The simulation-supervised package combines different sets of code needed to train a DNN policy to fly a drone.
- online_training: the tensorflow code used for training and running the DNN policy.
- drone_simulator: a simulated drone model for Gazebo. This is a copy of the original hector quadrotor model. OR
- bebop_autonomy: a copy of the bebop autonomy package of ROS. This package allows you to test the DNN in the real-world. The copy is only for ensuring stability while performing research. There are no significant modifications so it is probably best to use the original. If you are using the Doshico docker image, it is already installed globally.
This package is best build in a separate catkin workspace.
mkdir -p ~/simsup_ws/src
cd ~/simsup_ws && catkin_make
cd ~/simsup_ws/src
git clone https://github.com/kkelchte/simulation_supervised
cd ~/simsup_ws && catkin_makeYou will have to set the correct path to your tensorflow pilot_online package.
In case you are using different drone models, you will have to adjust the config.yaml file in order to set the correct rosparams.
Here are some common used setting combinations in order to remember them:
# Test online the performance of a heuristic defined in tensorflow/gt_depth_online/../rosinterface.py using for instance recovery cameras flying 3 times through a generated canyon
$ ./scripts/evaluate_model.sh -m auxd -s start_python_sing_gtdepth.sh -t test_depth_online -r true -w canyon -n 3
# Train online in canyon, forest, sandbox
$ ./scripts/train_model.sh -m mobilenet_025