Navigation and collision avoidance with human augmented supervisory training and fine tuning via reinforcement learning

Robust navigation and orientation under complex conditions is a must for autonomous drones operating in new and varied environments. Creating drones with adequate behaviors can be a challenge from both a training standpoint and a generalization standpoint. Using human expertise data is an option to help bootstrap the learning process; however, using the human data can lead to side consequences that are not immediately intuitive. This study focuses on applying varying levels of human input to an agent to determine how this input affects the agent's performance. The Unreal Engine and the Airsim plugin are used to train a quadcopter agent in an abstract "blocks world" type environment. Six agents in total are trained, with the first five having increasing amounts of human input and the sixth agent having no human input. A variety of metrics are looked at, including total goals achieved and time to achieve some number of goals.