acceptodds
Under review as a conference paper at ICLR 2027

EviNav: Evidence-Grounded Search and Active Verification for Training-Free Aerial Object Navigation

Abstract

Large-scale Aerial Object Navigation (Aerial ObjectNav) requires a UAV to locate semantically relevant regions and determine whether an observed candidate is the requested target instance under partial and viewpoint-dependent observations. Existing aerial ObjectNav methods primarily strengthen spatial memory, exploration, and target grounding, while candidate-level evidence is not typically maintained and reused as a unified persistent state across search, verification, and stopping. We present EviNav, a training-free framework that treats candidate evidence as a persistent navigation state connecting evidence-grounded search, persistent target hypotheses, and active verification. EviNav constructs a shared evidence specification that distinguishes search-supporting cues from acceptance-critical evidence, accumulates candidate evidence across observations, and actively acquires complementary views when the available evidence is insufficient. Unresolved evidence guides further observation, contradictory evidence supports candidate rejection, and STOP is authorized when the available evidence provides sufficient support for the target identity under the applicable acceptance conditions. On the complete 1,000-episode, 14-scene UAV-ON test protocol, EviNav achieves a Success Rate (SR) of 27.00%, an Oracle Success Rate (OSR) of 36.60%, and a Success weighted by Path Length (SPL) of 22.98%, improving the previous best full-benchmark SR by 7.50 percentage points without task-specific navigation training. These results indicate that persistent evidence and active verification improve target-acceptance reliability while maintaining effective search and path efficiency in large-scale aerial ObjectNav.

Primary area:
applications to robotics, autonomy, planning
Keywords:
aerial object navigation, embodied AI, UAV navigation, active verification, persistent memory, evidence-grounded navigation, vision-language models, training-free navigation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_oam8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EviNav is a training-free framework for aerial object navigation that combines evidence-grounded search, persistent candidate hypotheses, and active verification. It maintains candidate evidence across observations and uses unresolved requirements to guide further inspection and inform stopping decisions. On the UAV-ON benchmark, it reports 27.00% SR, 36.60% OSR, and 22.98% SPL. Its main contribution is a system design that connects search, candidate verification, and termination through persistent evidence.

Critical strengths
  1. The distinction between discovering a plausible candidate and establishing its identity is well motivated by the viewpoint-dependent nature of aerial observations. Keeping unresolved evidence separate from explicit contradictions is a sensible design choice.

  2. The reported full-benchmark improvements are substantial, and the primary evaluation retains all 14 scenes despite the identified environment issues. EviNav’s aggregate results are consistent with its reported breakdowns across target scales and seen/unseen environments.

Critical weaknesses
  1. The experiments do not directly validate the failure modes motivating the method. Section 4.4 and Table 5 analyze only successful episodes, without measuring incorrect acceptance, rejection of correct candidates, or repeated visits. The SR/OSR ratio also does not measure verification accuracy. Thus, evidence for improved target-acceptance reliability remains limited.

  2. The ablations do not isolate the evidence mechanism. Removing active verification also removes candidate approach and additional viewpoints, while removing persistent evidence also removes interaction history. A generic-inspection baseline with a comparable budget is needed. The exact acceptance and STOP rules, especially for conflicting or incomplete evidence, also require clarification.

  3. Candidate association is not validated. The matching rules may combine evidence from nearby or visually similar objects, potentially leading to incorrect acceptance. Tests with nearby same-category distractors should evaluate association accuracy and its effect on final decisions.

  4. Figure 2 needs larger text, clearer notation, and connections that match the implementation. The authors should distinguish illustrative elements from recorded observations and clarify any generative AI assistance in figure preparation. These presentation issues alone do not establish undisclosed AI use.

Questions for authors
  1. Can you provide evidence that evidence-gap-driven inspection outperforms a generic inspection policy with the same capabilities and a comparable budget? This would help distinguish the proposed reasoning mechanism from the benefits of additional observations.

  2. Across all evaluated episodes, does the proposed method reduce incorrect acceptances, rejections of correct candidates, and repeated visits compared with the relevant ablated variants? Please distinguish per-decision statistics from final candidate outcomes.

  3. The association procedure in Appendix B.2 uses spatial proximity and appearance similarity. What evidence shows that nearby, visually similar instances are reliably kept separate? How often do association errors affect accumulated evidence or final acceptance decisions?

6 comments
AuthorsOpenReview ↗

Q1. Comparison with Geometry-Driven Exploration

Response to Reviewer oam8

We thank the reviewer for recognizing EviNav's motivation, its distinction between unresolved and contradictory evidence, and the substantial improvements on the complete UAV-ON benchmark. We provide additional controlled comparisons, comprehensive termination analysis, and qualitative and quantitative evidence on persistent candidate verification.

Q1. Comparison with Geometry-Driven Exploration

To further examine the role of evidence-guided observation, we conducted an additional comparison with a geometry-driven exploration (GE) baseline on the same 200 UAV-ON episodes used in Table 3.

In GE, candidate approach and additional viewpoint acquisition are guided primarily by geometric considerations, such as target distance and viewpoint configuration, rather than by unresolved semantic evidence requirements. The remaining candidate verification and stopping mechanisms are retained.

Both methods use the same perception and navigation components, with identical action and observation budgets. Their difference lies in how additional observations are selected.

