VisualPatchWorld: Code World Models as Latent Structured Representations for Planning

Researchers introduced VisualPatchWorld, a code-based world model for planning from scene graphs.

It infers a qualitative dynamics form from short probes, then fits parameters from state-action traces.

The resulting programs support planning and replanning across control tasks like navigation, grasping, and pushing.

VisualPatchWorld reached 69.0% mean planning success, beating the best prior code baseline by 23.5 points and approaching ground-truth engine performance on several tasks.