==========================================================
== PACKAGE: haf_grasping ==
==========================================================
Author: David Fischinger, Vienna University of Technology
Version: 1.0
Date: 15.5.2015
HAF_GRASPING
is calculating grasp points for unknown and known objects represented by the surface point cloud data.
For scientific foundation see:
D. Fischinger, M. Vincze: "Learning Grasps for Unknown Objects in Cluttered Scenes", IEEE International Conference on Robotics and Automation (ICRA), 2013.
<a href="files/ICRA2013.pdf">[pdf]</a>
D. Fischinger, A. Weiss, M. Vincze: "Learning Grasps with Topographic Features", The International Journal of Robotics Research.
In a first step the point cloud is read from a ROS topic and a heightsgrid is created.
For each 14x14 square of the hightsgrid a featurevector is created.
Using SVM with an existing model file, it is predicted if the center of the square is a good
grasping point. For good grasping points the coordinates and the direction of the approach vectors
are published.
DOWNLOAD CODE
>> git
HOW TO USE HAF_GRASPING - GET STARTED
Start calculation server (does the work), haf_client (small programming incl. class that shows how to use haf_grasping) and a visualization in rviz:
>> roslaunch haf_grasping haf_grasping_all.launch
Publish the path of a point cloud to calculate grasp points on this object with the gripper approaching direction along the z-axis:
>> rostopic pub /haf_grasping/input_pcd_rcs_path std_msgs/String "$(rospack find haf_grasping)/data/pcd2.pcd" -1
(Alternatively, publish a point cloud at the ros topic: /haf_grasping/depth_registered/single_cloud/points_in_lcs)
EXPLANATION FOR THE RVIZ VISUALIZATION
RVIZ will now visualize the point cloud with corresponding frame (blue indicates the z-axis).
Bigger rectangle: indicates the area where heights can be used for grasp calculation
Inner rectangle: defines the area where grasps (grasp centers) are searched.
Long red line: indicates the closing direction (for a two finger gripper)
Red/green spots: indicate the positions where grasps are really tested for the current gripper roll (ignoring points where no calculation is needed, e.g. no data there)
Green bars: indicate where possible grasps were found. The height of the bars indicate an grasp evaluation score (the higher the better)
Black arrow: indicates the best grasp position found and the approching direction (for a parallel two finger gripper)
HAF-GRASPING CLIENT - CODE EXPLAINDED
In calc_grasppoints_action_client.cpp we subscribe to a point_cloud topic and start the following callback when a point cloud comes in:
== code start ==
//get goal (input point cloud) for grasp calculation, send it to grasp action server and receive result
void CCalcGrasppointsClient::get_grasp_cb(const sensor_msgs::PointCloud2ConstPtr& pc_in)
{
ROS_INFO("\nFrom calc_grasppoints_action_client: point cloud received");
// create the action client
// true causes the client to spin its own thread
actionlib::SimpleActionClient<haf_grasping::CalcGraspPointsServerAction> ac("calc_grasppoints_svm_action_server", true);
ROS_INFO("Waiting for action server to start.");
// wait for the action server to start
ac.waitForServer(); //will wait for infinite time
ROS_INFO("Action server started, sending goal.");
// send a goal to the action
haf_grasping::CalcGraspPointsServerGoal goal;
goal.graspinput.input_pc = *pc_in;
goal.graspinput.grasp_area_center = this->graspsearchcenter;
// set size of grasp search area
goal.graspinput.grasp_area_length_x = this->grasp_search_size_x+14;
goal.graspinput.grasp_area_length_y = this->grasp_search_size_y+14;
// set max grasp calculation time
goal.graspinput.max_calculation_time = this->grasp_calculation_time_max;
//send goal
ac.sendGoal(goal);
//wait for the action to return
bool finished_before_timeout = ac.waitForResult(ros::Duration(50.0));
if (finished_before_timeout)
{
actionlib::SimpleClientGoalState state = ac.getState();
boost::shared_ptr<const haf_grasping::CalcGraspPointsServerResult_<std::allocator<void> > > result = ac.getResult();
ROS_INFO("Result: %s", (*(result)).result.data.c_str());
ROS_INFO("Action finished: %s",state.toString().c_str());
}
else
ROS_INFO("Action did not finish before the time out.");
}
== code end ==
PARAMETER SETTING
There are a number of parameters that can be set (directly or via a service call). Set grasp search center (in m) to (x=0.1,y=0):
Grasp_center: the x-,y-position, that is the center of the area where grasps are searched.
Service call to change it:
>> rosservice call /haf_grasping/set_grasp_center "graspsearchcenter:
x: 0.10
y: 0.0
z: 0.0"
Grasp_area_size: the size of the area were grasps should be detected. Set rectangle to 16x10 centimeter:
>> rosservice call /haf_grasping/set_grasp_search_area_size "grasp_search_size_x: 16
grasp_search_size_y: 10"
Grasp_calculation_time_max: maximal time in seconds until a grasp has to be returned. Set max timt to 3 sec:
>> rosservice call /haf_grasping/set_grasp_calculation_time_max "max_calculation_time:
secs: 3
nsecs: 0"
== Input ==
A point cloud from objects
== Output ==
Grasp points and approach vectors which are detected using Support Vector Machines
(at the beginning the approach vectors are parallel to the z-axis)
== LIBSVM ==
We have included LIBSVM to work as our classifier (go to folder libsvm-3.12 and type "make" after checking out):
Chih-Chung Chang and Chih-Jen Lin, LIBSVM : a library for support
vector machines. ACM Transactions on Intelligent Systems and
Technology, 2:27:1--27:27, 2011. Software available at
http://www.csie.ntu.edu.tw/~cjlin/libsvm