minsj1225's picture
PartNeXt material and simulation-ready physical property annotations
2f20604
|
Raw History Blame Contribute Delete
25 kB
metadata
license:
  - cc-by-4.0
  - cc-by-sa-4.0
  - odbl
  - other
license_name: mixed-corpus
license_details: >-
  Annotations are CC BY 4.0, but individual property values inherit the licence
  of the corpus sources they were measured from, and those are not all
  compatible. Every row carries a `license_tier` column giving the strictest
  licence that applies to it. Filter on that column before redistributing or
  using commercially.
task_categories:
  - graph-ml
tags:
  - 3d
  - material-properties
  - physics-simulation
  - part-segmentation
  - simulation-ready
  - genesis
  - isaac-gym
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files: PartNeXt_material_annotations.csv

PartNeXt Material & Simulation-Ready Physical Properties

Per-part material class and simulation-ready physical properties $(\nu, E, \rho)$ for 333,152 parts across 23,221 textured 3D models of PartNeXt (NeurIPS D&B 2025).

Every property value carries per-quantity provenance back to a cited measurement source, and ships with metadata that Genesis and Isaac Gym can consume directly.

PartNeXt itself ships no material or physical-property labels — these annotations are new.


Quick start

pip install -U pip                       # older pip may fail to find a pyarrow wheel
pip install huggingface_hub datasets pandas pyarrow

hf download minsj1225/PartNeXt-Material-Properties \
    --repo-type dataset --local-dir ./partnext_mat
# on huggingface_hub < 1.0 the command is `huggingface-cli download` with the same arguments

cd partnext_mat
# prove the join against the official masks (no mesh download needed, ~3 min)
python3 join_with_partnext.py verify

Both scripts ship with a test that runs them against a 10-model subset picked to hit every edge case in the release — bare-int face indices, ? part names, abstained parts, a 4,747-part model, a non-LFS GLB — downloads those meshes for real, and checks the face indices against the actual geometry (50 checks, no HF token needed):

pip install trimesh pyarrow
python3 test_scripts.py     # 64 MB of downloads; 50 checks
annotations   333,152 parts / 23,221 models
AuWang/PartNeXt 23,519 models

model_id      ours-not-upstream 0   upstream-not-annotated 298
type_id       mismatches 0 / 23,221
part_id       333,152 / 333,152 resolve to a masks key  (100.0000%)
              models with an unresolvable part_id: 0

coverage      333,152 / 347,434 parts of the annotated models (95.9%)
              333,152 / 350,145 parts of all PartNeXt       (95.1%)

JOIN IS EXACT

How every row maps onto a PartNeXt asset

Three identifiers, all of which come straight from upstream:

Column Links to
model_id a row of AuWang/PartNeXt, and the GLB basename
type_id the bucket directory, so the mesh is glbs/{type_id}/{model_id}.glb in AuWang/PartNeXt_mesh
part_id a key of the official masks dict for that model_id

masks[part_id] is {mesh_id: [face_indices]} — the exact triangles of that part inside the GLB. So a row of this dataset says: these faces, of this mesh, in this file, are carbon_steel, with $\rho$ = 7850 kg/m³, $E$ = 203 GPa, $\nu$ = 0.29.

part_name and part_path are the leaf name and ancestor chain of the matching maskId node in the official hierarchyList, carried over verbatim.

