Phase II year
2022
(last award dollars: 2023)
Phase II Amount
$3,399,987
Robust collaboration is needed between different agents in various Unmanned Aerial/Ground Systems (UxS). ML/AI development has enabled development of a graph neural network (GNN)-based swarm control algorithm. Clostras GNN-Swarm uses a GNN framework for mapping relationships between all members of a UxS swarm, and deep reinforcement learning (DRL) for organizing, training, optimizing, and robustifying swarm behavior towards a specific goal. GNNs allow swarming agents to be gracefully added or dropped from the current swarm graph as GNNs function well with incomplete information. Critically, use of GNN-based graphs allow agents to quantitatively determine the informational validity (or likelihood of noise) of received data in contested RF or communication-poor environments. Current GNN control solutions are based on Laplacian matrices, which are hard coded and not responsive to real-time changes to the agent or graph (for example, a team member is added or lost). A more flexible, robust approach is needed, which easily takes into account adding/dropping agents from a swarm (and complex goals) without significant customization or lengthy, expensive training: Clostras GNN-Swarm. By the end of Phase II Clostra will have our GNN-Swarm algorithm installed and testing in a swarm of UAVs in various outdoor and/or indoor environments.
