KITE transfers robot task knowledge across hands zero-shot, hitting 7/10 on real hardware

qinzytech · x · 2026-10-08

KITE separates task decisions from motor control so robot skills transfer across grippers and dexterous hands without new demonstrations.

A shared policy predicts where and in what direction to engage an object; per-robot decoders, trained on synthetic data from each robot's geometry and joint models, turn intent into motor commands.
In simulation, cube-picking transferred from a Robotiq gripper to Barrett, Allegro, and Wuji hands at 100% success over 100 unseen episodes each; keyboard pressing and bottle pumping were also tested.
On a physical Wuji hand: 7/10 cube-picking (gripper source), 8/10 keyboard-pressing and 7/10 bottle-pumping (human-hand source), with no target-task demonstrations or real-world fine-tuning.
Caveats: "zero-shot" still uses the target robot's kinematic model, and transfer depends on whether that robot can achieve the requested contacts.

More from Embodied