Future Force: Impact of Autonomous Systems on the Defense Sector

In two conflict-ridden decades, unmanned combat and reconnaissance systems have gone from being exotic capabilities fielded by a handful of armed forces to ubiquitous staples of military conflicts across the globe by state and non-state actors alike.

Kratos XQ-58 Valkyrie. Image: Kratos.

We lie on the cusp of a new revolution in unmanned warfare as advances in sensors and artificial intelligence are poised to allow a growing range of uncrewed systems to perform their missions with much more limited direction from their human operators. The fruits of this welling sea change will affect small drones costing just hundreds or thousands of dollars, and aircraft, vessels and armored vehicles costing millions.

Already, traditional uncrewed systems reliant on remote control typically exhibit several game-changing characteristics: reducing risk to lives of human operators, lower procurement and operating costs, greater potential endurance, and capability to perform missions that may be impractical or impossible with traditional means.

Autonomy overlays the above with additional distinct qualities: the potential to carry out missions without access to satellite and communication links; the capacity to exceed human reaction speeds and mental tasking limitations, allowing a single human to control many drones; and for drones to cooperate with other drones and potentially even act in concert as a swarm. 

Of course, many remote-control and even manned platforms already incorporate autonomous functions such as automatic takeoff and landing systems, ground-collision avoidance, waypoint navigation, and emergency landing or return-to-base algorithms for when communication links are lost. But new AI agents enable much more complex, whole-mission tasks while requiring much less human input.

While advanced autonomy is not easy to develop, as a digital product, once perfected it may be highly reproducible and fielded in very small, cheap platforms as well as exquisite ones. Admittedly, certain autonomy enablers—advanced sensors for navigation and target identification and comms to support cooperative behavior—do have a physical footprint, though hardly a prohibitive one. And even when general-purpose AI agents are developed, adapting and testing them to work with specific platforms will require non-trivial efforts.

Shield AI Nova. Image: Shield AI.

Major factors impeding current military use of remotely controlled platforms are distance (for drones lacking satellite communication links), and signal interference from terrain, weather and enemy jamming and spoofing on the battlefield. Maintaining command links also may run contrary to the “silent running” needed to stealthily penetrate hostile airspace.

By doing away with the necessity of a command link, autonomous drones without satcoms may be able to fly further than the 110 or so miles practical with radio-frequency systems. A kamikaze drone could lock-on and guide itself to targets even when its plunging attack takes it too low for radio control. Above all, autonomous drones can prosecute their mission regardless of the defensive jammers proliferating in trenches and on armored vehicles in Ukraine. 

UAVs also often rely heavily on satellite navigation systems, access to which is vulnerable to jamming and spoofing, as well as environmental effects. But an autonomous drone with adequate sensors and AI can, like a human pilot, estimate its position using a combination of inertial navigation systems and optical sensors able to identify landmarks through image-matching and terrain-mapping technology.

The up-shot, then, is that one of the most effective counter-UAS methods today, jamming, may decline in effectiveness against purpose-built military drones—particularly as these systems become more and more capable of completing missions without command and navigation links.

Kratos Firejet. Image: Kratos.

Weathering the Perfect Storm

“True swarming is non-deterministic,” Kratos’s Steve Fendley told Inside Unmanned Systems. “The machines decide what they get to do. No human can accurately determine what specifically they’re going to do.” For that reason, he said Kratos has had little interest in pursuing swarming despite its earlier research into nature-informed swarming in a program called WolfPak, which mostly influenced Kratos’s autonomous trucks, not aircraft.

“We’re very cautious. What we see from users in the fighter, tanker community is we need to 100% trust in an uncrewed system to do what we expect it to do. We’ve basically drawn a box around deterministic approaches to AI.” Fendley added there are ways to try to make it work. “You can try to put guardrails on non-deterministic behavior. Say, for example, a manned asset flies within 200 meters to the right—the guardrail constrains the action.”

Manned-Unmanned Teaming 

Before militaries wholesale ditch manned aircraft and armored vehicles for autonomous uncrewed systems, they wisely seek assurance those robotic substitutes work. After all, plenty of early attempts didn’t. 

Two approaches for bridging that transition are optional manning (manned vehicles that can optionally be operated uncrewed) and manned-unmanned teaming (MUM-T), or uncrewed systems that act to support a manned platforms much like a loyal wingman. By the 2020s, MUM-T has arguably surpassed the optional manning paradigm, which sacrifices the important cost and engineering efficiencies of uncrewed-only systems.

Autonomy is vital to MUM-T because pilots of manned aircraft, especially single-seat fighters like the F-35, cannot pay attention to micro-managing drones while in the grip of aerial combat. That’s even more the case when controlling multiple drones. Thus, MUM-T platforms must be able to correctly interpret fairly simple commands to perform complex missions—and prosecute assigned tasks with minimal or no oversight. 

Shield AI V-BAT. Image: Shield AI.

