Inside the special F-16 the Air Force is using to test out AI

Inside the special F-16 the Air Force is using to test out AI
The X-62 performs a test flight over Edwards Air Force Base

The variable In-flight Simulator Aircraft (VISTA) flies in the skies over Edwards Air Force Base, California, on August 26, 2022. The aircraft was redesigned from NF-16D to the X-62A on June 14, 2021. (Kyle Brasier/US Air Force)

EDWARDS AIR FORCE BASE, Calif. — During a Dec. 9 flight over the Mojave Desert, an Air Force pilot handed control of a special, highly-modified F-16 over to an artificial intelligence system that autonomously piloted the plane. Less than two hours later, that same F-16 took to the skies again for a second flight test of a completely different AI.

With those first two flights, the Air Force has now unlocked a new capability that will allow the service to rapidly flight test autonomy software regardless of which organization or company has developed it, make quick improvements to the algorithms based on the results of that testing, and then reload the AI and fly again within a matter of hours.

Such is the power of the X-62A — also known as the NF-16D Variable In-flight Simulator Aircraft, or more simply as VISTA — a bespoke version of the F-16D that has been flying since the 1990s but recently was upgraded to be a testbed for different autonomy software cores.

With AI still a nascent technology, VISTA’s ability to churn through flight tests could be critical for the Air Force. The service could begin a program of record for a Collaborative Combat Aircraft, its term for an autonomous combat drone, as early as fiscal 2024, Air Force Secretary Frank Kendall said in September. With plans to field CCA drones in the mid to late 2020s,the service will need to do everything in its power to drive out risk and prove that drones can safely and effectively be controlled by AI.

“I can fly the [X-62A] aircraft twice a day,” said Chris Cotting, director of research for the Air Force Test Pilot School, which owns the X-62. “If I want to change something, I just go out to the aircraft and change it, and then the aircraft itself acts as its own ground simulator. So I can very rapidly change software.”

Cotting spoke to Breaking Defense during a Nov. 30 trip to Edwards Air Force Base, just a little more than a week before DARPA’s Air Combat Evolution (ACE) and the Air Force Research Laboratory’s Autonomous Air Combat Operations (AACO) programs demonstrated AI software that could autonomously control the X-62.

At the time, both teams were hard at work, huddled inside a hangar as the aircraft completed ground tests for flights that were originally scheduled for Dec. 1. (An Edwards AFB spokesman said the flights occurred “when all parties were ready to execute” but provided no other details for why testing was delayed to Dec. 9.)

As of Dec. 16, the ACE program had flown about eight sorties aboard the X-62, while the AACO program had completed roughly six sorties, said Terry Wilson, who manages the AACO program.

Both efforts are concurrently using the X-62 as a testbed to try out different autonomy cores that could be used in the future to pilot drones or tactical aircraft without needing a fighter pilot at the stick and throttle. However, each program has developed different AI focused on slightly different problem sets.

For instance, the ACE flights on VISTA have focused on autonomous dogfighting against simulated opponents that were within visible range of the aircraft. On AACO, flight tests centered around basic aviation and navigation tasks, as well as some simulated fights against a virtual enemy located beyond visual range, Wilson said.

So far, flight tests haven’t revealed any “major problems” with the ACE team’s autonomy core, but there have been “some differences compared to simulation-based results,” Lt. Col. Ryan Hefron, program manager for DARPA’s Air Combat Evolution program, said in an emailed statement on Dec. 17. “This highlights the importance of not only flight testing advanced autonomous capabilities, but doing so on testbeds like VISTA which allow us to rapidly learn lessons and iterate at a much faster rate than with other air vehicles.”

A Human Fail-Safe

VISTA began its life in the 1990s as an F-16 that can be programmed to simulate the flight controls and handling of other aircraft — allowing test pilots to understand how it feels like to fly a C-17 or a business jet while in the confines of a fighter, or being used to prove out the flight control laws of a demonstrator aircraft as it did during the Joint Strike Fighter competition.

VISTA gets a new look

After recently receiving a new look and modifications at the Ogden Air Logistics Complex, the NF-16D known as VISTA (Variable stability In-flight Simulator Test Aircraft), departs Hill Air Force Base, Utah, Jan. 30, 2019. (Cynthia Griggs/US Air Force)

That human pilot is one of the reasons the X-62 is an optimal aircraft for proving out AI algorithms, as the operator can step in at any time and retake control of the aircraft, thus reducing risk and allowing the ACE and AACO programs to test more frequently, Cotting said

The X-62 accommodates two people. In the back sits a highly skilled F-16 instructor pilot who has received special training to fly VISTA — a feat that only three people are currently qualified to do. At any point during a flight, the pilot can shut off all of the simulation or autonomous systems, causing the jet to revert back to normal F-16 flight controls.

