Grounded Acquisition of Containment Prepositions

We present a developmental approach towards a) pre-linguistic learning of grounded spatial schemas, and b) the acquisition of spatial prepositions based on association from these pre-linguistic concepts. We learn from sentential data, picking out the words most frequently associated with the concept, and show that simple associative structures are adequate for learning object names, or distinctions such as “in” or “out”. A synthetic model of visual attention is used to constrain the set of objects in current focus. We first learn perceptual-object labels from simple 2D multi-agent visual streams cooccurring with word-separated utterances. We show that a notion of proximity between perceptual objects is sufficient to obtain a pre-verbal notion of graded spatial poses. We also demonstrate that the spatial concepts learned on a single shape generalize to other shapes, and correlate the learned perceptual schema with human precepts through a simple psychological