Verified against upstream (re-runnable with join_with_partnext.py verify)

Check Result
part_id resolves to an official masks key 333,152 / 333,152 = 100.0000%
Models present in AuWang/PartNeXt 23,221 / 23,221 (none missing)
type_id agrees with upstream 0 mismatches / 23,221
GLB files present in AuWang/PartNeXt_mesh 23,221 / 23,221 (none missing)
Parts resolving to a non-empty face set 333,152 / 333,152, 367,486,735 faces total

Two ways to read a part's geometry

import pandas as pd, json
from datasets import load_dataset

ann = pd.read_csv(
    "hf://datasets/minsj1225/PartNeXt-Material-Properties/PartNeXt_material_annotations.csv",
    dtype={"model_id": str, "type_id": str, "part_id": str})   # keep these as str!

pnx   = load_dataset("AuWang/PartNeXt", split="train")
masks = dict(zip(pnx["model_id"], pnx["masks"]))

row   = ann.iloc[0]
faces = json.loads(masks[row.model_id])[row.part_id]   # {mesh_id: [face indices]}
glb   = f"glbs/{row.type_id}/{row.model_id}.glb"       # in AuWang/PartNeXt_mesh

dtype=str is not optional. type_id carries leading zeros ("000-029") and part_id is a string key upstream. Letting pandas infer types silently breaks the join.

Or let the bundled script do it — it streams the masks instead of materialising all 350k of them (which costs several GB of Python objects):

python3 join_with_partnext.py table  --out joined.parquet   # 1 row/part + glb_path + n_faces
python3 join_with_partnext.py export --out sim_json/        # 1 JSON/model, E already in Pa
python3 join_with_partnext.py show   --model-id 298ed9ae6f9b41dd8ee0c0c5b412f343
model 298ed9ae6f9b41dd8ee0c0c5b412f343   category Chair
GLB   AuWang/PartNeXt_mesh/glbs/000-029/298ed9ae6f9b41dd8ee0c0c5b412f343.glb
plan  rigid_soft_coupling   stiffness ratio 1691666.7

part_id part_name                  material   bucket                     rho     E[GPa]    nu   faces
-----------------------------------------------------------------------------------------------------
      0 Armrest Horizontal Rod     metal      carbon_steel              7850  203.00000  0.29   1,502
      2 Back Cushion               fabric     woven_textile__membran     307    0.00012  0.36  10,553
      6 Leg                        metal      carbon_steel              7850  203.00000  0.29     246
     10 Foot Pad                   rubber     -                         1015    0.00065  0.49     187

Meshes

python3 download_partnext_meshes.py --out ./partnext_mesh --workers 6 --token hf_xxx
python3 download_partnext_meshes.py --out ./partnext_mesh --check   # re-hash what is on disk

Files land at partnext_mesh/glbs/{type_id}/{model_id}.glb, mirroring the upstream repo — the same relative path that join_with_partnext.py export records, so the two line up.

Downloads exactly the 23,221 GLBs referenced by this release (55.10 GB), driven by manifest_models.csv, which carries the expected glb_bytes and glb_sha256 of every file. Each file lands as <name>.part and is renamed only once its size matches, so an interrupted run resumes correctly instead of leaving a truncated file that looks complete. --check re-hashes what is already on disk; --verify does HEAD checks only. A token is strongly recommended — anonymous requests get rate-limited (HTTP 429).


Scale

Annotated parts 333,152
…of the parts of the models we annotate 333,152 / 347,434 = 95.9%
…of all PartNeXt parts 333,152 / 350,145 = 95.1%
Models 23,221 (of 23,519 in AuWang/PartNeXt; 23,229 have a GLB)
Object categories 50 / 50
Distinct part names 451
All three properties present 329,289 (98.84%)
Abstained 3,863 (1.16%) — 3,701 sanitary whiteware with no traceable grade, plus 162 parts whose bucket is withheld (see cfrp below)
Objects with no property at all 624 (1,510 parts) — every part of these models abstained, so they cannot be loaded into a simulator as-is

16,993 PartNeXt parts carry no annotation. Most are degenerate upstream records, but 4,426 are a genuine coverage gap — see limitation 6:

parts
in the 298 models with no annotation at all 2,711 290 of those models have no GLB in AuWang/PartNeXt_mesh, so no visual or geometric evidence exists
geometrically degenerate (≤ 2 faces) 9,844 a 1–2 triangle sliver carries no simulable volume
maskId with no hierarchyList node 12 no part name to condition on
named, real geometry, still skipped 4,426 a pipeline gap, not a principled exclusion

Columns

Identity model_id type_id part_id part_name part_path object_category

Material material_class confidence bucket bucket_basis

material_class is drawn from a 17-class vocabulary; 10 classes occur in this release: metal plastic fabric wood glass rubber ceramic foam leather composite.

