This repository is implementations of both training and prediction of Grasp Quality CNN (GQ-CNN) with Dexnet3.0 dataset using Pytorch modules.
For more information, please visit original project website and the paper
- Project Website
- Paper: Dex-Net 3.0: Computing Robust Robot Suction Grasp Targets using a New Analytic Model and Deep Learning, Mahler et al., ICRA 2018
This repository features:
-
dsets.py - Script for pre-fetching dexnet3.0 dataset onto RAM, split train/validation sets and more.
-
model.py - Grasp Quality CNN model consists of torch.nn module.
-
training.py - Run this script to train your model. The default options are as below.
- The number of images: number_of_files x 1000 = 2,760,000 images.
- Learning rate: 0.001
- Momemtum: 0.99
- Epochs : 25
- Batch size: 64
-
predict.py - Prediction with the trained GQ-CNN.
To fetch full dexnet3.0 dataset, there should be at least 20GB free space on your RAM. If you want to train with a fewer images, reduce 'number_of_files' in the training.py
- You can download dexnet3.0 dataset in the project website dexnet3.0 dataset
- Using GQ-CNN trained with the Dexnet3.0 dataset, stable suction grasping point on a object can be determined. (Left) Predicted candidates. (Right) Final suction point.
