Model transformations support a model-driven design by providing an c translation of abstract models into more concrete ones, and eventually program code. Crucial to a successful application of model transformations is their correctness, in the sense that the meaning (semantics) of the models is preserved. This is especially important if the models not only describe the structure but also the intended behaviour of the systems. Reasoning about and showing correctness is, however, often impossible as the source and target models typically lack a precise definition of their semantics. In this paper, we take a first step towards provably correct behavioural model transformations. In particular, we develop transformations from UML Activities (which are visual models) to programs in TAAL, which is a textual Java-like programming language. Both languages come equipped with formal behavioural semantics, which, moreover, have the same semantic domain. This sets the stage for showing correctness...