AI Fighter Falco Beats Human Pilot in Simulated Dogfight – Implications for Military AI

Source: https://www.oschina.net/news/118195/ai-beats-f16-pilot

In September of last year, the U.S. defense research agency DARPA launched the "ACE" AI air‑combat program to train fighter jets to operate autonomously using AI algorithms, and the initiative has now achieved initial success.

According to the UK tech outlet The Register, DARPA recently held an AI‑vs‑human simulated air‑combat tournament using the U.S. Air Force’s flight simulator. In the competition, an AI‑controlled fighter named Falco defeated a seasoned U.S. Air Force instructor.

The tournament used a 1‑v‑1 format where each participant piloted a computer‑generated F‑16, shooting to deplete the opponent’s health bar while avoiding damage. Eight different machine‑learning algorithms from various companies and research groups competed, and the top‑ranked AI earned the right to face a veteran human pilot.

The human pilot, nicknamed "Banger," is an experienced Air Force instructor. Falco, employing a deep reinforcement‑learning agent with an aggressive strategy, proved a terrifying opponent, beating Banger 0‑5 despite his extensive combat experience.

DARPA cautioned that, although the result may raise concerns, such a dominant AI drone is not imminent in reality because the current simulated battles are far simpler than real combat scenarios.

ACE project manager Lt. Col. Dan Javorsek explained that AI must be able to make predictive, on‑the‑fly tactical decisions comparable to human pilots before the simulations become truly realistic.

Heron Systems, the U.S. defense contractor behind Falco, said their AI model has accumulated the equivalent of 30 years of a human pilot’s experience and is trained using the open‑source AdeptRL framework.

AdeptRL is an open‑source reinforcement‑learning framework designed to abstract engineering challenges and accelerate research. It supports single‑ or multi‑GPU training, uses PyTorch baseline models, processes up to 3000 steps/s and 12000 FPS (Atari), and includes modular interfaces for custom networks, agents, and environments, as well as logging, model saving, reloading, evaluation, and rendering capabilities.

An engineer from Heron Systems revealed that they plan to test Falco on real unmanned aerial vehicles in the future.

DARPA’s ACE initiative, launched in September, aims to create combat aircraft that can operate without human pilots, though the primary goal is to develop AI that assists pilots by handling rapid, low‑level maneuvers while humans focus on higher‑level strategic tasks such as issuing commands and launching missiles.

AdeptRL project URL: https://github.com/heronsystems/adeptRL