In the field of computational neuroscience, we develop distributed models of the cortex to account for perceptual and sensorimotor capabilities. Adopting a mesoscopic level of modeling with dynamic neural fields representing topologically organized populations of cortical columns, we propose various learning rules, competition mechanisms and interconnection schemes. Under the right conditions, these allow the emergence of spatially coherent bumps of activity yielding attentional properties and high robustness to noise. We focus in this poster on the distributed, continuous and unsupervised learning of multi-sensory representations and sensorimotor behaviors. Learning should lead to the emergence of coherent bumps of activity in cortical maps when correlated stimuli are presented, with the possibility to recall potentially missing modalities. This raises the issues of the necessity and nature of generalization in a multimodal context. Research projects on this topic range from the introduction of constrained self-organization with associative maps, adaptive learning rule parameters for multimodal selectivity, to local contingencies in high dimensional sensorimotor spaces.