Graph-based Collaborative Autonomy for Intelligent Agents

Barron Associates, Inc. has teamed with the Virginia Tech Sanghani Center for Artificial Intelligence and Data Analytics (VT) and BlackSwift Technologies (BST) to develop the Advanced Graph-enabled Network Technology for Collaborative Autonomous Agents (AGENTCA) solution for distributed collaborative control of heterogeneous swarms of UAS, UGV, USV, and UUV agents. Previous approaches to multi-vehicle swarming control typically rely on centralized algorithmic formulations, which can deliver optimal control performance but scale poorly when used with large numbers of agents. Poor scalability manifests in excessive computational and communication throughput requirements that are not supportable in practical applications. In contrast, AGENTCA offers an efficient decentralized approach designed specifically for scalability to swarms of hundreds or more agents. In AGENTCA, each agent computes its own control solution using an efficient Graph Neural Network (GNN) algorithm that can be run on modest computational hardware and provides near-optimal control performance. The formulation makes it possible to operate very large, autonomous swarms of small UxV with limited SWAP capabilities in contested areas with limited network connectivity. In AGENTCA, a GNN is trained off-line to implement a desired control algorithm, such as formation flight, intruder pursuit, or distributed area coverage. The resulting GNN is then copied to each agent, who then uses the GNN to compute their own control solution based on what they know about other agents in the swarm. AGENTCA offers three primary innovations: (1) agents require only local information about the locations of their neighbors, rather than global information about the entire swarm; (2) swarm composition and topology can change dynamically during a mission, allowing the swarm to smoothly adapt to the loss of individual agents or the addition of new ones; and (3) a modular training environment is included, which makes it straightforward to incorporate new swarm behaviors and to customize existing ones. This project builds on a highly successful Phase I effort, which demonstrated the ability to train GNN controllers that perform a variety of functions and highlighted key robustness properties of the approach. The Phase II effort will expand the available GNN training options, add new swarm behaviors, and explore a wider range of candidate operational scenarios. It will expand the GNN approach to incorporate heterogeneous mixtures of both UAV and UGV assets and conduct high-fidelity simulations involving hundreds of these agents in a single swarm. Finally, the effort will integrate the GNN solution onto commercially available UAV and UGV platforms, demonstrating the technology in a series of flight tests involving swarms of 15 to 20 agents in realistic outdoor environments. In parallel to this activity, the company will aggressively pursue a carefully crafted transition and commercialization plan.