The aim of cross-modal spatial representation is a latent space in which heterogeneous perceptual signals, images, 3D scenes, video, audio, and language are aligned so that the spatial structure each carries is preserved rather than averaged away. Where the previous chapter asked how much geometry a single modality holds on its own, the concern here is the opposite and harder one: how signals that encode space in incompatible formats- a pixel grid, a world-frame point set, a waveform, a sentence- can be brought into correspondence without the alignment itself distorting what it aligns.
No single recipe achieves this. Methods differ in what they treat as the unit of alignment, in the objective that binds one modality to another, in how tightly they anchor the space to physical geometry, and in the linguistic interface through which it is queried. These choices are not independent, and the chapter first makes them explicit as a small set of design axes before turning to the individual cross-modal pairings in which they play out.