acceptodds
Under review as a conference paper at ICLR 2027

EGNav: Evidence-Guided Memory Repair for Zero-Shot Goal-Oriented Navigation

Abstract

Zero-shot goal-oriented navigation requires more than exploring new regions: a previously visited location may still contain a target that was misidentified or remained outside the agent's view. This ambiguity makes it difficult to determine whether to continue exploring or revisit a potential target location. We propose EGNav, a zero-shot navigation framework that treats goal search as sequential evidence management. At its core is a task-conditioned evidence memory that associates spatially grounded candidates with direct, contextual, and negative evidence. Source- and visibility-aware updates maintain evolving belief and uncertainty, distinguishing contextual support from direct target matches and reducing the influence of negative observations obtained under limited visibility. A unified memory-repair policy then ranks unexplored frontiers and memory candidates using accumulated belief, expected information gain, and execution costs. Within this policy, direct repair re-examines previously observed target-like instances, while contextual repair revisits locations that may contain an unseen target. Both modes update the same evidence state and participate in a shared decision process. Experiments across object-, image-, and text-goal navigation on HM3D and MP3D show that EGNav outperforms the compared universal zero-shot methods, with further gains over the compared training-free baselines on GOAT-Bench. Ablations support the complementarity of the two repair modes and the contributions of evidence updating and candidate ranking.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
zero-shot navigation, memory

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_av5rOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EGNav addresses the ambiguity of failed observations in zero-shot goal-oriented navigation by maintaining a task-conditioned memory of direct, contextual, and negative evidence. It uses visibility-aware belief updates and a unified policy to balance memory revisitation against frontier exploration. Experiments demonstrate improvements over several zero-shot baselines, with ablations supporting the complementarity of direct and contextual repair.

Critical strengths
  1. The paper identifies a meaningful failure mode in zero-shot navigation.

  2. The framework is reasonable.

  3. Experiments across multiple navigation tasks and benchmarks show consistent improvements.

Critical weaknesses
  1. The methodological novelty appears incremental. Revisiting previously explored locations is not a new concept in embodied navigation, as prior methods such as 3D-Mem already leverage stored scene memories to guide subsequent exploration. Although EGNav explicitly frames revisitation as evidence repair and introduces visibility-aware evidence updates, the paper does not sufficiently establish that this formulation constitutes a substantial advance over existing memory-guided exploration and verification mechanisms.

  2. The ablations do not fully establish the value of the proposed evidence-management mechanism. EGNav lacks a direct comparison with a simpler revisitation baseline that retains memory-guided revisits but removes accumulated evidence modeling and information-gain-based ranking. Such a comparison is necessary to determine whether the gains stem from the proposed evidence-management mechanism or primarily from enabling revisitation.

Questions for authors

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

4 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer av5r

W1. Methodological Novelty Beyond Memory-guided Revisitation

We thank the reviewer for this insightful comment! We would like to clarify that EGNav differs from 3D-Mem in how stored spatial memories are represented, updated, and utilized for goal-oriented navigation. EGNav builds upon the memory-guided exploration paradigm by introducing an explicit evidence-management framework to address the ambiguity of failed observations under partial observability. Specifically, EGNav introduces three complementary designs:

1.Task-conditioned evidence memory: We maintain direct, contextual, and negative evidence associated with stored spatial candidates, enabling the agent to distinguish between observed targets, contextually supported hypotheses, and insufficiently verified locations.

2.Visibility-aware belief updates: We accumulate positive and negative evidence according to observation reliability, preventing missed detections under limited visibility from prematurely excluding potential targets.

3.Unified memory-repair policy: We jointly consider accumulated belief, expected information gain, and navigation costs to determine whether to revisit uncertain candidates or explore new frontiers.

Together, these designs enable EGNav to more effectively utilize stored 3D memories by reasoning about whether a candidate has been sufficiently verified, rather than simply whether its location has been explored. The key contribution therefore lies in managing and repairing uncertain goal evidence through a unified navigation policy. Furthermore, our existing comparison with 3D-Mem on GOAT-Bench Val-Unseen (Table 2) shows that EGNav improves SR/SPL from 28.8/15.8 to 39.1/23.7, providing additional empirical evidence of the effectiveness of our framework. We also introduce two additional controlled experiments in the following response to further demonstrate that the improvements arise from our evidence-management mechanisms rather than memory revisitation alone.

Q2. Contribution of Evidence Management Beyond Revisitation

We thank the reviewer for this valuable suggestion! To directly distinguish the benefits of simple memory revisitation from our proposed evidence-management mechanisms, we introduce an additional baseline, Memory + Nearest-first Revisit. This baseline just revisits the nearest eligible memory candidate without accumulated evidence modeling or information-gain-based ranking. Unsuccessfully revisited candidates are temporarily excluded until new relevant observations become available, and frontier exploration resumes when no eligible candidates remain. We compare this new baseline with EGNav and the existing component ablations from our original manuscript, including variants without visibility weighting or information gain.

