UToledo Researchers Developing AI-driven ‘Adaptive Adversaries’ to Make Army Vehicle Training Smarter, Safer

UToledo Researchers Developing AI-driven 'Adaptive Adversaries' to Make Army Vehicle Training Smarter, Safer

Researchers at The University of Toledo have secured new federal funding to help the U.S. Army rethink how it trains and tests autonomous ground vehicles and prepares the operators who oversee them. The project uses artificial intelligence that learns and adapts in real time rather than following a script.

The project, known as A3GENT, for Army Adaptive Adversary Generation Technology, is sponsored by the U.S. Army's DEVCOM Ground Vehicle Systems Center (GVSC) in Warren, Michigan, the Army's lead organization for research, development and engineering of military ground vehicle systems. It is contracted through the University of Michigan's Automotive Research Center (ARC).

"This project hits two of the strategic focus areas in our UToledo 'research without boundaries' strategy: leveraging artificial intelligence and advanced computing in defense, national security and autonomous systems," said Dr. Grace M. Bochenek, UToledo's vice president for research and innovation. "When you think about unmanned systems and national security, this project touches both."

Replacing scripted opponents with AI that learns

In current military training simulations, autonomous vehicles and their operators encounter "opposing forces" that behave in known, scripted ways. It's the digital equivalent of a well-studied playbook. The UToledo research is building AI agents that instead learn from an operator's own moves and respond with unpredictable, real-time tactics.

The work at UToledo, one of two university partners on the project, is led by Bochenek and Dr. Ahmad Y. Javaid, interim chair and associate professor in the Department of Electrical Engineering and Computer Science and the project's co-principal investigator and technical lead.

"If our systems only train against scripted scenarios, they're not going to learn how to handle the unexpected," Bochenek said. "We're building adversaries that adapt as they go, so the systems and the people operating them are ready for the curveballs, not just the plays they've already seen."

UToledo is partnering with the University of Central Florida, which is helping to build realistic, simulated environments using its Project GL simulation platform. The team plans to incorporate real-world terrain and geographic data, including data from conflict zones such as Ukraine and Afghanistan, to make the training environments as realistic as possible. The partnership combines UToledo's automotive and AI expertise with UCF's simulation and training capabilities.

The broader goal, Javaid said, is to use simulation to uncover an autonomous system's vulnerabilities long before it's ever deployed in the field.

"The real purpose of doing this in simulation is to make these systems trustworthy and reliable before they're deployed," Javaid said. "So, it's important to develop scenarios and simulations that mimic real-world situations and terrain and that include unexpected behavior."

That matters because systems tested only against predictable opponents may perform well in familiar scenarios but struggle when they encounter something new. By continuously challenging autonomous systems with an adversary that keeps learning, the project aims to catch and fix those weaknesses in simulation, well before a vehicle or system is ever built or fielded.

"This isn't about deploying a weapon"

Bochenek said the project is a direct example of what the University means when it talks about trustworthy, use-inspired AI, which is technology developed for a specific, real-world purpose rather than in the abstract.

"This isn't about using AI to deploy a weapon," Bochenek said. "It's about adding intelligence into ground vehicle systems in a way that's designed to reduce a soldier's cognitive load, the same way the car you drive today has intelligence built in to help keep you safe."

Running thousands of simulated iterations also lets researchers and the Army test design decisions before they become costly, Bochenek said, allowing far more experimentation than was possible when human testers had to be involved in every scenario.

Early recognition, and a pipeline for students

The team's work on the project has already been recognized nationally. UToledo researchers won a best paper award at the most recent Ground Vehicle Systems Engineering and Technology Symposium (GVSETS) for their work on intelligently deploying simulated forces based on terrain.

The project is also creating hands-on research opportunities for students. The research currently supports two Ph.D. students and one master's student at UToledo, with plans to add a research scientist and a software developer to the team as the two-year project continues.

"This ties into the UToledo Advantage by offering students work on cutting-edge research that supports the kind of experiential, real-world learning that is a hallmark of a UToledo education today," Bochenek said. "The University's distinctive 'applied learning for all' approach translates into career success."

What's next: dual-use potential

Beyond defense applications, Bochenek and Javaid point to commercial potential for the same underlying technology. One example could be adapting the technology to train commercial semi-truck drivers using simulators that respond dynamically instead of following a fixed script.

More broadly, the research team says the work being done to make autonomous systems more trustworthy and adaptable could extend to other safety-critical fields as well, including self-driving vehicles and aerospace.