Mobjaverse: A Large-Scale Rigged 3D Model Dataset with Skeletal Animations

Mobjaverse is a curated dataset derived from Objaverse-XL, specifically designed for research on skeletal animation understanding, motion generation, and articulated 3D shape analysis. It is curated in the paper TopoCap and Mobjaverse contains ~19k rigged 3D models spanning ~5k distinct skeletal topologies and ~2M motion frames, making it one of the largest publicly available datasets of its kind.

---

Dataset Overview

Property	Value
Total Models	18,914
Skeletal Topologies	5,006 distinct rig structures
Total Motion Frames	2,010,447
Data Format	NumPy .npz archives
Source	Curated from Objaverse-XL

---

Data Structure

Each model is stored under mobjaverse/{model_id}/raw_data.npz and contains:

Mesh Data

Field	Shape	Description
vertices	(N, 3)	Merged mesh vertices in edit/rest pose
faces	(F, 3)	Triangular face indices
mesh_names	(P,)	Names of individual mesh parts

Skeletal Data

Field	Shape	Description
joint_names	(J,)	Names of skeletal joints
parents	(J,)	Parent joint indices (-1 for root)
matrix_local	(J, 4, 4)	Local rest-pose transform of each joint (Blender convention)
matrix_basis	(T, J, 4, 4)	Per-frame joint animation transforms ($T$ frames, $J$ joints)

Skinning & Appearance

Field	Shape	Description
skin	(N, J)	Linear blend skinning (LBS) weights
uvs	(U, 2)	UV texture coordinates
texture	(H, W, 3)	Diffuse texture map (512×512 typical)
texture_slot	(U,)	Texture slot assignment per UV coordinate

(see render.py example to understand the order of uvs and texture)

---

Dataset Splits

Split files are organized by category under datalist/train/ and datalist/validate/, with each .txt file containing one model path per line.

---

Getting Started

Installation

# Requires Python==3.11
pip install -r requirements.txt

Dependencies:

• numpy==2.2.6 — Numerical computation
• bpy==4.2 — Blender Python API for rig/animation processing
• trimesh — 3D mesh loading and manipulation
• pyrender — Off-screen rendering
• imageio[ffmpeg] — Video/animation export

Loading a Model

import numpy as np
from src.rig_package.info.asset import Asset

# Load raw data
data = np.load("mobjaverse/{model_id}/raw_data.npz", allow_pickle=True)
asset = Asset(**data)

# Access mesh vertices
print(asset.vertices.shape)       # (N, 3)
# Access skeletal joints
print(asset.joint_names)          # ['Hips', 'Spine', ...]
# Access animation frames
print(asset.matrix_basis.shape)   # (T, J, 4, 4)

Rendering Animations

See render.py.

---

Exporting to glTF/GLB

See export.py.

---

Directory Structure

Mobjaverse/
├── README.md                     # This file
├── requirements.txt              # Python dependencies
├── render.py                     # Animation rendering script
├── export.py                     # glTF/GLB export script
├── datalist/
│   ├── train/                    # Training split (8 files, one per category)
│   │   ├── biped_128.txt
│   │   ├── quadruped_128.txt
│   │   ├── aquatic_128.txt
│   │   ├── avian_128.txt
│   │   ├── hexapod_128.txt
│   │   ├── octopod_128.txt
│   │   ├── serpentine_128.txt
│   │   └── others_128.txt
│   └── validate/                 # Validation split (16 files, seen/unseen × 8 categories)
│       ├── {category}_seen_128.txt
│       └── {category}_unseen_128.txt
├── mobjaverse/                   # Model data (18,914 directories)
│   ├── 000001/
│   │   └── raw_data.npz
│   ├── 000002/
│   │   └── raw_data.npz
│   └── ...
└── src/                          # Core library
    └── ...

---

---

Citation

If you use Mobjaverse in your research, please cite our paper:

@article{pu2026topocap,
  title     = {TopoCap: Learning Topology-Agnostic Motion Priors for Monocular Video-to-Animation},
  author    = {Pu, Cheng-feng and Zhang, Jia-peng and Guo, Meng-hao and Cao, Yan-Pei and Hu, Shi-Min},
  journal   = {ACM Transactions on Graphics (TOG)},
  volume    = {45},
  number    = {4},
  year      = {2026},
  doi       = {10.1145/3799902.3811159},
  isbn      = {979-8-4007-2554-8/2026/07},
  booktitle = {Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers (SIGGRAPH Conference Papers '26)},
  address   = {Los Angeles, CA, USA},
  month     = jul,
}

This dataset is also built upon Objaverse-XL. Please cite the original dataset as well:

@inproceedings{objaverse-xl,
  title     = {Objaverse-XL: A Universe of 10M+ 3D Objects},
  author    = {Deitke, Matt and Liu, Ruoshi and Wallingford, Matthew and Ngo, Huong and Michel, Oscar and Kusupati, Aditya and Fan, Alan and Laforte, Christian and Voleti, Vikram and Gadre, Samir Yitzhak and others},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2023}
}

---

License

This dataset follows the same licensing terms as Objaverse-XL (ODC-By v1.0 license). Please refer to the original dataset for detailed license information.

---

Acknowledgments

Mobjaverse is built upon the Objaverse-XL dataset by Allen Institute for AI (AI2). We thank the original authors for making their work publicly available.