acceptodds
Under review as a conference paper at ICLR 2027

SPM-Nav: Persistent Space–Portal Memory for Zero-Shot Object Navigation

Abstract

Zero-shot object navigation requires long-horizon search in unseen environments, where effective decisions depend on both target likelihood and accumulated search history. However, existing spatial representations often leave search evidence and progress fragmented, limiting the persistence and reuse of search experience across revisits. To address this issue, SPM-Nav presents a zero-shot object navigation framework centered on Persistent Space–Portal Memory. Each Space maintains a persistent regional search state that consolidates regional semantics, target evidence, historical observations, negative evidence, and search progress across locations and viewpoints. Portals represent traversal-validated inter-Space transitions and preserve reusable connectivity between regions. Building upon this Memory, Adaptive Space Search, which dynamically adjusts local search effort according to accumulated Space states, reducing redundant inspection while preserving essential inter-region connectivity. SPM-Nav is evaluated on HM3D-v1, HM3D-v2, and MP3D, achieving success rates of 70.6%, 83.8%, and 55.7%, respectively. Component ablations and mechanism analyses further validate the effectiveness of persistent search states, reusable inter-Space transitions, and memory-conditioned adaptive allocation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Zero-Shot Object Navigation, Vision-Based Navigationl, Embodied AI, ersistent Space-Portal Memory

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_urf8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SPM-Nav proposes a training-free zero-shot object navigation framework that maintains persistent region-level search states (Spaces) and traversal-validated inter-region connections (Portals). It uses accumulated search evidence and target relevance to adapt local exploration effort while preserving connectivity to unexplored regions. Experiments on HM3D-v1, HM3D-v2, and MP3D show strong success rates, with ablations highlighting the benefits of persistent portal connectivity.

Critical strengths
  1. The paper addresses a practical challenge in long-horizon object navigation by explicitly preserving search progress and inter-region connectivity.

  2. The proposed framework is reasonable.

  3. The method achieves strong success rates across three benchmarks, and the ablation studies provide evidence for the effectiveness of persistent portal connectivity.

Critical weaknesses
  1. Insufficient differentiation from existing spatial-memory methods. TopoNav already supports the accumulation, updating, and retrieval of spatial knowledge, while EvoMemNav maintains hierarchical room–view–object memory and continuously updates navigation priors through experience. The paper does not sufficiently explain why these existing memory mechanisms cannot achieve similar functionality or what unique advantages the proposed Space–Portal Memory provides.

  2. High potential tuning costs due to extensive hand-crafted heuristics. The framework relies on numerous manually specified thresholds, weights, and search budgets for Space construction, memory association, and adaptive exploration, potentially requiring substantial tuning across environments. The parameter sensitivity analysis is needed to demonstrate the robustness of these choices or assess their tuning requirements.

Questions for authors

Please see weaknesses. I'm willing to raise my score if these concerns are adressed.

5 comments
AuthorsOpenReview ↗

Response to Reviewer urf8 (R3-1)

We are very grateful to Reviewer urf8 for the positive assessment of our work and the constructive suggestions for further improving the paper. We appreciate the recognition of the practical challenge addressed by SPM-Nav, the reasonableness of the proposed framework, its strong success rates across three benchmarks, and the experimental evidence supporting the effectiveness of persistent Portal connectivity.

We understand the reviewer's two main concerns regarding the differentiation from existing spatial-memory methods and the potential tuning costs of hand-crafted heuristics. Accordingly, we address these concerns by (1) clarifying the unique advantages of Space–Portal Memory over existing spatial-memory representations, particularly in persistent regional search-state management; and (2) providing systematic parameter sensitivity analyses and cross-dataset parameter-reuse results to evaluate robustness and tuning requirements.

R3-1. Differentiation from Existing Spatial-Memory Methods

Response:

We acknowledge that TopoNav and EvoMemNav have introduced effective mechanisms for organizing and updating spatial knowledge. We do not claim that these methods cannot be extended to support similar functionalities. Rather, our distinction lies in their memory representations and decision mechanisms. SPM-Nav explicitly maintains persistent, target-conditioned search states at the physical-region level and uses them to guide subsequent exploration.

(1) Compared with TopoNav and EvoMemNav: Physical-region-level search-state management.

TopoNav organizes spatial and semantic information around robot-reached waypoints, while EvoMemNav adopts a Room–View–Object hierarchy to organize visual memories and update navigation priors. In contrast, SPM-Nav uses the physical region itself as a persistent memory unit. Each Space aggregates observations across multiple positions and viewpoints and maintains target relevance, accumulated negative evidence, and consumed local-search steps. These states determine the remaining search budget and early-stopping conditions, allowing target-specific search progress to be preserved and reused across revisits.

