Part 3: How AI Learns to Fight

1. Tens of Millions of Virtual Deaths

Altitude 25,000 feet. In a simulated sky, an AI-controlled F-16 goes into a steep dive. It turned its nose toward the enemy aircraft but lost speed too quickly. The aircraft stalled and plummeted straight toward the ground. The word "SHOT DOWN" appeared on screen, and the simulation reset.

This AI just died. But one second later, it reappeared in the same sky. And once again charged toward the enemy. This time at a slightly different angle, with slightly different timing.

This process repeats millions of times a day.

A human pilot needs thousands of hours of flight time to become a veteran. Graduate from the Air Force Academy, complete basic flight training, undergo fighter conversion training, get assigned to an operational unit, and serve for years before finally pulling their weight. And the total flight hours a pilot can accumulate in their lifetime amount to perhaps 2,000 to 3,000 hours at most. Because there are limits—physical limits, career limits, and above all, the limit of having only one life. One crash, and it is over.

But AI faces no such limits. In virtual space, AI can die as many times as it wants and be reborn as many times as it needs.

In August 2020, at the AlphaDogfight competition hosted by the U.S. Defense Advanced Research Projects Agency (DARPA), the winning AI from Heron Systems had fought approximately 4 billion virtual battles by the time of the competition. This is equivalent to roughly 31 years of flight experience for a human pilot.

Heron Systems engineer Ben Bell said: "Our AI was trained fighting against 102 different agents. That is what made it robust enough to handle any opponent."

The principle of reinforcement learning is surprisingly simple. It resembles how a baby learns to walk. No parent tells a baby, "Bend your knee at this angle and shift your center of gravity forward." The baby simply stands up and falls, stands up and falls, hundreds of times. Then one day it takes a step, the parents applaud, and it thinks, "Ah, that must be right." Falling hurts; walking gets praise.

AI reinforcement learning works the same way. When first dropped into a virtual world, the AI does not even know what an airplane is. It has no idea how to move the stick or when to fire missiles. It just tries things randomly. Most of the time it crashes or gets shot down. Each time, the system gives the AI a negative score. Conversely, when it gets on the enemy’s tail, lands a gun hit, or shoots down an opponent, it receives a positive score. The AI’s sole objective is to maximize its cumulative score.

The key here is the design of the Reward Function. If you simply say "shoot down the enemy," the AI might learn kamikaze tactics—charging suicidally in mutual destruction. So engineers must craft highly sophisticated reward systems. Shoot down the enemy but survive. Earn bonus points for getting on the enemy’s six o’clock. Lose points for wasting energy on reckless maneuvers. Receive heavy penalties for entering civilian zones.

Lockheed Martin’s AlphaDogfight team designed such a reward system with advice from a retired F-16 pilot. Decades of a pilot’s experience were translated into equations and weights.

In the AlphaDogfight final, Heron Systems’ AI defeated human pilot "Banger" 5 to 0. One moment was particularly shocking. The AI fired its cannon while flying head-on toward the human. This is called a "Head-on Gunshot." Human pilots instinctively avoid such maneuvers out of collision fear. Training regulations prohibit it. But through tens of millions of virtual deaths, the AI had discovered on its own that "if I fire precisely at 0.1 seconds before collision, I can shoot down the enemy before I die."

No human taught it this tactic. It was a formula for victory that the AI discovered through infinite trial and error.

This process is called "Curriculum Learning." If you pit the AI against a veteran pilot from the start, it learns nothing and just keeps dying. So it begins by learning to maintain level flight. Then it practices aiming at stationary targets. It gradually faces tougher opponents. In the final stage, it fights against itself—or past versions of itself. This is called "Self-Play." It is the same method AlphaGo used to conquer Go.

Ultimately, reinforcement learning begins with the "freedom to fail." A real pilot’s single failure means death. But in a virtual world, the AI becomes immortal through tens of millions of deaths. It compresses a generation’s worth of human experience into mere days.

This is the true reason the AlphaDogfight AI could overwhelm a veteran pilot. AI is not simply fast at calculations. It accumulated a volume of experience through tens of millions of deaths that no human could amass in a lifetime.

2. Transplanting Simulation Experience into Reality

If we put an AI that dominates virtual combat into a real F-16 cockpit, would it instantly rule the skies? The answer is no.

The sky in a simulation is mathematically perfect. Wind blows as calculated, air density is uniform, and the engine always runs at 100% efficiency. Sensors send clean data without noise, and communications have zero delay. But reality is entirely different. Gusts blow, clouds obscure sensors, communication signals drop, and the airframe vibrates unpredictably. An AI that boasted a 100% win rate in simulation might crash on its first real flight.