MethodSR ↑SPL ↑
Base navigator (w/o revisit)58.822.5
Memory + Nearest-first Revisit (new)60.223.1
EGNav w/o visibility weighting61.523.8
EGNav w/o information gain61.123.2
EGNav (Full)62.625.1

As shown above, simple memory revisitation improves SR/SPL from 58.8/22.5 to 60.4/23.1, confirming the benefit of revisiting stored candidates. EGNav further improves SR by 2.4 points and SPL by 2.0 points over this baseline, demonstrating that its improvements cannot be attributed solely to memory revisitation. Furthermore, our existing ablation results provide additional evidence for the proposed mechanisms. Removing visibility weighting reduces SR/SPL to 61.5/23.8, while removing information gain reduces them to 61.1/23.2, despite retaining memory revisitation. Together, these results demonstrate that EGNav benefits not only from revisiting stored candidates, but also from visibility-aware evidence updating and information-gain-based candidate selection.

We sincerely thank the reviewer for their valuable suggestions. We have incorporated the new results into the revised manuscript and would be happy to address any further questions and continue the discussion!

AuthorsOpenReview ↗

Follow-up Regarding Our Rebuttal for Submission #24757

Dear Reviewer av5r,

We hope you are doing well. Thank you again for your thoughtful comments and helpful suggestions.

We have posted a detailed response to your concerns. In particular, we clarified the methodological contributions beyond memory-guided revisitation and added a controlled comparison with a nearest-first revisitation baseline to isolate the benefits of evidence management. The additional results have been incorporated into the revised manuscript.

We would welcome your thoughts on whether these explanations and experiments adequately address your concerns. We remain happy to answer any further questions during the author–reviewer discussion period !

Best regards,
The Authors of Submission #24757

Reviewer_av5rOpenReview ↗

Response by Reviewer av5r

The rebuttal has clarified the intended contribution of EGNav as an integrated evidence-management framework for zero-shot goal-oriented navigation. The additional experiments convincingly demonstrate that this framework provides benefits beyond simple memory-guided revisitation. However, the contribution appears to lie primarily in the task-specific integration of existing concepts, including spatial memory, visibility-aware evidence accumulation, belief estimation, and information-gain-based candidate selection. This integration is useful and addresses a meaningful practical problem. But I still think the methodological novelty beyond the proposed system-level framework remains relatively modest.

However, I appreciate the effort and will raise my overall rating.

AuthorsOpenReview ↗

Thank You for the Updated Assessment

Dear Reviewer av5r,

Thank you very much for your thoughtful reassessment and for raising your overall rating ! We sincerely appreciate the time and care you have devoted to reviewing our work and considering the additional experiments and clarifications. Your constructive feedback has been valuable in helping us strengthen the paper and better articulate its contributions.

Best regards,
The Authors of Submission #24757

Reviewer_NNWEOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper focus on the sicuation when the goal target is misidentified or remained outside the agent's view during a navigation task. The proposed EGNav's core is a task-conditioned evidence memory that associates spatially grounded candidates with direct, contextual, and negative evidence.

Critical strengths

The research topic is interesting and important. In an obeject-goal navigation task, especially the goal is specified by certain qualifiers, it's common that the goal is missed or incorrectly selected. I believe that conducting research in this area is meaningful, and could enhance the performance.

Critical weaknesses

The idea of this paper is interesting, and the experiments are fairly comprehensive. However, I think the performance gain of the proposed method could be better explained. EGNav introduces several design choices for deciding when and where to revisit previous locations, including visibility-aware evidence updates and uncertainty-based candidate selection. Though the results suggest that these designs are helpful (in overall performance enhancement), it is still unclear whether the gains come from the proposed formulations themselves or simply from maintaining and revisiting candidate memories. For example, comparing the method against simpler revisit strategies, or testing different ways of updating candidate confidence under partial visibility, would help clarify the contribution of each design choice.

Questions for authors

See weakness.

3 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer NNWE

W1. Contribution of Evidence Updating and Memory Repair

We thank the reviewer for this valuable suggestion!To distinguish the benefits of simple memory revisitation from our proposed evidence updating and candidate selection mechanisms, we introduce two additional experiments:

1.Memory + Nearest-first Revisit: This baseline retains the same perception components, spatial candidate memory, navigation backend, and goal verification mechanism as EGNav, but revisits the nearest eligible memory candidate without visibility-aware belief updates or information-gain-based ranking. Unsuccessful candidates are temporarily excluded until new relevant observations become available, and frontier exploration resumes when no eligible candidates remain.

