DreamX Uses Video Diffusion for Universal Latent Action Learning
November 2, 2026
DreamX repurposes pretrained video diffusion models to extract compact latent actions by compressing video frames 300x. This method captures spatiotemporal features while remaining invariant to embodiment and appearance changes.
HOW THIS AFFECTS YOU
●
builderThis provides a way to extract efficient action tokens from video without training from scratch.
●
researcherYou can leverage generative priors to learn action representations that generalize across different robot embodiments.