This gap is called the "Reality Gap." The technology that bridges this gap is Sim2Real: "Simulation to Reality." It is the core challenge of AI fighter development.

The first Sim2Real strategy is "Domain Randomization"—deliberately messing up the simulation environment. During training, wind intensity varies randomly, the aircraft’s center of gravity shifts slightly, and engine thrust fluctuates between 80% and 120%. Sensor data gets intentional noise, and communication delays are injected. An AI trained in such "dirty" environments perceives the real world as just "another noisy simulation."

The second strategy is the "Digital Twin"—continuously feeding real data into the simulation. Real flight data from actual aircraft is collected in real time to update the virtual aircraft model. As this cycle repeats, the AI grows stronger in reality.

The U.S. Air Force operates the X-62A VISTA experimental aircraft to validate Sim2Real technology. This modified F-16D allows AI to actually control the aircraft. AI-piloted flight tests began in December 2022 over Edwards Air Force Base. In the first 12 flights, the AI performed both close-range dogfight and beyond-visual-range engagement scenarios. During over 17 hours of autonomous flight, the aircraft optimized performance while respecting actual airspace restrictions.

In 2023, the X-62A VISTA, under AI control, engaged in actual aerial combat against a manned F-16—the world’s first real "human vs. AI" aerial engagement.

In May 2024, U.S. Air Force Secretary Frank Kendall personally boarded the X-62A’s cockpit. After the flight, he said: "I think we could trust AI with weapons release authority."

A "Runtime Assurance" system monitors all AI actions in real time. If the AI’s command exceeds the aircraft’s structural limits or leads to a ground collision trajectory, the system immediately cuts AI control and recovers the aircraft to a safe state. In all X-62A test flights, the safety pilot never once had to forcibly shut down the AI.

Future AI fighters will know basic combat skills before takeoff, but will evolve on the spot by detecting enemy patterns during combat. This is called "Meta-Learning"—fine-tuning its neural network in real time based on changes in aircraft condition or enemy tactics.

Sim2Real is the process of a virtual intelligence putting on a physical body and learning the laws of reality. In robotics, this is called "Embodiment." For AI to become a true warrior, it must leave the clean skies of simulation and descend into the rough skies of reality.

3. Sensor Fusion: Integrating Radar, EO/IR, and ESM Data

The secret to winning in aerial combat is simple: see first, shoot first. But "seeing" is more complex than it sounds.

Modern air combat is decided hundreds of kilometers away—before a pilot can visually confirm an enemy. This is "Beyond Visual Range (BVR)" engagement. The problem is that no sensor is perfect.

Radar emits electromagnetic waves and detects their reflections, revealing position and speed. However, stealth fighters minimize radar reflections. Also, radar exposes the emitter’s own position.

EO/IR sensors detect heat from the enemy’s engine. They emit no waves, so they are covert. But range measurement accuracy is lower, and they are vulnerable to bad weather.

ESM passively listens to enemy emissions, analyzing frequency patterns to determine position and type. But if the enemy keeps radar off, ESM is useless.

It is nearly impossible for a human pilot to mentally integrate all this data in a split second. This is the limit of human "Cognitive Load."

AI’s true power emerges here: "Sensor Fusion." AI simultaneously analyzes and integrates data from radar, EO/IR, and ESM to create one comprehensive situational picture. When radar reports "object at 50 km" and IR reports "heat source at 48 km," AI determines whether this is the same target or two different objects, correcting discrepancies into a single "Track" through "Data Association."

The F-35 Lightning II is the crystallization of sensor fusion technology. Its DAS consists of six infrared cameras providing 360-degree imagery on the pilot’s helmet display. Pilots call this the "God’s Eye View."

When the enemy attempts jamming, AI cross-references ESM data with IR imagery. If radar shows a target but there is no heat signature, AI classifies it as a "decoy" and removes it. All of this happens in milliseconds.

Europe’s FCAS and GCAP projects aim for "Cloud Sensor Fusion"—all manned and unmanned aircraft in a formation sharing sensor information. Sensor fusion is the technology that lifts the fog of war. This is the true strength of fifth- and sixth-generation fighters. Not stealth. Fusion.

4. Target Detection, Tracking, and Identification: ATR and Multi-Target Tracking

Altitude 25,000 feet. The sky beyond the canopy stretches bright blue, but beneath that calm lies lethal tension. The radar warning receiver starts beeping. Someone is watching me. Six green dots appear on the tactical display. Friend or foe—I cannot tell. Flying at 1,000 km/h, if one of those dots is an enemy, I have only seconds.

