By Javier Rollon · 2023-12-10
In 2020, DARPA ran an experiment called AlphaDogfight. An AI system fought against a human F-16 pilot in a simulated dogfight. The AI won. Five to zero. And it changed how I think about everything I build.
The AI — developed by Heron Systems — flew a simulated F-16 against a real fighter pilot, an experienced instructor with thousands of hours. The AI was aggressive in a way human pilots aren't. It flew at the edge of performance limits continuously, pulling G's that would cause a human to black out, maintaining shot solutions that required reaction times below human capability. Five engagements, five kills, zero losses.
But here's the part the headlines missed: the simulation environment was simplified. No weather. No equipment failures. No electronic warfare. No wingmen. No rules of engagement beyond "shoot the other guy." Real combat is nothing like this, and the DARPA team was the first to say so. What AlphaDogfight proved wasn't that AI can fight better than humans. It proved that AI can optimize within a constrained physics simulation better than humans can. Which is both less scary and more interesting than the headlines suggested.
Because the AI trained in a simulator. Specifically, it trained in something functionally identical to what we build — a physics-based model of aircraft performance with blade-element aerodynamics and realistic control dynamics. The AI learned to fly by crashing millions of times in simulation and gradually learning what works.
That's exactly how X-Plane works. When I build the SF-260 for X-Plane, the flight dynamics are physics-first. The simulator doesn't tell the aircraft how to behave — it calculates forces from geometry. An AI training in X-Plane would learn from the same physics that teach human simmers. The AI's advantage isn't better physics. It's infinite patience and zero fear of crashing.
This connects to something I've been thinking about since I started playing submarine simulations. Naval warfare is the original domain of AI-assisted decision making. Torpedo firing solutions have been computer-calculated since World War II — the real innovation in the Silent Hunter series was modeling those analog computers accurately and letting players interact with them.
Modern submarine combat is even more automated. Sonar processing, torpedo guidance, tactical situation assessment — these are already AI-adjacent systems. The step from "computer assists human" to "computer decides, human supervises" is smaller in naval warfare than in air combat because the decision timelines are longer. A submarine engagement unfolds over hours. A dogfight unfolds in seconds. AI needs millisecond reaction times for dogfighting but merely good analysis for submarine warfare.
Two things. First, flight simulators are now dual-purpose tools. They train human pilots AND they train artificial pilots. X-Plane's physics engine was built for humans, but it's equally valid for machine learning. As AI flight becomes a real industry, the demand for accurate flight simulators will grow, not shrink. Better physics models mean better AI training, which means better autonomous systems.
Second, the line between "game" and "training tool" is disappearing. The simulation that trained AlphaDogfight's AI is functionally similar to DCS World, which is a commercial product anyone can buy. The simulation I build for hobbyists is architecturally similar to what professional training organizations use. That convergence is accelerating, and it validates what we've been doing for decades: building accurate, physics-based simulations because accuracy matters.
DARPA's AI won a dogfight in a simulator. The interesting question isn't whether AI will replace human pilots — it probably will for some missions. The interesting question is whether the simulators we build today are accurate enough to produce AI that works in the real world. And from what I've seen, the answer is getting closer to yes every year. For someone who builds aircraft models, that's simultaneously exciting and a little terrifying.
Javier Rollon develops aircraft for JRollon Planes. Follow on Twitter.