Doppel open-sourced: one video world model unifies 8 robot embodiments

mangahomanga · x · 2026-11-19

CMU, Johns Hopkins, and Toyota Research Institute have released and fully open-sourced Doppel, an action-conditioned video world model spanning 8 embodiments — robot arms, humanoids, human hands, and full bodies.

Its key idea is factorization: instead of re-learning how a body moves, Doppel renders the embodiment along the commanded trajectory from the camera's viewpoint, then conditions a video model on both the rendering and the action vector, so the model focuses on how the scene responds.

Code and weights are on GitHub (DoppelWM/Doppel, MIT license).

Related event: Doppel: one open video world model for eight robot embodiments(2 posts)→

Original post →

More from Embodied

Embodied channel →