PhotoReal-Robo

PhotoReal-Robo is a large-scale, photo-realistic 3D asset dataset designed for robotic manipulation research and embodied AI simulation. It contains 1,001 high-fidelity 3D objects spanning 313 fine-grained categories across 12 super-categories, all modeled at desktop/tabletop scale.

Dataset Overview

Property	Value
Total Objects	1,001
Fine-grained Categories	313
Super Categories	12
License	CC-BY-4.0
3D Format (Visual)	.glb (glTF Binary)
3D Format (Collision)	.obj (Wavefront OBJ)
Preview Format	.png
Metadata Format	.json

Super Categories

Super Category	# Objects	# Sub-categories
Decor and Figurines	301	96
Stationery and Office	216	68
General Merchandise	161	51
Toys and Models	125	39
Electronics	68	21
Kitchenware and Tableware	63	22
Beauty and Personal Care	14	4
Gardening	13	4
Lighting	13	4
Other	13	4
Hardware and Tools	7	3
Snacks and Confectionery	7	3

Material Distribution

Material	Count	Percentage
Plastic	274	27.4%
Metal	261	26.1%
Wood and Bamboo	251	25.1%
Ceramic	102	10.2%
Glass	63	6.3%
Paper	19	1.9%
Rubber and Silicone	12	1.2%
Fabric	6	0.6%
Leather	4	0.4%
Others	9	0.9%

Directory Structure

Each object category is organized as a top-level folder with the following structure:

<Category_Name>/
├── visual/
│   ├── <asset_name>.glb              # High-fidelity visual mesh (glTF Binary)
│   └── <asset_name>_meta.json        # Per-object metadata
├── collision/
│   └── <asset_name>.obj              # Simplified collision mesh
└── preview/
    └── <asset_name>.png              # Preview rendering

Per-Object Metadata

Each object includes a _meta.json file with the following fields:

{
  "meta_version": 1,
  "category": "Desktop_Miniature_Train",
  "name": "metal_miniature_train_silver_modern",
  "visual": "visual/metal_miniature_train_silver_modern.glb",
  "preview": "preview/metal_miniature_train_silver_modern.png",
  "collision": "collision/metal_miniature_train_silver_modern.obj",
  "meta": "visual/metal_miniature_train_silver_modern_meta.json",
  "preview_rendered": false,
  "collision_built": true,
  "bounds": {
    "min": [-0.139, 0.0, -0.628],
    "max": [0.138, 0.393, 0.526]
  },
  "density": {
    "method": "names",
    "water_reference_kg_m3": 1000.0,
    "ok": true,
    "predicted_kg_m3": 7000.0,
    "confidence": "medium",
    "material_guess": "Die-cast zinc alloy",
    "notes": "Typical density for die-cast metal model trains."
  }
}

Metadata Fields

Field	Description
meta_version	Schema version (currently 1)
category	Object category name
name	Unique asset identifier
visual	Relative path to the visual mesh (.glb)
preview	Relative path to preview image (.png)
collision	Relative path to collision mesh (.obj)
collision_built	Whether the collision mesh has been generated
bounds	Axis-aligned bounding box (min/max in meters)
density	Physical density estimation with material prediction

Additional Files

File	Description
embodiment_taxonomy_stats.json	Full taxonomy tree with per-category statistics
scales_global.json	Global scale normalization parameters for all objects
desktop_put_in_targets.json	List of 22 container-type objects suitable as "put-in" targets for manipulation tasks

Intended Use

This dataset is designed for:

• Robotic manipulation simulation — realistic object grasping, placement, and rearrangement tasks
• Embodied AI training — sim-to-real transfer with photo-realistic visual assets
• Scene generation — composing realistic tabletop/desktop scenes for training and evaluation
• Object recognition — fine-grained object classification with diverse materials and geometries
• Physics simulation — collision meshes and density estimates enable realistic dynamics

Evaluate asset quality

Please use the scripts "VISER_env/eval_quality.py" to evaluate the asset quality.

Citation

If you use this dataset in your research, please cite:

@misc{zhu2026visuallyrealisticsimulationbenchmark,
      title={Toward Visually Realistic Simulation: A Benchmark for Evaluating Robot Manipulation in Simulation}, 
      author={Yixin Zhu and Zixiong Wang and Jian Yang and Jin Xie and Jingyi Yu and Jiayuan Gu and Beibei Wang},
      year={2026},
      eprint={2605.06311},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.06311}, 
}

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC-BY-4.0).