SBIR-STTR Award

Toyon Research Corporation proposes the research and development of Artificial Intelligence (AI) to optimize UAS control for non-kinetic effects to counter enemy Integrated Air Defense (IAD) systems. We will focus on a single non-kinetic attack challenge problem; the generation of phantom tracks with small UAS against a modern networked set of threat radars. The problem requires coordination among the UAS team to cast a single coherent phantom track from the perspective of all adversary radar systems. To achieve this, the AI must optimize both the maneuvers of the UAS team and the controls of an onboard EW payload. Our selected approach combines deep reinforcement learning (DRL) with multi-agent training. The combination allows each agent to both learn and master complicated tasks from the experience gained while competing in a simulated environment. In Phase I, Toyon will design red-side IAD and blue-side UAS EW agent models. Toyon will also identify enhancements to its SLAMEM simulator to appropriately model the EW payload and missions of interest. In Phase II, the agents and simulation enhancements will be implemented and the agents will be trained over millions of simulated mission sets using Toyon’s Modular Autonomy Incubator (MAUI) framework and high-performance computing cluster.

Near-peer adversaries with advanced Integrated Air Defense Systems (IADS) have created anti-access area denial (A2AD) zones for US Air Forces. A2AD scenarios continue to push conventional strategic platforms to further stand-off ranges, decreasing their mission effectiveness. Toyon proposes the research and development of Artificial Intelligence (AI) to optimize swarming of Electronic Warfare (EW)-enabled munitions. The hypothesized swarm will employ cooperative maneuvers and non-kinetic effects to counter IADS in A2AD zones. To achieve this, the AI must optimize both the joint maneuvers of the swarm and the parameters of onboard EW payloads. Our approach combines Deep Reinforcement Learning (DRL) with multi-agent training. The combination allows each agent to learn complicated tasks (through deep reinforcement learning) and to master them (through training against an adversarial AI) from the experience gained while competing in a high-fidelity simulated battlefield. In Phase I, a pulse-level EW simulator was developed and integrated with Toyon’s Modular Autonomy Incubator (MAUI) framework to train a proof-of-concept AI to automatically jam an enemy radar (1-v-1). In Phase II, we will develop a preliminary digital twin of a real EW payload being developed by Toyon (a parallel effort) and expand our focus to harder scenarios combining the digital twin, enhanced EW simulator, and realistic threat models. Agent models will be designed for both red-side and blue-side AI agents and large-scale multi-agent training will be performed on a high-performance computing cluster to optimize their underlying policy networks. Our objective is the development and demonstration of novel, robust, and deployable autonomy to support Phase III production of Swarming Electronic Attack Munitions (SEAM).