Reinforcement Learning on a custom car using ROS

Package which trains a robot to reach a goal position without colliding using RL. Experiments are carried out on a car simulation in Gazebo.

Install dependencies

sudo pip install gym
sudo apt-get install python-skimage
sudo pip install h5py
pip install tensorflow-gpu (if you have a gpu if not then just pip install tensorflow)
sudo pip install keras
cd ~
git clone https://github.com/erlerobot/gym-gazebo
cd gym-gazebo
sudo pip install -e .

Clone this repository

cd catkin_ws/src
git clone "https://github.com/YugAjmera/rl-car-ros.git"
cd ..
catkin_make

Launch

roslaunch rl-car-ros simple.launch

Start Q-Learning agent with 𝜖 -greedy action selection :

roslaunch rl-car-ros start_qlearning.launch

Start Sarsa-Learning agent with 𝜖 -greedy action selection :

roslaunch rl-car-ros start_sarsa.launch

Start Expected Sarsa-Learning agent with 𝜖 -greedy action selection :

roslaunch rl-car-ros start_expected_sarsa.launch

Environment Specifications

State : LaserScan + Odometry(x and y coordinates)
Actions : Forward, Right, Left
Episodic Task: An episode ends when the robot hits a wall (Reward = -200) or when the robot reaches the goal position (Reward = +5000). A living reward of -1 is given at each step which encourages the robot to end the episode as soon as possible.
The world is show below:

Below are the results obtained for 50 episodes: (if someone could run this for 1000 episodes, I would be happy to accept a pull request !)