Despite years of development, truly autonomous technologies such as self-driving vehicles and delivery drones continue to remain elusive, largely due to safety and reliability concerns.
Meanwhile, thanks to advances in artificial intelligence, neural networks—a type of machine learning algorithm—have become an indispensable component of modern autonomous systems, especially in their perception capabilities. In autonomous vehicles, neural networks use data from sensors to map the environment and recognize objects such as traffic signs.
However, these machine learning components greatly increase the complexity of the overall autonomous system, making it much more difficult to design the system and analyze its behavior.
Xiangru Xu
That’s why Xiangru Xu, an assistant professor of mechanical engineering, will use his National Science Foundation CAREER Award to establish a theoretical framework and design control algorithms to ensure the safety of autonomous systems with machine learning components. He is focused on safety-critical systems in which any safety failure will break the product or cause human injury.
“Results from surveys show that the American public is very skeptical of autonomous systems,” Xu says. “But, in the future, if there were a self-driving car that came with provable performance guarantees for safe and reliable operation on our streets, it could boost the public’s trust in this technology, which is a very important factor in determining the role of autonomous systems in our future society.”
He says there are many fundamental problems in machine learning that he will need to consider as he designs control algorithms. For one, neural networks are very sensitive, so there’s a risk that a small change to the input can lead to an incorrect result for image recognition. If there’s a random sticker affixed to a stop sign, for instance, that small alteration could cause a neural network to mistake the sign for something else, potentially causing a self-driving vehicle to miss it entirely.
Xu is leveraging his expertise in control theory to tackle this challenge.
“A neural network is basically a function that maps input data to output data, and the areas of controls and optimization have a lot of principled methods for analyzing complicated dynamics and functions,” Xu says. “This expertise will allow me to look at the system in a holistic way so we can design safe trajectories more efficiently. Ultimately, I expect to be able to provide insight on how to design better autonomous systems.”
As part of the educational and public outreach effort associated with his CAREER Award, Xu is developing a new active learning-based course for UW-Madison students focused on safety control in robotics. He will also help develop a summer workshop to introduce high school students to robotics and controls. And he plans to participate in outreach events like Engineering EXPO and the Wisconsin Science Festival to help increase public literacy, awareness and trust in safety-related technologies for autonomous systems.
“I think it’s important to educate the public about autonomous technology, like autonomous vehicles, because people are skeptical of things they don’t understand,” he says. “It’s not only a technical challenge but also a social acceptance problem, so I’m excited to do outreach and help inform the public.”
Xu’s NSF CAREER Award provides $603,643 of support over five years.
Featured image caption: Members of the UW Autonomous & Resilient Controls Laboratory (ARC Lab), directed by Assistant Professor Xiangru Xu, test a drone. Credit: Joel Hallberg.
ME Faculty Video – Assistant Professor Xiangru Xu