Reproduce “FoldingAgent: Inferring Parametric Origami Procedures from Demonstration Videos” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Maya Moriya, Sigal Raab, Yael Vinker, Tali Dekel
FoldingAgent, developed by researchers at Weizmann Institute of Science and MIT, infers explicit, parametric origami folding procedures directly from demonstration videos. The framework successfully translates video keyframes into executable action sequences, achieving significantly higher compilation validity (96%) and geometric similarity compared to ablated baselines, and completing 100% of sequences on a new PurelandFold dataset.
Key insights
- Combining a VLM with structured computational tools and a simulator enables robust inference of complex procedural programs directly from unstructured visual demonstrations.
- An explicit parametric action space and a detailed state representation that accounts for layer ordering are crucial for generating physically valid and precise origami folds.
- Agentic self-correction capabilities, particularly the ability to roll back to earlier states, are essential for mitigating compounding errors in long, sequential folding tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.