Autonomy Software Evolves From Military Platforms To Missions

Four jet-powered drones take off in quick succession from a wind-swept, 8,000-ft. runway in the remote Chihuahuan Desert.

The drones coordinate among themselves and quickly form a perfect defensive counter-air patrol over the tumbleweed-filled sprawl of the Anduril Texas Test Site.

• Anduril drones practice air-to-air roles
• Collaborative combat aircraft autonomy project moves forward in secrecy

Then the plot thickens. A human test operator inserts a simulation of a Chinese Shenyang J-35 fighter into the digitized operational picture shared among the flying drones and their controllers on the ground. It is coming to attack the base.

The Anduril operator orders drone No. 3 to break away from the defensive pattern and intercept the intruder. Automatically, the other three drones adjust their flightpaths, filling the gap created by their missing partner. As the ground-based human controller observes, the interceptor drone acquires the J-35 and—after receiving the human operator’s permission—fires a missile and kills the enemy.

Anduril recently staged this demonstration for journalists to show how far software-based capabilities have come. The scenario came from the company’s Hyperion experimental series, which adapts the internal Lattice autonomy software to perform air-to-air missions. As one of the two companies building operational prototypes for the U.S. Air Force’s Collaborative Combat Aircraft (CCA) program, Anduril’s internally funded autonomy software experiments offer a glimpse into the future of air combat.

Autonomy software is no longer limited to enabling a single aircraft to execute the commands of a remote human controller. Software code can be trusted instead to orchestrate a mission involving multiple platforms. Anduril’s Hyperion project focuses solely on coordinating aircraft, but other industry and military efforts are connecting platforms across multiple domains to carry out missions, with only policy dictating a human’s explicit approval for the release of lethal effects.

The premise that “mission autonomy” is ready for combat guides several major U.S. Defense Department programs. The Air Force aims to introduce the first CCA by the end of the decade. The Navy’s Project Overmatch is developing multidomain technology for sea and air platforms, mostly in secret. The Army’s Project Convergence is pursuing a similar result on and over land, along with the Air Force Special Operations Command’s Adaptive Airborne Enterprise. Meanwhile, the Pentagon’s Project Replicator seeks to support all of those initiatives with a package of cheap drones and mission autonomy software.

As far as mission autonomy software has progressed, however, gaps are still apparent. Anduril’s demonstration followed a tight script. When asked afterward about several contingencies—what if communications between the drones are disrupted? What if the J-35 shoots down the first interceptor?—Anduril’s answer was clear: Software development is needed to address every possible permutation.

The capabilities of Anduril’s choice of test drones also limited the realism of the demonstration. In April, the Air Force selected the company to develop the Fury as one of the first two operational prototypes for the CCA program. But the Fury is still months away from achieving first flight. Instead, the Anduril Texas Test Site operates Class 3-size—up to 1,320 lb.—-“clay pigeon” jets as surrogates for the future CCA fleet. These aircraft lack the payload and computing capacity to be aware of their own surroundings. That would require carrying cameras and radars as well as a sensor fusion processor to interpret the data and a mission systems processor to act on it.

Several programs are underway to close the gaps in the software and computing resources needed to enable mission autonomy in combat. These programs are running in parallel to the more visible hardware prototyping efforts, but in some ways they are more important.

In the last year, the Air Force and DARPA launched separate programs, with each assigning several contractors to focus on developing mission autonomy software for future CCA. Meanwhile, industry heavyweights and small startups are working independently on their own software applications. Lockheed Martin’s Skunk Works continues experiments of its Enhanced Collaborative High-Frequency Orientation System. Startup companies also are refining software for mission autonomy roles, including Anduril’s Lattice, Shield AI’s Hivemind and EpiSci’s Tactical AI.

While the CCA operational prototypes catch public attention, the autonomy software that will define their combat capabilities is proceeding quietly. The Air Force selected five companies to develop mission autonomy software for the CCA program in July but withheld their identities as classified information. The project is being led by the Air Force Experimental Operations Unit, which is tasked with translating autonomy experiments into combat capabilities.

The program builds on the Air Force Research Laboratory’s yearslong Skyborg program and DARPA’s Air Combat Evolution (ACE) program. The latter produced algorithms that have been tested aboard various aircraft, including the Air Force Test Pilot School’s X-62 Variable Inflight Stability Test Aircraft, General Atomics Aeronautical Systems Inc. MQ-20 Avenger, and Kratos XQ-58 Valkyrie and UTAP-22 Mako.

Slightly more visible are the contractors working on the follow-on to DARPA’s ACE program. In June, the agency confirmed that six companies—Northrop Grumman Mission Systems, Lockheed Martin Missiles and Fire Control, BAE Systems, EpiSys Science, Systems & Technology Research and Strategy Robot—are working on Phase 1 of the AI Reinforcements (AIR) program. Whereas the ACE program was limited to one-on-one, within-visual-range dogfights, AIR is developing the algorithms to support engagements involving multiple aircraft on both sides in beyond-visual-range scenarios.

To date, technology demonstrations have shown the limitations of state-of-the-art autonomous algorithms. The 2020 AlphaDogfight Trials showed that a simulated, autonomously piloted fighter could easily shoot down a human pilot at the controls of a simulated F-16. However, the simulator provided the autonomous pilot with perfect situational awareness of the location of the F-16 throughout each engagement. By the time the first CCA are supposed to be fielded—-before 2030—situational awareness must migrate to the sensors and processors aboard each aircraft.

Anduril officials say they have a straightforward path to achieving CCA autonomy. The Lattice system is already used on the ground for base defense systems, which package a network of sensors with sensor fusion algorithms to develop a real-time situational awareness picture. The task now is to repackage those sensors and fusion processors into hardware compatible with the size, weight and power limitations of a small aircraft.

“Sensor fusion type problems was one of the original core problems that we were trying to understand,” says Kevin Chlan, Anduril’s senior director for air dominance and strike. “We just have really matured those capabilities. But it’s also not to say that we solved sensor fusion, and now we will not have challenges. No, absolutely not. We’re gonna continue to learn, but we’re also not starting from zero.”