Buckets:
| Name | Size | Uploaded | Xet hash |
|---|---|---|---|
| dataset_chunks | 3 items | ||
| dataset_chunks_v1 | 3 items | ||
| dataset_chunks_v1.1 | 2 items | ||
| .gitattributes | 2.63 kB xet | 1c19e241 | |
| .gitignore | 6 Bytes xet | c69bc03b | |
| README.md | 6.82 kB xet | e228bc0e | |
| pack_dataset_chunks.py | 27.8 kB xet | a3cef49f | |
| releases.json | 3.29 kB xet | 2c30b190 | |
| split.json | 329 kB xet | 586ec703 |
Artiverse: A Diverse and Physically Grounded Dataset for Articulated Objects
Disclaimer: this repository contains data that is still under development and is continuously being verified and improved. We are committed to making it as clean and useful as possible for the community. If you find any issues with the data or have any feedback/requests, please file an issue on Hugging Face or the data viewer interface, or send us an email. Your feedback is an important part of the community effort to bring more high-quality articulated objects data to everyone. Thank you!
This repo contains a subset of pre-release Artiverse assets. We are actively working on finalizing the data. We will soon release the remaining models and then the cleanup pass resolving all the remaining problems. The data viewer is available online.
Our data structure is organized by semantic category, source, and model id:
data/
{category}/
{source}/
{model_id}/
{model_id}.segmented.glb
{model_id}.articulations.json
material.json
imgs/
urdf_w_collider/
Where:
{category}is the object class, for examplearmoire,bookcase,chest_of_drawers,desk,sideboard,tv_stand, orwall_cabinet.{source}is the upstream asset source:fpModel,3dw,3dfModel,objaverse_xl,texverse,objaverse,wss,3dc200,partnextorabo.{model_id}is the unique model identifier within that source.- Note - this structure will be updated in the full release.
Each model folder contains:
{model_id}.segmented.glb: segmented geometry with part structure and metadata used by our processing pipeline.{model_id}.articulations.json: articulation annotations, including articulated part ids, joint types, axes, and motion ranges.material.json: inferred material assignment and density per part, keyed by part id. Parts are the fine-grained segmentation; entries also cover every part the URDF links to, so each link has a material for mass computation.imgs/: reference renderings for the object.urdf_w_collider/: URDF export with visual meshes and collision geometry for simulation use.
Splits
split.json gives the train / val / test assignment for the released models, keyed as
{category}/{source}-{model_id}:
{"train": ["aerosol_can/3dw-256cab538afb0e2c24c1dae961c37e57", ...],
"val": [...],
"test": [...]}
Versions
Artiverse is published as versioned archive sets. Older versions stay available and unchanged, so results produced against them remain reproducible.
| Version | Kind | Contents | Categories | Download |
|---|---|---|---|---|
v0.1 |
pre-release | 3,544 models | 84 | 65.3 GB, 2 chunks |
v1 |
release | +1,858 models (total 5,402) | 88 | 47.5 GB, 2 chunks |
v1.1 |
patch | 683 entries fixing v0.1 content | 53 | 6.1 GB, 1 chunk |
Every version extracts into the same data/ tree, applied in registry order
(v0.1 → v1 → v1.1).
v0.1 — the original pre-release. 3,544 models across 84 categories.
v1 — new models. Adds 1,858 model folders that were not in v0.1, bringing the dataset to
5,402 models across 88 categories. It is purely additive: no folder shipped in v0.1 is touched.
v1.1 — corrections to v0.1. Adds no new models; it repairs content that shipped in v0.1.
What v1.1 fixes
- Materials. Every part now has exactly one material entry, and some implausible assignments were corrected.
- Interior geometry. Regenerated for armoires that were missing it.
- Kinematic graphs. Models whose URDF did not form a single rooted tree now export correctly.
Reconstructing data/
# what is available
python pack_dataset_chunks.py versions
Starting from nothing — unpack builds the tree from the base release up to the version you name (defaults to latest).
# everything, up to the newest release (recommended)
python pack_dataset_chunks.py unpack
# same, verifying checksums as it goes
python pack_dataset_chunks.py unpack --verify-sha256
# pin an older state: base + v1, without the v1.1 corrections
python pack_dataset_chunks.py unpack --version v1
# just the original pre-release
python pack_dataset_chunks.py unpack --version v0.1
Already have the data — upgrade in place instead of re-downloading everything. Nothing on disk records which version you have, so you tell it:
# one step up: apply v1.1 to a tree that is already at v1
python pack_dataset_chunks.py unpack --version v1.1 --patch
# you are on v0.1 and want to be current: applies v1 then v1.1
python pack_dataset_chunks.py unpack --from v0.1
# you are on v0.1 and want to stop at v1
python pack_dataset_chunks.py unpack --from v0.1 --version v1
--patch and --from never clear data/ — they merge onto the tree they are upgrading, and
--overwrite is ignored for them. Only a from-scratch build may clear data/ first. Applying a
version you already have is harmless: it rewrites the same files.
| You want | Command |
|---|---|
| Everything, from scratch | unpack |
| Everything up to version X | unpack --version X |
| One version bump onto existing data | unpack --version X --patch |
| Upgrade existing data from A to newest | unpack --from A |
| Upgrade existing data from A to B | unpack --from A --version B |
pygltftoolkit can be used to load the GLBs and parse the segmentation annotations. The annotation formats largely follow S2O, so please check out the documentation.
- Total size
- 119 GB
- Files
- 14
- Last updated
- Sep 10
- Pre-warmed CDN
- US EU US EU