Feasibility of Neural Control Certificates for Rare Events

Istari, partnering with the Reliable Autonomous Systems Lab at MIT (REALM), proposes to study the feasibility of adapting the autonomous control certification algorithms being developed at MIT for self-driving cars to apply to rare events for autonomous flight control algorithms in aircraft and integrate into Istari’s commercially available digital engineering platform that is used to develop new aircraft digitally before they are produced in the physical world.  The Reliable Autonomous Systems Lab at MIT (REALM) at MIT has developed methods to certify that control algorithms for self-driving cars are guaranteed safe based upon the hardware and software models provided.  The work has centered around three types of certificate functions from control theory: Lyapunov Functions, Barrier Functions, and Contraction Metrics.  Thus far they have been proven to work on aircraft within narrow operational limits (e.g. normal flight) and achieved local optimization.  For this research we want to see if these control theory certificate functions can be extended to apply to global optimization and specifically applied to rare events that pose high risk for flight (e.g. at the edge of the flight envelope).  These certificate functions will be integrated as modules into the overall Istari Digital Engineering platform (likely via APIs) to allow for new aircraft developed to be evaluated to see if they satisfy these control certificates.