DexUMI — Dexterous Universal Manipulation Interface

DexUMI is a dexterous manipulation dataset developed by the Stanford University REAL Lab using the human hand as a universal interface for collecting robot training data. Released in 2025, it uses the Inspire Hand as the target robot platform and captures human hand demonstrations that are retargeted to the robot's kinematic structure. The dataset covers dexterous manipulation tasks including cube picking and kitchen manipulation tasks with tactile and force sensing data, stored in Zarr format on the Stanford REAL Lab server. The Inspire Hand is a priority embodiment in the Geppetto dataset catalogue. DexUMI represents an important methodological contribution: by using the human hand directly as the data collection interface rather than a teleoperation device, it dramatically increases demonstration diversity and naturalness. A 519 MB sample dataset is freely available for download. The full dataset is in the dataset directory subdirectory of the project server. Episode count and total hours are unconfirmed from primary sources.

Dataset specifications
Year2025
EmbodimentsInspire Hand
Modalitiesrgb, depth, proprioception, tactile
View typeegocentric
Task categoriesdexterous-manipulation
Data formatzarr
LicenseMIT — confirm in dataset directory before activation
Accessopen — commercial use permitted
MaintainerStanford University REAL Lab
Origin countryUS

What is it?

DexUMI (Dexterous Universal Manipulation Interface) is a dexterous manipulation dataset developed by the Stanford University REAL Lab that uses the human hand as a universal interface for collecting robot training data. Released in 2025 under MIT license, it targets the Inspire Hand as the robot platform and captures human hand demonstrations retargeted to the robot's kinematic structure. The dataset covers dexterous manipulation tasks including cube picking and kitchen manipulation with tactile and force sensing data, stored in Zarr format. A 519 MB sample dataset is freely available; the full dataset is accessible from the Stanford REAL Lab project server.

Who is it for?

DexUMI is designed for researchers working on dexterous manipulation with the Inspire Hand platform, and for those interested in the Universal Manipulation Interface (UMI) methodology — using human hand demonstrations as the primary data collection mechanism rather than teleoperation devices. It is valuable for researchers studying tactile sensing integration and fine-motor manipulation tasks requiring 6-DoF finger dexterity.

Key specifications

How it compares

DexUMI is methodologically closest to UMI (Universal Manipulation Interface) from Columbia/Stanford, using the human hand directly as the collection device. DexWild (CMU) covers more environments (93) but uses a different dexterous hand. The Inspire Hand is a notable embodiment — it appears in several supply chain datasets as a key dexterous hand component for humanoid robots. DexUMI provides one of the only open datasets specifically for this platform.

Limitations and access notes

Episode count and total hours are unconfirmed from primary sources — figures should be treated as estimates. The full dataset requires access via the Stanford REAL Lab project server rather than a standard repository like Hugging Face. The Zarr format is well-supported in the robot learning ecosystem (compatible with LeRobot and standard Python tooling). MIT license permits unrestricted commercial use.

Linked professions

Frequently asked questions

What is the Universal Manipulation Interface (UMI) method?

UMI is a data collection methodology that uses the human hand directly as the manipulation interface rather than a robot teleoperation device. Human demonstrations are recorded and then retargeted to the robot's kinematic structure, providing more natural and diverse demonstrations than teleoperation.

What robot platform does DexUMI target?

DexUMI targets the Inspire Hand, a 6-degrees-of-freedom dexterous robotic hand used in humanoid robot development. The Inspire Hand appears in several humanoid robot supply chains as a key dexterous manipulation component.

How do I access DexUMI?

A 519 MB sample dataset is freely available at real.stanford.edu/dexumi. The full dataset is accessible from the dataset directory subdirectory of the Stanford REAL Lab project server. No Hugging Face or GitHub download is available for the full dataset.

Can DexUMI be used commercially?

Yes. DexUMI is licensed under MIT, which permits commercial use, modification, and redistribution with no restrictions beyond preserving the license notice.

What tasks does DexUMI cover?

DexUMI covers dexterous manipulation tasks including cube picking and kitchen manipulation tasks requiring fine motor control. All tasks require 6-DoF finger dexterity and use the Inspire Hand as the target platform.