Doppel world model renders the robot body and beats action-only baselines across 8 embodiments

mangahomanga · x · 2026-11-19

Doppel factorizes video world modeling: instead of learning motion from raw action vectors, it renders the robot along its planned motion from the camera's viewpoint and conditions one shared video model on those renders plus actions. One model trained on 600 hours across 8 datasets (robot arms, bimanual, humanoid, human hands/bodies) beats action-only baselines by 1.00 dB PSNR on unseen trajectories and 1.20 dB out of distribution; trained on just three arm datasets it matches the baseline with 63.8% fewer training draws, and lifts moving-object PSNR from 21.50 to 24.85 dB on a held-out kitchen scene. Missing camera poses are recovered from videos.

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

Original post →

More from Embodied

Embodied channel →