Autonomous systems are becoming more and more impressive, from smart homes that learn a family’s routine to delivery robots that can safely cross an intersection on their own.
However, the type of control systems that run these gadgets aren’t powerful enough to run next-generation autonomous systems, which will need to handle changing environments and complex uncertainties, all in real time while guaranteeing safety—like an autonomous car headed for an interstate pileup or a delivery drone caught in a flock of geese.
But Jeremy Coulson, who started as an assistant professor in electrical and computer engineering at the University of Wisconsin-Madison in the spring semester of 2023, is finding ways to make these smart systems even smarter. He specializes in data-driven control, a framework that enables autonomous and robotic systems to learn from their past and make intelligent decisions in real time.
“Control theory has been built on the use of mathematical models describing exactly how things work,” Coulson says. “But now, as these systems become more and more complex—for example, autonomous driving or complex robots—you don’t necessarily have access to these models or the expert knowledge needed to understand how these systems work under the hood. So you want to try to learn how to control these systems without such models. And the way you do that is through data.”
Coulson grew up near Toronto in Ontario, Canada, and received a bachelor’s degree in mechanical engineering and applied mathematics and a master’s degree in mathematics and engineering from Queen’s University in Kingston, Ontario. He then traveled to ETH Zurich in Switzerland, where he recently completed his PhD in the Automatic Control Lab, focusing on data-driven control.
Much of Coulson’s work developing algorithms is mathematical. However, he says it’s exciting to see how that work makes its way down the pipeline, eventually controlling real systems. “In my PhD, we developed this nice theory for data-driven control and developed a very powerful algorithm based on it,” he says. “Then we actually implemented this algorithm on real physical machines, like quad copters, autonomous excavators and power grids, where you actually get to see the theory put into practice.”
At UW-Madison, he says he wants to continue and expand this work. In the past few years, these sorts of control problems have been approached from two different communities: the control theory side, which Coulson is a part of, and the computer science and machine learning/artificial intelligence community. He hopes to bring these groups together and also involve the robotics community at UW-Madison. “Together we can tackle new problems, open up new research directions and introduce disruptive ideas that might not be possible without such collaborations,” he says. “Madison has a dynamic environment which is conducive to these types of collaborations.”
Coulson is also excited to continue teaching at UW-Madison. In Zurich, he was nominated for a teaching award for an undergraduate course involving quad copters. During the COVID-19 pandemic, he redesigned the course from the bottom up under the concept: if the students cannot come to the lab, the lab can come to their homes! He modified the drones so that the students could take them home and perform hands-on experiments. Students worked in teams developing control systems that would allow the drones to fly autonomously. Coulson hopes to bring similar hands-on experiences to students in Wisconsin.
Featured image caption: Jeremy Coulson specializes in data-driven control. Credit: Joel Hallberg.