From Anywhere to the Destination: Visual Reinforcement Learning for GNSS-Denied Drone Navigation

Unmanned aerial vehicles (UAVs), also referred to as drones, typically rely on the global navigation satellite system (GNSS) for positional awareness and navigation. However, the GNSS can prove unreliable in certain environments because of signal obstruction or degradation. Recently, vision based navigation has emerged as a promising alternative to GNSS. Despite its potential, this visual method necessitates a significant investment in expert-guided demonstrations, resulting in a cumbersome and passive collection process, especially for large-scale outdoor navigation tasks. In this work, we propose AceFormer, a vision based active exploration transformer framework that enables agent navigation from any starting point to a destination without expert guidance. Specifically, this framework introduces an intra instance feature encoder in tandem with a semantic guidance decoder, which empowers the drone agent to extract global instance-level semantic information. Deep reinforcement learning is further employed to train the drone agent, enabling it to explore the environment dynamically, and learn navigation tasks in a trial-and-error way. To empirically evaluate AceFormer, a challenging drone navigation simulated flight environment (DroEnv) is designed by leveraging Google Earth as a navigation platform, which contributes to a standardized benchmark for the drone navigation field. Extensive experiments conducted on DroEnv with randomly sampled starting points demonstrate the effectiveness of the proposed AceFormer in both ideal and noisy environments. Our code will be made publicly available.