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: >-
The dataset-authored curation, metadata, converted representations, and
packaging are licensed under CC BY-NC 4.0 for non-commercial research use.
Underlying geometry, textures, scene components, and runtime dependencies
retain their applicable upstream terms; see provenance/third_party_notices.md
and provenance/scene_third_party_notices.md. By requesting access you agree to
non-commercial research use and to comply with all applicable upstream license
and attribution requirements.
extra_gated_button_content: Agree and request access
SpatialAct-Bench-Assets (SPAB)
SpatialAct-Bench-Assets is the object-and-scene release for SpatialAct-Bench, a target-centric benchmark for embodied spatial reasoning and manipulation. It combines a curated object library, portable tabletop scene bundles, and machine-readable benchmark metadata so that environment construction, target selection, and task evaluation can be reproduced across machines.
1. Dataset Overview
At a glance
- 701 movable 3D object assets spanning 66 basic categories and 80 canonical classes.
- 32 selected Infinigen-derived tabletop scenes packaged for Franka manipulation in Isaac Sim.
- One representative preview per object.
- Per-object semantic, geometry, appearance, benchmark-task, provenance, and graspability metadata.
- Manually gated research access on Hugging Face.
Repository structure
README.md
benchmark_32_scenes.json scene selection and validation metadata
spab_objects_v1.json canonical metadata for 701 objects
data/
train-00000.parquet tabular object metadata
assets/
SPAB-OBJ-000001/
asset.usdc object USD entry point
textures/ local textures, when required
...
previews/
SPAB-OBJ-000001.jpg one representative preview per object
metadata.jsonl
scene/
seed_45/
scene.usdc portable scene entry point
textures/ scene-local textures
materials/ scene-local custom materials
...
provenance/
third_party_notices.md
scene_third_party_notices.md
LICENSE
SCHEMA.md
files.sha256
tools/
verify_release.py
The complete assets/SPAB-OBJ-xxxxxx/ directory is the relocation unit
for an object. The complete scene/seed_<seed>/ directory is the
relocation unit for a scene. Do not move an entry-point USDC file without
its accompanying local resource directories.
2. 3D Object Asset Library
Content and taxonomy
The object library contains 701 stable public object IDs of the form
SPAB-OBJ-xxxxxx. Its source composition is:
| Source | Assets |
|---|---|
| RoboCasa | 409 |
| OmniObject3D | 229 |
| Google Scanned Objects | 13 |
| GraspXL | 20 |
| Direct Objaverse selection | 2 |
| Legacy pool | 28 |
| Total | 701 |
The legacy pool contains 21 Objaverse-derived objects and 7 YCB objects.
See provenance/third_party_notices.md
for source-level attribution and license details.
Each record in spab_objects_v1.json provides:
- a stable object ID and relative USD path;
- basic category, canonical class, subtype, and natural-language caption;
- physical size, plausible mass range, and the spawn-time uniform scale;
- material/color information and a representative preview;
- eligibility and accepted answers for category, subtype, color, and pointing tasks; and
- an empirically measured graspability tier.
Category QA accepts either the basic category or the canonical class when
the two differ. Subtype QA requires an exact subtype match. Color answers
use a closed eight-color vocabulary plus multicolor. See
SCHEMA.md for the complete field definitions.
Portable USD packages
Every object contains a binary asset.usdc entry point and all required
local texture files. Of the 701 objects, 677 use packaged texture files
and 24 require no texture files. The packages contain no external USD
layer references or machine-specific absolute paths.
Material shaders may reference Isaac Sim/Omniverse built-in MDL modules,
including OmniPBR.mdl and, for some sources, gltf/pbr.mdl. These modules
are runtime dependencies supplied by Isaac Sim and are not redistributed
with this dataset. Geometry, UVs, and physics schemas remain readable in a
generic USD viewer, but materials may fall back when its MDL runtime is not
available.
