Almost half of Americans will hop in their cars for a Thanksgiving trip this year. But if you were being very precise—if you were a team of Massachusetts of Technology researchers who study human-machine interactions—you wouldn’t say that all those Americans are "driving," exactly. The new driver assistance systems on the market—like Tesla's’s Autopilot, Volvo's’s Pilot Assist, and Jaguar Land Rover’s InControl Driver Assistance—mean that some of those travelers are doing an entirely new thing, participating in a novel, fluid dance. The human handles the wheel in some situations, and the machine handles it in others: changing lanes, parking, monitoring blind spots, warning when the car is about to crash. Call it…piloting? Shepherding? Conducting? We might need a new word.
Fully autonomous cars won't swarm the roads en masse for decades, and in the meantime, we'll have these semiautonomous systems. And scientists need to figure out how humans interact with them. Well, actually, the first thing to know is that most humans don't: Preliminary research by the Insurance Institute of Highway Safety noted that, of nearly 1,000 semiautonomous vehicles studied, 49 percent had their systems turned off. The warnings were annoying, owners said.
If you could actually watch those drivers—sit inside the car and eyeball them while they drive—you might get a better understanding of how these systems are helpful and how they're not. Maybe drivers find one of kind of warning sound frustrating, but another (a bloop instead of a bleep?) helpful. Maybe they get more comfortable with the system over time, or stay mystified even as the odometer rolls over. That spying would be really helpful for people who build and design semiautonomous systems; for those who want to regulate them; and for those expected to evaluate the risks of using these systems, like insurers.
That's why MIT researchers are announcing this week a gigantic effort to collect data on how human drivers work with their driver assistance systems. They outfitted the cars of Boston-area Tesla, Volvo, and Range Rover drivers with cameras and sensors to capture how humans cooperate with the new technology. They want to understand what parts of these systems are actually helping people—keeping them from crashing for example—and what parts aren't.
If designers knew more about how drivers' eyes and bodies move in specific driving situations—when they're making a left turn, when it's raining—they can design machines that better respond to people and get them where they need to go safely.
“Our goal is to try to understand how the act of driving is beginning to transform from one where the human has primary oversight responsibility to one where the human is actively engaged in a robotic interaction with the vehicle,” says Bryan Reimer, part of the MIT team who studies interactions between humans and robots. “The transformation from manual control to automation is probably more complex than any aspect of vehicle development to date.” Driving, he says, hasn't changed this much since Ford rolled out the Model T. A century after the Tin Lizzy made driving widely accessible, humans are suddenly being asked to share the task with computer systems.
The concept of peeping in on drivers is not a new one, research-wise. Scientists have used cameras to conduct so-called naturalistic driving studies for over a decade, studying, for example, what sorts of factors lead up to crash events. Are most caused by mechanical failures? Texting? Looking away from the road to adjust the radio, or yell at the kids in the back?
Using cameras and sensors to actually watch real-life driving is a more reliable way to collect info on driver behavior than simulations in the lab, or even reading police reports. Naturalistic driving studies, can, for example, parse the differences between teenage, middle age, and senior drivers, providing justification for policies like drivers' training education. (I've asked many scientists involved in this kind of research whether the presence of the camera skews the results—aren't people more careful if they know they're being watched? They always say something like, "If you've seen the recordings of the stuff I've seen, you'd know our lab rats forget about the cameras real quick.")
For this study, funded by a consortium of carmakers, insurers, and consumer watchdog groups, the MIT researchers equip vehicles with at least one onboard computer, multiple webcams, a GPS unit, lidar, and bluetooth sensors. Some of these cars are owned by their drivers; in these cases, the researchers will observe the vehicles for a year. They're interested in understanding, long-term, how drivers adapt to driver assistance technologies. Do they start depending on the machines more frequently? With more confidence? For different driving tasks? With more, ahem, irresponsibility? (Witness the Tesla owners who crawl into their backseats with Autopilot on.)
Scientists will also lend out cars for a month at a time. In this case, they want to figure out how drivers react to a totally new driver assistance system. How long does it take them to learn to use it? Do they always use it correctly—say, only on approved highways? Are there some elements of the tech they never understand at all? The answers to these questions could help engineers build safer cars.
So far the MIT team has collected 21 months’, 7,146 days’, 78 drivers’, 275,559 miles’, and 3.48 billion video frames’ worth of data. As they gather more information, the researchers will analyze it with the same sorts of machine learning techniques used to train self-driving cars to pick up on wide-scale patterns. The footage in many past driving studies was categorized by actual people, but the scale of this work—the scientists hope to collect many years of driving data—means these researchers have to turn to computers to get where they want to go. Just like drivers.
No, you’re not crazy—a whole lot of people are using their cellphones while driving.
It would be great to banish distracted drivers forever. But it will never work that way, researchers say. To crack distracted driving, figure out how to design cars that work with humans.
If all this semiautonomous talk is freaking you out, here’s a more calming thought: Stick your kids in a self-driving school bus.
