Breaking News
U.S. Air Force’s X-62A Aircraft Hosts Lockheed Martin’s Tactical AI for Autonomous Missile-Evasion Tests.
Lockheed Martin’s Skunk Works and the U.S. Air Force Test Pilot School flew a tactical artificial intelligence agent aboard the X-62A VISTA F-16 to demonstrate autonomous missile-evasion maneuvers. The milestone moves AI from simulation into live fighter operations, signaling a potential shift in how the U.S. Air Force prepares for high-threat, integrated air defense environments.
On 23 February 2026, Lockheed Martin announced that its Skunk Works division, in collaboration with the U.S. Air Force Test Pilot School, had successfully flown a tactical artificial intelligence agent aboard the X-62A VISTA to demonstrate autonomous missile-evasion maneuvers. The announcement describes how an AI system trained through extensive simulation was transitioned onto a live manned testbed, allowing it to control a fighter-class aircraft in dynamic flight conditions. Far from a mere technical showcase, the campaign is framed as a pivotal advance in human-machine teaming: software capable of perceiving, deciding, and adapting in real time alongside pilots, rather than passively automating preset tasks. For air forces preparing to operate in increasingly contested and integrated air defense environments, the ability to delegate critical split-second defensive actions to a validated AI agent signals profound operational and strategic implications.
Lockheed Martin and the U.S. Air Force Test Pilot School flew a tactical AI agent on the X-62A VISTA F-16 to demonstrate autonomous missile-evasion maneuvers in live flight (Picture Source: Lockheed Martin / U.S. Air Force)
The core of the programme is the Have Remy Test Management Project, a joint effort between Skunk Works engineers and U.S. Air Force Test Pilot School (TPS) students using the X-62A VISTA as a high-performance autonomy testbed at Edwards Air Force Base. Under Have Remy, TPS students were brought into the full development loop, from defining the threat scenario to evaluating the behaviour of autonomous agents in flight. The scenario chosen was deliberately demanding: a missile-evasion profile requiring split-second decision-making and three-dimensional manoeuvring at the limits of the VISTA’s flight envelope, with the AI assuming direct control of the aircraft when a simulated surface-to-air missile launch was detected. In these test points, the onboard autonomy stack sensed the threat cue, generated an appropriate defensive response and flew the manoeuvre without the pilot touching the control column, while the safety crew remained ready to intervene.
To make such behaviour acceptable on a manned aircraft, Lockheed Martin relied on an intensive simulation-to-flight pipeline. The tactical AI agent was first trained in hours using billions of simulated missions run on Skunk Works’ Supermassive simulation engine, exposing it to a wide variety of engagement geometries, missile performance assumptions and initial conditions. A high-fidelity F-16 simulator, tuned to mirror both the aerodynamics and the modified flight control system of the VISTA, then allowed engineers and TPS students to validate the agent’s decision logic and handling-qualities impact before committing to real sorties. Only after this virtual test campaign did the AI move onto the X-62A for live evaluation, where it was flown through more than one hundred test points that demonstrated consistent transfer of the missile-evasion behaviour from simulation to the real aircraft.
The X-62A VISTA itself is central to this approach. Based on a two-seat F-16D, it has been heavily modified as a Variable In-flight Simulator Test Aircraft, with programmable flight-control laws, additional sensor suites and high-performance onboard computing that allow it to emulate other aircraft and host experimental autonomy software. Although the VISTA has previously flown AI agents under U.S. Air Force Research Laboratory and DARPA programmes, this campaign marks the first time a Lockheed-developed tactical AI system has had direct control authority over the aircraft. In the missile-evasion tests, the AI was effectively integrated into the fly-by-wire loop, commanding control surfaces and thrust to execute three-axis manoeuvres that stayed within structural and handling limits while aggressively changing the aircraft’s aspect relative to the simulated threat.
