LEGS

Yu 2024 · Paper

One-line summary — Language-Embedded Gaussian Splats: a mobile robot incrementally builds a room-scale 3DGS map with a scale-conditioned CLIP feature field online as it drives, matching LERF’s open-vocabulary query success while training 3.5x faster.

Problem

Semantic 3D maps are valuable for “searching for objects of interest in offices, warehouses, stores, and homes”, but LERF — the reference method for language-embedded 3D — requires offline per-scene NeRF optimization with all poses known in advance (as does LangSplat, which must train a scene-specific autoencoder over all images first). A robot needs the opposite workflow: build the semantic map while traversing a previously unseen environment (>= 750 sq ft in this paper), from constrained trajectories where cameras run near-parallel to shelves and walls, with accumulating pose drift, and answer queries immediately. Point-cloud approaches (ConceptFusion, CLIP-Fields) fuse a single feature per point, losing the multi-scale object-vs-part reasoning that LERF’s scale-conditioned field provides.

Method & architecture

A Fetch robot with a forward RGB-D Realsense D455 and two side-facing ZED 2 stereo cameras (known extrinsics); processing streams to a desktop with two RTX 4090s (one for LEGS training, ~15 GB; one for DROID-SLAM, up to 18 GB):

Results

Why it matters for SLAM

LEGS is a concrete step from “SLAM as geometry” toward Spatial AI: a robot maps a store or warehouse and can immediately be asked “where is the first aid kit?”. It was among the first systems to combine online 3DGS training with language-aligned feature supervision, showing that the LERF idea survives the move from capture-then-optimize to incremental robot mapping — with classical SLAM machinery (DROID-SLAM tracking + periodic global BA) supplying the pose consistency that makes online splat training work at room scale.