ThinShellLab: Thin-Shell Object Manipulations With Differentiable Physics Simulations

---

Official repo for the paper:

Thin-Shell Object Manipulations With Differentiable Physics Simulations

ThinShellLab is a fully differentiable simulation platform tailored for robotic interactions with diverse thin-shell materials.

Installation

You can create a Conda environment for this simulator first:

conda create -n thinshelllab python=3.9.16
conda activate thinshelllab

And install the package with its dependencies using

git clone https://github.com/wangyian-me/thinshelllab.git
cd thinshelllab
pip install -e .

Render

• Here are two ways to render our scene, Taichi GGUI and LuisaRender Script. Taichi GGUI renders real-time image in GUI windows with low resolution, and LuisaRender Script generates meta-data script files for high-resolution and more realistic rendering outputs. This can be specified using the option --render_option.
• To run LuisaRender Script, necessary assets should be loaded. Run git submodule update --init --recursive to load the submodule AssetLoader and run export PYTHONPATH=$PYTHONPATH:${PWD}/data/AssetLoader to add the asset path to PYTHONPATH.
• For seeing the rendering results of LuisaRender Script, you should setup LuisaRender and use the command `` to get the outputs.

Usage example

We put running scripts under code/scripts, you can simply run

cd thinshelllab
cd code
sh scripts/run_trajopt_folding.sh

to train a trajectory optimization policy for the folding task, or use other scripts to train on different tasks.

Citation

If you find this codebase/paper useful for your research, please consider citing:

@inproceedings{wang2023thin,
  title={Thin-Shell Object Manipulations With Differentiable Physics Simulations},
  author={Wang, Yian and Zheng, Juntian and Chen, Zhehuan and Xian, Zhou and Zhang, Gu and Liu, Chao and Gan, Chuang},
  booktitle={The Twelfth International Conference on Learning Representations},
  year={2023}
}