Researchers debut new study for safe learning-enabled systems
Grainger Engineering Professor Naira Hovakimyan is leading a new study that seeks to develop safe learning architecture for autonomous systems. Her NSF-funded study, “SLES: Guaranteed Tubes for Safe Learning across Autonomy Architectures,” is a collaboration with Shenlong Wang from UIUC’s Department of Computer Science and electrical engineering professor Xiaofeng Wang at the University of South Carolina. The latter Wang is a former postdoc in Hovakimyan’s lab.
“This award is very prestigious and we’re very excited,” said Hovakimyan, who is the W. Grafton and Lillian B. Wilkins Professor of Mechanical Science and Engineering.
Safe learning-enabled systems (SLES) represent autonomous, or artificial intelligence-driven, systems that can adapt to extreme events, environmental hazards, and irregular behaviors to perform tasks safely and effectively. Safe operation in developing technologies such as self-driven cars and airplanes depends on a viable SLES framework. Hovakimyan’s study seeks to develop a data-enabled Simplex (i.e., DeSimplex) architecture that will ensure end-to-end safety by transferring control authority between high-performance learning-based components and high-assurance solutions.
“I firmly believe that building learning-enabled autonomous systems with safety guarantees will be a key milestone toward a safer and more efficient future in transportation,” said Shenlong Wang, whose research will focus on uncertainty quantification in robot perception.
This work builds on Hovakimyan’s extensive body of research by leveraging her celebrated L1 adaptive control system. This new study also connects to her collaboration through the Center for Autonomous Vehicles in Air Transportation Engineering (AVIATE) to investigate safe learning architectures that can be certified through the Federal Aviation Administration for unmanned aerial vehicles (UAVs).
The team proposes to develop novel methods for three desirable SLES characteristics: on-policy, closed-loop learning to boost performance; reliable uncertainty quantification to provide data-driven adaptability; and verifiable observability and controllability. With year one of the award dedicated to development and planning, the team hopes to experiment with electric cars and quadrotors by year three. However, applications for their framework extend far beyond unmanned vehicles.
“This project endeavors to forge a vital link between learning-based methodologies and traditional approaches, creating a harmonious framework that seamlessly integrates both, such that a system can enjoy both intelligence and safety,” Xiaofeng Wang said. “The far-reaching impact of this project extends to a multitude of safety-critical domains including transportation systems, space exploration, health care and power infrastructure, among others.”