The U.S. Air Force showcases its first AI fighter jet with high profile! The minister personally conducted the test drive without interfering during the whole process, and 100,000 lines of code were tested for 21 times.

Recently, the military circle has been overwhelmed by the news: US military fighter jets can now complete fully automatic air combat using AI.

Yes, just recently, the US military’s AI fighter jet was made public for the first time and the mystery was unveiled.

The full name of this fighter jet is the Variable Stability Flight Simulator Test Aircraft (VISTA). It was personally flown by the Secretary of the US Air Force and simulated a one-on-one air battle.

On May 2, U.S. Air Force Secretary Frank Kendall took off in an X-62A VISTA at Edwards Air Force Base

Note that during the one-hour flight, all flight movements are completed independently by AI!

Kendall said -

For the past few decades, we have been thinking about the infinite possibilities of autonomous air-to-air combat. Potential, but it always seemed out of reach. Now, however, we have reached a moment of transformation, made possible by the breakthrough achievements of the ACE team.

The Secretary of the Air Force personally conducted a test flight

Under the blazing sunshine at noon, an aircraft The ship takes off with a roar - it's controlled by AI, not a human pilot. The F-16 fighter jet took off with a roar and flew into the distance.

During the test flight, U.S. Air Force Secretary Frank Kendall sat in the front seat of the aircraft and, accompanied by a safety pilot in the back seat, was able to operate the X-62A without touching the controls. Completed a set of test tasks.

AI controls the F-16 to fly at a speed of more than 550 miles per hour, which makes Kendall's body directly withstand five times the center of the earth. Gravitational pressure.

It approached an F-16 flown by a human pilot. The two aircraft were racing within a distance of only 304 meters apart, trying to force the other side into a weak position by twisting and circling.

After the one-hour flight, Kendall walked out of the cockpit with a smile.

He said that he had seen enough evidence during the flight that he was willing to trust this AI and let it decide whether to launch nuclear weapons in a war.

The AI ​​software on the aircraft will first learn from millions of data points in the simulator, and then verify the conclusions during actual flights. These real-world performance data are then re-entered into the simulator, where the AI ​​processes the data for in-depth learning.

VISTA’s military operator claimed that the United States alone has similar AI aircraft in the world.

The first dog fight between AI and humans

As early as April 2024, the U.S. Air Force Test Pilot School and DARPA announced , they have completed the first AI-human dog fight.

The dog fight was completed by an F-16D (Block 30) two-seat aircraft code-named X-62A VISTA and a human pilot flying an F-16 fighter jet.

In the dog fight, the X-62A demonstrated defensive maneuvers, attack trembling and other skills, while the human pilot sitting in it could take over the AI ​​system without activating the safety switch.

The two aircraft demonstrated a "high-angle nose-to-nose battle". The relative speed directly reached 1,200 miles per hour, and the closest distance reached about 610 meters.

The X-62A’s safe and autonomous dogfighting aboard another manned aircraft is a major milestone not only for ACE, but for the entire A milestone in autonomous flight.

Moreover, DARPA and the Air Force emphasized that although dog fighting is the core of this test, the goal of ACE is far from being as simple as dog fighting.

Bill Gray, chief test pilot of the USAF TPS of the U.S. Air Force, explained: Dog fighting is a problem that needs to be solved, so they began to test autonomous AI systems in the air.

Every lesson we learn applies to any task you can give an autonomous system.

X-62A and F-16 during simulated dog fight

21 test flights, 100,000 lines of code

VISTA, also known as X-62A, is a modified version of the General Dynamics F-16D.

The U.S. Air Force has been using it to test advanced technologies since the 1990s.

Today, VISTA has been integrated into DARPA’s Air Combat Evolution Program ACE, equipped with machine learning and specialized software.

VISTA was developed by Lockheed Martin, a defense technology company.

VISTA can be regarded as an AI pathfinder, which uses a new driverless design.

This work is crucial to realizing distributed teams.

The efficiency of the team is also amazing. In less than a year, the initial real-time AIAgent was installed in the X-62A system and demonstrated the first AI Dogfight with humans, and over 21 test flights, more than 100,000 lines of critical code changes to flight software were completed.

X-62A cockpit during last year’s simulated dog fight

Lockheed· Vice President Martin said that X-62A VISTA is an important platform for them to develop, test, integrate AI, and establish AI certification standards.

Moreover, X-62A VISTA will completely change the future of aerospace.

The hardware and software architecture in the platform are proven and provide a safe and controllable environment for AIAgent and advanced algorithms, allowing them to quickly prototype and develop.

This open architecture enables highly complex testing by leveraging Skunk Works’ Model Tracking Algorithm (MFA) and Simulated Autonomous Control System (SACS).

