Uncertainty Management Framework for Space-based Autonomy

Uncertainty management within autonomous systems is imperative to improving trust in multi-agent cyber-physical-human systems. These multi-agent systems require not only a robust uncertainty quantification strategy, but also an intelligible representation of uncertainty to human agents to elucidate confidence in the autonomous portions of the multi-agent systems. Here, we propose to develop a general framework for integrating uncertainty management within a variety of missions relevant to NASA. This framework will include key elements of deep learning for autonomous decision-making, quantification of data and model uncertainties, model explainability, and effective representation of system status to human agents. To achieve the goals of this project, CFD Research is partnering with the University of Michigan and Draper Laboratory to transition recent advancements in the areas of deep learning, autonomy, uncertainty quantification, and mission management into an operational setting. In this program, space rendezvous will serve as the primary design reference mission for which to demonstrate the uncertainty management framework. In Phase I, we will perform a proof-of-concept study where perception-based deep learning models are trained to perform pose estimation of a target spacecraft. These models will be initially trained on publicly available spacecraft imagery datasets (e.g., SPEED), then subsequently refined with renders of spacecraft which are of specific interest to NASArsquo;s current objectives (e.g., Gateway outpost). Uncertainty quantification and model explainability approaches will be used to analyze the resulting model predictions. In Phase II, as the technical approach is further refined, we will work with our collaborators at Draper Laboratory to develop an uncertainty management software interface capable of being inserted into their existing workflows for the application of rendezvous with the Gateway.