The virtual contest raises the stakes for weaponized AI and it's future use in the wars.
An artificial intelligence claimed victory over several other AIs in simulated dogfights and then went on to trounce a real human Air Force pilot, beating him 5-0.
The AlphaDogfight contest, coordinated by the Defense Advanced Research Projects Agency (Darpa), shows the potential for AI to take on mission-critical military tasks that were once exclusively done by humans. It might be impossible to write a conventional computer program with the skill and adaptability of a trained fighter pilot, but an AI program can acquire such abilities through machine learning.
The Heron Systems-developed AI beat algorithms from Lockheed Martin and Aurora Flight Sciences and eventually advanced to a final matchup against a skilled District of Columbia Air National Guard pilot. Ironically, the AI’s lack of a formal air combat education likely contributed to the win. All participating AI algorithms relied on deep reinforcement learning, in which they develop an understanding to perform their task with efficiency by indulging in it over and over again.
An F-16 Fighting Falcon.On 20 August, in an event organised by the US’ Defense Advanced Research Project Agency (DARPA) to understand how AI and machine learning can be employed in air-to-air combat. This event was the final dogfight of the US military’s AlphaDogfight challenge, which was organised to “demonstrate the feasibility of developing intelligent autonomous agents capable of defeating adversary aircraft in a dogfight”.
A fighter wins a dogfight when it manages to place itself behind the enemy and remain in that position long enough to get a kill shot.
The round robin tournament was the culmination of a yearlong effort by the Defense Advanced Research Projects Agency (DARPA). Known as AlphaDogFight, the effort recruited several defense contractors, startups, and university laboratories to develop an artificial intelligence capable of dogfighting in a F-16 Fighting Falcon jet. Before that, One of the teams Heron System AI defeated was from Lockheed Martin, the world’s largest defence contractor, which settled for the second runner up position.
The final dogfight of the event was between the Heron Systems AI and “Banger,” an F-16 pilot assigned to the Washington D.C. Air National Guard and a veteran of the Air Force’s Weapon Instructor Course. Heron beat Banger five times in a row.
A screenshot from the AlphaDogfight challenge. (DARPA/PATRICK TUCKER)The Heron vs. Banger dogfights took place after a “merge,” in which two opposing fighter jets race toward one another head-on. The duels took place in JSBSim, a multi-platform open source flight dynamics model, with Banger utilizing a VR headset. The duel involved using the F-16’s onboard M61 Vulcan 20-millimeter Gatling gun only, with no missiles allowed. According to the rules of engagement, gun shots were assumed to automatically hit the target, with no misses.
A screenshot from Heron AI vs F-16 Pilot Virtual Dogfight.Heron’s AI was very aggressive throughout the tournament, quickly maneuvering into position to get a shot off against its opponent. The AI made the most of the F-16’s flight characteristics, position, and speed in ways that trained pilots might not, often to execute some shots that were unlikely to score hits in real life, but were registered as “kills” in-tournament.
This is particularly apparent in the duel against Banger. While the pilot took time to set up a shot with a high kill probability, the Heron AI went for any shot it could conceivably take. In each case, Heron reached a position where it could get a shot off faster than Banger could—but in reality, Banger might have actually won the duel.
“All of the simulated fighter battles were restricted to allow use of the nose cannon only, and after the AI vs. AI matches, an anonymous human pilot entered the competition, wearing a VR helmet,” a report on Engadget said.
In other words, AI-based algorithms make errors and learn from them, and in the process associate cost and benefits with each manoeuvre and update these based on every new experience simulation after simulation.
“What you were basically watching was the AI agents learning to fly the plane,” Col Dan Javorsek, a former F-16 aviator and test pilot who managed this event for DARPA, was quoted by Popular Science as saying. “A lot of them killed themselves on accident—they would fly into the ground, or they would just forget about the bad guy altogether, and just drive off in some direction,” he added.
Here's a video from that AI event by DARPA.
Heron’s victory is a landmark win for artificial intelligence in warfare, and it highlights the advantages of AI-controlled weapons systems. Unencumbered with a human body that feels disorientation, discomfort, and even pain, the AI pushed its virtual F-16 to greater limits, including higher g-forces, to position itself for a kill—any kill. Those physical factors simply don’t exist for an AI.
For example, pulling hard turns during flight produces dramatic G forces which, among other things, can damage the fighter jet. The AIs had to develop an “understanding” to deal with such tricky situations. Meanwhile, this was not the first time an AI agent defeated a human fighter pilot. In a simulated combat in 2016, an AI agent named Alpha had managed to beat an experienced human combat flight instructor.
In the future, artificial intelligence will have to deal with engagements where a kill is not guaranteed. If a specific maneuver could lead to the AI being in position with only a 30 percent probability of kill, should it take it, or should it continue seeking an advantage in the dogfight in hopes of a higher probability of kill in the future? How will an AI dogfight with missiles?
USAF F-16 In Their Formation.