Flight Test of a Collision Avoidance Neural Network with Run-Time Assurance

D. Cofer, R. Sattigeri, I. Amundson, J. Babar, S. Hasan, E. Smith, K. Nukala, D. Osipychev, M. Moser, J. Paunicka, D. Margineantu, L. Timmerman, J. Stringfield

Digital Avionics Systems Conference, September 2022

Our team is developing assurance technologies that can support the use of machine learning in the design of safety-critical aircraft systems. These capabilities have been integrated on Boeing's Autonomy Testbed Aircraft to show that they can provide evidence of correct operation and safety guarantees needed by real aircraft. We have applied run-time assurance along with formal methods synthesis, modeling, and analysis tools to an airborne collision avoidance system based on a neural network. This system was demonstrated in flight and shown to correctly monitor neural network operation and intervene when needed to prevent violation of the "remain well clear" safety requirement relative to an intruder aircraft.