The construction of a Bayesian network requires the exact specification of local conditional probability distributions for all the variables in the network. In case of limited data or disagreeing and/or partial expert opinions, this requirement is clearly unrealistic and renders the resulting model arbitrary. The goal of this talk is to explain how imprecise probability theory, which, basically, is the theory of sets of probability distributions, allows us to deal with this type of model uncertainty in a robust manner. I intend to give an overview of recent developments on this topic, at an introductory level, ranging from computational challenges to foundational questions.