Reward-driven U-Net training for obstacle avoidance drone

摘要

Along with the fast progress in deep learning, an autonomous drone with obstacle avoidance capability has been studied mainly by two machine learning paradigms, Le. supervised learning, and reinforcement learning. The former has some advantages since the trained network is light and fast, but it needs a large amount of data that requires laborious manual labeling. With the latter, such a drawback can be overcome as an agent learns by itself in a simulated environment, although the gap between the real and simulated one has to be minimized in the end. This study proposes a new framework where a supervised segmentation network is trained with labels made by an actor-critic network in a reward-driven manner, wherein this U-Net based network infers the next moving direction from the sequence of input images. For the actor-critic part, several recent policy gradient algorithms have been tested for controlling the drone with the continuous action space. After training in the Airsim simulation environment, the model is transferred to a Bebop drone flying in the real environment, built as a reconfigurable maze using panels and a hoop. The result suggests that our network enables the drone to navigate through the obstacles using only monocular RGB input in the trained environment as well as in the reconfigured ones without retraining. (C) 2019 Elsevier Ltd. All rights reserved.