Different Kinds of AI

Computer actors harnessing AI to make decisions are known as agents. A taxonomy developed by scientists Stuart Russel and Peter Norvig in 2003 is commonly used to rate more sophisticated AIs.

Simple reflex agents simply gather data and carry out actions when that data fits certain criteria driven by simple “if-then” logic.

Model-based reflex agents retain information from its internal condition, allowing it to update its internal model and fill in missing data gaps.

Goal-based models also retain externally collected information to update its assumption, but is designed to achieve specific goals or outcomes, not merely evaluate whether to execute actions according to if-then logic.

Utility-based agents measure success in more nuanced fashion by weighing multiple success variables and selecting the course of action calculated to maximize those variables.

Learning agents have a dedicated learning element that actively evaluates its own performance of objectives and uses that knowledge to update how it acts in pursuance of predefined goals. Such an AI can learn how to act in entirely novel situations based on the feedback of an internal critic, which punishes failed actions and rewards successful ones.

The brute force “fail and fail until you succeed” nature of iterative learning techniques may seem crude, but has succeeded in gradually refining agents that outperform humans in complex contests: for example, defeating an F-16 pilot in a guns-only dogfight, or beating champion human players in games of Star Craft.

Many AI platforms are single-agent, focusing on an individual protagonist. A multi-agent supports cooperation between several peer AI agents.

A hierarchical AI instead features leaders that process data from subordinates and instruct those lower-level agents what to do.

The Magic of MOSA

Military computer, communications and sensor systems built by competing corporations historically faced a Tower of Babel problem in service: unless they’re purpose-built to speak the same language, they simply end up talking past each other. Ruthless imposition of common standards by clients, or expensive cross-integration requiring installation of new physical parts, seemed the only way to bring order to chaos.

Today, the language problem is being overcome not by convergence on common systems—they’re more diverse than ever—but rather a Modular Open Systems Approach (MOSA) design philosophy that fosters interoperability. According to the Defense Standardization Program, doing things the MOSA way involves: 

• Compartmentalizing the fixed primary systems of a platform from its more adaptable modular components.

• Making sure the platforms’ interface and subsystem syntax is software-defined (and thus software redefinable) in a format readable by common standards.

• Supplying the documentation with functional descriptions needed to understand the syntax.

• Ensuring open systems architectures that expressly supports incremental upgrades or even swapping out one system for another.

• Accepting product transparency and non-prohibitive legal regime on data rights.

Standardization thus happens at the level of digital systems that translate unique outputs of sensors and AI agents into a mutually intelligible lingua franca.

MOSA rewards willingness to fluidly interoperate with other components—rather than defaulting to a company’s walled-off ecosystem. The results, ideally, are products developed more quickly, less expensively, and with greater adaptability and longevity thanks to upgradability.

Swarming Warfare

In Guardians of the Galaxy 2, there’s a scene in which the protagonists are chased by hundreds of drone fighters, each remote-controlled by an individual pilot sitting at a console in a faraway galactic capital. Arguably the most outlandish aspect of the scene wasn’t the talking racoon, but rather that each of these numerous, seemingly expendable fighters was individually piloted by these technologically advanced aliens.

These days, “drone swarm” gets casually invoked to describe nearly any formation of drones acting in proximity. But genuine swarming capability involves drones actively cooperating with one another in a hivemind-like fashion to complete their mission. This can be for wide-area ISR purposes, piranha-like swarming kamikaze attacks that oversaturate and disorient defenses, or both at once. While such tech has been tested, operational use has been limited so far.

Swarming behavior is communications reliant, which may be extended via a mesh network using relaying or software-defined radios. But that leaves swarm autonomy vulnerable to electronic warfare again and faces complex challenges in managing bandwidth between numerous transmitters in a confined area.

However, swarming algorithms also can be trained to operate like flocks of birds in nature, which maintain formations and act as a group without radio communication through mutual observation, signaling behaviors and instinctual algorithms.

Surface Sensing

Anduril Dive. Image: Anduril.

Breaking the command-link tether also has major implications for ground, sea and underwater drones (UGVs, USVs and UUVs respectively) as surface interference restricts maintenance of reliable quality command links to short distances. However, the greater complexity of ground environments has slowed adoption of autonomous military UGVs for combat. Despite its unique challenges, the maritime domain falls in the middle ground of complexity, and major strides in USV and UUV autonomy are in progress. There are huge potential economies in transitioning from manned to unmanned platforms, and from ship-tethered remote-control UUVs to free-ranging autonomous ones.

Collaborative Combat Aircraft

The U.S. Air Force’s attempt to bring autonomy to the edge of aerial warfare is embodied in the Collaborative Combat Aircraft (CCA), intended to accompany manned supersonic stealth fighters as “loyal wingmen.” The Air Force tapped Anduril, Boeing, General Atomics, Lockheed Martin and Northrop Grumman for proposals, and this Spring will down select two or three to build prototypes for evaluation. 

