SOMEWHERE IN WEST TEXAS—Four drones take off one by one into clear skies, then join in formation, performing a brief airshow for a dozen visitors here to a lonely West Texas airfield followed by a combat air patrol demonstration. A simulated enemy approaches, the visitors are told, then one of the four drones veers off to make the intercept, “firing” a single simulated missile at the imaginary aggressor.
A single technician—using a combination of voice and keyboard commands—controls the entire operation.
Anduril is an unusual defense venture, a startup funded like a Silicon Valley tech firm with venture capital funding, and its aim is to disrupt the defense marketplace, just as SpaceX shook the ground in space launch and Tesla changed the automobile industry. With a valuation of some $14 billion, the company was a surprise finalist in the contest to build the Air Force’s first autonomous Collaborative Combat Aircraft, where it is locked in a competition that has already narrowed to Anduril vs. General Atomics Aeronautical.
Anduril flew defense reporters to its West Texas test facilities to demonstrate its software technology, using Class 3 “Clay Pigeon Jets,” which are smaller and less sophisticated than Anduril’s planned Class 5 “Fury” drones, and simulations to prove its prowess. Anduril paid travel costs for Air & Space Forces Magazine and other journalists to attend the demonstration.
The showcase provided a rare glimpse into how Anduril is doing in the contest to build USAF’s first CCAs. The first public demonstration of its autonomy software showed a single human providing relatively simple directions for multiple “robotic wingmen” to execute combat missions.
“We believe that driving towards delivering this capability is what’s going to be disruptive and is what’s going to be required to achieve victory in combat in the future,” said Kevin “Shaka” Chlan, a retired Navy fighter pilot who is now Anduril’s senior director for air dominance and strike.
Hyperion
Anduril has been working on the underlying software for autonomous aircraft to take on the air dominance mission for more than a year, company officials said. A testing campaign, codenamed Hyperion after the titan from Greek mythology, has included more than 200 live flights and 40,000 simulations, testing dozens of scenarios and mission profiles like the one shown to reporters.
Company officials declined to discuss government contracts during the demonstration, but Air Force officials have said the very first Collaborative Combat Aircraft will be focused on the air-to-air missions, and Chlan praised the Air Force for “outlining a very specific capability need, a capability gap with the CCA program, and going after it quickly.”
“In 2023, hearing what Air Force and other customers were talking about when it comes to manned/unmanned teaming with large Group 5 aircraft to go out and conduct air operations, we realized that we needed to invest some significant resources in maturing that,” Chlan said.
In addition to Anduril and General Atomics, which both have contracts to develop CCAs, the Air Force has also contracted five unidentified firms to work on autonomy software. Anduril declined say whether it was involved in that part of the CCA program.
Anduril officials say they have been working on their Lattice autonomy software for years. The software takes in sensor data, establishes a common operating picture, and presents users with decisions at key moments. The air dominance version, called Lattice for Mission Autonomy, controls decision-making, coordination, and action for multiple aircraft simultaneously.
At the heart of it all is a “quarterback,” which provides intent that the software interprets into discrete tasks.
“No one wants to be flying around in their F-22 or their F-35 or their F/A-18 telling, like, ‘OK, robot A, I need you to go do that.’ You want to be able to give teams intent,” said Chlan.
Using voice commands is a new feature. Anduril officials say voice input is better than manual inputs because of the information overload pilots face from a bevy of screens and the shaky atmosphere in a fighter cockpit.
The voice system aims to build confidence by helping pilots feel as if they’re talking to another human, not a robot. Like trained pilots, the software answers using “brevity codes” followed by the U.S. military and NATO, using text-to-voice and voice-to-text software, and referencing context from the mission plan.
“You can imagine that if there is a human pilot out there interfacing with crewed and uncrewed platforms, that if they can address and speak to the uncrewed platforms in the exact same manner that they speak to the manned platforms, that is the least cognitive burden we’re going to introduce in the interface,” said mission software engineer Andrew Burke.
During the demonstration, the engineer at the controls spoke only a few words to get the entire four-ship drone formation to spring into action:
“Mustang, fence in.”
“Authorization requested for approval,” the software “speaks” with a flat, emotionless tone, asking for permission to intercept the simulated approaching aircraft.
“Authorized, commit all,” the engineer said.
“Authorization requested for approval,” the software replied later, asking permission to engage an aggressor after it entered a predetermined “no-fly” zone.
“Mustang 11, engage!” the engineer commanded.
