We develop on a model for multi-sensory integration to perform sensorimotor tasks. The aim of the model is to provide missing modality recall and generalization using cortico-inspired mechanisms. The architecture consists in several multilevel cortical maps with a generic structure. Each map has to self organize with a continuous, decentralized and unsupervised learning which provides robustness and adaptability. These self-organizations are constrained by the multi modal context to obtain multi-sensory generalization. More precisely, each modality is represented by a perceptive map and each perception by an activity bump, that emerges thanks to a competition mechanism, in the corresponding map. All perceptive maps are reciprocally and laterally connected to an unique associative map. A competition takes place in the associative map to generate an activity bump which represents the multi-sensory perception. Multi modal constraints are relaxed, thanks to the lateral connections, to converge towards coherent perceptions within a multi modal context. We will present a model of perceptive map using a modulated BCM (Bienenstock Cooper Munro) learning rule to create a self-organization at the map level, which can be influenced by the multimodal context. An unlearning mechanism adds robustness and plasticity to the architecture and makes the self-organization smoother.