A paper posted to arXiv on March 20, 2026, proposes an architecture for an AI-based system to automate Course of Action (CoA) planning in military operations — the process by which commanders develop and evaluate possible plans in response to a situation. The authors are Ji-il Park, Inwook Shim, and Chong Hui Kim.
The problem the paper addresses
The paper’s stated motivation is operational speed. As maneuver speeds increase, surveillance ranges extend, and weapon ranges grow, the authors argue that traditional manned CoA planning becomes harder to sustain at the pace modern environments demand. Multiple countries and defense organizations are known to be developing AI-based CoA systems, but, as the paper states, “due to security restrictions and limited public disclosure, the technical maturity of such systems remains difficult to assess.”
Approach and contribution
Because classified operational systems are not available for study, the paper works from publicly available military doctrine. The authors identify three contributions: introducing relevant military planning doctrines drawn from public sources, mapping applicable AI technologies to each stage of the CoA planning process, and presenting a proposed system architecture for automated CoA planning.
The paper includes 15 figures and 2 tables illustrating the proposed architecture. The contribution is architectural and survey-like rather than empirical — the paper does not describe a working system or report experimental results.
The classification problem as a research constraint
The authors acknowledge directly that the classification of actual military CoA planning systems prevents the kind of comparative research that would normally be possible in a technical field. Research that would proceed by building on or comparing against prior systems cannot do so when those systems are not disclosed. The paper’s response is to treat publicly available doctrine as the ground-truth specification and build from there — a choice that allows the work to be shared and reviewed, at the cost of being disconnected from whatever operational systems are in use. The paper explicitly states this limitation.
Scope
Automated CoA planning sits at the intersection of multi-step planning under uncertainty, autonomous decision-making in adversarial environments, and AI use in high-stakes operational contexts. The paper addresses all three at the architectural level, describing what a system should be able to do and how components might fit together. Whether the proposed architecture matches what actual military CoA systems look like is, by the paper’s own account, not verifiable from public sources.