# SpatialLM

A 3D large language model that turns point-cloud data from monocular video, RGBD images or LiDAR into structured indoor scene descriptions, including walls, doors, windows and labeled object bounding boxes.

SpatialLM is a 3D large language model that processes point-cloud data and produces structured 3D scene understanding, identifying architectural elements such as walls, doors and windows along with oriented bounding boxes for objects and their semantic categories. Unlike methods that need specialized capture equipment, it accepts point clouds from monocular video, RGBD images or LiDAR.

The multimodal architecture is intended to bridge unstructured 3D geometric data and structured 3D representations for uses such as embodied robotics, autonomous navigation and other 3D scene-analysis tasks. Released models include SpatialLM-Llama-1B and SpatialLM-Qwen-0.5B variants, and inference is tested with Python, PyTorch and CUDA.

Website: https://manycore-research.github.io/SpatialLM