High Assurance Run-Time Monitoring Architecture for Autonomous Control
Yi Chou, Aditya Zutshi, Matthew Clark · AIAA Scitech 2021 Forum · 2021
View Video Presentation: https://doi.org/10.2514/6.2021-0790.vid We present run-time assurance approach SARAS to provide safe recovery when using complex learning enabled autonomous components in safety critical systems. SARAS is based on the Simplex architecture, that was designed to enable run-time safety by implementing a fallback mechanism which uses trusted components. SARAS achieves the goal of system safety by associating conservatively designed recovery control actions and switching to them in the case of a failure, where a failure is observed via a set run-time monitors. Existing Simplex based approaches use a single recovery controller for each complex controller. In such an arrangement, safety often comes at the expense of performance as the safe controller limits the operating region of the complex controller. In order to balance safety and performance of learning enabled controllers, SARAS uses a fine grained approach by splitting the system controller into modes and control primitives and letting the engineer associate multiple recovery actions and a switching scheme to dynamically select the best action to maintain safety while minimally affecting the performance. We present our results on the F-16 flight system consisting of a 6-DOF 13 dimensional non-linear plant model and a two layer controller: an outer-loop autopilot coupled with a tracking inner-loop.