Apply geometry.scale from spab_objects_v1.json as a uniform
xformOp:scale at spawn time, before physics initialization. This value is
not necessarily baked into the object USDC.
Graspability tiers
Each object was evaluated in five fixed tabletop layouts, one trial per
layout and five trials per object (3,505 trials total). The grasp
field is a difficulty tier derived from the number of successful trials,
not a policy-independent physical property.
| Successful trials | Tier | Assets | Share |
|---|---|---|---|
| 0 / 5 | low |
182 | 26.0% |
| 1–2 / 5 | mid |
440 | 62.8% |
| 3–5 / 5 | high |
79 | 11.3% |
3. Scene Library
Selection
The scene release contains 32 tabletop environments selected from a
40-scene candidate pool. Every retained scene achieved at least 2
successes in 6 trials during the common screening protocol. Six scenes
also carry the expanded_eval_validated status in the manifest.
Included seeds:
45, 86, 152, 182, 196, 794, 813, 7777,
7801, 7812, 7845, 7897, 7909, 7927, 7944, 7983,
7986, 7996, 8001, 8017, 8062, 8064, 8121, 8168,
8178, 8188, 8192, 8383, 8393, 8402, 8415, 8473
benchmark_32_scenes.json is the authoritative scene index. For each
scene it records the benchmark scene ID, generation seed, portable path,
screening result, and validation status. Screening success rates describe
one fixed evaluation protocol and should be treated as environment
difficulty signals, not model-independent scene scores.
Packaging and portability
Each scene/seed_<seed>/ bundle contains:
scene.usdc binary USD scene entry point
textures/ local scene textures
materials/ local custom MDL materials
The scene entry point has no machine-specific absolute paths and no active Replicator or post-processing runtime graph. Scene-local dependencies are relative to the bundle directory. Isaac Sim's built-in shader library is still required for standard Omniverse MDL modules.
The release validation covers two complementary layers:
- Static package validation: binary-USDC structure, absence of
symlinks and absolute/remote asset paths, removal of runtime-only
/Replicatorand/Rendergraphs, and resolution of every packaged local dependency. - Isaac Sim validation: standalone loading, required camera checks, material resolution, and non-constant 640×360 renders from both the external and wrist cameras.
Move or download the entire seed_<seed> directory to preserve material
and texture resolution.
4. Other
Download
This repository is manually gated. Sign in to Hugging Face, accept the terms, and wait for author approval before downloading files.
hf auth login
# Download the complete dataset.
hf download nerako/SpatialAct-Bench-Assets \
--repo-type dataset \
--local-dir SpatialAct-Bench-Assets
# Or download only the scene release and its manifest.
hf download nerako/SpatialAct-Bench-Assets \
scene/ benchmark_32_scenes.json \
--repo-type dataset \
--local-dir SpatialAct-Bench-Scenes
For the object release, run python3 tools/verify_release.py after a full
download to perform the provided integrity checks.
Access status
The dataset page is publicly discoverable, but file access is manually gated. Access requests are reviewed by the authors and must be approved before download. Gating does not replace compliance with the license and third-party terms.
License and third-party terms
The dataset-authored selection, normalization, converted representations,
and metadata are released for non-commercial research under CC BY-NC
4.0 as described in LICENSE.
Underlying geometry, textures, and other third-party components remain
subject to their applicable upstream licenses and attribution
requirements. Inclusion in this repository does not relicense those
components or grant rights beyond their original terms. Object-level
source information is documented in
provenance/third_party_notices.md.
Scene bundles and their embedded/runtime dependencies must likewise be
used in accordance with the terms of their upstream scene-generation and
simulation components; see
provenance/scene_third_party_notices.md.
Review the applicable upstream terms before use or redistribution.
Citation
A formal citation will be added with the accompanying SpatialAct-Bench paper.
Changelog
- v1.1 (2026-07-31): Added the 32-scene portable benchmark release and scene manifest.
- v1.0 (2026-07-21): Initial 701-object release with unified five-scene graspability evaluation.