Paper: "YUBI: Yielding Universal Bidigital Interface for Bimanual Dexterous Manipulation at Scale" — arXiv 2606.10244 (Jun 8 2026) Authors: Takehiko Ohkawa, Jumpei Arima, Yuki Noguchi, … Floris Erich, Yukiyasu Domae, Tatsuya Matsushima, Kei Ota (19 authors) · AIST / University of Tokyo et al. (Japan consortium) What it is: a handheld, finger-driven gripper (a UMI successor) that captures bimanual manipulation data at massive scale and deploys without retargeting — an L3 capture interface that sidesteps the L4 bridge. See the Dexterous-Hand Data Pyramid.

Companions: Dexterous-Hand Data Pyramid · DexEXO (wearable-exo counterpart) · DexUMI · Do As I Do · Dexterous Manipulation.

1. A finger-aligned handheld gripper, not a glove. YUBI uses "yielding, finger-driven actuation that directly maps human finger movements to gripper jaw motion," paired with VR-based 6-DoF tracking of the gripper for high-fidelity trajectories — an ergonomic alternative to bulky pistol-grip UMI designs.
2. Deploy without retargeting. The same gripper is mounted on each robot (UR, Franka, ELEY), so data collected by hand transfers directly — no human→robot hand mapping (no L4 bridge).
3. Massive scale. The released dataset is 8,434 hours across 1.20M episodes and 119 tasks — one of the largest dexterous-manipulation corpora, and a single policy transfers across multiple bimanual robots.
4. Bidigital, by design. "Bidigital" = a two-finger interface — YUBI trades full five-finger dexterity for scale, ergonomics, and zero-retargeting deployment.

• It maximizes the L3 tier by collapsing L4. Because the capture device is the deployed end-effector, YUBI removes the retargeting-fidelity ceiling the data pyramid §4 flags — the tradeoff being that the "hand" is a 2-finger gripper, not a 5-finger hand.
• Scale-first UMI lineage. Where DexUMI pushes a sensorized-glove route and Do As I Do a device-free video route, YUBI pushes the handheld-gripper route to 8,434 h — a reproducible, open path to large-scale bimanual data.
• Direct evidence for "interface = embodiment." A single policy transferring across UR/Franka/ELEY confirms that fixing the interface (not the arm) is enough for cross-robot transfer — a deployment-side cross-embodiment result (cf. Single-Checkpoint Multi-Robot).

flowchart LR
  F[Human fingers] -->|yielding, finger-driven actuation| G[YUBI handheld gripper<br/>finger motion → jaw motion]
  VR[VR 6-DoF tracking] --> G
  G -->|collect handheld| D[Bimanual dataset<br/>8,434 h · 1.20M ep · 119 tasks]
  G -->|mount on robot, no retarget| R[UR / Franka / ELEY]
  D --> P[Single policy] --> R

• Actuation: yielding, finger-driven — human finger movement maps directly to the gripper jaw; finger-aligned ergonomics vs pistol-grip.
• Tracking: VR-based 6-DoF pose of the gripper for trajectory capture.
• Transfer: mount-on-robot → no retargeting; the same physical gripper collects human data and executes on the robot.

• Dataset: 8,434 hours · 1.20M episodes · 119 tasks (bimanual).
• Cross-robot: a single policy trained on the YUBI dataset transfers across multiple bimanual robots (UR, Franka, ELEY).
• Vs UMI: claimed advantages in versatility for complex bimanual tasks, dexterity, and operational efficiency over the UMI gripper.

Significance. YUBI is the scale champion of the handheld-interface branch and a clean demonstration that fixing the interface removes the retargeting problem — the most direct answer to the pyramid's L4 bottleneck, at the cost of finger count.

Limitations.

1. Bidigital (2-finger) — not a five-finger hand; in-hand reorientation and high-DoF dexterity are out of scope. For a humanoid 5-finger target, YUBI is a scale-and-recipe reference, not a same-morphology data source.
2. No tactile/force capture noted; grounding relies on vision + gripper pose.
3. The paper emphasizes the interface + dataset over an explicit limitations analysis; downstream 5-finger transfer is not its claim.

• Paper: arXiv 2606.10244
• Pyramid placement: L3 handheld capture, L4 bypassed (mount-on-robot) — Dexterous-Hand Data Pyramid
• Branch siblings: DexUMI (glove) · DexEXO (exoskeleton) · Do As I Do (video) · AnyDexRT (retargeting)
• Dexterous Manipulation · Cross-Embodiment

← Back to Reviews · Home