SpatialLM

Mao 2025 · Paper

One-line summary — Fine-tunes an off-the-shelf LLM to consume 3D point clouds and emit structured indoor models — walls, doors, windows, and oriented object boxes — as Python-dataclass scripts, grounding LLM reasoning in 3D spatial structure.

Problem

Structured indoor modeling — recovering the architectural layout (walls, doors, windows) plus 3D object bounding boxes from raw RGBD scans — has been the domain of task-specific networks (RoomFormer’s polygon queries, V-DETR’s detection heads) or bespoke autoregressive decoders with domain-specific tokens (SceneScript). Meanwhile LLMs excel at reasoning and code generation but have no grounding in 3D space. Two things blocked simply fine-tuning an LLM for the task: no large, high-quality dataset pairing point clouds with structured 3D annotations, and no empirical understanding of how to align point-cloud input with an LLM. SpatialLM (tech report: arXiv 2506.07491, NeurIPS 2025) tackles both.

Method & architecture

F=E(P),PRN×6,  FRK×D,  KN,\mathbf{F} = \mathcal{E}(\mathbf{P}), \qquad \mathbf{P} \in \mathbb{R}^{N \times 6},\; \mathbf{F} \in \mathbb{R}^{K \times D},\; K \ll N,

where each point carries XYZ + RGB. The chosen encoder is Sonata (an encoder-only, self-supervised Point Transformer V3 variant), a two-layer MLP projector, and Qwen2.5-0.5B as the base LLM (603.5M params total; released checkpoints also include a Llama-1B variant).

Results

Why it matters for SLAM

SpatialLM illustrates where semantic mapping is heading: from label-annotated point clouds (OpenScene, ConceptFusion) and scene graphs (ConceptGraphs) toward representations an LLM can directly parse and manipulate. A SLAM system provides the point cloud — the paper’s own demo pipeline runs MASt3R-SLAM on monocular video — and SpatialLM turns it into a compact, editable, machine-readable description of the space, useful for scene editing, AR, robot navigation, digital twins, and embodied AI.