The DARPA Air Combat AI project will soon exit the development stage and start working in the real world.
The Air Force AI project is one step closer to implementation. As part of a series of virtual tests, F-16 fighters, controlled by AI , worked as a team to destroy the enemy. The experiments were carried out as part of the first phase of the Air Combat Technology Development (ACE) program. This program is run by DARPA, and through it, the agency wants to understand how AI and machine learning models can help automate various aspects of aerial combat.
DARPA recently announced that the first phase was half completed and that aerial combat simulations were conducted at the Johns Hopkins Applied Physics Laboratory last month.
Until now, the Air Combat Evolution research program has focused on virtual air combat, but this will soon change.
Using the simulation environment developed at the Johns Hopkins Laboratory of Applied Physics, test battles were conducted in the "2 versus 1" format. They were attended by two F-16 fighters (friendly), which fought against a red (enemy) aircraft belonging to an unnamed enemy.
According to DARPA, the ambitious research program aims to develop “a robust, scalable, and AI-driven aerial combat automation system. This system will have to be based on data from combat, in which people used various systems for automating combat. "
The February tests of artificial intelligence air combat were the first since the AlphaDogfight system tests conducted in August last year. Eight teams participated in the competition, providing AI systems that piloted the F-16 in one-on-one aerial combat. The team with the winning AI ran five more aerial combat simulations against an experienced F-16 fighter pilot in the simulator, defeating a human 5: 0 - a very clear demonstration of the potential of AI, you can read more about it here .
“At the end of Phase 1, we are putting a lot of emphasis on moving AI algorithms from simulations to the real world as we prepare for scaled-down tests at scale at the end of 2021,” said Colonel Dan Yavorsek, Program Manager at DARPA's Strategic Technology Office. “This transition to the real world becomes a critical test for most AI algorithms. Models from past attempts have been limited because they over-relied on digital artifacts in the modeling environment. ”
Compared to the AlphaDogfight Trials, which used only cannons, the Scrimmage 1 tests used missiles for distant targets.
“The addition of weapons and new aircraft models enhances the dynamics that were not achieved in the AlphaDogfight tests,” added Javorsek. “These tests represent an important step in building confidence in algorithms, as they allow us to assess how AI agents are coping with restrictions put in place to prevent friendly fire. This is extremely important when dealing with offensive weapons in a dynamic and challenging fighter environment. We can also increase the complexity of enemy aircraft maneuvers and test how the model reacts to them. "
A pilot fights an AI opponent during AlphaDogfight trials.
So far, the ACE has demonstrated advanced AI-assisted virtual aerial combat. In particular, within the framework of these tests, the use of both weapons requiring the presence of the enemy in the field of view and weapons free from these restrictions were practiced, and simulations of real flights were tested and analyzed in terms of the pilot's physiology and his trust in the AI.
Throughout the program launched last year, DARPA has emphasized the importance of working to build confidence in AI for human pilots. Pilots are expected to allow the system to maneuver while they themselves concentrate on overall battle management decisions.
In the process of “collecting trust data,” test pilots flew an L-29 Delfin trainer jet at the Iowa Institute of Technology Operator Performance Laboratory. Sensors were installed in the cockpits of these aircraft to measure the physiological reactions of the pilot, which allow them to understand whether the pilot trusts the AI. In these missions, the L-29 was operated by a front-seat reserve pilot who entered data for AI-based flight control. The pilot, whose performance was assessed, had the impression that the aircraft was being controlled by an AI.
Test pilots on the L-29 Delfin jet trainer evaluate the pilot's physiological responses to actions carried out by artificial intelligence.
The second phase of the ACE, scheduled for the end of this year, will include aerial combat using real scaled-down aircraft, both propeller driven and jet-powered. In this way, it will be possible to ensure that AI algorithms can be transferred from the virtual environment to the real world. Calspan has also begun work on modifying the L-39 Albatros to incorporate airborne AI. The modified aircraft will be used in the third phase of testing, in which real flights with test fights will be carried out. Phase 3 is scheduled for late 2023 and 2024.
The L-39 Albatros will serve as an onboard AI platform for Phase 3 testing of the research program.
Once this concept is tested, DARPA plans to introduce AI technology into unmanned aerial vehicles such as Skyborg , working in conjunction with manned fighters. Thus, drones will be able to automatically take part in aerial battles, while a human pilot in a manned aircraft will focus primarily on combat control.
Ultimately, this AI could be critical in realizing the dream of a fully autonomous unmanned combat aircraft capable of conducting aerial combat and striking ground targets.... While this device will be able to perform most of the functions of manned aircraft, its "brain" will be able to make key decisions based on much more information in a shorter period of time faster and more accurately, without getting lost in the chaos of combat conditions. Also, these algorithms can be adapted to allow drones to form "flocks" working together. In this way, they will be able to maximize their combat effectiveness, while decisions in such flocks will be made much faster than in formations of aircraft piloted by real people.
A similar artificial intelligence technology is also used as a "virtual co-pilot"- a concept being developed under R2-D2, a program run by the Autonomy Capability Team 3 (ACT3) of the US Air Force Research Laboratory (AFRL). Thus, the software and other systems that emerged from the ACE could potentially provide new types of assistance to the crew of manned aircraft.
It is clear that ACE has the potential to participate in various Air Force programs in the field of autonomous and semi-autonomous unmanned aerial vehicles, as well as accelerate decision-making in manned aircraft. While AI algorithms have proven their ability to win in virtual aerial combat, we should be able to see how this technology works in the real world later this year.
