Exploring DeepQ Learning for Micro UAV Tree Avoidance

Abstract

In this paper, we focus on tree avoidance with a quadcopter in a search and rescue scenario. Search and rescue is time critical and usually involves complex environments such as forests, or destroyed buildings. Because of their speed, size, and potential autonomy, micro UAV's are ideal for supporting search in rescue operations. The controllers for these micro UAV's must be efficient, reliable, and also tolerant enough to handle noisy data from various lightings, blurs, and orientations caused by flight. Current obstacle avoidance controllers do not meet the speed requirements in complicated environments to be practical in search and rescue. We explore the use of a deep convolutional, recurrent Q network to shorten the pipeline from observation to reaction during flight and handle noisy data.