Eventually we would like the network to be the brains of a robot which exists in a playpen-like world of blocks and other simple objects. In the short run, we must simulate the world. This requires an account of what regularities there are in the world, including linguistic regularities. Even when Playpen is embedded in a robot, a theory of the development of spatial cognition must include an account of what is out there to be learned. For the purposes of this paper, we will remain vague in this regard; details will come out when we consider particular linguistic relations, as we will do in the next report on Playpen. We do assume, however, that the linguistic input to the child makes a fundamental distinction between expressions for things and expressions for relations. This fact imposes constraints on the network: it must have the capacity to represent objects as collections of features and relations as pairs of objects (which are collections of features).
The world that infants are exposed to includes language, as well as other sorts of stimuli, from the start. But because infants have not yet figured out the phonology of the target language, linguistic input is probably completely irrelevant for the learning of spatial concepts. Once children are capable of recognizing particular words, normally early in the second year, they are in a position to learn the spatial concepts associated with words. Thus it is convenient to divide children's task into two phases, a pre-linguistic phase in which they learn about space as they observe and manipulate the things around them, and a linguistic phase in which they also receive linguistic input, input within which they are capable of distinguishing particular words. In this latter phase, non-linguistic learning about space continues, but it is supplemented by the co-occurrence of words and phrases with particular scenes.