RL-Kirigami: AI designs and laser-cuts deployable prototypes in 8 minutes

Kirigami—the art of cutting and folding paper—is gaining traction as a fabrication method for shape-programmable metamaterials. However, inverse design is notoriously difficult because the deployment is nonlinear, and cut layouts must satisfy discrete compatibility rules, avoid overlap, and map precisely to target shapes. Existing methods require hundreds of forward simulator evaluations and often produce infeasible designs.

In a new paper on arXiv, Milad Yazdani and colleagues present RL-Kirigami, which marries optimal-transport conditional flow matching (OT-CFM) with reinforcement learning. The OT-CFM prior generates compatible ratio fields for parallelogram quad kirigami, while a marching decoder enforces geometric compatibility. Then Group Relative Policy Optimization (GRPO) fine-tunes the generator using nondifferentiable rewards for silhouette accuracy, feasibility, and regularity. Results show a single sample from the prior reaches 94.2% sIoU—surpassing solver baselines and reducing simulator calls from hundreds to just 1. With GRPO, accuracy climbs to 94.91%, and adding a regularity term cuts the total variation of the ratio field from 0.95 to 0.81 while maintaining 94.83% sIoU. The team exported designs to DXF and laser-cut them in 50μm polymeric sheets, producing deployable prototypes in 8.0 ± 1.0 minutes per part. This work establishes a manufacturing-aware inverse design pipeline for kirigami metamaterials under hard geometric constraints.