RL-Kirigami

Code for Reinforcement learning for inverse structural design and rapid laser cutting of kirigami prototypes. RL-Kirigami performs inverse design for compact reconfigurable parallelogram quad kirigami: an OT-CFM generator proposes ratio fields conditioned on a target deployed silhouette, and GRPO aligns the generator to non-differentiable rewards (silhouette match, feasibility, ratio-field regularity). Decoded layouts can be exported as DXF for laser cutting.

Setup

pip install -e .   # optional: installs the CLI wrappers declared in pyproject.toml

The three entry points below can also be invoked directly as python <script>.py.

1. Generate the dataset

python -m data_generator.generator --config configs/data_generator.yaml

Outputs:

• data_generator/kirigami_x_dataset.pkl - dataset pickle
• data_generator/preview.png - sample grid
• data_generator/gifs/ - per-sample deployment animations

A prebuilt 5000 / 500 / 500 split is available on Google Drive. If the pickle is not at the default path, set data.pickle_path in configs/training.yaml.

2. Train the OT-CFM prior

python fm_training.py --config_path configs/training.yaml --resume last

Outputs:

• checkpoints/<run_name>/ - checkpoints
• checkpoints/tb/ - TensorBoard logs

Use --resume last to continue from the most recent checkpoint, or omit it to start fresh.

3. RL fine-tune with GRPO

python rl_training.py --config_path configs/training.yaml --init_from last --resume last

--init_from last loads the latest OT-CFM checkpoint as the starting policy. RL checkpoints are written to checkpoints/<run_name>_RL/.

Configuration

File / block	What it controls
configs/data_generator.yaml	grid size, mask resolution, split sizes, x range, sampler, seed
configs/training.yaml -> model_config	backbone and tensor shapes
configs/training.yaml -> data	dataset and generator references
configs/training.yaml -> training	shared training settings (optimizer, batches, epochs)
configs/training.yaml -> rl_training	GRPO-only overrides (group size, reward weights, temperature)

Keep the two YAML files consistent: training.yaml reads grid_rows, grid_cols, x_min, x_max, and the mask size from data_generator.yaml.

Cite us

If you use this code, please cite:

@misc{yazdani2026rlkirigami,
  title={Reinforcement learning for inverse structural design and rapid laser cutting of kirigami prototypes},
  author={Yazdani, Milad and Shalileh, Shahriar and Shahriari, Dena},
  year={2026},
  eprint={2605.08098},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  doi={10.48550/arXiv.2605.08098},
  url={https://arxiv.org/abs/2605.08098}
}