A principled data engine with 3 million spatial samples significantly improves model performance on spatial reasoning tasks—showing 19% average improvement—and provides the infrastructure needed to build spatially-aware AI systems.
OpenSpatial is an open-source system for generating high-quality spatial data at scale. It uses 3D bounding boxes to create a dataset of 3 million samples across five spatial reasoning tasks, enabling AI models to better understand 3D scenes, measurements, relationships, and camera perspectives.