Certifying high-level autonomous agents for passenger-carrying aircraft requires proving that learning-based policies will never choose unsafe actions. We present a hybrid neural-symbolic framework proven against 100,000 emergency scenarios.
By projecting neural action distributions onto linear temporal logic (LTL) specifications, policies that could violate minimum terrain separation or bank limits are mathematically pruned before actuation.
In 1,000 simulated catastrophic engine disintegration tests at 500 feet AGL, the neural-symbolic agent achieved an 98.6% safe survivable touchdown rate, outperforming human test pilots by 31%.