F-22 autonomous drone test shows pilot directing AI wingman

The two jets launched separately from Edwards, the Air Force’s big test base in California, then met up. With the pair in formation, the Raptor pilot transmitted revised orders to the drone using a tactical data link, a secure radio network military aircraft use to share information.

The Fury’s onboard autonomy software then interpreted those orders and executed them. The Air Force described the pilot’s role as an airborne quarterback: deciding what the drone should do, while the drone decided how to fly it. The service announced the result on October 16, and Anduril says the flight lasted several hours and met its goals.

It was a first for both the F-22 and a pre-production Fury. An F-22 had already directed a General Atomics MQ-20 Avenger, used as a stand-in drone, in a similar test back in February, but this was the first time a Raptor worked with one of the actual aircraft the program is buying.

A fighter pilot’s hands and head are already full. They fly their own aircraft, read sensor data, judge threats and decide when to release weapons, often all at once. Asking them to also hand-fly a second aircraft would be close to impossible.

So the plan is delegation. The pilot could assign the drone a job, such as watching for enemy air defenses and going after them, while staying focused on the main task, like engaging enemy fighters. The drone handles the flying and the details.

The software doing that work on the Fury is Hivemind, made by Shield AI, chosen for the Fury earlier this year. The service set a common, open design rule for that software so the code can move between different aircraft, so the service is not locked into any one supplier. In June, the Air Force also set up a pool of six companies that can compete to supply mission autonomy software over the next six years.

A clean flight with one Raptor and one drone is a long way from combat. Neither the Air Force nor Anduril has explained exactly how the F-22’s cockpit was set up to send the orders, although earlier budget papers mentioned adding tablets and cabling. It is also unclear how well a pilot can manage several drones while flying hard in a real fight.

Several big questions are still being tested separately:

• How much freedom the drone gets. The limits on its tactical choices, safe flight control and weapons authorization each need their own trials.
• What happens when the link breaks. The drone has to keep working sensibly under radio jamming, or when it can’t hear from its pilot at all.
• Day-to-day support. The aircraft must fit into maintenance, weapons handling and operations from small, bare-bones forward bases.

On weapons, the Air Force has kept a person in charge so far. When a Fury fired an AIM-120 missile at a simulated target over the Mojave Desert in July, a human operator held weapons-release authority the whole time.

For most people, the clearest impact is cost. The idea behind the Collaborative Combat Aircraft program is to add cheaper, semi-autonomous jets alongside expensive crewed fighters, so one pilot can bring extra sensors, weapons and electronic warfare gear into a mission without risking more human crews. Air Force leaders call this a force multiplier.

The scale is large. The service wants at least 500 first-increment drones by 2032 and an eventual fleet of 1,000 or more. Production contracts for Anduril’s Fury and General Atomics’ Vengeance were awarded in June, months ahead of schedule.

There is also a wider lesson for anyone following AI. This is the same “give it the goal, let it work out the steps” model behind AI agents in office software and self-driving features, only with far higher stakes. How the military sets boundaries, keeps humans in charge of lethal choices and handles lost connections is likely to be studied well beyond defense. Other countries, including Sweden and China, are testing similar one-operator, many-drone ideas.

The next step is scale. Fury and Vengeance are scheduled to take part in Emerald Flag in December, their first large-force exercise, where many aircraft work together in one scenario. Neither drone has flown in an event like that before.

The F-22 is the first fighter the Air Force plans to pair with these drones. Older F-16s, the coming F-47, the F-15EX and F-15E, and the F-35A are all being studied as partners too, and the long-term goal is for one Raptor to direct several drones at once. Data from the October flight will feed into further software work and developmental testing.

The program builds on years of experiments. In 2023, the Air Force described AI flying a modified F-16 through tactical drills, and in 2024 DARPA revealed that its AI-flown X-62A VISTA had faced a human pilot in a real dogfight, with safety pilots aboard who never had to take over.