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

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.

💡 Research Summary

Vision‑Language Navigation in Continuous Environments (VLN‑CE) requires an agent to translate high‑level natural language instructions into low‑level continuous actions while maintaining safety and long‑horizon consistency. Existing topological‑map approaches provide explicit spatial memory but suffer from a “Granularity Rigidity” problem: they rely on a fixed distance threshold (γ) to decide whether a newly predicted waypoint should be merged into the existing graph or added as a new node. This static rule leads to two opposite failures. In simple, low‑uncertainty regions (e.g., straight corridors) the graph becomes overly dense, wasting computation, whereas in complex or open areas the graph becomes too sparse, depriving the planner of sufficient candidate waypoints and increasing collision risk. Moreover, conventional Graph Transformers use only static geometric distances as bias terms, causing the planner to attend to geometrically close but semantically irrelevant nodes—a phenomenon the authors call “Navigational Myopia”.

DGNav (Dynamic Graph Navigation) addresses both issues with two complementary mechanisms.

1. Scene‑Aware Adaptive Strategy

   • At each navigation step t, a waypoint‑prediction module generates a set of candidate ghost nodes Cₜ from the depth map. For each candidate i, the relative heading angle θᵢ (with respect to the agent’s forward direction) is computed.
   • The angular dispersion σₜ = sqrt( (1/N₍c₎) Σᵢ (θᵢ – ȳθ)² ) quantifies scene complexity: high σₜ indicates a spread of candidates (e.g., an intersection), low σₜ indicates a narrow forward‑focused distribution (e.g., a corridor).
   • A linear inverse control law γₜ = Clip(α – β·σₜ, γ_min, γ_max) dynamically adjusts the merging threshold. When σₜ is high, γₜ shrinks, forcing the graph to become denser; when σₜ is low, γₜ expands, keeping the graph sparse.
   • The coefficients α and β are not hand‑tuned; they are derived from a statistical calibration on a fully converged ETPNav baseline run over the Val‑Seen split. The distribution of σₜ is approximately Gaussian, justifying the linear mapping and providing a principled way to set α, β, γ_min, and γ_max.

2. Dynamic Graph Transformer

   • Instead of static Euclidean distance bias, DGNav fuses three modalities for each edge (i, j): (a) geometric distance dᵢⱼ, (b) visual similarity sᵢⱼ (computed from RGB‑D panoramic features), and (c) linguistic relevance lᵢⱼ (derived from attention between node visual embeddings and the instruction embedding).
   • These cues are concatenated and passed through a shallow MLP to produce a dynamic edge weight eᵢⱼ = MLP(