Robert to research drivers' takeover performance in self-driving cars

On March 18, 2018, an autonomous vehicle (AV) struck and killed a woman in Tempe, Arizona. It was believed to be the first pedestrian death associated with self-driving technology.

There was a driver in the vehicle, but somehow, he did not intervene. Why not? 

Lionel Robert

That is perhaps the key question right now with AVs – and one that Lionel Robert, associate professor at the University of Michigan School of Information (UMSI), along with a team of U-M researchers, will be studying in depth.

He has received a $300,000 grant from MCity, University of Michigan, for the study “Predicting drivers’ takeover readiness and designing adaptive in-vehicle alert system.” Robert is a member of a team of five researchers in the project led by Jessie Yang in Industrial and Operations Engineering (IOE). MCity is housed at the University of Michigan and brings together partners from industry, government and academia to develop a foundation for an ecosystem of connected, automated vehicles.

The three-task study will investigate drivers’ takeover performance in what is called “highly automated driving” (HAD); predict, via computer analysis, drivers’ takeover readiness; and design and evaluate an alert system for AVs that adapts to drivers’ relative ability to take control of an AV when needed – and, hopefully, in time.

Overall, Robert says, “this project is designed to help the AV assess the ability of the driver to take over the driving and to reduce error. To accomplish this, the AV needs to be able to assess the driver’s state and readiness to take over.”

AVs have arrived; nothing will change that. But the relationship between vehicle and driver is a bumpy road right now.

“In highly automated driving, the driver’s role transforms from operator to system supervisor,” says Robert, one of the nation’s top AV experts. “This makes it possible for him to engage in non-driving-related activities.” When the AV reaches a certain “system limit,” the driver must take over.

But, too often, Robert says, “the drivers become increasingly out-of-the-loop” once they start reading, emailing, or just relaxing while the AV does all the work. Once “decoupled” from the operational level of control, drivers’ can have difficulty taking over, especially in situations the AV cannot handle.

At least 25 studies in recent years have examined human factors in AV driving, especially what is called the “optimal takeover-request lead time” for drivers to resume control of the vehicle.

“These studies shed light on the design of in-vehicle alert systems,” Robert says. However, he adds, they don’t take into full consideration the “contextualization and personalization of in-vehicle alert systems, despite the fact that drivers’ takeover performance can be influenced by lots of factors.”

These include drivers’ engagement in non-driving tasks, drivers’ ages and the “driving scenario complexity,” as Robert describes it.
 
In other words, all AVs may be alike; all human drivers are not. As a result, Robert says, the in-vehicle alert system should provide adaptive instead of static alerts.

In order to design such alert systems, the researchers will conduct human subject experiments to study drivers’ takeover performance, while manipulating non-driving task difficulty and driving scenario complexity; physiological data also will be collected.

The team will then develop a computational model to predict drivers’ takeover readiness – in real time. Then, they will design an in-vehicle alert system which will act adaptively in response to that readiness.

Sheryl James, PR Specialist

Posted March 30, 2018