DARPA’s AlphaDogfight Tests AI Pilot’s Combat Chops

WASHINGTON: DARPA’s AlphaDogfight Trials, which culminate Wednesday with a simulated aerial dual pitting the winning AI F-16 ‘pilot’ against a human pilot, aims to determine the viability of relying on machines in a fast-paced, unpredictable air-to-air combat environment.

“Before AI can find its place, really, in the cockpit in a formation, it kind of has to earn its wings first,” DARPA Deputy Director David Honey said as he opened the virtual contest this morning.

“It’s got to show that it can actually fly, that it can compete, and that it can contribute,” he explained. “A human pilot goes through pretty much the same thing, you know. You learn to fly, you have to demonstrate you have certain level of skill, then you can join a formation, be part of the squadron and go out and be part of the part of the team.”

But just as importantly, Honey and other officials involved with DARPA’s Air Combat Evolution (ACE) program say, the three-day virtual games will help flesh out how human and machine pilots can share operational control of a fighter jet to maximize the chances of mission success.

“When we put this particular part of the program together, there were some really critical questions that have to be answered and we’ll see how that comes out this week. We need to understand: can we actually develop autonomy algorithms that can operate in the very demanding environment of air-to-air combat?” explained Honey.

“And we’ll need to understand — you know later on once we have those algorithms developed — when you do the human-machine teaming, how do you share the workload in the dynamic situations of aerial combat where aircraft are getting damaged and whatnot? How do you shift responsibilities back and forth so that you can have the air dominance that you want? Those are very complex problems,” he added.

As Breaking D readers know, Air Force acquisition head Will Roper is very keen on the concept of teaming unmanned and manned aircraft to undertake a variety of missions in future all-domain fights against peer adversaries — for example, championing Air Force Research Laboratory’s Skyborg effort to mate AI-driven drones with fighter jets to perform traditional ‘wingman’ roles, as well as other operations.

ACE Program Manager Col. Dan Javorsek (call sign “Animal”) said that the bottom line for the Air Force’s fighter jocks is the issue of trust; i.e. ceding life-or-death control to an artificial intelligence.

“We are very reluctant — the fighter pilot community is very reluctant — towards any sort of change. There’s almost an institutional requirement to not accept change,” he aid. “When it comes to autonomy, this has been avoided almost like the plague.”

Outgoing head of Air Combat Command, Gen. Mike Holmes, told the Air Force Association last Friday that he believes one of the first areas the service could use AI pilots is in training.

“As we look at how we’re going to bring those pilots’ assistants in, and then how we might move toward a more autonomous vehicle, one of our ideas is we’d like to try it in red air first,” he said. “So, we’d like to take it and see if there are ways to do some unmanned adversaries against our manned aircraft, and then let them learn and train their algorithms and see how good they get at fighting against us.”

Indeed, Javorsek explained, that the overarching goal of the ACE program is to gradually build up trust  in increasingly capable autonomous systems that eventually can take on the full gamut of air combat missions. (Building trust that AIs can understand human intentions, as Kelsey reported last week, also is the goal of another DAPRA program, called Adaptive Distributed Allocation of Probabilistic Tasks (ADAPT), using the popular Minecraft video game.)

“The idea of the AlphaDog trials was essentially to be an accelerator, or an incubator, for a lot of the technologies that are flowing into this larger program,” he said. “In ACE, we’re really demonstrating trusted, scalable, human level, AI-based and AI-driven autonomy for air combat.”

This week’s trials are the result of work that began in September 2019 that served as a kind of “pre-season” of football or baseball of any other sport, during which the AI teams essentially learned to fly their simulated F-16 fighters, Javorsek said. Two subsequent trials in November 2019 and January 2020 allowed the teams to hone their AI pilot’s skills.

The eight teams comprise a wide variety of participants, from traditional DoD contractors to academic institutions to gaming companies, each of which demonstrating divergent methodologies in building their algorithms.

According to DARPA’s AlphaDogfight website, the eight teams are led by: Aurora Flight Sciences, EpiSys Science, Georgia Tech Research Institute, Heron Systems, Lockheed Martin, Perspecta Labs, PhysicsAI, and SoarTech.

“These eight performers have gone down very different paths, chosen very different approaches in AI, and for us that’s a bonus, because we not only get to see how these algorithms perform, but we get to see how these different areas of AI work,” he said. This can inform later investments, he added, “if it turns out that there are particular areas that need further development or if there are things that offer, you know, new advantages that we previously hadn’t anticipated.”

Today, the eight AI pilots face simulated ‘adversaries’ developed by John Hopkins Applied Physics Laboratory, which is hosting the virtual games, each based on a typical scenario involving offensive and defensive maneuvers. This includes, Javorsek said, one called “Zombie” that mimics a cruise missile or an unsophisticated drone.

(Interestingly, Colin reported wayyyy back in 2016 that a teeny company called PSIBERNETIX said in a journal article that it had created an AI program, called ALPHA, that allowed drones to consistently outgun a manned pilot. Since then the European defense giant Thales bought the company.)

Tomorrow, the eight AI pilots will battle each other, four of which will ‘live’ to fight again on Wednesday morning. The ultimate winner of the round-robin contests will then face off against a human F-16 pilot.

According to Honey, the virtual games have attracted an enormous amount of attention within DoD, the various services, industry, academia and from outside the US. Some 10,000 people from 93 countries have signed up to watch, maxing out DARPA’s computer capacity, with another 5,000 or so still hoping to be able to join in.

“We’re trying to see if there are some other solutions available to let everybody in. But again, this shows the depth and breadth of interest in this particular area, and we’re very glad to see that kind of interest in this program,” he said.

(We trust, however, that his invitation is not open to the Russians and Chinese.)