COrigami: AI pipeline for co-designing flat-foldable origami

COrigami, an end-to-end AI-driven pipeline for co-designing flat-foldable, visually recognisable origami, was submitted to arXiv on 24 Jun 2026 as arXiv:2606.26299. The paper, authored by Tom Zahavy and 17 other authors, presents a system that generates crease patterns from natural language and refines them through algorithmic optimisation and reinforcement learning.

What is COrigami and what does it do?

COrigami is an AI pipeline that assists artists and designers by producing mathematically valid, flat-foldable crease patterns from natural-language inputs and then refining those patterns with an autonomous aesthetic critique. The system produces a semantic stick figure, computes a base packing, solves for a flat-foldable crease pattern, shapes the flat-folded crease pattern, and refines the generated model using reinforcement learning driven by an autonomous aesthetic evaluation loop.

The paper frames computational origami as a domain with strict geometric constraints and subjective visual objectives. COrigami aims to bridge those objectives by combining algorithmic optimisation for physical correctness with automated aesthetic scoring, so the output acts as a structural starting point human artists can further expand and shape.

How does the pipeline generate and refine origami?

COrigami runs a multi-stage pipeline: first it converts text to a semantic stick figure, then computes a base packing, solves for a flat-foldable crease pattern, shapes the folded crease pattern, and finally applies reinforcement learning guided by an autonomous aesthetic evaluator. The pipeline order is explicit in the paper and underpins the system's ability to meet both geometric and visual goals.

Each stage tackles a distinct constraint. The semantic stick figure encodes the user intent; base packing arranges paper regions to support that intent; the crease-pattern solver enforces flat-foldability; shaping adapts the folded surface toward the target appearance; and reinforcement learning iteratively improves models using an autonomous loop that evaluates aesthetics. The authors position the system as a collaborative assistant that generates structural starting points rather than finished artworks.

Why does this matter?

COrigami demonstrates that an AI system can satisfy multi-objective physical constraints while addressing subjective aesthetics within a mathematically rigid domain. By integrating geometric solvers with an autonomous aesthetic critic and reinforcement learning, the pipeline shows a path for AI to produce reliable, mathematically grounded artifacts that still respond to human visual goals. That combination is notable because prior generative systems often target either verifiable solutions or stylistic output, not both simultaneously.

The paper underscores the value of a hybrid approach: algorithmic optimisation ensures flat foldability, while autonomous aesthetic evaluation guides creative refinement. The authors argue this supports co-creativity, supplying human practitioners with structurally valid starting points they can further shape.

What to watch

Look for code, data, and demos linked from the paper: the arXiv entry includes sections labelled "Code, Data and Media Associated with this Article" and toggles for demo platforms such as Replicate and Hugging Face Spaces. The arXiv identifier is arXiv:2606.26299 and the DOI link listed is https://doi.org/10.48550/arXiv.2606.26299, which will help locate any released artifacts.

If the authors publish implementation artifacts or interactive demos, those will reveal how effectively the autonomous aesthetic loop and reinforcement learning produce visually recognisable results that remain flat-foldable. The next milestone to watch is any public code or demo release tied to the arXiv entry that demonstrates the pipeline end to end.

References and source details: the submission was posted to arXiv on 24 Jun 2026 and lists Tom Zahavy and 17 other authors as contributors to the work.