The authors built RL-Kirigami, a system that uses reinforcement learning to design kirigami (paper-cutting patterns) that fold into target 3D shapes. The challenge is that valid designs must satisfy hard geometric constraints (no overlaps, compatible ratios) that aren't differentiable. They combine optimal-transport conditional flow matching (a generative model) with Group Relative Policy Optimization to align the generator with non-differentiable rewards for shape matching and feasibility. The system produces designs that can be laser-cut and physically fabricated in about 8 minutes per part.
Main takeaways:
- Inverse design problem: given a target 3D shape, find a 2D cutting pattern that folds into it
- Uses flow matching to generate candidate designs, then RL (GRPO) to optimize for shape accuracy, geometric feasibility, and smoothness
- A single sample from the pretrained model achieves 94.2% shape match, outperforming solver baselines while using far fewer simulator evaluations
- Adding RL improves accuracy to 94.91% and makes designs smoother (lower total variation)
- Generated designs were successfully laser-cut and physically deployed as working prototypes