CCA is a follow-on to the preceding Skyborg program and will use the Autonomous Core System developed in that program. Tech feeders include the Defense Advanced Research Project Agency (DARPA) Air Combat Evolution and the Air Force Research Laboratory (AFRL), as well as Autonomous Air Combat Operations programs and Australia’s Airpower Teaming System (ATS).

The CCA will have the speed and range to fly alongside or ahead of their manned fighter while carrying a mission-relevant modular payload (sensors, weapons, electronic attack) with an initial focus on offensive counter-air (OCA). Not all CCAs may be armed, but enemy assets won’t be able to distinguish them.

Though receiving commands from pilots and ground controllers, they’ll be able to complete missions without handholding in GNSS and signal-denied conditions, and in split-second combat scenarios will react with robotic speed unimpeded by the friction imposed by human reflexes and remote-control signal delays.

The Air Force’s plan is to spend $6 billion developing CCA through 2028, then begin production of at least 1,000 Increment 1 drones, though a greater total is likely. Operationally, 300 of the service’s F-35A stealth fighters will each be assigned two CCAs. Another two each will be dedicated to 200 forthcoming sixth generation Next Generation Air Dominance (NGAD) fighters, though theoretically up to five CCAs may be controllable from one NGAD jet. It’s also suggested larger aircraft—say B-21 stealth bomber, KC-46 tankers and E-7 AWACS jets—could control CCAs serving as escorts. Meanwhile, by 2025 a still mysterious CCA Increment 2 will begin development. It will be open to foreign participation.

Not So Cheap, Not So Expendable

Air Force officials have made increasingly clear they want something higher on the capability and cost spectrum than attritable subsonic jets costing single-digit millions of dollars. Air Force Secretary Frank Kendall characterized CCA as costing one-quarter to one-third the price of an F-35 stealth jet, implying a figure ranging from $20-35 million each. The Air Force’s requirement for engines generating between 3,000-8,000 pounds of thrusts—more than most pre-existing candidates—implies an interest in supersonic flight capability, perhaps to keep up with super-cruising NGAD fighters transiting to the battlespace. 

Such an aircraft is too pricey to dismiss as attritable, according even to standards set by Congress, which notionally suggested a price cap of $10 million on attritable UAVs and $25 million on exquisite UAVs, even if its loss is preferable to that of a manned aircraft. Overall, it seems the service has less faith in the highly attritable loyal wingman and wants something closer to an unmanned F-35 than an unmanned A-4 Skyhawk. This more conservative approach unsurprisingly has critics in the aerospace industry.

Shield AI V-BATs. Image: Shield AI.

Shield AI

In nine years, Shield AI has risen to the forefront of autonomous military systems, integrating its Hivemind AI solutions with partners including Kratos and Boeing, as well as building its own autonomous platforms. Furthermore, after Heron System’s AI triumphed over a human pilot in the famous Alpha Dogfight tests in 2020, Shield AI bought out Heron to incorporate its AI, and Martin UAV and its V-BAT VTOL fixed-wing drone.

Hivemind

“Hivemind is a standardized package that enables future autonomous platforms,” company co-founder Brandon Tseng said. “As a software company, we’ve set out, from day one, to deliver a platform-agnostic product thanks to our open and modular architecture designed to integrate third-party systems. Recently, we integrated Hivemind with the Kratos MQM-178 Firejet, and went from the first meeting to flight ready in 165 days.”

“AI architecture starts with foundational behaviors, essential for both administrative and tactical operations. These are then enhanced to support advanced, collaborative tactics among multi-agent systems utilizing expert systems and reinforcement learning. This enables precise, autonomous coordination of multiple sensor-equipped or shooter-equipped uncrewed aircraft, marking a significant advancement in collaborative tactical behaviors.”

Tseng emphasized the challenges of developing effective autonomy. “Autonomy is freaking hard. It’s not just like ‘put autonomy on it,’ that’s like saying ‘just land a rocket.’ It took SpaceX a decade and massive investments to do that. Shield AI has raised more capital and can work on it longer than any other company in the world in the defense sector.”

Confronting One’s Own Mortality

Tseng formerly served as a Navy SEAL and surface warfare officer—experiences that informed his vision for the company. He recalls a pre-deployment training mission drove home to him that certain missions simply couldn’t be performed without unavoidably serious risks to life and limb.

“I was going through pre-deployment Close-Quarters Battle training, force-on-force movements through the house, with weapons loaded with paint rounds. I’m executing the Tactics, Training and Procedures and I remember getting shot in the face. And as I’m moving through a doorway, the instructor puts me down and says, ‘You’re dead’. I remember thinking to myself, this is BS. I did everything I’m supposed to do, and I’m still dead or a liability for my teammates.”

Nova, Shield AI’s first major autonomous platform, now in its second generation, has come to offer a combat-tested tool aimed at reducing the odds of unpreventable death clearing built-up combat zones.

Ethics and Autonomous Killer Robots

General Atomics Avenger. Image: General Atomics.

