Toward Low-Flying Autonomous MAV Trail Navigation using Deep Neural\n Networks for Environmental Awareness

We present a micro aerial vehicle (MAV) system, built with inexpensive\noff-the-shelf hardware, for autonomously following trails in unstructured,\noutdoor environments such as forests. The system introduces a deep neural\nnetwork (DNN) called TrailNet for estimating the view orientation and lateral\noffset of the MAV with respect to the trail center. The DNN-based controller\nachieves stable flight without oscillations by avoiding overconfident behavior\nthrough a loss function that includes both label smoothing and entropy reward.\nIn addition to the TrailNet DNN, the system also utilizes vision modules for\nenvironmental awareness, including another DNN for object detection and a\nvisual odometry component for estimating depth for the purpose of low-level\nobstacle detection. All vision systems run in real time on board the MAV via a\nJetson TX1. We provide details on the hardware and software used, as well as\nimplementation details. We present experiments showing the ability of our\nsystem to navigate forest trails more robustly than previous techniques,\nincluding autonomous flights of 1 km.\n