ImagineNav++: Prompting VLMs as Embodied Navigator through Scene Imagination

ABSTRACT

Visual navigation in home environments often fails because textual planning cannot capture scene geometry. We propose ImagineNav++, which uses a VLM to 'imagine' future viewpoints from candidate robot views, turning navigation into a 'best-view' selection problem. Our Selective Foveation Memory mechanism integrates keyframe observations into a compact representation for long-term spatial reasoning. ImagineNav++ achieves SOTA performance in mapless settings, outperforming most traditional map-based methods.