Towards a Statistical Learning Theory of Nonlinear Control
Abstract
The proliferation of unmanned aerial vehicles (UAVs) in combat scenarios poses a national security threat- while autonomous aerial defense systems that can respond at scale will allow the U.S. Air Force (USAF) to maintain air-space superiority in the face of this changing battlefield landscape, current technology cannot meet this demand, as the USAF does not have enough pilots to meet its staffing targets for UAVs, and the incident rates for UAVs are several times higher than those for manned aircraft, limiting the current scope of UAV deployment. (Semi-)Autonomous UAVs offer an appealing solution to these challenges. However, in order to achieve high-performing yet robust autonomy in combat airspace, future autonomous UAVs will perform complex tasks under continuously evolving, uncertain, and dynamic conditions, using information gleaned from complex, high-dimensional sensors. Due to this ubiquitous and dynamic uncertainty, feedback control loops incorporating learning-enabled components will be pervasive. While impressive demonstrations of learning-enabled UAVs abound, they lack the strong guarantees of robustness, performance, and safety that USAF applications demand, and are very data-hungry. To address this gap, this proposal develops novel learning-enabled nonlinear control algorithms and corresponding analysis techniques that synergistically integrate tools from stochastic processes, statistical learning theory, and nonlinear control to achieve data-efficient, safe, and robust autonomy. Three complementary challenges are considered- (i) single trajectory learning for nonlinear control systems, which focuses on developing a contemporary learning theory tailored to nonlinear control systems, (ii) representation learning for nonlinear control systems, which initiates a study of representation learning for control through the lens of nonlinear robust control so as to enable rapid adaptation in the face of changing environmental conditions, and (iii) characterizing the fundamental limits of learning to control nonlinear systems, so as to inform how to design nonlinear control systems for which learning-enabled control is provably easy.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 06, 2025
- Source ID
- FA95502410102
Entities
People
- Nikolai Matni
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Pennsylvania