Phase II Amount
$1,699,936
Neuroevolutionary Optimization of Aggregation-GNNs for Scalable Swarm Tactics (NO-AnGSST) enables the learning of rich sets of complex behaviors over large heterogeneous swarms of low-SWaP hardware by capitalizing on aggregation graph neural networks (GNNs) to learn scalable control policies that are robust to both unit attrition and sparse, dynamic communications environments --each a familiar thorn in the sides of traditional control and multiagent learning methods. NO-AnGSST's architecture achieves these goals by fusing: GNN structures from SSCI's prior work in multivehicle autonomous kill chains via Event Horizon Processing (EHP); comms- and uncertainty-aware modifications enabled by the recent work of Tolstaya et al.; and a novel, hybrid neuroevolutionary approach for training GNNs that strives to avoid the major pitfalls of gradient-based optimizers, while admitting significant parallelization opportunities and a more accessible expansion of the behavior library. Phase 1 will leverage SSCI's multi-sim reinforcement learning suite (Bourbon) to linearize our algorithms research and facilitate all subsequent development. Bourbon wraps an extended OpenAI Gym interface around a variety of simulations, ranging from highly-efficient in-house simulations to Airsim/Unreal's realistic flight dynamics, high visual fidelity environments, and tremendous feature richness. The Phase 2 demonstration will transition these learned behaviors onto a modest swarm of UAVs in a live flight test, in conjunction with a significant expansion of NO-AnGSST's simulation results, targeting swarm sizes of up to two hundred mixed UAVs and UGVs, with a focus on counter-swarming and flocking based transit use cases in urban environments.
