|
|
--- |
|
|
license: odbl |
|
|
task_categories: |
|
|
- robotics |
|
|
- video-classification |
|
|
- image-classification |
|
|
- object-detection |
|
|
tags: |
|
|
- dexterous-manipulation |
|
|
- hand-object-interaction |
|
|
- motion-capture |
|
|
- physics-simulation |
|
|
- rgbd |
|
|
- contact-forces |
|
|
- computer-vision |
|
|
size_categories: |
|
|
- 10K<n<100K |
|
|
--- |
|
|
|
|
|
# DexCanvas: Dexterous Manipulation Dataset v0.1 |
|
|
|
|
|
**⚠️ TEST RELEASE**: This is a preview version containing 1% of the full dataset. Contact force data is not included in v0.1. |
|
|
|
|
|
DexCanvas is a large-scale hybrid dataset for robotic hand-object interaction research, combining real human demonstrations with physics-validated simulation data. |
|
|
|
|
|
## Dataset Statistics (v0.1 Test Release) |
|
|
|
|
|
- **Total Frames**: ~30 million multi-view RGB-D frames |
|
|
- **Total Duration**: ~70 hours of dexterous hand-object interactions |
|
|
- **Real Demonstrations**: ~0.7 hours of human mocap data (1/100 of collected data) |
|
|
- **Expansion Ratio**: 100× from real to simulated data |
|
|
- **Manipulation Types**: 21 types based on Cutkosky taxonomy |
|
|
- **Objects**: 30 objects (geometric primitives + YCB objects) |
|
|
- **Capture Rate**: 100 Hz optical motion capture |
|
|
|
|
|
## Manipulation Coverage |
|
|
|
|
|
The dataset spans four primary grasp categories: |
|
|
|
|
|
- **Power Grasps**: Full-hand wrapping grips |
|
|
- **Intermediate Grasps**: Mixed precision-power combinations |
|
|
- **Precision Grasps**: Fingertip-based manipulation |
|
|
- **In-Hand Manipulation**: Object reorientation and repositioning |
|
|
|
|
|
All 21 manipulation types follow the Cutkosky grasp taxonomy. |
|
|
|
|
|
## Data Modalities |
|
|
|
|
|
Each frame includes: |
|
|
|
|
|
- **RGB-D Data**: Multi-view color and depth images |
|
|
- **Hand Pose**: MANO hand parameters with high-precision tracking |
|
|
- **Object State**: 6-DoF pose and object wrenches |
|
|
- **Annotations**: Per-frame labels and metadata |
|
|
|
|
|
**Note**: Contact force data is not included in v0.1. Contact forces will be available in future releases. |
|
|
|
|
|
## Data Pipeline |
|
|
|
|
|
The dataset is generated through three stages: |
|
|
|
|
|
1. **Real Capture**: Optical motion capture of human demonstrations at 30 Hz |
|
|
2. **Force Reconstruction**: RL-based physics simulation to infer contact forces |
|
|
3. **Physics Validation**: Verification of contact points, forces, and object dynamics |
|
|
|
|
|
This hybrid approach provides contact information impossible to observe directly in real-world scenarios while maintaining physical accuracy. |
|
|
|
|
|
## Installation |
|
|
|
|
|
```bash |
|
|
pip install datasets huggingface_hub |
|
|
``` |
|
|
|
|
|
For image processing and visualization: |
|
|
|
|
|
```bash |
|
|
pip install pillow numpy torch |
|
|
``` |
|
|
|
|
|
Authenticate with HuggingFace (required for private datasets): |
|
|
|
|
|
```bash |
|
|
huggingface-cli login |
|
|
``` |
|
|
|
|
|
Or set your token as an environment variable: |
|
|
|
|
|
```bash |
|
|
export HF_TOKEN="your_token_here" |
|
|
``` |
|
|
|
|
|
## Quick Start |
|
|
|
|
|
### Data Structure |
|
|
|
|
|
```json |
|
|
{ |
|
|
"trajectory_meta_data": { |
|
|
"generated_data": "int", |
|
|
"data_fps": "int", |
|
|
"mocap_raw_data_source": { |
|
|
"operator": "str", |
|
|
"object": "str", |
|
|
"gesture": "str" |
|
|
}, |
|
|
"total_frames": "int", |
|
|
"mano_hand_shape": "(10,)" |
|
|
//... |
|
|
}, |
|
|
"sequence_info": { |
|
|
"timestamp": "(T,)", |
|
|
"hand_joint": { |
|
|
"position": "(T, 3)", |
|
|
"rotation": "(T, 3)", |
|
|
"finger_pose": "(T, 48)" |
|
|
}, |
|
|
"object_info": { |
|
|
"pose": "(T, 6)" |
|
|
}, |
|
|
"mano_model_output": { |
|
|
"joints": "(T, 63)" |
|
|
} |
|
|
} |
|
|
} |
|
|
``` |
|
|
### Visualization |
|
|
|
|
|
Visualize trajectories using the **mocap_loader**: |
|
|
|
|
|
```bash |
|
|
# Install dependencies |
|
|
pip install open3d trimesh scipy |
|
|
|
|
|
# Visualize trajectory |
|
|
python -m hand_trajectory_loader.examples.visualize_trajectory \ |
|
|
dataset.parquet 0 \ |
|
|
--mano-model assets/mano/models/MANO_RIGHT.pkl \ |
|
|
--object assets/objects/cube1.stl \ |
|
|
--show-joints |
|
|
``` |
|
|
|
|
|
Controls: **SPACE** pause/resume, **M** toggle hand mesh, **O** toggle object, **Q** quit |
|
|
|
|
|
## Version Information |
|
|
|
|
|
**v0.1 (Test Release)** includes: |
|
|
- 1% of collected real human demonstration data |
|
|
- MANO hand parameters |
|
|
- Object pose data |
|
|
- Manipulation type annotations |
|
|
|
|
|
**Coming in future releases**: |
|
|
- Complete dataset (100× larger than v0.1) |
|
|
- Contact force data with physics validation |
|
|
- Additional objects and manipulation types |
|
|
- Extended annotations and metadata |
|
|
|
|
|
## Contact |
|
|
|
|
|
**Research Collaboration** |
|
|
Academic inquiries: lyw@dex-robot.com |
|
|
|
|
|
**Business Inquiries** |
|
|
Business collaboration: info@dex-robot.com |
|
|
|
|
|
**Website** |
|
|
https://www.dex-robot.com/en |
|
|
https://dexcanvas.github.io/ |
|
|
|
|
|
## Citation |
|
|
|
|
|
```bibtex |
|
|
@article{dexcanvas2025, |
|
|
title={DexCanvas: A Large-Scale Hybrid Dataset for Dexterous Manipulation}, |
|
|
author={DexRobot Team}, |
|
|
year={2025}, |
|
|
eprint={2510.15786}, |
|
|
archivePrefix={arXiv}, |
|
|
url={https://arxiv.org/abs/2510.15786} |
|
|
} |
|
|
``` |
|
|
|
|
|
## License |
|
|
|
|
|
This dataset is released under the Open Database License (ODbL). |
|
|
|
|
|
--- |
|
|
|
|
|
**Developed by DexRobot Team** |
|
|
Last Updated: October 2025 |
|
|
|