For voice commands to be effective requires preflight preparation. Using the Lattice interface, mission planners can set a boundary area for the drones to patrol, require them to ask for permission before engaging, and choose from prebuilt behaviors and formations to fly and execute. Once those are set, the brevity codes spoken by the technician start the software executing.
Anduril officials said this is similar to preflight mission briefings for human pilots—except that the software can execute over and over again without getting bored, tired, or emotional.
“These are the kinds of things that we expect out of a wingman,” Chlan said. “This idea that they’re going to unemotionally accomplish these important tasks—the dull, the dirty, the dangerous.”
Next Steps, Challenges
In the pristine conditions of the Sept. 10 demonstration, Anduril’s software worked exactly as planned, quickly defeating the simulated adversary. But it’s impossible to know how pre-programmed the whole effort was. Achieving real-world performance in challenging combat conditions against unpredictable, complex threats is still to come. How will the system deal with broken or jammed communications channels, degraded or disabled sensors, or confusing target sets?
“We’re playing around with figuring out which asset should go after which threat,” said engineering director Spencer Fishman. “And we go through and say, let’s play from one to 10 threats coming in from like 100 different approach paths. This is just one of the very early examples. And so using that to then evaluate which ones fail and which ones work.”
In the real world, challenges are perpetual. “Perfect GPS and perfect comms is not reality,” added Diem Salmon, vice president of air dominance and strike. “As we get more mature, those are the things we’re going to start stress-testing a lot more. And those are the things that you have to figure out.”
Right now, the surrogate drones being used for flight testing lack the sophisticated sensors that would go on a hypothetical CCA drone, which is why Anduril simulates them in tests. Offiials projected confidence that their software will be able to fuse sensor data effectively in the future.
“We’re constantly investing in the core algorithms where you can take detections of different modalities and be able to start testing with them, and then change either how we’re conditioning that data from a new sensor or the core algorithms,” said Fishman.
The question of when the company will move on to testing the software on the large Group 5 drones like Fury temains open-ended. When asked, Chlan answered that the current CPJ drones are “a pretty good surrogate … which has allowed us to be really focused on the software maturation aspect.”
Simulations make that possible and accelerate learning. Anduril has run tens of thousands of them already, and expects to ramp up the pace of testing to discover the limits of the system.
“The thing that we’re already doing now, and we just need to continue doing and expanding is large scale simulation,” said Fishman. “Like, OK, let’s tweak those models. Don’t just trust the model. Maybe let’s change it. What would happen if some standard deviation started to increase or decrease? What are the effects of not just doing one asset coming in, but doing many threats coming in? Are we even thinking of the problem correctly at all? Are our models and the spaces that we’re changing, is that even correct? That gets tricky.”
Chlan said the company believes live flight testing is important too, and its remote test facility is not shared with any partners. That enables the company to conduct frequent tests, adjusting and testing again in hours or days, not weeks.
How Soon?
With more and more data from simulations and testing, intense interest in autonomous military technology, and billions of dollars in funding, Anduril faces pressure to turn Lattice for Mission Autonomy into something in the Pentagon inventory, rather than one in a line of experiments.
That could be discovered “over the course of the next couple years,” Chlan said. The initial aim is a minimally viable product—a version of the software that accomplishes its tasks well enough to be fielded and help warfighters, but with room for future updates and improvements.
“Looking at where the technology is today, and if we really invested resources and focus on a specific use case and marshaled resources, hardware and software today, it’s a couple years out,” said Salmon. “It doesn’t need to be a 10-year development effort.”
Salmon said her own goal is to have the software ready before the end of the decade—a timeline roughly in line with the Air Force’s own stated ambition to buy around 100 CCA drones by 2029.
The Air Force has said it plans to make a production decision on CCAs by fiscal 2026. General Atomics, which has a long history now of developing and flying unmanned aircraft, is working on autonomy software of its own, as are others.
But Anduril leaders are confident they will win a long-term place in this arena.
“We’ve built our software to be open and hardware agnostic, because the goal is to get it on platforms in the future,” Salmon said. “And maybe it’s an Anduril platform, maybe it’s not an Anduril platform.”
The company will show off a life-size model of its CCA offering at AFA’s Air, Space & Cyber Conference, starting Sept. 16, in a booth managed by the Air Force Research Laboratory. A second AFRL booth will show off a model of an the General Atomics offering. But GA is also bringing a real XQ-67, on which its CCA offering is based.
The race is on.