intelligolabs/Le-RNR-Map

Combines CLIP embeddings with neural radiance field maps to enable spatial querying of 3D scenes using natural language, supporting both single-object localization and multi-object discovery. The architecture integrates a pretrained autoencoder (GSN) for encoding rendered views into compact spatial representations, which are then indexed by CLIP embeddings for semantic matching. Built on Habitat simulator and Gibson 3D scene datasets, it generates queryable maps through navigation-based scene exploration and produces heatmap outputs indicating object locations.

No commits in the last 6 months.