AI Bests Human Fighter Pilot in AlphaDogfight Trial at Johns Hopkins APL

APL’s AI agents gave teams a run on the first day with all eight teams virtually flying their algorithms against five APL-developed computer adversaries. APL’s algorithms were developed to perform at different skill levels, allowing the industry agents to train similarly to actual pilots, explained Kerry Neace, a former fighter pilot and APL’s program area manager for Air Dominance.

“From the control of the environment to the displays that allowed us humans to view what was happening virtually, the APL team delivered continuous and comprehensive simulations to the industry team,” Neace explained. “This allowed the industry teams to progressively develop and train their AI agents against constantly updated adversaries, drastically reducing the time to develop such complex AI capability.”

Teams flew against one another in a round-robin tournament on the second day, complete with live play-by-play and analysis from the “Control Zone” commentary desk at APL’s Intelligent Systems Center. Christopher “Disco” DeMay, program manager for the DARPA ACE program at APL, and DARPA cohosts, including Col. Dan “Animal” Javorsek, program manager in DARPA’s Strategic Technology Office, and Lt. Col. Justin “Glock” Mock, recapped and analyzed each dogfight.

The top four teams advanced to a single-elimination tournament on day three, with Heron emerging to square off against Banger. Overwhelming demand prompted organizers to move the final event to a public YouTube livestream, which ultimately garnered over 300,000 views.

DARPA employed APL’s expertise to develop the simulation environment and adversary agents for the trials. The Lab, which also hosted the first two ADT events, leveraged its capabilities in AI and software development, modeling and simulation, and aircraft dynamics and controls to create the “arena” for all three ADT competitions.

APL started on the fundamental technology used for many of the more successful agents in the DARPA ACE program under an internally funded research project called “Athena Inspired.”

“Athena Inspired leveraged the promise that reinforcement learning has to revolutionize the design of future warfare systems and discover new and novel combat tactics,” said Tom Urban, the lead of APL’s technical group responsible for advancing AI capabilities to revolutionize air, strike and electronic warfare.

As AI and advanced autonomy systems reach the field, the future of warfare will move to machine speeds and, as a result, “we must also enable warfighters to fight these future battles with machine precision,” Urban said. “Thus, we need to also research and develop intelligent virtual assistants to work with warfighters to think, reason, react, respond and plan at machine speeds with machine precision.

“APL is investing in the fundamental research that will drive the AI technology of the future and empower both the intelligent autonomous systems and intelligent virtual assistants,” he added.

ADT was a significant achievement that offered fascinating glimpses into the future of warfare, noted commentators DeMay, Javorsek and Mock.

“The objectives of the program are really about bringing together the fighter pilot and AI communities,” DeMay said. “While we’re proud that, in the span of a year, we were able to take eight separate companies, give them a challenge problem that’s very relevant for DoD, get them involved with the fighter pilot community and reach an outcome where we’re seeing credible maneuvers, credible behaviors and, actually, very good performance, there’s still a lot of work to be done.

“That’s why ADT is really a risk-reduction effort for ACE,” he continued. “We’re looking to take this from simulation to getting it out on subscale, and ultimately into live aircraft.”

APL will continue to serve as a core part of the government team on the ACE program. As the experimentation integration team, the Lab brings together various technical performers to demonstrate trusted, scalable, human-level autonomy for air combat. APL will continue to provide the modeling and simulation environment, subscale aircraft as well as adversary AI capability at both the tactical and battle-management level to mature performer algorithms while minimizing risk — with the ultimate goal of flying on full-scale combat representative aircraft.