DexSeed is a new framework that transforms existing human hand‑object datasets into a wide range of physically executable robot demonstrations. This allows robots to learn dexterous manipulation from cheap, pre‑recorded human data instead of relying on costly teleoperation.

The core idea is to keep the local spatial relationships between hand and object that are present in the human recordings. DexSeed uses an InteractionMesh representation to encode these relationships and then retargets them to the robot hand. It also adapts the hand to different object shapes and to new pick‑and‑place locations. A reinforcement‑learning finger‑control policy then turns the retargeted sequences into grippable robot rollouts, while a pre‑computed arm trajectory keeps the robot’s arm motion realistic.

When the retargeted interactions were evaluated against three other popular methods—Dex‑Retargeting, SPIDER, and Wuji—DexSeed showed the best performance. In terms of matching the hand‑object contacts it achieved a 3.326 mm contact error and a 7.636° angular error, while the other methods were significantly higher on both metrics. Penetration depth, a measure of how much the hand intrudes into the object, was similar across all methods (DexSeed 3.371 mm, best competitor 3.418 mm).

DexSeed currently covers eight common grasp modes and focuses on pick‑and‑place tasks. Future work will need to extend the approach to in‑hand manipulation, tasks that involve both hands, and integration of tactile sensors to allow the robot to detect slip and adjust its grip during execution.

These results show that DexSeed can produce high‑quality, diverse robot demonstrations cheaply, which could accelerate learning of dexterous vision‑language‑action models.