SpatialLM

SpatialLM looks worth evaluating as an emerging spatial-understanding model, but the evidence does not yet support treating it like a mature general-purpose platform. What the sources do support is a clear product direction: it takes point clouds, reconstructed 3D scenes, and in some posts video/LiDAR-related inputs, then outputs structured scene understanding such as walls, doors, windows, oriented object boxes, and semantics. A better analogy is a 3D scene parsing and spatial reasoning model, not a chatbot and not a full CAD/BIM authoring tool.

Its practical value is in turning messy 3D captures into machine-readable scene structure for robotics perception pipelines, indoor digitization, interactive scene understanding, and downstream retrieval or planning. Some sources mention newer 1.1/1.5 versions adding handheld video reconstruction, LiDAR, synthetic mesh sampling, and robustness to incomplete point clouds. Those claims mainly come from explainers and repost-style summaries, so they indicate capability direction rather than guaranteed production readiness.