SpatialGrammar: A DSL for LLM-Based 3D Indoor Scene Generation

A team of researchers has introduced SpatialGrammar, a specialized language designed to create interactive 3D indoor environments from natural language inputs. This system visualizes gravity-aligned configurations as BEV grid placements, ensuring a deterministic process that leads to valid 3D geometries, which allows for reliable constraint verification. Two distinct models have been created: SG-Agent, an iterative closed-loop system that utilizes compiler feedback for refinement, and SG-Mini, a 104M-parameter model developed using synthetic data. This methodology effectively tackles spatial inaccuracies and collisions that frequently occur in LLM-driven scene generation. The findings are documented in a paper available on arXiv, identified by ID 2604.27555.