acceptodds
Under review as a conference paper at ICLR 2027

ConsistNav: Improving Action Consistency in Zero-Shot Object Navigation with Semantic Executive Control

Abstract

Zero-shot object navigation has advanced rapidly with open-vocabulary detectors, image–text models, and language-guided exploration. However, even after current methods detect a plausible target hypothesis, the agent may still oscillate between exploration and pursuit, or abandon the object near success. We identify this failure mode as an action consistency gap: semantic evidence is repeatedly reinterpreted at each step without persistent commitment across the episode. We introduce ConsistNav, a training-free zero-shot ObjectNav framework built around a semantic executive composed of three coordinated modules: Finite-State Executive Controller stages target pursuit through guarded semantic phases; Persistent Candidate Memory accumulates cross-frame target evidence into stable object hypotheses; and Stability-Aware Action Control suppresses rotational stagnation, ineffective pursuit, and unverified stopping. This design changes neither the detector nor the low-level planner; instead, it controls when semantic evidence should influence navigation and when it should be suppressed or revisited. We conduct extensive experiments on HM3D and MP3D, where ConsistNav achieves state-of-the-art results among compared zero-shot ObjectNav methods and improves SR by 11.4% and SPL by 7.9% over the controlled baseline on MP3D. Ablation studies and real-world deployment experiments further demonstrate the effectiveness and robustness of the proposed executive mechanism.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Deep Learning, Embodied AI, Vision-and-Language Navigation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tbApOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that zero-shot ObjectNav agents often fail even after a plausible target has been detected: they oscillate between exploration and pursuit, abandon the target near success, or stop without verification. It proposes ConsistNav, a training-free control layer between a fixed perception stack and planner, with a persistent candidate memory, a seven-state controller, and stability-aware action control. The paper also introduces an Action Consistency Score (ACS), and reports 84.2/41.2 SR/SPL on HM3Dv2 and 50.6/25.7 on MP3D, plus ablations and real-robot experiments.

Critical strengths
  • The motivation is clear. Most zero-shot ObjectNav work focuses on better semantic evidence, while this paper points out that keeping the target and stopping at the right time is a separate problem.

  • The improvement on MP3D is large. On this noisier benchmark with more categories, SR improves by 11.4 points and timeout failures drop by about 10 points.

Critical weaknesses
  • The "Non-executive baseline" seems to be ApexNav itself, but the paper does not state which components are reused or changed. After subtracting the reported gains, the baseline numbers exactly match ApexNav's results on all three benchmarks. So the method is only tested on top of one existing system, and testing it on another backbone such as VLFM would strengthen the claim of a general executive layer.
  • The technical novelty is limited. The method is basically a finite-state machine, a hand-designed candidate score, and stall detection, with dozens of manually specified thresholds and no sensitivity analysis. CogNav already models ObjectNav as a state machine with search, target observation, and candidate verification states, and ApexNav already fuses target evidence across frames, but the paper does not discuss how its design differs from them.
  • The ablation shows that the components work together, but it does not separate the effect of semantic commitment from generic anti-stalling fixes. FSEC alone slightly lowers performance, and there is no SAAC-only variant, so it is unclear how much of the gain comes from SAAC, which mainly fixes spinning and stalling. The ablation is also only done on HM3Dv2, not on MP3D where the main gain appears.
  • Fig. 4 does not show that the targeted failure mode is reduced. Unstable Commitment barely changes on HM3Dv2 (11.80% to 10.90%) and slightly increases on MP3D (19.90% to 20.59%), while most of the MP3D gain comes from fewer timeouts and missed targets. This makes it unclear how much of the improvement comes from better action consistency.
  • ACS is not yet convincing as a contribution. The persistence and release events have no operational definition, and the 12-action window for valid stopping is shorter than the 14-action verification turn budget, which may partly explain the low valid-stop rate of the FSEC variants; a sensitivity analysis on the window would help. Also, FSEC-only gets a higher ACS than PCM-only but has the lowest SR, mostly because it almost never stops at invalid opportunities (111/112), so ACS may over-reward not stopping.
  • The 150-episode ACS experiment is hard to interpret. SR there is only 9–19% across variants, far below the 76–84% reported on HM3Dv2, and it is not explained how these episodes and the 22 scenes were selected. If they were drawn mainly from baseline failures, the comparison may be biased.
Questions for authors
  1. Are the exploration-related parameters, such as the room-prior frontier bonus and the frontier shortlist, also used in the Non-executive baseline?
  2. For each ablation variant, how do the failure categories in Fig. 4 change? For example, does FSEC mainly reduce Unstable Commitment?
  3. Could you add a per-episode timeline showing detection confidence, PCM evidence, and FSEC state, so that readers can see the consistency problem and how the controller fixes it?
  4. Why are recent training-free methods such as CogNav, UniGoal, and WMNav not included in Table 1?
  5. The paper says the Non-executive baseline is run on your own stack (Habitat-Sim v0.3.1, YOLO-World), but its six SR/SPL numbers on the three benchmarks are exactly the same as those reported in the ApexNav paper. Getting identical numbers with a different simulator and detector is very unlikely. Did you re-run the baseline, or take these numbers from the ApexNav paper?
  6. In Fig. 4, the baseline's Unstable Commitment rate on MP3D is 19.90%, which cannot be obtained from 2,195 episodes (436/2,195 = 19.86% and 437/2,195 = 19.91%). All other numbers in the figure can. How was this number computed?
4 comments
AuthorsOpenReview ↗

Response to Reviewer tbAp (1/4): Contributions, Comparison with ApexNav, and Ablations(W1-W3)

Response overview

Thank you for recognizing that detecting a target and pursuing it reliably are distinct problems. We now add a second backbone and MP3D component experiments, connecting this evidence to the proposed decision rules. Below, we address transferability, decision rules, failure modes, ACS interpretation, and the sources of our comparisons.

