GPD-ROS-Docker image

A docker container for using Grasp Pose Detection in combination with ROS (melodic)

Based on pcl docker by DLopezMadrid

Adapted for GPD-ROS use (GPD Lib , ROS Wrapper)

Overview

This image is based on Ubuntu 18.04 and has the following packages installed:

• CUDA-10.2.89
• nvidia-docker
• CMake-3.10.2
• VTK-8.2.0
• PCL-1.11.2
• OpenCV-4.5.2
• GPD-2.0.0
• Eigen
• Flann
• Boost
• gdb
• ROS Melodic (Desktop Full)
• terminator

Build the image

To build the GPD image, run:

./build_image.sh

Creating & starting a container

GPD Container

To start a GPD container, run (NOT using sudo):

./gpd_setup_container.sh

If you get an error like (you probably ran the setup with sudo earlier):

touch: setting times of '/tmp/.docker.xauth-n': Permission denied
# or
xauth:  unable to link authority file /tmp/.docker.xauth, use /tmp/.docker.xauth-n

Please first delete the file (that was just created), using:

sudo rm /tmp/.docker.xauth 

Then you will probably enter into bash into the container you just created. From there you can for example start terminator. You can also exit whenever you want and start a new container using:

./gpd_start_container.sh

This should start a terminator window.

Test GPD installation

cd /home/gpd/gpd/build, then

./detect_grasps ../cfg/eigen_params.cfg ../tutorials/krylon.pcd

PAL Container

To start a PAL container, run (NOT using sudo):

./pal_setup_container.sh

Open up a terminator window by running the container start script:

./pal_start_container.sh

By default, ROS isn't sourced in this container yet. To source ROS for every new startup window run:

echo "source /opt/pal/ferrum/setup.bash" >> ~/.bashrc

To only source it in the current window, run:

source /opt/pal/ferrum/setup.bash

GPD ROS Wrapper

This image comes with the full ros desktop installation. The example ros ws is found in /home/gpd/docker_dir/ws_gpd_tiago. If ROS commands aren't found, be sure to source the /ros_entrypoint.sh, or run source "/opt/ros/$ROS_DISTRO/setup.bash" (alternatively, put this into the .bashrc).

This workspace contains:

• A modified version of the gpd ROS Wrapper adopted for use with the TIAGo robot.
• the PCL TIAGo tutorial from the PAL tiago tutorials repo
• the robot_description to be able to visualize the robot in rviz.

Modifying Config Files

Adopted from ROS Wrapper for GPD:
First, you need to modify the config file in your gpd folder (the one that is used is defined in the gpd_tiago.launch file): /home/gpd/docker_dir/gpd/cfg/ros_eigen_params.cfg. Search for parameters that have absolute file paths and change them to actual paths on your system.

In ros_eigen_params.cfg, that means editing L32 to:

weights_file = /home/gpd/docker_dir/gpd/models/lenet/15channels/params/

Starting the GPD ROS Pacakge

Find the workspace (/home/gpd/docker_dir/ws_gpd_tiago), build it with catkin_make (might need to build multiple times as it auto-generates some of the header files). source the workspace, and run the gpd package using:

roslaunch gpd_ros gpd_tiago.launch

Start TIAGo gazebo sim

If used in combination with a TIAGo docker image, start a gazebo simulation in its docker container:

launch tiago_gazebo with the world: Good options are objects_on_table, and tabletop_cube, and with model steel

roslaunch tiago_134_gazebo tiago_gazebo.launch public_sim:=true robot:=steel world:=tabletop_cube

Note: For the tabletop_cube world, you'll likely need to move the cube back slightly so it is captured in the pointcloud.

To look down at the objects:
Load up the pcl motions parameters into the play_motion namespace (in the gpd-ros container)

rosparam load `rospack find tiago_pcl_tutorial`/config/pcl_motions.yaml

• To check if they are loaded, find them using rosparam list | grep look_down

In the TIAGo Docker container:

rosrun actionlib axclient.py /play_motion

and run the look_down motion.

motion_name: 'look_down'
skip_planning: True
priority: 0

PCL Options

The topic from which the point cloud is read is defined in the gpd_tiago launch file.

You can either feed the complete pointcloud to GPD package, make sure to set the cloud_topic line

    <param name="cloud_topic" value="/xtion/depth_registered/points" />

Alternatively, we can modify what point cloud data is fed to GPD. The TIAGo PCL Tutorial package provides some examples on how to do this.

Table segmentation

We can filter out the table, start the launch file:

roslaunch tiago_pcl_tutorial segment_table.launch

• To prevent rviz from loading, add show_rviz:=false

The points published on /segment_table/nonplane are the objects that we found on the table. To feed this data to the gpd package, modify the gpd_tiago.launch cloud_topic param to look for this topic.

Cylinder Detector

If we specifically want to pick up a cylinder shape, launch the launch file:

roslaunch tiago_pcl_tutorial cylinder_detector.launch

• To prevent rviz from loading, add show_rviz:=false

There are some topics published under the /cylinder_detector namespace, which are also visualized in rviz. To feed this data to the gpd package, modify the gpd_tiago.launch cloud_topic param to look for this topic.

Region Segmentation

We can segment the point cloud into regions, so that we can specify a certain region to feed to gpd. Launch the launch file:

roslaunch tiago_pcl_tutorial pcl_region.launch

• To prevent rviz from loading, add show_rviz:=false

Also start a rqt_reconfigure window. In this window you can play with the parameters untill you get the desired results.

Find the topic of the region you want to use and feed this data to the gpd package, modify the gpd_tiago.launch cloud_topic param to look for this topic.