Hugging Face's logo

Datasets:
nerako
/
SpatialAct-Bench-Assets

Modalities:
Image
Text
3D
Formats:
parquet
Size:
< 1K
Tags:
robotics
3d
manipulation
benchmark
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
SpatialAct-Bench-Assets / README.md
nerako's picture
nerako
Point viewer config at parquet for image thumbnails
ad4dfdb verified
preview code
|
Raw
History Blame Contribute Delete
7.3 kB
metadata 
license: other
license_name: cc-by-nc-4.0
license_link: LICENSE
pretty_name: SpatialAct-Bench-Assets
tags:
  - robotics
  - 3d
  - manipulation
  - benchmark
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.parquet
extra_gated_heading: Request access to SpatialAct-Bench-Assets
extra_gated_description: >-
  This is an early research release. Access requests are reviewed manually by
  the authors.
extra_gated_prompt: >-
  This dataset is licensed under CC BY-NC 4.0 for non-commercial research use
  only. It is a derivative work built from RoboCasa, OmniObject3D, Google
  Scanned Objects, GraspXL, Objaverse, and the YCB Object and Model Set — see
  provenance/third_party_notices.md for the complete per-source attribution
  required by each original license. By requesting access you agree to use this
  dataset for non-commercial research purposes only and to comply with all
  license terms referenced above.
extra_gated_button_content: Agree and request access

SpatialAct-Bench-Assets (SPAB)

A curated, de-identified 3D object asset library for the SpatialAct-Bench target-centric "acknowledge" benchmark (built for pi0.5 / DROID-style manipulation policies: grasp a target object, then answer questions about its category, subtype, or color).

At a glance

  • 701 assets, 66 basic categories, 80 canonical (detailed) classes.
  • Every asset ships as one self-contained USD entry point (assets/SPAB-OBJ-xxxxxx/asset.usdc + textures/) with no external dependencies.
  • One representative preview image per asset (previews/).
  • Every asset has a physical graspability tier (low / mid / high, see below) from a unified evaluation.
  • Single license for the whole package: CC BY-NC 4.0.

Access status

This repository is currently private, shared with a small set of collaborating labs for internal use and feedback. It is planned to later switch to gated public access (same repository, same content — no rebuild) once the accompanying paper is ready. Feedback and bug reports from this private phase will be folded in before that switch.

License

The whole package (asset selection, deduplication, rescaling, USD conversion, and all authored metadata) is licensed under CC BY-NC 4.0 — see LICENSE.

The underlying geometry and textures of individual assets originate from third-party open datasets (RoboCasa, OmniObject3D, Google Scanned Objects, GraspXL, Objaverse, and the YCB Object and Model Set) and remain attributed to their original creators under their own licenses. See provenance/third_party_notices.md for the complete, per-source breakdown and required attribution statements. In short: 651 of 701 assets trace to CC BY 4.0 sources, 20 trace to CC BY-NC 4.0 (GraspXL), and 30 trace to an Objaverse-linked legacy pool whose per-object upstream terms are documented but not all independently re-verified — see the notices file for exact per-source figures and caveats.

Grasp quality tiers

Each asset was evaluated with a fixed manipulation policy across 5 independent tabletop layout scenes (one trial per scene, 5 trials per asset, 3,505 trials total over all 701 assets), and camera-validated after the run. The grasp field in spab_objects_v1.json is a tier derived from the count of successful trials out of 5, not a raw success percentage:

Successes / 5 Tier Assets Share
0 low 182 26.0%
1–2 mid 440 62.8%
3–4 high 79 11.3%

This is a measurement from one evaluation round with one checkpoint and one fixed scene protocol — it reflects physical graspability under that setup, not an immutable property of the object, and should be treated as a difficulty signal rather than a benchmark ground truth.

Dependency scope: what's self-contained and what isn't

Every asset ships as exactly two things: asset.usdc (one USD crate file holding mesh, UVs, material bindings, and physics API schemas — rigid body, mass, collision — all in one layer) and its textures/ folder. This is a deliberate, verified-complete package, not a partial one: across all 701 assets there are zero external-layer references and zero absolute paths pointing outside each asset's own directory (every reference is relative and resolves correctly after moving or re-downloading the package anywhere).

The one intentional exception: material shaders reference OmniPBR.mdl (or gltf/pbr.mdl for some sources) by bare name, which Isaac Sim / NVIDIA Omniverse resolve from their own built-in shader library at runtime — the same way a PDF viewer supplies a font it wasn't shipped with. These are not packaged with this dataset. In Isaac Sim, materials render correctly. In a generic USD viewer without that shader library, geometry, UVs, and physics schemas will still load correctly, but the material may fall back to a default/untextured look.

Layout 

spab_objects_v1.json        canonical per-asset metadata (see SCHEMA.md)
assets/SPAB-OBJ-xxxxxx/     one self-contained USD asset per directory
  asset.usdc                 entry point (crate binary), no external refs
  textures/                  local textures referenced by asset.usdc
previews/SPAB-OBJ-xxxxxx.jpg  one representative image per asset
previews/metadata.jsonl      per-image metadata for automatic dataset preview
provenance/third_party_notices.md   required third-party attribution
LICENSE                      CC BY-NC 4.0 notice for this package
SCHEMA.md                    full field-by-field schema reference
files.sha256                 checksum manifest for every packaged file
tools/verify_release.py      standard-library integrity verifier

Download and verify 

# download (private repo — requires an invited HF account + token)
huggingface-cli download <org>/SpatialAct-Bench-Assets \
  --repo-type dataset --local-dir spab_release_v1

# verify checksums and structural integrity
cd spab_release_v1
python3 tools/verify_release.py

Isaac Sim smoke test 

from pxr import Usd, UsdGeom
stage = Usd.Stage.Open("assets/SPAB-OBJ-000001/asset.usdc")
assert stage.GetDefaultPrim().IsValid()
bbox_cache = UsdGeom.BBoxCache(Usd.TimeCode.Default(), [UsdGeom.Tokens.default_])
print(bbox_cache.ComputeWorldBound(stage.GetPseudoRoot()).ComputeAlignedRange())

Remember to apply geometry.scale from spab_objects_v1.json via an xformOp:scale at spawn time, before physics initialization — it is not baked into the USD file (see SCHEMA.md).

Schema

See SCHEMA.md for the complete field reference and an example record. Two scoring/vocabulary rules apply dataset-wide (see taxonomy_policy in spab_objects_v1.json): category QA accepts either the basic category or the detailed canonical_class as correct, while subtype QA requires an exact subtype match; all color labels are drawn from a fixed 8-word palette plus multicolor.

Citation

TBD — will be added once the accompanying paper has a citable reference.

Changelog

  • v1.0 (2026-07-21): Initial private release. 701 assets, unified 5-scene grasp-tier evaluation, de-identified schema, single CC BY-NC 4.0 license.