MetricFull EviNavGE
SR ↑27.50%25.00%
OSR ↑43.00%44.50%
SPL ↑0.24700.2207
Collision Rate ↓20.50%27.00%
Failed-STOP Rate ↓21.43%18.03%
Avg. Additional Geometric Views3.2051.825
Avg. Sub-actions per High-level Step ↓2.543.00

The average number of sub-actions per high-level step is computed over episodes successfully completed by both methods.

Both methods achieve comparable target reachability, with OSR values of 43.00% and 44.50%, respectively. Full EviNav achieves a modestly higher SR (27.50% vs. 25.00%) and better SPL (0.2470 vs. 0.2207), while reducing the collision rate by 6.50 percentage points. EviNav also acquires more complementary geometric views (3.205 vs. 1.825) and uses fewer sub-actions per high-level step on jointly successful episodes (2.54 vs. 3.00).

These results suggest that incorporating unresolved evidence requirements into viewpoint selection can provide additional benefits beyond geometry-driven observation alone. More broadly, EviNav is designed as an integrated navigation framework in which evidence-guided search, persistent hypothesis maintenance, active verification, and stopping decisions work together. This complementary comparison further illustrates the role of evidence-guided observation within the overall system.

AuthorsOpenReview ↗

Q2. Incorrect acceptance, rejection, and repeated verification

Q2. Incorrect acceptance, rejection, and repeated verification

We extend the original success-only analysis to the complete 1,000-episode evaluation and examine target-acceptance outcomes across the Table 3 ablation variants.

Across all 1,000 episodes, EviNav achieves 270 successful STOPs, while the remaining 730 episodes terminate through 87 unsuccessful STOPs, 255 collisions, and 388 budget-exhaustion events. This provides a complete accounting of navigation outcomes beyond the successful trajectories analyzed in the original Table 5.

Target-acceptance outcomes across ablations. We further examine terminal STOP decisions on the same fixed 200 episodes used in Table 3, distinguishing successful acceptance, premature stopping at the correct target, and acceptance of an incorrect target instance.

Table Y. Target-acceptance outcomes across Table 3 ablation variants.

VariantSuccessful STOPsPremature STOPs on Correct TargetsWrong-Instance AcceptancesTotal Unsuccessful STOPs
Full EviNav55/70 (78.57%)9/70 (12.86%)6/70 (8.57%)15/70 (21.43%)
w/o Evidence-Grounded Search40/54 (74.07%)8/54 (14.81%)6/54 (11.11%)14/54 (25.93%)
w/o Persistent Evidence49/65 (75.38%)10/65 (15.38%)6/65 (9.23%)16/65 (24.62%)
w/o Active Verification22/33 (66.67%)6/33 (18.18%)5/33 (15.15%)11/33 (33.33%)

All percentages use the total number of STOP-terminated episodes in each variant as the denominator. Premature STOPs denote correct-target acceptances outside the benchmark's 20 m success radius. Wrong-instance acceptances were identified through candidate-identity checks of unsuccessful STOPs.

Full EviNav achieves the highest successful-STOP fraction and the lowest wrong-instance acceptance and premature-STOP fractions among all four variants. Compared with the active-verification ablation, successful STOPs increase from 22 to 55, while wrong-instance acceptance decreases from 15.15% to 8.57% (−6.58 percentage points) and premature STOP decreases from 18.18% to 12.86% (−5.32 percentage points). The overall unsuccessful-STOP fraction decreases from 33.33% to 21.43% (−11.90 percentage points).

These results provide direct quantitative evidence for the improved target-acceptance outcomes of EviNav's integrated active-verification process. The framework combines continued candidate inspection, persistent evidence, and explicit acceptance checks to support reliable terminal decisions.

Decision-level and candidate-history analysis. We additionally examine intermediate verification behavior across the complete evaluation. The archived logs contain 1,631 rejection events associated with 1,600 hypothesis IDs, allowing individual controller decisions to be distinguished from persistent hypothesis histories.

We further analyze 85 unsuccessful episodes that reached the target neighborhood without collision but did not issue STOP. Among these episodes, 48 contain subsequent rejection records, 67 contain deferral records, and 27 include deferred hypotheses without subsequent same-ID verification. In total, 56/85 episodes (65.88%) exhibit rejection and/or deferred-candidate follow-up patterns.

These complementary analyses characterize three levels of EviNav's verification behavior: individual inspection decisions, persistent candidate histories, and final terminal-acceptance outcomes. They illustrate how candidate rejection, deferral, and continued exploration interact with the final navigation result.

AuthorsOpenReview ↗

Q3. Ablation Study on the Stopping Mechanism. and Q4. Candidate association and instance-level evidence consistency

Q3. Ablation Study on the Stopping Mechanism.

To isolate the contribution of evidence-based stopping, we conducted an additional ablation on the same 200 UAV-ON episodes used in Table 3. Specifically, we replaced the evidence-based STOP authorization with a simplified stopping rule, while retaining the original search policy, candidate approach, and viewpoint acquisition mechanisms. The results are summarized below.

