YUBI gripper collects 1.2M robot training episodes across 119 tasks
Finger-driven gripper enables 8,434 hours of bimanual manipulation data at scale.
A team of 19 researchers from OMRON SINIC X and other institutions has released YUBI (Yielding Universal Bidigital Interface), a novel gripper designed to make bimanual dexterous manipulation data collection more ergonomic and scalable. Traditional handheld systems like the Universal Manipulation Interface (UMI) use bulky pistol-grip designs that limit fine-grained tasks. YUBI replaces this with a finger-driven, yielding mechanism that directly maps human finger movements to gripper jaw motion, enabling intuitive and fatigue-free operation.
Using two YUBI devices (one per hand) with VR-based 6 DoF tracking, the team amassed a massive dataset: 8,434 hours spanning 1.2 million episodes across 119 distinct tasks. Remarkably, a single imitation learning policy trained on YUBI data transferred zero-shot to multiple bimanual robot platforms—UR, Franka, and ELEY—simply by mounting the gripper. The project is fully open-sourced, including hardware designs, data-collection software, and the dataset, providing a reproducible path for the robotics community to scale training data for foundation models.
- YUBI uses finger-driven actuation instead of pistol grip, improving ergonomics for bimanual tasks
- Dataset includes 8,434 hours, 1.2M episodes, and 119 tasks, an order of magnitude larger than prior UMI datasets
- Single policy trained on YUBI data transfers directly to UR, Franka, and ELEY robots without additional tuning
Why It Matters
YUBI's open-source stack could dramatically lower the barrier to collecting high-quality robot training data at scale.