jaybhagiya/FAST-assembly-meshes

FAST Assembly Meshes

The part meshes and contact graphs of the 2,132 assemblies of the FAST simulation labels (jaybhagiya/FAST-assembly-labels, release fast-labels-v1). With both datasets, the labels and the code (ZeMA-gGmbH/FAST, release v2.0.0) can be used without a separate download of the ASAP assembly dataset.

This dataset is a portion of the Fusion 360 Gallery Dataset (Autodesk), in the processed form distributed with ASAP. It is not the Fusion 360 Gallery Dataset itself. Use is restricted to non-commercial research; see LICENSE.md.

Provenance

Meshes from the Fusion 360 Gallery Dataset (Autodesk), in the processed form distributed with ASAP (Tian et al., 2024), downloaded from the ASAP project page (dataset_2404/training_assembly.zip and test_assembly.zip).

Every published file is byte-identical to the corresponding file in those two ASAP archives. The processing (from the Fusion 360 Gallery assemblies to one OBJ mesh per part in the assembled pose, at centimetre scale, and the contact graphs) was done by the ASAP authors; this dataset applies no further processing: no subdivision, rescaling, repair, or re-export. ZeMA gGmbH selected the 2,132 assemblies of the FAST labels, arranged them by the label splits, and added the manifest.

These are the exact files used to generate the FAST labels: for every assembly, input_sha256 computed from these files equals the value in the labels dataset.

Contents

Split	Assemblies	Files	Archive	Bytes
train	1,703	15,068	data/train.tar.gz	535,296,399
validation	191	1,696	data/validation.tar.gz	68,581,555
test	237	4,674	data/test.tar.gz	124,717,794
excluded	1	14	data/excluded.tar.gz	188,725

The splits are those of the labels. The excluded split holds test_assembly/07386, an exact duplicate of training_assembly/07385, which the labels keep as excluded.

Each archive holds <split>/<assembly_id>/<part>.obj (one triangle mesh per part, Wavefront OBJ) and <split>/<assembly_id>/contact_graph.json (the contacting part pairs, NetworkX node-link JSON). Nothing else is included: no signed-distance (.sdf) caches, pose caches, features, or planner outputs, and no other assembly of the ASAP or Fusion 360 Gallery data.

Other files:

• manifest.jsonl: one line per assembly with assembly_id, source_split (training_assembly or test_assembly, ASAP's folder), dataset_split, included, tier, part_count, input_sha256, archive, and the SHA-256 of every file (files).
• place_meshes.py: unpacks the archives into the layout of the FAST code and verifies every assembly (standard library only).
• SHA256SUMS: SHA-256 of every file of this dataset.

Use with FAST

hf download jaybhagiya/FAST-assembly-meshes --repo-type dataset --local-dir data/hf-meshes
(cd data/hf-meshes && sha256sum -c SHA256SUMS)
python data/hf-meshes/place_meshes.py --dataset-dir data/hf-meshes --source-root "$FAST_SOURCE_ROOT"

place_meshes.py writes $FAST_SOURCE_ROOT/training_assembly/<id>/ and $FAST_SOURCE_ROOT/test_assembly/<id>/, the layout that the FAST commands read through --source-root, and checks every assembly's input_sha256. --verify-only checks an existing folder.

input_sha256 identifies an assembly's input: for each *.obj, contact_graph.json, and config.json present in the folder, sorted by name, it hashes the name, a zero byte, the file's SHA-256 in hex, and a newline (_hash_assembly in the FAST code; input_sha256 in place_meshes.py). These assemblies have no config.json; every mesh is stored in its assembled pose.

Units

The meshes have no unit field. ASAP uses centimetres, and the FAST simulator settings assume them. Each assembly's largest dimension is at most 10 cm: 1,982 assemblies measure exactly 10 cm and 150 are smaller (the smallest is 0.42 cm).

SDF caches

The simulator builds a signed-distance field (<part>.sdf) for every mesh on first use and stores it next to the mesh; later runs reuse it. Delete the .sdf files after changing a mesh. The label pipeline builds a separate, immutable cache per assembly (python -m plan_sequence.prepare_sdf_cache --assembly-dir <assembly> --cache-dir <cache-root>/<input_sha256> --expected-input-sha256 <input_sha256>).

Limitations

• One CAD source: all assemblies come from the Fusion 360 Gallery assembly dataset; part counts are 3 to 70.
• The meshes are ASAP's processed version. Parts are rigid; joints and materials of the original CAD models are not included.
• ASAP recommends subdividing the meshes for denser contact points. FAST's labels were generated on these original, not subdivided, meshes.

Citation

If you use this dataset, please cite the source datasets and FAST:

@inproceedings{willis2022joinable,
  author    = {Willis, Karl D. D. and Jayaraman, Pradeep Kumar and Chu, Hang and Tian, Yunsheng and Li, Yifei and Grandi, Daniele and Sanghi, Aditya and Tran, Linh and Lambourne, Joseph G. and Solar-Lezama, Armando and Matusik, Wojciech},
  title     = {{JoinABLe}: Learning Bottom-up Assembly of Parametric {CAD} Joints},
  booktitle = {2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {15828--15839},
  year      = {2022},
  doi       = {10.1109/CVPR52688.2022.01539}
}

@inproceedings{tian2024asap,
  author    = {Tian, Yunsheng and Willis, Karl D. D. and Al Omari, Bassel and Luo, Jieliang and Ma, Pingchuan and Li, Yichen and Javid, Farhad and Gu, Edward and Jacob, Joshua and Sueda, Shinjiro and Li, Hui and Chitta, Sachin and Matusik, Wojciech},
  title     = {{ASAP}: Automated Sequence Planning for Complex Robotic Assembly with Physical Feasibility},
  booktitle = {2024 IEEE International Conference on Robotics and Automation (ICRA)},
  pages     = {4380--4386},
  year      = {2024},
  doi       = {10.1109/ICRA57147.2024.10611595}
}

@mastersthesis{bhagiya2026fast,
  author = {Bhagiya, Jaykumar},
  title  = {Feasibility Aware Assembly Sequence Planning using Graph Transformer},
  school = {Saarland University},
  year   = {2026}
}