Learning visual-based deformable object rearrangement with local graph neural networks

This repository is the code of the paper "Learning visual-based deformable object rearrangement with local graph neural networks", published in Complex & Intelligent SystemsL.

If you find this code useful in your research, please consider citing:

@article{deng2023learning,
  title={Learning visual-based deformable object rearrangement with local graph neural networks},
  author={Deng, Yuhong and Wang, Xueqian and Chen, Lipeng},
  journal={Complex \& Intelligent Systems},
  pages={1--14},
  year={2023},
  publisher={Springer}
}

Table of Contents

• Installation
• Generate Data
• Train Model
• Evaluate Model

Installation

This simulation environment is based on pybullet. You can follow the instructions in DeformableRavens to setup the simulator.

1. Clone this repository.

2. Install the following packages in the created conda environment:

   • pytorch and torchvision: `
   • tensorflow:
   • tensorflow_addons
   • pybullet:
   • torch_geometric

Generate Data

• Run the following script to generate demonstrations for model training:

  python collect.py -eposide_num 1100 -task_type "rope_line"
  

  where -task_type specifies the deformable object rearrangement task type (rope_line|rope_l|rope_v|rope_n|ring_circle|ring_square|ring_move|cloth_fold|cloth_flatten|cloth_fold_a)

  

Train Model

• Train Model:

  python main.py --mode train -task_type "rope_line"
  

  python main_multi.py --mode train 
  

  where main.py represent single task learning fashion, main_multi.pyrepresent multi-task learning fashion

Evaluate Model

• Evaluate Model:

  python main.py --mode test -task_type "rope_line"
  

  python main_multi.py --mode test