- November 12, 2026
- By John Tucker
During a 25-year career as a naval aviator that included several “Spain to Maine” flights, Donald “Bucket” Costello regularly refueled his speedy EA-18G aircraft in midair to avoid losing time island hopping across the ocean, even though the maneuver posed risks.
Zooming along at 300 mph, sometimes through turbulence, Costello steered his jet toward a tanker plane carrying a 100-foot hose and “drogue”— a 2-foot funnel serving as a combination of a wing and a fuel cap — then nudged a 4-foot probe inside. With just a 4-inch margin for error, the stakes of midair refueling are high for an experienced pilot: “If you don’t practice, you can crash.”
But how do you teach this skill — one that military pilots have taken pains to learn since jets swooped onto the scene in the 1950s — to future squadrons of unmanned fighter aircraft that the U.S. military aims to fly alongside conventional aircraft? Being able to refuel autonomous aircraft without landing would greatly extend their range and endurance, which is critical for carrier-based jets.
That capability is just what Costello, now a University of Maryland assistant professor of aerospace engineering, is attempting to make possible. In a study accepted last month by the AIAA Journal of Aerospace Information Systems, he and colleagues demonstrated the feasibility of probe-and-drogue refueling with unmanned fighter planes using data from artificial intelligence-enhanced videos and a virtual reality studio. Funded by a grant from the Lockheed Martin Corp., the research was a collaboration between the A. James Clark School of Engineering MATRIX Lab, which specializes in autonomous robotics, and UMD’s Institute for Health Computing (IHC).
Autonomous airborne technology has become a cornerstone of modern warfare, and in Russia's invasion of Ukraine, fighting has even featured drones that remove humans from so-called "kill chain" decision-making entirely. But deploying autonomous fighter-sized aircraft will be a more complex challenge. Within the next decade, the Navy intends to dramatically increase the number of unmanned aircraft within a carrier air wing, requiring autonomous aerial refueling.
But the economic hurdles to developing self-piloting algorithms are high. While outfitting a Waymo with autonomous capabilities can be done relatively cheaply through ground-level data collection, gathering data in the air costs the government $45,000 per flight hour, said Costello, the director of test and evaluation of autonomous systems at the MATRIX lab, who has studied autonomous aerial refueling for the past five years
Considering the conditions pilots deal with — different light patterns, cloud cover and weather, for example — “you may need 50,000 images to train an algorithm to refuel,” he said. “The amount of flight hours you’d need is almost ‘unobtanium.’”
Costello and his team devised a workaround. Rather than take to the skies, the researchers built a drogue-detection model by superimposing images of an actual drogue onto video images of a tanker jet and fuel hose whipping through the air.
To do that, the researchers employed UMD’s HoloCamera, an advanced setup with 300 cameras that can produce 3D renderings of objects and pair them with background videos. After imaging a F/A-18 drogue in the HoloCamera, they combined the apparatus with video footage of midair refueling captured during a flight test, adding in various light, weather and cloud patterns — finally building an autonomous refueling algorithm out of the virtual but highly realistic scene they’d put together. Their testing indicated self-piloting aircraft could likely be trained for real-world fueling with the synthesized data.
Artificial data collection could save the government big money over five years, Costello said. “Instead of spending hundreds of thousands of dollars on flight tests, I can spend 65 cents on electricity to have computer models train with augmented reality videos,” he said.
In collaboration with the IHC, which manages the HoloCamera, Costello’s team might soon expand the identification method beyond drogues to include entire aircraft. In the future, autonomous jets known as collaborative combat aircraft are expected to fly alongside piloted planes, conducting missions in foreign airspace without access to GPS or radio communications or during other sensitive operations.
“If you’re trying to prevent yourself from being seen, passive sensing is more stealthy,” said Ryan Lowe M.Eng. Robotics ’26, the study’s first author. He studied robotics under Costello as an undergraduate at the U.S. Naval Academy, where Costello was an assistant professor prior to his retirement in 2024 and appointment to UMD.
Now training to be a naval aviator in Pensacola, Fla., Lowe hopes for a future flying F-35s, coordinating missions carried out with autonomous jets on his wing.
"The use of machine learning will enable the collaborative combat aircraft to coordinate with the pilot and each other, increasing situational awareness through passive sensing with minimal communications," he said. “It would help support missions including reconnaissance, strikes, patrols, really anything.”
Inside UMD’s HoloCamera, the researchers built a 3D model of an F/A-18 drogue, then superimposed images of the tool onto videos of actual and synthetic refueling missions.