(2) Inter-Space Portals: Explicit transition nodes and reusable connectivity.

Unlike conventional topological representations that directly connect spatial nodes through edges, SPM-Nav represents each inter-region transition as an independent Portal node between two Space nodes, forming an explicit Space–Portal–Space structure. Each confirmed Portal records its physical transition location, connected Spaces, navigable anchors on both sides, and traversal-confirmation state. Importantly, traversal confirmation determines when the active Space switches: observations remain associated with the source Space before confirmation and are assigned to the destination Space afterward. Confirmed Portal nodes can subsequently be reused for inter-region Transit and navigation recovery, preserving access to downstream regions even after local inspection terminates.

(3) Memory-conditioned Adaptive Space Search and experimental evidence.

Building upon Space–Portal Memory, Adaptive Space Search jointly considers target relevance, previously consumed search effort, and accumulated negative evidence to determine local-search budgets and early-stopping conditions. It further separates the decision to continue local inspection from the decision to traverse the current region toward unexplored or relevant Spaces.

Taken together, the contribution of SPM-Nav is not merely the combination of spatial memory, topological connectivity, and heuristic budget allocation. Rather, it establishes a persistent regional search-control mechanism in which target-conditioned search experience is maintained across revisits and directly governs both local inspection and inter-region navigation. In this framework, memory is not merely used to retrieve spatial knowledge, but serves as an explicit state representation for subsequent search and navigation decisions.

We have revised Related Work to clarify the differences from TopoNav and EvoMemNav, and further specified Space identity maintenance, Portal construction and validation, and memory-conditioned search control in Sections 3.2 and 3.3 and Appendix A.

Overall, SPM-Nav uses persistent regional search states as explicit control variables, coupling accumulated search experience with reusable physical connectivity to guide subsequent exploration.

AuthorsOpenReview ↗

Response to Reviewer urf8 (R3-2)

Question:

High potential tuning costs due to extensive hand-crafted heuristics. The framework relies on numerous manually specified thresholds, weights, and search budgets for Space construction, memory association, and adaptive exploration, potentially requiring substantial tuning across environments. The parameter sensitivity analysis is needed to demonstrate the robustness of these choices or assess their tuning requirements.

Response:

We agree that the manually specified parameters in SPM-Nav require sensitivity analysis to assess their impact on navigation performance and the need for environment-specific tuning. Accordingly, we have added a systematic parameter sensitivity study on the full 1,000-episode HM3D-v2 benchmark (see Appendix E). We evaluate ten representative parameters or coupled parameter groups across three modules, with two alternative configurations for each, resulting in 20 additional configurations. All other parameters remain fixed at their default values.

The results reveal different levels of sensitivity across the three modules:

(1) Space Construction and Association. Performance variations are relatively small, with SR ranging from 82.7% to 83.3%, SPL from 36.6% to 37.2%, and average executed steps from 222.31 to 233.42 (shown in Table 10).

(2) Relevance Estimation and Adaptive Search. Performance variations are moderate, with SR ranging from 82.5% to 84.2%, SPL from 36.7% to 37.3%, and average executed steps from 207.51 to 232.18 (shown in Table 11).

(3) Negative Evidence and Early Stopping. These parameters exhibit relatively larger variations, with SR ranging from 80.9% to 84.3%, SPL from 36.0% to 37.4%, and average executed steps from 198.35 to 240.62 (shown in Table 12). These results highlight the trade-off between early termination and sufficient regional inspection.

Notably, the default configuration achieves an SR of 83.8%, an SPL of 37.6%, and 218.17 average executed steps, providing a favorable balance within the tested parameter ranges.

Regarding cross-environment tuning, we have explicitly clarified in Section 4 Implementation Details that "All algorithmic parameters are fixed across HM3D-v1, HM3D-v2, and MP3D, without scene-specific or dataset-specific hyperparameter tuning." We have also reorganized the parameter table (see Table 5) to distinguish algorithmic hyperparameters, geometric and perception thresholds, and execution and evaluation settings.

Overall, these additional experiments demonstrate that SPM-Nav exhibits limited-to-moderate performance variations across most tested parameter settings, with greater sensitivity to negative-evidence and early-stopping parameters. The cross-dataset evaluation (see in Table 2 and Section 4.2 Ablation Studies) further confirms that a shared default configuration can be used across all three benchmarks without additional environment-specific tuning.

##Table 2. Comparison with a matched LLM-free frontier baseline and core-component ablations on HM3D-v2 and MP3D.

