---

📖 Table of Contents

• 📖 Table of Contents
• 📍 Overview
• :key: Key Features
  • Allsight Sim
  • SightGAN
• 📂 Contents
• 🚀 Getting Started
  • 🔧 Installation
  • 🤖 Running allsight_sim2real
• 📄 License
• 👏 Acknowledgments
• :clipboard: BibTeX

---

📍 Overview

This repository contains the code and resources related to the paper "Augmenting Tactile Simulators with Real-like and Zero-Shot Capabilities" (under review). The paper introduces SightGAN, a cutting-edge solution for enhancing tactile perception in high-resolution tactile sensors such as allsight sensor.

---

:key: Key Features

Allsight Sim

The simulation data was obtained using tha allsight_sim package that utilizing TACTO a physics-engine simulator for optical-based tactile sensors.

For more information about the simulation package please see the link provided above.

• Example for data collection session

SightGAN

A bi-directional Generative Adversarial Network which built upon CycleGAN and designed to bridge the reality gap between simulated and real tactile data, particularly for high-resolution tactile sensors.

SightGAN introduces contact-specific consistency losses:

• Spatial Contact Consistency loss
• Pixel-wise Contact Region Consistency loss

The GAN diagram and training process can be represented as follow:

• Example of SightGAN integrated in the simulation

---

📂 Contents

This repository contains several directories with the details as follow:

Root

File	Summary
requirements.txt	Dependencies file
train.py	Train the GAN model
test.py	Test the GAN model
train_regressor.py	Train the regressor model only
train_regressor_finetune.py	Train a pre-trained regressor model with additional data

Options

Python files forked from CycleGAN repo

Models

File	Summary
diff_cycle_gan_model.py	The class of SightGAN model with its auxilliary losses

All the rest of the files were forked from CycleGAN and helps for building and managing the models during training and test procedures.

Data_process

File	Summary
merge_json_sim.py	Merging all the dataset from simulation into main json file database
sim2gan_json.py	Updata the json file with generated image pathes
filter_real_images.py	Filtering the real data and creating a json file database
transfer_images.py	Transfering iamges from sim and real databases to the train/test folders
add_diff_frame.py	update the json file with the substracted images

Train_allsight_regressor

A Python package for training the spatial contact estimator and inherit the models.

Util

Python files forked from CycleGAN repo

---

🚀 Getting Started

Dependencies

Please ensure you have the following dependencies installed on your system.

Project was tested on:

• Ubuntu 18/20
• python >= 3.8

🔧 Installation

1. Clone the allsight_sim2real repository:

git clone https://github.com/RobLab-Allsight/allsight_sim2real

2. Change to the project directory:

cd allsight_sim2real

3. Install the dependencies:

pip install -r requirements.txt

🤖 Running allsight_sim2real

NOTE: Be aware for the pathes adjusts and adaptation for your custom dataset.

NOTE: Please Be aware for the arguments adjusts needed for every script.

Assuming you have a dataset folder and updated package with the relevant pathes to your dataset:

Filter the real data:

python3 data_process/filter_real_images.py 

Merge sim data:

python3 data_process/merge_json_sim.py 

Add diff frame:

python3 data_process/add_diff_frame.py

Transfer images to gan folders:

python3 data_process/transfer_images.py 

Train CycleGAN:

python3 train.py <arguments> 

---

📄 License

This repository is licensed under the MIT License. Feel free to use, modify, and distribute the code as per the terms of this license.

---

👏 Acknowledgments

• Our code is inspired by CyclGAN.
• Simulation was implemented using TACTO.

---

:clipboard: BibTeX

@misc{azulay2023augmenting,
      title={Augmenting Tactile Simulators with Real-like and Zero-Shot Capabilities}, 
      author={Osher Azulay and Alon Mizrahi and Nimrod Curtis and Avishai Sintov},
      year={2023},
      eprint={2309.10409},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}

↑ Return

---