ESEM: A Visual Topological Navigation Method Integrating Edge Semantic Enhancement in Challenging Environment

Image-goal navigation ranks among the critical tasks in embodied robotic visual navigation, requiring the robot to navigate to the goal position indicated by the image in an unknown environment. While recent works have made progress in image-goal navigation by constructing image-based topological maps, the complexity of the real world, along with the lack of directional control and semantic information in topological maps, still pose significant challenges to robustness and reliability. To solve this problem, we propose a new strategy that utilizes a semantic topological map to help the robot's navigation in unknown environments. Specifically, we quantitatively analyze the directional relationship between the current and goal nodes to construct the semantic edges of the topological map. This approach enables the robot to make informed decisions by dynamically assigning physically meaningful directional semantics to each edge. We deploy this method on a real-world four-wheeled ground robot, relying solely on visual input (RGB), to realize the image-goal navigation task. The experimental results demonstrate that the proposed approach significantly outperforms other high-performance baseline approaches in terms of navigation success rate, demonstrating excellent robustness.