YUBI: Yielding Universal Bidigital Interface for Bimanual Dexterous Manipulation at Scale

Researchers introduce YUBI, a finger-aligned gripper that improves upon existing data collection systems for robotic manipulation by enabling more ergonomic, intuitive bimanual control. The team released an unprecedented 8,434-hour dataset across 1.20M episodes and demonstrated that policies trained on YUBI data transfer successfully across multiple robot platforms, advancing the development of robotic foundation models.

YUBI represents a significant engineering advancement in robotic data collection infrastructure, addressing practical limitations that have constrained progress in dexterous manipulation research. The finger-driven design eliminates the ergonomic friction of pistol-grip systems, enabling researchers to collect higher-quality data more efficiently during extended sessions. This matters because large-scale, high-fidelity datasets form the foundation upon which modern robotic foundation models are built, and collection bottlenecks directly limit the pace of capability development.

The breakthrough demonstrates a clear trend: as robotics researchers pursue foundation models analogous to large language models, they increasingly recognize that data quality and collection ergonomics directly impact model performance. YUBI's dataset—8,434 hours across 1.20M episodes—represents one of the largest curated robotic manipulation datasets available, a critical resource for the field. The VR-based 6 DoF tracking ensures high-fidelity trajectory capture that previous systems struggled to achieve consistently.

The cross-platform transferability of trained policies across UR, Franka, and ELEY robots validates YUBI's claim that its collected data directly transfers as policy supervision. This eliminates robot-specific data silos and suggests the gripper captures manipulation principles at a generalizable level. The open-source release of hardware, software, and dataset creates network effects within the research community, potentially accelerating the timeline for robotic foundation models.

Looking ahead, this work establishes YUBI as infrastructure likely to influence downstream robotics development. The generalization results suggest future foundation models trained on this data could achieve broader applicability, reducing the need for custom task-specific training. Monitoring adoption patterns within robotics research labs will indicate whether this infrastructure becomes a standard component of large-scale robotic learning pipelines.

• →YUBI's finger-aligned ergonomic design enables collection of 8,434 hours of high-fidelity robotic manipulation data across 1.20M episodes.
• →Policies trained on YUBI data transfer directly across multiple robot platforms without robot-specific retraining, validating cross-platform generalization.
• →The open-source release of hardware, software, and dataset removes infrastructure barriers for large-scale robotic learning research.
• →VR-based 6 DoF tracking ensures high-fidelity trajectory acquisition superior to previous handheld data collection systems.
• →This work establishes foundational infrastructure for developing robotic foundation models at scale comparable to language model training.