Lyapunov-Based Deep Learning for Safe and Reliable Autonomy

Modern autonomous systems increasingly rely on deep learning for perception, prediction, and control, yet ensuring safety and reliability remains a fundamental challenge. This talk presents a principled framework that integrates Lyapunov stability theory with deep neural networks to provide formal guarantees of convergence, robustness, and safety during online learning and control. By shaping adaptation dynamics through Lyapunov functions, analogous to enforcing dissipation in physical systems, learning evolves according to stability and passivity principles rather than unconstrained gradient updates. This energy-consistent perspective unifies deep learning and adaptive control, enabling real-time system identification and control with provable stability under uncertainty. The resulting framework supports structure-preserving autonomy in complex robotic environments, bridging physics-based modeling and data-driven learning.

Dr. Omkar Sudhir Patil is a Research Scientist at the University of Florida whose work spans nonlinear control, robotics, and learning-enabled autonomy. His research integrates Lyapunov-based adaptive control, machine learning, and nonlinear dynamics to develop provably stable, safe, and real-time learning methods for uncertain robotic systems. He has pioneered extensions of Lyapunov theory to deep neural networks, enabling online system identification and control with formal guarantees in complex environments. Dr. Patil received his Ph.D. (2023) and M.S. (2022) in Mechanical Engineering from the University of Florida, and his B.Tech. in Production and Industrial Engineering from the Indian Institute of Technology (IIT) Delhi in 2018.