Studying the Safety of Self-Driving Cars

Apr 3, 2019

Self-driving cars appear to be on the horizon. But will that mean increased safety on roadways? Three researchers in Texas Tech’s Department of Psychological Sciences, one who is now at Rice University, studied how well drivers in self-driving cars stayed vigilant to roadway hazards.

Assistant professor Eric Greenlee, the study’s lead author, says the research is important in gauging drivers’ ability to act as a safeguard in automated vehicles.

“If there’s a hazard that comes up, it’s ultimately the driver’s responsibility to be paying attention, step in and take control. My background is in human, automation interaction like this and what a lot of that research has shown is that people are really bad at doing this. They’re really bad at paying attention for long periods of time, bad at detection really rare and unpredictable events like an automation failure. So, I thought it was pretty unlikely that they would be able to do what they’re expected to do to act as a fail-safe for the automation,” Greenlee says.

Sixty Texas Tech students each spent 40 minutes seated in a simulator. Inside a closet-sized space, participants had a steering wheel, and brake and accelerator pedals. Each viewed a two-lane roadway on three 27-inch monitors that spanned 135 degrees around them.

Graduate student David Newton, a co-author on the study, says the trio of researchers was trying to figure out if drivers’ vigilance decreased due to a complacency or if the vigilance declined because the subjects felt overwhelmed.

Vigilance decrement is when there’s an increase in error rates as an effect of time-on-task during tedious monitoring tasks.

“We designed this study to manipulate the demands of the monitoring tasks. You can see by which direction it went, based on how the demands were manipulated, what was the mechanism. So, we found that by increasing the demands, both the temporal demands, which are the frequency of the events over time, and the spacial demands which are how predictable the locations of the potential events was—by increasing those demands it actually caused a decline in performance greater than lower demands. So that demonstrated that what was causing the detriment was the high attentional demands of the task,” he explained.

Participants were told to monitor the roadway for vehicles that stopped dangerously at intersections, creating a hazard by intruding into the driver’s lane, which in the simulator constituted a hazard not detected by the automated car. They also had to discern between vehicles that were safely stopped and dangerously stopped at intersections.

The longer the students sat in the simulator, the worse their performance. That surprised Greenlee.

“One of the things that surprised me most from this line of research is then how big the vigilance detriment has been, so how much performance suffers over time. We’ve seen a range of things depending on how difficult a task is, how complex it is. But anywhere between 10 and 40 percent, we’ve seen drop. People are good to begin with, you know 80 or 90 percent, they’re detecting what they’re supposed to, they’re seeing these hazards that might constitute an automation failure, but within 30 or 40 minutes they’re missing quite a few more.”

Pat DeLucia, a former professor at Texas Tech who’s now at Rice University, was another of the study’s co-authors. Newton says he believes the increase in self-driving will be exponential in the next five years.

“I think what will become interesting is when we start mixing cars with this technology, with older cars, because one of the systems that these automated cars use, or could potentially use is vehicle to vehicle communication. They could see, based on this network, if there’s another vehicle that could potentially be involved in a collision, but if older vehicles are on the road, that network is not really there,” Newton says.

Another study about self-driving cars is ahead for Greenlee and Newton. They want to compare vigilance issues between self-driving vehicles and those that are manually driven.

“I think that’s a concern, so that’s why we do are best to simulate what the task would actually be like so that it’s psychologically valid. Typically the simulator studies align pretty closely to the way people would behave in real life, if not better,” he says.