Computing and Informatics

Vision-and-language navigation faces critical challenges, including insufficient long-horizon reasoning, misalignment between instruction segments and visual observations during navigation, and hallucination interference in large model reasoning. These limitations hinder the ability of large models to accurately interpret complex scenes through textual summaries alone, particularly in understanding spatial distributions of environmental elements and dynamically adapting to the demands of target instructions. To address these challenges, we propose a novel historical-topological vision-and-language navigation framework based on multi-curvature spatiotemporal Chain-of-Thought reasoning. Leveraging an encoder-decoder architecture, our model encodes historical information from input features, effectively resolving the lack of global topological guidance in long-horizon reasoning tasks. To mitigate hallucination in large model reasoning, we exploit the inherent data characteristics of vision-and-language navigation tasks. By leveraging multi-curvature manifolds, we identify salient spatiotemporal references in visual observation sequences, enabling the model to ground its reasoning in these references for accurate situational understanding and future prediction. To enhance spatial structure perception, we introduce a graph self-attention mechanism that jointly models node distances and visual similarity in historical topological graph embeddings, effectively capturing spatial relationships between nodes. This work introduces novel insights for large model-based vision-and-language navigation in indoor scenes, advancing their applicability in complex environments.