UMI (Universal Manipulation Interface)
The Universal Manipulation Interface (UMI) is an open-source manipulation data-collection interface in the form of a handheld parallel-jaw gripper equipped with a GoPro camera. It was designed by researchers from Stanford, Columbia and the Toyota Research Institute (Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, Shuran Song).
An operator holds the gripper in hand and performs manipulation tasks while the system records video, inertial data and gripper state. The collected in-the-wild demonstrations are used to train policies (e.g. Diffusion Policy) that are then deployed on physical robot arms (including UR5/UR5e and Franka).
The gripper is 3D-printed (95A TPU) with soft fingers; the GoPro Hero 9 uses a 155-degree fisheye lens, and a pair of mirrors in the camera peripheral view creates implicit stereo. GoPro’s built-in IMU (accelerometer and gyroscope) supports trajectory reconstruction (ORB-SLAM3). Presented at RSS 2024 (arXiv:2402.10329) and released under the MIT license.
Classification
Embodied data-collection rig · serves as: Manipulation, Grasping, Research and Education.
Which group UMI (Universal Manipulation Interface) belongs to and how it is built
A subcategory covering portable or wearable platforms for industrial-scale collection of multimodal demonstration data (vision, touch, audio, action trajectories), often “embodiment-free”, used to train robot policies without the robot itself.
A component type: a complete hardware rig for embodied data collection — combining cameras (first-person view), tactile and audio sensors and UMI-standard grippers, with frame-level synchronization. Enables fast, low-cost collection of human demonstrations without a robot.
Technical specification
Basic physical properties of UMI (Universal Manipulation Interface) — dimensions, weight and materials