This repository contains the implementation of the Robotic Manipulation Network (RoManNet) which is a vision-based model architecture to learn the action-value functions and predict manipulation action candidates. More details can be found in the following paper:

Sulabh Kumra, Shirin Joshi, Ferat Sahin

Xplore | arxiv

If you use this project in your research or wish to refer to the baseline results published in the paper, please use the following BibTeX entry:

@article{kumra2022learning, 
author={Kumra, Sulabh and Joshi, Shirin and Sahin, Ferat}, 
journal={IEEE Robotics and Automation Letters},  
title={Learning Robotic Manipulation Tasks via Task Progress Based Gaussian Reward and Loss Adjusted Exploration},   
year={2022},  
volume={7},  
number={1},  
pages={534-541},  
doi={10.1109/LRA.2021.3129833}
}

• Checkout the repository

$ git clone https://github.com/skumra/romannet.git

• Create a virtual environment

$ python3.6 -m venv --system-site-packages venv

• Source the virtual environment

$ source venv/bin/activate

• Install the requirements

$ cd romannet
$ pip install -r requirements.txt

• Install the CoppeliaSim simulation environment

• Run CoppeliaSim (navigate to your CoppeliaSim directory and run ./sim.sh). From the main menu, select File > Open scene..., and open the file romannet/simulation/simulation.ttt from this repository.

• In another terminal window, run the following:

python main.py <optional args>

Note: Various training/testing options can be modified or toggled on/off with different flags (run python main.py -h to see all options)

Some parts of the code and simulation environment has been borrowed from andyzeng/visual-pushing-grasping for fair comparison of our work in simulation.