We propose SAGE (Semantic Attribute Graph Extraction), a framework for robust room classification and semantic mapping in home robots. SAGE integrates exploration-based visual data with Large Language Model (LLM) reasoning to infer room functionality and construct a Semantic Topological Graph. Unlike conventional single-image or object-based methods, SAGE applies multi-image fusion and probabilistic integration for improved accuracy and consistency. By leveraging the world knowledge embedded in LLMs, SAGE can infer room functionality even in visually ambiguous or partially occluded environments. Experiments in Matterport3D simulations show that SAGE achieves 84.7% accuracy, outperforming YOLOv10 and single-image LLM baselines.
Paper
Full text
SAGE: Semantic Attribute Graph Extraction for Room Understanding
Semantic Scholar · Computer Science · 2025
Abstract
We propose SAGE (Semantic Attribute Graph Extraction), a framework for robust room classification and semantic mapping in home robots. SAGE integrates exploration-based visual data with Large Language Model (LLM) reasoning to infer room functionality and construct a Semantic Topological Graph. Unlike conventional single-image or object-based methods, SAGE applies multi-image fusion and probabilistic integration for improved accuracy and consistency. By leveraging the world knowledge embedded in LLMs, SAGE can infer room functionality even in visually ambiguous or partially occluded environments. Experiments in Matterport3D simulations show that SAGE achieves 84.7% accuracy, outperforming YOLOv10 and single-image LLM baselines.