FoldingAgent: Inferring Parametric Origami Procedures from Demonstration Videos

FoldingAgent is a new agentic framework that infers explicit parametric origami procedures directly from demonstration videos. Instead of predicting a static crease pattern, it watches a folding sequence, reasons about the geometry, and outputs an executable folding plan that can be checked against physical constraints.

The system combines a pre-trained vision-language model with specialized tools for simulating geometric transitions, verifying plausibility, retrieving and comparing visual content, and scoring its own predictions. That tool use matters: the agent can re-plan as it goes, which helps reduce the compounding errors that usually show up in multi-step visual reasoning tasks.

The work also introduces PurelandFold, a curated benchmark of Pureland origami videos with ground-truth geometry and action labels. The benchmark gives the approach a concrete testbed for comparing unstructured demonstrations against structured, parametric plans. For developers, the broader takeaway is that video-to-procedure extraction is getting closer to something you can actually build around, not just classify.

While this is an origami-focused system, the underlying pattern is relevant to game tech: sequential inference, tool-assisted validation, and physically plausible reconstruction are all ideas that map cleanly onto animation tooling, procedural content generation, and authoring assistants. If this class of system matures, it could reduce the gap between human-made demonstrations and machine-readable pipelines.