Semantic learning of action-related words (e.g. cut, kick) was modelled in brain-constrained neural networks fashioned according to the neuroanatomical structure of relevant parts of the human cortex. These networks were applied to simulate infants uttering action words when performing related actions, where word meaning is learnt and grounded in the world. During learning neuronal circuits emerged each interlinking information about a word form and its related action type, so that word form perception-simulation in the network brought about immediate activation of semantically-related action information. We discuss this model in view of established evidence from cognitive neuroscience and argue that a mechanism similar to the network’s word-action associations and resultant topographically-organised memory representations may provide the biological basis of understanding and mentalising action-related concepts in humans. Further, the results shed light on how perceptual and action-related information may shape the structure of semantic circuits across cortical areas.