Lockheed Martin presents Have Remy as more than a one-off demonstration; it is described as a model for the accelerated development of tactical AI. Engineers used a fly-fix-fly approach, replaying flight data in simulation to understand how the agent behaved, updating the autonomy stack and then redeploying refined versions back to the aircraft within hours. This real-to-sim and sim-to-real loop is intended to give AI engineers the kind of iterative, agile cycle common in software development while preserving the rigour expected in flight-test campaigns. For TPS students, being embedded in that loop provided hands-on exposure to monitoring, interpreting and constraining AI behaviour, including the development of new monitoring algorithms to quantify how closely real-world performance matches what was seen in synthetic environments. From an operational perspective, the work focuses squarely on one of the most demanding tasks in air combat: surviving modern surface-to-air missiles and air-to-air weapons launched within a severely compressed decision window, and showing that an autonomous agent can manage such engagements under realistic constraints.
Looking ahead, the tactical AI architectures demonstrated on the X-62A VISTA could, hypothetically and through careful, incremental validatio, transition onto other U.S. Air Force platforms as onboard tactical autonomy layers. In the near term, selected decision logics may be deployed as optional, tightly bounded assistance modes within current fourth- and fifth-generation fighters. These could recommend or autonomously execute pre-authorized evasive maneuvers when specific threat envelopes are detected, always preserving pilot authority and override capability. Over the longer horizon, similar AI agents might be integrated more deeply within collaborative combat aircraft concepts, forming the cognitive core of self-protection suites for uncrewed escorts operating alongside crewed fighters in contested airspace.
Even larger and more specialized assets, such as tankers, ISR aircraft, or stand-off jamming platforms, could, in principle, gain from derivative, mission-tailored implementations of this technology. Such systems would provide optimized last-resort escape options if a platform were unexpectedly drawn into a hostile engagement zone. While any operational deployment would depend on future policy, safety, and certification frameworks, the X-62A program already provides a tangible model for how a standardized, flight-proven autonomy stack could be adapted across a broad spectrum of air assets, enhancing survivability, efficiency, and mission resilience while keeping human operators decisively in command.
This latest milestone on the X-62A VISTA sits within a broader push in the United States toward collaborative combat aircraft and advanced human-machine teaming concepts. Previous autonomous trials on the same platform under other programmes showed that AI agents could safely conduct demanding within-visual-range manoeuvring in structured trials, while Have Remy shifts the emphasis explicitly to missile survivability and the transfer of complex, safety-critical behaviours from classroom to cockpit. For the U.S. Air Force, the underlying idea is that future pilots will fly in highly networked environments where threats emerge faster than a human can manually process sensor data, evaluate options and command the aircraft. In such a context, a trusted AI that can assume temporary control to execute an optimal escape manoeuvre, then hand back to the crew, offers a way to extend the pilot’s decision loop without removing the human from overall mission command.
By transferring a missile-evasion AI agent from massive simulation runs into more than a hundred live test points on a manned X-62A VISTA, Lockheed Martin Skunk Works and the U.S. Air Force Test Pilot School have demonstrated that tactical autonomy is no longer confined to laboratories or small uncrewed demonstrators. The Have Remy project shows that an AI system can be trained, validated in high-fidelity simulators and then entrusted with direct control of a fighter-class aircraft for one of the most demanding tasks in air combat, while remaining fully supervised and overrideable by human pilots. As integrated air defences grow more capable and reaction times shrink, this combination of human judgment and machine-speed decision-making points toward a future in which survivability and mission effectiveness depend increasingly on tactical AI that is designed from the outset to decide, act and adapt alongside aircrews in contested skies.
Written by Teoman S. Nicanci – Defense Analyst, Army Recognition Group
Teoman S. Nicanci holds degrees in Political Science, Comparative and International Politics, and International Relations and Diplomacy from leading Belgian universities, with research focused on Russian strategic behavior, defense technology, and modern warfare. He is a defense analyst at Army Recognition, specializing in the global defense industry, military armament, and emerging defense technologies.