Title: A Universal Framework for Generalized Run Time Assurance with JAX Automatic Differentiation

Abstract: With the rise of increasingly complex autonomous systems powered by black box AI models, there is a growing need for Run Time Assurance (RTA) systems that provide online safety filtering to untrusted primary controller output. Currently, research in RTA tends to be ad hoc and inflexible, diminishing collaboration and the pace of innovation. The Safe Autonomy Run Time Assurance Framework presented in this paper provides a standardized interface for RTA modules and a set of universal implementations of constraint-based RTA capable of providing safety assurance given arbitrary dynamical systems and constraints. Built around JAX, this framework leverages automatic differentiation to populate advanced optimization based RTA methods minimizing user effort and error. To validate the feasibility of this framework, a simulation of a multi-agent spacecraft inspection problem is shown with safety constraints on position and velocity.