What if city navigation models learned invisible geometry instead of what things look like?

Navigation agents need to predict where they can move, not what buildings look like. This work replaces flat occupancy grids with 3D isovists—360° depth maps capturing free space around an agent. Trained on Manhattan and Paris data, a single city-blind model unexpectedly encodes city identity in its learned dynamics rather than visual appearance, making the spatial signature linearly decodable from latent trajectories. Open dataset and code included.