LMBuild

LMBuild asks a system to build a real, working object as an assembly of parts. The object can be an office desk, a wheelchair, a scissor car jack, a 3D printer or a LEGO fire engine. The system places catalogue parts or creates new ones, and declares what each part is, what it is made of, where the object moves and how it is put together. It works from the object's name and an in-the-wild photo, through a tool environment. The result is then scored as a physical product on twelve metrics in four levels: Soundness, Affordance, Design and Realization.

Code (environment, agents, evaluator, table scripts): https://github.com/Lumos-Jiateng/LMBuild

Contents

Folder	What	Size
core/	LMBuild-Core, the 200 evaluated objects: task specs (6 prompt conditions), in-the-wild images, reference assemblies with joints, mixed subpart pools, Wikipedia/Wikidata-grounded part, attribute, subsystem and kinematics annotations with evidence	~12 GB
full/	LMBuild-Full, 2,549 objects (Core plus 2,350 extended objects) in the same format	~85 GB
results/	Everything behind the paper's tables: all 6,858 scored designs of the 30 systems in the main setting (re-scorable), per-task evaluator fixtures and both VLM judges' answers, score tables, the human study, and example agent trajectories	~80 GB

Paths inside core/core_v2.json are relative to a release root that contains benchmark/. The GitHub repository's scripts/download_data.py downloads core/ (and optionally full/, results/) into the expected places.

Sources and licences

LMBuild repurposes BrickNet, BrickComposer, PartNeXt, the Fusion 360 Gallery assembly dataset, Artiverse, and complete assemblies from 27 open-source hardware projects. Because the Fusion 360 Gallery data is CC BY-NC-SA 4.0, the dataset as a whole is released under CC BY-NC-SA 4.0; every object keeps its upstream licence string in reference.license, and the per-source terms are listed in DATA_LICENSES.md in the code repository.