Several neuronal models for autonomous robot behavior show the relevance of biologically inspired approaches. Examples of fundamental capabilities of biological beings successfully emulated in artificial systems can be found in the literature, as conditioned reinforcement or stimulus prediction.\\ Our work aims at providing an autonomous robot with an efficient action control mechanism. Basically, the robot perceives its environment from multimodal inputs, has a set of multimodal actions at its disposal, and must keep fundamental variables (as energy, fitness...) within a viability zone. Our implementation is strongly inspired from spatial and temporal processing of elementary units of the cerebral cortex called cortical columns. First, any perceptive or motor event can be represented by a cortical column, columns being gathered in cortical maps according to their modality. Associative maps combine modal maps on the basis of biological data, enabling management of multimodal events. \\ Second, temporal regularities are extracted to keep traces of sequences of events that often lead to satisfaction of the fundamental needs. This sequence learning mechanism occurs only within a modal map, ensuring a capability of generalisation and transfer of learning when needed.\\ The consistent use of the learnt regularities is provided by propagation of specific cortical activations, calling activations, inside themaps. These calls, seen as hypothesis propagation, are constrained bycurrent perceptions and learnt sequences to achieve the most relevant action.