MetricFull EviNavw/o Evidence-based STOP
SR ↑27.50%23.50%
OSR ↑43.00%29.50%
SPL ↑0.24700.2161
Successful STOPs ↑5547
Failed STOPs ↓1581
Failed-STOP Rate ↓21.43%63.28%

The results show that the full EviNav achieves a higher success rate (27.50% vs. 23.50%), while OSR increases from 29.50% to 43.00% and SPL improves from 0.2161 to 0.2470. More importantly, removing the evidence-based stopping mechanism leads to 81 unsuccessful stops out of 128 STOP decisions, corresponding to a failed-STOP rate of 63.28%. In contrast, the full EviNav produces only 15 unsuccessful stops out of 70 STOP decisions, reducing the failed-STOP rate to 21.43%, a decrease of 41.85 percentage points.

Furthermore, removing the evidence-based stopping mechanism leads to substantially more frequent STOP decisions (128 vs. 70), accompanied by a sharp increase in unsuccessful stops (81 vs. 15). Premature termination at incorrect candidates prevents the UAV from continuing its exploration, potentially leaving unexplored regions containing the true target. This explains the substantial decline in OSR from 43.00% to 29.50%, indicating that unreliable stopping decisions not only compromise final target confirmation but also limit the agent's ability to reach the correct target. These findings highlight a key advantage of EviNav: evidence-driven exploration and stopping are tightly integrated rather than treated as independent processes. By maintaining and updating candidate evidence throughout navigation, EviNav can defer uncertain stopping decisions, continue exploration when evidence is insufficient, and terminate only when sufficient evidence has been accumulated. This unified mechanism improves both exploration coverage and final stopping reliability.

This controlled ablation further distinguishes the contribution of STOP authorization from candidate approach and viewpoint acquisition, demonstrating that reliable stopping decisions independently contribute to navigation success and exploration effectiveness.

Q4. Candidate association and instance-level evidence consistency

EviNav maintains candidate-specific hypotheses across observations, allowing nearby objects to be represented and verified independently. Its association mechanism combines spatial proximity, category compatibility, and appearance information to maintain persistent candidate histories. Target-specific semantic and structural evidence is subsequently used to determine whether each candidate satisfies the requested description.

To illustrate this capability, we provide a new qualitative example in Figure 5(b) using original RGB observations and corresponding controller logs from an archived diagnostic navigation task.

Discriminating visually related playground objects (CityStreet, task 39). The scene contains a nearby arm-and-seat apparatus and the requested circular platform. Both exhibit blue visual elements and share the same playground context. At step 10, EviNav rejects the arm-and-seat apparatus because its structure lacks the required circular platform. At step 11, it accepts the separately maintained target candidate, whose blue circular platform and yellow handrails/supports match the requested description.

This case demonstrates that shared visual appearance and spatial context do not override instance-specific structural evidence during target verification.

Persistent candidate evidence is extensively incorporated into final target verification. In the Table 3 Full EviNav evaluation, 64 of 69 terminal STOP records with available history fields (92.75%) include multi-view candidate observations. This complements the qualitative example in Figure 5(b), where EviNav maintains separate hypotheses for neighboring candidates, preserves their observation histories across viewpoints, and uses target-specific evidence to reject distractors and accept the requested target.

Together, these results highlight the strength of EviNav's instance-aware verification framework: the system maintains separate hypotheses for neighboring candidates, accumulates evidence across viewpoints, rejects visually plausible distractors, and uses target-specific evidence to support reliable final acceptance. By linking persistent candidate association with evidence-based verification, EviNav enables consistent and discriminative target reasoning throughout navigation.

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

EviNav presents a training-free aerial object-navigation framework that treats candidate-level evidence as a persistent state shared across exploration, hypothesis maintenance, active verification, and stopping. It separates cues useful for searching from evidence required to accept a target, and acquires additional views when verification is incomplete. On the full 1,000-episode UAV-ON benchmark, it reports 27.00% success rate and 22.98% SPL, with further module-level ablations.

Critical strengths
  1. The separation between finding a plausible object and verifying that it satisfies the instruction is a well-motivated problem formulation. Persistent candidate-level evidence provides a coherent mechanism for linking exploration and stopping decisions.
  2. The evaluation covers the full benchmark rather than a small selected subset, and the authors provide ablations of search, evidence persistence, and active verification. This makes the observed system-level improvement reasonably credible.
Critical weaknesses
  1. The active-verification ablation does not isolate verification itself. As noted in the description of Table 3, removing this component also disables candidate-specific approach and viewpoint acquisition. Therefore, the large SR/OSR reduction combines changed navigation behavior with changed acceptance behavior, making the claimed causal contribution of verification harder to determine.
  2. The analysis of final stopping decisions is incomplete. Table 5 categorizes the 270 successful acceptances, but does not quantify false acceptances and false rejections in failed episodes. These failure statistics are important because reliable target acceptance is one of the paper's main claimed advantages.
Questions for authors
  1. Would a more controlled ablation that changes only evidence-based STOP authorization, while retaining candidate approach and viewpoint acquisition, lead to similar gains?
  2. Among unsuccessful episodes, how many failures are due to accepting the wrong candidate versus rejecting or failing to revisit a correct candidate?
3 comments
AuthorsOpenReview ↗

Q1. Ablation Study on the Stopping Mechanism.

