Much of the initial reporting focused on the unambiguous final result when AI defeated the human pilot in each of their five dogfights. Here, as in the past, when such a decisive result occurs, some herald it as the end of an era and the dawn of a new one, like the shift from cavalry to tanks.
Conversely, skeptics highlight the unrealistic conditions that applied to the test, such as the fact that the Alpha Dogfight Trials (ADT) used “perfect” data during the scenario conditions, a fact that any experienced pilot or controller would identify as unrealistic. In the ADT, this meant that a kill was adjudicated by reaction time in close quarters, which gives a significant inherent advantage to the AI.
The art of dog-fighting can be distilled into four steps: Observe your opponent for both obvious and subtle cues, Orient the bandit’s maneuvering in relation to your jet, Decide what to do, and then Act to defeat him. You can define the rapidly hanging dynamics associated with fighting called a thinking enemy, particularly in a highly maneuverable fighter, the OODA Loop, for Observe, Orient, Decide and Act.
The software loaded into computers we now refer to as AI can readily handle the last two steps, but what feeds that “intelligent” system with the ever-changing cues for the “observe” and “orient” steps? The only operational fighter with built-in technology that can “see” other aircraft is the F-35, and while it has sensors that can optically capture an opposing platform in any quadrant, it sees with an acuity of only 20-30.
Even then, it cannot discern the type of bandit it is fighting—its configuration, aspect, heading-crossing angle, or the “rate” at which its nose is tracking—unless the bandit is in front of the F-35 where its radar can track the target. Even then, it gets only some of what it needs.
Lacking the ability to independently observe and orient means AI cannot feed itself the inputs required for the programmed elements to kick in and defeat an adversary in the decide and act steps. So, how did the AI simulation gain those details in this fight with Banger?
That information came not from some visual interpretation of the other aircraft on the simulator’s screen, but from “perfect information” supplied by the simulator. The exact range, altitude, airspeed and nose position of the manned fighter is calculated and immediately fed into the AI simulation. That level of clarity can never be gained in a dynamic, neutral fight — but with it, even humans are hard to beat.
Unless you have super vision, it’s really hard to keep track of two aircraft that are maneuvering to kill you, so if everything else is equal in a two versus one (2-V-1) scenario, the “1” always gets killed. On that particular day, we had problems with one of the radios we would normally use, which forced adversaries to talk “amongst themselves” on the same radio pilot was using. To successfully prosecute a 2-V-1 attack, coordination and communication is critical. Every time one fighter would take his nose off of me to get more airspeed and maneuvering space, that pilot would tell the other pilot so he could pitch back in to attack.
Cues received in real-world combat sorties pale in comparison to the perfect information the DARPA simulation fed to Banger’s AI opponent in their virtual fights. There was no need for the machine to “look outside” and find Banger, then try to assess how much airspeed he had, when his afterburner was cooking, when he went to idle with his speed brakes deployed, or how much “G” he was pulling. The simulation fed all that info to the AI fighter in real time.
While those might seem like petty elements, the observe and orient steps are the heart and soul of dogfighting – the two most critical elements in the OODA sequence and there is no system in the world that can touch a human’s ability to capture and process those tasks.
These artificialities aside, DARPA appropriately chose a technically challenging but simplified tactical problem for this cutting-edge experimentation in air warfare. What then should we learn from the experiment?
High-profile DARPA experiments, like the ADT, are critical catalysts to stimulating technology and industrial ecosystems, while also pushing the boundaries of the state-of-the-art, embracing competition and learning, and inspiring the wider technology community.
Services would benefit from more consistently embracing this type of approach to promote innovation and progress, along with more acquisition programmatic “on ramps,” so mature technologies can be included in critical weapon system upgrades.
Ideas of individual combat loom large in the military aviation community’s ethos of “aerial knights” dueling in the sky, using quick reaction maneuvers in close proximity to win. However, 1 v. 1 aerial gun-based dogfighting, or even short-range missiles, are increasingly a relic of air-to-air combat from the age before sensors and missiles grew in range, sophistication and lethality.
For several decades now, technology advancements have already enabled air warfare to evolve from dogfighting to beyond visual range missile engagements. Future air warfare scenarios are unlikely to include 1 v. 1 dogfighting between aircraft using guns, due to advances in sensor ranges and fusion, along with network-enabled weapons and cooperative teaming, that are already fielded on 5th generation fighters.
Weapons and deployable platforms using increasingly sophisticated combinations of these technologies will be able to more easily kill a target or team together to achieve more advantageous positioning for a successful kill.
Further efforts to incorporate uncertainty, such as fuzzy logic controllers, will make simulated combat conditions and performance with AI more realistic and enable effective transition to real-world conditions.
Although excited speculation continues about how AI will replace a pilot in the cockpit and thus enable an unmanned fighter aircraft to pull many more g's than a human, this advantage already exists today: a human deployed missile can pull many more g’s than a fighter aircraft.
