Gazebo environment for testing reinforcement learning and evolutionary robotics.
Demonstrates a neural network controlled robot evolving to navigate collision free in a maze environment.
This was a small project during my MSc in 2017, the slides for the presentation I gave related to this repo can be found here.
HARD CODE 'kobuki.launch.xml' in 'turtlebot_gazebo' to point to /opt/ros/kinetic/share/turtlebot_description/robots/kobuki_hexagons_hokuyo.urdf.xacro
During evolution (left) and after (right)
Opposite direction (left) and escaping a trap (right)
Failure cases: open space and unfamiliar environments
A lot can be improved, and it's clearly overfit to the "maze" environment, but hey, it's a proof of concept. Of course, you could use a classic controller or optimise the weights another way.
This repo and all contained was for a short project/assignment and was never intended to be publicly released (hence the mess, lack of documentation and general unusability), so everything here is provided as-is. This is unlikely to be worked on in the forseeable future.