DiffTactile: A Physics-based Differentiable Tactile Simulator for Contact-rich Robotic Manipulation

This package provides a tactile simulator with differential physics for contact-rich manipulation tasks . It models soft tactile sensors, objects with various material properties, and contact between sensors and objects. For more information, please refer to the project webpage or corresponding paper.

Installation

You can create a Conda environment for this simulator:

conda create -n difftactile python=3.9.16
conda activate difftactile

And install the package with its dependencies using

git clone https://github.com/DiffTactile/DiffTactile.git
cd DiffTactile
pip install -r requirements.txt -e .

File Structure

• meshes includes all object and sensor mesh models.
• object_model includes soft (elastic, elastoplastic), rigid, multi-material, and cable object models.
• sensor_model includes FEM tactile sensor model and parallel-jaw gripper model.
• tasks includes gradient-based skill learning for manipulation tasks.
• baseline includes baseline methods (CMA-ES, PPO, SAC, RNN) implementation.

Usage example

Under the tasks, run

python box_open.py --use_state --use_tactile

• use_state means using state rewards
• use_tactile means using tactile rewards

This will optimize the trajectory of the box opening task with differential physics.

License

This project is licensed under MIT license, as found in the LICENSE file.

Citating DiffTactile

If you use DiffTactile in your research, please cite

@inproceedings{
si2024difftactile,
title={{DIFFTACTILE}: A Physics-based Differentiable Tactile Simulator for Contact-rich Robotic Manipulation},
author={Zilin Si and Gu Zhang and Qingwei Ben and Branden Romero and Zhou Xian and Chao Liu and Chuang Gan},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=eJHnSg783t}
}