Following the overwhelming success of artificial intelligence in dog fights against human pilots last year, the US Defense Advanced Research Projects Agency (DARPA) started preparations for the flight of a full-scale jet trainer that will be piloted by AI by 2023.
There are, however, several milestones that the agency’s Air Combat Evolution (ACE) program has to achieve before the flight can take place.
The program is currently half way through Phase 1 and the agency expects live subscale aircraft dogfights to take place later this year.
ACE has so far achieved advanced virtual AI dogfights involving both within visual range (WVR) and beyond visual range (BVR) multi-aircraft scenarios with updated simulated weapons; live flights of an instrumented jet to measure pilot physiology and trust in AI; and initial modifications to the first full-scale jet trainer scheduled to host an onboard AI “pilot” in Phase 3 of the program.
“Our biggest focus at the end of Phase 1 is on the simulation-to-real transition of the AI algorithms as we prepare for live-fly sub-scale aircraft scenarios in late 2021,” said Col. Dan “Animal” Javorsek, program manager in DARPA’s Strategic Technology Office. “Managing this transition to the real world is a critical test for most AI algorithms. In fact, prior efforts have been brittle to just these types of transitions because some solutions can be over reliant on digital artifacts from the simulation environment.”
The goal of the ACE program, which kicked off last year, is to develop trusted, scalable, human-level, AI-driven autonomy for air combat by using human-machine collaborative dogfighting as its challenge problem.
In August 2020, the Johns Hopkins Applied Physics Laboratory (APL) executed the ACE program’s AlphaDogfight trials, a competition of eight teams whose AIs flew simulated F-16s in 1-v-1 aerial dogfights, developed by APL. The champion AI then flew five simulated dogfights against an experienced F-16 fighter pilot in a simulator, beating the human 5-0.
In February this year, the ACE algorithm-development teams completed the next level of simulated AI dogfights in Scrimmage 1 at APL. APL has continued to design and extend the simulation environment for this phase of the ACE program. Teams demonstrated 2-v-1 simulated engagements with two friendly “blue” F-16s fighting as a team against an enemy “red” aircraft. This marked the first AI scrimmage following the AlphaDogfight trials and introduced more weapons into the mix – a gun for precise, shorter-range shots, and a missile for longer-range targets.
“Adding more weapon options and multiple aircraft introduces a lot of the dynamics that we were unable to push and explore in the AlphaDogfight Trials,” Javorsek said. “These new engagements represent an important step in building trust in the algorithms since they allow us to assess how the AI agents handle clear avenue of fire restrictions set up to prevent fratricide. This is exceedingly important when operating with offensive weapons in a dynamic and confusing environment that includes a manned fighter and also affords the opportunity to increase the complexity and teaming associated with maneuvering two aircraft in relation to an adversary.”
Another major focus of the ACE program is measuring pilot trust in the AI’s ability to conduct combat maneuvers while the human on board focuses on higher-cognitive battle manager decisions.
To begin capturing this trust data, test pilots have flown several flights in an L-29 jet trainer at the University of Iowa Technology Institute’s Operator Performance Laboratory. The two-seat jet is outfitted with sensors in the cockpit to measure pilot physiological responses, giving researchers clues as to whether the pilot is trusting the AI or not. The jet is not actually flown by an AI; rather a safety pilot in the front cockpit acts as a “human servo actuator” executing flight control inputs generated by an AI. To the evaluator pilot in the backseat, it appears as if the AI is performing the aircraft maneuvers.
“In an analog to the miles per disengagement used in self-driving cars, we are recording the time per disengagement which serves as a major metric for the program. In addition, we’ve started looking at measurement techniques to see where the evaluation pilot’s head is pointing, as well as where their eyes are looking around the cockpit,” Javorsek said. “This enables us to see how much the pilot is checking on the autonomy by looking outside the window, and comparing that to how much time they spend on their battle management task.”
The ACE program is also researching two independent frameworks for an AI battle manager in what has been deemed the AlphaMosaic agent for BVR and campaign scale command and control. Scrimmage 1 saw these two agents engaging in more complex multi-domain cruise missile defense (CMD) scenarios created by APL. The two contractor teams continue to refine their agents and architectures in preparation for their Scrimmage 3 head-to-head competition at the end of the phase.
Looking beyond the subscale aircraft tests in late 2021, ACE performer Calspan has begun modifying the first L-39 full-scale jet trainer that will be piloted by the AI in live-fly team dogfights during Phase 3 of the program in late 2023 and 2024.
The first step is to create an accurate aero performance model of the L-39 that the AI algorithm can use to make predictions as well as tactical maneuver decisions. Once the aero model is complete, the L-39 will be fully modified so that the AI will be able to take control of the aircraft.
DARPA awarded Calspan $14.1 million in July 2020 for the modification of up to four Aero Vodochody jet trainers with autonomous fly-by-wire flight control system technology to allow implementation and demonstration of advanced human machine interfaces (HMI) and AI algorithms.