Headlines about autonomous weapons often evoke visions of The Terminator or The Matrix: machines independently deciding whom to kill. The reality is far less dramatic and far more consequential.
Take the collaborative combat aircraft, or CCA, that militaries around the world are developing to assist existing and future crewed fighter jets. They could be equipped for missions ranging from air engagements to intelligence, surveillance and reconnaissance. These aircraft would be tasked and controlled by a human operator, but navigate and operate semi-autonomously using AI-enabled software.
CCA and other emerging semi-autonomous and autonomous systems are not a sudden leap into science fiction. They are the latest step in a technological evolution that has been underway for more than half a century, steadily shifting more battlefield tasks from human operators to increasingly capable machines.
Having spent part of my career designing and testing missiles, drones and unmanned aircraft, I’ve watched that evolution firsthand. Today’s systems are the product of decades of incremental advances in guidance systems, automation, machine learning and artificial intelligence. However, while the technologies have advanced, the question of human control has remained unresolved.
A look back
From my experience, there is an important distinction between automation and autonomy. Automation executes preplanned commands when predefined conditions are met. Autonomy goes further, allowing a system to evaluate its environment, adapt to changing conditions and carry out assigned missions with limited human direction.
Initial projectiles were unguided and simply ballistic, with neither automation nor autonomy. Throw a rock, shoot an arrow, fire a bullet, drop a bomb. Aim them and boost their velocity and they are on the way to the target. Early rockets like the Hydra and Zuni extended range, but remained unguided.
The late 1950s introduced Lock-on-Before-Launch (LOBL) guidance for air-to-air heat seeking missiles, like the Sidewinder, where very basic automation guided the weapon toward a designated target. In the late 1960s, air-to-ground weapons like the Maverick incorporated LOBL guidance requiring the weapon to lock on to a target leveraging visual, infrared, electro-optical, or radar seekers. The information enabled by inexpensive signal processors would ensure accurate guidance.
Later, using GPS with digital signal processing along with flight path guidance would allow “smart bombs” like the dual-mode Enhanced Paveway to augment precision targeting, even when line of sight wasn’t available.
Next came missiles like the Phoenix in the 1970s and AMRAAM in the 1990s that could engage targets beyond visual range. Rather than simply flying toward a designated target, they could update their course during flight and complete the engagement using onboard sensors and software.
This was a notable leap forward. The weapon was no longer just following instructions — it was making limited tactical decisions within carefully defined parameters. That meant the next generation of Lock-on-After-Launch weapons could be carried internally and launched more quickly with minimal data from the launch platform due to their advanced software. Consider the Tomahawk missiles, which used digital scene-matching algorithms to refine their navigation during cruise to greatly improve target location accuracy. Once the operator launches the weapon, they no longer have control to change targets or abort the mission if conditions change.
By the 1980s, that evolution had reached programs I would eventually help develop. I was the test director on the millimeter-wave Maverick program, a follow-on to the Wasp mini-missile demonstration. These were attempts at early autonomy that relied on high-resolution radars and pattern matching to search for and engage massed armor or air defense units.
In the early 2000s, I was chief engineer of one contender for the joint DARPA-U.S. Army Unmanned Combat Armed Rotorcraft program, an effort to field a “robotic rotary wingman” to augment manned platforms. Our proposal included complete autonomous operation of a patrolling triad of UCAR that would use advanced signal processing and data analytics, along with electro-optical sensors, to triangulate and engage threats to clear an area prior to deploying troops.
DARPA embraced the technical feasibility, but the Army was reluctant to delegate so much battlefield decision-making to software that lacked the maturity, transparency and reliability commanders required. The obstacle was less whether the aircraft could fly autonomously than whether military leaders were prepared to trust autonomous systems with greater operational and lethal authority.
Advancements in autonomy provided an opportunity for a captive-carry demonstration project I managed in the 2000s. By then, autonomy was considered mature enough to search, classify and prioritize targets over a large area in real-world conditions. Based upon the airframe of a small loitering munition, the mission concept was to find Soviet-era Scud launchers, tanks and military vehicles in an area and attack the highest-value target.
The test aircraft, using a large parallel processor with advanced pattern correlation, machine learning and automated decision logic, successfully discriminated between military and civilian vehicles in real time over a 50-square-kilometer search area. Once the aircraft reached the initial point, human involvement was limited to navigation cues provided to the captive flight test pilot. In an operational scenario, there would be no time for a human to intervene in the lethality process as target selection to target prosecution would only be seconds.
What’s ahead
While many of these programs never entered production, the technologies they proved were matured, migrated into other programs and steadily became part of today’s operational systems.
For autonomous technologies, the real question is no longer whether machines can perform increasingly complex combat tasks — they can. It’s where we choose to draw the line between machine autonomy and human judgment, especially when lethal force is involved.
The Pentagon already has a directive focused on autonomy that requires autonomous and semi-autonomous weapon systems to “be designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force.” International Humanitarian Law, which instructs combatants to operate within defined rules of engagement, requires judgment about the particular circumstances at the time of attack. Can a weapon operate autonomously while still allowing the people responsible for the attack to make the context-dependent judgments that IHL requires? Who is responsible if an autonomous weapon can’t distinguish that battlefield conditions have changed after the weapon is launched? Will the speed of the engagement preclude human intervention?
In July, a bipartisan group of U.S. lawmakers introduced legislation that would require human control over autonomous weapons. “The bill would establish a clear legal requirement that commanders and operators retain appropriate levels of human judgment in decisions involving the use of lethal force,” according to a press release from Rep. Don Beyer (D-Va.), one of the co-sponsors.
The United States may insist that humans remain responsible for decisions involving lethal force; whether our adversaries adopt the same restraint is less certain. In the compressed decision cycle of future combat, commanders may face pressure to delegate more authority to autonomous systems to gain or maintain a battlefield advantage.
Technology will continue advancing. Whether policy, law and ethics keep pace is the decision that still belongs to us.
Aerospace America publishes a rich variety of opinions relevant to the future of aerospace. The views expressed are those of the author(s) and do not necessarily reflect those of our publisher, AIAA.
