GR-Dexter

A ByteDance Seed VLA from late 2025 for bimanual dexterous hands, built together with its own hand hardware and teleoperation system.

GR-Dexter was released as a technical report by ByteDance's Seed team in late December 2025. Most VLA (vision-language-action) models only control a two-finger gripper; switching to a pair of high-DOF dexterous hands enlarges the action space, the hands frequently occlude the object, and real-robot data becomes more expensive too. GR-Dexter tackles hardware, data, and model together: a self-developed, 21-degree-of-freedom, linkage-driven ByteDexter V2 dexterous hand with piezoresistive tactile sensors in the fingertips, mounted on two Franka Research 3 arms; data collected through bimanual teleoperation using a Meta Quest headset plus Manus data gloves; and a model that carries over GR-3's hybrid Transformer structure at about 4 billion parameters, trained on a mix of teleoperation trajectories, vision-language data, curated cross-embodiment data, and more than 800 hours of human hand-trajectory data.

ExampleOn the long-horizon “organize cosmetics” task, success rate is 0.97 under standard layouts and still 0.89 across 5 unseen layouts; on a general pick-and-place task with 23 unseen objects, success rate is 0.85.

Also called

ByteDance Seed VLA for bimanual high-DoF dexterous-hand robots

Related

Dexterous Manipulation · Bimanual Manipulation · Dexterous Hand · Seed GR-3 · Vision-Language-Action Model · Data Glove

Sources

GR-Dexter Technical Report (arXiv 2512.24210)
GR-Dexter 技术报告 HTML 全文 (Chinese)

As of

2025-12