Robotics Project

This repo hosts the code relative to my Robotics project of the posgraduate program in applied Artificial Intelligence hold at Erasmushogeschool Brussel (EhB). In this project, we try to solve the following problem: a person is allowed to hold a Raspberry Pi in front of a laptop’s front camera. The goal is to train a RL agent so that the red light on the Pi’s LED matrix is as close a possible to the webcam’s image center despite the Pi being moved and/or rotated in front of this camera.

Installation

In order to run the code, simply create a new conda environment using:

conda env create --file env.yml

Then, make sure to run the setup.py file for local dependencies:

pip install -e .

Now you're good to go! Be aware that in order to run the code, you must provide access to the Pi by typing its IP address in the Network class. To enjoy a trained agent, simply run the main.py script with the following lines:

from src import Agent
from stable_baselines import PPO2

if __name__ == "__main__":
    agent = Agent(model=PPO2)
    agent.evaluate(tensorboard_log='./evaluation/', model_path='/checkpoint_2')

Feel free to edit this file according to your needs.

Code Structure

The code is organized around the following parts:

src

Source code of the project. It contains the support for the agent as well as the associated gym environment. The network and processing scripts are aimed to respectively provide access to the server and to process the image captured by the PC's webcam. The calibrate.py file allows to play around with OpenCV so as to find the optimal filters requird to retrieve the position of the red dot. Finally, the checkpoints and the trial directories respectively store intermediate versions of an agent as well as the learning curves associated to it. The model checkpoints directory contains the current version of the agent being trained and the evaluation directory allows to monitor an agent while being tested.

notebooks

This folder contains the notebook required to load the data from tensorboard logs into pandas for visualization. A typical learning curve obtained with this notebook is the one using PPO2 in checkpoint_2:

doc

The doc folder contains the pdf version of the report of this project. It is stored in the main.pdf file. There you will find a detailed description of this project along with the results obtained so far.

Demonstration

To visualize a demonstration of the trained PPO2 agent of stable-baselines, check out the following video:

https://youtu.be/VGtoMPvjIwM

Feel free to leave some comments and open up issues if you run into troubles :=)