Embodock: A robust embodied language-driven framework for vision-based AUV docking in turbid waters with zero-shot generalization

Autonomous Underwater Vehicle (AUV) docking in turbid, unstructured environments demands deliberative reasoning beyond reactive pixel-to-action mappings—existing geometry-based and end-to-end learning methods fail when visual features degrade or unforeseen scenarios occur. This paper proposes Embodock , an embodied language-driven framework that introduces Large Language Models (LLMs) as a cognitive core within a closed-loop Perceive-Reason-Act-Observe cycle. We formalize the docking task as a structured decomposition f = f act ∘ f reason ∘ f perceive and introduce three specific technical innovations: (i) a semantic persistence mechanism that maintains operational intent during visual dropout through structured token history; (ii) a vision-dominant multimodal prompt grammar that encodes temporal decision history and auxiliary telemetry alongside visual tokens, enabling context-aware compensation for external disturbances; and (iii) a three-tier Sim2Real pipeline combining knowledge distillation and LoRA fine-tuning to adapt foundation models to data-scarce subsea domains. Extensive experiments validate the framework: in simulated turbid waters ( c = 1.5 m −1 ), Embodock achieves a 92.5% docking success rate with graceful degradation, while real-world lake trials under controlled project-acceptance conditions demonstrate a 4.1 cm terminal Mean Arrival Error and zero-shot transfer to the tested unseen docking geometry; component-level ablations attribute these gains to the three contributions individually.