Site navigation

AI Pilot Defeats Human in DARPA Fighter Combat Simulation

Michael Behr

,

F-16 AI pilot

The winning AI pilot was created by Heron Systems, which beat out competition from larger companies including Lockheed Martin.

An AI pilot has successfully defeated a top US Air Force F-16 fighter pilot as part of a series of simulated aerial dogfights.

The three-day AlphaDogfight Trials were organised by the US Defense Advanced Research Projects Agency (DARPA) as part of its Air Combat Evolution (ACE) programme. The first two days saw eight AIs battle each other, being narrowed down to four before a final winner emerged.

The AI pilots were pitted against simulated opponents developed by John Hopkins Applied Physics Laboratory, which hosted the trial. Scenarios reflected typical aerial missions, such as dealing with an unmanned cruise missile.

During the trial, the AI systems proved they could manoeuvre an aircraft in simple one-on-one combat scenarios and fire the vehicle’s forward guns. The rules of the trials meant that the AIs were not allowed to learn from their experience during the missions.

The tests were the result of months of preparation that started in September 2019 when the eight teams taught their AIs how to fly simulated F-16 fighters. The teams honed their AI’s skills with two additional trials in November 2019 and January 2020.

The eight teams were backed by companies from a range of industries, including defence contractors, academic institutions and gaming companies – Aurora Flight Sciences, EpiSys Science, Georgia Tech Research Institute, Perspecta Labs, PhysicsAI, and SoarTech.

The eventual winner was a small Maryland-based AI developer, Heron Systems, which beat an AI developed by defence powerhouse Lockheed Martin.

On the final day of the event, Heron Systems’ AI was pitched against a human fighter pilot using a virtual reality headset and won five out five of the simulated dogfights.

Heron’s AI, with the callsign Falco, flew aggressively and was consistently able to turn and outmanoeuvre the US Air Force pilot, codenamed Banger, before scoring hits on their simulated F-16. DARPA’s Justin Mock, commenting on the trial, said that the AI was able to make nanosecond decisions and adjustments as well as exhibiting “superhuman aiming ability”.

While Heron’s AI was not allowed to learn, the human pilot was able to shift tactics across the five trials to survive for longer with each mission.

The algorithm had to perform within the limits of what a human could perform in an F-16, such as ensuring that it did not pull excessive amounts of g-force

The program used deep reinforcement learning to teach the AI which actions brought the greatest reward, along with several undisclosed innovations from Heron, with the program running through over 4 billion simulations.

“Using deep reinforcement learning, a novel training pipeline, and other innovations, we trained an AI agent that placed first in the three trials held to date,” the company stated. “The agent performs advanced tactical planning on long time horizons. It is able to operate within the latency requirements of the target system.”

Heron Systems said that its team will publish some of the details about its how it trained its AI later this year.

Recommended

The tests aim to determine whether machines are capable of performing in aerial combat situations, where fast decisions need to be made in an unpredictable environment.

DARPA is ultimately investigating ways for human and machine pilots to share control of a fighter jet and to build trust in the use of autonomous systems.

Eventually, ACE aims to evolve the pilot’s role from operating a single aircraft to managing teams of unmanned drones.

Mock noted that the limited nature of the trials meant that they were not definitive proof of an AI pilot’s viability, but from “what we saw was that in this limited area, in this specific scenario, we’ve got AI that works”.

DARPA’s next move will be to test the AIs against other Air Force pilots and testing the AI pilots’ ability to perform other missions.

AlphaDogfight Trials were not the first time an AI defeated a human pilot in trials – the 2016 Alpha demonstration saw the Alpha AI beat an experienced human combat flight instructor.

However, the test comes on the tail of a Human Rights Watch report calling countries to ban autonomous weapons and halt their development.

Michael Behr

Senior Staff Writer

Latest News

%d bloggers like this: