SFPathFormer for spatial frequency robot navigation and obstacle avoidance

In complex and dynamic environments, robot path planning and obstacle avoidance often suffer from significant perception noise, severe feature redundancy, and insufficient global environment structure modeling. To address these challenges, this paper proposes a vision-driven robot navigation and obstacle avoidance framework that integrates spatial-frequency perception, dynamic feature selection, and global environment modeling. The proposed framework aims to enhance path safety and planning stability in complex scenarios by constructing a unified modeling pipeline from environmental perception and feature representation to path decision-making. First, at the methodological level, we design a spatial-frequency joint perception module based on parameterized wavelet downsampling. By introducing a global spatial-frequency attention mechanism, the proposed module effectively reduces feature redundancy and computational complexity while enhancing the salient representation of obstacle regions. Second, we propose a dynamic domain feature selection mechanism for path decision-making, where contrastive learning is employed to guide the model to adaptively select feature subspaces that are highly correlated with navigational safety, thereby improving generalization and robustness under diverse environmental conditions. Furthermore, a vision Transformer is introduced to model the global structural relationships of the environment, capturing long-range obstacle correlations and the overall spatial layout of traversable regions, thus providing consistent and structured high-level representations for path planning. Extensive experiments are conducted on multiple typical robot obstacle avoidance scenarios and complex environmental datasets. The experimental results demonstrate that the proposed method significantly outperforms state-of-the-art approaches in terms of path safety, obstacle avoidance success rate, and planning stability, while exhibiting stronger robustness under complex backgrounds and dynamic disturbances. Additional analysis further validates the synergistic effects of spatial-frequency perception, dynamic feature selection, and global modeling, proving that the proposed framework effectively improves robot path planning and obstacle avoidance performance in complex environments. This work provides a scalable and effective solution for vision-driven autonomous robot navigation.