Properties density_kg_m3 youngs_modulus_GPa poissons_ratio, each with _lo / _hi.

interval_kind says what the interval means, and governs whether bucket is populated:

interval_kind bucket n Meaning
bucket_IQR set 247,490 a specific grade bucket was resolved; interval is its P25–P75
reference_range set 586 no corpus grade matches this material; the value and interval are a published handbook range (see basis_note)
class_range_p5_p95 null 81,213 no discriminative evidence — class-level estimate, treat as a range
null null 3,863 abstained; all property and sim_* columns are null

Provenance {rho,E,nu}_provenance ∈ {measured, reference_value, class_imputed, not_retrievable}; {rho,E,nu}_source_ids → source_id of material_corpus_sources.csv (82 sources, 42 used). license_tier gives the strictest licence among the sources that produced a row's values — see the Licence section before redistributing.

reference_value (5,918 quantities) marks a number taken from a published handbook range rather than retrieved from the corpus, used where the corpus holds nothing representative for that material. It appears in two forms. Where the corpus does characterise the bucket's other two quantities, source_ids lists those sources — this is the case for nu in engineering_rubber and polymer_foam__soft. Where the corpus holds no grade for the material at all — the three reference_range buckets — source_ids is empty, because claiming corpus support for a number the corpus does not contain would be false. Each bucket's basis_note in buckets_properties.json names the reference.

Materials with no corpus grade. Three buckets carry interval_kind = reference_range: their value and interval are a published handbook range because the 82-source corpus contains no grade for that material.

Bucket Parts rho E nu What it covers
polyurethane_wheel 325 1180 40 MPa 0.48 cast polyurethane caster and skateboard wheels, Shore A 85–95
flexible_pu_foam__upholstery 169 40 40 kPa 0.30 open-cell flexible upholstery foam — mattresses, seat and back cushions, ear pads
cable_jacket__plasticized_pvc 92 1350 25 MPa 0.40 plasticized PVC cable and wire jacket compound (connectors and plugs are rigid moulded plastic and sit elsewhere)
copper_alloy 5,213 8600 128 GPa 0.27 copper and copper-alloy parts. The eight corpus records all give 8600 kg m⁻³ and nu spans 0.27–0.33; pure copper is 8960 kg m⁻³ at nu 0.34, so the bucket is named for the family, not the pure metal

A source id may carry a proxy:: prefix (e.g. proxy::nist_ceramics_elasticity, 1,392 rows). That means the value is taken from a closely related grade in that source rather than the exact material — strip the prefix to look the source up. Used only where no exact grade exists in the corpus (whiteware).

The cfrp bucket is withheld. Its density rested on one record (n = 1) carrying the density of an epoxy matrix (1200 kg m⁻³) rather than of a laminate, which is 1550–1600 kg m⁻³; no corpus source supplies a laminate density and none can be derived from corpus values, so the quantities are withheld rather than substituted. The 162 parts routed there keep material_class = composite and carry a sim_caveat. Most were a poor match in any case — 104 are skateboard Grip Tape and 7 Rest of Deck, which are abrasive-coated paper and maple ply.

