US Air Force says AI-controlled F-16 fighter jet has been dogfighting with humans
Robo-plane was made to restrain itself so as not to harm pilot or airframe
Video The US Air Force Test Pilot School and the Defense Advanced Research Projects Agency (DARPA) claim to have achieved a breakthrough in machine learning by demonstrating that AI software can fly a modified F-16 fighter jet in a dogfight against human pilots.
The claims rest on the USAF and DARPA implementing machine learning in an X-62A VISTA, a plane built as a testbed as it can mimic the performance of other aircraft, and recognition of their work as one of four finalists for the National Aeronautic Association's 2023 Robert J. Collier Trophy, an annual award for exceptional feats of aeronautics or astronautics in America.
"The potential for autonomous air-to-air combat has been imaginable for decades, but the reality has remained a distant dream up until now," said Secretary of the Air Force Frank Kendall. "In 2023, the X-62A broke one of the most significant barriers in combat aviation. This is a transformational moment, all made possible by breakthrough accomplishments."
DARPA has been testing AI agent software for piloting simulated planes for several years. Its Air Combat Evolution (ACE) program dates back to 2020, when AlphaDogfight trials pitted human pilots in a flight simulator against an AI opponent.
The AI software won that competition but had an edge – it was allowed to fly at speeds that would have overstressed a real aircraft and generated g-forces that would harm a human pilot.
Heuristic or rules-based autonomy has been a common approach in military and space applications. These sorts of expert systems boil down to if-then statements that specify condition-based triggers that lead to specific actions. But this approach is less useful when there are too many variables and rules to account for.
"The machine learning approach relies on analyzing historical data to make informed decisions for both present and future situations, often discovering insights that are imperceptible to humans or challenging to express through conventional rule-based languages," explained Daniella Rus, director of MIT CSAIL, in a DARPA video you can see below. "Machine learning is extraordinarily powerful in environments and situations where conditions fluctuate dynamically making it difficult to establish clear and robust rules."
Dogfights between military aircraft are very much a dynamic scenario. But machine learning has a downside that needs to be overcome. It needs to be explainable and verifiable enough that military personnel will trust it and that aviation authorities will certify systems that implement such code.
The X-62A is essentially an F-16 that's been integrated with a flight simulator, so that machine learning agents can operate the plane.
- Boston Dynamics' humanoid Atlas is dead, long live the ... new commercial Atlas
- US Air Force secretary so confident in AI-controlled F-16s, he'll fly in one
- What if AI produces code not just quickly but also, dunno, securely, DARPA wonders
- DARPA worried battlefield mixed reality vulnerable to 'cognitive attacks'
"The incredible accomplishment of this year was to take these machine learning agents and place them into the X-62A in a real world environment," said Col. James Valpiani, commandant of the Test Pilot School.
In December 2022, machine learning agents controlled the flight path of the X-62A, a first for AI piloting. Testing and improvements continued over the next few months, until in September 2023, the AI software flew the X-62A in a mock dogfight against a human-piloted F-16. It did so without violating human safety norms, and without leading the on-board pilots to intervene and take control.
According to DARPA, the X-62A team's accomplishments would be viewed similarly to AlphaGo Zero's impact on Chess, Shogi and Go, as a validation of autonomous aviation for both military and commercial applications.
But the team’s efforts were not enough to win the 2023 Collier Trophy, which was awarded to NASA and the OSIRIS-REx asteroid sample capture-and-return team. ®