AmaraSpatial-10K: A 3D Dataset for Spatial Computing and Embodied AI

Researchers have unveiled AmaraSpatial-10K, a collection of over 10,000 synthetic 3D models designed for uses in areas like embodied AI, robotics simulation, game development, and AR/VR. This dataset stands out from typical web-scale collections because each asset is metric-scaled and comes with semantic references. They are provided as .glb files, which include separate PBR material maps, a convex collision hull, reference images, and detailed multi-sentence metadata. The assets cover a range of categories, including indoor items, vehicles, architecture, creatures, and props, all following a consistent spatial standard. Moreover, it features an evaluation suite with a continuous Scale Plausibility Score (SPS) and an LLM Concept Density score to tackle common issues in 3D asset libraries.