Chapter 6. DARPA's Challenge: Can AI Defeat a Human Pilot?

DARPA’s challenge: Can it beat human pilots? Somewhere in the Pentagon building, in the fall of 2019, a question was asked: Can machines really beat human fighter pilots in the sky? This question was not just curiosity. This question posed by the U.S. Defense Advanced Research Projects Agency, commonly known as DARPA, was as fundamental a challenge as Alan Turing's question 70 years ago: "Can machines think?" DARPA is a strange organization. On the surface it looks like a bureaucratic government agency, but in reality it's more of a mad scientist's playground. It is here that the Internet was created and GPS was introduced to the world.

It was they who sowed the seeds of stealth fighters. It's not wrong to say that it's the only government agency that bets on the seemingly impossible and says it's okay to fail. This time, they challenged an area dominated by human instinct and intuition: aerial combat. It was the beginning of a program called ACE (Air Combat Evolution) in English. The program manager was Colonel Dan Javorsek. His call sign is ‘Animal’. As a former F-16 pilot, he knew the psychology of his fellow pilots better than anyone else. What kind of people are fighter pilots?

These are people who take pride in holding the steering wheel and fighting against their enemies. The idea of ​​entrusting one's life to a machine is almost an insult to them. Colonel Animal decided to break through this resistance head on. He explained the meaning of this challenge with a historical analogy. In 1939, U.S. Army Chief of Staff General George Marshall asked Cavalry Chief John Herr how he would respond to the German blitzkrieg. General Herr's answer was as follows. They said they would put horses on trailers and move to the front lines to save stamina and then overwhelm tanks on the battlefield.

Of course, history did not go as he expected.

The cavalry was gone, and tanks dominated the battlefield. Animal warned. If today's fighter pilots do not want to become the cavalry of the 21st century, they must embrace a new tank called AI. But how can we break the pilots' distrust? It is better to show once than to explain a hundred times. Isn’t there an old saying that a hundred words are worth a sight? Animal decided to show AI beating the best human fighters in a one-on-one dogfight. That is the background of the birth of Alpha Dog Fight Trial. Why did it have to be a dogfight, or a close-range dog fight?

Modern air combat is increasingly involving firing missiles from long, invisible distances. This is called BVR, or Beyond Visual Range. But why did they choose a World War II-style close-range dogfight? There was a deep reason for this. First, dogfights are a closed world problem. Like Baduk, the rules are clear, but the number of possible cases is almost infinite. This environment is optimal for AI to develop its skills through reinforcement learning. DARPA saw Dogfight as a gateway to more complex air combat missions. Second, dog fights are an extreme test bed for the OODA loop.

A dogfight is a place where the decision-making process of observing, setting a direction, deciding, and acting takes place in a split second. Here, surpassing humans means proving that AI's computational speed and judgment ability have exceeded human physiological limits. Third, it is also a starting point for building trust. Pilots develop their basic skills through dog fights starting from training camp. If AI surpasses humans in this most basic yet most instinctive area, pilots will have no choice but to acknowledge AI. The Johns Hopkins Applied Physics Laboratory, or APL for short, played a key role in this competition.

They created an AI arena named Colosseum. It is named after the amphitheater where gladiators fought in ancient Rome. This system, which combines open source flight dynamics software JSBSim with middleware, autonomous algorithms, and visualization software developed by APL, can simulate faster than real time.

I was able to turn it. AI agents learned to fight by dying and being reborn tens of millions of times in this virtual sky. DARPA selected eight teams in August 2019. The so-called Elite 8 came from diverse backgrounds. They included defense giants such as Lockheed Martin and Boeing, small AI specialized companies such as Heron Systems and Physics AI, and university research teams such as Georgia Tech Research Institute. Their task was to fully understand the flight dynamics of the F-16 fighter jet, perform basic combat maneuvers to outsmart the opponent, and develop algorithms for shooting down enemies with cannons.

Missiles have been ruled out. It was the most primitive form of fighting, with only the aircraft's maneuverability and machine gun aiming ability determining victory or defeat. The competition took place in three stages. The first trial in November 2019 was an exhibition match. At this time, the AIs could not even fly properly and often fell headlong into the ground. Researchers recalled, "In the beginning, AI didn't even know how to fly an airplane. It was just lucky if it didn't crash." But in just a few months, things changed dramatically.

In the second trial in January 2020, the AIs have already started imitating the basic tactics used by human pilots. As the program was extended due to the COVID-19 pandemic, the 2.5th trial was added virtually in May, and by the finals in August, they began to perform maneuvers that humans could not have imagined.

DARPA Deputy Director David Honey said in his opening remarks at the convention: “When putting together this program, there were important questions that had to be answered: We need to understand whether AI autonomous algorithms can actually operate in the very challenging environment of air-to-air combat.” The competition attracted tremendous interest. About 10,000 people from 93 countries registered to watch, and an additional 5,000 wanted to participate. What is unique about Alpha Dog Fight Trial is that it adopts the format of an e-sports competition.

The APL team created a corner called the Control Zone and prepared a commentary area modeled after ESPN's Sports Center. Air combat and autonomy experts discussed the basics of AI and dogfighting, how to train AI and human pilots, and provided educational and interesting analysis and commentary. This approach is

It reflected DARPA's intention to engage the broader public beyond the general defense community. This was not just a skills competition. It was the first official test in the battle for control of the sky, which humanity had monopolized for a long time. Can AI surpass humans in the speed race of the OODA loop? Can machine calculations win in areas where human intuition dominates? In August 2020, the world was waiting for the answer.