Development of a Distributed and Scalable Testbed for UAVs using Reinforcement Learning 

Abstract The aim of this project is to develop a Testbed for designing and training Multi-agent Reinforcement Learning (RL) algorithms for cooperative and self-organizing Unmanned Aerial Vehicles (UAVs). The main purpose of the development of a scalable and distributed testbed based on Multi-agent RL algorithms is to enable UAVs to make decisions using real-time data and perform tasks autonomously. In this project, a novel testbed is developed that allows the integration of different Multi-agent RL algorithms with a flight simulator. This testbed supports UAVs that learn to fly and coordinate together in the simulated environment to accomplish the objective of target tracking. It employs novel techniques that enable faster learning and higher performance as compared to conventional Multi-agent RL methods. FlightGear is the flight simulator used in this project. This testbed can be used to train control models for a wide variety of use cases. As a proof of concept, a problem is formulated regarding target tracking of UAVs. The tracking aircraft follows the path of the target aircraft. Both tracking and target aircraft are controlled by different Multi-agent RL models and fly on a common flight simulator. This testbed can also scale up the number of tracking aircraft and can be distributed to several systems.