VariantPSMPortalASSHM3D-v2 SR ↑HM3D-v2 SPL ↑HM3D-v2 Avg. Steps ↓MP3D SR ↑MP3D SPL ↑MP3D Avg. Steps ↓
Matched LLM-free Frontier Baseline✗✗✗70.730.2263.3938.314.8351.47
w/o Persistent Memory✗✓✓82.137.2231.5753.624.2322.83
w/o Portals✓✗✓79.135.6234.1949.722.5330.16
w/o Adaptive Search✓✓✗81.836.4226.1752.423.6315.59
Full SPM-Nav✓✓✓83.837.6218.1755.724.7301.86

We hope that the clarified distinctions from TopoNav and EvoMemNav, together with the additional parameter sensitivity analyses and cross-dataset parameter-reuse experiments, adequately address the two concerns raised by the reviewer. We appreciate the reviewer's time and constructive feedback, which have helped us improve the clarity and completeness of this work.

If the clarifications and additional evidence address your concerns, we hope you will consider updating your overall assessment of the paper accordingly.

Reviewer_urf8OpenReview ↗

Response by Reviewer urf8.

Thank you for your response. My concern has been addressed. I will maintain my positive score.

Reviewer_3Z2mOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets zero-shot object navigation, where search evidence tied to transient frontiers is fragmented or lost at revisits. SPM-Nav introduces a Persistent Space-Portal Memory: a Space is a region-level unit holding room semantics, EWMA target relevance, independent observations, negative evidence, and consumed budget, while a Portal is a traversal-validated transition with navigable anchors that survives revisits. Adaptive Space Search uses hysteretic LOW/MEDIUM/HIGH priorities, remaining budget, and negative evidence to switch among Peek, Probe, Search, Transit, and Backtrack. The pipeline is training-free and LLM-free, built on Places365, RAM, D-FINE, Grounding DINO, RedNet, BLIP-2, and DD-PPO. It reports 70.6/34.8 on HM3D-v1, 83.8/37.6 on HM3D-v2, and 55.7/24.7 on MP3D; on HM3D-v2, removing Portals costs 2.0 SPL and 16 steps, and removing the persistent state or adaptive search costs 13.4 and 8.0 steps. The contribution is a region-level memory that separates whether a region merits further search from how to reach other regions, replacing frontier bookkeeping with persistent states. The results suggest a modular, LLM-free pipeline can exceed LLM-based training-free systems on success rate, at some cost in path efficiency.

Critical strengths
  1. The Space-Portal abstraction is coherent: separating persistent regional state from reusable connectivity addresses a real weakness of frontier-based memory, and the Peek/Probe/Search vocabulary keeps the policy interpretable.
  2. Success rates of 70.6, 83.8, and 55.7 from a training-free, LLM-free pipeline are strong, and the 5.2-point gain over ASCENT on HM3D-v1 shows memory can substitute for LLM reasoning.
  3. The evaluation is more rigorous than typical here: paired bootstrap intervals and McNemar tests for ablations, plus mechanism-level diagnostics of state retention, Portal reuse, and budget allocation.
Critical weaknesses
  1. The statistics do not support the central claim, since the SPL effects of persistent memory (-0.40 [-1.11, 0.29]) and adaptive search (-1.20 [-2.49, 0.07]) include zero, SR is tested only for Portals, and all ablations are HM3D-v2 only.
  2. Efficiency does not hold against memory baselines, since SPL trails ApexNav (38.0) and EvoMemNav (39.4) on HM3D-v2 while the ablation gains are almost entirely in step count.
  3. Portal evidence is thin, as 22 of 50 diagnostic episodes contain no Portal and only 19 of 53 Portals are reused, while w/o Portals also removes Transit and Backtrack, confounding memory with behavior.
  4. Table 1 needs verification, because InstructNav carries identical 58.0/20.9 values for both HM3D splits, the MTU3D and OVSegDT rows are empty, and recent methods that report on the same MP3D split, such as DRIVE-Nav, are not compared.
  5. The memory is never isolated from the base pipeline, and the hand-set Space thresholds have no sensitivity analysis. Besides, the provenance of the target-room prior is unclear.
Questions for authors

See the weakness.

12 comments
AuthorsOpenReview ↗

Response to Reviewer 3Z2m (R2-1)

We appreciate Reviewer 3Z2m's careful assessment of our work and the valuable suggestions for strengthening the paper. We particularly appreciate the recognition of the coherent Space–Portal design, which separates persistent regional search states from reusable inter-region connectivity. We also appreciate the positive assessment of SPM-Nav's success rates across three benchmarks, especially the 5.2-percentage-point improvement over ASCENT on HM3D-v1, as well as the paired statistical tests and mechanism-level analyses supporting the proposed design.

