README.md

Sample Factory

The original code base is from: https://github.com/alex-petrenko/sample-factory

Our contributions:

Design two policies to make quadrotors can avoid collisions, mean embedding policy and histogram policy.

gym_art

The original code base is from: https://github.com/amolchanov86/gym_art Our repo is: https://github.com/Zhehui-Huang/gym_art/tree/cs527

Our contributions:

Extended single-drone simulator, gym_art, to multi-drone simulator

Added obstacles to the simulator

Implemented 6 interesting scenarios, circular configuration scenario, static goal scenario, dynamic goal scenario, pattern composing scenario, pursuit-evasion scenario, dynamic obstacle scenario

Created a camera for global visualization

Create the collision model

Visualized the drone collisions

Since our group closely collaborate with the author of the above two projects, most of our contributions/modifications have already been accepted by the author and merged to their code base.

Goal

Create a robust control policy which can allow quadrotors to avoid collisions & obstacles

Use Deep Reinforcement Learning Algorithms to train the policy in a simulated environment.

Motivation

Multi-drone collaboration can be used in many fields, such as search and rescue, surveillance and mapping missions, attack and defense mission

Most of previous research are control-based methods can not handle complicated environments and they are computational expensive.

Our website:

https://sites.google.com/view/multi-drone-collaboration/home