While there’s a cottage industry churning out academic articles on the potential that humanity will be destroyed by self-aware AI, a more tangible and proximate quandary is the proliferation of autonomous systems perhaps less sophisticated than Terminator T800 but nonetheless empowered to execute lethal attacks.

Yes, human beings will set the parameters for what targets these killer robots attack—all of the kinetic systems described in this article require human authorization for kinetic attacks. But when operating on the edge beyond assured reach of comms, or in numbers too great to practically control, AI agents will classify possible targets and assess if they’re authorized to kill. Russia and Ukraine have both already begun fielding kamikaze drones mating automatic target recognition AI with terminal electro-optical guidance, albeit with mixed results.

In truth, a narrow form of such lethal autonomy already existed on some Cold War-era missiles and torpedoes with target classification capability. That these didn’t cause protests decades ago suggests not all forms of lethal autonomy are equally controversial. After all, there are unlikely to be civilian warships, tanks or jet fighters in a warzone.

However, risks of error multiply when targeting dismounted human beings or lighter, plausibly civilian vehicles. Entrusting automation based on technologies such as facial recognition known to be unreliable and affected by biased datasets has particularly large risks. After all, humans often fail to accurately distinguish civilians from adversaries, and robots may have an even harder time. To be fair, it’s also possible AIs—not being susceptible to combat stress and disobedience—might eventually achieve lower rates of misidentification than humans. But even then, the inevitable accidents will pose complex moral issues given diffused and unclear responsibility for actions performed by autonomous systems.

Bear in mind self-imposed ethical restrictions on uses of killer AI aren’t bound to be observed by foreign actors, absent an arms control treaty. While seeking to restrict proliferation of ethically problematic technology, defense planners must also prepare for the use of autonomous systems in morally objectionable ways by adversaries and third parties, much as some cyberwarfare tools have been repurposed by U.S. allies for repressive purposes.

Lastly, just as long-endurance drones enabled the U.S. to embark on a sprawling campaign of surveillance and targeted assassination in the 2000s and 2010s, autonomous systems will de-risk and enable operational concepts that were impractical or risky before. That opens the door to diverse ways to improve force protection, lethality and cost-efficiency. But not every concept made possible by new technology is actually a good idea.

Nova 2 Quadcopter

Nova 2’s edge in the world of tactical quadcopters is its ability to fully autonomously navigate inside multi-level buildings, or underground tunnels, all without need for GNSS access, waypoints or operator inputs. Thanks to its computer vision and pathfinding algorithms, it can detect occupants and report their position in real-time when communications are open. 

Tseng said Nova had been deployed by U.S. forces on “most important national-level missions,” referring to Tier 1 SOF units, and that Israeli units successfully employed it to rescue civilian hostages inside buildings at the onset of the Israel-Hamas war on October 7-8. “We have stories from the users, text messages from them thanking us,” Tseng said.

MQ-35A V-BAT

The V-BAT is a medium-sized (Class 3) fixed-wing ISR drone with VTOL capability thanks to its thrust-vectoring TOA-288 engine, with ducted fans allowing operators to remain close even when its propeller is turning. Tseng argues such mobility is key in the Pacific, where the U.S. military relies on aircraft carriers and island bases: “We have to learn how to be mobile and get off those forward operating bases; turn every ship into an aircraft carrier, instead of having it all centralized.”

In 2023, Shield AI unveiled a new V-BAT Teams capability allowing cooperative swarming operations, initially in an unarmed maritime domain awareness role, though strike, suppression of enemy air defenses, escort and logistical missions are envisioned. Currently four drones can act cooperatively, increasing to 16 drones by 2026.

“It’s the only Group 3 UAS with a military type designation,” Brendon said. “We’ve done 17 deployments with the Navy and Marine Corps, helped interdict millions of dollars of drugs in Southern Command, identified targets in EUCOM, CENTCOM, PACOM.” 

Recently integrated payloads include Synthetic Aperture Radar and kinetic munitions, and notably, Northrop Grumman’s 6-pound Hatchet mini glide bomb. But Tseng primarily sees the V-BAT as a targeting asset for more powerful shooters. “If a V-BAT is looking at you, maybe a guided missile submarine or F-35 is too.”

“I think we should never cede a machine the moral decision to use lethal force. That’s the NATO policy–and the moral one.”

Brandon Tseng, CEO, Shield AI

The Economies of Drone Warfare

V-BATs allegedly cost in the mid-six figure range. Tseng argues it hits a sweet spot balancing affordability with effectiveness and resilience. 

“It’s the cheapest aircraft you can build that warrants a SAM shot, compared to our fighters or Collaborative Combat Aircraft. We’d much rather have a $1 million drone shot out of the sky by a $1 million missile.” Drones like V-BAT, he argues, allow missions to be performed at “much cheaper and distributed costs and risks” and “create many more dilemmas for adversaries because V-BAT is tied to the kill chain [posing a real threat], and flies high enough that you can’t shoot it down [cheaply] with anti-aircraft guns.”

Kratos Valkyrie. Image: Kratos.