We understand the reviewer's concerns regarding statistical significance and cross-dataset validation, path efficiency, Portal reuse and ablation configurations, benchmark comparison completeness, and parameter robustness. Accordingly, we address these concerns in five parts: (1) strengthening statistical validation and cross-dataset component ablations while carefully qualifying component contributions and path-efficiency claims; (2) clarifying Portal reuse statistics and ablation configurations; (3) verifying benchmark results and incorporating recent methods, including DRIVE-Nav; (4) introducing a matched LLM-free frontier baseline to isolate the overall contribution of the proposed memory mechanisms; and (5) providing systematic parameter sensitivity analyses and clarifying the provenance and usage of the Target--Room Prior.

Question:

The statistics do not support the central claim, since the SPL effects of persistent memory (-0.40 [-1.11, 0.29]) and adaptive search (-1.20 [-2.49, 0.07]) include zero, SR is tested only for Portals, and all ablations are HM3D-v2 only.

Response:

We sincerely thank the reviewer for the careful reading of our manuscript and for providing thoughtful and constructive suggestions. We have added two-sided exact McNemar tests for SR and 10,000 paired bootstrap resamples for 95% confidence intervals of SPL and average executed steps on HM3D-v2 (see Appendix B). We have also extended all three core-component ablations to MP3D (see Table 2). Meanwhile, we have revised the descriptions of individual component contributions according to the statistical results and no longer characterize SPL differences whose confidence intervals include zero as statistically significant improvements (see Section 4.2, Table 2, and Appendix B).

Table 7. Paired statistical analysis of core-component ablations on HM3D-v2.

VariantΔSPL [95% CI]ΔSteps [95% CI](n₁₀, n₀₁)McNemar p
w/o Persistent Memory−0.40 [−1.11, +0.29]+13.40 [+9.44, +17.48](41, 24)0.0464
w/o Portals−2.00 [−2.79, −1.23]+16.02 [+11.29, +20.84](70, 23)1.10 × 10⁻⁶
w/o Adaptive Search−1.20 [−2.49, +0.07]+8.00 [+0.82, +14.99](68, 48)0.0773

#ΔM is computed as M_ablated − M_Full. n₁₀ and n₀₁ denote episodes where full SPM-Nav succeeds but the ablated variant fails, and vice versa, respectively.

As shown in Table 7, on HM3D-v2, the paired SR tests yield the following results:

  • Persistent Memory: Removing this component decreases SR by 1.7 percentage points, with 41 versus 24 discordant episodes (unadjusted McNemar p = 0.0464).

  • Adaptive Search: SR decreases by 2.0 percentage points, with 68 versus 48 discordant episodes (p = 0.0773).

  • Inter-Space Portals: SR decreases by 4.7 percentage points, with 70 versus 23 discordant episodes (p = 1.10 × 10⁻⁶).

For SPL, the differences after removing Persistent Memory and Adaptive Search are -0.40 percentage points (95% CI: [-1.11, +0.29]) and -1.20 percentage points (95% CI: [-2.49, +0.07]), respectively. Both confidence intervals include zero. We therefore do not claim statistically significant SPL improvements for these two components. In contrast, removing Portals decreases SPL by 2.00 percentage points (95% CI: [-2.79, -1.23]) and increases average executed steps by 16.02 (95% CI: [11.29, 20.84]).

As shown in Table 2, on MP3D, removing Persistent Memory, Portals, and Adaptive Search decreases SR by 2.1, 6.0, and 3.3 percentage points, and SPL by 0.5, 2.2, and 1.1 percentage points, respectively. Average executed steps increase by 20.97, 28.30, and 13.73, respectively. These results show consistent performance trends across MP3D and HM3D-v2, with Portal removal causing the largest degradation.

Accordingly, we have refined the contribution claims in the revised Section 4.3: Persistent Memory primarily reduces redundant inspection through historical-state reuse, Adaptive Search improves local search-resource allocation, and Portals provide the clearest contribution to overall navigation performance. We have avoided overstating the statistical significance of individual component contributions.

AuthorsOpenReview ↗

Response to Reviewer 3Z2m (R2-2)

Question:

Efficiency does not hold against memory baselines, since SPL trails ApexNav (38.0) and EvoMemNav (39.4) on HM3D-v2 while the ablation gains are almost entirely in step count.

Response:

We acknowledge the path-efficiency gap identified by the reviewer. SPM-Nav achieves an SPL of 37.6% on HM3D-v2, below ApexNav (38.0%) and EvoMemNav (39.4%). We therefore do not claim superior path efficiency over these methods. Instead, we have clarified in Section 4.1 that the primary advantage of SPM-Nav lies in achieving the highest SR on both HM3D splits and the best overall performance on MP3D, while maintaining competitive path efficiency, as shown in Table 1.

