Dynamic Topology Awareness: Breaking the Granularity Rigidity in Vision-Language Navigation

TL;DR

This work tackles Granularity Rigidity in Vision-Language Navigation within Continuous Environments by introducing DGNav, a framework that dynamically adapts topological maps and connectivity. A Scene-Aware Adaptive Topology modulates graph density in response to scene complexity, using a linear, conditional mechanism to adjust the merging threshold $\gamma_t$ based on waypoint dispersion $\sigma_t$. A Dynamic Graph Transformer reconstructs edges by fusing geometric priors with visual semantics and instruction relevance, encoded into a dynamic adjacency $\mathbf{E}_{dynamic}$ that guides graph attention. Empirical results on R2R-CE and RxR-CE demonstrate superior navigation performance and robustness, with ablations confirming the effectiveness of adaptive topology and multi-modal edge fusion for long-horizon, instruction-guided navigation.

Abstract

Vision-Language Navigation in Continuous Environments (VLN-CE) presents a core challenge: grounding high-level linguistic instructions into precise, safe, and long-horizon spatial actions. Explicit topological maps have proven to be a vital solution for providing robust spatial memory in such tasks. However, existing topological planning methods suffer from a "Granularity Rigidity" problem. Specifically, these methods typically rely on fixed geometric thresholds to sample nodes, which fails to adapt to varying environmental complexities. This rigidity leads to a critical mismatch: the model tends to over-sample in simple areas, causing computational redundancy, while under-sampling in high-uncertainty regions, increasing collision risks and compromising precision. To address this, we propose DGNav, a framework for Dynamic Topological Navigation, introducing a context-aware mechanism to modulate map density and connectivity on-the-fly. Our approach comprises two core innovations: (1) A Scene-Aware Adaptive Strategy that dynamically modulates graph construction thresholds based on the dispersion of predicted waypoints, enabling "densification on demand" in challenging environments; (2) A Dynamic Graph Transformer that reconstructs graph connectivity by fusing visual, linguistic, and geometric cues into dynamic edge weights, enabling the agent to filter out topological noise and enhancing instruction adherence. Extensive experiments on the R2R-CE and RxR-CE benchmarks demonstrate DGNav exhibits superior navigation performance and strong generalization capabilities. Furthermore, ablation studies confirm that our framework achieves an optimal trade-off between navigation efficiency and safe exploration. The code is available at https://github.com/shannanshouyin/DGNav.

Figures (9)

• Figure 1: The overall framework of DGNav. DGNav dynamically adjusts its navigation strategy according to the estimated scene complexity $\sigma$. Specifically, higher scene complexity leads to denser topological graph construction, while simpler environments favor sparser representations. The graph merging threshold $\gamma$ controls the resulting graph granularity and is inversely correlated with $\sigma$, enabling adaptive trade-offs between navigation safety and efficiency.
• Figure 2: Illustration of the Scene-Aware Adaptive Strategy. The process begins with Waypoint Prediction, generating raw ghost nodes from the depth map. These nodes are then passed to the Adaptive Graph Update module. Based on the angular dispersion ($\sigma$) of the candidates, the controller dynamically adjusts the merging threshold $\gamma$. The figure demonstrates three scenarios: in simple environments (low $\sigma$), a larger $\gamma$ yields a sparse graph for efficiency; in complex environments (high $\sigma$), a smaller $\gamma$ results in a dense graph for safety.
• Figure 3: Statistical distribution of angular dispersion ($\sigma_t$) for parameter calibration. Data is collected from the ETPNav baseline on the R2R-CE Val-Seen split. The distribution exhibits a Gaussian-like profile, which serves as the empirical basis for calibrating $\alpha$ and $\beta$.
• Figure 4: The architecture of Multimodal Encoding and Dynamic Edge Fusion. The visual encoder and instruction encoder extract node features ($\mathbf{V}$) and word features ($\mathbf{W}$), respectively. The Dynamic Edge Fusion module constructs the graph connectivity by fusing the geometric map ($\mathbf{E}_{\mathbf{geo}}$), pairwise visual similarity ($\mathbf{E}_{\mathbf{sem}}$), and instruction relevance ($\mathbf{E}_{\mathbf{inst}}$ derived from $\mathbf{W}_\mathbf{L}$). The resulting matrix $\mathbf{E}_{\mathbf{dynamic}}$ guides the Graph Transformer to perform context-aware planning
• Figure 5: Distribution of Dynamic Thresholds ($\gamma_t$) under Different Mapping Strategies on R2R-CE Val-Unseen. (a) Global Linear Mapping exhibits a dispersed distribution, indicating frequent but potentially noisy adjustments. (b) Conditional Linear Mapping shows a long-tailed distribution, reflecting a selective intervention strategy. Note: To clearly visualize the distribution of dynamic adjustments, the dominant peak at the baseline ($\gamma=0.5$) is omitted; data at the lower bound ($\gamma=0.25$) is also excluded to maintain visual symmetry.