Makes sense, thanks for the explanation. The compressed map representation + test-time training part sounds especially interesting.

Does the approach hold up well when the environment changes over time (lighting, objects moved, etc.), or does it assume mostly static scenes?