Regarding component ablations as shown in Table 2, removing Persistent Memory and Adaptive Search increases average executed steps by 13.40 (95% CI: [9.44, 17.48]) and 8.00 (95% CI: [0.82, 14.99]), respectively. Although both corresponding SPL confidence intervals include zero, the step-count differences are statistically supported. These results suggest that the two components primarily reduce redundant inspection and improve local search-resource allocation, rather than providing statistically significant SPL gains. We have revised the corresponding claims in Section 4.2 accordingly.

Table 2. Comparison with a matched LLM-free frontier baseline and core-component ablations on HM3D-v2 and MP3D.

VariantPSMPortalASSHM3D-v2 SR ↑HM3D-v2 SPL ↑HM3D-v2 Avg. Steps ↓MP3D SR ↑MP3D SPL ↑MP3D Avg. Steps ↓
Matched LLM-free Frontier Baseline✗✗✗70.730.2263.3938.314.8351.47
w/o Persistent Memory✗✓✓82.137.2231.5753.624.2322.83
w/o Portals✓✗✓79.135.6234.1949.722.5330.16
w/o Adaptive Search✓✓✗81.836.4226.1752.423.6315.59
Full SPM-Nav✓✓✓83.837.6218.1755.724.7301.86

Table 1. Comparison with prior methods on HM3D-v1, HM3D-v2, and MP3D.

MethodTraining-freeVenueVision Foundation ModelLanguage Foundation ModelHM3D-v1 SR ↑HM3D-v1 SPL ↑HM3D-v2 SR ↑HM3D-v2 SPL ↑MP3D SR ↑MP3D SPL ↑
Uni-NaVid✗RSS 2025————73.737.1——
ImagineNav✗ICLR 2025PolyOculusGPT-4o-mini53.023.8————
CompassNav✗ICLR 2026————61.627.842.017.5
TrajRAG✗CVPR 2026————78.140.242.618.0
CoW✓CVPR 2023——————9.24.9
L3MVN✓IROS 2023—RoBERTa-large50.423.136.315.734.914.5
VLFM✓ICRA 2024BLIP-2—52.530.463.632.536.417.5
SG-Nav✓NeurIPS 2024LLaVA-1.6-7BGPT-454.024.949.625.540.216.0
InstructNav✓CoRL 2024GPT-4VGPT-4——58.020.9——
BeliefMapNav✓NeurIPS 2025——————37.317.6
ApexNav✓IEEE RA-L 2025BLIP-2DeepSeek-V359.633.076.238.039.217.8
WMNav✓IROS 2025——58.131.272.133.345.417.2
TopoNav✓arXiv 2025GLEEGPT-4o60.134.6——45.516.8
MSGNav✓CVPR 2026GPT-4oGPT-4o63.031.474.133.4——
DRIVE-Nav✓IEEE RA-L 2026SAM3Qwen3-VL-8B55.535.972.441.341.822.6
ASCENT✓IEEE RA-L 2026BLIP-2Qwen2.5-7B65.433.5——44.515.5
EvoMemNav✓arXiv 2026YOLOv8-World + SAMQwen3-VL-8B59.233.663.839.4——
SPM-Nav (Ours)✓—BLIP-2—70.634.883.837.655.724.7

#Best results among training-free methods are highlighted in bold.

AuthorsOpenReview ↗

Response to Reviewer 3Z2m (R2-3)

Question:

Portal evidence is thin, as 22 of 50 diagnostic episodes contain no Portal and only 19 of 53 Portals are reused, while w/o Portals also removes Transit and Backtrack, confounding memory with behavior.

Response:

We appreciate the reviewer's concern regarding the interpretation of Portal statistics and the ablation configuration. We clarify the Portal statistics and ablation configuration to address these concerns.

First, among the 22 episodes without confirmed Portals in the 50-episode diagnostic subset, 4 terminate before establishing a valid Space, while the remaining 18 visit only one Space. Thus, no inter-region transition occurs in these episodes. Every episode visiting multiple Spaces establishes at least one confirmed Portal (Appendix D).

Second, among the 53 confirmed Portals, 19 are reused at least once, corresponding to a reuse rate of 35.85%, with 59 reuse events in total (see Table 4). The reuse rate measures the proportion of confirmed Portals reused at least once, whereas reuse events count the total number of repeated traversals.

Regarding the w/o Portals ablation, it removes persistent Portal representations and Portal-conditioned navigation, but does not disable general inter-region movement or Backtracking. Movement through the current map and frontiers, as well as trajectory-based navigation recovery, remains available. Consistent with this configuration, Backtracking events increase from 10 to 18, and Backtrack Steps increase from 175 to 332 after Portal removal (Table 4). Thus, this ablation evaluates the joint contribution of persistent Portal connectivity and its associated navigation control, rather than Portal storage alone.