Anyone who can pass a flight physical can sit in the front seat of the X-62 — the seat requires no special flight qualifications — but typically that position is taken by a flight engineer who controls the simulation systems, or more recently, by someone overseeing autonomous operations of the aircraft.

Before the ACE and AACO programs could use VISTA as a testbed for its AI algorithms, the Air Force spent more a year upgrading VISTA with new hardware and software.

Over the past year, the service replaced the VISTA Simulation System, which allows the plane to mimic the flight controls of other aircraft, with a new, modernized version of the capability. The Air Force also upgraded the X-62 with the System for Autonomous Control Simulation (SACS), a Lockheed Martin-made computer architecture that allows the X-62 to host different autonomy cores developed by various organizations.

“That now allows us to do AI work. It allows us to attach sensors to the aircraft,” Cotting told Breaking Defense during the November visit to Edwards. “It really is a Swiss army knife that we can use to attach lots of different things to the airplane.”

In a typical test flight, the pilot gets the X-62 to the right altitude and speed. Then the person in the front seat loads up the simulation system for whatever aircraft flight controls are going to be used during the exercise, and the pilot engages the simulation. (Even if the F-16 is the aircraft being evaluated, a simulated version of the jet is used in order to eliminate the variances in performance that a specific F-16 might have compared to the rest of the fleet, Cotting said.)

Those same steps are followed when the X-62 flies tests in support of the ACE and AACO programs. Then, after the simulation is engaged, the person in the front seat uses a tablet to load a specific autonomy solution into SACS. Once that’s ready, transitioning to autonomous flight is as easy as pushing a button in the cockpit, Cotting said.

“We have the autonomy test set up so that we’re going to fly pattern one, pattern two, pattern three, pattern four. And then we just select which one we want to do,” he said. During those tests, the AI agent is also able respond to virtual enemies and wingmen. Currently, those entities are all controlled by a computer, but future flights could involve human pilots in a simulator flying a virtual aircraft alongside the X-62.

The end goal would be to autonomously fly the X-62 in a demonstration with human pilots in real fighter jets, acting either as adversary forces or wingmen, Cotting said. “But that’s still quite a ways out,” he added. “I don’t want to give any dates yet, because it really kind of depends on how we mature.”

Updating VISTA 

VISTA’s new autonomy test mission is another lifeline for the aircraft, which narrowly escaped being sent to the boneyard in 2001, when the test pilot school took custody of the aircraft. In recent years, its systems “began to degrade” and needed to be modernized, Cotting said.

Over the past couple of years, the Air Force Research Laboratory invested about $15 million into upgrading VISTA — which was redesignated as the X-62 in 2021 to highlight its role as an experimental aircraft. Aside from adding the SACS system, the Air Force overhauled the X-62’s computers, which “were the equivalent of the old gumdrop iMacs,” Cotting said, referencing the brightly colored Macintosh desktop computers that were sold from 1998 to 2003. VISTA also moved off the VxWorks operating system — which is no longer widely used or supported — to a Linux-based operating system that is more frequently updated.

The Air Force added a programing environment known as Simulink, which allows the development team to do model-based design. It also refreshed the X-62’s simulation capabilities, adding a “model following algorithm” that allows the aircraft to better replicate a wider variety of aircraft. Importantly, the new simulation system is open source and government owned, allowing the Air Force to add to the list of aircraft the X-62 can simulate without having get the permission of the third-party company that owns the software.

After modification work on the aircraft wrapped up this summer, the X-62 proceeded into flight tests this fall to ensure that the F-16’s flight control system still worked as expected with the addition of the new VISTA Simulation System (VSS) and SACS.

“The way the VSS and the SACS work, they are completely independent of the flight control system on VISTA,” Cotting said. “If we violate some constraints that we’ve set up that kind of keep us in a safety sandbox, the VSS automatically disengages” and the X-62 reverts back to typical F-16 flight controls. “The testing we did early on was proving that none of that broke.”

After that, the team set about evaluating the performance of the new VSS, testing whether the aircraft could replicate the flight controls of General Atomics MQ-20 drone, a Learjet 25 business jet and a simulated version of an F-16, Cotting said. Replicating the F-16’s control system with a simulated version essentially acts as a control variable, allowing the team to easily compare the performance of the simulation with the performance of the actual F-16.