sand3d-articraft

9 209 3D objects written as code. Each object is a short Python program in sand3d, the OpenUSD file that program exports, and a render of it. Everyday objects, articulated mechanisms and mechanical CAD parts, in one language: every part is named, every dimension a parameter, and articulated objects declare their joints.

source	objects	articulated	joints
Articraft	9 209	9 139	33 444

Every .usda opens in the Hub's 3D viewer; an articulated object's file carries its joints. One per source:

• Articraft: usda/0002/articraft-rec_a-compact-on-camera-flash-with-a-narrow-body-a-r_20260426_045857_759417_bff0fdec.usda

What a row holds

The Parquet files under data/ hold one row per object; the Dataset Viewer shows them.

column
id	the object's id in this dataset
program	its sand3d program
entry	the program's function that builds the object; null when the program builds it as it runs
render	one picture of the object, as the program builds it
usda_path	its OpenUSD file in this repository: metres, Z up, materials; an articulated object's joints as UsdPhysics
n_joints	its joints, 0 for a rigid object
f_score	how close the program's shape is to the source object's (Quality)
family	a coarse category (furniture, hand tool, container…)
split	train or val, the train and validation splits (Splits)
source, source_group, asset_id	where the object comes from, and its id there
license, source_license, attribution, license_status	Licence
sand3d_version	the sand3d version the program was run under to write the file
usda_bytes, usda_sha256, program_sha256	the files' size and digests

Load it

One environment holds both the loading and the rebuilding below, on Python 3.13:

pip install "bpy==5.2.*" sand3d datasets huggingface_hub

from datasets import load_dataset
from huggingface_hub import hf_hub_download

ds = load_dataset("Sandflow/sand3d-articraft", split="train")
row = ds[0]
print(row["program"])     # the sand3d program
row["render"]             # a PIL image
usda = hf_hub_download("Sandflow/sand3d-articraft", row["usda_path"], repo_type="dataset")

Rebuild an object with sand3d (Apache-2.0; it runs on Blender's bpy):

import bpy
import sand3d as s3d

bpy.ops.wm.read_factory_settings(use_empty=True)      # an empty scene
ns = {}
exec(row["program"], ns)
if row["entry"]:                                      # the function that builds the object
    obj = ns[row["entry"]]()
else:                                                 # the program built it as it ran
    roots = [o for o in bpy.context.scene.objects if o.parent is None]
    obj = roots[0] if len(roots) == 1 else s3d.group("object", *roots)
s3d.export(obj, "object.usda")                        # or .glb, .urdf

Every file here was written that way, under sand3d 1.5.0 (a pre-release) and bpy 4.5: an object whose program did not run, or did not pass a static check, is not in the dataset. The programs rebuild with sand3d ≥ 1.6 on bpy 5.2 into the same objects, though not always into byte-identical files.

Splits

train and val are a tag on every row, and both are released. Objaverse objects are split by Sketchfab author and SimReady assets by NVIDIA product, within each family, so that no author or product sits on both sides; about 3 % of each family is in val, and every family is on both sides. CAD parts and Articraft objects are split by whole groups: for CAD, held-out shape families, for Articraft, 12 held-out categories.

source	train	val
Articraft	8 737	472

Quality: f_score

f_score is the F-score between the surface the program builds and the source object's, at a distance of 1 % of the object's bounding-box diagonal. It is the pipeline's own measurement of its output, taken when the program was written, not an independent judge, and nothing here re-measures it. An articulated object is scored whole, at rest; each of its moving parts and joints was checked too. The floors kept: 0.70 for Objaverse and the static SimReady assets (scored against the whole mesh, where a lower score is often a program that leaves out small details), 0.90 for CAD parts and Articraft's rigid programs, and for articulated objects every joint right and every part at 0.90.

How the programs were written

The programs come from Sandflow's iterative agentic reconstruction pipeline: an agent writes a sand3d program, runs it, compares what it builds with the source object, and revises the program until it rebuilds the object. The pipeline itself will be described in upcoming releases, not in this dataset. The Articraft and Zero-To-CAD source objects were themselves generated by their authors' own agentic systems, as their papers describe.

Licence

The sand3d programs, renders and USD files in this dataset are released under CC-BY-4.0. Each record is derived from a third-party object; to the extent that it is an adaptation of that object, the object's own licence also applies, and its attribution must be kept: see the source, source_license and attribution columns. Every source is under CC-BY-4.0, Apache-2.0 or CC0; none is non-commercial or share-alike. The sand3d library that runs these programs is released separately, under Apache-2.0 (https://github.com/Sandflow-ai/sand3d).

Every row carries both layers: license (ours, CC-BY-4.0) and source_license (the object's). attribution names the object, its author where there is one, its licence, and that it was modified.

source	objects	source licence	what reuse must keep
Articraft	9 209	CC-BY-4.0	the attribution "Articraft Data, Matt Zhou et al." and the note that it was modified

Removal requests

If you hold rights in an object in this dataset and want it removed, write to thibault.neveu@sandflow.ai with its asset_id.

Citation

If you use this dataset or the sand3d library, please cite:

@misc{sand3d2026,
  author = {Neveu, Thibault and Loy, Andr{\'e}a},
  title  = {sand3d},
  year   = {2026},
  note   = {Version 1.6.2},
  url    = {https://github.com/Sandflow-ai/sand3d}
}

Please also cite the sources its objects come from.

@misc{zhou2026articraftagenticscalablearticulated,
      title={Articraft: An Agentic System for Scalable Articulated 3D Asset Generation},
      author={Matt Zhou and Ruining Li and Xiaoyang Lyu and Zhaomou Song and Zhening Huang and Chuanxia Zheng and Christian Rupprecht and Andrea Vedaldi and Shangzhe Wu},
      year={2026},
      eprint={2605.15187},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.15187},
}