Such considerations aside, the types of AI approaches that ADT demonstrated are also a valuable way for DoD to define for industry how to further enhance a missile’s ability to dogfight with a target. Using AI’s strength to assess aircraft maneuvers and transitions will allow pilots to have a higher probability of kill from a wider range of conditions, including at the boundaries of a weapons engagement zone. AI's maturation has been, and will likely continue to be, much more evolutionary than revolutionary.
As aviation technology advanced, complex mechanical systems were replaced by analog switchboards in the cockpit, requiring an aircraft second-seater. Subsequently, digitized cockpits with sensor readouts, autopilot, and automated navigation followed. While current systems even include automated take-off and landing, integrated displays to aid with mission planning and weapons selection, as well as coordination across flight groups.
Correspondingly, the preponderance of pilot workload for advanced aircraft continues to shift from being primarily about how to best fly the plane, instead of using the aircraft’s own sensors and weapons, in conjunction with other offboard assets most effectively in support of the Joint Force.
This crucially and fundamentally shifts pilots' emphasis from manning the equipment (the fighter) to equipping the pilot to perform a wider range of functions in more lethal and effective ways.
The ultimate goal of using AI in warfare is to provide decisive advantage in an engagement to achieve victory. The ADT and broader DARPA ACE program is a crucial catalyst to spawn advancements for “pilot assist” technologies, just like driver-assist technologies continue to lay a crucial foundation for future driverless cars.
Today, AI can become the virtual “second seater,” able to navigate and perform complex flying functions, while the human pilot retains more of a Weapon Systems Officer (WSO) focused role. The AI virtual second seater function will enable algorithms to learn from the human operator, further building trust, sophistication, and capability in the near-term. To further build trust with human pilots, ensure algorithm explainability, and enhance AI learning and capability, the Services should establish a version of mission debrief for AI in both virtual and real-world employment.
The use of AI based pilot assist technologies will enable a human pilot’s role to shift to a weapons version of air traffic control: a combination of local air battle manager and platoon leader in the sky. As a local aerial mission commander, the human pilot’s focus can be on directing a pack of other assets– either as platforms, or network-enabled weapons with smart control, such as DARPA’s Gremlins or the Air Force Golden Horde programs.
As a local air warfare conductor, the pilot can be close to the action, but from a safe vantage point, allowing for unmanned platforms and weapons to synchronize and improvise. This will necessitate a radical shift in pilot training to focus on cultivating mission command at a higher echelon level, but at a much earlier stage in a human pilot’s career.
Organizationally, this will also require a reformulation of the concept of the Air Force “Flights” echelon to one in which a human oversees and directs each Flight but delegates AI control over each of the corresponding supporting "Elements." This approach parallels and can leverage ongoing developments in other sectors, such as industrial robotic manufacturing, autonomous shipping flotillas, and autonomous trucking and taxi fleets.
In each of these cases, automation and AI enables lower-level tasking to be performed, while abstraction enables a human to oversee the coordination and conduct of multiple assets in a feasible and constructive way.
The Services continue to have revolutionary visions of blended flight groups of manned fighters and loyal wingmen, or fully autonomous multi-role fighters. The challenges to achieving these visions are not insurmountable, but fully realizing them will not be as quickly as desired.
However, a significant evolution of air warfare is already underway. Over the last decade, the rapid convergence of key technologies is occurring beyond AI and machine learning. These technologies include fusion engines, miniaturized active electronically scanned array (AESA), onboard processing, advanced displays, and interfaces, swarming coordination, sophisticated multivariate recommendations, palletized munitions, controllers for many vs. many, abstraction layers, and systems to enable conversion of legacy platforms to unmanned (e.g., the F-16 to a QF-16).
The combination of these technologies will soon allow a pilot to lead a force of previously obsolete converted unmanned aircraft into battle, loaded with network-enabled weapons, and re-enforced by palletized munitions deployed from rearward air bastions of relative safety.
Select unmanned squadrons, or swarms of munitions, will be able to engage in specific deconflicted sectors of operation while operating at appropriate levels of balanced trust and risk. In parallel to these technologies, a more pragmatic conceptualization is required for how AI can be best used.
A.I is not a pilot replacement but can be a virtual second-seater, enabling rapid and continuous evolution of human-machine cooperation for decisive advantage. Instead of considering options for either human pilots or AI-based replacements, DARPA and the Services, through a symbiosis of commercial and military partnership, should focus on advancing virtual second-seaters for air warfare victories into the next decade and beyond.
1. Collapse Human-Machine decision/action loop operates across time/space
2. Transform Dated Platforms into auto, manoeuvre, attritable
3. Swarm Potential coordinates operation in contested airspace
4. Individual Systems/Platforms capability overmatched by networks
5. Ability to discriminate action decisions to counter adversary systems
6. Fuses/Integrates data to identify/simulate web of probabilites/confidence
7. Extraction of topical hierarchies control systems with reinforcement
8. Enables Simulation-based prediction and search
9. Interaction of intel/autonomous blocks makes upgrades possible
10. Must fast-track AI applications into existing and new platforms