To further isolate the role of Transit, we conduct a separate w/o Connectivity Preservation analysis (Appendix C), which retains the complete Space–Portal Memory but disables exit-oriented Transit after local inspection terminates. Downstream-Space reachability consequently decreases from 73.42% to 2.13%, while the number of distinct downstream Spaces reached decreases from 36 to 1.

Together, these results clarify the complementary roles of persistent Portal connectivity and connectivity-preserving Transit, while distinguishing their contributions from general navigation recovery.

Reviewer_BECaOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SPM-Nav, a training-free framework for zero-shot object-goal navigation that aims to reduce redundant exploration by maintaining persistent region-level search states and reusable inter-region connectivity. The proposed framework consists of three main components: 1) Persistent Space–Portal Memory: The environment is represented using persistent regions (Spaces) and traversal-validated connections (Portals). 2) Adaptive Space Search: The agent allocates local search budgets according to the estimated relevance of each Space to the target. 3) Connectivity-Preserving Transit: Even when further local inspection is deemed unnecessary, the agent can continue traversing a Space to reach unexplored regions through previously established Portals. The method builds upon an existing navigation pipeline and combines pretrained perception models, online geometric mapping, and a pretrained PointNav policy without task-specific training or fine-tuning on evaluation environments.

The reported success rates (SR) are 70.6% on HM3D-v1, 83.8% on HM3D-v2, and 55.7% on MP3D. The results suggest that persistent spatial memory and relevance-conditioned exploration can improve the reliability of zero-shot object navigation.

Critical strengths

1.‌The paper tackles a well-motivated problem: existing exploration methods tend to attach their search evidence and progress to transient frontiers or individual observations, which produces redundant exploration and ineffective revisits. A reasonable way to address this is to associate persistent search states with spatial regions while maintaining inter-region connectivity separately.

‌2.‌The responsibilities of Spaces, Portals, and Adaptive Space Search are clearly separated, which makes the design easy to interpret. It is also conceptually useful that the decision to stop inspecting a region is distinguished from the decision to traverse it, since this separation allows the agent to avoid unproductive local exploration without unnecessarily giving up access to unexplored regions. The overall design is therefore clear and well structured.

Critical weaknesses
  1. The two central ideas about persistent region-level memory and explicit inter-region connectivity are intuitive extensions of structured spatial memory and topological navigation. Combining them with relevance-conditioned search-budget allocation is reasonable, but the paper does not sufficiently establish how this combination differs fundamentally from existing methods. In particular, the paper discusses related approaches such as TopoNav, EvoMemNav, and ASCENT, but the distinctions are primarily presented in terms of system functionality. Several important questions remain insufficiently addressed:
  • Compared with existing hierarchical spatial memories involving rooms, viewpoints, or objects, what fundamentally new capability does the persistent Space state provide?
  • Beyond recording validated transitions and navigable anchors, what algorithmic novelty does a Portal introduce compared with an edge in an existing topological graph?
  • Does Adaptive Space Search introduce a principled decision mechanism beyond heuristic budget allocation based on target relevance?

As it stands, the main contribution appears to be a combination of reasonable engineering choices rather than a clearly isolated methodological innovation. The authors may identify the closest existing approaches and articulate the precise algorithmic differences, ideally supported by targeted comparisons.

2.Several key decisions, including historical-evidence scoring, inspection-mode selection, and Space identity maintenance, also need clearer formal definitions before the method can be reproduced. These components are currently described at a level of detail that leaves their exact behaviour open to interpretation and makes it hard to implement them faithfully, so stating them more precisely would considerably strengthen the reproducibility of the reported results.

3.The experimental validation is likewise incomplete. Some ablations show improvements in action efficiency, but they do not establish statistically significant SPL gains, and the mechanism analyses rest on a relatively small diagnostic subset. These two gaps together make it difficult to judge how robust the claimed advantages actually are, and the evaluation would benefit from broader and more rigorously tested evidence.

Questions for authors

1.What is the fundamental algorithmic novelty compared with existing topological navigation and spatial-memory methods?

2.Can the authors provide complete definitions of the key decision rules?

3.Can additional experiments better isolate the contributions of Persistent Memory and Adaptive Search?

6 comments
AuthorsOpenReview ↗

Response to Reviewer BECa (R1-1)

We sincerely thank Reviewer BECa for the careful evaluation and constructive feedback. We particularly appreciate the recognition of the long-horizon search challenge addressed in this work and the complementary roles of Spaces, Portals, and Adaptive Space Search.

We understand the reviewer's concerns regarding algorithmic novelty, reproducibility, and experimental validation. Accordingly, we organize our response into three parts: (1) clarifying the distinctions from existing methods; (2) formalizing key decision rules and state transitions; and (3) strengthening experimental validation through statistical tests, cross-benchmark ablations, mechanism analyses, and parameter sensitivity experiments.

