Senior Design Team sdmay21-33 • Reinforcement Learning with Graph Neural Networks for Drone Collision Avoidance

Project Description

The problem the team was faced with for this project was designing a reinforcement learning (RL) model that will be trained in an environment in which the networks must control a drone that will explore an environment while performing obstacle avoidance. Furthermore this RL model needs to be designed so that it can utilize two different network architectures, a convolutional neural network (CNN) policy and a graph neural network policy (GNN). This is because our advisor’s end goal with this project is to use the networks we develop to formulate a paper on the comparisons between the CNN and GNN and how well the drone handles new environments depending on which one is in control. The technologies we are using to achieve this outcome are Unreal Engine, Microsoft Airsim, and Google's machine learning framework Tensor Flow. Unreal Engine is used in this project for developing simulated enviornments that we can train the ML model on prior to deploying it to a physical drone. Airsim is used to communicate between the model in tensorflow and the world created in Unreal Engine. Feel free to check out our news report as well!

Image of Drone Training in Simulated Enviornment

Image of Physical Drone the model will be deployed to