Simulation

Column Meaning
sim_body_type rigid (E ≥ 1 GPa) / semi_rigid (0.1–1 GPa) / deformable (1–100 MPa) / soft (< 1 MPa)
sim_shear_modulus_GPa $G = E/(2(1+\nu))$
sim_bulk_modulus_GPa $K = E/(3(1-2\nu))$
sim_lame_lambda_GPa $\lambda = E\nu/((1+\nu)(1-2\nu))$
sim_p_wave_speed_m_s $c = \sqrt{(\lambda+2\mu)/\rho}$
sim_cfl_dt_us_at_5mm explicit time step $dt = dx/c$ at dx = 5 mm
sim_self_weight_strain $\varepsilon = \rho g L/E$, with $L$ mapped to real-world object size
sim_object_solver_plan per object: rigid_soft_coupling / implicit_or_substep / single_domain
sim_object_stiffness_ratio $E_{max}/E_{min}$ within the object
sim_caveat dimensional-basis, near-incompressibility or withheld-bucket warning (46,907 parts — read the limitations)

Units are GPa for moduli and kg/m³ for density. join_with_partnext.py export converts $E$ to Pa so the JSON drops straight into a solver.

Other files

File What it is
manifest_models.csv 23,221 models: glb_relpath, glb_bytes, glb_sha256, part count, solver plan, stiffness ratio
buckets_properties.json all 84 property buckets (82 flagged usable, 2 withheld, 23 referenced here, 22 supplying values) with per-quantity value, P25/P75, source ids and row counts
material_corpus_sources.csv the 82 corpus sources with citation, DOI/URL, license, redistribution clearance
pipeline_config.json every fusion weight and threshold used to produce this release
join_with_partnext.py verify / table / export / show
download_partnext_meshes.py fetch exactly the annotated GLBs, with checksum resume
test_scripts.py end-to-end test of both scripts on a 10-model edge-case subset

Quality

Material classification

Weak labels come from 7 regex rules over unambiguous part names (Blade→metal, Keycap→plastic, Shelf Board→wood …). Parts are split into calibration / tuning / report thirds by model_id hash; during evaluation the leaf-name prior of those parts is masked so the semantic channel cannot read off the answer.

micro accuracy (report third, n=37,888) 92.8%
macro accuracy 83.1%
majority baseline 31.5%
split spread ≤0.4 pts

Per class: glass 97.0 · plastic 96.2 · fabric 95.4 · wood 93.7 · metal 85.7 · rubber 63.9 · ceramic 50.0

Plausibility (GT-free, four hard constraints)

Check Score
P1 within-object material composition 99.98% (76 / 333,152 violate manufacturing sense)
P2 same-name consistency 80.1%
P4 physical validity 100.00%
P5 symmetric-part consistency 98.1%
Composite 94.5%

Property values vs engineering reference

metal 1.00× / 1.01× · plastic 1.04× / 0.98× · ceramic 0.98× / 1.07× · wood 0.93× / 0.92× · rubber 0.91× / 0.88× · glass 0.89× / 0.89× — 8 / 10 material classes within tolerance.

Simulation failure risk (every part checked)

Mode Criterion Result
Self-weight collapse $\varepsilon \geq 0.10$ 0 parts
Volumetric locking $\nu \geq 0.45$ 5,378 parts (1.61%) — all flagged in sim_caveat; use MPM or a stable-neo-Hookean FEM model
CFL explosion $dt < dx/c$ min dt 0.802 µs
Buckling aspect > 50 and E < 100 MPa 81 parts (0.02%)
Overall OK 99.98% · MEDIUM 0.02% · HIGH 0 · CRITICAL 0

