HRG_Net

To reduce the spatial dimensional inaccuracy due to upsampling in the traditional CNN framework, we develop a novel grasping visual architecture referred to as High resolution grasp nerual network (HRG-Net), a parallel-branch structure that always maintains a high-resolution representation and repeatedly exchanges information across resolutions.

This repository contains the implementation of the High Resolution Grasp Nerual Network(HRG-Net) from the paper: A Robotic Visual Grasping Design: Rethinking Convolution Neural Network with High-Resolutions

arxiv|video

Contact

Any questions or comments contact mail.

Installation

conda create -n hrgnet python==3.8
conda activate hrgnet
pip install -r requirements.txt / conda env create -f environment.yaml

Datasets

Cornell | Jacquard | multiobject(Just for test)

Training

python train_hrgnet.py

Evaluation/Visualisation

For validation and visualization purposes, we provide our previously trained model

python evaluation_grasp.py    # For Cornell and Jacquard dataset

python evaluation_heatmap.py  # For Cornell and Jacquard dataset

python multi_grasp_visualization.py # For multiobject dataset

Acknowledgement

Code heavily inspired and modified from https://github.com/dougsm/ggcnn. The code for the experiments related to the robot in the physical environment will be released later