The languages of the world use a wide range of different categorization schemes to specify space and spatial relations. This suggests that when children learn the spatial terms of their native languages, they must search through a hypothesis space of considerable size. Yet children seem to learn the spatial categories of their native languages quickly and accurately. This proposal seeks to determine how this might happen, by exploring a model of that process. The model includes three foundational principles: (a) the creation during learning of language-specific semantic features that are shared across words within a language; (b) mutual exclusivity, the tendency to interpret different words as having different meanings; and (c) endpoint emphasis, a posited attentional bias toward the endpoints of events, rather than their beginnings. The model makes predictions concerning pre-linguistic perception, word learning, and ultimately, the nature of spatial language. These predictions are investigated experimentally and linguistically. The model is also tested computationally, in simulations of the spatial term learning of child language learners. The data on which the simulations rely is to be collected in the home, from families speaking three different languages, namely, English, Korean, and Arabic. The project as a whole has the potential to identify the constraints, and also the language-malleable mechanisms, at play in children's learning of spatial terms. It may also uncover the effect of such mechanisms on the kinds of spatial meanings that language may carry.