IndusGrasp

This is the official PyTorch Implementation of IndusGrasp: A Novel Robotic Grasp Detection Method Using Synthetic Data for Disordered Industrial Scenarios

Video

• The video is available in YouTube

The pipeline:

• The Comparison studies in the real-world environment
  

Requirements: Software

sudo apt-get install libglfw3-dev libglfw3  

sudo apt-get install libassimp-dev   

pip install --pre --upgrade PyOpenGL PyOpenGL_accelerate 

pip install  cython cyglfw3 pyassimp==3.3 imgaug progressbar

pip install  -r requments.txt`

Generate dataset

• Using the tools/pybullet_dataset/pybullet_dataset.py to create render poses with your own 3D mesh(xx.urdf)

• Using the tools/creat_grasp.py to create grasp representation with your own 3D mesh(xx.ply)

• Using the dataset_generate.py to create your dataset

Train the IndusGrasp

python train.py

Test the IndusGrasp

python test.py

Comparison Studies

• Our Method in dense scenarios

• GR-CNN, in sparse scenarios

• GraspNet, which doesn't work for industrial parts

Generalization Studies

• Our method can generate to unseen objects with different shapes and textures (Note: the network is only trained by Synthetic Data generated by object 1. )