Known limitations

1. 46,907 parts (14.1%) carry a sim_caveat. Three different warnings share the column. (a) Near-incompressibility — 5,378 parts with nu >= 0.45. At nu = 0.49 the bulk modulus is about 50x the shear modulus and low-order tetrahedral FEM volumetrically locks. Use a corotational MPM with enough substeps — about 250 at dt = 1e-2 and a 15.6 mm grid — or an FEM stable-neo-Hookean model; size the substep count from sim_p_wave_speed_m_s. The elastomer nu = 0.49 is a deliberate compromise: real filled elastomers are 0.4985–0.4995, but that raises the explicit substep count roughly tenfold. Linear elasticity at 0.49 still loses about 8% volume under 30% compression — for seals and pinching contacts, reach for a hyperelastic model rather than using these constants directly. (b) Dimensional basis — 41,367 parts. Woven-textile E (1.2e-4 GPa) is an ARCSim cloth membrane modulus (physically N/m, not Pa) and rho (307 kg/m³) is a bulk packing density. Use a cloth/shell solver — not valid for volumetric FEM/MPM. Foam E is an effective compressive modulus and nu is apparent, not the cell-wall value. (c) Withheld bucket — 162 parts. The cfrp bucket supplies no values: its density rested on a single record carrying an epoxy-matrix density rather than a laminate density. These parts keep material_class = composite but have null properties and cannot be simulated as-is.

2. 9,813 objects (42%) mix rigid and soft bodies. A sofa's steel frame (203 GPa) and its upholstery (1.2e-4 GPa) differ by 1.7 × 10⁶. This is physically real, but a single explicit MPM solver is throttled by the stiffest part's CFL (1 s of simulation ≈ 1.25 M steps). Follow sim_object_solver_plan.

3. Evaluation covers 11.4% of the deliverable. The 92.8% figure is measured on weak-label parts only; 64.7% of parts have ambiguous names (Leg, Frame, Other) with no independent label. Their confidence is lower (median 0.72 vs 0.996) — treat confidence as a usable signal.

4. Weak labels are authored priors, not measurements. Cross-checked against UniPhys-Bench's 2,255 human-verified parts: 12 of 14 testable rules agree (86%).

5. ceramic (50.0%) and rubber (63.9%) are the weak classes. Ceramic has only 64 weak-label samples; rubber grips often share a GLB material with the metal shaft they wrap.

6. 4,426 parts are missing for no good reason. They have a name and real geometry and should have been annotated. They sit in 906 models — 80% of those lose only one or two parts, but 26 lose many: one laptop has 495 parts of which 132 are annotated, and the 304 missing ones are all Keys. The skip does not follow part order or mesh size (kept Keys span 12–176 faces, skipped ones 2–244), so it is a gap left by feature extraction rather than a deliberate filter. Treat coverage as 95.15% of PartNeXt parts, and expect keyboards and laptops to be the worst affected. model_id + part_id absent from the CSV always means not annotated — never a failed join.

7. Upstream quirks carried through. 23 parts have part_name ? and 33 have a null part_path: their maskId exists in masks but has no node in hierarchyList. Four parts upstream store a lone face index as a bare int rather than [int]; join_with_partnext.py normalises this, hand-written loaders should too.


How the annotations were produced

Two stages. Stage 1 fuses four evidence channels into a material class; stage 2 resolves a property bucket and emits per-quantity $(\nu, E, \rho)$.

Channel Weight Alone Role
Part-name prior (391 exact category|part_name entries) 1.0 82.2% backbone — covers 97.32% of parts
Discriminative head (44-d: texture, geometry, topology, ancestor path) 1.8 in-distribution / 1.5 cap out-of-distribution 90.6% refinement
Geometry likelihood (relative extent, aspect) 2.0 48.8% refinement
Visual PBR (4-band metalness likelihood) 1.2 28.1% refinement
Visual colour (HSV histogram, out-of-domain gated) 0.4 — 74.7% skipped as out-of-domain
Fusion 95.9%

Plus material-group pooling and a realism gate (a load-bearing part may not be assigned a bucket with E < 0.5 GPa) that rejected 2,983 physically impossible pairings.

The part-name prior is the backbone because the discriminative head extrapolates on the 64.7% of parts whose names never appear in its training set: measured evidence span for Chair|Leg was 27.7 versus a prior span of 5.1, which produced "chair leg = rubber". Capping out-of-distribution evidence below the prior's span restores realistic labels.

