boardd/dexwild-dataset
Tony Tao*, Mohan Kumar Srirama*, Jason Jingzhou Liu, Kenneth Shaw, Deepak Pathak Robotics: Science and Systems (RSS) 2025 [Project page] [Video] [ArXiv] [Main Code Repo] [Policy Training] [Hardware Guide] [Data Collection Guide] Data Breakdown The DexWild dataset is one of the largest, high quality human dexterous hand datasets with a huge diversity of objects and environments. In total, we collect 9,505 Episodes across 5 tasks and 93 environments. 33+ Hours… See the full description on the dataset page: https://huggingface.co/datasets/boardd/dexwild-dataset.
<p align="center"> <img src="website_assets/imgs/title.png" width="100%"> </p>
<div align="center">
Tony Tao<sup>\</sup>, [Mohan Kumar Srirama](https://www.mohansrirama.com/)<sup>\</sup>, Jason Jingzhou Liu, Kenneth Shaw, Deepak Pathak
Robotics: Science and Systems (RSS) 2025
[[Project page]](https://dexwild.github.io/) [[Video]](https://youtu.be/oMaamSkcl5E) [[ArXiv]](https://arxiv.org/abs/2505.07813)
[[Main Code Repo]](https://github.com/dexwild/dexwild) [[Policy Training]](https://github.com/dexwild/dexwild-training)
[[Hardware Guide]](https://resisted-salad-9e6.notion.site/DexWild-Hardware-Setup-Guide-20eee3f68d27801b8eb8dfde3a5bb7c4?source=copylink) [[Data Collection Guide]](https://resisted-salad-9e6.notion.site/DexWild-Data-Collection-Guide-20fee3f68d27803b9a88ef9847e292d4?source=copylink)
</div>
Data Breakdown
The DexWild dataset is one of the largest, high quality human dexterous hand datasets with a huge diversity of objects and environments. In total, we collect 9,505 Episodes across 5 tasks and 93 environments. 33+ Hours equating to 3.5M+ Transitions!
Florist Human - 1,030 Episodes, 310,571 Transitions Robot - 248 Episodes, 92,954 Transitions
Clothes Folding Human - 1,123 Episodes, 294,367 Transitions Robot - 295 Episodes, 123,385 Transitions
Spray Human - 2,820 Episodes, 886,460 Transitions Robot - 388 Episodes, 162,862 Transitions
Toy Cleanup Human - 2,323 Episodes, 1,251,649 Transitions Robot - 546 Episodes, 316,289 Transitions
Pouring Human - 621 Episodes, 181,541 Transitions Robot - 111 Episodes, 30,880 Transitions
Data Structure
Data is saved in HDF5 Format in the following general structure. Note that single arm tasks are missing a few of these folders.
data
├── ep_0000
│ ├── intergripper
│ │ └── intergripper.pkl
│ ├── left_leapv2
│ │ └── left_leapv2.pkl
│ ├── left_manus
│ │ └── left_manus_full.pkl, left_manus.pkl
│ ├── left_pinky_cam
│ │ └── timestamp1.jpg, timestamp2.jpg, ...
│ ├── left_thumb_cam
│ │ └── timestamp1.jpg, timestamp2.jpg, ...
│ ├── left_tracker
│ │ └── left_tracker_cam_frame_abs.pkl, ...
│ ├── right_leapv2
│ │ └── right_leapv2.pkl
│ ├── right_manus
│ │ └── right_manus_full.pkl, right_manus.pkl
│ ├── right_pinky_cam
│ │ └── timestamp1.jpg, timestamp2.jpg, ...
│ ├── right_thumb_cam
│ │ └── timestamp1.jpg, timestamp2.jpg, ...
│ ├── right_tracker
│ │ └── right_tracker_cam_frame_abs.pkl, ...
│ ├── timesteps
│ │ └── timesteps.txt
│ ├── zed
│ │ └── zed_ts.pkl
│ └── zed_obs
│ └── timestamp1.jpg, timestamp2.jpg, ...
├── ep_0001
├── ep_0002
├── ...Data Recombination
Since HuggingFace has a limit on single file size, you must recombine the data to obtain the full dataset for certain HDF5 files.
After downloading the split files, follow this guide.
Merging Instructions:
- Navigate to the Download Directory:
cd path_to_downloaded_files- Merge Files Using `cat`:
Use the cat command to concatenate the split parts. Replace filename.part_* with your actual file name.
cat filename.part_* > filename.tarExample:
cat DexWild_Data.part_* > DexWild_Data.tar- Verify the Merged File:
Ensure that the merged file size matches the original file size before merging. You can use the ls -lh command to check file sizes.
ls -lh DexWild_Data.tar- Extract the Dataset:
Once merged, extract the dataset using the tar command:
tar -xvf DexWild_Data.tarCitation
If you find this project useful, please cite our work:
@article{tao2025dexwild,
title={DexWild: Dexterous Human Interactions for In-the-Wild Robot Policies},
author={Tao, Tony and Srirama, Mohan Kumar and Liu, Jason Jingzhou and Shaw, Kenneth and Pathak, Deepak},
journal={Robotics: Science and Systems (RSS)},
year={2025}}