Response to Reviewer iYbG

We thank the reviewer for the careful assessment and constructive suggestions on stopping reliability and failure analysis. In response, we have added a controlled STOP-authorization ablation that retains candidate approach and viewpoint acquisition. We have also extended our analysis to include failure outcomes across the complete 1,000-episode evaluation and provided terminal-acceptance statistics across the Table 3 variants on the same fixed 200 episodes. These additions clarify the role of evidence-based stopping and strengthen the empirical evaluation of EviNav's unified navigation and verification framework.

Q1. Ablation Study on the Stopping Mechanism.

To isolate the contribution of evidence-based stopping, we conducted an additional ablation on the same 200 UAV-ON episodes used in Table 3. Specifically, we replaced the evidence-based STOP authorization with a simplified stopping rule, while retaining the original search policy, candidate approach, and viewpoint acquisition mechanisms. The results are summarized below.

MetricFull EviNavw/o Evidence-based STOP
SR ↑27.50%23.50%
OSR ↑43.00%29.50%
SPL ↑0.24700.2161
Successful STOPs ↑5547
Failed STOPs ↓1581
Failed-STOP Rate ↓21.43%63.28%

The results show that the full EviNav achieves a higher success rate (27.50% vs. 23.50%), while OSR increases from 29.50% to 43.00% and SPL improves from 0.2161 to 0.2470. More importantly, removing the evidence-based stopping mechanism leads to 81 unsuccessful stops out of 128 STOP decisions, corresponding to a failed-STOP rate of 63.28%. In contrast, the full EviNav produces only 15 unsuccessful stops out of 70 STOP decisions, reducing the failed-STOP rate to 21.43%, a decrease of 41.85 percentage points.

Furthermore, removing the evidence-based stopping mechanism leads to substantially more frequent STOP decisions (128 vs. 70), accompanied by a sharp increase in unsuccessful stops (81 vs. 15). Premature termination at incorrect candidates prevents the UAV from continuing its exploration, potentially leaving unexplored regions containing the true target. This explains the substantial decline in OSR from 43.00% to 29.50%, indicating that unreliable stopping decisions not only compromise final target confirmation but also limit the agent's ability to reach the correct target. These findings highlight a key advantage of EviNav: evidence-driven exploration and stopping are tightly integrated rather than treated as independent processes. By maintaining and updating candidate evidence throughout navigation, EviNav can defer uncertain stopping decisions, continue exploration when evidence is insufficient, and terminate only when sufficient evidence has been accumulated. This unified mechanism improves both exploration coverage and final stopping reliability.

This controlled ablation further distinguishes the contribution of STOP authorization from candidate approach and viewpoint acquisition, demonstrating that reliable stopping decisions independently contribute to navigation success and exploration effectiveness.

AuthorsOpenReview ↗

Q2. Failure-episode analysis and target-verification behavior

Q2. Failure-episode analysis and target-verification behavior

We extend the original success-only analysis to the complete 1,000-episode UAV-ON evaluation, providing a detailed characterization of failure outcomes and candidate-verification behavior. EviNav achieves 270 successful STOPs, while the remaining 730 episodes are categorized according to their termination outcomes and recorded target-neighborhood visits.

Table X. Failure outcomes and candidate-verification records on UAV-ON.

GroupOutcome or recorded eventEpisodesPercentage
A. All failed episodes (n=730)Unsuccessful STOP8711.92%
Target neighborhood reached; no STOP or collision8511.64%
Target neighborhood reached; collision without STOP405.48%
No recorded target-neighborhood entry; no STOP51870.96%
Total730100%
B. Post-arrival inspection analysis (n=85)Rejection recorded after entering the target neighborhood4856.47%
Deferral recorded after entering the target neighborhood6778.82%
Deferred candidate without subsequent same-ID verification2731.76%
Rejection and/or deferred candidate without same-ID follow-up5665.88%

The post-arrival analysis further reveals how candidate verification influences navigation outcomes. Among 85 collision-free episodes that reached the target neighborhood but did not issue STOP, 48 contain subsequent rejection records, 67 contain deferral records, and 27 include deferred hypotheses without later same-ID verification. In total, 56/85 episodes (65.88%) exhibit rejection and/or deferred-candidate follow-up patterns. These findings provide a detailed characterization of verification and continued-search behavior after the agent has reached a promising target region, complementing the terminal-acceptance analysis below.

Navigation failures can involve interacting factors, including observation quality, verification decisions, and subsequent exploration. Attributing an unsuccessful episode to a single verification event may therefore oversimplify the underlying process and yield cause-specific percentages with limited interpretability. To provide a more direct assessment of target-acceptance behavior, we complement the full-benchmark outcome analysis with terminal-acceptance statistics across the Table 3 ablations on the same fixed 200 episodes. Specifically, we distinguish wrong-instance acceptance from premature stopping at the correct target and report their counts and proportions among STOP-terminated episodes. These clearly defined outcomes provide a focused and interpretable comparison of terminal-acceptance reliability across variants.

Target-acceptance reliability across ablations. We further examined every STOP-terminated episode in the same fixed 200 tasks used by Table 3, distinguishing successful termination, wrong-instance acceptance, and premature STOP at the correct target.

Table Y. Target-acceptance outcomes across Table 3 ablation variants.

