EXPLORING DEEP LEARNING APPROACHES FOR QUADROTOR OBSTACLE AVOIDANCE

Abstract

This thesis seeks to address the speci c problem of quadrotor obstacle avoidance for search and rescue applications. Unmanned Aerial Vehicles (UAV's), speci cally quadrotors, are a robotic platform capable of high speed agile ight while carrying sen- sor and processing payloads. Their capabilities have potential for many applications including inspection, photography, and search and rescue. While they are a capable platform for many tasks, the algorithms designed for them still do not to utilize these systems e ectively. Obstacle avoidance is an integral part of robotic navigation and a basis for many tasks. This thesis approaches this problem from a learning perspective. Deep learning speci cally has recently gained popularity for various robotics tasks, be- cause it is e ective at addressing noisy data and is fast once trained. Due to quadrotors' limited on board computing capabilities and noisy sensor input, deep neural networks are an attractive solution to the obstacle avoidance problem. This thesis will formulate and evaluate di erent approaches to deep learning for obstacle avoidance.