Sand3D
Provider:Sandflow
Sand3D is Sandflow's open-source 3D-as-code library: compact Python programs describe 100K+ objects (with meshes and semantic materials), exportable to URDF/USD for robot simulation.
At a glance
- Readiness
- L2 · Semantic & material ready
- Rationale
- Quality varies across the library: every object ships as a Sand3D Python program plus a mesh, with each part carrying a semantic name and a named material (e.g. painted_metal, steel, walnut), reaching L2 semantic material; over 10K articulated objects additionally declare joints (axis and limits), but mass, inertia and friction are defined by the user in code rather than auto-estimated, so the bulk stays at L2 rather than L3 behavior Profile validation.
- Scale
- About 100K 3D objects across five Hugging Face datasets: Objaverse 67K, CAD 24K, Articraft 9.2K, SimReady 2.3K, AutoMate 3.1K; the landing page also advertises a 78K-asset library. One Python program describes one object. (Scale figures conflict: the landing page header says 78K assets while the same page also claims 100K 3D objects in one collection, and the five source datasets sum to roughly 105.6K. Use the ~100K figure; all five HF datasets are CC BY 4.0.)
- License
- CC BY — Datasets are CC BY 4.0 (commercial use with attribution); the sand3d library itself is Apache-2.0 (see GitHub LICENSE). The sand3d-simready subset rewrites NVIDIA SimReady, so for that subset still verify the upstream source-asset license.
- Access
- Open
- Formats
- gltf · urdf · usd · obj · stl
- Runtimes
- isaac-sim · omniverse · other
- Physics properties
- Semantic labelsMDL materialJoint structure
- Domain
- Other
- Published
- 2026-11-08
- Updated
- 2026-11-08
Why it is listed
Raw Blender Python from an LLM is long and brittle. Sand3D flattens CAD-as-code into an s3d.* verb surface plus a prompt-sized API card (about 3,600 tokens), so agents can emit runnable 3D programs. Every object is one deterministic Python program — the same program yields a bitwise-identical mesh — with each part named and every dimension a parameter.
Readiness assessment
Each asset ships as a Sand3D program and a mesh, and every part carries a semantic name and a named material (painted_metal, steel, walnut, and so on), which places the library at L2 semantic material. Over 10K articulated objects additionally declare their joints with axis and limits. Meshes export to GLB / URDF / USDA / OBJ / STL, so they feed Omniverse, Isaac and web viewers. Mass, inertia and friction are user-defined in code rather than auto-estimated, so the bulk does not reach L3.
Practical notes
The collection aggregates five Hugging Face datasets — sand3d-objaverse (67K), sand3d-cad (24K), sand3d-articraft (9.2K), sand3d-simready (2.3K, a rewrite of NVIDIA SimReady) and sand3d-automate (3.1K) — all under CC BY 4.0, with the library itself Apache-2.0. Token cost is roughly 4x lower than raw bpy for the same part. Note the landing-page scale figures conflict (78K vs 100K+), so treat 100K as approximate.
Disclaimer: license and usage terms follow the official source — verify them yourself before use. This site indexes and links only; it does not host or mirror any asset file.
License and usage terms are those of the official source — verify before use. No asset file is hosted or mirrored on this site.