SynManDex project page is live

SynManDex is a synthetic data-generation pipeline for human-like dexterous robot grasp synthesis. The key design choice is to treat human pre-grasps as semantic seeds, not as executable robot grasps.

The pipeline separates the mechanism:

An object-conditioned diffusion model samples MANO pre-grasp proposals.
GeoRT-calibrated retargeting transfers those proposals to the robot embodiment.
Robot-native optimization enforces contact, collision, and joint constraints.
Dynamic rollout validation converts accepted keyframes into policy demonstrations.

h_{MANO} \rightarrow \hat{q}_{XHand} \rightarrow q^*_{XHand}=\arg\min_q E_{contact}+E_{penetration}+E_{joint}+E_{self}

The human prior proposes intent; the robot objective decides executability.

SynManDex system figure — SynManDex grounds human-functional priors into robot-native bimanual dexterous grasps.

SynManDex rollout strips — Trajectory strips make the validation claim testable: accepted grasps must survive dynamic execution.

SynManDex hardware rollout

Zero-shot hardware validation on a bimanual UR5e-XHand system.

SynManDex project page assets main/public/SynManDex

Read the full project page: SynManDex.