SOAR TECHNOLOGY, LLC — Department of Defense SBIR Phase I: DTRA192-002

Description

SoarTech proposes DeepCounter: a system that automatically learns UAS and C-UAS behaviors across multiple tasks using DRL within a 3D simulation environment. Our system addresses several gaps in applying state of the art DRL to learn C-UAS and UAS behaviors. First, our existing DRL framework extends the state of the art to work with variable numbers of heterogenous agents. Our work in Phase I will extend the framework to scale across multiple tasks and multiple agents. Second, we will design efficient state and action representations which improve data efficiency, allowing DeepCounter to deliver behaviors faster than conventional DRL approaches. Third, we will use flexible exploration algorithms (agnostic to simulation type) which allow DeepCounter to learn complex behaviors with an infrequent reward. On the Phase I effort, we will develop a 3D simulation environment with C-UAS scenarios, modify our existing DRL framework to simultaneously learn multiple tasks in the new simulation, learn both UAS and C-UAS behaviors within the simulation, and evaluate learned behaviors across multiple dimensions.