argonrobotics-ai/alloy-library

Alloy object library

5,423 physics-ready objects for the Alloy robot simulator. Each object is a card (YAML: shape plus sourced material values: density, friction, restitution, stiffness, each with its source and uncertainty) together with its physics mesh, convex-hull collision sidecar, textured visual mesh and a rendered thumbnail (and a look where the library has one). Every object passed Alloy's card checks before it was added; those an importer flagged (an implausible mass or size, a failed physics check) say why in flags.

The dataset is public: anyone can download it, with no access request, account or token. Every object is under a permissive licence that allows commercial use (CC BY 4.0: 4,976 objects, MIT: 393, CC0 1.0: 29, Apache 2.0: 25); credits are in ATTRIBUTION.md (see Licences and attribution).

Library	Source	Objects	Size	Licences
alloy	Alloy's own models	30	0.01 GB	Apache-2.0 (25), CC0-1.0 (5)
gymanything	Gym-Anything for Robotics	393	0.14 GB	MIT (393)
objaverse	Objaverse 1.0 (CC BY and CC0 objects only)	1,989	6.06 GB	CC-BY-4.0 (1965), CC0-1.0 (24)
robocasa	RoboCasa kitchen assets (Lightwheel packs)	1,900	1.79 GB	CC-BY-4.0 (1900)
ycb_gso	YCB Object and Model Set and Google Scanned Objects	1,111	1.85 GB	CC-BY-4.0 (1111)
all		5,423	9.85 GB

• alloy: Authored props with sourced dimensions.
• gymanything: Objects of 28 simulations, with the colours and materials of their renders.
• objaverse: Robot-relevant categories (food, office, kitchen, tools, containers, packaging, lab).
• robocasa: Kitchen objects and fixtures as MJCF with convex hulls and textures.
• ycb_gso: Household scans: watertight physics mesh, convex hulls, textured glTF; measured or estimated mass.

Use it with Alloy

pip install huggingface_hub          # once; Alloy downloads with it (no account or token needed)
alloy fetch --list --library ycb_gso                  # what is there (nothing downloaded)
alloy fetch ycb_025_mug                               # one object
alloy fetch --library robocasa --category blenders    # a category
alloy fetch --library objaverse                       # a whole library

Objects go to ~/.cache/alloy/imports/<library>/ (or $ALLOY_CACHE/imports/<library>/); files already there are not downloaded again. A Hugging Face login (hf auth login or HF_TOKEN) is optional: the Hub gives logged-in downloads a higher rate limit, which helps when fetching whole libraries. Alloy finds fetched objects by card name:

import alloy

scene = alloy.Scene(dt=0.01)
scene.add_ground()
scene.add_object("mug", "ycb_025_mug", pos=(0, 0, 0.1))
sim = alloy.build(scene)

alloy compose "a mug and a bowl on a table" picks from every fetched object. ALLOY_ASSETS_REPO=<owner>/<name> points Alloy at another copy of this dataset (default argonrobotics-ai/alloy-library); ALLOY_ASSETS_REVISION pins a revision.

Layout

index.jsonl                      one row per object
objects/<library>/<name>/        card.yaml, material.yaml, meshes, physics.hulls.obj, textures, thumb.png
libraries/<library>/             files shared by a library's objects (manifest.json, shared materials, looks)
libraries/textures/              CC0 texture sets (Poly Haven, ambientCG) for scene variants; credits in its README
LICENSES/  ATTRIBUTION.md        licence texts and every object's credit line

An index.jsonl row: name, library, category, title, licence, attribution, source (upstream URL), showcase_ok, mass_basis, estimate, sources, flags (only when flagged), size (bytes), sha256 (the object's tree digest) and files (each file's name in the object folder, its path under the library root, sha256 and bytes).

Without Alloy:

from huggingface_hub import snapshot_download
snapshot_download("argonrobotics-ai/alloy-library", repo_type="dataset", allow_patterns=["objects/ycb_gso/ycb_025_mug/*"])

Values: sourced or estimated

Card values carry their source. mass_basis says where an object's mass comes from (measured, sourced, datasheet, authored, derived, or estimate); sources gives the kind of each material value (catalog: a cited datasheet or handbook value, fit: fitted to a measured mass, model: a stated model or typical value). 5,180 of 5,423 objects have estimate: true: their mass or a material value is an estimate, so check those values before you rely on them.

Licences and attribution

Every object keeps its upstream licence (CC BY 4.0: 4,976 objects, MIT: 393, CC0 1.0: 29, Apache 2.0: 25), and all of them allow commercial use. Each object's licence is the licence field of index.jsonl; the licence texts are in LICENSES/ (Apache-2.0, CC-BY-4.0, CC0-1.0, MIT, MIT-gymanything, ODC-By-1.0). CC BY 4.0 objects need credit when you share them or images of them: ATTRIBUTION.md lists every object's credit line (also the attribution field of the index). Alloy changed the objects (repaired and simplified meshes, convex hulls, re-exported visuals, estimated physical values); the upstream files are linked in source.

Brands. Some scans show third-party trademarks (product packaging, logos). A licence covers the scan, not the brand. showcase_ok: false marks the 2,364 objects whose name or label shows a brand (or may); keep them out of marketing images, docs and launch material. They remain usable in simulation.

Citation

If you use this library, cite Alloy and the upstream datasets you use:

@misc{alloy_library,
  title  = {Alloy object library},
  author = {{Alloy contributors}},
  year   = {2026},
  howpublished = {Hugging Face dataset},
}
@article{calli2015ycb,
  title   = {Benchmarking in Manipulation Research: Using the {Yale-CMU-Berkeley} Object and Model Set},
  author  = {Calli, Berk and Walsman, Aaron and Singh, Arjun and Srinivasa, Siddhartha and Abbeel, Pieter and
             Dollar, Aaron M.},
  journal = {IEEE Robotics \& Automation Magazine},
  year    = {2015},
}
@inproceedings{downs2022gso,
  title     = {Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items},
  author    = {Downs, Laura and Francis, Anthony and Koenig, Nate and Kinman, Brandon and Hickman, Ryan and
               Reymann, Krista and McHugh, Thomas B. and Vanhoucke, Vincent},
  booktitle = {ICRA},
  year      = {2022},
}
@inproceedings{nasiriany2024robocasa,
  title     = {RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots},
  author    = {Nasiriany, Soroush and Maddukuri, Abhiram and Zhang, Lance and Parikh, Adeet and Lo, Aaron and
               Joshi, Abhishek and Mandlekar, Ajay and Zhu, Yuke},
  booktitle = {RSS},
  year      = {2024},
}
@inproceedings{deitke2023objaverse,
  title     = {Objaverse: A Universe of Annotated 3D Objects},
  author    = {Deitke, Matt and Schwenk, Dustin and Salvador, Jordi and Weihs, Luca and Michel, Oscar and
               VanderBilt, Eli and Schmidt, Ludwig and Ehsani, Kiana and Kembhavi, Aniruddha and Farhadi, Ali},
  booktitle = {CVPR},
  year      = {2023},
}