Coverage path planning with Reinforcement Learning. Using ROS, controlling a TurtleBot3 with a UV lamp to cover an area with a minimum UV energy to kill coronavirus. Gym environment implementation were based on the implementation from gym-turtlebot3

ROS is required for the simulation. Requires ROS-UV-Cleaning-Robot, can be pulled as a submodule with git submodule update --init --recursive

I used a docker container for Tensorflow and Baselines. Using the Baselines zoo CUDA image, I installed the required ROS packages with rospypi, so that the ROS installation is not needed in the container. I launched my container with docker run --runtime=nvidia --env="DISPLAY=${localEnv:DISPLAY}" --volume="/tmp/.X11-unix:/tmp/.X11-unix" --net=host -v /home/$USER/cppRL:/usr/share/cppRL -it stablebaselines/rl-baselines-zoo This exposes the network adapter for ROS communication, adds a shared folder where the project is, as well as a display.

In addition, the UVC lamp workspace needs to be sourced in the docker container source ROS-UV-Cleaning-Robot/devel/setup.sh.

To launch, a gazebo simulator is needed, a pre-made map must be published with map_server, and the uvc_lamp node needs to be launched. If angle control(instead of angle rate control) is used, the angle_controller in ROS-UV-Cleaning-Robot needs to be run.

Two maps have been tested with the RL agent, small-room and small-house.

Launch Gazebo with a map: roslaunch turtlebot3_gazebo turtlebot3_small_room.launch

Launch map_server with the same map: rosrun map_server map_server maps/small_room_40cm.yaml

Launch uv_lamp: roslaunch uvc_lamp uvc_grid_map.launch

Launch angle_controller: rosrun angle_contoller angle_controller.py