VariantSuccessful STOPsPremature STOPs on Correct TargetsWrong-Instance AcceptancesTotal Unsuccessful STOPs
Full EviNav55/70 (78.57%)9/70 (12.86%)6/70 (8.57%)15/70 (21.43%)
w/o Evidence-Grounded Search40/54 (74.07%)8/54 (14.81%)6/54 (11.11%)14/54 (25.93%)
w/o Persistent Evidence49/65 (75.38%)10/65 (15.38%)6/65 (9.23%)16/65 (24.62%)
w/o Active Verification22/33 (66.67%)6/33 (18.18%)5/33 (15.15%)11/33 (33.33%)

All percentages use the total number of STOP-terminated episodes in each variant as the denominator. Premature STOPs refer to correct-target acceptances outside the benchmark's 20 m success radius. Wrong-instance acceptances were identified through candidate-identity checks. The three outcome categories are mutually exclusive.

Full EviNav achieves the highest successful-STOP fraction and the lowest wrong-instance acceptance and premature-STOP fractions among all four variants. Compared with the active-verification ablation, EviNav increases successful STOPs from 22 to 55, improves the successful-STOP fraction from 66.67% to 78.57%, and reduces the wrong-instance acceptance fraction from 15.15% to 8.57% (−6.58 percentage points). Premature STOPs on correct targets decrease from 18.18% to 12.86% (−5.32 percentage points), while the overall unsuccessful-STOP fraction decreases from 33.33% to 21.43% (−11.90 percentage points).

Together, these results strengthen the empirical evaluation of EviNav's evidence-driven navigation framework. The full-benchmark analysis characterizes failure outcomes and candidate-verification behavior, while the ablation analysis demonstrates the benefits of EviNav's integrated active-verification process in improving target acceptance and successful termination.

Thank you for the clarifications. The rebuttal satisfactorily addresses the majority of the concerns raised in my initial review.

Reviewer_gN4eOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces EviNav, a training-free framework for goal-oriented aerial navigation that integrates evidence-based search, persistent target hypotheses, and active verification mechanisms. Its core concept lies in distinguishing between "evidence that helps locate potential targets" and "evidence sufficient to confirm a target's identity." The system maintains a history of candidate targets across multiple observations and, when existing evidence is insufficient, acquires additional viewpoints to perform further verification.

This evaluation covers the UAV-ON test set, which consists of 1,000 segments across 14 scenarios. EviNav achieves an SR of 27.00%, an OSR of 36.60%, and an SPL of 22.98%, with the SR metric outperforming the strongest known baseline by 7.50 percentage points. Furthermore, the paper presents component ablation studies, sensitivity analyses regarding the large language model (LLM) used for navigation, and qualitative examples.

Critical strengths
  1. A problem of practical significance, with a coherent contribution at the system level.

In the field of aerial navigation, it is crucial to distinguish between "detecting a plausible target" and "verifying the specific identity of that target," as factors such as distance, occlusion, and changes in viewing angle often render individual observations ambiguous. Integrating persistence assumptions, active observation, and verification results into a closed-loop navigation framework constitutes a valuable contribution. This approach is clearly articulated in Sections 3.2 through 3.5, and the implementation details provided in the supplementary material are highly thorough.

  1. Promising experimental results were achieved on a substantial benchmark dataset.

The evaluation utilized the full UAV-ON test set, rather than a smaller subset. Significant improvements were observed in both SR and SPL metrics, with these performance gains extending across various target sizes and both seen and unseen environments (see Tables 1–2). Table 3 further demonstrates the contribution of key components—most notably the "active verification" mechanism—to overall performance.

Critical weaknesses

Clarify how explicit "evidence gaps" influence the active verification process.

Section 3.4 defines "evidence states" and "unresolved evidence gaps" at the requirement level (Equations 3–4). Appendix B.2 further notes that these gaps dictate what observational information should be acquired next (see Table 8). However, the relationship between these explicit states and the actual "observation selection strategy" could be further elaborated. Specifically, it would be highly beneficial to explain how a "specific unresolved requirement" drives the system to take actions—such as "moving closer to a candidate target," "changing the viewpoint," or "deferring verification"—rather than relying solely on suggestions from a Vision-Language Model (VLM) or geometry-based heuristics. Providing a concrete example that tracks the evolution of evidence states and the corresponding subsequent actions would further strengthen this part of the paper.

Questions for authors
  1. Clarify the budget constraints regarding navigation steps and action execution.

Section 4.1 and Appendix A.1 specify that each episode consists of at most 100 navigation steps. The implementation provided in this paper supports "macro-actions"—sequences of low-level actions that may involve observing intermediate obstacles during their execution. Could the authors clarify how these actions are counted toward the navigation step budget used in the benchmarks? Furthermore, are the control and observation budgets for the baseline methods comparable to those used for the method proposed in this paper? Providing the average number of sub-actions executed per navigation step would help establish the comparability between different methods.

  1. Further isolating the contribution of "persistent evidence."

Table 3 shows that, in an ablation study conducted over 200 episodes, removing "persistent evidence" causes the success rate (SR) to drop from 27.50% to 24.50%. This indicates that this component makes a positive contribution, though the improvement is relatively limited compared to the gains provided by "active verification." In this ablation experiment, "cross-observation evidence" and "interaction outcomes" were also removed. A more detailed analysis—disentangling the respective roles of multi-view historical information, structured requirement-level evidence, and search feedback—would help clarify which factors are the most critical. Furthermore, pairwise statistics at the episode level would also provide valuable insights.

