Registering two captures of the same indoor space taken at different times underpins persistent spatial memory for robots and augmented-reality systems. The abstract frames the egocentric setting as the realistic version of this task, and the paper proposes a training-free method that uses vision-language models to perform the registration.

The approach, called PROSE, avoids task-specific training, which the authors position as a more scalable route to scene registration. The abstract does not include quantitative results, so the contribution is described in terms of the problem framing and the proposed method.

The significance is that a training-free, vision-language-model-based approach could make persistent spatial memory more practical for robots and AR systems operating in indoor environments.