These important updates not only enhance VISTA's capabilities, but also maintain its advantage for rapid prototyping.

As a result, the team was able to quickly change the software and conduct frequent flight tests.

Test after test has proven that this architecture is powerful enough to reliably transition to third-party distributed hardware and replicate the safe and controllable flight test goals demonstrated by VISTA.

AI has achieved a breakthrough impact in autonomous air combat

This April’s first test between the X-62A and the manned F-16 The battle can be described as a milestone event.

It was this incident that made machine learning in 2023 a reality in the air.

The future U.S. drone program and the U.S. Air Force's collaborative fighter program will be directly affected.

Shield AI, a participant in the ACE program, acquired Heron in 2021.

The AI ​​pilot developed by Heron had won DARPA’s dog fighting test the previous year.

As early as 2022, DARPA worked with the Air Force and Lockheed Martin to integrate AIAgent into the X-62A system, and in December of that year, the jet was used for the first time with these algorithms. autonomous test flight.

X-62A/VISTA’s flight system can be configured to imitate any other aircraft model, which allows it to complete a variety of testing purposes, making it an ideal platform to support work such as ACE .

Que Harris, Lockheed Martin’s chief flight control engineer, said that the team has an integration space in VISTA’s flight control that allows AIAgent Send commands to VISTA as if they were sending commands to VISTA's simulation model.

This can be seen as an "autonomous sandbox" inside the jet.

Dr. Chris Cotting, director of USAF TPS, described it this way: It's like you have a simulator laboratory in a research institution.

"We've stuffed the entire simulator lab into the F-16."

As we mentioned above, in the subsequent 21 test flights of the X-62A, Agents needed to be reprogrammed almost every day, and the final code modifications reached 100,000 OK.

The X-62A can support a variety of flight tests by rapidly training and retraining algorithms in a fully digital environment.

No pilot, no communication, fully realized swarm concept

Co-founder and CEO of Shield AI President Brandon Tseng said that the ideal autonomous aircraft is like this-

No remote pilot, no communications, and no GPS. These aircraft implement the concept of a formation or swarm that can carry out the commander's intent.

They can execute missions, working together dynamically, reacting to each other, as well as to the battlefield, hostile threats and civilians on the ground.

And another value reflected by this technology is to make the entire system and the entire fleet better.

At all times, the best AI pilots can be on board the aircraft. When learning on a fleet, there are them on every aircraft.

"You will always have the best quadcopter pilots, the best V-BAT pilots, the best CCA pilots. They will dominate. Then, you will Able to win battles with an incredibly high winning rate.”

Big challenge: you need to trust the algorithm to achieve it in reality

However, this process also presents difficult challenges.

First of all, it is rule-based autonomy.

Dr. Daniela Rus of MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) explains that if you write rules in an "if-then" way, they must be robust. Great sex. Therefore, in order for the system to work properly, a team of experts is needed to generate the code.

Dr. Rus explained that because machine learning relies heavily on analyzing historical data when making decisions, they often find insights that humans cannot detect, or that are unconventional and not rule-based. language expression.

#Machine learning is very powerful in environments and situations where conditions fluctuate dynamically, making it difficult to establish clear and strong rules.

The so-called "environment and situation" are the unknowable independent decisions made by the human opponent in a dog fight.

Given this unpredictability, conducting this type of simulated engagement is extremely dangerous even for well-trained pilots.

Accident and fatality data in dog fighting training for F-16 and F/A-18 fighter jets from 2000 to 2016

U.S. Air Force TPS Commander, Air Force Col. James Valpiani said that many elements involved in using machine learning are not fully understood.

"Understandability and verification hinder us from exploring this field." Currently, there is no civilian or military way to certify machine learning agents for flight-critical systems.

And this is indeed where ACE and the real-world X-62A test flight come into play.

The most important thing about the machine learning agent on the VISTA jet is to learn to prevent the aircraft from performing dangerous and unethical behaviors, including defining the code that allows the flight range to be avoided in the air or on the ground. of collisions, as well as preventing the unauthorized use of weapons.

The U.S. military insists that in the operation of future autonomous weapons systems, humans will always be somewhere in the cycle, but where exactly are they in that cycle? , will evolve over time. This has been the subject of much debate.

"We must be able to trust these algorithms in order to use them in the real world."

## On April 1, 2024, the first batch of F-16 fighter jets will be converted into autonomous test beds in accordance with the VENOM program and arrive at Eglin Air Force Base

With the exception of the ACE program, the X-62A is not The only aircraft used by the U.S. Air Force in recent years with advanced autonomous technology.

Currently, the Air Force is modifying six additional F-16 aircraft, turning them into test aircraft to support a program called Project VENOM for large-scale collaborative autonomy. test.