mikeogezi / turtlebot3_rl

Teaching a TurtleBot 3 to drive in Gazebo.

TurtleBot3 RL 🤖⚙️

Teaching a TurtleBot 3 to follow a track using reinforcement learning. Credits to Kalvin, and Professor Matt Taylor for the track, and the original 3d assets.

Key info

• Trained a DQN on a Linear Annealled Policy based on Epsilon Greedy
• Fed the TurtleBot's camera feed to a CNN
• Negative reward for being outside the middle of the track
• Large magnitude negative reward for veering far off the track
• Large positive reward for reaching the finish line

Replicating training

• Install ROS Noetic
• Install Gazebo
• Install OpenAI Gym and others with pip install -r requirements.txt
• Pull this repo to your ~/catkin_ws/src folder
• Run catkin_make in your ~/catkin_ws folder
• Run roscore in one terminal
• Run roslaunch turtlebot3_rl main.launch gui:=false headless:=true in another (configure the arguments accordingly to render the environment in Gazebo)
• Run VISUALIZE=false rosrun turtlebot3_rl main.py in yet a third (configure VISUALIZE env var accordingly to render the camera stream)
• 🚀

Further work

• Running a proper headless setup on more capable machines (This was trained in an Ubuntu VM on my 2020 1.4GHz Macbook Pro)
• Providing other sensory input (odometry, or laser scan) from the robot to the RL algorithm