Recovering hand-object interactions from 3D point motion, retargeted to Allegro robot hand

mangahomanga · x · 2026-09-17

A research thread demonstrates using estimated 3D point motion to recover joint types and trajectories of everyday object interactions — laptop screens rotating, drawers sliding, chairs turning — and reconstructing the objects involved.

The key challenge is aligning recovered hand motion with reconstructed objects, since small scale or pose errors break contact. The team also shows the recovered human interactions can be retargeted to an Allegro robot hand in MuJoCo simulation, pointing toward learning robot manipulation from video.

Related event: JHU's Track, Articulate, Act Rebuilds Articulated Objects from Monocular Video for Robots(5 posts)→

Original post →

More from Embodied

Embodied channel →