DexWild — In-the-Wild Dexterous Manipulation Dataset
DexWild is a large-scale dexterous manipulation dataset developed at Carnegie Mellon University for learning robust in-the-wild dexterous manipulation policies via human-robot co-training. Released in 2025 under the MIT license, it contains human hand demonstrations and corresponding robot data collected across 93 diverse real-world environments, spanning kitchens, workshops, offices, and outdoor settings. The dataset focuses on 6-degrees-of-freedom dexterous hand manipulation and includes egocentric video with proprioceptive data. DexWild addresses a critical gap in the training data ecosystem: most existing dexterous manipulation datasets are collected in controlled lab settings, whereas DexWild captures the full diversity of real-world manipulation contexts. The dataset is distributed in ImageFolder format on Hugging Face with auto-converted Parquet. Full episode count is unconfirmed from the preview viewer. DexWild is a priority dataset for Geppetto covering the dexterous manipulation task category, which has the fewest publicly available datasets relative to its importance for humanoid robot deployment.
| Year | 2025 |
|---|---|
| Embodiments | 6 DoF dexterous hand |
| Modalities | rgb, proprioception |
| View type | egocentric |
| Task categories | dexterous-manipulation, human-robot-interaction |
| Data format | imagefolder, parquet |
| License | MIT |
| Access | open — commercial use permitted |
| Maintainer | Carnegie Mellon University |
| Origin country | US |
What is it?
DexWild is a large-scale dexterous manipulation dataset developed at Carnegie Mellon University for training robust in-the-wild dexterous manipulation policies via human-robot co-training. Released in 2025 under the MIT license, it contains human hand demonstrations and corresponding robot data collected across 93 diverse real-world environments including kitchens, workshops, offices, and outdoor settings. The dataset focuses on 6-degrees-of-freedom dexterous hand manipulation with egocentric video and proprioceptive data. DexWild addresses the critical gap between lab-setting manipulation datasets and the full diversity of real-world manipulation contexts.
Who is it for?
DexWild is designed for researchers working on dexterous manipulation policies, human-robot co-training, and in-the-wild generalisation. It is particularly valuable for teams developing policies that must work across varied environmental conditions and object types not seen during training. The MIT license makes it suitable for both academic research and commercial robotics development.
Key specifications
- Environments: 93 diverse real-world settings
- Data type: Human hand demonstrations + corresponding robot data
- Task categories: Dexterous manipulation, human-robot interaction
- View type: Egocentric
- Format: ImageFolder, Parquet (Hugging Face auto-converted)
- License: MIT — commercial use permitted
- Access: Open — freely downloadable via Hugging Face
- Origin: United States (Carnegie Mellon University)
How it compares
DexWild is distinguished by its environmental diversity (93 environments vs typical lab datasets with 1-5 settings) and its human-robot co-training methodology. ALOHA and ACT datasets cover bimanual manipulation but in controlled lab settings. DexUMI (Stanford) covers a similar dexterous manipulation focus using the Inspire Hand specifically. DexWild's broader environmental scope makes it more useful for generalisation research, while DexUMI is more appropriate for Inspire Hand-specific fine-tuning.
Limitations and access notes
DexWild uses a 6-DoF dexterous hand rather than a specific named commercial robot platform, which may limit direct transfer to particular hardware. Episode count from primary sources is unconfirmed — the preview viewer shows partial data only. The ImageFolder format requires conversion for use with standard robot learning pipelines. MIT license permits unrestricted commercial use.
Linked professions
Frequently asked questions
What makes DexWild different from other manipulation datasets?
DexWild covers 93 diverse real-world environments — far more than typical lab-setting datasets. It uses a human-robot co-training approach where human hand demonstrations are collected alongside corresponding robot demonstrations, enabling policies that generalise to novel environments and objects not seen during training.
Can DexWild be used commercially?
Yes. DexWild is licensed under MIT, which permits commercial use, modification, and distribution with no restrictions beyond preserving the license notice.
How do I access DexWild?
DexWild is available on Hugging Face at huggingface.co/datasets/boardd/dexwild-dataset and via the project GitHub at github.com/dexwild/dexwild. No registration is required.
What type of robot hand does DexWild use?
DexWild uses a 6-degrees-of-freedom dexterous hand. The dataset is not tied to a single named commercial platform, making it more broadly applicable to research on dexterous manipulation across different hardware.
How does DexWild compare to DexUMI?
Both DexWild and DexUMI focus on dexterous manipulation, but with different emphases. DexWild prioritises environmental diversity across 93 real-world settings using human-robot co-training. DexUMI from Stanford REAL Lab targets the Inspire Hand specifically and focuses on using the human hand directly as a data collection interface. DexWild is stronger for generalisation research; DexUMI is more specific to the Inspire Hand platform.