Artificial Intelligence/Machine Learning,
Phase I

Graph Neural Networks (GNN) for UxS Collaborative Agent Control

Release Date: 11/16/2021
Open Date: 11/30/2021
Application Due Date: 01/04/2022
Close Date: 01/04/2022
Solicitation: 21.4
Topic Number: A214-045
Duration: Up to 6 months
Amount Up To: $250K


The purpose of this topic is to use recent advancements in Artificial Intelligence, specifically, GNN to be able to collaborate between different AI agents, Unmanned Aerial/Ground Systems (UxS). Such collaboration must be demonstrated in Airsim due to its large adaptation; however, other game engines can also be utilized.


The purpose of this topic is to create an AI GNN framework for collaboration between swarming agents. Currently, collaboration is done using Laplacian matrices which can be used to find useful properties of a graph but it has to be hard coded and thus will be not be robust, since many lines of codes are need to program the behavior of each agent/graph. Current methods limit the behavior changes if a team member is added or lost. Since the hard coded method is used, when a new member is added, the user must account for it, which makes it harder to add new member, the same goes for losing a member, thus a new matrix will be needed to be added or the formation will not be robust. Having AI for each member will make the collaboration faster, more robust and will take the need for pre-determined behaviors. Communication, control and collaboration must be dynamic for large numbers of graph. Adjusting intelligently for unforeseen circumstances i.e adding/loosing members. Swarming for defensive and offensive fires will be able to utilize such Artificial Intelligence. This technology can also be utilized by Ballistic Low Drone Engagement (BLADE) and other C-UAS systems where they can communicate with each other and give suggestions to the user. C-swarming will also benefit from this research as it will give a testbed as to how the swarms of the future will look like and what it takes to counter them. If successful, the GNN will be easier to implement thus will make scaling up very easy and efficient thus will reduce the time it takes for the user to pre-program each new agent. It will also reduce the communication time between agents thus making it faster. If successful, having this assessed will also help in defending against future swarms.

Phase I

Phase I will consist of the demoing of communication for graphs/agents and creating the foundation of GNN in Python. It should also include the demoing of swarm control in a game engine of choice for UxVs where they dynamically change behaviors due to obstacles and/or mission goals.

Phase II

Phase II should consist of a continuation of Phase I as well as a GNN for large number of graphs (200+) in a game engine, approximately. This GNN should be implemented on NVIDIA Jetson platforms for a small number of graphs (20+); it could be a mix of ugv/uavs. .

Phase III

Phase III will continue off of Phase I and Phase II work while proceeding into commercialization. This final software should be one that can be implemented to any device on edge where communication, control and collaboration between agents can be achieved for UxVs

For more information, and to submit your full proposal package, visit the DSIP Portal.

  1. Tolstaya, E., Gama, F., Paulos, J., Pappas, G., Kumar, V., & Ribeiro, A. (2021, March 24). Learning decentralized controllers for robot swarms with graph neural networks.
  2. Kallenborn, Z. (2018, October). The Era of the Drone Swarm Is Coming, and We Need to Be Ready for It. Modern War Institute at West Point. Retrieved from
  3. Anh-Duc Dang and Hung M. La and Thang Nguyen and Joachim Horn “Distributed Formation Control for Autonomous Robots in Dynamic Environments” arXiv: preprint arXiv:1705.02017 (2017)
  4. Nguyen, and H. M. La. “Distributed Formation Control of Nonlonolomic Mobile Robots by Bounded Feedback in the Presence of Obstacles.” arXiv preprint arXiv:1704.04566 (2017).
  5. Makiko Okamoto and Maruthi R. Akella. Avoiding the local-minimum problem in multi-agent systems with limited sensing and communication. International Journal of Systems Science, pp 1-10, Oct. 2014.
  6. H. Yang and F. Zhang, “Geometric formation control for autonomous underwater vehicles”, IEEE Intl. Conf. on Robotics and Automation, pp. 4288-4293, May, 2010.
Scroll to Top