Objaverse

A massive open library of 3D object models, led by the Allen Institute for AI.

Objaverse is an open 3D object dataset led by the Allen Institute for AI (Ai2). The 2022 version 1.0 contains more than 800,000 models with titles and tags; the 2023 Objaverse-XL, built with several partner institutions, expanded this to over 10 million, sourced from human-made models, photogrammetry, and scans of physical artifacts. 3D data had previously been far smaller in scale than image and text data, which limited 3D vision and generative models. In embodied AI, it's mainly used as a source of simulation assets — for example, Holodeck uses a large language model to plan a room's layout, then retrieves objects from Objaverse to populate the scene. Most models come with only shape and texture, so before they can be used in a simulator, they usually need collision geometry and physical parameters like mass and friction added.

ExampleGiven a text description like “a study crammed with books,” Holodeck retrieves models such as bookshelves, a desk, and a lamp from Objaverse and arranges them in an AI2-THOR scene for navigation and manipulation training.

Also called

Objaverse 1.0, Objaverse-XL

Related

Simulation Assets · Holodeck · ShapeNet · Single-Image 3D Reconstruction · PartNet-Mobility · Procedural Generation

Sources

Objaverse: A Universe of Annotated 3D Objects (arXiv 2212.08051)
Objaverse-XL: A Universe of 10M+ 3D Objects (arXiv 2307.05663)
Holodeck: Language Guided Generation of 3D Embodied AI Environments (arXiv 2312.09067)

As of

2023-07