Enabling Reliable Navigation for Lightweight VLMs Via a Semantically-Gated Visual Servoing Framework

While Vision-Language Models (VLMs) have advanced embodied navigation, deploying lightweight models on resource-constrained platforms remains challenging due to perception hallucinations and weak command compliance. To address these issues, we propose the Semantically-Gated Visual Servo (SGVS) framework, a dual-system cognitive architecture tailored for lightweight VLMs. The framework integrates a Visual Object Interception (VOI) mechanism as a reflexive system to trigger rapid maneuvers upon target recognition, enhancing execution efficiency. Concurrently, a Closed-Loop Semantic Error Correction (CL-SEC) mechanism acts as a reflective system. By combining multi-prompt verification with dynamic negative sample mining, CL-SEC translates perceptual errors into real-time episodic memory and employs action pruning to block hallucination-driven failures at the source. Experimental results on the HM3D dataset demonstrate that SGVS improves the navigation success rate of the Qwen2.5-VL model from 45.7% to 48.0% while reducing the dangerous false positive rate by 30.4%. This study indicates that system-level architectural interventions can partially compensate for the perceptual limitations of lightweight foundational models.