R1-1. Fundamental Algorithmic Novelty

Question:

What is the fundamental algorithmic novelty compared with existing topological navigation and spatial-memory methods?

Response:

We acknowledge that structured spatial memory, topological connectivity, and relevance-guided exploration have been investigated in prior work. Rather than claiming these individual components as entirely new, the contribution of SPM-Nav lies in integrating them into a persistent, target-conditioned regional search-control framework. We clarify this contribution through three aspects.

(1) Persistent Space: Target-conditioned physical-region-level search states.

TopoNav constructs a topological graph based on robot-reached waypoints, associating each node with local spatial and semantic information to guide exploration and waypoint selection. EvoMemNav adopts a Room–View–Object hierarchy, in which Room nodes organize and index associated View and Object memories, while experience-driven updates refine navigation priors. In contrast, SPM-Nav uses the physical region itself as a persistent memory unit. Each Space aggregates observations across multiple positions and viewpoints and maintains target relevance, accumulated negative evidence, and consumed local-search steps. These states determine the remaining search budget and early-stopping conditions, allowing target-specific search progress to be preserved and reused across revisits.

(2) Inter-Space Portals: Explicit transition nodes and reusable connectivity.

Unlike conventional topological graphs that directly connect spatial nodes through edges, SPM-Nav represents each physical inter-region transition as an independent Portal node between two Space nodes, forming an explicit Space–Portal–Space structure. Each Portal maintains its own transition location, connected Spaces, navigable anchors on both sides, and traversal-confirmation state. Importantly, traversal confirmation determines when the active Space switches, while confirmed Portal nodes can be reused for subsequent inter-region Transit and navigation recovery. Thus, Portals explicitly represent and manage physical transitions rather than merely encoding spatial adjacency.

(3) Adaptive Space Search: Memory-conditioned search control beyond relevance-based budget allocation.

Adaptive Space Search goes beyond allocating budgets according to current target relevance. It jointly considers persistent relevance, previously consumed search effort, and accumulated negative evidence. A hysteretic priority mapping determines local-search budgets, while historical consumption and negative evidence govern remaining effort and early stopping. Crucially, terminating local inspection does not terminate inter-region exploration: the agent can still follow confirmed Portals toward unexplored or relevant downstream Spaces. This separates two coordinated decisions: whether to continue searching within a region and whether to traverse it to reach other regions.

Relationship to ASCENT. SPM-Nav builds upon the open-source ASCENT navigation pipeline and retains its pretrained perception and low-level navigation components, but replaces its LLM-driven coarse-to-fine exploration with the proposed Space–Portal memory and Adaptive Space Search. A component-level comparison of the original ASCENT and full SPM-Nav is provided in Table 6 (Appendix B).

Taken together, the contribution of SPM-Nav is not merely the combination of spatial memory, topological connectivity, and heuristic budget allocation. Rather, it establishes a persistent regional search-control mechanism in which target-conditioned search experience is maintained across revisits and directly governs both local inspection and inter-region navigation. In this framework, memory is not merely used to retrieve spatial knowledge, but serves as an explicit state representation for subsequent search and navigation decisions.

We have revised Related Work to clarify the differences between SPM-Nav and TopoNav, EvoMemNav, and ASCENT, and further specified Space identity maintenance, Portal construction and validation, and memory-conditioned search control in Sections 3.2 and 3.3 and Appendix A.

AuthorsOpenReview ↗

Response to Reviewer BECa (R1-2)

Question:

Several key decisions, including historical-evidence scoring, inspection-mode selection, and Space identity maintenance, also need clearer formal definitions before the method can be reproduced. These components are currently described at a level of detail that leaves their exact behaviour open to interpretation and makes it hard to implement them faithfully, so stating them more precisely would considerably strengthen the reproducibility of the reported results. Can the authors provide complete definitions of the key decision rules?

Response:

We agree that these key decisions require explicit computational rules and state-transition conditions for reproducibility. Accordingly, we have further formalized four aspects in the revised manuscript and Appendix A:

(1) Historical Evidence Scoring. We specify historical observation counting, the scoring formula, and observation deduplication, distinguishing candidate-level historical evidence from Space-level negative evidence (Section 3.3.1 and Appendix A).

(2) Inspection-mode Selection. We define the threshold-based selection of Peek, Probe, and Search, including the Peek-first procedure, post-Peek relevance reassessment, and Probe execution and fallback conditions (Section 3.3.1 and Appendix A).

(3) Space Identity Maintenance. We clarify provisional Space initialization, association, stabilization, and merge-back conditions, including the spatial and semantic criteria for identity maintenance and the traversal-confirmation requirement for Active Space switching (Section 3.2.1 and Appendix A).

