Tap the brakes: Self-driving vehicle technology is 'absolutely not ready,' says robotics researcher

When the U.S. Senate Commerce Committee held a hearing on Capitol Hill this month to assess America's readiness for the arrival of self-driving cars, it summoned a who's who of industry executives: Chris Urmson, director of Google's self-driving car project, plus executives from Delphi, General Motors and Lyft, all of which are racing to bring self-driving cars to market.

Then, as if to splash cold water on their ambitions, the committee called Missy Cummings, an engineering professor and human-factors expert at Duke University who argued self-driving cars are "absolutely not ready for widespread deployment."

Cummings, 49, who was one of the Navy's first female fighter pilots from 1988 to 1999 and managed a $100 million Navy program to build a sensor-laden robotic helicopter, is director of Duke's Humans and Autonomy Lab.

As a professor, she is leading a National Science Foundation-funded study of how pedestrians interact with self-driving cars. Cummings spoke with Staff Reporter Gabe Nelson on March 18, three days after visiting Capitol Hill.

Q: How good are humans at working with robots?

A: That's a big question. It depends what the machine is trying to do, whether the people have lots of training -- as in the case of aviation -- and the complexity of the environment. Humans certainly can adapt to a high-complexity environment. The question is how much training a person needs to do it.

How complex is driving?

Driving is one of the most complex domains. It's even more complex than aviation.

Even though you're moving in three dimensions in aviation, road environments are a lot denser. When you're in the sky there aren't many planes near you, but there can be a lot going on near a road in an urban setting: cars, people, bicycles.

And the people are significantly less trained. It doesn't take much training to get a driver's license in this country, and we're not going to move to a society where you have to go to school for six months just to operate a driverless car. We're going to need to be sure everyone from ages 16 to 96 can operate these things.

What can we learn from aviation in bringing self-driving cars to market?

I think the auto industry could learn a lot from how airlines and airplane manufacturers worked to automate their planes, and tested them to be sure that they would work in all conditions.

We would have never allowed people to fly in airplanes when the industry was still trying to figure out automated landing. The planes had to be tested, and manufacturers had to prove they could land under all sorts of different conditions.

I believe that before we take drastic steps such as taking steering wheels out of cars, the car manufacturers need to prove that a human will never need to intervene. We're simply not going to go to a car with no steering wheel overnight. We will get there eventually -- I just think it's not going to happen as quickly as Google might want.

Can humans be trusted to snack or nap in the car and take the wheel when needed? 

No. 

If the human is a problem, why not go straight to fully autonomous cars? 

That would be my first choice -- for everyone to go into an autonomous car. We're just not ready, and I haven't seen any test data to suggest we are. 

So how can the industry solve that problem? 

We have to make sure the automation is good enough to at least get the car into a safe position. The car can never assume that when it needs to hand off control, the human will be ready at that instant. 

Context is important. If a traffic policeman is gesturing and a car can't interpret the gesture, it could slow down and vibrate the seat and ask a human to take over. It's not critical that a human take over in that case. If they don't, the car can stop. 

But if a car is going 65 mph and a car is having trouble deciding whether to get off the interstate, it can't just say: "Three, two, one, now take over." A car would need to say: "Click this button if you're ready to take over," and if [the driver doesn't], the car will need to be able to come to a safe stop in some way. 

Does a similar problem exist in aviation? 

In aviation we have something called mode confusion. Sometimes pilots don't know what mode they're in because it's not clearly communicated who is doing what. 

There can be a mismatch between what the car thinks a human is capable of and what the human can do, and vice versa. So it's really, really important that humans understand what an automated system is going to do, and what it's likely to do next. 

Should we be designing self-driving cars to mimic human drivers or to drive in a uniform, predictable, robotlike way? 

They should be as predictable as possible. Of course, so should we. I don't think it's a big deal for humans to learn how to drive in and around autonomous cars. 

What about other people besides drivers? Will pedestrians get used to them? 

Check back with me in three months. Right now we're running a study at Duke around pedestrian behavior and autonomous cars. We're putting real cars and real people on the road together, and asking people to cross the road. 

How should self-driving cars be regulated? 

[The National Highway Traffic Safety Administration] needs to start setting guidelines for testing and set certain levels of certification so companies know what's expected of them, and know what is a safe enough system under certain conditions. 

The most important thing from my perspective is that we do move down this path. Self-driving cars are the future. There's no question that is where we should go. The question is: How do we get there? And how fast do we need to get there?