Vision-Based Obstacle Avoidance in Drone Navigation using Deep Reinforcement Learning

Abstract

Drones are soon getting into action in many commercial and non-commercial missions because of their low cost and easy deployment. Nonetheless, this so-called easy deployment requires a human operator to be safe and reliable. With the use of drone platforms as means of package delivery or search and rescue in dangerous sites, full autonomy is now attracting the robotic community. Specifically, using drones in urban areas requires some degrees of autonomy because of potential dynamic obstacles. This paper provides an architecture inspired by natural livings such as insects to mitigate the navigation challenge autonomously. We conduct our study in a standard simulation environment, and results will be compared to a human expert. Using deep reinforcement learning, we manage to generalize the learned policy in this specific domain. Finally, our simulated results imply that reasonable collision avoidance in urban environments is achieved compared to a human pilot.