(4) Space Priority and Search Control. We formalize hysteretic priority updates, priority-dependent search budgets, consumed-step accounting, negative-evidence validity, and early-stopping conditions, while distinguishing local-search budgets from Transit budgets (Section 3.3.2 and Appendix A).

These revisions provide explicit decision rules and state-transition definitions in the main text, with detailed thresholds and implementation specifications in Appendix A. We are also organizing and documenting the project code, which we plan to release on GitHub upon completion to facilitate reproduction.

AuthorsOpenReview ↗

Response to Reviewer BECa (R1-3)

Question:

Can additional experiments better isolate the contributions of Persistent Memory and Adaptive Search?

Response:

To further assess the contributions of Persistent Memory and Adaptive Space Search, we have clarified the corresponding ablation configurations, added paired statistical tests for the HM3D-v2 results, and extended the component ablations to MP3D. We also report mechanism-level evidence from the 50-episode HM3D-v1 diagnostic subset to examine how these components affect search-state reuse and local search-resource allocation.

(1) Statistical significance of component ablations.

  • On HM3D-v2, we perform 10,000 paired bootstrap resamples for SPL and average executed steps, and two-sided exact McNemar tests for SR (as shown in Table 7).
  • Persistent Memory: Removing this component increases average executed steps by 13.40 (95% CI: [9.44, 17.48]), with a nominally significant SR difference (p = 0.0464). Its SPL difference is not statistically significant.
  • Adaptive Space Search: Removal increases executed steps by 8.00 (95% CI: [0.82, 14.99]), while the SPL difference is not significant (95% CI: [-2.49, 0.07]; SR p = 0.0773). We therefore avoid claiming statistically significant SPL improvements for Persistent Memory and Adaptive Search.

Table 7. Paired statistical analysis of core-component ablations on HM3D-v2.

VariantΔSPL [95% CI]ΔSteps [95% CI](n₁₀, n₀₁)McNemar p
w/o Persistent Memory−0.40 [−1.11, +0.29]+13.40 [+9.44, +17.48](41, 24)0.0464
w/o Portals−2.00 [−2.79, −1.23]+16.02 [+11.29, +20.84](70, 23)1.10 × 10⁻⁶
w/o Adaptive Search−1.20 [−2.49, +0.07]+8.00 [+0.82, +14.99](68, 48)0.0773

(2) Ablation configurations and cross-dataset validation.

We further extend all three component ablations to MP3D (Table 2). Specifically, Removing Persistent Memory and Adaptive Search decreases SR by 2.1 and 3.3 percentage points, and SPL by 0.5 and 1.1 percentage points, respectively, while increasing average executed steps by 20.97 and 13.73. All three components exhibit consistent performance trends across MP3D and HM3D-v2, with Portal removal causing the largest degradation.

Table 2. Comparison with a matched LLM-free frontier baseline and core-component ablations on HM3D-v2 and MP3D.

VariantPSMPortalASSHM3D-v2 SR ↑HM3D-v2 SPL ↑HM3D-v2 Avg. Steps ↓MP3D SR ↑MP3D SPL ↑MP3D Avg. Steps ↓
Matched LLM-free Frontier Baseline✗✗✗70.730.2263.3938.314.8351.47
w/o Persistent Memory✗✓✓82.137.2231.5753.624.2322.83
w/o Portals✓✗✓79.135.6234.1949.722.5330.16
w/o Adaptive Search✓✓✗81.836.4226.1752.423.6315.59
Full SPM-Nav✓✓✓83.837.6218.1755.724.7301.86

(3) Mechanism-level evidence.

The 50-episode HM3D-v1 subset is used to examine actual mechanism behavior rather than serve as the sole evidence for overall performance (as shown in Section 4.3). The analyses quantify the effects of individual mechanisms: Persistent Memory retains historical search states across all 65 Space revisits; and disabling Connectivity-Preserving Transit reduces downstream-Space reachability from 73.42% to 2.13% (Appendices C and D).

Overall, these additional experiments strengthen the statistical and cross-benchmark evidence while allowing us to characterize component contributions more precisely: Portals provide the clearest contribution to overall navigation performance, whereas Persistent Memory and Adaptive Space Search primarily reduce redundant inspection and improve local search-resource allocation.

We hope that these revisions clarify the algorithmic contributions of SPM-Nav, improve the reproducibility of the proposed method, and provide more comprehensive experimental evidence for the roles of its individual components. We appreciate the reviewer's constructive feedback, which has helped us substantially strengthen the manuscript.

If the clarifications and additional evidence address your concerns, we hope you will consider updating your overall assessment of the paper accordingly.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.