Kratos and Valkyrie

San Diego-based Kratos was the first company to make the jump from concept to material reality, coming well ahead of the world’s aerospace giants in developing the world’s first practical Loyal Wingman drone fighter, the XQ-58A Valkyrie, which first flew in March 2019. Valkyrie’s attractive price and five-year-maturity has garnered it orders from the Air Force, Navy and Marine Corps for use in extensive tests.

But so far, the leap for large-scale procurement for an operational role has been elusive, and many aviation observers were shocked when the Air Force didn’t include Kratos in the lineup of CCA competitors.

Steve Fendley, president of Kratos’s Unmanned Systems Division, said, “First, we weren’t identified as a prime on that system.” He also pointed out CCA had drifted toward becoming more expensive, which didn’t fit the Valkyrie’s attritable concept. 

Kratos UTAP-22 Mako. Image: Kratos.

“Attritable has gone in and out of fashion. [But] we are pretty focused on keeping a missionized Valkyrie at the sub $10 million mark. If you look at the wargames, and even Mitchell Institute studies, that sweet spot keeps coming back as the answer for what turns the tide [by providing affordable mass.]”

Kratos still has an eye on the follow-up CCA Increment 2, but Fendley said he had little firm knowledge of the Air Force’s intentions for that phase. “We’ll look hard at it from a Prime position. But if our ‘Prime-wind’ isn’t sufficiently good, we’ll see if we can provide value some other way. We’re hopeful that the spec is something that’s a good fit for our capabilities for a particular aircraft.”

Perhaps Valkyrie’s first break may come via the Marine Corps, for which Kratos announced in April it’s developing an operational MQ-58B multi-mission variant capable of Suppression of Enemy Air Defense (SEAD) for that service’s PAACK-P program. In a February 2024 Marine Corps trial over Florida, a Valkyrie teamed with two F-35s detected, classified and geolocated hostile emitters, relayed their coordinates to friendly assets and then jammed them successfully.

Five Kratos Valkyrie Variants are in Production or Development. Here’s How They’ll Differ.

Kratos 5GAT. Image: Kratos.

“The Valkyrie everyone sees daily is the rocket-assisted takeoff (RATO) model able to operate from remote locations without a runway and be recovered by parachute, which can solve some potential [basing] problems in the Pacific,” said Steve Fendley. But there are operational areas [like Europe] where maybe runways are available and won’t be tied up with manned assets, so you don’t need RATO/parachute system.”

Forthcoming Valkyrie subvariants will support more and less stringent takeoff and landing conditions tailored to different deployment contexts. However, mission payloads will remain interchangeable between them.

The economics of cheap robot fighters: Currently, each Valkyrie costs $5.5 million, implying 15-18 could be purchased for the price of a single F-35A. But Kratos claims that price could drop to just $2 million if it receives sufficient orders to scale up the annual production rate of 100 aircraft per year. That’s in 2016 dollars, Fendley warned, so adjust accordingly.

However, these prices are for a “naked” Valkyrie devoid of mission systems. Those bump the price back to $10 million. “We’re trying to be very selective and find high capability for cost payloads,” Fendley explained. Valkyrie is nonetheless capable of beyond-visual range air-to-air combat (usually requiring heavy radar and missiles) and moreover has demonstrated it can execute maneuvers and tactics exceeding human G-Force tolerance, Fendley said.

The Loop Within the Autonomy Loop: Kratos’s Approach to MOSA and Autonomy

Anduril and Shield AI both emphasized they are creators of software and AI agents first, and specific platforms second. But Kratos’s Fendley didn’t hesitate to stake out the opposite stance.

“We are not pursuing AI and machine-learning as a company. We focus on basic autonomy, the easy-to-prove deterministic aspect, such as under what conditions to turn on a certain sensor system. We made a strategic decision that we would develop internally the flight control laws tied to the airplane’s unique configuration, and constraints keeping the airplane from crashing itself,” he explained. 

“We developed an interface to connect this ‘inner loop’ of autonomy with an ‘outer loop’ of control so we can rapidly evolve it. Basically, the inner loop flies the airplane and introduces guardrails to keep it from crashing itself. The outer loop tells it where to go [and what to do.] Those can be manual or autonomous.

“So we do our own Kratos autonomy, which has controlled most of the flights until recently, and now we’ve begun incorporating autonomy from the government and Shield AI. These different autonomous AIs each have their own feature sets, missions they’re most optimal for. We wanted to ensure our system is flexible. You could have one AI autonomy solution, a different one on a different platform; that’s doable because the structure and definition tends to match the standards.” 

Without getting into swarming behaviors, interoperability remains essential for the basic functioning of UAVs. “You don’t have to determine the answers the same way, but you have to share information the same way, particularly at the navigation level so as to remain within a safe ‘tube’ in which they won’t collide with each other.”

Kratos UTAP-22 Mako. Image: US Air Force.

