Exploratory UAV Autonomous Visual Navigation With Spatiotemporal Cognition

In extreme environments such as military conflicts or natural disasters, GPS signals may become completely ineffective or unreliable, rendering navigation systems utterly useless. Unmanned aerial vehicles (UAVs) are instantly at risk of getting lost, straying, or crashing. Vision-based navigation technology provides a solution for this task scenario. However, the currently mainstream Vision-and-Language Navigation (VLN) technology mainly relies on strict pairing of language instructions and visual observations to guide UAVs to their target locations. Clearly, this method is not suitable for extreme environments, as in unfamiliar wilderness environment, people cannot provide detailed step-by-step guidance and clear navigation directional instructions. UAVs need to engage in exploratory autonomous navigation based on spatio-temporal cognition. Therefore, this study aims to develop an Exploratory Navigation system, UAV E-Nav. The main idea of this system is to simulate the spatio-temporal cognition and flight experience of Senior Navigator. It proposes an Exploratory Navigation framework consisting of three large trained models: the Navigation Language Model (NLM), Navigation Positioning Model (NPM), and Navigation Steering Model (NSM). This framework supports natural language interaction and dynamic path planning. The Exploratory Navigation Framework enables UAV E-Nav to better handle complex navigation cognition in the real world, thereby effectively making action decisions in exploratory navigation. We have instantiated a real-world UAV navigation system, UAV E-Nav, by turning off the satellite positioning function of navigation and demonstrating remote exploratory navigation in complex unknown environments through natural language instructions. This study provides a new thinking approach and paradigm for UAV Exploratory Navigation, enhancing the practicality of Human-UAV interaction.