Next-generation technologies—including self-driving vehicles, intelligent robots, personalized healthcare devices, and smart infrastructure—are already reshaping industries and becoming increasingly woven into daily life. But to be successful, these advances, and many others, need safe, reliable and flexible control systems.
Jeremy Coulson, the Mark and Jenny Brandemuehl Assistant Professor in electrical and computer engineering at the University of Wisconsin-Madison, will use a five-year National Science Foundation-funded CAREER award to develop novel data-driven control systems that can handle these new, complex systems safely and efficiently.
Control systems are everywhere. These software systems control relatively simple processes, like a thermostat that turns on a heater when it’s cold, to more advanced systems, like an insulin pump that needs to constantly regulate blood glucose, or a prosthetic limb adjusting to its user’s gait.
Traditionally, these control systems are based on mathematical models and physics that guide each of their decisions. That’s fine when the underlying system is easy to model. But when it comes to really complex systems—such as an autonomous taxi navigating a dark, slick road through unpredictable traffic—traditional methods hit a wall.
“These systems we are dealing with are becoming more and more complex. It is becoming nearly impossible to do the modeling step by hand,” says Coulson. “My research in this CAREER project is to skip that modeling step entirely and use historical data from sensor measurements to design controllers directly, without ever needing to write down equations for how these systems behave.”
To accomplish this, Coulson will first use mathematical tools to develop principled uncertainty modelling and robust control methods directly from data sets, with a focus on ensuring safety and performance. The next step will be incorporating streaming data, like information coming from an autonomous vehicle’s various sensors, into this framework. The goal is to create an algorithm capable of continuous learning and adaptation in complex, changing environments. Once this algorithm is complete, Coulson plans to validate the algorithm through high-fidelity simulations. He’ll partner with colleagues at UW-Madison to test the algorithms in robotics experiments.
At the end of the project, Coulson aims to have a powerful, adaptable, data-driven control system that can be implemented in all sorts of automated technologies. “It’s a general-purpose system; that’s the beauty of the theory, it can be applied so widely that it’s really just up to individuals to decide if it’s applicable in their specific cases,” he says.
In fact, Coulson says early versions of his data-driven algorithms have already been applied in healthcare, power systems, quadrupedal walking robots and even growing cherry tomatoes in a greenhouse. He says he’s excited to see what other systems adapt his framework as he improves his algorithm. “There’s nothing more practical than a good theory,” he says. “If you build the foundation and you build the theory, then the applications are limitless.”
Photo by Joel Hallberg