Scalable Physics-Aware Runtime Assurance for Autonomy

Autonomous systems and semi-autonomous systems are integral to a number of domains like aerospace, transportation, and robotics. Market pull to deploy new features at an unprecedented time to market scale, requires experimenting with and deploying untrusted and partially tested technologies. However, this is at odds with the need to avoid catastrophic failures especially in safety critical systems. Runtime assurance (RTA) addresses this tension. The  idea is to introduce a  decision module that somehow chooses between a well-tested safety controller (SC) and the experimental untrusted controller (UC), assuring safety of the overall system while also allowing experimentation with the new untrusted technology where and when possible. RTA is seen as a key enabler for autonomy in safety-critical aerospace systems. In Phase I of this project, we introduced a novel formulation of the RTA problem where the goal is to identify a decision logic that guarantees hard safety constraints and at the same time optimizes performance metrics. We should how this problem can be reduced to one where the goal is simply to optimize a single reward (instead of the competing goals of safety and performance) using a technique called reward shaping. Having reduced the RTA problem to a simple optimization problem, a number of off-the-shelf techniques can be brought bear, including approaches like reinforcement learning which can find optimal policies in a model-free manner. We demonstrated the effectiveness of this approach a number of design problems that arise in the context of defense-relevant aerospace applications, such as, formation flight and wingman aircraft following a leader while assuring safety with respect to ground and air collision requirements. In Phase II of this project, we plan to extend this novel approach to accomplish three main thrusts. (a) Algorithm development of the optimal RTA design with hard constraints that go beyond safety, including liveness and temporal logic constraints, and stochastic and nondeterministic plant models. (b) Development and implementation of RTA systems for multi-agent scenarios including distributed and de-centralized communication and control across heterogeneous teams of agents that is robust to system failures. (c) Evaluation of new algorithms and design strategies via more realistic simulation of fighter-class aircraft models and commercial grade aircraft simulators such as X-Plane.