turtlebot_transfer_learning

Contains ROS stack for zero-shot transfer experiment using cycle-GAN images for turtlebot navigation.

Creating a python3 catkin workspace

We followed these instructions to build a catkin workspace in melodic compatible with python3: https://www.miguelalonsojr.com/blog/robotics/ros/python3/2019/08/20/ros-melodic-python-3-build.html
We used conda to install the python3 packages necessary to run the ppo network.

Installing packages

Pull this repository into the /src folder of a catkin workspace. Run catkin_make in the root of the workspace to build the package. In the python3 catkin workspace, pull the repository at https://github.com/bkesari1998/turtlebot_transfer_learning_python3 into the /src folder of the workspace. Run catkin_make in the root of the python3 workspace to build the package.

Generating a data set

In config/dataset.yaml, assign the ROS parameters for images per episode and total number of episodes. In config/actions.yaml, assign float values to each of the four primitive move actions. On the turtlebot's computer, launch both image.launch and turtlebot.launch. Run generate_dataset.py using rosrun or a python interpreter.

Running experiment

In config/actions.yaml, assign float values to each of the four primitive move action values and assign an integer to the timesteps_per_episode parameter. On the turtelbot's computer, on a ROS melodic catkin workspace built on python 2.7, launch turtlebot.launch. Then on a different catkin workspace with python3 support, either on a remote computer or the turtlebot's compute, run ppo_manager.py using python3 found in the turtlebot_transfer_learning_python3 package. Make sure to add the correct path to the weights file (i.e weights.pth) to model_net.py (line 67). Finally, on the original catkin workspace, run manager.py. Make sure to source the python3 catkin_ws in the new terminal.