W1. ApexNav inheritance and transfer to VLFM

ConsistNav extends ApexNav's navigation pipeline with PCM/FSEC/SAAC. We also implemented four VLFM variants, all evaluated on the same 2,195 MP3D tasks and settings (SR/SPL, %):

VLFM configurationSRSPL
Original, cited36.417.5
+PCM38.218.7
+FSEC37.215.0
+PCM+FSEC42.222.8
+PCM+FSEC+SAAC45.625.7

This extends evaluation to a second backbone. Among the comparable new variants, PCM+FSEC exceeds PCM by +4.0/+4.1 points and FSEC by +5.0/+7.8. The original VLFM row is cited; the comparisons among our new variants do not depend on treating it as a rerun. These results provide initial evidence that PCM and FSEC can transfer across backbones, supporting their potential as general-purpose executive components.

W2. Methodological novelty and threshold design

1.Our contribution is a revisable commitment policy for the same object instance. PCM retains identity and evidence; FSEC combines them with verification progress and recovery feedback; SAAC constrains execution.

ApexNav already fuses observations and handles missed detections; CogNav uses LLM-selected states; UniGoal uses graph matching, verification and blacklists. Our executive couples three decisions around the retained candidate:

  • Retention: viewpoint-induced ranking fluctuations need not replace an established commitment; provisional SUSPECT candidates can still be released.
  • Verification: a valid, in-range candidate remains pending until bounded verification finishes; incomplete verification differs from rejection.
  • Recovery/STOP: recovery may change the subgoal while retaining identity; rejection or exhausted recovery permits release; STOP requires verification of that candidate.

The contribution lies in these coupled transition rules. At the module level, PCM+FSEC improves both metrics over either module alone with SAAC disabled.

2.Thresholds should be assessed by function, selection protocol and sensitivity. Five relevant systems explicitly use them:

SystemDocumented example
ApexNavSemantic/geometric switching: max/mean ratio 1.10; score standard deviation 0.015.
CogNavFrontier landmark assignment within 1 m of a frontier.
UniGoalMatching threshold 0.9; stage thresholds 0.5/0.9.
VLFM codeDetection defaults 0.8/0.4 and a stop radius.
SG-Nav codeSpatial matching threshold 0.01.

Training-free does not mean parameter-free. These examples serve different functions. Benchmark criteria, sensor/geometry tolerances and executive guards must be distinguished: camera/robot height and depth error can require internal calibration, while the benchmark success radius (e.g. ApexNav's 0.2 m) stays fixed. Explicit disclosure makes these assumptions reproducible and testable. We will clarify inherited versus added parameters, their roles and selection protocol.

W3. SAAC-only and the source of the gains

We added SAAC directly to ApexNav without PCM/FSEC (SR/SPL, %; ✗: absent, ✓: present):

DatasetSAACSRSPL
HM3Dv2✗76.238.0
HM3Dv2✓78.139.6
MP3D✗39.217.8
MP3D✓42.317.7

In the new SAAC-only results, SR exceeds the cited baseline on both benchmarks; HM3Dv2 SPL also improves, while MP3D SPL remains nearly unchanged. This trend is consistent with SAAC helping task completion. Furthermore, in the shared VLFM setting, adding SAAC to PCM+FSEC raises SR/SPL from 42.2/22.8 to 45.6/25.7, a gain of +3.4/+2.9 points, providing direct evidence of SAAC's additional contribution to success rate and path efficiency in this setting.

AuthorsOpenReview ↗

Response to Reviewer tbAp (2/4): Failure Analysis, ACS Interpretation, and Evaluation Scope (W4–W6)

W4. Failure analysis from 500 additional runs

1.We evaluated five configurations on the same 100 prespecified HM3Dv2 tasks across 36 scenes: 500 runs, each with a 500-action budget. The baseline was rerun. Counts are mutually exclusive; percentages use each row's failures.

ConfigurationFailuresIncorrect STOP¹No frontierStuck, gave upStep limit
Baseline5620 (35.7%)6 (10.7%)2 (3.6%)28 (50.0%)
PCM6539 (60.0%)8 (12.3%)3 (4.6%)15 (23.1%)
FSEC8315 (18.1%)1 (1.2%)0 (0%)67 (80.7%)
PCM+FSEC7719 (24.7%)6 (7.8%)1 (1.3%)51 (66.2%)
ConsistNav6924 (34.8%)5 (7.2%)3 (4.3%)37 (53.6%)

¹ The agent issued STOP to claim success, but the environment recorded failure.

PCM+FSEC reduces incorrect STOPs relative to PCM. The count falls from 39 to 19: 20 fewer per 100 tasks. In paired comparisons, 21 tasks avoid PCM's incorrect STOP and only one changes in the opposite direction. This directly supports fewer false success claims.

Full reduces timeouts and completes more tasks than PCM+FSEC. With SAAC present, step-limit failures fall from 51 to 37 and successes rise from 23 to 31 (+8). Ten PCM+FSEC timeouts become Full successes, while two successes become failures. The paired outcomes confirm more tasks completed.

These results support the practical value of action-consistent execution. PCM+FSEC reduces incorrect STOPs relative to PCM, while Full reduces timeouts and completes more tasks relative to PCM+FSEC. Together, these gains support coordinating verified stopping with execution progress.

W5. Event definitions, timing and ACS interpretation

1. How persistence and release are judged. Both concern an instance the agent has already pursued. Persistence tests whether pursuit of the same supported instance continues within the response window. Changing the approach waypoint can preserve that commitment; a pass must also satisfy the evaluator's progress criterion. Release tests whether pursuit ends after clear counterevidence against that instance. A single missed detection or unfinished verification scan is insufficient counterevidence. Appendix C.1 will specify the actual evidence criteria, cross-frame instance association, response windows and event deduplication. Event denominators reflect the opportunities encountered by each trajectory.

2. Does the 12-action window underestimate valid stopping? The 12-action window starts when a valid-stop opportunity appears; the 14-action maximum verification-turn budget starts upon entering verification. Their onsets can differ. A valid opportunity during scanning followed by a correct STOP on action 13 or 14 can therefore fail the current window. We plan to hold trajectories and event onsets fixed and rescore at 12, 14 and 16 actions. For each variant, we will report passed/eligible valid-stop events and how many previously failed events subsequently ended in a voluntary STOP validated as successful by Habitat. This will quantify the window's effect on FSEC results.

