Knowledge Graph-Driven Reinforcement Learning for Zero-Shot Vision-Language Navigation

To address the limitations of zero-shot generalization in Vision-Language Navigation (VLN), this paper proposes a novel knowledge graph-driven reinforcement learning approach. Our method constructs a hierarchical, dynamically updated knowledge graph online during the agent’s real-time interaction with the environment, seamlessly aligning external semantic priors with continuous visual perception. By leveraging a Chain-of-Thought (CoT) prompting mechanism, the agent performs multi-hop reasoning to precisely locate target objects. Furthermore, we design an end-to-end optimized reinforcement learning framework that fuses multi-modal features and employs a task-oriented composite reward function. Extensive experiments in the AI2-THOR simulation environment demonstrate that the proposed method significantly improves navigation success rates in zero-shot settings. The results validate its robust generalization capabilities, particularly for unseen object categories and complex scene layouts.