2.Binary Visibility Gating: We replace the continuous visibility weighting in our confidence updates with a binary threshold-based mechanism, where evidence is incorporated only when the visibility score exceeds a predefined threshold, while keeping the remaining components unchanged.

MethodSR ↑SPL ↑
Base navigator (w/o revisit)58.822.5
Memory + Nearest-first Revisit60.223.1
EGNav w/o visibility weighting61.523.8
EGNav w/ Binary Visibility Gating61.924.3
EGNav w/o information gain61.123.2
EGNav (Full)62.625.1

As shown above, simple memory revisitation improves SR/SPL from 58.8/22.5 to 60.2/23.1, confirming the value of revisiting stored candidates. EGNav further improves SR by 2.4 points and SPL by 2.0 points over this baseline. Regarding evidence updating, binary visibility gating achieves 61.9/24.3, outperforming the variant without visibility weighting (61.5/23.8), while our continuous visibility weighting further improves performance to 62.6/25.1. This suggests that fine-grained visibility modeling is more effective than simple threshold-based updates under partial observability. Moreover, removing information gain reduces SR/SPL to 61.1/23.2, despite retaining memory repair. These results demonstrate that the improvements arise not only from memory revisitation itself, but also from the proposed visibility-aware evidence updating and information-gain-based candidate selection.

We sincerely thank the reviewer for recognizing the value of our work. We have incorporated all new experimental results into the revised manuscript, with the corresponding updates highlighted in blue. We welcome any further questions or concerns and would be happy to continue the discussion and address them!

Thank the author for the detailed reply. Please include the additional experiments in the fianl version of the paper.

AuthorsOpenReview ↗

Thanks for your review

Dear Reviewer NNWE,

Thank you very much for your thoughtful reassessment and encouraging feedback! We sincerely appreciate the time and effort you have devoted to carefully reviewing our work and considering our additional experiments and clarifications. Your constructive feedback has been invaluable in helping us improve the rigor and presentation of our work. We also appreciate your remaining concerns and will continue to address them in the revised manuscript.

Best regards, The Authors of Submission #24757

Reviewer_SGJjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of goal ambiguity in object search. The key idea is to maintain a task-conditioned evidence memory that accumulates observations relevant to disambiguating the target. Building on this memory, the planner jointly considers belief, information gain, and cost when selecting candidates. The method is evaluated on three benchmarks, HM3D, MP3D, and GOAT-Bench, and demonstrates superior performance.

Critical strengths
  1. The objective is clear: the paper tackles the goal ambiguity issue in existing object navigation tasks. The memory design is well explained and intuitively aligned with this goal.

  2. The agentic system leverages recent advances in vision-language foundation models to address practical robotics problems, and its modular design provides interpretability into the decision-making process.

  3. The overall pipeline is clearly presented and easy to follow. The writing, figures, and ablations clearly support the claims, and the comparisons with state-of-the-art methods together with the sensitivity analysis help demonstrate the effectiveness of the method.

Critical weaknesses
  1. The contribution to disambiguation is not clearly established. Since the overall system is complex and its components are tightly coupled, it is difficult to attribute improvements specifically to better goal identification. Directly measuring target identification precision and recall would provide more convincing evidence.

  2. Comparisons with recent VLM-based approaches to goal identification would further strengthen the case for the proposed method.

  3. The paper lacks comparison with more recent works that outperform EGNav, as summarized below (SR / SPL):

MethodHM3DMP3D
EGNav (this paper)60.0 / 32.943.1 / 18.2
LongNav-R1 [1]76.0 / 44.363.0 / 30.2
LightNav-0 [2]74.5 / 43.953.3 / 21.2
Uni-NaVid [3]73.7 / 37.1–

[1] LongNav-R1: Horizon-Adaptive Multi-Turn RL for Long-Horizon VLA Navigation, RSS 2026.

[2] LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation, arXiv 2026.

[3] Uni-NaVid: A Video-based Vision-Language-Action Model for Unifying Embodied Navigation Tasks, RSS 2025.

  1. Real-world experiments would strengthen the paper. It would also be interesting to see how the method performs across environments ranging from small to large spaces.
Questions for authors
  1. What is the motivation behind the information gain (IG) design in Equation 5? It is somewhat difficult to follow, and a high-level intuitive explanation of each term in the formula would be very helpful.

  2. What are the computational cost and inference time of the system? This would help readers understand its overall efficiency. Please also clarify the sensor requirements.

5 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer SGJj

W1. Target disambiguation and identification accuracy

Thanks! To directly evaluate target disambiguation and identification accuracy, we reproduced UniGoal and analyzed the target identification results across 1,000 episodes.

