A model that tackles the Multiple Object Manipulation task computationally solves a higly complex cognitive task. It needs to learn how to identify and predict the dynamics of various physical objects in different contexts in order to manipulate them. MOSAIC is a model based on the modularity hypothesis: it relies on multiple controllers, one for each object. In this paper we question this modularity characteristic. More precisely, we show that the MOSAIC convergence during learning is quite sensitive to parameter values. To solve this issue, we deﬁne another model (CARMA) which tackles the manipulation problem with a single controller. We provide experimental and theoretical evidence that tend to indicate that non-modularity is the most natural hypothesis.