Artificial intelligence offers major potential to expand the operational capabilities of unmanned aerial systems in defence and safety missions. Yet their practical use in contested and rapidly changing environments remains limited by the functional reliability of AI-based system models under battlefield conditions.
With state-space–based interpretation methods that enables targeted, iterative generation of training data, improving model accuracy particularly in safety-critical edge states we introduce an approach to overcoming these constraints by enhancing the trustworthiness of neural-network–based dynamic models. Additionally, an online reliability monitor evaluates model trust during flight and an adaptive mechanism updates the network when previously unseen conditio ...