4 comments
AuthorsOpenReview ↗

Q1. How do evidence gaps influence observation acquisition?

Response to Reviewer gN4e

We thank the reviewer for recognizing EviNav's coherent system contribution and the distinction between finding a plausible object and verifying the requested target. Your suggestions help clarify the evidence-to-action mechanism, execution budget, and role of persistent feedback.

Q1. How do evidence gaps influence observation acquisition?

EviNav translates unresolved target evidence into active observation decisions. Rather than simply following the VLM's recommendations, the controller combines missing target attributes with observation distance, viewpoint diversity, and persistent candidate consistency to determine whether to approach, acquire another viewpoint, defer verification, or authorize STOP. Newly acquired observations continuously update the same candidate hypothesis, forming an evidence-driven inspection loop.

We have added a new qualitative figure on page 10 of the revised manuscript to further illustrate how evidence gaps guide active observation and target verification. Figure 5(a) illustrates this process through a recorded navigation traces. In WesternTown task 9, the UAV maintains the same candidate hypothesis across successive observations. At step 3, several distinctive features, including the rounded ends, top cap, and rear structure, remain unverified. Given the observation distance of 14.30 m, EviNav approaches the candidate to obtain a clearer view.

At step 7, the closer observation reveals additional target features, but the rear details and surface texture remain uncertain, while the available viewpoints provide insufficient angular diversity and exhibit an instance-consistency discrepancy. Although the VLM proposes STOP, EviNav continues verification by acquiring a complementary viewpoint instead of prematurely accepting the candidate. At step 8, the new observation resolves the remaining visual uncertainties, with the verifier reporting complete attribute support, and EviNav authorizes STOP.

This example demonstrates the central role of evidence in EviNav's active verification: evidence not only determines whether a candidate can be accepted, but also guides when and how additional observations should be acquired. By connecting persistent candidate evidence, adaptive observation, and final target acceptance, EviNav turns incomplete visual information into a coherent sequence of navigation and verification decisions.

AuthorsOpenReview ↗

Q2. How are navigation steps and macro-actions counted?

Q2. How are navigation steps and macro-actions counted?

EviNav follows the parameterized motion setting of UAV-ON and adopts a hierarchical navigation architecture. The original UAV-ON benchmark supports motion actions with variable translation distances and rotation angles, allowing navigation policies to operate at different motion granularities.

EviNav separates high-level semantic decision-making from local motion execution. When a new navigation decision is required, the navigation LLM selects a spatially grounded waypoint. During execution, the local controller follows the selected subgoal through physically executed motion primitives without repeatedly invoking the navigation LLM for each substep. Active inspection, recovery, and continuation of existing subgoals can also be handled directly by the corresponding controllers. Intermediate depth observations support obstacle avoidance and safe execution.

Our execution statistics further demonstrate the fine-grained nature of this hierarchical control. Across the complete 1,000-episode evaluation, EviNav travels 92,660.86 m over 56,694 high-level navigation steps, averaging only 1.63 m of actual movement per high-level step, despite executing an average of 2.477 motion substeps per high-level record. Thus, multiple substeps primarily support local planning and safe execution rather than substantially extending the distance traveled per navigation decision.

Existing UAV-ON methods also employ different motion granularities:

MethodMotion magnitude per actionAction design
AOA-F (UAV-ON)5 m horizontal movementFixed-step actions
AOA-V (UAV-ON)Variable distanceLLM-selected motion magnitude
APEX10 m forward movementRL-based action controller
OctMem-Agent3 units forward movementFixed-magnitude discrete actions
EviNav1.63 m actual movement per high-level step (average)LLM waypoint selection + local motion execution

The baseline values describe nominal action magnitudes, whereas EviNav's value is measured from its complete execution trajectories. The APEX value refers to the forward action in its public execution implementation.

These results show that EviNav operates within the continuously parameterized motion setting of UAV-ON, with a modest measured displacement per high-level decision. All Table 3 variants share the same 100-step high-level navigation budget, and physical execution, collisions, and trajectory length are explicitly accounted for. Within this parameterized-action setting, EviNav makes effective use of local motion substeps for fine-grained execution and obstacle avoidance while maintaining the stated 100-step high-level navigation budget. This provides a transparent basis for evaluating EviNav's hierarchical execution alongside methods with different action granularities.

AuthorsOpenReview ↗

Q3. Component-wise Ablation of Evidence Mechanisms

Q3. Component-wise Ablation of Evidence Mechanisms

Following the reviewer's suggestion, we refined our ablation study to more clearly distinguish the contributions of three evidence components. We conducted two additional ablations and clarified the scope of the original persistent-evidence ablation. All experiments use the same fixed 200 UAV-ON episodes as Table 3.

The three ablations are defined as follows:

  1. w/o Cross-Observation Evidence: Removes accumulated candidate evidence across observations, including multi-view histories and persistent verification information.
  2. w/o Interaction-Outcome Feedback : Removes interaction-outcome feedback and search/rejection memory while retaining cross-observation candidate verification. This more precisely describes the original w/o Persistent Evidence setting.
  3. w/o Structured Requirement Evidence : Removes structured target-requirement evidence used to guide candidate inspection and verification.

