“Messy and confused” is Xintao Yan’s summation of the two-lane roundabout just south of I-94. The U-M civil and environmental engineering PhD student is one of the thousands of people who drive through it regularly—in his case, to get to Costco. And he realized that the ten-year-old roundabout’s difficulty made it ideal for training autonomous vehicles.

The problem with teaching autonomous vehicles anything is collecting enough data for them to make extremely well-informed decisions instantly. So Yan and Henry Liu, the U-M civil engineering prof who directs Mcity, the U-M–led public-private mobility research partnership, developed SAFE TEST—the Safe AI Framework for Trustworthy Edge Scenario Tests. In late 2021, they installed sensors on light poles around the intersection to observe how drivers negotiate it.

Xintao Yan says the “messy and confused” roundabout is “a very good demonstration case for the AI technology.” | Photo: Mark Bialek

“We use all the data to build the simulation model to basically try to replicate what is happening in the real world,” Yan explains. “Normally we will drive hundreds of thousands of miles until you encounter safety-critical events. This is too time-consuming and not efficient, so we developed some testing methodology to accelerate this process to make sure it can still guarantee the accuracy but reduce the testing miles.”

The goal is to let autonomous vehicle manufacturers “efficiently and accurately evaluate the safety performance” of their systems, Yan says. “The real-world traffic environment is very, very complicated, because human drivers are pretty diverse and their human behaviors are very hard to model. We have been working on this for years, and we think there’s still a long way to go.” But with lots of data and advancements in AI technologies, he predicts it “could be done in years rather than decades.”

“The roundabout is a very good demonstration case for the AI technology,” Yan concludes. “This could potentially improve the traffic autonomous vehicle and potentially improve traffic safety [and] efficiency.”

According to state records, the intersection averaged nineteen crashes annually before the roundabout went in ten years ago and 111 annually after—and that includes the 2020 pandemic year’s atypically low sixty-nine. But Pittsfield Township police chief Matt Harshberger says it’s nonetheless both more efficient and safer than when the intersection was controlled by stoplights.

“In normal situations, when everything is open and traffic is flowing properly, it gets a little congested at rush hour,” Harshberger says. “But nothing compared to what it did before. And that’s when traffic volumes were much less than what we have now.” And despite the overall increase in accidents, he says, it improved safety by “minimizing or eliminating the likelihood of a serious injury crash.”

The state stats show that angle crashes jumped from an average of three a year to fifty-six, and sideswipes from two to thirty-four. But as Harshberger notes, those are the crashes “you wouldn’t typically see injuries with.”

Drivers clearly still find the roundabout difficult. “You would think after almost ten years that people would become aware,” Harshberger comments. He suspects it may take “multiple decades” for people to get used to driving multilane roundabouts. Still the chief believes “that we’ll get crashes below a hundred. I’m hoping down the road aways that we’ll see it get down below eighty.”

Harshberger says he doesn’t have an “educated opinion” on autonomous vehicles navigating roundabouts. “But I was also thinking, ‘Wouldn’t it be exciting if we got to a point where you could, especially through this roundabout, but any and all roundabouts, have an autonomous vehicle manage and navigate it for you!’”