Kratos’s Target Drones

Prior to Valkyrie, Kratos’s aerospace bona fides rested on the jet-powered BQM-167 and BQM-177 target drones used by the Air Force and Navy respectively to simulate enemy missiles and aircraft. These were originally products of Sacramento-based Composite Engineering Inc., which Kratos acquired in 2012. 

Advances in autonomy and engineering not only enable these drones to do their primary job better and more cheaply but open a pathway to evolving them into operational platforms capable of performing ISR and attack missions.

Indeed, Fendley said the company’s Firejet and 5GAT drones may potentially evolve into low-end and high-end solutions relative to the “medium-weight” Valkyrie. “We acquired Sierra Tactical Systems last October [2023]. Their 5GAT is 40% larger than Valkyrie. Then you have the tactical Firejet at very low cost—sub $500,000, can pull 12 Gs, and can carry and deploy Switchblade loitering munitions. If you’re a Switchblade operator, you’d sure rather be launching those 400 miles away from the enemy rather than 16.”

Kratos UAS

XQ-56A Valkyrie – Cruising Speed: 552 mph
Max Speed: 652 mph
Service Ceiling: 45,000 feet
Max Range: 3,450 miles
Weight: 1.25 tons empty
Payload Capacity: 600 lbs. internal + 600 lbs. external mid-wing

Takeoff: RATO-assisted
Landing: Parachute recovery
Propulsion: Williams FJ33 2,000-lb thrust turbofan

UTAP-22 MAKO BQM-167A
Target drone transformed into UCAV 
The Mako is an adaptation of the BQM-167 target drone to also support operational use in a fast ISR role at near supersonic speeds. The Marine Corps has successfully tested teaming Makos with Harrier jump jets, including autonomous formation flying. 

Max Speed: 698 mph
Cruising Speed: 260 mph
Service Ceiling: 50,000 ft.
G-Force limits: +9/-2 Gs
Weight: 650 lbs. empty MTOW: 2,050 lbs.
Max Range: 1,600 miles
Endurance: 3 hours
Max Payload: 351 lbs. internal + 500-800 lbs. underwing, + 99 lbs. on each wingtip. Options include jamming pods, chaff/flare dispensers. Comms: UHF + tactical datalinks.
Propulsion: MicroTurbo Tri 60-5 turbojet with 990-lbs. thrust
Takeoff: RATO-assisted
Landing System: parachute recovery
Unit Cost: $2 to $3 million
Straight Talk from Fendley: “Compared to Valkyrie, it’s substantially less expensive, and less capable. It has no internal weapons capability, though from the speed/survivability perspective it’s similar. If you wanted to deploy something for a distributed sensor operation, UTAP makes sense if you wanted to deploy large numbers—say thousands.”

MQM-178 Firejet – Manufactured of lightweight carbon fiber at a facility in Oklahoma City, the Firejet is dramatically smaller and lighter than its stablemates. It’s also cheaper to operate because its pneumatic launch system removes the need for rocket-assisted takeoff, and even enables shipboard launch. Up to eight Firejets can be controlled by one operator.

The diminutive target drone’s excellent performance for cost led Kratos to develop the Tactical Firejet/Air Wolf prototype in 2019—a “mini-Loyal Wingman” capable of carrying payloads for ISR, target acquisition and light precision ground attack missions. Several have already been delivered to customers.

Maximum Speed: 529 mph
Service Ceiling: 35,000 feet/Minimum 20 feet
G Force Limits: +9/-2 Gs
Weight: 130 lbs. empty
MTOW: 320 lbs.
Payload: 71 lbs. internally, 35 lbs. on each wing station and 20 lbs. on each wingtip pod MQM-178 Payloads: towed targets, proximity sensors, radar and infrared augmentation systems, fixed flares
Known Air Wolf Payloads: BAE Systems “Tactical Mission System,” Switchblade loitering munitions (The weight tolerance of wing stations may support the longer-range 33-lb. Swithcblade-600 model)
Propulsion: 2x JetCat C81 turbojets generating 81.5 lbs. of thrust each.

Anduril Ghost. Image: Anduril.

Anduril and Lattice AI

Competing with Shield AI is the other big new player in autonomy, Anduril, co-founded by tech billionaire Palmer Luckey. Luckey has attracted controversy for his embrace of military contracts—shunned by many in Silicon Valley—and prominent support for Donald Trump. Luckey bills himself a Silicon Valley-style disruptor focused on agile, software-first development of AI for diverse autonomous systems.

Anduril’s “killer app” is Lattice, a software that a spokesperson wrote “…has been integrated with an order of magnitude more [of] non-Anduril products than Anduril products, including both modern uncrewed and autonomous systems and legacy crewed systems, across all domains.” Lattice’s open API and development kit facilitate rapid integration of new systems, contributing to a networked common operating picture.

Compared to other autonomy agents Anduril claims, “Lattice enables teams of unmanned heterogeneous systems to work together to autonomously and dynamically execute a given mission. This is fundamentally different from ‘self-driving’ or single-platform-level autonomy, which is typically focused on getting from point A to point B. Mission autonomy is focused on automating an effective and reliable team of assets to achieve a desired end state.”

