3DSGrasp

IEEE ICRA 2023 - 3DSGrasp: 3D Shape-Completion for Robotic Grasp [ Youtube Video ]

We present a grasping strategy, named 3DSGrasp, that predicts the missing geometry from the partial Point-Cloud data (PCD) to produce reliable grasp poses. Our proposed PCD completion network is a Transformer-based encoder-decoder network with an Offset-Attention layer. Our network is inherently invariant to the object pose and point's permutation, which generates PCDs that are geometrically consistent and completed properly. Experiments on a wide range of partial PCD show that 3DSGrasp outperforms the best state-of-the-art method on PCD completion tasks and largely improves the grasping success rate in real-world scenarios.

💻 Quick Start

⤵️ After installing follow the appropriate instructions if you want to:

run the full pipeline (from camera depth input to kinova grasping the object) ➡️ Full Pipeline
run only the completion network to generate shape completion on a partial.pcd ➡️ Completion Network
run only GPD to generate grasps for PCD of either partial.pcd or complete.pcd ➡️ Test GPD
🚋 use our model? ➡️ Completion Network
🚦 use the same train-test split of the YCB dataset? ➡️ Completion Network

🔑 Installations

To begin, clone this repository locally

git clone git@github.com:NunoDuarte/3DSGrasp.git
$ export 3DSG_ROOT=$(pwd)/3DSGrasp

This repo was tested on Ubuntu 20.04 and with ROS Noetic

🔑 Install requirements for Completion Network:

$ cd $3DSG_ROOT
$ conda create -n 3dsg_venv python=3.8  # or use virtualenv
$ conda activate 3dsg_venv
$ sh install.sh

🔑 Install requirements for full Pipeline

install ROS, moveit, ros_kortex, and ros_kortex_vision

Check oficial documentation for ROS
Check oficial documentation for Moveit or just do this
Check oficial documentation for ROS_kortex and ROS_kortex_vision

install gpd

Check oficial documentation for GPD (:warning: gpd repo was tested on Ubuntu 16.04; if you have trouble installing on Ubuntu 20.04 send an issue to us and we'll help)

replace files src/gpd/grasp_detector.cpp src/gpd/util/plot.cpp include/util/plot.h with the ones provides in folder GPD on this repo
add the correct path of your pipeline

  ofstream fw("'PATH_TO_THE_DIR/gpd_grasp_scores.txt", std::ofstream::out);

compile GPD again

🔑 Install requirements to test GPD (see grasps generated of your partial.pc or complete.pc)

Check oficial documentation for GPD (:warning: gpd repo was tested on Ubuntu 16.04; if you have trouble installing on Ubuntu 20.04 send an issue to us and we'll help)

📄 Step by step to run 3DSGrasp Pipeline

Open terminals:

Open a terminal and start kinova gen3 ros

source catkin/devel/setup.bash
roslaunch kortex_driver kortex_driver.launch

(optional) open our rviz environment in 3DSG_ROOT/ROS/rviz/. It includes our table for moveit

open config file grasp_kinova.rviz

Open a new terminal and start kinova gen 3 intel sense camera ros package

source catkin/devel/setup.bash
roslaunch kinova_vision kinova_vision_rgbd.launch device:=$IP_KINOVA

Place an object on top of the table and in front of the robot. The object should be visible from rviz using the Point Cloud data
Open a new terminal and run main pipeline with GPD

source catkin_ws/devel/setup.bash
cd 3DSG_ROOT/
python main_gpd.py

Windows will pop out showing the progress of segmentation, the completion network, merging the completion output in the scene, and the gpd output on the completion output
After main_gpd.py finishes a complete_final_frame appears on rviz. It should be pointing at the object.
Open a new terminal and run the ros node to grasp the object

source catkin/devel/setup.bash
roslaunch kortex_examples reach_approach_grasp_pose.launch

The node asks first if you want to move the robot to the initial position. If not type "n", otherwise it will bring you to a predefined initial position. You can change this initial position by following the steps in How to set the initial position of Kinova
After the first question, it then asks if you want to execute the grasp pose. Type "y" to execute or "n" to quit. If you press "y" moveit should try to find a feasible path to control the robot to the desired grasp pose. It it fails you should see an output in the terminal of kortex_driver stating No solution was found . This could be for various reasons: out of reach for the robot; there is an obstacle in the robot's path, or moveit timeout before finding a solution. You could try to change to another grasp pose by changing the index of the gpd output in main_gpd.py:

#######  ---------------------------// ---------------------------------- #####
# change grasp [k]
# pose = coord_to_transform(grasp_poses[0])
k = grasp_poses[0]
pose = coord_to_transform(k)
approach = approach_coord_to_transform(k)

#######  ---------------------------// ---------------------------------- #####

You can pick from 0-5 grasp poses generated by GPD. 0 - best quality score to 5 - worst quality score.

Set initial position on ros kinova gen3

set an initial pose for kinova manually or (optional) set it as a .npy file and load it in reach_approach_grasp_pose.py

cd 3DSG_ROOT/ROS/src/

(optional) open reach_approach_grasp_pose.py and set location of initial_state.npy of kinova

        # Load initial state of robot (joint angles)
        initial_state = np.load('location_of_initial_state.npy')

set location of final_pose.npy and final_approach.npy (These are the best grasp and approach from GPD)

        print('Load grasp pose')
        final_pose = np.load('location_of_final_pose.npy')
        final_approach = np.load('location_of_final_approach.npy')

ℹ️ Information:

When grasping the closure of the gripper is predefine, if you want to change open reach_approach_grasp_pose.py and change variable

approach.example_send_gripper_command(0.3)

For acquiring the point cloud and segmenting it run:

python main_agile.py

it saves the acquired_point cloud as original_pc and the segmented as partial_pc in tmp_data

📄 GPD for Point Cloud

To run GPD on .pcd file

cd $GPD_ROOT/build
./detect_grasps ../cfg/eigen_params.cfg $LOCATION_OF_FILE.PCD

Completion Network

🎉 The pre-trained model is here (around 500 MB)!!!

To train point cloud completion model

cd Completion
python3 main.py --config  ./cfgs/YCB_models/SGrasp.yaml

To test point cloud completion model

python3 main.py  --test --ckpts /PATH_TO_pre_trained_MODEL/MODEL.pth --config  ./cfgs/YCB_models/SGrasp.yaml

Note that the input of the network for the real-world grasping experiment should be a single sample

Citation

If you find this code useful in your research, please consider citing our paper. Available on IEEE Xplore and ArXiv:

@INPROCEEDINGS{10160350,
  author={Mohammadi, Seyed S. and Duarte, Nuno F. and Dimou, Dimitrios and Wang, Yiming and Taiana, Matteo and Morerio, Pietro and Dehban, Atabak and Moreno, Plinio and Bernardino, Alexandre and Del Bue, Alessio and Santos-Victor, José},
  booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)}, 
  title={3DSGrasp: 3D Shape-Completion for Robotic Grasp}, 
  year={2023},
  volume={},
  number={},
  pages={3815-3822},
  doi={10.1109/ICRA48891.2023.10160350}
}

NunoDuarte / 3DSGrasp