Property values come from a provenance-traceable corpus of 82 sources (CC0 / CC BY / CC BY-SA / ODbL / MIT / US-Government, plus one non-profit-use cloth database and one manufacturer datasheet — see license_tier), name-grounded into 84 buckets so that a bucket named brass_bronze actually contains C86300/CuZn grades rather than whatever has a similar density.


Citation

Cite PartNeXt alongside this dataset.

@misc{partnext_material_properties,
  title  = {PartNeXt Material and Simulation-Ready Physical Property Annotations},
  year   = {2026},
  url    = {https://huggingface.co/datasets/minsj1225/PartNeXt-Material-Properties}
}

Attribution you are required to give

Most of the corpus is under licences that make attribution a condition of use, not a courtesy. If your work uses parts in the corresponding license_tier, cite the source:

Source Rows using it Licence Required citation
Global Wood Density Database v.2 67,026 CC BY 4.0 Fischer, F.J., Chave, J., Zanne, A.E. et al. (2026), GWDD v.2 v2.2, Zenodo 10.5281/zenodo.20815517; article: New Phytologist 10.1111/nph.70860
ARCSim cloth material database 41,177 non-profit use only, citation required Wang, H., Ramamoorthi, R., O'Brien, J.F. (2011), Data-Driven Elastic Models for Cloth: Modeling and Measurement, ACM TOG 30(4), 71:1–11
CIRAD wood density database 67,026 CC BY-SA 4.0 Vieilledent et al., github.com/ghislainv/wood-density-Cirad — share-alike
SciGlass 28,791 ODC-ODbL SciGlass database, EPAM Systems release — share-alike
Soft Robotics Materials Database 916 ODC-ODbL 1.0 Marechal, L. et al. (2021), Toward a Common Framework and Database of Materials for Soft Robotics, Soft Robotics 8(3), 284–297, 10.1089/soro.2019.0115 — share-alike
StressEng 262,628 CC BY 4.0 Kumar, Kabra & Cole (2024), Scientific Data 11:1273
USDA FPL Wood Handbook, MIL-HDBK-5J, NIST 103,451+ US Government works no restriction; cited in material_corpus_sources.csv

Per-source citation, DOI, licence and clearance for all 82 corpus sources are in material_corpus_sources.csv; the {rho,E,nu}_source_ids columns say which ones a given row actually used.

License

The annotation layer — part-to-material assignment, bucket resolution, solver metadata — is CC BY 4.0. The property values are not uniformly licensed: each one inherits the licence of the corpus sources it was measured from, and those sources are not mutually compatible. PartNeXt meshes and part masks are CC BY 4.0 (AuWang/PartNeXt, AuWang/PartNeXt_mesh).

A row blends several corpus sources, so it cannot be assigned to a single-licence file. Instead every row carries license_tier, the strictest licence among the sources that produced its values. Filter on it:

license_tier parts share what it means for you
PERMISSIVE 85,763 25.7 % CC0 / CC BY / MIT — attribution only
PUBLIC_DOMAIN 92,825 27.9 % US Government works — no restriction
COPYLEFT 96,733 29.0 % CC BY-SA 4.0 (CIRAD wood) and ODbL (SciGlass, Soft Robotics Materials Database) — share-alike; these two are also incompatible with each other
NONCOMMERCIAL_OK 41,177 12.4 % all fabric parts. The ARCSim cloth material database is released for non-profit use only and requires citation — not usable commercially
SPEC_DATA_CITED 12,205 3.7 % Eastern Alloys Zamak manufacturer datasheet — published specification data, cited not relicensed
NO_SOURCE 4,449 1.3 % abstained rows, no property values

183,037 parts (54.9 %) are PERMISSIVE, PUBLIC_DOMAIN or NO_SOURCE and carry no share-alike or non-commercial obligation.

Per-source citation, DOI, licence and clearance for all 82 corpus sources are in material_corpus_sources.csv.