The United States Air Force (USAF) has made public the first-ever real-world aerial duel simulation featuring artificial intelligence (AI) and a human pilot. This is a ground-breaking development. In a high-stakes combat test last year, a modified F-16 known as the X-62A faced off against a human pilot.
The X-62A was fully AI-controlled, even though a human safety pilot was present. Years of research and development culminated in this momentous occasion, which might completely change the future of unmanned aerial vehicles (UAVs) such as those used in the USAF’s Collaborative Combat Aircraft (CCA) effort.
The battle footage is from a recently published film that highlights the Air Combat Evolution (ACE) program of the Defense Advanced Research Projects Agency (DARPA). With support from the Air Force Research Laboratory (AFRL) and the USAF Test Pilot School, ACE, which was launched in 2019, has made considerable progress. Partners from business and academia, such as Shield AI, have also been extremely important. By virtue of its 2021 acquisition of Heron Systems, Shield AI created the AI “pilot” that won DARPA’s fully virtual 2020 AlphaDogfight Trials.
These earlier successes, meanwhile, lacked realism since the AI pilots were unconstrained by actual constraints, like pilot welfare and aircraft safety. This most recent duel was a big step forward because it addressed that worry.
“2023 was the year ACE made machine learning a reality in the air,” declared Air Force Lieutenant Colonel Ryan Hefron, the ACE program manager, in the video. This dogfight served another critical purpose – fostering trust in AI-powered aircraft for human crews, a key objective for both DARPA and the USAF.
The X-62A, also known as the Variable-stability In-flight Simulator Test Aircraft (VISTA), is a heavily modified two-seat F-16D. Its configurable flight systems allow it to mimic various aircraft, making it ideal for real-world testing. This versatility also makes VISTA perfect for the ACE program.
“We have a space within VISTA’s flight controls that allows AI agents to send commands as if they were in the simulated model,” explains Que Harris, the lead flight controls engineer for the X-62A at Lockheed Martin. Harris referred to this as a “sandbox for autonomy” within the aircraft.
To support ACE, the X-62A completed 21 test flights from Edwards Air Force Base in California across three testing periods between December 2022 and September 2023. These flights necessitated frequent reprogramming of the “agents,” resulting in over 100,000 lines of code being modified. The AFRL has previously emphasized the potential for further enhancing such testing by rapidly training and retraining algorithms in digital environments.
The US military acknowledges the continued need for human oversight of autonomous weapon systems. However, the extent of human involvement in decision-making is likely to evolve over time, sparking ongoing discussions and debate.