This is a solution baseline for the AI Driving Olympics competition using Reinforcement Learning & Imitation Learning via Supervised Learning (a.k.a. Behavioral Cloning) in PyTorch, Tensorflow, and Tensorflow's Keras for the challenge aido_LF.

The online description of this challenge is here.

For submitting, please follow the instructions available in the book.

Most of the code is explained within its script as well as in the corresponding folder's README.

Go ahead and

You can train a reinforcement learning agent (expert) that learns to drive perfectly within an environment. Then you can run the agent on a bunch of different maps/environments to collect data (observation & action pairs) to imitate the expert's behaviour, a.k.a. Imitation Learning, Behaviour Cloning. Finally, you have an agent that navigates within an environment using only one single sensor, the camera.

Follow the installation steps explained in this GitHub repository, which is the official repository of the simulator used at the competition.

Note: You do not need to install anything on your local PC to use the notebook on Colab! That means, without any installation you can train both networks: The RL agent and the IL agent, which at the end yields a self-driving agent! Therefore, you would only need to install the required packages to your local PC if you want to evaluate, visualize how the trained agents work.

This repository can be used by anyone who would like to ground his/her knowledge in Reinforcement Learning, Imitation Learning, PyTorch, Tensorflow, Keras, and Self-Driving Cars.

• Simulations in general and how to use them

• Image processing methods, use-cases for Self-Driving Cars such as; Canny Edge Detection, Lane Line Detection with Hough Lines, etc.

• Reinforcement Learning and one method of RL, namely DDPG and its implementation in PyTorch

• Applying DDPG to:

  • the "Hello World" of RL, namely CartPole Problem a.k.a. Inverted Pendulum

  

  • a Self-Driving Car that learns itself how to drive well in different environments

  

• How Imitation Learning can be applied to Self-Driving Cars by training neural network models with both Tensorflow and Keras

• Submission to a world-wide competition using Docker

This is a project within the curriculum of MSc. Intelligent Systems and supervised by Dr.Andrew Melnik at University Bielefeld. If you are a student at University Bielefeld and interested in this project, Dr.Melnik would be happy to work with you!