How AI Learned to Fly a Fighter Jet: Inside DARPA's Autonomous F-16

High above the Mojave Desert, two F-16 fighter jets merge at 1,200 miles per hour, closing the distance to a mere 2,000 feet. In one cockpit, a human pilot grips the stick, pulling aggressive G-forces to gain the tactical advantage. In the other, the pilot's hands are resting in their lap. The jet is flying itself.[1][2]

This is not a pre-programmed autopilot following a set of GPS waypoints. The aircraft—a heavily modified F-16 known as the X-62A VISTA—is being flown entirely by an artificial intelligence agent. It is reading sensor data, calculating aerodynamics, and making split-second tactical decisions in the middle of a within-visual-range dogfight.[1][4]

Announced in early 2024, this milestone marks the culmination of the Defense Advanced Research Projects Agency’s (DARPA) Air Combat Evolution (ACE) program. It represents one of the most significant breakthroughs in aerospace engineering in decades: the moment machine learning successfully crossed the "sim-to-real" gap, moving from the sterile environment of a computer simulator into the chaotic, high-stakes physics of the real sky.[2][6]

To understand how an AI learns to fly a fighter jet, you have to look back to August 2020 and a virtual competition called the AlphaDogfight Trials. DARPA invited eight teams to develop AI agents capable of flying a simulated F-16 in a classic, World War II-style dogfight.[5]

The agents trained using reinforcement learning, a process where the AI learns by trial and error, receiving mathematical "rewards" for good maneuvers and "penalties" for crashing or getting shot down. After billions of simulated flights, the winning agent, developed by Heron Systems, faced off against an active-duty Air Force F-16 Weapons School instructor.[3][5]

The result was a 5-0 sweep for the machine. The AI exhibited superhuman aiming ability and executed aggressive, high-aspect maneuvers that human pilots are typically trained to avoid due to physiological limits and safety regulations. But while the simulator victory was a massive leap, a computer model is a perfectly predictable environment. The real world is not.[3][5][6]

In a simulator, there is no turbulence, no sensor degradation, and no subtle variations in engine thrust. To bridge this gap, the engineers utilized a specialized architecture known as Hierarchical Reinforcement Learning (HRL).[3]

Rather than relying on a single, monolithic neural network to handle everything from strategy to rudder control, HRL divides the problem into a hierarchy of sub-tasks. At the top level, a "policy selector" acts as the brain's executive, deciding on the overarching strategy—whether to play offensively, defensively, or evade.[3][6]

Once the high-level policy makes a strategic choice, it hands the execution down to a set of low-level, specialized policies. These lower-level algorithms are trained specifically to excel in narrow regions of the flight envelope, translating the strategic command into the continuous, high-dimensional micro-adjustments of the jet's ailerons, elevators, and throttle.[3]

This hierarchical approach allows the AI to balance exploration with exploitation, adapting dynamically to the unpredictable variables of real-world aerodynamics. But to test this software safely in the air, the military needed a very specific kind of hardware.[3]

Enter the X-62A VISTA (Variable In-flight Simulator Test Aircraft). Originally built in the early 1990s as the NF-16D, the VISTA is a one-of-a-kind experimental testbed operated by the U.S. Air Force Test Pilot School at Edwards Air Force Base.[4]

The VISTA is unique because its flight control systems can be programmed to mimic the aerodynamic characteristics of virtually any other aircraft. In 2021, the jet underwent a massive upgrade, receiving the System for Autonomous Control of Simulation (SACS). This upgrade effectively turned the X-62A into a flying sandbox for artificial intelligence.[4][6]

The SACS architecture allows AI algorithms to plug directly into the jet's flight controls. Crucially, it maintains a strict safety boundary. A human safety pilot always sits in the cockpit, monitoring the AI's decisions. If the algorithm attempts a maneuver that violates safety parameters, or if the jet approaches the ground too quickly, the human pilot can instantly disengage the AI with the flip of a switch.[1][2][4]

With the hardware and software ready, the ACE program began real-world flight testing in December 2022. Over the next year, engineers modified more than 100,000 lines of flight-critical software, iterating the AI's code based on the data gathered from the actual sky.[1][2]

The testing culminated in September 2023, when the X-62A engaged in a series of live dogfights against a standard, human-piloted F-16. The engagements started with basic defensive maneuvers before escalating to high-aspect, nose-to-nose passes at closing speeds of 1,200 miles per hour.[1][2]

Throughout the intense combat maneuvering, the AI agent maintained complete control of the X-62A. The human safety pilots on board never once had to activate the safety switch to override the system. While the military has not declassified the win-loss record of these real-world engagements, the fact that the AI safely and aggressively navigated the physical environment was a historic victory in itself.[1][2][6]

The ultimate goal of the ACE program, however, is not to replace human fighter pilots with machines. The goal is to build trust.[1][5]

As the military moves toward a future defined by Collaborative Combat Aircraft—swarms of autonomous drones flying alongside crewed fighters—human pilots will need to shift their focus. Instead of spending their cognitive energy on the mechanical act of flying and dogfighting, pilots will become "mission commanders," orchestrating the broader battle space.[5][6]

For a pilot to comfortably hand over the stick and focus on a tablet screen, they must have absolute faith that the AI will not crash the jet or make a fatal error. By proving that an AI can handle the most complex, dynamic, and dangerous task in aviation—a within-visual-range dogfight—DARPA and the Air Force have laid the foundation for that trust.[1][2][5]