Visual End-to-End Autonomous Navigation System for UAV

This paper constructs a deep reinforcement learning navigation framework for indoor unknown scenes, which takes visual information and drone motion information as inputs. By extracting and integrating visual features, motion features, and temporal features, the adaptability of drones to complex environments and their ability to transfer between different environments have been improved. Based on the AirSim simulation environment, a discrete action set of unmanned aerial vehicles was designed and experimentally validated for target point navigation in different indoor environments. The experiment shows that the navigation network in this article can effectively complete various navigation tasks and has a certain degree of generalization.