RynnWorld-Latent Enables Cross-Embodiment Video World Modeling
October 26, 2026
RynnWorld-Latent is a video world model that uses 3D-grounded latent actions to enable joint pretraining across human videos, robot demonstrations, and simulations. It leverages a pretrained video transformer and a shared conditioning interface to overcome incompatible action spaces.
HOW THIS AFFECTS YOU
●
builderThis enables training more robust world models using heterogeneous datasets from different robotic platforms.
●
researcherThis provides a path toward universal latent action spaces in multimodal world models.