Collision avoidance for a small drone with a monocular camera using deep reinforcement learning in an indoor environment

Collision avoidance of drones in a complex environment, especially in an indoor environment, is a challenging task. This paper develops an obstacle avoidance system for small multi-rotor drones based on a deep reinforcement learning algorithm using only a monocular camera. The proposed method comprises two steps: depth estimation and navigation decision-making. For the depth estimation step, a pre-trained depth estimation algorithm based on a CNN (Convolutional Neural Network) is used. In the navigation decision-making step, a dueling double deep Q-network is employed. The entire training procedure is performed in a Gazebo simulation environment using a robot operating system. To validate the robustness of the proposed approach, various simulations and experiments are conducted using a Parrot Bebop2 drone in an indoor corridor. We demonstrate that the proposed algorithm successfully navigates through a narrow corridor comprising a texture-free wall, people, and boxes. A supplementary video clip of the experiments can be found at https://youtu.be/ oSQHCsvuE-8. © ICROS 2020.