PaperlayerView exact revision on arXiv

Source-linked AI summary

A Sparse Topometric Graph for Scalable Cross-Session Indoor Navigation

Eun Sun Lee, Minkyun Seo, Gwangtak Bae, Junho Kim, Seunggu Kang, Jaesik Park, Young Min Kim

arXiv:2610.31432v1cs.ROcs.CV

TL;DR

Indoor cross-session map reuse must accommodate scene and sensing variation, partial overlap, and unknown relative poses. The paper represents traversable space as a sparse topometric graph and matches graphs using local geometry and connectivity. It reports the highest Gibson success rate under both odometry conditions and the highest success rate and SPL on 3D-FRONT, while identifying clear evaluation-scope boundaries.

  • Problem

    Indoor map reuse must handle scene changes, different sensor configurations, and unknown relative poses, while detailed geometry can require extensive spatial memory.

    Source: §1 · p. 2

  • Method

    The method builds a sparse graph with local 3D geometry descriptors and distance-weighted traversable connectivity, then matches nodes using geometry and graph structure for localization and navigation.

    Source: §1 · p. 2, §3.1 · p. 4, §3.2 · p. 5

  • Results

    On Gibson, the method achieves the highest SR under both odometry conditions and best SPL under noiseless odometry; on 3D-FRONT, it achieves the highest SR and SPL.

    Source: §4.3 · pp. 6–7

  • Takeaways & Limitations

    The experiments support the graph as an abstraction for cross-session indoor localization and goal-directed navigation under odometry noise, limited overlap, large-scale scenes, and environmental changes.

    Source: §6 · p. 8

    The evaluation covers single-floor 2D top-down point-goal navigation, does not systematically test severe connectivity changes, and uses independently captured real-world data rather than live onboard deployment.

    Source: §5 · pp. 7–8

Abstract

from arXiv · show

A conventional embodied agent in an indoor environment builds a map by accumulating noisy real-world measurements and results from heavy inference to guide navigation. While the detailed measurements may differ due to scene changes or sensor settings, the unique topological structure of navigable passages or the connectivity between rooms may provide training-free guidance that complements local observations. We propose a sparse topometric graph as an efficient structural representation for robust indoor navigation, alleviating the need for fine-grained matching, which is computationally expensive, memory-intensive, and prone to real-world variations. The map consists of nodes that encode local 3D geometry and distance-weighted edges that capture traversable connectivity. A learned graph-matching network finds consistent subgraph structures across sessions by considering both node-level geometric descriptors and neighborhood connectivity. The holistic contextual matching is robust to local ambiguities caused by weak distinctiveness or repetitive structure, and establishes stable correspondences despite local scene changes, different sensor modalities, or limited spatial overlap. We analyze the practicality of the graph through cross-session point-goal navigation, where a goal specified in a prior reference map must be reached from an unknown initial pose in a later session. Experiments show that our method achieves the highest success rates (77.1% on Gibson and 40.0% under 3D-FRONT scene changes), while reducing memory by more than two orders of magnitude compared to point-level representations. The proposed approach further supports cross-session alignment and goal-directed navigation with strong scalability and robust real-world transfer.

1. Introduction

Indoor map reuse is difficult when scenes, sensors, and starting poses vary, while dense maps are costly and other compact representations can be sensitive to change. The paper proposes a sparse topometric graph and learned graph matching for cross-session alignment and navigation.

1. Introduction

  • The method targets map reuse despite scene changes, different sensor configurations, and unknown relative poses between sessions.Indoor mapping also faces complex concavities, occlusions, and confined passages.Source: §1 · p. 2
  • Sparse topometric graphs encode traversable space with fixed-radius local geometry descriptors and distance-weighted connectivity, keeping map size proportional to graph nodes.The descriptors summarize local point sets at graph nodes rather than tying geometry to raw point samples.Source: §1 · p. 2
  • The paper evaluates whether one sparse representation supports both cross-session alignment and downstream point-goal navigation.The task starts from an unknown pose and specifies the goal on a reference map built in a prior session.Source: §1 · p. 2
  • A learned graph matcher predicts cross-session node correspondences from local geometry and neighboring connectivity, reducing ambiguity from repeated structures and partial overlap.The correspondences support relative pose estimation and transfer of the reference-map goal for waypoint navigation.Source: §1 · p. 2
  • The contributions combine a compact geometry-only graph, fixed-radius node descriptors, learned graph matching, and evaluation across simulated, large-scale, changed-scene, and real-world settings.The same graph supports cross-session alignment and waypoint selection.Source: §1 · p. 2

2. Related Work

Prior navigation maps trade detailed geometry against compactness and resilience to change. This work combines sparse geometry anchored to traversable regions with graph structure for cross-session localization and planning.

2.1. Map Representations for Navigation

  • Metric maps preserve geometric detail but incur increasing storage and matching costs, while appearance and semantic representations have distinct sensitivities or dependencies.Appearance-based maps can degrade under viewpoint and scene changes; semantic and hierarchical scene graphs often depend on detections, labels, or predefined categories.Source: §2.1 · pp. 2–3
  • The proposed sparse topometric graph anchors sparse 3D geometric descriptors to navigable areas and supports both cross-session localization and goal-directed planning.It is presented as a compact alternative between fine-grained point-level geometry and higher-level map abstractions.Source: §2.1 · p. 3

2.2. Graph-based Correspondence

  • Graph-based methods use relational structure to help resolve ambiguities that local features alone cannot resolve.Prior examples include factor graphs for multi-session SLAM and learned topological maps with nodes for visited locations and edges for navigability.Source: §2.2 · p. 3
  • The proposed graph represents traversable regions and their reachability, enabling holistic matching that is robust to local-feature perturbations and moderate topological deviations.This is framed as a response to object-level graph dependence on stable semantic detection and point-level methods that do not operate over navigable spatial regions.Source: §2.2 · p. 3

3. Method

The method builds a sparse graph of navigable anchors, matches reference and current graphs using geometry and connectivity, and reuses correspondences for localization and waypoint selection. Its deployment loop combines transferred goal information with frontier exploration.

