A framework for improving UAV decision of autonomous navigation from training to application migration under perceptual uncertainty

Abstract

Unmanned Aerial Vehicles (UAVs) autonomous navigation based on reinforcement learning (RL) usually requires training agents in simulation scenarios and then transferring the trained agents to application scenarios. However, due to serious distribution mismatch between the idealized simulation scenario and the application environment and the inevitable uncertainty perception problem of airborne sensors in complex scenarios, the navigation performance of UAV under migration applications is not ideal. This work fully analyzes the factors that affect UAV navigation performance, including algorithm performance, training strategy, and state awareness. Based on the analysis results, this article proposes a framework to improve the autonomous navigation performance of UAVs in the migration process from training to application, which consists of three parts: "scenario-perception-algorithm". In addition, this paper proposes improvement strategies for each part from the perspectives of spatial features, temporal features, and perceptual denoising. We combine the proposed framework with navigation algorithms to improve the navigation decision-making performance of UAVs in migration applications under uncertainty perception. Many simulation experiments demonstrate the effectiveness of the proposed framework and its robustness to uncertainty perception.