The Kratos Valkyrie demonstrates the separation of the Anduril ALTIUS-600. Image: U.S. Air Force.

Altius Vehicle-Launched Swarming Drone

Resembling a flying baseball bat with pop-out wings and pusher propeller, Altius was devised by Area-1, purchased by Anduril in 2021, and is in production at an Atlanta factory. Carrying modular payloads in its nose, Altius can launch from Common Launch Tubes (CLTs) or PILS pneumatic launchers on air and surface vehicles. Non-attack models are recovered via belly landing. 

In a 2022 Army trial, a single operator controlled 28 Altius-600s divided between four cooperative swarms via Lattice. The software also enables ISR and attack types to team-up: at the Edge-23 exercise, an Altius-600 located and identified a hostile SAM site, authorized an Altius-600M to attack it, and then conducted a Bomb Damage Assessment. 

Altius have been used operationally, including in Ukraine where Luckey claims its open-architecture software was rapidly updated to defeat evolving Russian countermeasures. Currently, it’s being pitched to the Army’s Air-Launched Effects and Launched Effect-Medium Range programs.

Anduril Fury. Image: Anduril.

Fury

When Anduril purchased Raleigh-based Blue Force Technologies in September 2023, it came with its Fury multi-mission stealth UAV. Made of composite materials, Fury was intended for use as a ‘Red Team’ stealth fighter aggressor in adversarial training, and is claimed to have substantial fuel economy, long range and reduced noise signature. Autododyne has developed a complimentary GCS capable of controlling 4-8 Furies simultaneously for oversight of adversarial training. 

While Fury may court the Air Force’s ADAIR-UX unmanned adversary program, Anduril is evolving it into its likely CCA candidate with modular open-architecture systems able to integrate COTS sensors and government-furnished AI cores. Lattice will allow Fury to operate in autonomous, collaborative swarming and manned-unmanned teaming scenarios. Anduril is reportedly investigating ways to attain Mach 1+ surge speed, followed by super-cruising using a next-gen engine. Anduril claims use of digital prototyping, adaptable open systems, and tests on the XF-62 test aircraft ensure agile design, development and scalable production of Fury.

Anduril Roadrunner. Image: Anduril.

Anduril’s Drone Interceptors: Anvil and Roadrunner-M

The 11.6-pound Anvil is aimed at “low-collateral” defeat of Class 1 and 2 drones by ramming from below at an angle protecting vital systems from damage. (According to Anduril, Anvils are potentially reusable, but couldn’t share information on reusability rates.) There’s also an explosive-discharging Anvil-M to counter faster Group 2 drones. Anvils are cued by Lattice to autonomously intercept possible threats at up to 200 mph using radar and computer vision; the operator then judges from video imagery whether to attack. 

The Roadrunner-M VTOL interceptor (unveiled December 2023) is half drone, half surface-to-air missile. Nested in automated mini-hangars, the “high subsonic” Roadrunners launch to intercept larger, higher-flying UAVs and even manned aircraft. But if interception proves unnecessary, it can land vertically for recovery. This allows launch of multiple Roadrunners to ensure a kill without “wasting” redundant shots (costing in the “low 100,000s of dollars” per Luckey). SOCOM has requested $19 million to develop Roadrunner.

Anduril UAS

Altius – Mission Payloads: C-UAS jammers, Comms Relays, SIGINT, EW and cyberwarfare systems, radar decoys. Attack models: high-explosive, thermobaric aerosol, shaped charge anti-tank and wall- penetrating warheads, plus various seeker types.

Tested launch platforms: AC-130J gunships, JLTV and MRZR trucks, P-3 and P-8 maritime patrol planes, UH-60 helicopters, drone vessels (USVs), and MQ-1C and XQ-58 Valkyrie drones.

Unit costs: The article “SOF Swarm” (Air University) expresses concern that Altius-600 unit costs (allegedly $150,000 for ISR payload plus $100,000 for the drone) may be too high for expendable use.

Fury (projected characteristics) Max Speed: Mach .95
Service Ceiling: 50,000 feet
MTOW: 5 tons

Max G-Force: +9 Gs/-3 Gs
Endurance: 5 hours, adequate for two 40-minute training dogfights per sortie, or three closer to base Sensors: radar and/or IRST, electronic attack (jamming) system
Max Payload: 400 lb.
Propulsion: William FJ44-4M turbofan
(4,000 lb. thrust)
Target Cost: $2-20 million (“leaning toward the lower end”). Low thousands of dollars per flight hour.

Ghost: Quiet Tactical UAV – The rail-like Ghost UAV features acoustically stealthy and climate-resilient, it can be assembled in 2 minutes and has bays supporting multiple simultaneous mission payloads including ISR, targeting and force protection. Its sensors can automatically detect and classify objects. Integration with Lattice allows simple point-and-click mission tasking and control of large Ghost teams by one operator. Computer-vision navigation and frequency-hopping command link give it resilience in signal/navigation denied environments. It’s been used by the U.S. military and Royal Navy, and has seen combat use in Ukraine.

