Monocular Vision Obstacle Avoidance UAV: A Deep Reinforcement Learning Method

Abstract

In this paper, a method based on deep reinforcement learning (DRL) is proposed, which allows unmanned aerial vehicles (UAVs) to complete obstacle avoidance tasks only through vision in an environment full of common indoor obstacles. This technology is very important for indoor UAVs, due to the limited GPS signal and overcrowding of obstacles compared to the outdoor environment. We use Variational Autoencoder (VAE) to compress image information combined with the policy-based DRL model to implement the visual obstacle avoidance of VAVs. Simulation experiments have demonstrated that this method can make the UAV master obstacle avoidance in a continuous action space with a fixed direction. Compared with the traditional policy-based DRL visual obstacle avoidance algorithms, it can converge faster.