Physics-Informed, Data-Driven Dynamic Run-Time Assurance using Online Reachability

EM&A proposes a Phase II project to enhance to Technology Readiness Level (TRL) 6 the proof-of-concept runtime assurance (RTA) technology that it created during the Phase I project. The method uses fast, real-time reachability computations to provide dynamic boundaries of operation for the advanced controller and a decision module with stored command sequences and dynamic run-time checks to enable use of unverified controllers. Our Black Box Simplex with Online Reachability (BBSOR) RTA system aims to let the advanced controller act most of the time but applies safety control action when its reachability and safety checks find that the advanced controller would result in unsafe trajectories.   Intelligent, fully autonomous air, ground, surface, underwater, and space vehicles and stationary systems can enable performance, accuracy, repeatability, and safety levels that are either impossible or too costly to accomplish with their manned counterparts. A critical challenge for intelligent autonomy systems such as those based on reinforcement learning (RL) is that they must guarantee their correct performance in all scenarios, but design-time verification for autonomous systems is often intractable. In standard RTA, an advanced, not fully verified controller is designed to operate at most times and a safety controller to replace it when needed to ensure safety.   Our proposed approach enables expansion of the portion of the state space where the advanced, higher performing controller can operate. In Phase I, our team demonstrated our BBSOR RTA method and software on simple problems including aircraft formation flight examples that used RL for the advanced controller and simple models of the aircraft dynamics. Our online reachability features could only handle linear dynamics.   In Phase II, we will adapt our method to black- and gray-box models of system dynamics. We will develop a framework for gray-box aircraft 6-DOF dynamics models. We will populate them based on simulation data and use them for controller training. These models will be of higher dimension than the original nonlinear models but will be equivalent in the system-theoretic sense. We will also explore training controllers using nonlinear simulations directly. The main benefit of linear models is that they are fast and compatible with many system-theoretic and computational tools We will also modify our fast reachability algorithm and efficient safety checking as necessary be compatible with the new models.   To enable software-only and hardware-in-the-loop demonstrations, we will convert our code from MATLAB to software C and C++.  We will integrate it into a modular software architecture. We will then evolve the software functionality. When the software is mature, we will develop an approach for implementing it in embedded hardware. That will culminate in a real-time, hardware-in-the loop demonstration using high-fidelity flight simulation of aircraft flight dynamics.