the flow of vehicles.
Engineers are addressing all these problems, while rolling out intermediate technologies that will give cars a bit of autonomy, while the driver stays firmly in control. Most big automakers, for instance, have vowed to install emergency braking systems on all their cars by 2022, making standard a feature that automatically slows the car or slams on the brakes if sensors detect a collision is about to happen. Another feature would sense when drivers nod off and trigger an alert that wakes them up.
Nobody’s really sure how long it will take before there’s a car that can transport passengers on its own, without any need for a driver. But automotive engineers like to quote Bill Gates, who famously said, “We always overestimate the change that will occur in the next two years and underestimate the change that will occur in the next 10.” So, someday.
Rick Newman’s latest book is Liberty for All: A Manifesto for Reclaiming Financial and Political Freedom. Follow him on Twitter: @rickjnewman.