Table X. Component-wise ablation results on the same 200 UAV-ON episodes.

VariantSR ↑OSR ↑SPL ↑Failed STOPs / All STOPs ↓
Full EviNav27.50% (55/200)43.00%0.247021.43% (15/70)
w/o Cross-Observation Evidence23.00% (46/200)36.50%0.204522.03% (13/59)
w/o Interaction-Outcome Feedback24.50% (49/200)41.50%0.222724.62% (16/65)
w/o Structured Requirement Evidence25.00% (50/200)40.00%0.208825.37% (17/67)

The failed-STOP fraction is calculated using all STOP-terminated episodes within each variant.

Full EviNav achieves the best SR, OSR, SPL, and STOP reliability across all four configurations. Removing cross-observation evidence produces the largest reductions in SR (−4.50 percentage points) and OSR (−6.50 percentage points), highlighting the importance of accumulating candidate evidence across viewpoints. Removing interaction-outcome feedback reduces SR by 3.00 percentage points, indicating the value of incorporating previous search and verification outcomes into subsequent navigation decisions. Removing structured requirement evidence reduces SR by 2.50 percentage points and SPL from 0.2470 to 0.2088, while increasing the failed-STOP fraction from 21.43% to 25.37%, demonstrating the contribution of explicit target requirements to efficient exploration and reliable acceptance.

These results highlight the complementary roles of the three evidence components within EviNav's unified framework. Cross-observation evidence supports persistent candidate understanding, interaction-outcome feedback informs subsequent exploration, and structured requirement evidence connects target descriptions with inspection and acceptance decisions. Their integration enables EviNav to coordinate evidence-driven search, active verification, and reliable stopping, leading to consistently stronger navigation performance.

Thank you for the clarifications and additional experiments, which have addressed my concerns. I will maintain my original rating.

Reviewer_fpV9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EviNav addresses aerial object navigation by maintaining candidate evidence across search, verification, and stopping. It separates search cues from acceptance requirements, tracks supported, unknown, and contradicted evidence, and acquires additional views when needed. Without task-specific navigation training, it reports 27.0% SR and 22.98% SPL on the complete 1,000-episode UAV-ON benchmark, with component ablations and analyses of successful acceptance pathways.

Critical strengths
  1. The distinction between unknown and contradictory evidence directly addresses viewpoint-dependent ambiguity. Combining this distinction with persistent candidate histories gives the proposed search and verification loop a clear motivation.
  2. The complete 14-scene evaluation reports substantial SR and SPL gains. Results across target scales and seen/unseen environments, together with controlled module ablations, provide useful evidence of overall system effectiveness.
Critical weaknesses
  1. Evidence for improved acceptance reliability remains indirect. Table 3 keeps perception and planning fixed, but removing active verification also removes candidate approach and viewpoint acquisition; OSR falls from 43.0% to 21.5%. Thus, the SR drop cannot isolate better identity decisions from improved reachability. Table 5 describes only successful episodes. Existing incorrect-acceptance statistics or a control retaining approach while varying verification would strengthen this claim; otherwise, it should be narrowed.
  2. The final acceptance rule needs clarification. Equations (3)-(4) define evidence states and gaps, while Section 3.5 permits acceptance with incomplete evidence. Appendix D.3 allows roughly half the distinctive attributes to support a positive verifier response, but Appendix C.3 states that this alone cannot authorize STOP. Which unresolved requirements the controller permits, and how historical conflicts are resolved, remain unclear in the manuscript. Precise pseudocode or a supplementary-code pointer would make this central mechanism assessable.
Questions for authors
  1. Do existing logs provide incorrect-acceptance counts and denominators for the Table 3 variants? If so, please report them; otherwise, clarify which measured quantity supports the acceptance-reliability claim.
  2. Can you state the final STOP conditions for both acceptance pathways, including how unknown requirements and historical conflicts are handled, or identify the relevant function in the supplementary code?
3 comments
AuthorsOpenReview ↗

Q1. Acceptance behavior across episodes and ablations and Q2. STOP Authorization under Incomplete Observations

Response to Reviewer fpV9

We thank the reviewer for recognizing EviNav's motivation and the value of persistent candidate histories. We provide a detailed analysis of termination outcomes and clarify the evidence-based STOP authorization mechanism.

Q1. Acceptance behavior across episodes and ablations

Full EviNav achieves the highest successful-STOP fraction and the lowest incorrect-acceptance rates across all four variants. We examined every STOP-terminated episode in the Table 3 experiments on the same fixed 200 tasks. Beyond benchmark success or failure, we further classified unsuccessful STOPs into two categories: accepting an incorrect target instance and prematurely stopping at the correct target.

Table 3 variantSuccessful STOPsPremature STOPs on Correct TargetsWrong-Instance AcceptancesTotal Unsuccessful STOPs
Full EviNav55/70 (78.57%)9/70 (12.86%)6/70 (8.57%)15/70 (21.43%)
w/o Evidence-Grounded Search40/54 (74.07%)8/54 (14.81%)6/54 (11.11%)14/54 (25.93%)
w/o Persistent Evidence49/65 (75.38%)10/65 (15.38%)6/65 (9.23%)16/65 (24.62%)
w/o Active Verification22/33 (66.67%)6/33 (18.18%)5/33 (15.15%)11/33 (33.33%)

All percentages use the total number of STOP-terminated episodes in the corresponding variant as the denominator. Premature STOPs refer to correct-target acceptances outside the 20 m success radius. Wrong-instance acceptances were identified through candidate-identity checks. The two failure categories account for all unsuccessful STOPs.

