Brian Plancher, assistant professor of computer science, will work with the Toyota Research Institute on a long-term project to help make human interactive driving and autonomous vehicles safer and more responsive, particularly in inclement conditions.
Plancher is among a new cohort of researchers announced May 18 who will receive three-year grants through TRI's University Research Program to work on technologies that advance artificial intelligence, robotics, advanced driving, and material science. The program, TRI, which is headquartered in Los Altos, Calif., with an office in Cambridge, Mass., pairs university researchers with a Toyota co-investigator "to ensure that fundamental research and real-world application evolve together."
Plancher's team at Dartmouth will work with TRI Senior Research Scientist Thomas Lew in using advanced computing tools to design new technologies that will enhance the ability of autonomous vehicles to make reliable maneuverability and reaction decisions, particularly under challenging road conditions.
"Current autonomous vehicles are often restricted to conservative maneuvers because standard algorithms fail at the physical limits of handling, particularly in challenging conditions like wet or icy roads," Plancher says. "If you hit black ice and you're sliding out, this is a moment where it would be great if the car could help you avoid a crash."
Skidding out on black ice, it turns out, is very similar to drifting or racing a car. So Plancher and his team are developing robust new algorithms that they will deploy to autonomously control racecars on a test track.
Plancher’s team at Dartmouth will work with TRI to design safer technologies for next-generation vehicles using advanced computing tools. (Photo By Katie Lenhart)
"The goal is to build a controller that can enable the racecar to autonomously drift and race through challenging conditions like water puddles," says Plancher. Such a system could then be adapted into passenger cars to reliably support human drivers as they navigate unexpected, dangerous conditions.
What makes these conditions difficult to design controls for is that the friction between the tires and the road is highly variable and challenging to estimate perfectly in real-time based on the precipitation on the road. This makes it hard to model the physics accurately for a computer to solve.
Using graphics processing units, specialized hardware that can handle many complex calculations in parallel, the team is building numerical solvers that compute an optimal response that is inherently robust to the wide range of friction values—on snow, ice, water, or pavement—that the car might encounter.
"Historically, we just didn't have enough computational power available on vehicles, or the high-performance software that can run on such vehicles, to attempt this. We are directly addressing this challenge," says Plancher. While the control software is being designed for this scenario, it can also be applied to other similar problems, such as flying drones or humanoid robots, he says.
"We are trying to build algorithms that are inherently safe, by construction," says Plancher. "Traffic fatalities are still a huge issue today, so it's really important that we try to make our driving systems safer."