Mary “Missy” Cummings at the 2017 GeekWire Summit. (Photo by Dan DeLong for GeekWire)
Fully automated cars are looked at by many in the tech community as a near certainty, maybe even in the next few years, but a former U.S. Navy fighter pilot and a professor at Duke University says the technology is not quite ready for primetime.
Mary “Missy” Cummings, director of the Humans and Autonomy Laboratory at Duke University, spoke at the GeekWire Summit, where she discussed the limitations on the technology behind driverless cars as well as some of her flight stories.
Cummings’ interest in automated systems goes back to her Navy days, where she was the branch’s first female fighter pilot and one of the first in the military in general. While sitting in the plane waiting for takeoff she and her fellow pilots had to keep their hands up to show they weren’t interfering with the computer, like small children showing parents they aren’t touching anything.
“I am the best of the best, certainly that’s what Top Gun says, and I am not allowed to touch anything because the computer always does a better job, particularly on the takeoff,” Cummings said. “So what does that say about the future and where my career is going?”
That’s when she got into academia, where she focused on autonomous systems. As a professor, Cummings has developed a handy rubric for areas where automation works and where it doesn’t. The short-hand of it is machines don’t deal well with high levels of uncertainty, and those situations are best handled by people.
She cited as an example the time in 2009 when Chesley “Sully” Sullenberger landed US Airways Flight 1549 safely in New York City’s Hudson River in after it struck a flock of Canada Geese shortly after takeoff. That was as high of a degree of uncertainty as there is, and it is unclear if a computer could have made the same correct decisions as the pilots.
It is the uncertainty that makes cars the most difficult form of transportation to automate, Cummings said. Airplanes have been mostly automated for years, with humans stepping in for only three to seven minutes per flight and during emergencies. Many countries around the world have automated trains, with the U.S. as an exception. But cars are an entirely different beast.
“When you’re in an aircraft, there are not 50 other people of questionable intelligence driving closely to you on their cellphones at all times,” Cummings said.
An Uber self-driving car in Pittsburgh. (Uber Photo)
Cummings showed off several examples highlighting the limitations of the systems that power driverless cars. She showed a simple moving van, something people would recognize instantly, and then flashed to a computer vision analysis, which thought the van was actually an amalgamation of several different vehicles, a person and other objects.
What you see vs. what the computer sees. Professor @missy_cummings describes strengths, limits of AI. #GWSummit pic.twitter.com/w5Y0guvtiq
— GeekWire (@geekwire) October 10, 2017
She also cited a recent study where the team put several black and white markers on a stop sign, and that led a vehicle’s computer to read it as a 45 mph sign.
At least for now, Cummings thinks the best solution is for people and machines to work together. An example is what she called the R2D2 program — which had to be changed to R2A2, as George Lucas apparently didn’t like the name. It replaces co-pilots with robots to help people fly planes. That will help with pilot shortages around the world, shifting first officers into the pilot chair.
“My core research is about bringing systems together, letting the humans and computers, the autonomous systems work together as a team,” Cummings said.