Supported Payloads: Gimbaled electro-optical/infrared sensors, lasers, electronic warfare systems, spotlights, comms relays, loudspeakers.

General Atomics Aeronautical Systems Inc.

For the first two decades of the 21st century, General Atomics’ large MQ-1 Predator and MQ-9 Reaper combat drones (UCAVs) embodied what drone warfare looked like in the public imagination. But while the MQ-9 continues to find new roles and users, the U.S. military is now more interested in platforms that strike a balance of being more expendable and more survivable in a high intensity conflict facing modern air defenses.

GA-ASI chief communications strategist Mark Binkley wrote to Inside Unmanned Systems that the company’s proprietary “Autonomy Ecosystem” is an end-to-end framework encompassing “data collection, autonomy training, autonomy evaluation, autonomy analytics, and autonomy deployment. We are using government open standards-based methods like Open Mission System (OMS) to integrate the wide variety of elements into the Autonomy Ecosystem. This is to demonstrate our commitment to a vendor agnostic open autonomy ecosystem.”

While in 2011 the Air Force passed on the MQ-9’s proposed turbofan-powered successor, the MQ-20 Avenger, the nine prototypes built have busily led a second career testing diverse payloads and autonomous and collaborative capabilities. Brinkley wrote, “As we look at the future of UAS and AI, we believe that using a jet powered platform gives us a better baseline for our test flights as a surrogate test platform for future designs.”

The company’s AI has proven able to act as single-agents capable of reactively avoiding threats while pursuing mission objectives, or collaborating with other drones to chase an aerial target. Meanwhile, an overlooking hierarchical AI analyzed the sensor data of its subordinates and issued commands in real time based on its assessment of the world state without human intervention.

Meanwhile, the company is retrofitting autonomous capabilities onto MQ-9A Reaper drones including autonomous flight and sensor systems, as well as automatic track recognition (ATR) of ground targets in conjunction with an improved inverse Synthetic Aperture Radar (ISAR). General Atomics has also teased an unnamed swarming aircraft with a 40-pound payload.

GA-ASI Gambit

General Atomic’s CCA concept is apparently based on a previously unveiled “systems of systems” called Gambit. The four subvariants share 70% of components in common for efficiency but are configured for very different missions.

The unarmed Gambit 1 is designed to serve a long-range, long-endurance air domain awareness role, helping detect and illuminate targets so that manned jet fighters don’t expose themselves doing so.

Gambit 2, meanwhile, is armed with missiles for air-to-air combat. Gambit 3 will be like Gambit 2 but serve as an unarmed adversary for training dogfights. Special software will allow it to mimic performance characteristics of various aircraft to improve the fidelity of the training experience. 

Finally Gambit 4 exhibits a stealth blended-wing body intended to enable long-endurance penetration of defended airspace in a “combat reconnaissance” role.

“These and the other aircraft in the series are highly common and the use of a common core and other common components, as well as advanced manufacturing techniques, mean they will be comparatively less expensive and quicker to produce,” Brinkley wrote. “That’s important because it’ll enable building a comparatively large force of them quicker than an equivalent force of human-piloted fighters and without the need for a complex and expensive training pipeline for a cadre of human fighter pilots.”

This multi-platform approach seems aimed at distributing costs rather than wrapping a bunch of capabilities in one pricier platform, enabling CCAs in training or sensor-bearing roles to be built and operated more cheaply than those specialized in penetrating strike and air-to-air combat.

Mystery Resolved? Offboard Sensing Station (OBSS) and General Atomics XQ-67A

In 2021, the Air Force Research Laboratory awarded $17 million each to General Atomics and Kratos for a sensor-wielding jet-powered drone veiled in secrecy. This project was a follow-on to three prior Low-Cost Attritable Aircraft programs (LCAAT, LCDAS and LCAAPS), which included the initial purchase of Valkyire. For OBSS, Kratos offered an ISR variant of the XQ-58A called Demogorgon, while General Atomic’s XQ-67A design concept appeared very similar to its Gambit, matching conceptually with its Gambit 1.

The XQ-67A won the competition and a prototype made its first flight on February 29—though its raison d’etre remained unclear. Afterall, didn’t a loyal wingman with air domain sensors, presumably radar or IRST, functionally overlap with CCA?

But more recent AFRL statements imply OBSS was as much aimed at testing the ability for industry to rapidly generate “affordable mass’” of attritable drones at low R&D costs in a very short timeframe by developing a “genus’”airframe using mature subsystems and hull components. This could be rapidly evolved into new models by swapping in modular mission kits. 

Indeed, it was revealed an armed Off Board Weapon System was also conceived—supposedly faster and with greater range but less endurance than OBSS. However, little else is known about OBWS and its current status. Likewise, whether the initial OBSS variant is bound for an operational role or will solely serve as a demonstrator is unclear.