Complex technical applications often show severe uncertainties, which may vary over time, e.g., situation dependent sensor inaccuracies or anomalies and faults. In order to ease the engineering process for such systems, organic computing principles, e.g., self-adaptation and self-optimization, offer a solution. Hence, machine learning paradigms are needed which work online and which can cope with such dynamically varying uncertainties, but still operate safely all the time. In this work, such a learning paradigm is developed based on the Organic Robot Control Architecture and the incremental learning scheme Directed Self-Learning. It is combined with an explicit uncertainty representation. The core idea is to use a strategy blending scheme to show a good performance and improve it by self-optimizing learning on the one hand in case of high trust, or low uncertainty, respectively. On the other hand, a robust fallback-system is used to ensure safety in situations of high uncertainty. Of course, in such situations learned knowledge has to be protected from corruption. The feasibility of this approach is demonstrated in a simulated pick-and-place scenario with unknown, but changing load masses.