This project aimed to leverage a deep learning model to predict the robotic arm grasp poses from a depth image scene. According to the objects' spatial information, the deep learning model can output the grasp position, grasp angle, and gripper width to finish the grasp work. The GGCNN 1 was implemented in this project since it is a simple and widely used model..
The goal of this project is not to create novel models or algorithms but to familiarize yourself with the entire workflow of network training, environment construction, and grasping based on the predicted results in the simulated environment.
-
Install anaconda
-
Create and activate a conda virtual environment with python3.7.
sudo apt update conda create -n env_name python=3.7 conda activate env_name
-
Download this repository.
git clone https://github.com/BoceHu/vision_based_grasp.git cd vision_based_grasp
-
Install PyTorch (Please choose the suitable version according to the CUDA version and the system.)
conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia
-
Install OpenCV
conda install -c conda-forge opencv
-
Install other requirement packages
conda install -r requirements.txt
In this project, I used the Cornell Dataset to train our model. The relabeled dataset 2 can be found Here.