3. How to interpret FSEC-only's higher ACS. It mainly reflects stronger invalid-stop avoidance: 111/112, versus 5/20 timely valid stops. PCM-only achieves 6/85 and 17/18, respectively. The original formula weights the two termination rates equally; 112 opportunities do not receive extra weight merely because they are more numerous.

W6. Purpose and interpretation of the 150-episode comparison

Thank you for asking about the scope and purpose of this experiment.

1. Shared episodes and the SR gap. The four configurations were evaluated on the same 150 episodes across 22 scenes. Baseline completed 20 episodes and Full ConsistNav completed 29. Both SRs are substantially below the main HM3Dv2 evaluation; this comparison alone does not identify the cause of the gap. We plan two additional random draws of 150 episodes using different seeds and a common prespecified sampling procedure. For each draw, we will report ACS, its components, SR and SPL to test whether the observations recur across samples.

2. Examining ACS on shared episodes. The experiment asks whether ACS captures decision changes introduced by the executive and whether those changes accompany better navigation outcomes. From Baseline to Full, ACS rises from 0.5292 to 0.7493. For example, when counterevidence calls for releasing a candidate, the passed/eligible release count rises from 31/45 to 51/51 (Table C.3). On the same episode set, SR rises from 13.33% to 19.33%, and SPL from 6.01% to 9.43% (Table 4). In this comparison, Full completes more tasks and improves performance accounting for path length. ACS adds a behavioral account of these accompanying changes, including more reliable release of disconfirmed candidates.

AuthorsOpenReview ↗

Response to Reviewer tbAp (3/4): Exploration Settings, Failure Analysis, and Episode Timeline(Q1-Q3)

Thank you for asking how exploration, component effects and recorded behavior support our claims.

Q1. Were the exploration settings shared?

The two systems use different exploration rules. We report ApexNav's published SR and SPL for the baseline. We did not rerun ApexNav with our room-prior bonus or our 12-frontier shortlist.

  1. ApexNav. Room context enters image-text matching. With strong semantic cues, the planner selects a variable-sized frontier set by relative semantic score and plans a tour; otherwise it chooses the nearest reachable frontier. This differs from a separate room-prior bonus and a top-12 path-check shortlist. Room prompt · Frontier selection

  2. ConsistNav. When the target-room prior is known and no provisional, active or sufficiently supported candidate is present, frontiers receive a prior-dependent bonus of up to 22.0. The bonus is disabled when such a candidate exists. Ranking combines information gain relative to distance, semantic value, this bonus, and revisit/failure penalties. Only the top 12 receive detailed path checks. These settings can change explored locations and encountered targets. Table B.2 specifies our settings; Table 1 therefore compares complete systems. The component comparisons in Q2 use variants within the same experiments.

Q2. How do failures change across ablations?

We ran all five configurations on the same 100 prespecified HM3Dv2 tasks (36 scenes, 500 runs). The baseline here was run on these tasks. The table counts how failed episodes ended. Parentheses give the percentage of failures in that row. With 100 tasks per row, each count also gives the percentage of all tasks.

ConfigurationFailuresIncorrect STOP¹No frontierStuck, gave upStep limit
Baseline5620 (35.7%)6 (10.7%)2 (3.6%)28 (50.0%)
PCM6539 (60.0%)8 (12.3%)3 (4.6%)15 (23.1%)
FSEC8315 (18.1%)1 (1.2%)0 (0%)67 (80.7%)
PCM+FSEC7719 (24.7%)6 (7.8%)1 (1.3%)51 (66.2%)
ConsistNav6924 (34.8%)5 (7.2%)3 (4.3%)37 (53.6%)

¹ The agent issued STOP to claim success and the environment recorded failure.

FSEC and incorrect STOPs. On 21 of the 100 tasks, PCM issued an incorrect STOP that PCM+FSEC avoided. Six of those runs succeeded with PCM+FSEC and 15 reached the step limit. Only one task changed in the opposite direction, giving 20 fewer incorrect STOPs overall. Total successes fell from 35 to 23 on this subset. FSEC checks target evidence before authorizing STOP, and the agent made fewer false claims of success with this check in place.

PCM and SAAC. Adding PCM to FSEC raised successes from 17 to 23 and reduced step-limit endings from 67 to 51. With SAAC added to PCM+FSEC, 10 tasks that previously reached the limit succeeded. Two previously successful tasks failed, giving eight more successes overall (23 to 31). Incorrect STOPs rose from 19 to 24. With PCM retaining candidate evidence across views, more tasks succeeded within the step budget, and adding SAAC then turned 10 previous timeouts into successes.

Evidence for better commitment retention is less conclusive. We observed release of a supported target without contrary evidence in 0 PCM and 4 PCM+FSEC failures, with 24 and 33 failures unassessable, respectively.

Q3. Can you show an episode timeline?

We provide a successful logged run from HM3Dv2 scene 4ok3usBNeis: sofa, source episode 5 (runtime ID 10), negative-update coefficient 0.12. The table uses controller indices; confidence is PCM's internal score.

IndexRecorded event for candidate 2
124SEARCH to SUSPECT to APPROACH; candidate selected
125–126Confidence increases: 0.60 to 0.80 to 1.00
156APPROACH to VERIFY
157–170Same candidate retained through 14 verification actions
170Verification confirmed; VERIFY to SUCCESS
171STOP issued (action 0)

The environment log independently records execution of STOP at environment step 135, success=1, distance to goal 0.02958 m, and SPL 0.62716. Controller indices and environment steps are different counters.

The trace connects accumulated candidate evidence, sustained approach, bounded verification and successful stopping. Candidate 2 remains selected throughout approach and verification; the controller confirms verification before issuing STOP, and the environment confirms task success. This provides an observable link from the internal decision sequence to the navigation outcome.

AuthorsOpenReview ↗

Response to Reviewer tbAp (4/4): Method Comparisons, Baseline Sources, and Percentage Calculation(Q4-Q6)

Q4. Why were CogNav, UniGoal and WMNav missing from Table 1?

CogNav uses LLM-driven cognitive states, UniGoal graph reasoning for multiple goal types, and WMNav a VLM-based world model. Their reasoning, memory and action selection make them relevant comparisons.