MethodPrecision ↑Recall ↑ITIR↓
UniGoal (reproduced)70.2%58.0%24.6%
EGNav (Ours)78.7%+8.5_{+8.5}62.6%+4.6_{+4.6}16.9%−7.7_{-7.7}

Incorrect Target Identification Rate (ITIR): The proportion of instances in which the agent incorrectly identifies a non-target object as the intended target. Lower is better. EGNav improves precision by 8.5% and recall by 4.6 %, while reducing the incorrect target identification rate by 7.7%. These improvements demonstrate the effectiveness of our complementary designs for evidence accumulation, visibility-aware updating, and memory repair.

W2 & W3. Comparison with recent VLM-based methods

Good question! LongNav-R1 introduces horizon-adaptive multi-turn reinforcement learning to improve long-horizon VLA navigation. LightNav-0 develops a generalist navigation model through VLM-based spatial reasoning and navigation-specific training. Uni-NaVid unifies diverse embodied navigation tasks within a video-based VLA framework trained on large-scale navigation data.

MethodTraining-freeZero-shotExtra DatasetHM3DMP3D
LongNav-R1✗✗✓76.0 / 44.363.0 / 30.2
LightNav-0✗✗✓74.5 / 43.953.3 / 21.2
Uni-NaVid✗✗✓73.7 / 37.1–
ApexNav✓✓✗59.6 / 33.039.2 / 17.8
EGNav (Ours)✓✓✗60.0 / 32.943.1 / 18.2

As shown, these methods achieve strong performance through navigation-specific training and additional datasets, whereas EGNav operates in a training-free, zero-shot setting. Compared with the similarly setting method ApexNav, EGNav achieves improvements on both HM3D and MP3D.

W4. Real-world demonstrations and environment scale

Thanks! We have already included real-world navigation demonstrations in the supplementary videos of our original submission, and we kindly invite the reviewer to refer to these materials. We also compare EGNav with UniGoal across small, medium, and large environments.

ScaleUniGoal SR (%)↑EGNav SR (%)↑
Small63.3868.73$_{+4.35}$
Medium61.6972.24$_{+10.55}$
Large53.0055.65$_{+2.65}$

EGNav consistently improves SR across all three environment scales, particularly in medium-sized, where our method can more effectively explore and utilize accumulated evidence. In large environments, some episodes terminate due to the predefined time limit, this motivates us to investigate more efficient exploration strategies in future work.

Q1. Motivation and Interpretation of Information Gain

Thanks! Eq. 5 estimates the expected reduction in uncertainty about a candidate target location after visiting it, rather than relying solely on its current target likelihood. Specifically, the first term represents the current belief uncertainty (Beta variance in Eq. 3), while the remaining terms estimate the expected uncertainty under three observation outcomes: positive, negative, and uninformative evidence. These outcomes are weighted by target relevance and visibility, with updates accounting for observation reliability. Together with navigation costs in Eq. 6, this allows EGNav to balance exploring new locations and revisiting uncertain candidates.

Q2. Questions about cost

Good question! Computational Cost: The average inference time is approximately 80 s per episode with 27.1 model calls, compared to approximately 100 s per episode for UniGoal. Inference Time: Goal parsing takes 1.11 s, contextual evidence scoring 1.10 s, and visual target matching 3.18 s. IIN goal-image attributes are extracted once (7.07 s) and cached. Sensor Requirements: RGB-D observations and relative pose information.

We have added all new results and revisions to the manuscript, highlighted in blue. We would be happy to address any further questions or concerns from the reviewer.

AuthorsOpenReview ↗

Follow-up Regarding Our Rebuttal for Submission #24757

Dear Reviewer SGJj,

We hope this message finds you well. Thank you again for your time and constructive feedback on our work.

We have posted a detailed response addressing your concerns. Specifically, we provided direct measurements of target-identification precision and recall, comparisons with recent VLM-based approaches, a detailed explanation of the information-gain formula, the system’s computational cost and inference time, and clarification of the real-world experiments.

We would greatly appreciate hearing whether these clarifications and additional results have addressed your concerns. Should any questions remain, we would be pleased to provide further information during the author–reviewer discussion period !

Best regards,
The Authors of Submission #24757

AuthorsOpenReview ↗

Follow-up Before the Discussion Deadline

Dear Reviewer SGJj,

With fewer than 12 hours remaining in the discussion period, we would like to kindly follow up on our rebuttal. We would be grateful if you could share your assessment of whether our responses and additional results have addressed your concerns. If any questions remain, we would welcome the opportunity to clarify them before the discussion ends.

Thank you again for your time and careful consideration of our work.

Best regards,
The Authors of Submission #24757

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.