VisualPatchWorld: Code World Models as Latent Structured Representations for Planning
VisualPatchWorld turns learned dynamics into inspectable code that can be used directly for planning.
The paper says VPW probes a task to choose a qualitative dynamics form, then fits its parameters from state-action traces. Those programs can be rolled forward like simulators, inspected as source, and used in model-predictive control. In comparisons with earlier code-based world models, it reports 69.0% mean planning success, 23.5 points above the strongest code baseline. It comes closest to ground-truth engines on navigation and grasp-heavy control, while contact-rich pushing still needs engine checks to close most of the gap. HF Daily Papers' note
The paper says VPW probes a task to choose a qualitative dynamics form, then fits its parameters from state-action traces. Those programs can be rolled forward like simulators, inspected as source, and used in model-predictive control. In comparisons with earlier code-based world models, it reports 69.0% mean planning success, 23.5 points above the strongest code baseline. It comes closest to ground-truth engines on navigation and grasp-heavy control, while contact-rich pushing still needs engine checks to close most of the gap. HF Daily Papers' note
score 5