The updated table below adds all three and SysNav (2026). It reports published ObjectNav SR/SPL in percent. Bold marks each column's highest value, underlining its second highest, and a dash an unreported result. Y/N mean Yes/No. Each cited row is a literature reference for the complete system.

MethodUnsupervisedZero-shotHM3Dv2 SRSPLHM3Dv1 SRSPLMP3D SRSPL
PONI (2022)NN————31.812.1
OpenFMNav (2024)YY——54.924.437.215.7
InstructNav (2024)YY58.020.9————
VLFM (2024)YY63.632.552.530.436.417.5
TriHelper (2024)YY——56.525.3——
SG-Nav (2024)YY49.625.554.024.940.216.0
ApexNav (2025)YY76.238.0‾\underline{38.0}59.633.0‾\underline{33.0}39.217.8
CogNav (ICCV 2025)YY——72.526.246.616.1
UniGoal (CVPR 2025)YY——54.525.141.016.4
WMNav (IROS 2025)YY72.233.358.131.245.417.2
SysNav (RA-L 2026)YY80.8‾\underline{80.8}37.263.7‾\underline{63.7}30.550.718.1‾\underline{18.1}
ConsistNavYY84.241.263.234.850.6‾\underline{50.6}25.7

Q5. Were the six baseline numbers rerun or quoted?

The six values were directly quoted from ApexNav. They are 76.2/38.0 on HM3Dv2, 59.6/33.0 on HM3Dv1 and 39.2/17.8 on MP3D. This explains the numerical identity.

Table 1 already cites ApexNav and identifies prior rows as literature-reported results. However, Sec. 4.2 describes the corresponding differences as a controlled same-stack comparison. Table 2 repeats 76.20/38.00 under "Non-executive baseline" with the same-stack caption. Those descriptions are inconsistent with the actual source of the numbers.

Habitat-Sim v0.3.1 and YOLO-World/GroundingDINO describe our implementation. The quoted ApexNav row follows its source evaluation. We will label Table 2's cited row "ApexNav (reported)", specify the scope of our implementation settings, and correct the abstract, Secs. 4.2/4.4 and conclusion. Comparisons with the quoted row measure overall system performance and cannot isolate executive control.

There is also positive evidence from the new runs. Across the four added VLFM variants, tasks and settings are shared. With SAAC disabled, adding FSEC to PCM raises SR/SPL from 38.2/18.7 to 42.2/22.8 on 2,195 MP3D tasks. This supports the component's contribution in that configuration. We will ground component claims in these within-experiment comparisons and explicitly identify the provenance of every baseline row.

Q6. How was 19.90% obtained?

We took 19.90% directly from ApexNav Fig. 7, where it labels category C (False Positive) on MP3D. You are correct that no integer count out of 2,195 episodes rounds to 19.90% at two decimal places. ApexNav, Fig. 7

1. Inferring the count. The other five categories have published percentages of 39.23%, 13.85%, 4.92%, 14.67% and 7.43%. On a denominator of 2,195 episodes, these imply counts of 861, 304, 108, 322 and 163, respectively. If the six categories partition all episodes, the count left for C is

nC=2,195−(861+304+108+322+163)=437n_C = 2{,}195 - (861 + 304 + 108 + 322 + 163) = 437

2. Checking the displayed value. Subtracting the five rounded percentages from 100% reproduces the figure label.

100% − (39.23% + 13.85% + 4.92% + 14.67% + 7.43%) = 19.90%

Using the inferred episode count directly gives

4372,195\displaystyle\frac{437}{2{,}195} × 100% = 19.9088838…%

The two calculations explain how the displayed percentages can sum to 100%. ApexNav does not specify its rounding procedure or publish a raw count for C, so taking the remainder is one possible explanation. We will credit ApexNav Fig. 7 in our caption and restore its original label, False Positive.

Summary and thanks

The added VLFM comparisons, failure analysis and episode trace clarify our contribution and its limits. Thank you for your careful review and constructive suggestions.

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

This paper addresses the lack of sustained target pursuit, insufficient verification, and unreliable stopping in zero-shot ObjectNav, and proposes ConsistNav, a training-free framework. While keeping the underlying perception and planning modules unchanged, the method combines Persistent Candidate Memory, a Finite-State Executive Controller, and Stability-Aware Action Control to coordinate target pursuit, verification, recovery, and stopping. The paper further proposes the Action Consistency Score to evaluate commitment persistence, commitment release, and stopping behavior. The authors validate the method through performance comparisons, component ablations, and behavioral analysis on HM3Dv1, HM3Dv2, and MP3D, and demonstrate deployment feasibility on a real robot.

Critical strengths
  1. The method is designed around a well-identified navigation failure mode, and the candidate memory, staged verification, and recovery mechanisms are sensibly coordinated.
  2. Improvements in Success Rate (SR) and Success weighted by Path Length (SPL) are reported on three simulation benchmarks, and the ablation studies support the complementarity and effectiveness of PCM and FSEC. Real-world robot experiments demonstrate the deployment feasibility of the system.
Critical weaknesses
  1. The criteria for the core behavioral evaluation are not specified clearly enough. The paper presents action consistency as one of its main contributions, but the event trigger conditions, evaluation window, and pass criteria for commitment persistence (P) remain unclear. The appendix provides the control procedure and aggregate event counts, but I would like to see a more detailed specification of these rules.
  2. The relative gains in real-world settings lack quantitative support. The real-world experiments report the success rate of the full method, but the comparison against the non-executive baseline is presented only qualitatively.
Questions for authors
  1. Please clarify the evaluation criteria for commitment persistence (P) in Table C.3: when is an evaluation event triggered, how long is the observation window, and what counts as a pass?
  2. Could you additionally report the baseline's success rate under the same real-world task settings, so that the practical benefit of the added executive mechanism can be assessed? The full number of trials is not necessary. Time is limited, so a reasonably sized small-scale experiment would suffice.
  3. Is the system built upon ApexNav? In the ablation study, the numbers reported for the non-executive baseline are identical to those of ApexNav.
2 comments
AuthorsOpenReview ↗

Response to Reviewer wHqG (1/2): Overview and Responses to Weaknesses(W1-W2)

1. Response overview

Thank you for your careful review and constructive suggestions, and for recognizing our motivation, the coordination of memory, verification and recovery, and the simulation and robot evaluations. Below, we address your concerns by clarifying the persistence criteria, providing the real-world comparison, and making the ApexNav foundation explicit.

