FoldingAgent: Inferring Parametric Origami Procedures from Demonstration Videos
Maya Moriya, Sigal Raab, Yael Vinker, Tali Dekel
cs.CV · cs.AI · cs.GR
Quick verdict
First system inferring executable Pureland folding programs from origami videos; strong agentic gains, but small self-built benchmark, thin baselines.
Summary
FoldingAgent converts instructional origami videos into executable, parametric folding programs. Focusing on Pureland origami — single mountain/valley creases plus unfold, rotate, and flip — the method represents each paper state as a FOLD-format planar graph extended with face orientation and plane-based layer ordering, and exposes five primitives (add_vertex, fold, unfold, rotate, flip) through a custom simulator. A pretrained VLM (Gemini 3.1 Pro Preview) acts as an agent equipped with tools to view frames, render states, save and restore checkpoints, and consult a separate VLM critic that compares rendered states against target photos; bounded exploration (five attempts per keyframe, 300 tool calls) with a fallback selector keeps runs tractable. The authors also curate PurelandFold, 27 self-captured sequences (5–21 keyframes) annotated with ground-truth geometry and actions. FoldingAgent completes 100% of sequences with 96% compilation validity, reaching TSS 0.90, GS 0.58, CS 0.77, FFS 0.39 and CPD 189, versus 85% completion and clearly worse similarity for a non-agentic action-prediction ablation; naive crease-pattern prediction compiles only 8% of the time. Gains hold across GPT-5.5, Claude-Opus-4.8 and Gemini backbones, and users preferred its renders in 84%/80% of 600 judgments. Limitations include a small single-source self-annotated benchmark, evaluation restricted to completed sequences, no statistical testing, an unexplained regression when the critic is added non-agentically, and reliance on a proprietary preview model with no stated code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- FoldingAgent is the first framework to translate unstructured origami demonstration keyframes into executable procedural folding programs (Pureland scope).
- The parametric state/action representation is the primary driver of output validity: compilation validity rises from 8% (naive CP prediction) to 96% (VLM-S, Table 1).
- The full agentic loop with external critic, checkpoints and rollback outperforms non-agentic variants on all reported similarity/dissimilarity metrics (e.g., TSS 0.90, CPD 189 vs 216–268).
- The framework's gains generalize across GPT-5.5, Claude-Opus-4.8 and Gemini-3.1-Pro backbones (Table 2: e.g., GPT similarity 0.37→0.73).
Strengths
New task with a clean, well-scoped formulation
Inferring executable procedural programs from video keyframes is genuinely new: the paper distinguishes it from before/after crease prediction (Kato et al. 2025) and CP-to-sequence generation (Learn2Fold, Huang et al. 2026), and the Pureland restriction (Sec. 4.2) keeps the action vocabulary to five verifiable primitives.
Representation choice is validated, not just asserted
Table 1 isolates the representation: naive CP prediction (VLM-CP) compiles only 8% of steps, while the same VLM working in the proposed state/action space (VLM-S) reaches 96% CV — the single most convincing number in the paper. The plane-based layer ordering (Sec. 4.1, App. B.3) replacing FOLD faceOrders with a sortable DAG is a sensible engineering contribution.
Unusual transparency
Backbone generalization and an informative negative result
Weaknesses
No external baseline on any metric
Evaluation protocol favors the full method
Small, single-source, self-annotated ground truth
Statistical and internal-anomaly reporting gaps
Reproducibility hinges on a proprietary preview model
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper poses a new task — converting instructional origami videos into executable, parametric folding programs — scoped to Pureland origami (single mountain/valley creases plus unfold, rotate, flip). Four contributions are claimed: the task; a parametric action space and simulator; an agentic VLM framework with critic verification, checkpoints and rollback; and the PurelandFold benchmark (27 sequences with ground-truth geometry and actions). The novelty claim is scoped carefully and is defensible: Kato et al. [2025] handles before/after image pairs in a controlled setting, and Learn2Fold [2026] generates sequences from a given crease pattern — the inverse direction. Both agent and critic run on Gemini 3.1 Pro Preview; keyframes are manually extracted, so the pipeline is video→program only after human segmentation.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does FoldingAgent do?
It takes keyframes from an instructional origami video and outputs an executable sequence of parametric folding actions (fold, unfold, rotate, flip, and precise vertex placement) that reproduces every intermediate paper state in a custom simulator. The scope is Pureland origami, where every fold is a single mountain or valley crease along a straight line.
What is the PurelandFold benchmark?
A dataset of 27 origami folding sequences filmed by the authors (iPhone 12, black Bristol board, ring light) with manually annotated ground-truth geometry and folding actions for each keyframe. Sequences average 12 frames (range 5–21) and are based on the Easy Origami instructions on the OrigamiWay website.
How accurate is FoldingAgent and what does it cost to run?
On PurelandFold it completes 100% of sequences with 96% of outputs passing the Flat-Folder compiler, and it beats its own non-agentic ablations on all similarity metrics (crease-pattern dissimilarity 189 versus 216–268; naive crease-pattern prediction compiles only 8% of the time). Each sequence averages about 140 VLM queries, roughly $8.9, 4M input tokens, and 31 minutes.
Is code or data available for FoldingAgent?
The paper publishes full prompts, controller/simulator/critic design details, and a project page linking the dataset and complete agent traces, but it does not state a code release. The system also depends on the proprietary Gemini 3.1 Pro Preview model, which is a deprecation and reproducibility risk.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure images (Figs. 1–15) are not present in the extracted text, so all qualitative assessments here rest on captions and the quantitative tables only.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.