The Defense Advanced Research Projects Agency (DARPA) have just released a video of an AlphaDogfight Trials Competition where a simulated dogfight between the AI and an Air Force F-16 pilot had a very interesting result.
Not only did the AI win, but it won 5-0.
The DARPA AlphaDogfight Trials aim to demonstrate the feasibility of developing effective, intelligent autonomous agents capable of defeating adversary aircraft in a dogfight.
If a country is able to leverage AI to pilot their planes, there are some massive opportunities from not only exceed human capabilities, but perhaps the best outcome is that we could avoid risking human life on the battlefield.
A really interesting insight from the commentary of this event is that the behaviour of the two planes really appeared as if 2 humans were behind the controls, meaning the AI has done a great job of learning how humans maneuver their planes in their aerial chess battle.
Regardless of the pilot (AI or human), physics is the absolute limiting factor. Interestingly the capabilities of the F-16 fighter jet, far outweigh the capabilities of the human body. During the fight the commentators reference the fact that the AI starting taking turns and making maneuvers that humans try and avoid.
The problem with g-forces on the human body is that you’re either really uncomfortable or even risk blacking out. Of course, these elite pilots train to prevent this and may even wear pressurised suits to assist, but there’s always a limit. Naturally, if you replace the human and use AI to pilot the plane, these limitations are removed and your available trajectories are only limited by the forces in which the plane can withstand.
DARPA’s AlphaDogfight Trials seeks to advance the state of artificial intelligence (AI) technologies applied to air combat operations. The trials are a computer-based competition designed to demonstrate advanced AI algorithms that can perform simulated within-visual-range air combat maneuvering, otherwise known as a dogfight. The goal is to use the dogfight as the challenge problem to increase performance and trust in AI algorithms and bring together the AI research and operator communities.
In August 2019, DARPA selected eight technically and organizationally diverse teams to compete in the AlphaDogfight Trials with the purpose to energize and expand a base of researchers and developers applying AI technologies to complex operational problems.
- Trial #1 in November 2019 was an exhibition match with the opportunity for teams to compete against different APL developed adversary agents and test the simulation environment at scale.
- Trial #2 held in January 2020, was the first competition where teams were ranked against each other and tested their agents against more challenging adversary agents.
- Trial #3 is the final competition where teams competed against each other in a bracket-style competition. The top team (Heron) advanced to fight against a USAF fighter pilot in a simulated dogfight on Aug. 20.
AlphaDogfight Trials is a precursor to the DARPA Air Combat Evolution (ACE) program, which involves AI development and demonstration in three program phases – modelling and simulation, sub-scale aircraft, and full-scale aircraft testing. Ultimately, ACE will be flying AI algorithms on live aircraft to demonstrate trusted, scalable, human-level autonomy for air combat.
Watch the whole whole thing below, or skip to the end to see the final human vs AI battle.