Title: Fly-by-pixels: End-to-end Learning for Autonomous FPV Drone Flight
Abstract: This project develops a system for autonomous flight of drones in a photorealistic simulator environment. The AirSim quadrotor simulator is used to collect training data, using a human pilot, in the form of images from the first-person view (FPV) camera. Each image is annotated with corresponding control inputs for throttle, roll, yaw and pitch. The end-to-end deep learning architecture combines Convolutional Neural Networks and Recurrent Neural Networks to learn human pilot behavior.
Committee Members: Madhur Behl, Sebastian Elbaum