For world-model pretraining, DreamDojo calculates latent proxy actions from human frame pairs. During target-robot adaptation, DreamDojo resets the action conditioner and grounds controls with target-robot data. Distillation then trains an autoregressive student.
The author made this diagram from the source figure. Human transitions supply proxy actions. Target-robot adaptation grounds the command interface before autoregressive distillation. The training phases are different. The playback is an example. It does not show measured training time or generated-video performance. Open full-size SVG