Robotic-grasping-cornell

In this project, Deep Convolutional Neural Networks (DCNNs) is used to simultanously detect a grasping point and angle of an object, so that a robot arm can pick the object. In general, this is the implementation of a small model of the model presented in this paper https://arxiv.org/abs/1802.00520. We do not implement the Grasp Proposal Networks, so instead of using 2-stage DCNNs, we use one-stage DCNNs.

Platform

• python 3.6.8
• pytorch 1.0.0

Codes

1. Data preprocessing
2. Training
3. Demo

1. Data preprocessing

• Download Cornell Dataset
• Run dataPreprocessingTest_fasterrcnn_split.m (please modify paths according to your structure)

2. Training

$ python3 train.py --epochs 100 --lr 0.0001 --batch_size 8

3. Demo

• Download the pretrained model GoogleDrive
• Put in the folder ./models
• Run demo:

  $ python3 demo.py

Acknowledgment

This repo borrows some of code from https://github.com/ivalab/grasp_multiObject_multiGrasp