cobot-atlas

A large-scale 3D mesh dataset for robot simulation, manipulation research, and embodied AI.

  

cobot-atlas provides 2,023 unique 3D meshes (glTF 2.0 Binary) spanning industrial components, geometric primitives, household items, YCB benchmark objects, robot hardware, and more — designed for collaborative robot simulation, VLA model training, and robotics benchmarking.

Dataset

The full dataset is hosted on Hugging Face:

torusprime/cobot-atlas — 2,024 GLB files, 33.5 GB, MIT license.

Category	Count	Examples
Industrial	734	ICs, connectors, bearings, gears, motors, sensors, fasteners
Primitives	228	Cubes, spheres, cylinders, cones, tori in various colors
Household	179	Kitchen utensils, containers, bottles, furniture
Manipulation	163	Stacking blocks, peg-in-hole, Hanoi tower, shape sorters
Robotics	88	Robot arms, grippers, depth cameras, mobile bases
YCB Objects	72	Mustard bottle, spam can, cracker box, bowls, cups
Food	66	Fruits, vegetables, prepared items
+ 24 more	493	Fabric, tools, toys, vehicles, lab equipment, medical, art supplies
Total	2,024

This Repository

This repo contains the generation pipeline — scripts, manifests, and tools used to produce the dataset.

scripts/       — Mesh generation (fal.ai text-to-image → image-to-3D)
manifests/     — Item definitions with prompts (10 manifest files)
outputs/       — Generated meshes (GLB + reference PNG + metadata)
docs/          — Documentation and prompt reference

Usage

pip install fal-client requests

# Generate meshes from manifests
python scripts/generate_meshes.py

# Verify outputs against manifests
python scripts/verify_and_document.py

Sister Project

relief-atlas — 10,000+ meshes for disaster relief, humanitarian aid, and civil protection (Germany DRK/THW/Feuerwehr, EU Civil Protection, Ukraine recovery, global disaster response).

Citation

If you use cobot-atlas in your research, please cite:

Dataset:

@misc{sepehr_mahmoudian_2026,
  author       = {Sepehr Mahmoudian},
  title        = {cobot-atlas (Revision 425eef2)},
  year         = 2026,
  url          = {https://huggingface.co/datasets/torusprime/cobot-atlas},
  doi          = {10.57967/hf/9199},
  publisher    = {Hugging Face}
}

Generation pipeline:

@software{cobot_atlas_code_2026,
  author    = {Mahmoudian, Sepehr},
  title     = {cobot-atlas: Generation Pipeline},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20697491},
  url       = {https://doi.org/10.5281/zenodo.20697491}
}

License

MIT