Deep Reinforcement Learning for Motion Planning of Quadrotors Using Raw Depth Images

In this work, we introduce a novel, end-to-end motion planner for quadrotor navigation. Informed by a rough path to goal in partially unknown environments, our method creates desirable motion plans using raw depth images from a front-facing camera. It exploits correlations between local spatial portions of these images to generate desirable motion primitive sequences on the fly without conducting explicit sensing-reconstructing-planning. We evaluate our method through an extensive comparison with three competitor algorithms over ten different environments in AirSim simulations. Our method outperforms its competitors in terms of safe navigation distance, navigation time, and crash rate over 50 flights. We also deploy our method for real flight tests with DJI F330 Quadrotor equipped with Intel RealSense D435, and demonstrate its real-time ap-plicability. Our method successfully performs 15 real flights in three different environment settings with increasing complexity. The experiments can be found at https://youtu.be/hw0sxNwliqs.