Video
what the camera sees
kept as memory
Past
Future
the world-modeling target
State
Action
what the hand does
Language
another instruction,
another future
Flow matching
joint denoising
One sequence
language
image
video
action
Mixture of experts