The ability to correctly respond to a sudden adverse event is critical to ensuring a system’s safety and performance in complex and changing environments. In response, classical control methods largely attempt to design strategies which enable the system to complete its original task even after such an event. However, catastrophic events – such as physical damage through natural disasters or partial adversarial system takeover – may simply render the original task impossible to complete. In that case, design of any strategy that attempts to complete the task is doomed to be unsuccessful. Instead, the controller should recognize the task as impossible to complete, propose an alternative that is certifiably completable given the current knowledge, and formulate a strategy that drives the system to complete this new task. To do so, in this talk I will present the emergent twin frameworks of quantitative resilience and guaranteed reachability. Combining methods of optimal control, online learning, and reachability analysis, these frameworks first compute a set of performance specifications that can be satisfied by all systems consistent with the current partial knowledge, possibly within a time budget. These tasks can then be pursued by online learning and adaptation methods. The talk will consider three scenarios: degradation of control capabilities, loss of authority over a part of the system, and structural change in system dynamics. I will briefly present several applications to autonomous vehicles and infrastructure. Finally, I will identify promising future directions of research, including real-time safety-assured mission planning, resilience of networks, and perception-based task assignment.
Melkior Ornik is an assistant professor in the Department of Aerospace Engineering at the University of Illinois Urbana-Champaign, also affiliated with the Coordinated Science Laboratory, as well as with the Discovery Partners Institute. He received his Ph.D. degree from the University of Toronto in 2017. His research focuses on developing theory and algorithms for control, learning and task planning in autonomous systems that operate in uncertain, changing, or adversarial environments, as well as in scenarios where only limited knowledge of the system is available. His work has been funded by several NASA grants and Department of Defense programs, and he is currently a co-PI in the DPI’s Net Zero Transportation Infrastructure science team.