2. Responses to the two weaknesses

W1. Commitment-persistence evaluation

  1. Scope and reported counts. We appreciate your request for clearer evaluation criteria. P measures whether pursuit is maintained while evidence supports the target. Appendix C.1 pools events: CP=nPpass/(nPpass+nPfail)C_P=n_P^{pass}/(n_P^{pass}+n_P^{fail}). Existing Table C.3 reports:

    ConfigurationPassFailEligibleCPC_P
    Non-executive104140.7143
    PCM-only7071.0000
    FSEC-only83110.7273
    ConsistNav234270.8519
  2. New Appendix Table C.4. To address your concern about the commitment-persistence (P) criteria in Table C.3, we add an appendix explanation of the event triggers, evaluation window and pass criteria.

    ItemPredeclared rule
    TriggerFirst joint occurrence: pursuing a valid, reachable target gg; support in at least 2 of the last 3 observations; at least 18 actions remaining. No prior persistence requirement.
    WindowHP=12H_P=12 environment actions: at0,…,at0+11a_{t_0},\ldots,a_{t_0+11}. One observation per action; no wall-clock or FSM-tick window.
    PassAll actions stay linked to pursuing, verifying or recovering toward gg, with zero unjustified switches/abandonments. Require executed-action/subgoal records, not memory or state labels alone.
    FailFirst unjustified switch/abandonment while support holds, or invalid STOP. Returning later does not erase failure.
    Early endingVoluntary STOP for gg validated by Habitat before failure: pass. Undecided interruptions/incomplete logs: censored/unscorable. An earlier decided outcome takes priority.
    DeduplicationFirst eligible onset per episode/object only; the same logging/association rules for every method, without FSEC-state prerequisites.
    Robustness/reportingReuse identical onsets at 6/12/18 actions. Report passed, failed, censored and unscorable counts; freeze detector, association and scorer versions.

W2. Quantitative real-world gains

  1. Comparison setting. Thank you for suggesting this comparison. We now provide an independent evaluation under the same real-world task settings: 15 trials per category and 60 trials per method.

  2. Measured results. Each cell reports successes/trials and SR; gains are absolute percentage points (pp).

    TargetApexNavConsistNavGain (pp)
    Couch8/15 (53.33%)10/15 (66.67%)+13.33
    Plant10/15 (66.67%)12/15 (80.00%)+13.33
    Chair12/15 (80.00%)14/15 (93.33%)+13.33
    Toilet7/15 (46.67%)10/15 (66.67%)+20.00
    Overall37/60 (61.67%)46/60 (76.67%)+15.00
  3. Supported conclusion. The results show a net increase of 9 successes, with higher SR in all tested categories. We interpret this as evidence of a practical benefit over ApexNav in the tested setting.

Summary and Thanks

Thank you for your careful review and constructive suggestions. Compared with the ApexNav backbone, ConsistNav adds an execution layer that coordinates persistent evidence, commitment, verification and recovery to support consistent target pursuit and STOP decisions. In response to W1-W2, we specify P's evaluation criteria in the appendix and provide a comparison under the same real-world task settings, where SR rises from 61.67% to 76.67% (+15.00 pp). We sincerely thank you again for helping us clarify our contribution and supporting evidence.

AuthorsOpenReview ↗

Response to Reviewer wHqG (2/2): Direct Answers to Questions(Q1-Q3)

Thank you for carefully reviewing our manuscript and posing three insightful questions. Your feedback has helped us identify details needing further clarification. Below, we address Q1–Q3 in turn with relevant experimental results and analysis.

Direct answers to the three questions

Q1. When is P triggered, how long is its window, and what is a pass?

  1. Evaluation criteria. Thank you for asking us to make these criteria explicit. Appendix Table C.4 (Reply 1) specifies the event triggers, evaluation window and pass criteria for commitment persistence (P):

    ItemP-v1 rule
    TriggerFirst joint occurrence of pursuing a valid, reachable gg with support in at least 2 of the last 3 observations and at least 18 actions remaining.
    Window12 environment actions from t0t_0 through t0+11t_0+11; sensitivity checks at 6/12/18 on the same onsets.
    PassMaintain goal-linked pursuit/verification/recovery throughout; zero unjustified switch/abandonment. Valid voluntary STOP for gg before failure is an early pass.
  2. Boundary rules. Invalid STOP or an unjustified switch/abandonment is failure, even if pursuit later resumes. Undecided interruptions/incomplete logs are excluded and counted separately; an earlier decided outcome takes priority. Only the first eligible episode/object event is used.

Q2. Can you report baseline SR under the same real-world settings?

  1. Comparison under the same settings. Yes, and thank you for this suggestion. Both methods were evaluated on 15 trials per category, covering Couch, Plant, Chair and Toilet, for 60 trials per method.

  2. Category-wise and overall results. Cells show successes/trials and SR; gains are absolute percentage points (pp).

    TargetApexNavConsistNavGain (pp)
    Couch8/15 (53.33%)10/15 (66.67%)+13.33
    Plant10/15 (66.67%)12/15 (80.00%)+13.33
    Chair12/15 (80.00%)14/15 (93.33%)+13.33
    Toilet7/15 (46.67%)10/15 (66.67%)+20.00
    Overall37/60 (61.67%)46/60 (76.67%)+15.00

    The absolute SR gain is 15.00 pp, a net increase of 9 successes.

  3. Results and interpretation. Under the same real-world task settings, ConsistNav achieves higher SR than ApexNav in all four target categories, demonstrating better navigation performance. These results support the practical effectiveness of our method in this real-world evaluation.

Q3. Is ConsistNav built on ApexNav, and why do the baseline numbers match?

  1. Implementation relationship. Yes, ConsistNav is built on ApexNav. Thank you for highlighting this relationship; we apologize for the unclear attribution. ApexNav supplies the base navigation system; our contribution coordinates persistent candidate evidence, commitment, staged verification, execution feedback and STOP authorization.