3. Method

  • The deployment task uses a prior-session graph and a goal in its coordinate frame, while the current session may have partial overlap, odometry drift, or changed local arrangements.The agent receives depth and odometry, incrementally builds geometric maps, and converts them into a current-session graph every K=20 steps.Source: §3 · p. 4

3.1. Sparse Topometric Graph

  • Each topometric graph node is a navigable anchor with a 2D position and sparse 3D descriptor, while edges encode adjacent traversable connectivity with local-distance weights.Nodes are extracted from reachable free space at uniform intervals of 5 cells (0.25 m); descriptors use points within a fixed radius and a frozen rotation-equivariant encoder.Source: §3.1 · p. 4

3.2. Correspondence Estimation

  • Graph matching predicts partial node assignments by jointly using local geometry and connectivity, allowing unmatched nodes under partial overlap.Training pairs come from partial traversals of the same scene, use spatial proximity for correspondence supervision, and perturb edge distances to simulate odometry drift.Source: §3.2 · p. 5, Figure 3 · p. 5

3.3. Localization and Waypoint Selection

  • RANSAC-filtered matches estimate the relative session transform, and the navigation policy propagates reference-goal scores to current graph nodes while balancing them with frontier exploration.The policy initially uses exploration alone, then combines goal-directed and exploration scores once cross-session evidence is reliable; a Fast Marching Method planner executes the selected waypoint.Source: §3.3 · p. 5

4. Experiments

The experiments evaluate cross-session point-goal navigation across odometry noise, environment scale, scene change, compactness, and real-world data. The results report strong navigation performance alongside a compact representation and graph-based correspondence under ambiguity.

4.1. Experimental Setup

  • Evaluation spans Gibson under noiseless and noisy odometry, HM3D at large scale, and 3D-FRONT under scene changes, using goals on a prior-session map and actions in the current session.The graph-matching network is trained on Gibson training scenes and applied without retraining across datasets; 3D-FRONT uses a 1.0 m success threshold.Source: §4.1 · p. 6

4.2. Baselines

  • Baselines cover dense occupancy grids, top-down visual maps, appearance retrieval, scene or object graphs, and point-level geometry in the compactness analysis.Several baselines are adapted to the cross-session protocol while preserving their original representation and decision logic as much as possible.Source: §4.2 · p. 6

4.3. Navigation Performance

  • Our method achieves the highest Success Rate under both Gibson odometry conditions, the best SPL under noiseless odometry, and the highest SR and SPL on 3D-FRONT scene changes.Table 2 covers Gibson under noiseless and noisy odometry; Table 3 covers 3D-FRONT scene changes.Source: Table 2 · p. 6, §4.3 · pp. 6–7, Table 3 · p. 7
  • On HM3D, Figure 4 evaluates cross-session navigation by plotting cumulative successful episodes out of 100 as geodesic separation between reference- and current-session start positions increases.This separation is a controlled difficulty proxy; larger separation generally makes useful overlap harder to recover early in the episode.Source: §4.3 · p. 6, Figure 4 · p. 7

4.4. Compactness

  • The stored representation is substantially smaller than dense occupancy grids, dense feature maps, and point-level geometry because it stores geometry at sparse graph-node anchors plus traversable connectivity.Table 4 compares stored reference-session representations on Gibson.Source: §4.4 · p. 7, Table 4 · p. 7

4.5. Real-World Data Demonstration

  • A real-world correspondence study uses sessions recorded weeks apart with handheld LiDAR and phone LiDAR, collected at different times of day with only partial overlap.In a repeated office layout, graph-structured matching identifies the correct correspondence where geometry-only and DINOv2 appearance-only cues leave multiple plausible candidates.Source: §4.5 · p. 7

5. Limitations and Scope

The evaluation covers single-floor indoor navigation with a 2D top-down point-goal formulation, including partial overlap and moderate scene changes. Multi-floor connectivity, severe changes to traversable structure, and live onboard deployment remain outside the demonstrated scope.

5. Limitations and Scope

  • Multi-floor environments and explicit vertical connectivity are outside the evaluation, which uses single-floor 2D top-down point-goal navigation.Extending the representation to multi-floor settings is identified as future work.Source: §5 · p. 7
  • Closed doors, blocked corridors, and large structural reconfigurations were not systematically evaluated.The study focuses on partial overlap, sensing variation, and moderate scene change.Source: §5 · p. 7
  • Qualitative examples include cross-session navigation in HM3D and 2D-3D-S, plus scene-change examples in Gibson and 3D-FRONT.The examples depict partial overlap and local environmental changes.Source: Figure 5 · p. 8
  • The real-world evaluation demonstrates independently captured data, not full onboard deployment with live control, sensing, and recovery behaviors.The authors describe this as a boundary on what the current experiments prove.Source: §5 · p. 8

6. Conclusion

The sparse topometric graph is presented as an effective abstraction for cross-session indoor localization and goal-directed navigation, with reliable performance under odometry noise, limited overlap, large-scale scenes, and environmental changes.

6. Conclusion

  • The representation supports cross-session localization and goal-directed navigation, with reliable performance under odometry noise, limited overlap, large-scale scenes, and environmental changes.The conclusion presents it as a practical abstraction for cross-session map reuse.Source: §6 · p. 8

Related papers

Loading 2610.31432v1…