Aurora Flight Sciences has successfully demonstrated its Fast Adaptation and Learning for Control Online (FALCON) technology, a machine learning-based control architecture designed to help vehicles adapt to unforeseen hazards and environmental shifts in real time.
Developed in collaboration with the Massachusetts Institute of Technology (MIT) Aerospace Controls Laboratory and the MIT Marine Autonomy Laboratory, this technology is part of the Defense Advanced Research Projects Agency’s (DARPA) Learning Introspective Control (LINC) program. The system allows land, maritime, and aerial vehicles to modify their control laws on the fly, serving as either an assistant to a human operator or as the primary method of vehicle control.
During a demonstration held in late 2025, researchers utilized a 1.5-meter-long Uncrewed Surface Vessel (USV) paired with a 5-meter-long USV. The exercise simulated underway replenishment, a scenario where two vessels must maintain a consistent relative position to transfer supplies. The technology was tested against challenges including wind loading, thruster failure, and simulated Venturi effects to determine how effectively the vessels could stay within a designated safe zone.
The trial compared manual human control against two levels of AI integration. In AI-assisted mode, the system compensates for hazards to help a human operator maintain stability. In AI-guided mode, the human operator sets high-level parameters like speed and position while the AI guides the vehicle in real time.
Data collected during the maneuvers showed that a manual operator remained within the safe zone 63% of the time. This figure rose to 82% in AI-assisted mode and reached 94% under AI-guided control. Furthermore, when hazards were introduced, the time required to regain control and return to safe operation was reduced by an average of 61% when using the AI-guided system compared to manual piloting.
As a subsidiary of Boeing, Aurora Flight Sciences continues to focus on autonomous systems and advanced vehicle configurations. The team is currently refining the FALCON algorithms in preparation for a follow-up demonstration scheduled for this summer.
