IN Brief:
- The X-62 completed 27 AI-controlled intercepts against a live T-38 target across eight flights.
- An operational Legion Pod supplied real-time infrared tracking data to the autonomous agent.
- Integration and ground testing took three months before the flight-test series began.
Lockheed Martin Skunk Works and the US Air Force Test Pilot School have completed 27 artificial-intelligence-controlled intercepts across eight flights, using live sensor information to guide the X-62 Variable In-flight Simulation Test Aircraft against a T-38 target aircraft.
The trials connected a Legion Pod infrared search-and-track sensor to an autonomous agent aboard the X-62. Instead of relying on simulated target information, the software consumed real-time onboard sensor feeds and commanded the aircraft through the intercept manoeuvre.
Lockheed Martin integrated and ground-tested the agents with the X-62 in three months using its Supermassive AI-agent generation capability. The flight series covered simulation, training, integration, and airborne execution within one development cycle.
Live sensor data drives the aircraft
Earlier autonomy demonstrations have shown that an AI agent can control an aircraft or perform tactical behaviours in a constrained test environment. This series adds a more demanding connection by feeding the agent information from an operational sensor tracking a live aircraft, requiring perception data and flight control to remain linked throughout the manoeuvre.
The Legion Pod tracked the T-38 and supplied targeting information to the autonomous agent. The agent then piloted the X-62 into an intercept position, closing the sensor-to-action loop described by Lockheed Martin. The release does not state that weapons were simulated, controlled, or employed during the tests.
An intercept is a defined flight-control and mission-systems task rather than proof of a complete autonomous combat capability. Target classification, rules of engagement, weapons management, communications, and human authority would each require separate integration and assurance before an operational system could be fielded.
Even within the narrower trial, the engineering task is substantial. Infrared tracking data must be processed, passed through secure interfaces, interpreted by the agent, and converted into aircraft commands while both aircraft continue to move. Latency, track quality, software timing, and flight-envelope protection can each affect whether the behaviour remains predictable.
The X-62 is suited to this work because its open hardware and software architectures allow new control laws, mission systems, and autonomy packages to be installed without building a dedicated test aircraft for every experiment. It exposes software to real aerodynamics, sensor noise, communications behaviour, and safety constraints that cannot be reproduced perfectly in simulation.
Repeatability will determine operational value
Twenty-seven intercepts across eight flights provide more evidence than a single demonstration, but the release does not disclose the range of starting conditions, target manoeuvres, weather, sensor geometry, or number of agent variants tested. Those details are needed to judge whether performance was robust beyond a controlled scenario.
The programme’s next step is tied to the X-62 Mission Systems Upgrade, which is intended to support the integration of combat systems, sensors, and airborne AI agents within a next-generation mesh network. That would extend the aircraft from largely self-contained autonomy trials towards experiments in which information and decisions are distributed across several platforms or nodes.
Networked autonomy introduces further engineering problems. Information may arrive from sensors with different accuracy and delay, communications links may be interrupted, and several agents may need to coordinate without producing conflicting actions. Assurance must cover the behaviour of the complete network when data is incomplete, late, or incorrect.
Rapid agent generation is another part of the programme’s claim. Completing integration and ground testing in three months suggests that reusable software tools and architectures can shorten the route from an algorithm to a flight experiment. Speed remains useful only when it is matched by independent test, cyber assurance, safety review, and documentation that allows results to be reproduced.
For crewed aircraft, the nearer-term application may be assistance rather than replacement. An autonomous system could manage positioning, sensor employment, or routine tactical tasks while a pilot retains authority over wider mission decisions. That division still requires interfaces that show what the agent is doing, why it is acting, and when control should be transferred.
The trial therefore represents a defined integration milestone without settling the wider questions surrounding operational autonomy. It demonstrates that live infrared sensor data can drive an onboard AI agent through repeated airborne intercepts, while leaving weapons employment, contested-network performance, human command, and certification for later work.
Lockheed Martin and the Test Pilot School have moved the experiment beyond synthetic targets and into a live aircraft encounter. The next measure of progress will be whether the architecture can handle broader scenarios, degraded information, additional mission systems, and networked partners without losing the predictable behaviour required for flight safety and military use.
