FastUMI

A hardware-independent handheld-gripper data-collection system and dataset from Shanghai AI Lab, improving on UMI.

FastUMI was proposed in September 2024, led by the Shanghai Artificial Intelligence Laboratory together with Shanghai Jiao Tong University, Fudan University, the University of Hong Kong, and others, as an improvement on UMI (Universal Manipulation Interface: a person holds a camera-equipped gripper to perform demonstrations, and the data is then transferred to a robot). The original UMI computes the gripper's pose by running visual-inertial odometry on GoPro footage, which is a fairly complex pipeline; FastUMI instead uses an off-the-shelf RealSense T265 tracking module to output pose directly at 200Hz, leaving the GoPro fisheye camera to handle only video capture, and adds standardized, swappable fingertips and camera mounts so the same device works across different robot arms and grippers. The team also open-sourced a dataset of 22 everyday tasks with over 10,000 trajectories; in October 2025 they released FastUMI-100K, covering 54 household tasks with more than 100,000 trajectories.

ExampleA collector holds FastUMI and performs a single pick-and-place demonstration in a real home; the T265 records the gripper's trajectory, the GoPro records the video, and the resulting demonstration can be used to train any robot arm fitted with the matching fingertip.

Also called

FastUMI: A Scalable and Hardware-Independent Universal Manipulation Interface, FastUMI-100K

Related

Universal Manipulation Interface · Handheld Gripper Data Collection · Robot-free (Embodiment-free) Data Collection · AgileX Pika · Demonstration Data · Visual-Inertial Odometry

Sources

FastUMI: A Scalable and Hardware-Independent Universal Manipulation Interface with Dataset (arXiv)
FastUMI-100K: Advancing Data-driven Robotic Manipulation with a Large-scale UMI-style Dataset (arXiv)

As of

2025-10