  2. Numerical correspondence. The reported HM3Dv2 rows are:

    Reported rowSR (%)SPL (%)
    Non-executive baseline76.2038.00
    ApexNav76.238.0

    We agree that these matching values call for a clear account of their source; the numbers alone do not establish independent reruns.

  3. ApexNav and the Ablation Baseline. ConsistNav builds on ApexNav by adding three executive components (PCM, FSEC, and SAAC) and reusing ApexNav’s detector and mapping module. In Table 2, the non-executive baseline corresponds to ApexNav with all three ConsistNav components disabled. Its HM3Dv2 SR/SPL scores (76.20/38.00) therefore match those reported for ApexNav in Table 1. We will clarify this relationship in Section 4 and explicitly label the baseline as ApexNav in Table 2.

Summary and Thanks

Thank you for your thoughtful questions and careful review. Building on ApexNav, ConsistNav coordinates candidate evidence, commitment, staged verification, recovery and STOP authorization to keep navigation decisions consistent over time. Our answers to Q1-Q3 specify the P protocol, provide the real-world baseline comparison, and clarify the implementation relationship and the provenance information to be added. We sincerely thank you again for your time and constructive feedback.

Reviewer_9pVhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies action consistency in zero-shot object navigation. The authors propose ConsistNav, which uses PCM to store observation evidence and execution history, FSEC to manage candidate approach, verification, and release, and SAAC to monitor motion progress, trigger recovery, and constrain stopping. The authors also define the Action Consistency Score (ACS) to evaluate target persistence, target release, and stopping behavior. The paper reports navigation results on HM3Dv1, HM3Dv2, and MP3D, along with module ablations, behavioral analyses, and real-world experiments.

Critical strengths
  1. The motivation focuses on interrupted target pursuit after target discovery. After semantic exploration identifies a candidate target, detection fluctuations caused by viewpoint changes may lead to target switching or interrupted pursuit. The paper studies how cross-frame evidence and execution states can maintain target pursuit and guide decision adjustments when counterevidence emerges.
  2. The experiments include simulation comparisons, module ablations, and real-world validation. The authors evaluate on HM3Dv1, HM3Dv2, and MP3D, compare different module combinations under shared perception, mapping, and planning configurations, and report task metrics and action consistency components. The paper also presents 60 real-world trials across four target categories, providing navigation results in physical environments.
Critical weaknesses
  1. The novelty and contributions relative to prior work need further clarification. The paper combines candidate memory, staged pursuit, and execution control, but multi-frame evidence fusion, target verification, and failure recovery have already been explored in prior work. The current description does not clarify whether the contribution primarily comes from new decision mechanisms or from combining and modifying existing mechanisms. The authors may specify the additions beyond prior work, rather than only describing each module’s functionality.
  2. The ablation experiments do not yet distinguish the contributions of individual designs. Table 2 compares combinations of PCM, FSEC, and SAAC, but these modules jointly involve target persistence, verification, recovery, and stopping rules. The existing results therefore do not establish which changes account for the performance gains. The authors may further separate these rules, test them individually while keeping other settings unchanged, and report the corresponding behavioral changes.
  3. The generation and interception of STOP requests need clarification. Algorithm B.1 specifies that STOP is issued only in the SUCCESS state, whereas Appendix B.2 describes STOP interception in other states. The paper does not explain where these additional requests originate or how the interception mechanism relates to the stopping conditions. The authors may clarify this part of the design.
  4. The statistical definitions and interpretation of ACS require further detail. In Table 4, ACS increases while the pass rate for valid stopping opportunities decreases from 20/24 to 11/44. An increase in the aggregate score therefore does not indicate improvement in every behavioral component. The authors may provide the statistical rules and analyze why valid stopping opportunities are missed during navigation.
Questions for authors
  1. Compared with existing methods for cross-frame evidence fusion, target verification, and execution recovery, what specific methodological contributions does this paper aim to emphasize? Which additional rules directly address interruptions in target pursuit caused by viewpoint changes?
  2. Equation (5) uses different confidence update magnitudes for matched observations that support or do not support the target category. What is the rationale for setting the coefficient to 0.12?
  3. Which component generates STOP requests outside the SUCCESS state, and how does the system decide whether to continue approaching, re-verify, or recover after intercepting them?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 9pVh (1/3): Contributions, Comparison with ApexNav, and Ablations(W1-W2)

1. Overview

Thank you for your careful review and for recognizing the motivation, multi-benchmark experiments, and real-robot validation. Your four concerns clarify what needs strengthening: concrete decision rules, module versus rule ablations, STOP proposals versus execution, and ACS components and statistical definitions. We address these points below.

2. W1: Specific contribution beyond prior mechanisms

2.1. Evidence-to-action coordination.

Our contribution is persistent, revisable target commitment: accumulated evidence constrains identity retention, verification, recovery, and STOP. With a fixed planner, perceptual evidence ete_t and execution feedback rtr_t update memory; the executive maintains candidate identity and phase in ξt\xi_t:

(\xi_t,a_t)=\mathcal F(\xi_{t-1},\mathcal C_t,\mathcal M_t,x_t,r_t;\mathcal P).$$ **2.2. Difference from prior work.** 2. W1: Specific contribution beyond prior mechanisms | Method | Existing focus | Our distinction | | --- | --- | --- | | [ApexNav](https://arxiv.org/html/2504.14478v2) | Volume-weighted fusion, missed-detection handling, action costs. | Explicitly link identity, pending verification, and STOP authorization. | | [CogNav](https://arxiv.org/html/2412.10439v3) | LLM-selected states, verification and recovery. | Inspectable guards for retaining/releasing a candidate and stopping. | | [UniGoal](https://arxiv.org/html/2503.10630v1) | Graph matching, verification and blacklisting. | Coordinate active identity with execution state and recovery feedback. | ApexNav Eq. (2) weights confidence by observation volume ($n$: accumulated; $v$: current): $$c_{\rm new}=\frac{n c_{\rm old}+v c_{\rm obs}}{n+v}.$$ Fusion or blacklisting alone does not specify whether a transient view change should cancel an ongoing pursuit. We address this control gap, rather than claim the first memory, verification, or recovery mechanism. **2.3. Concrete rules (Sec. 3.4; Appendix B).** Keep an established identity despite ranking fluctuations; provisional SUSPECT commitments may still be released. Keep a valid, in-range unfinished scan in VERIFY/pending. During permitted local recovery, change the subgoal without necessarily changing identity; rejection or exhausted recovery updates memory and enables release. ## **3. W2: Module evidence and rule diagnostics** **3.1. Separate cited baselines from measured ablations.** The 76.20/38.00 baseline was quoted from ApexNav, not rerun on our stack. We correct the same-stack wording. Table 2 reports (SR/SPL, %): | Configuration | SR | SPL | | --- | --- | --- | | ApexNav (cited baseline) | 76.20 | 38.00 | | PCM | 77.10 | 39.60 | | FSEC | 75.50 | 37.75 | | PCM+FSEC | 81.00 | 40.75 | | Full | 84.20 | 41.24 | This cited baseline cannot isolate executive gains or factorial interaction. For $v=(SR,SPL)$, the reported difference $v_{\rm Full}-v_{\rm PCM+FSEC}=(3.20,0.49)$ pp describes module settings, not individual rules. Table 1 gives literature context: | Benchmark | ApexNav | ConsistNav | | --- | --- | --- | | HM3Dv2 | 76.2/38.0 | 84.2/41.2 | | HM3Dv1 | 59.6/33.0 | 63.2/34.8 | | MP3D | 39.2/17.8 | 50.6/25.7 | A separate real-robot comparison under matched task settings gives ApexNav 37/60 versus ConsistNav 46/60 successes ($\Delta SR=15$ pp). **3.2. Added paired rule experiments.** We added 500 runs on 125 fixed HM3Dv2 tasks (36 scenes, six categories), excluding the six earlier pilot tasks. Same stack, 500 steps, raw Habitat scores; this separate current-code cohort does not reproduce the submitted tables. | Variant | SR % (k/n) | SPL % | VERIFY actions; windows | | --- | --- | --- | --- | | Full | 24.80 (31/125) | 10.22 | 14.00; 403 | | Relaxed commitment | 24.00 (30/125) | 9.81 | 14.00; 397 | | Early-exit verification | 22.40 (28/125) | 9.18 | 9.00; 434 | | No local recovery* | 24.00 (30/125) | 10.31 | 14.00; 413 | VERIFY lists mean actions and completed-window counts. Relaxed commitment removes only the subgoal gate (2 s or 0.5 m); early exit keeps confirmation predicates and allows positive completion before 14 actions. Logs show 1538 gate bypasses and 196 early exits. **3.3. What these results support.** *Local recovery never triggered in these 500 paired runs; this row cannot identify the effect of removing recovery. Full minus relaxed commitment gives $\Delta SR=+0.80$ pp (95% paired scene-bootstrap CI [-2.42, 4.24]). Single executions retain runtime variability; the observed differences do not establish reliable per-rule gains. STOP hold was not tested. New ACS/false-confirmation rates remain N/A without the event scorer and independent labels. Historical counts stay separate. The new runs quantify rule activation and behavior beyond the submitted module table.
AuthorsOpenReview ↗

Response to Reviewer 9pVh (2/3): STOP Handling and ACS Interpretation(W3-W4)

W3. STOP generation, interception, and completion

Thank you. B.1 describes STOP for task completion; B.2 handles intermediate STOP returns. The missing distinction is between a helper's return value and an action accepted for execution.

3.1. Execution interfaces and their roles.

ExplorationFSM implements the execution layer. SUCCESS is its internal state; STOP is an action value. Its relevant interfaces are:

Class/functionRole
ExplorationFSMMaintains execution state and coordinates actions.
callActionPlanner()Dispatches by state, then checks returned actions.
planActionFromSubgoal()Converts a subgoal/path into a motion proposal.
executeVerifyState()Checks the candidate and runs verification.
runApproachRecovery()Produces the next local recovery action.
executeFailoverState()Handles recovery or candidate reselection.

3.2. Where extra STOPs originate.

The subgoal helper returns STOP for an invalid subgoal/empty path; the verification handler does so for a missing/failed candidate; the recovery helper does so when inactive/exhausted. These returns do not establish that the target has been reached.

3.3. How interception relates to stopping conditions.

Target evidence, distance and temporal checks determine entry into SUCCESS. The dispatcher reads the resulting state. For SEARCH_OBJECT STOPs outside SUCCESS, APPROACH/VERIFY/FINAL_APPROACH enter FAILOVER and invoke its handler; FAILOVER invokes it again. Any residual STOP becomes TURN_LEFT; EXPLORE STOPs do too. SEARCH_OBJECT/EXPLORE are flow labels.

For example, a missing candidate in VERIFY makes verification enter FAILOVER and return STOP. The dispatcher then requests recovery/reselection instead of ending the task as completed. Thus completion checks decide whether SUCCESS is justified; interception prevents a helper return from bypassing that decision. We will expose this order in B.1/B.2 and distinguish failure/runner termination.

W4. ACS statistics and missed valid stopping opportunities

Thank you. Full ACS rises from 0.5292 to 0.7493 while valid-stop passes fall from 20/24 to 11/44. We agree that continuing when stopping is invalid and stopping promptly when it is valid must be distinguished.

4.1. Statistical rules and the aggregate increase.

Each rate is passed/eligible events. Use method-independent triggers and fully observed fixed windows; exclude/report censored or unscorable events. Pool event counts, not episode ACS. Shared tasks can yield different denominators because trajectories differ.

T+ passes only for voluntary STOP within 12 actions that satisfies Habitat's success criterion (goal distance ≤0.2\leq0.2 m). Timeout/forced termination/internal SUCCESS do not pass. T- passes only if STOP is withheld throughout the invalid opportunity's window.

ACS=(CPCRCT+CT−)1/3.\mathrm{ACS}=\left(C_P C_R\sqrt{C_{T^+}C_{T^-}}\right)^{1/3}.

