SBIR-STTR Award

This is a prototype system that will take raw observation data, detect and characterize maneuvers, and use reinforcement learning to understand and react to evasive RSO behaviors in near real time. This will provide a framework to evaluate autonomous behavior strategies, such a safety, effectiveness, and robustness to manipulation. This work will enable surveillance operators to distinguish between uncertainty in RSO statethe more directly measurable quantities, and uncertainty in underlying intent.The proposed system will also enable assessment beyond what behavior is occurring, and speak to why in an operational sense, and quantitatively show the uncertainty of such categorizations. Understanding how intent is reflected in behavior and state space will also allow the detection of any technological gaps, the unobservable behavior space region where an RSO could hide its behavior or intent. Framing the surveillance problem as a multiplayer game enables the use of Information Space Receding Horizon Control (ISRHC) to discover likely intent, as well as other reinforcement learning techniques, and allows for longer lookahead windows. This may allow for the identification of a variety of unexpected events, such as sudden break-ups, or aggressively evasive maneuvers.