-
First copy/move the
sensor_stick
folder to the~/catkin_ws/src
directory of your active ros workspace. -
Make sure you have all the dependencies resolved by using the
rosdep install
tool and runningcatkin_make
:
$ cd ~/catkin_ws
$ rosdep install --from-paths src --ignore-src --rosdistro=kinetic -y
$ catkin_make
- If it's not already there, add the following lines to your
.bashrc
file
export GAZEBO_MODEL_PATH=~/catkin_ws/src/sensor_stick/models
source ~/catkin_ws/devel/setup.bash
Launch the training.launch
file to bring up the Gazebo environment:
$ roslaunch sensor_stick training.launch
You should see an empty scene in Gazebo with only the sensor stick robot.
Next, in a new terminal, run the capture_features.py
script to capture and save features for each of the objects in the environment. This script spawns each object in random orientations (default 5 orientations per object) and computes features based on the point clouds resulting from each of the random orientations.
$ rosrun sensor_stick capture_features.py
The features will now be captured and you can watch the objects being spawned in Gazebo. It should take 5-10 sec. for each random orientations (depending on your machine's resources) so with 7 objects total it takes awhile to complete. When it finishes running you should have a training_set.sav
file.
Once your feature extraction has successfully completed, you're ready to train your model. First, however, if you don't already have them, you'll need to install the sklearn
and scipy
Python packages. You can install these using pip
:
pip install sklearn scipy
After that, you're ready to run the train_svm.py
model to train an SVM classifier on your labeled set of features.
$ rosrun sensor_stick train_svm.py
If everything went well you now have a trained classifier and you're ready to do object recognition!
$ rosrun sensor_stick robot_spawn.launch
In another terminal
$ chmod +x object_recognition.py
$ ./object_recognition.py