How UMI turns handheld grippers and SLAM into robot training data, in 5 steps

gabriberton · x · 2026-11-03

A five-post thread walks through Universal Manipulation Interface (UMI), the low-cost data collection method many robotics startups build on:

Human demos with handheld grippers: hold two grippers (each with a camera and IMU), perform tasks like washing dishes, touching objects only with the grippers;
Trajectory recovery: IMU alone is too noisy, so use SLAM fusing video and IMU (tracking error under 1 cm / 4°);
Deployment: attach the grippers to any robotic arms, train a model to predict gripper trajectories, and use those predictions to drive the arms;
Scale: collect lots of data across many tasks.

The paper figures show UMI is a portable, intuitive, low-cost framework transferring diverse human demonstrations to visuomotor policies, covering dynamic, precise, bimanual and long-horizon tasks across embodiments (6DoF/7DoF arms).

Original post →

How UMI turns handheld grippers and SLAM into robot training data, in 5 steps

More from Embodied