CISMG-Nav: A Cross-Task Incremental Semantic Memory Graph-Driven Vision-and-Language Navigation Framework

Vision-and-Language Navigation (VLN) is a crucial indicator for evaluating embodied intelligence. However, existing VLN methods face several challenges. First, the agent typically relies only on short-term memory, making it difficult to reuse accumulated experiences and environmental perceptions. Second, the semantic topological graph exhibits an inherent contradiction between representation granularity and computational efficiency, while its relational reasoning capability remains fundamentally limited. To address these issues, this letter proposes a Cross-task Incremental Semantic Memory Graph-Driven Vision-and-Language Navigation Framework (CISMG-Nav). The agent stores and updates environmental perception information by maintaining a Cross-task Incremental Semantic Memory Graph (CISMG). Specifically, CISMG integrates topological structures with semantic attributes. The observed semantic information is structurally encoded as nodes with semantic attributes, while edges represent reachability relations. Additionally, commonsense co-occurring objects are introduced to enhance semantic expressiveness and relational reasoning capability. We further propose a global-local hierarchical navigation strategy. At the global level, the agent performs semantic-guided path planning based on the CISMG, while at the local level, it achieves fine-grained goal location through a large language model. Experiments on R2R-CE dataset and real-world environments demonstrate that CISMG-Nav significantly improves navigation success rates. Compared with existing training-free methods, our framework maintains a zero-shot setting while achieving superior performance.