New System Combines SLAM and Language Models for Online 3D Scene Mapping

This article surveys advances in mobile robot mapping and semantic scene understanding, highlighting the limitations of voxel grids, NeRFs, and closed-vocabulary systems. It introduces LEGS, a novel framework that combines online 3D Gaussian Splatting with language-aligned feature supervision, enabling robots to build real-time, open-vocabulary semantic maps and localize objects through natural language queries in large-scale environments.