DexWild: Human and Robot Dexterous Hand Demonstrations in the Wild

Provider
Tony Tao, Mohan Kumar Srirama et al. (Carnegie Mellon University)
License
MIT (source)
Size
108 files
Stored at
Hugging Face (https://huggingface.co/datasets/boardd/dexwild-dataset)
Access
Free and public. No account with the data provider is needed.
Category
Robotics
Source
https://arxiv.org/abs/2505.07813

About DexWild: Human and Robot Dexterous Hand Demonstrations in the Wild

DexWild (RSS 2025, CMU) is a dataset of human dexterous hand demonstrations collected in diverse real environments, paired with robot demonstrations, for co-training dexterous robot policies (Tao et al., arXiv:2505.07813). The card reports 9,505 episodes across 5 tasks and 93 environments, more than 33 hours and 3.5 million transitions. Tasks are florist, clothes folding, spray, toy cleanup and pouring, each split into human and robot folders. Episodes store left and right LEAP hand states, Manus glove data, tracker poses, palm-mounted pinky and thumb camera images (JPEG), ZED stereo camera observations and timesteps, in a per-episode folder layout with pickled arrays, plus HDF5 files that must be recombined from split parts as described on the card. Measured 2026-11-14 at about 4.4 TB. Read with h5py, pickle and the dexwild training code on GitHub. Useful for learning from human video and motion, cross-embodiment co-training and dexterous manipulation. The card carries an MIT licence badge; cite the DexWild paper.

How to load DexWild: Human and Robot Dexterous Hand Demonstrations in the Wild in Python

from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="boardd/dexwild-dataset", filename="clothes_data/human/human_clothes_data.part_00", repo_type="dataset")
print(path)

Data structure

Files

Example file: clothes_data/human/human_clothes_data.part_00

Analyze this dataset without downloading it

On Scigantic, DexWild: Human and Robot Dexterous Hand Demonstrations in the Wild opens in a hosted JupyterLab notebook with the data already mounted, so you can read files where they are stored instead of copying them first. Each notebook includes an AI assistant that can see the mounted files and write and run analysis code. Open the dataset page to launch a notebook, or browse the full catalog.

This page is also available as Markdown and as JSON.

Related Robotics datasets