Persistence, release and termination have equal weight; T+/T- split termination. For nonempty event sets, any zero rate gives ACS=0. Any empty set gives component/overall N/A. No smoothing is used.

ModelPRT+T-
Base10/1431/4520/2411/101
Full23/2751/5111/44121/124

Base denotes Non-executive. T- improves from 11/101 to 121/124, raising the termination geometric mean from 0.3013 to 0.4939 despite lower T+. Together with better P/R, this raises ACS.

4.2. What these changes mean for navigation.

Accumulated evidence can sustain pursuit when the target briefly leaves view; verification counterevidence can release a candidate; STOP interception can prevent premature completion. These rules govern retention, release and stopping. Corresponding P/R/T- rates improve. On the same 150 tasks, SR rises from 13.33% to 19.33%, and SPL from 6.01% to 9.43%.

4.3. Why valid opportunities can still be missed.

Full misses 33 of 44 valid opportunities. Later episode success does not change those failures. Two control-flow situations can produce this mismatch:

(1) The scan is still pending.

The robot may already be at a valid stopping position, with a supporting observation, while the bounded scan and accumulated confirmation checks remain incomplete. If the response window expires first, the opportunity fails even if the target is confirmed later.

(2) The scan ends without sufficient confirmation.

Occlusion or a view-edge target can leave insufficient evidence after scanning. The controller may reposition for another observation or reselect after rejection, leaving the valid stopping position. Scan completion is therefore not the same as candidate confirmation.

The ACS gain reflects P/R and invalid-stop rejection outweighing the timely-stop loss under this aggregation; it does not imply uniformly improved stopping.

AuthorsOpenReview ↗

Response to Reviewer 9pVh (3/3): Responses to Q1–Q3

Q1. Which rules preserve tracking across viewpoint changes?

ConsistNav extends ApexNav. PCM stores identity and cross-frame evidence; FSEC uses these, verification and recovery outcomes to retain/release the same instance and authorize STOP.

MethodExisting designWhat ConsistNav adds
ApexNavVolume-weighted fusion, missed-detection handling, and action costs.Links evidence and pending verification to candidate-specific STOP authorization.
CogNavLLM state selection, verification and recovery.Defines guards for retention, release, and STOP.
UniGoalGraph matching, verification, and blacklisting.Recovery feedback guides subgoal revision or candidate release.

1.2. Decision rules.

PCM retains identity and cross-frame evidence. FSEC uses both for commitment, verification and recovery during loss of view or path obstruction. The table links rules to modules and states.

ConditionModuleDecision
Temporary missed detection / rank changePCM + FSECRetain memory/established identity; allow provisional SUSPECT release.
Valid candidate in verification range; scan incompleteFSEC (VERIFY)Keep VERIFY/pending until the bounded scan ends.
Supported target, blocked local pathFSEC (FAILOVER)Revise subgoal, retain identity; allow release after rejection/failed recovery.

Other navigation decisions (App. B.1–B.2)

Other components improve navigation decisions and task success through these functions.

ConditionModuleDecision
No active candidateFSEC (SEARCH)Explore frontiers and rank candidates; strong/persistent evidence or budget-triggered selection leads to SUSPECT.
Provisional candidateFSEC (SUSPECT)Reconfirmation/sustained top rank: APPROACH; invalidation/timeout: SEARCH.
Valid candidate, approach neededFSEC (APPROACH) + SAACSelect a reachable subgoal; penalize visited/failed regions. In range, enter VERIFY.
Spin limit exceeded with little translationSAACRequire minimum translation for the current or new subgoal.
Confirmed target, further approach neededFSEC (FINAL_APPROACH)Move closer; stale evidence: VERIFY; exhausted non-progress budget: FAILOVER.
Phase STOP gate satisfiedFSEC (SUCCESS) + SAACAuthorize STOP; otherwise intercept unexpected STOPs for recovery/heading correction.

Q2. Why use 0.12 in Eq. (5)?

2.1. Confidence update for matched observations:

c+=min⁡(1,c+α),c−=max⁡(0,c−λα),α=0.20.c^+=\min(1,c+\alpha),\quad c^-=\max(0,c-\lambda\alpha),\quad\alpha=0.20.

At λ=0.12\lambda=0.12, nominal updates +0.20/−0.024+0.20/-0.024 (ratio 8.338.33) slow confidence loss. Deduction requires prior target evidence; missed detection alone does not trigger it. λ=0\lambda=0 removes only this term; other negative evidence, verification rejection, decay and release remain active.

2.2. Four-coefficient comparison. We tested 0.24, 0.12, 0.06 and 0 on the same initial 20 HM3Dv2 tasks (80 runs); each used the Full controller with a 500-step budget.

λ\lambdaSR (%)SPL %
025.007.66
0.0635.0010.40
0.1240.0012.29
0.2435.0011.42

In this early panel, 0.12 has the highest SR and SPL of the four coefficients.

Q3. STOP sources and actions after interception

STOP sources

Outside SUCCESS, three ExplorationFSM routines can return STOP.

  1. Verification. Missing candidate/failed verification: enter FAILOVER, return STOP.
  2. Path action. Invalid subgoal/empty path: return STOP.
  3. Recovery. Disabled or out of budget: can return STOP.

Actions after interception

  1. Completion checks. Check completion first; if satisfied, enter SUCCESS and authorize STOP.
  2. Recovery. In APPROACH/VERIFY/FINAL_APPROACH, a SEARCH_OBJECT STOP enters FAILOVER and calls its handler; if already there, call it again.
  3. Fallback. Replace remaining STOP with TURN_LEFT; also replace premature EXPLORE STOPs.

After interception:

Candidate status and execution feedbackNext action
Valid, outside verification rangePursue it in APPROACH.
In range, scan incomplete or confirmation staleFinish or repeat VERIFY.
Supported, stalled or unreachable subgoalRecover/replan in FAILOVER; retain candidate if permitted.
Rejected, failed, or recovery exhaustedReselect; if none qualifies, resume SEARCH.

Summary and thanks

Thank you for the constructive feedback. ConsistNav adds an execution layer to ApexNav for consistent tracking and STOP decisions. Q1–Q3 clarify contributions, 0.12 and STOP processing;

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.