A Sparse Topometric Graph for Scalable Cross-Session Indoor Navigation
This paper tackles cross-session indoor navigation, where an agent must reach a prior-session goal despite scene changes, sensor differences, and partial overlap. The key idea is a sparse topometric graph whose nodes compress local 3D geometry around traversable anchors and whose edges encode traversable distance, enabling both memory-efficient storage and learned graph matching for alignment. The approach matters because it promises to slash map memory by over two orders of magnitude while outperforming dense grids, appearance maps, and semantic graphs on standard benchmarks.
The paper presents a clean and pragmatic representation for cross-session indoor navigation. The topometric graph effectively balances metric detail and topological abstraction, and the empirical gains on Gibson ($77.14\%$ SR) and 3D-FRONT scene changes ($40.0\%$ SR) are substantial. However, the real-world validation is only a correspondence demonstration, not closed-loop control, and the scope is explicitly limited to single-floor settings with moderate topological stability. It is a solid step toward scalable map reuse, but the boldest claims about real-world transfer remain partially unsubstantiated.
The fixed-radius node descriptor is a strong design choice: by anchoring geometry to graph nodes rather than raw points, the representation achieves impressive compactness while keeping enough local context for matching. Memory use falls to $0.02$ MB, more than two orders of magnitude smaller than point-level and grid baselines. The graph-matching pipeline cleanly exploits both node geometry and connectivity to disambiguate repetitive structure, and the evaluation spans odometry noise, large-scale HM3D scenes, and 3D-FRONT furniture changes.
The real-world study highlights correspondence quality but does not test navigation, and the paper admits that severe topological changes such as blocked corridors were not evaluated. Baseline comparisons are uneven: ObjectReact uses a different action space, making its SPL incomparable, yet it appears in the same table without adjustment. The method also trains only on Gibson and applies zero-shot to HM3D and 3D-FRONT without reporting an in-dataset training baseline, so it is hard to gauge how much cross-dataset generalization costs. Additionally, the online system still maintains a dense occupancy map for local planning, so the dramatic memory reduction applies only to the reference-session graph, not the full runtime footprint.
The results generally support the claim that graph-structured matching outperforms local geometry or appearance alone, especially under repetitive structure. On Gibson the method leads the baselines under noiseless odometry, but the SPL drops sharply from $47.71$ to $23.64$ under noisy odometry while SR only falls to $60.00\%$, indicating that path efficiency suffers even when the representation still enables eventual success. The 3D-FRONT scene-change evaluation uses a tighter $1.0$ m threshold justified as testing local reusability rather than planning, which changes the task objective relative to the $2.0$ m threshold on Gibson and HM3D and complicates direct cross-dataset conclusions.
Reproducing the work is hindered by the absence of code, training splits, and full hyperparameters in the main text; several critical details are deferred to an unreleased supplementary material. Key choices such as the voxel downsampling resolution, Gaussian noise parameters for edge-weight augmentation, and exact loss formulation for IPCA-GM training are omitted. Baseline adaptations are only summarized as being faithful to the original representation logic, but without specifics an independent replication of the comparison pipeline would be difficult. The reliance on standard Habitat and LoCoBot settings helps, but the full experimental stack is not yet open.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.