AI-Enabled Command and Control for Formation-Based Layered Protection

Project #5.29 begins 2026.

Protecting the Army’s units and operations in all domains (land, air, sea and space and cyberspace), known as Formation-Based Layered Protection (FBLP) [1] is key to mission success. FBLP spans across the following axes: (1) data-driven decision making (which refers to evaluating the effectiveness of advanced data analytics and artificial intelligence (AI) to provide real-time situational awareness for rapid, informed decisions), (2) expanded maneuver (which refers to enhancing the ability to maneuver and engage adversaries across domains), and (3) forging seamless interoperability (which refers to refining the ability of allied forces to operate together seamlessly across all domains).

The project will develop a FBLP that embraces AI-enabled Command-and-Control (C2) to design and evaluate the performance of teams of heterogeneous UxVs (Unmanned Ground and/or Aerial Vehicles; assets), with different capabilities in terms of sensing, communication, computation, for given missions. Our exemplar case study is the design of an ISR mission, for which we consider attributes (e.g., coverage area, mission time), returns (e.g., quality of surveillance, risk to assets), capabilities and constraints of the assets (computational, sensing, communication) as well as that the operational conditions and their associated uncertainty is continually changing.

The first objective is to develop methods for quantifying the uncertainty of the predictions of AI/ML components (such as NNs) that operate in-the-loop of the UxV system, so that this uncertainty can be considered for accurate trade-off analysis at the design phase of a UxV system (for FBLP). Such in-the-loop NNs could be models predicting the motion of detected adversaries, or the class/type of detected objects in the domain of interest. If the predictions are accompanied with high uncertainty, then the NN models should be refined online, and with acceptable risk to the UxV team. To this end, the UxV team should collect and incorporate new, informative data to adapt the NN weights. The second objective is to develop C2 methods to reduce the uncertainty of the NN outputs online, by deciding how to collect new data to refine the NN predictions. These two goals are captured in the research objectives below.

Research Question and Objective 1: How can the UxV team reason about the validity of the predictions of the AI/ML components, typically NNs, and how it can safeguard against unsafe NN outputs?

Research Question and Objective 2: How should the UxV team collect new data for reducing uncertainty of the NN predictions, even at the expense of increased individual risk or attrition for some agents?

The uncertainty quantification of AI/ML models, and the subsequent uncertainty reduction will be directly applicable to offline design and tradeoff analysis for FBLP, and to online adaptable deployment and execution. We will develop techniques and tools for optimizing the use of heterogeneous autonomous agents (UxVs) for given missions under complex classes of uncertainty compared to classical tools and approaches, such as those encountered in predictive NN models.

5.29