Safe Autonomy from Runtime Reachability for Learning-Enabled Systems
Abstract
This project will develop theory and algorithms for enabling safe autonomy of learning-enabled systems via reachability analysis conducted at runtime. An emerging challenge is to certify or guarantee correct behavior of nonlinear control systems with learning-enabled components. Such systems are generally not amenable to traditional offline analysis methods due to their scale, complexity, or adaptable design. An alternative to offline safety verification in these cases is to enforce at runtime that the system does not evolve into unsafe conditions, which requires predicting possible future system states and intervening at the controller if needed to maintain safety, an approach referred to as runtime assurance. Reachability analysis is concerned with computing possible future system states of a dynamical system and is thus a critical subcomponent in classical offline verification and to runtime assurance. However, most reachability methods are generally slow and do not scale well, making them ill-suited for runtime computations on complex systems. This project aims especially to leverage and significantly extend recent results in efficient and scalable reachability analysis for nonlinear systems to develop a comprehensive theory of runtime assurance for learning-enabled systems. In particular, the project will develop a theory of mixed monotone systems with respect to partial orders defined by general cones and seek to generalize connections between monotone and contractive systems. These advances will enable efficient computation of high-confidence reachable sets for learning-enabled systems that are then incorporated in barrier-based runtime assurance mechanisms. The project will also extend these methods to hybrid statespaces appropriate for models that combine discrete logical components with a dynamic system.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 29, 2024
- Source ID
- FA95502310303
Entities
People
- Samuel Coogan
Organizations
- Air Force Office of Scientific Research
- Georgia Tech Research Corporation
- United States Air Force