Abstract: To address the problems of decreased LiDAR observation consistency and unstable mapless navigation decision states caused by noise, occlusion, and temporal anomalies, this paper proposes a frequency-structural guided mapless navigation method, termed frequency-structural guided navigation (FSG-Nav). This method organizes continuous LiDAR scans into time-angle domain observation units, introduces a frequency-domain structural prior via wavelet decomposition, and learns latent environment representations under degraded conditions with self-supervised constraints; the learned representations are then fused with the target position and robot motion state, and a soft actor-critic policy network generates continuous control commands. In simulation experiments, the proposed method achieves navigation success rates of 84%–98% and 60%–96% under moderate and severe degradation conditions, respectively. In addition, in real-vehicle deployment, the robot completed angular-domain masking navigation experiments and closed-loop latency tests in an underground parking lot. The real-vehicle experiments and simulation results jointly show that the proposed method can still operate stably under conditions of missing local observations and meets real-time control requirements, while also indicating that the frequency-domain structural prior helps mitigate the adverse effects of observation degradation on navigation decisions, thereby improving the robustness of mapless navigation.