EviNav consistently improves terminal acceptance outcomes. Compared with the active-verification ablation, Full EviNav increases successful STOPs from 22 to 55, improving the successful-STOP fraction from 66.67% to 78.57%. More importantly, the wrong-instance acceptance rate decreases from 15.15% to 8.57% (−6.58 percentage points), while premature STOPs on correct targets decrease from 18.18% to 12.86% (−5.32 percentage points). The overall unsuccessful-STOP fraction decreases by 11.90 percentage points.

These results provide direct quantitative evidence that EviNav's integrated active-verification process supports more reliable target acceptance, reducing both incorrect-instance acceptance and premature termination while achieving substantially more successful navigation outcomes.

Q2. STOP Authorization under Incomplete Observations

EviNav employs two complementary pathways for final target acceptance.

  1. Direct acceptance: When the current observation provides strong evidence of target identity, the controller authorizes STOP based on positive category and description matching, sufficient confidence and attribute support, visible key attributes, and valid contextual and distance checks.

  2. Persistent-evidence acceptance: When the current observation remains partially incomplete, EviNav uses accumulated candidate evidence to support target confirmation. STOP is authorized when semantic evidence is sufficient and the candidate satisfies range validity, depth reliability, multi-view hypothesis consistency, instance consistency, and applicable contextual checks.

EviNav maintains supported, unknown, and contradicted requirements in persistent candidate histories. Unknown attributes can be supplemented through subsequent observations, allowing complementary views to jointly support target confirmation. Explicit identity, attribute, description, or contextual conflicts block acceptance under the corresponding checks. Historical observations are retained to inform subsequent verification, while final STOP authorization follows the applicable acceptance conditions.

The implementation is provided in target_confirmation.py::_maybe_confirm_final_target, with direct acceptance checks in target_evidence.py::_target_confirmation_high_quality_direct_stop_allowed and multi-view consistency checks in target_hypotheses.py.

AuthorsOpenReview ↗

Q3. Ablation Study on the Stopping Mechanism.

Q3. Ablation Study on the Stopping Mechanism.

To isolate the contribution of evidence-based stopping, we conducted an additional ablation on the same 200 UAV-ON episodes used in Table 3. Specifically, we replaced the evidence-based STOP authorization with a simplified stopping rule, while retaining the original search policy, candidate approach, and viewpoint acquisition mechanisms. The results are summarized below.

MetricFull EviNavw/o Evidence-based STOP
SR ↑27.50%23.50%
OSR ↑43.00%29.50%
SPL ↑0.24700.2161
Successful STOPs ↑5547
Failed STOPs ↓1581
Failed-STOP Rate ↓21.43%63.28%

The results show that the full EviNav achieves a higher success rate (27.50% vs. 23.50%), while OSR increases from 29.50% to 43.00% and SPL improves from 0.2161 to 0.2470. More importantly, removing the evidence-based stopping mechanism leads to 81 unsuccessful stops out of 128 STOP decisions, corresponding to a failed-STOP rate of 63.28%. In contrast, the full EviNav produces only 15 unsuccessful stops out of 70 STOP decisions, reducing the failed-STOP rate to 21.43%, a decrease of 41.85 percentage points.

Furthermore, removing the evidence-based stopping mechanism leads to substantially more frequent STOP decisions (128 vs. 70), accompanied by a sharp increase in unsuccessful stops (81 vs. 15). Premature termination at incorrect candidates prevents the UAV from continuing its exploration, potentially leaving unexplored regions containing the true target. This explains the substantial decline in OSR from 43.00% to 29.50%, indicating that unreliable stopping decisions not only compromise final target confirmation but also limit the agent's ability to reach the correct target. These findings highlight a key advantage of EviNav: evidence-driven exploration and stopping are tightly integrated rather than treated as independent processes. By maintaining and updating candidate evidence throughout navigation, EviNav can defer uncertain stopping decisions, continue exploration when evidence is insufficient, and terminate only when sufficient evidence has been accumulated. This unified mechanism improves both exploration coverage and final stopping reliability.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Additional Evidence

Dear Reviewer,

Thank you again for your thoughtful review and constructive suggestions.

In response to your concerns, we have conducted an additional controlled STOP ablation, provided detailed incorrect-acceptance statistics across the ablation variants, and clarified the final STOP authorization mechanisms. We hope these additional experiments and explanations provide stronger evidence addressing your main concerns.

As the discussion period is approaching its end, we would sincerely appreciate hearing whether you have any remaining questions or concerns. We would be more than happy to provide further clarification.

If our additional evidence has addressed your key concerns, we would be very grateful if you could reconsider your assessment in light of these new results.

Thank you again for your time and consideration.

The Authors

Comments on the paper

1 comment

As we near the end of the discussion period for this submission, please check whether the authors have answered questions or responded to critiques. If they have, we encourage the reviewers to leave a comment for the authors. If reviewers have raised new points, a gentle nudge to the authors to respond would also be appreciated.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.