GiRT represents a scene as a sparse field of latent 3D chunks. Each chunk pairs a representation with an uncertainty state. The intended perception loop compares observations with the stored scene, then updates the relevant parts of that scene.
Architecture
Four operations organize the proposal: observe, imagine, fade, and revise. A conditional diffusion prior is intended to support these operations at different noise levels. A tokenizer maps geometry into latent chunks, while rendering and grounding connect the internal state to camera observations.
The prototype defines the field, module boundaries, training order, and runtime interfaces. That structure does not by itself establish a useful learned world model. The tokenizer, prior, and grounding components require training, and their dependencies must work together in an end-to-end loop.
The main unresolved question is whether a persistent uncertain scene representation can be maintained consistently through new observations, missed information, and revision. Before evaluating that claim, the component training and the data transfers between stages need a demonstrated working path.