In the old days, fighter pilots relied on the "Mk.1 Eyeball"—the human eye. Even if radar caught something, the only way to confirm was visual identification. The problem was that by then, enemy missiles might already be airborne.

This is where ATR (Automatic Target Recognition) enters. At its core lies deep learning—AI viewing millions of images and learning patterns autonomously. It learns turret shapes, wing angles, exhaust heat patterns of enemy assets.

Synthetic Aperture Radar (SAR) can image the ground through clouds, at night, and in rain. Trained analysts take hours to decode SAR images. AI answers in 0.1 seconds: "Hiding under that shadow is a T-72 tank. Turret rotated 15 degrees. Engine running."

The U.S. Department of Defense’s "Project Maven" demonstrated this power. AI analyzes drone surveillance footage 24/7, flagging suspicious scenes for humans, allowing analysts to focus on the judgments that truly matter.

Multi-Target Tracking handles multiple enemies. In modern warfare, dozens of drones swarm like bees, with real fighters hidden among them. The human brain struggles to track more than three or four objects simultaneously. AI can track hundreds simultaneously, assigning each a unique ID, calculating trajectories, and predicting positions.

In July 2024, the U.S. Air Force launched the "ATA-AI" project—$99 million to develop next-generation tracking technology for stealth aircraft, hypersonic weapons, and drone swarms.

One key algorithm is the Kalman Filter, developed in the 1960s for predicting moving object positions. Modern AI goes further—learning enemy behavior patterns to predict tendencies.

China is integrating electro-optical tracking into the J-20 and developing algorithms to detect U.S. stealth aircraft. Europe’s FCAS pursues the "Combat Cloud" concept—every airborne platform connected in a single network, sharing what each sees in real time.

The laws of air combat remain unchanged: First Look, First Shoot, First Kill. What has changed is speed. Human judgment takes hundreds of milliseconds. AI operates in microseconds. Yet humans remain essential—the final decision belongs to a person. A hunter commanding a powerful hunting dog called AI—that is the future fighter pilot’s role.

5. Explainable AI (XAI)

July 3, 1988. The Persian Gulf. The USS Vincennes’ computer system classified an unidentified track as an Iranian F-14. The captain ordered missile launch. But it was Iran Air Flight 655, a civilian airliner. All 290 passengers and crew perished. The system had displayed incorrect information, and under extreme stress, the crew trusted it blindly. The system never explained why it reached that conclusion. This is the "Black Box" problem.

For combat pilots, trust equals life. If my AI wingman suddenly charges into enemy territory and my radar shows nothing there, I am gripped by fear. Incomprehensible behavior is indistinguishable from betrayal.

In 2020, Heron Systems’ AI "Falco" defeated human pilot "Banger" 5-0 at DARPA’s AlphaDogfight. Falco performed maneuvers no human would attempt. Banger said: "Its gunnery was unbelievably accurate. But I could not predict why it maneuvered that way."

XAI (eXplainable AI) explains in human-understandable language why AI made a particular decision. DARPA launched its XAI program in 2016. Where previous AI would say "Enemy tank, 97%," XAI says: "T-90 tank. Turret shape matches database at 95%. Engine heat shows diesel characteristics. Escort formation matches Russian armored doctrine."

In April 2025, the U.S. Air Force published AFDN 25-1, mandating "transparent and explainable algorithms" in military AI.

"Trust Calibration" is critical—neither over-trusting nor under-trusting AI. AI must honestly disclose its confidence level. French defense company Thales adopted the slogan "TrUE AI"—Transparent, Understandable, and Ethical.

In May 2024, Secretary Kendall boarded the X-62A VISTA while AI performed combat maneuvers. His trust was built only after the AI could explain its maneuvers.

XAI has limitations—generating explanations requires computation that combat cannot afford. Adversarial attacks could exploit revealed reasoning. Ethical questions of responsibility remain.

The principle of "Meaningful Human Control" is emphasized by both the ICRC and the U.S. Department of Defense. But in split-second scenarios with hypersonic missiles and drone swarms, human intervention time may not exist.

XAI is about testing across thousands of scenarios before deployment. It is about enabling AI to say: "This situation is beyond my training. Human, you decide."

Intelligence without explanation is, on the battlefield, indistinguishable from madness. We need not a ghost inside a black box, but a comrade we can trust with our backs. XAI is our effort to create that comrade.