duckie_town_reinforcement

lane following using reinforcment

Paper: here

source code :here

documentation: here

Introduction

We propose the use of Duckietown, an open-source platform for autonomy education and research, as a hardware implementation for our lane-following task in reinforcement learning. Duckietown includes autonomous vehicles called "Duckiebots" that are equipped with powerful onboard computers, such as Raspberry Pi, and a variety of sensors, including cameras and odometry sensors. The Duckiebots are capable of performing complex single- robot and multi-robot behaviors, making them an ideal platform for autonomy education and research. we start with The most basic task of the Duckiebot is lane following.

• This task is implemented using a realistic computer vision pipeline that contains these steps:
  1. Capture image
  2. Filter noise (Threshold/Masking)
  3. Detect line lanes 4.lane centroid is calculated, and error is deduced.
  4. input these calculation in reinforcment learning to start learning and take the next best action

Model pipline

We need to solve this problem by using reinforcement learning. OpenAI ROS provide a way to build a reinforcment learning algorithm by using ROS to control the agent(i.e car) To make the agent train by itself we will build some parts:

• Gazebo environment
• Robot environment
• Task environment
• learning algorithm(i.e. training script)

Results

after 100 epoch the episode reward reached 6000