Towards rigorous neural control

Abstract

Learning-based and data-driven approaches are becoming an indispensable part of robotic systems. Control and planning components based on neural networks challenge existing methods for ensuring reliability and safety of these systems. By taking a numerical and statistical perspective on synthesis and verification, we believe it is possible to still prove strong properties for highly nonlinear systems with highly nonlinear control laws. Interestingly, we often need to make use of inductive certificates that are themselves neural networks. I will survey some of our ongoing work in these directions towards rigorous neural control for nonlinear systems with "provably approximate" safety and reliability guarantees.