AssemblyBench

2,789 assemblies, 18,687 normalized source part meshes and 130,853 embedded PNGs, with multimodal manuals, poses and complete motions. One assembly per row. All payloads are embedded in Parquet.

from datasets import load_dataset
objects = load_dataset("AssemblyWorld/assemblybench", split="all")
sample = objects[0]

Requires datasets >= 5.0.1 and Pillow. Pin revision for reproducibility; use the prepared local directory instead of the Hub ID for offline loading.

Contents and coordinates

• parts: source IDs, names, float64 vertices, polygons, normals, corner normal indices and similar-part IDs. Meshes are the supplied objs_normalized.
• assembly_order, steps: original order, cumulative part IDs and indices into manual and manual_pages.
• manual: original JSON records, including original shape IDs and every text detail level. Missing optional descriptions stay absent; no embeddings are stored.
• poses: original view/step/part JSON mapping, preserving all source values.
• motions: per-part frames, preserving the full original sequence of poses.
• manual_pages: all diagram and arrow PNGs, source filenames, dimensions, kind, view and step IDs. PNG bytes are unchanged and decode to PIL images.
• source_splits, source_metadata: original membership and provenance.

We do not transform the supplied geometry. Upstream generation centers, PCA-aligns and scales part meshes. Both poses and motion frames use [tx, ty, tz, qw, qx, qy, qz]: apply quaternion rotation to a source mesh vertex, then add translation to obtain its position in the source world frame. The source provides pose view "0"; do not infer poses for the additional image views. No physical unit conversion is assumed. Motions are indexed frames, not timestamped robot commands. In this release each part has 12 frames.

The source image folders include extra PNGs beyond the regular step/view grid. They are retained with their original filenames; an unparseable filename has null view/step IDs, and pages outside the assembly step range have no step reference. No page numbers or geometry are repaired. Source point clouds, masks, embeddings, checkpoints, the unrelated point-cloud screenshot in object 8005, and macOS metadata are excluded.

Splits

The stored split is full; all selects it. source_splits preserves the official partition: train 2,231, val 278, test 280. Filter using source_splits; no new random split is created.

Source and rights

Converted from AssemblyBench v1, archive SHA-256 8f46d59dc7692822043efa28de77f98996aa736b76a97982d51c255ea6c06c5b. See official code revision a1ed6b8b70b963fc582744ae06eaad4de5633cfa for coordinate conventions. New annotations and modifications are CC BY-SA 4.0, copyright MERL 2026. Original Assemble-Them-All data retains its MIT terms and underlying Fusion 360 Gallery assets retain their source restrictions. This is a converted portion/modified set of the Fusion 360 Gallery Dataset, not the complete original Gallery Dataset. Underlying asset access, use, modification and redistribution are limited to non-commercial research under the Fusion 360 Gallery Dataset License. Those terms apply to downstream recipients and redistribution, including retained notices, Autodesk warranty exclusions and indemnification obligations. The annotation license does not remove underlying asset restrictions. The converter's MIT code license does not relicense these assets. This release changes storage format and embeds the selected source assets.

Please cite Li, Zhang, Egger, Chatterjee, Lohit, Marks and Cherian, AssemblyBench: Physics-Aware Assembly of Complex Industrial Objects, CVPR 2026. Schema version 1; conversion version 1.