Aurora Flight Sciences Demos DARPA-Funded AI Control
Aurora Flight Sciences, a Boeing subsidiary, has demoed FALCON, an AI control system for adaptive maritime autonomy. The system, funded by DARPA, integrates sensor fusion and introspective flight controls. The demonstration showcases the application of advanced AI in enabling autonomous operations in complex environments.
- The FALCON (Fast Adaptation and Learning for Control Online) system is a "computationally lightweight" and platform-agnostic control architecture, indicating its design for deployment in resource-constrained edge environments. - This work is part of DARPA's Learning Introspective Control (LINC) program, which aims to develop machine learning technologies that allow systems to adapt in real-time to unforeseen events not accounted for in their initial design. - The LINC program's goal is to create AI-powered controllers that require minimal computing power, on the scale of a cell phone, to enhance the safety and resilience of various platforms, including ground vehicles, ships, and drone swarms. - The recent maritime demonstration utilized 1.5-meter and 5-meter uncrewed surface vessels (USVs) in collaboration with MIT's Aerospace Controls Laboratory and Marine Autonomy Laboratory. - Test scenarios included underway replenishment (UNREP) and navigating a simulated Suez Canal, where the system had to overcome induced disturbances like thruster failure, wind loading, and the Venturi effect. - Aurora's adaptive controller successfully compensated for these disturbances where a conventional controller failed, enabling the USV to maintain safe operation. - This adaptive control technology is designed to be applicable across a wide range of platforms, including both air and land vehicles, and can function as either an assistant to a human operator or as the primary vehicle control system. - The challenge of certifying such adaptive, AI-based systems for aerospace applications will likely involve new approaches to meet the objectives of standards like DO-178C, which traditionally rely on deterministic and verifiable software behavior.
