This project was realized by the REDS institute @ HEIG-VD.

Authors: Guillaume Chacun, Mehdi Akeddar, Thomas Rieder, Bruno Da Rocha Carvalho and Marina Zapater
REDS. School of Engineering and Management Vaud, HES-SO University of Applied sciences and Arts Western Switzerland
Email: {guillaume.chacun, mehdi.akeddar, thomas.rieder, bruno.darochacarvalho, marina.zapater}@heig-vd.ch

This Jupyter notebook is used to train a DDPG (reinforcement learning) model to control a drone, specifically to guide it to follow a predefined line on the ground. A separate deep learning model is used to identify and track the line from images captured by the on-board camera. The pilot module processes real-world normalized coordinates of the line and outputs forward, lateral, and angular velocities for the drone.

• train_drone_pilot.ipynb : Notebook used to train the reinforcement learning model.
• Agents.py : DDPG related classes (OUActionNoise, abstract Agent class, LearningAgent, ExplorationAgent).
• DDPGNetworks.py : DDPG related classes (CriticNetwork, ActorNetwork).
• SharedReplayBuffer.py : DDPG replay buffer shared across processes.
• Drone.py : Class to simulate the behavior of the drone.
• Line.py : Class to generate a random line for the drone to follow.
• Environment.py : Class to handle the simulation (episodes). Contains an instance of Drone and Line.

Install dependencies:

poetry install

Activate the virtual environment of Poetry:

poetry shell

Start your Jupyter Lab server:

jupyter lab

Most of the DDPG code is from @philtabor's GitHub (last visited on March 5th, 2024).