[AI Library] Chapter 9. ACE Program and X-62A VISTA: From Simulation to the Real Sky
Chapter 9. ACE Program and X-62A VISTA: From Simulation to the Real Sky
ACE Program and X-62A VISTA: From simulation to real skies One day in August 2020, aviation experts around the world were holding their breath in front of their computer screens. I was watching as Heron Systems' artificial intelligence defeated a human veteran pilot 5-0. But the test pilots at Edwards Air Force Base shook their heads. “It’s a video game.” Their cynicism was valid. The sky in the simulator is always clear. The wind blows according to mathematical formulas, sensors don't lie, and communication links don't break. The most important thing is that even if you make a mistake, you can just hit the reset button.
There, AI fights based on perfect information. You can determine the enemy's location, speed, and direction in milliseconds. But the actual battlefield is different. Noise falls like snow on the sensors, communication is cut off by enemy electronic warfare, and the atmosphere shakes the aircraft. Gravity acceleration, which does not exist in the virtual world, puts pressure on the pilot's neck, and falling does not mean game over, but death. The Defense Advanced Research Projects Agency, commonly known as DARPA, has moved to fill that gap. The ACE program launched in 2019, its full name is the Air Combat Evolution program.
The goals of this program go beyond simply having AI fly an airplane. The goal was to ensure that human pilots would trust the AI and have the ability to fight together as a team. DARPA has identified air combat as a “challenge problem.” The calculation was that if trust could be built in air combat, where speed and uncertainty are extreme, other areas would become easier. The ACE program consisted of three phases: The first step was to validate and develop core competencies in a simulation environment over 18 months. The Alpha Dog Fight Trials were part of this phase.
The second step was to translate the algorithm into a small unmanned aerial vehicle by conducting 16 months of unmanned aerial vehicle flight testing.
The third step was testing on a real fighter-scale aircraft. And at the heart of this plan was one unique fighter plane. This is the X-62A VISTA. VISTA stands for Variable In-flight Simulator Test Aircraft. Translated, it means something like a variable flight simulator test aircraft. At first glance, it looks like an ordinary F-16D two-seat fighter. However, the inside of this aircraft is a completely different world. This oddball plane, created through a collaboration between Lockheed Martin's Skunk Works and Carlspan, can mimic the flight characteristics of aircraft other than the F-16 simply by changing software settings.
If you want the control feel of an F-35, set it that way, and if you want to recreate the movement of an MQ-9 drone, change it that way. Just like a chameleon changes color depending on its surroundings, VISTA was able to change its identity in the sky. DARPA equipped this one-of-a-kind aircraft owned by the U.S. Air Force Test Pilot School at Edwards Air Force Base with a system called SACS. System for Autonomous Control of Simulation, an autonomous control simulation system. This was an interface that allowed the AI agent to directly access the aircraft's flight control computer to control the control surfaces and engine power.
AI can now directly handle the control surfaces of wings and tails called ailerons, rudders, and elevators, as well as the thrust of jet engines. Of course, safety measures were also needed. It was impossible to hand over control of a fighter jet worth tens of billions of won to a newborn AI. The AI could have bugged and plummeted to the ground, or tried to maneuver beyond the limits of what the aircraft could withstand. For this purpose, a human safety pilot was seated in the back seat. If the AI attempted a dangerous maneuver or went out of control, this pilot could immediately intervene and take back control.
A safety envelope was set in the flight control computer, and software blocked the AI when it attempted to make an unreasonable maneuver exceeding 9G or rush to the ground. In December 2022, a historic flight took place over Edwards Air Force Base in California. From December 1 to 16, the X-62A conducted 12 flights piloted by AI agents. In total, I spent more than 17 hours in the sky. Learned through tens of millions of trials and errors in the simulator.
Those chunks of code controlled the thrust of the actual jet engine, handled the lift, and began to physically fly in the sky. This process was not always smooth. When the AI, which was perfect in the simulation, encountered turbulence in the real sky, it trembled slightly or had a delayed response due to sensor data delay. On some days, the AI was too passive, which was a problem, and on other days, it tried to maneuver beyond the limits of what the aircraft could tolerate, triggering a safety cutoff. However, the true value of testing using VISTA lies in repetition and mastery.
Typically, it takes several months to verify new fighter software. This is because the code needs to be modified, verified on the ground, cleared for flight, and then flown again. However, the ACE team updated the code daily through VISTA. If we discovered an AI error during a morning flight, we modified the coding during lunch, ran the simulation, and applied it immediately to the afternoon flight.
Instructors at the test flight school at Edwards Air Force Base testified that they “witnessed fixes that would previously have taken a year to be resolved in just one day.” Engineers were able to swap out autonomy algorithms on board the X-62A in minutes, and pilots were able to test different teams' AIs every few hours. In September 2023, the day has finally arrived. In the sky above Edwards Air Force Base, an AI-controlled X-62A and a human-controlled F-16 fighter jet engaged in a real-life dogfight. Imagine this. At an altitude of 600 meters, two fighter jets rush toward each other at a relative speed of 1,900 kilometers per hour.
The moment the distance between each other narrows to just 600 meters, the human pilot feels instinctive tension, but the AI calculates the optimal position without hesitation. It started out as a defensive maneuver, but soon developed into an offensive maneuver to chase each other. The climax of the battle was a high-angle nose-to-nose engagement where the two aircraft crossed each other face to face. This was no simple flight following a scripted flight. The two fighter planes maneuvered furiously, intent on shooting each other down for real.
During the flight, the X-62A had human pilots in the front and rear seats, but they did not touch the controls. They monitor the AI's flight
He only served as a safety officer and prepared for emergencies. AI judged the situation and controlled the aircraft on its own without human intervention. Over the course of 21 test flights, the team modified more than 100,000 lines of flight-related software. The AI flew with greater precision, operated within the limits of safety regulations, and reacted faster thanks to the computer's faster observation-orientation-decision-action loop. While the human pilot attempted to maneuver for a follow-up shot but did not attempt a head-on or opportunistic shot when given the opportunity, the AI fired as quickly as possible.
This ultimately made the difference. The U.S. Air Force did not disclose the specific winner or loser of this confrontation for national security reasons. However, testimonies from those involved in the test suggest that the AI's performance was surprisingly stable. Chief test pilot Bill Gray said, “The X-62A proved that non-deterministic AI can be safely applied to aerospace systems by solving the very complex problem of dogfighting.” From a pilot's perspective, this change is palpable. When fighting AI in a simulator, you are left with the feeling that your opponent is trapped within the rules of the game.
But when you move to the X-62A, AI is no longer just pixels and vectors. It's a multi-ton chunk of metal passing by me, and the vortices and angles of closure it creates could cause my aircraft to stall. Speed comes as a shock rather than a number, and G is felt on the neck bone rather than on the screen. Lt. Col. Ryan Heffron, DARPA ACE program manager, emphasized: “In all areas, research progresses only as fast as the tools allow. VISTA’s recent upgrades enable rapid integration and safe testing of AI-based autonomy, making it a much more effective testbed.
This brings us at least a year closer to full-scale flight testing of AI-based autonomy.” The message from the ACE program and the X-62A VISTA experiment is clear. If Alpha Dog Fight showed possibility, ACE created verifiable reality. The Turing Test of air combat is not just about whether you win or lose, but whether you can conduct the fight while passing real-world safety and reliability standards. The X-62A really steps on that threshold. It was the moment when the genius in the simulator was reborn as a warrior in the real world.
Kim Kyung-jin
Attorney · Former Member of the National Assembly · AI Policy Researcher
© 2026 Kim Kyung-jin. All rights reserved.