�鶹ӰԺ

Afrah Ghedira, Lorenzzo Mantovani and Mark Stephenson

A team of three aerospace PhD students were awarded second place in an autonomous satellite systems competition.

Afrah Ghedira, Lorenzzo Mantovani and Mark Stephenson competed in the organized by the American Institute of Aeronautics and Astronautics.

The goal? To design an automated approach for a simulated maneuvering satellite during non-cooperative space operations. It is a well-known game theory application for orbital dynamics called a “lady, bandit, guard problem.”

“Your satellite is the bandit. You’re trying to get as close as possible to the lady satellite, but the guard satellite is trying to keep you away,” Stephenson said.

Critically, the satellites had to act completely on their own. Once programmed and unleashed, no human intervention was allowed.

All three students are studying under Hanspeter Schaub, distinguished professor and chair of the Ann and H.J. Smead Department of Aerospace Engineering Sciences. Schaub’s lab focuses on complex orbital mechanics and created an open-source program for space guidance, estimation and control solutions.

Reinforcement Learning

Using the software, the team developed algorithms, based on the rules of the game, to model both their satellite and potential guard behaviors.

“We had to train our agent against a wide range of possible guards,” Mantovani said. “It’s the hardest part of any game. You need to have good behavior for yourself and also the other side. It’d be easy to play chess with someone if you knew what their move would be.”

Through the reinforcement learning aspect, they ran thousands of simulations for the agent to improve performance.

Unexpected Results

During those repeat simulations, their agent developed an unexpected approach. The premise of the game was to rendezvous for an extended period with the lady satellite, as though for maintenance, but the formal rules were more general, requiring only a momentary approach.

“Our agent learned kind of a funny strategy,” Ghedira said. “If you’re near the lady for too long, it gives lots of opportunity for the guard to come close. So, our agent would fake out the guard, fly far away and then apply full thrust to zoom close by the lady really fast and keep flying away.”

Through the reinforcement learning, their model realized something the team members had not.

“Initially we thought something was wrong, but based on the scoring, getting close to the guard was more penalized than getting close to the lady was rewarded, so it was a pretty reasonable strategy,” Ghedira said.

Fifty-two teams registered for the AIAA competition, which used the Kerbal Space Program, a computer game highly regarded for accuracy in simulating challenging orbital dynamics situations.

“This is a realistic environment. With any sort of autonomy and reinforcement learning project, the big question is can it work in an environment you don’t control. In our lab, we had a lot of the tooling to approach this problem in a really interesting way,” Mantovani said.

Future View

The use of autonomy in orbiting satellites is still in its infancy, and Stephenson said the competition was a good opportunity to experiment in a new domain.

“I’ve been looking at industry research proposals that are adjacent to this kind of problem and am glad to get experience on this type of thing. The industry is just opening its eyes to autonomy and what we can do if we enable these technologies for space like is being done on Earth in robotics and self-driving cars,” Stephenson said.

The competition was held during the 2026 AIAA SciTech Conference in Orlando, Florida, on Jan. 12-16.