Door-to-Floor: LLM-Driven Floorplan Reconstruction From 3D LiDAR Scans

Floorplan reconstruction captures the structure and layout of a three-dimensional space, which is essential for numerous applications such as robot localization, indoor navigation, and path planning. In recent years, with the continuous development of large language models, some models have been trained to reconstruct indoor environment structures. However, a major limitation of these models is their capability to address only small-scale, room-level scenes, making it challenging to cope with large-scale scenes at the floor level. To address this challenge, we introduce a hierarchical spatial segmentation method based on door frame recognition. By identifying door frames, large-scale floor scenes are divided into multiple small and regular rooms. Subsequently, semantic inference is applied to each subspace using a large language model, thereby achieving an accurate floorplan reconstruction. This approach overcomes the limitation of current large language models, which are restricted to handling small, room-level scenarios. We conducted an experiment on door frame recognition within a complex, large-scale environment containing nine door frames. The method successfully detected all the door frames, which allowed the precise division of the entire space into nine distinct rooms. Additionally, our floorplan reconstruction method is tested on both public and self-collected datasets, showing a 56.3% improvement over state-of-the-art (SOTA) methods.