A3GENT: Army Adaptive Adversary Generation Technology

Principal Investigators

• Ghaith Rabadi (PI), Sean Mondesire (Co-PI), University of Central Florida (UCF)
• Grace Bochenek (Co-PI), Ahmad Javaid (Co-PI), University of Toledo (UToledo)

Students

• Jasen Blevins, Jason Nguyen, UCF
• Sydni Paniagua-Fittro, Mohammed Hamza Sahaf, Md Sakib Galib Sourav, UToledo

Government

• Victor Paul, U.S. Army GVSC

Industry

• TBD

Project #2.A127 began in 2025.

Modern multi-domain military operations require training and evaluation of environments that mirror the complexity and unpredictability of real-world adversaries. Current simulation systems are a critical gap, as they primarily rely on scripted, predictable Red Forces (REDFOR), which limits their realism and effectiveness. This deficiency directly impacts the development of next-generation AI systems like Virtual Crew Members (VCMs) and Digital Advisors (DAs), as these systems are not rigorously challenged during their evaluation phase, leading to potentially brittle performance in the field.

The primary objective of this project is to develop ADTs (Adaptive Digital Twins) that can simulate intelligent and adaptive adversaries in real-time, providing a robust evaluation environment for VCMs and DAs. The project’s scope has not changed from the original proposal. The research is structured around three fundamental questions:

Research Task 1 (RT1): Adaptive AI Algorithms: How can we synthesize RL, DL, and GANs to create an adversary that learns optimal tactics while also generating novel, unpredictable behaviors to prevent VCMs from overfitting to specific strategies?

Research Task (RT2): Real-Time Data Integration: How can we design a data pipeline to integrate live intelligence feeds into the simulation, allowing the environment to automatically create and modify scenarios that remain challenging and relevant?

Research Task 3 (RT3): Multidimensional Threat & Game Theory: How can we use multi-agent systems and game theory to model a coordinated REDFOR team that makes strategic, long-term decisions and executes cross-domain threats?

Publications:

• Soykan, B., Rabadi, G., Bochenek, G. and Paul, V.J. (2026) “Multi-Objective Reward Shaping For Autonomous Ground Vehicle Control in Adversarial Simulations”, The Annual Modeling and Simulation Conference, May 4-7, 2026, Orlando, Florida
• Soykan, B., Rabadi, G., Bochenek, G. “CombatEnv (2026): A Gymnasium-Compliant Simulation Framework for Adversarial Stress-Testing of Autonomous Control Logic”, The Annual Modeling and Simulation Conference, May 4-7, 2026, Orlando, Florida
• Soykan, B., Rabadi, G., Bochenek, G., And Paul, V. J. (2026) Adaptive Adversarial Autonomy: Scaling Stackelberg Equilibria to Continuous Ground Combat Domains Via Latent-Context Hypernetworks, selected as Best Paper, 2026 NDIA Michigan Chapter Ground Vehicle Systems Engineering and Technology Symposium, Applied Artificial Intelligence (AAI) Technical Session, AUG. 11-13, 2026 - Novi, Michigan
• Soykan, B., Rabadi, G., Bochenek, G., And Paul, V. J. (2026). Physics-Constrained Procedural Generation of Synthetic Environments for The Robust Validation of Autonomous Maneuver and Logistics, 2026 NDIA Michigan Chapter Ground Vehicle Systems Engineering and Technology Symposium, MSPV Technical Session, AUG. 11-13, 2026 - Novi, Michigan

#2.A127