{"id":"45cd3f60-04fb-4975-b5d5-5ed153915898","arxiv_id":"2605.09869","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"ConsistNav is a new training-free framework that uses a semantic executive controller, persistent candidate memory, and stability-aware action control to close the action consistency gap in zero-shot object navigation, reporting SOTA results on HM3D and MP3D with 11.4% SR and 7.9% SPL gains on MP3D.","lead":"ConsistNav adds a training-free semantic executive with finite-state phases, persistent memory, and stability controls to make zero-shot object navigation more consistent. Smart readers should note it for practical robot deployment in homes or offices where retraining is costly.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Improvements may stem from auxiliary effects of memory and stability controls rather than specifically closing the action consistency gap","rationale":"This directly matches the reader's weakest assumption that the gap is dominant and closable by the modules without side effects. Full-text experiments and ablations would need to include the proposed oscillation diagnostics to move beyond indirect performance gains. No other internal inconsistency (e.g., in zero-shot framing or dataset choice) appears load-bearing given the training-free design and reported SOTA comparisons.","tokens_in":1731,"tokens_out":378,"duration_ms":40796,"concrete_test":"On the MP3D validation episodes, compute the per-episode oscillation count (number of explore-to-pursue or pursue-to-explore transitions) and abandoned-candidate rate for both the controlled baseline and ConsistNav; if the reduction in these metrics explains >70% of the SR variance via regression and the gains vanish when oscillations are artificially suppressed in the baseline, the gap-closure attribution holds.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that the action consistency gap (repeated reinterpretation of semantic evidence without persistent commitment) is the dominant failure mode and that the three modules close it without introducing new exploration or stopping failures. The Persistent Candidate Memory and Stability-Aware Action Control could instead improve results via longer evidence accumulation and reduced rotational waste, independent of enforcing cross-step commitment. The Finite-State Executive Controller adds phase gating, but without episode-level diagnostics (e.g., counts of explore/pursue switches or abandoned hypotheses before/after the modules), attribution of the 11.4% SR and 7.9% SPL gains to gap closure remains indirect. The design leaves detector and low-level planner unchanged, yet the executive effectively inserts a higher-level policy whose benefits might be replicable by simpler heuristics.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper identifies an 'action consistency gap' in zero-shot object navigation, where agents repeatedly reinterpret semantic evidence without persistent commitment, leading to oscillation between exploration and pursuit or premature abandonment of targets. It introduces ConsistNav, a training-free framework with a semantic executive consisting of three modules: a Finite-State Executive Controller that stages target pursuit through guarded phases, Persistent Candidate Memory that accumulates cross-frame evidence into stable hypotheses, and Stability-Aware Action Control that suppresses rotational stagnation and unverified stopping. The approach leaves the detector and low-level planner unchanged. Experiments on HM3D and MP3D report state-of-the-art results among zero-shot methods, with 11.4% higher success rate (SR) and 7.9% higher SPL over a controlled baseline on MP3D, plus supporting ablations and real-world deployment.","tokens_in":1912,"tokens_out":658,"duration_ms":35125,"significance":"If the central claim holds, the work would offer a modular, training-free method to improve consistency in open-vocabulary navigation without retraining core perception or planning components. The explicit separation of executive control from the detector/planner, combined with real-world validation, strengthens potential for broader adoption in embodied AI. The identification of a specific failure mode and the provision of ablations are positive elements.","major_comments":[{"comment":"§4 (Experiments) and associated tables: The reported 11.4% SR and 7.9% SPL gains on MP3D are presented as evidence that the modules close the action consistency gap, but the manuscript provides no episode-level diagnostics such as counts of explore/pursue switches, abandoned hypotheses, or rotational stagnation events before versus after adding the executive. Without these, it remains possible that gains arise from auxiliary effects of memory accumulation and stability filtering rather than enforced cross-step commitment, weakening attribution to the identified gap.","section":"§4 (Experiments)"},{"comment":"§3 (Method), description of the three modules: The Finite-State Executive Controller and Persistent Candidate Memory are presented as directly addressing reinterpretation without commitment, yet the design inserts a higher-level policy layer. A direct comparison to simpler non-executive heuristics (e.g., fixed hysteresis thresholds on detection confidence) would be needed to establish that the full three-module coordination is necessary for the observed gains rather than replicable by lighter mechanisms.","section":"§3 (Method)"}],"minor_comments":[{"comment":"Abstract and §1: The phrase 'guarded semantic phases' is introduced without a concise definition or diagram reference at first mention; a brief inline clarification or pointer to Figure 2 would improve readability.","section":"Abstract"},{"comment":"§4.3 (Ablations): The ablation table would benefit from explicit reporting of standard deviations or confidence intervals across the N runs, consistent with the main result tables.","section":"§4.3 (Ablations)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits well within robotics and embodied AI venues; citation of prior zero-shot ObjectNav work appears balanced. No obvious scope mismatch."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which help clarify the attribution of our results to the action consistency gap. We respond to each major comment below and indicate where revisions will be made.","responses":[{"response":"We agree that explicit episode-level diagnostics would strengthen direct attribution to reduced oscillation and premature abandonment. The current ablations isolate module contributions and the overall SR/SPL gains align with fewer consistency failures, but without per-episode switch counts the link remains indirect. In the revised version we will add these diagnostics, reporting average explore/pursue transitions, abandoned hypotheses, and rotational stagnation events for the baseline versus ConsistNav on MP3D.","revision_made":"yes","referee_comment":"[§4 (Experiments)] §4 (Experiments) and associated tables: The reported 11.4% SR and 7.9% SPL gains on MP3D are presented as evidence that the modules close the action consistency gap, but the manuscript provides no episode-level diagnostics such as counts of explore/pursue switches, abandoned hypotheses, or rotational stagnation events before versus after adding the executive. Without these, it remains possible that gains arise from auxiliary effects of memory accumulation and stability filtering rather than enforced cross-step commitment, weakening attribution to the identified gap."},{"response":"The three modules are coordinated: the finite-state controller stages commitment, memory accumulates evidence across frames, and stability control suppresses ineffective actions. A simple hysteresis threshold on confidence would address only part of the reinterpretation problem and would not stage pursuit phases or suppress rotational stagnation. Our module ablations already show that removing any component degrades performance. Nevertheless, to address the request we will add a controlled comparison against a hysteresis-only variant in the revised experiments.","revision_made":"yes","referee_comment":"[§3 (Method)] §3 (Method), description of the three modules: The Finite-State Executive Controller and Persistent Candidate Memory are presented as directly addressing reinterpretation without commitment, yet the design inserts a higher-level policy layer. A direct comparison to simpler non-executive heuristics (e.g., fixed hysteresis thresholds on detection confidence) would be needed to establish that the full three-module coordination is necessary for the observed gains rather than replicable by lighter mechanisms."}],"tokens_in":1559,"tokens_out":483,"duration_ms":22135,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"ConsistNav layers a semantic executive on top of standard zero-shot object navigation stacks. The core idea is to stop the agent from constantly reinterpreting what it sees and instead commit to a target once evidence builds up. They do this with a finite-state controller that moves through guarded phases, a memory that keeps candidate objects alive across frames, and action rules that cut rotational waste and bad stops. This setup is new in its coordinated packaging. Most prior work either tweaks the detector or the exploration policy. Here the base pieces stay fixed and the executive just decides when semantics should drive the planner. That keeps things training-free and easy to drop in. The reported numbers on MP3D show an 11.4 point success rate gain and 7.9 point SPL gain over the controlled baseline, with similar patterns on HM3D. Real-world tests add some credibility that the approach survives outside simulation. The main question is whether the gains truly come from closing the action consistency gap. The memory and stability controls might simply let the agent gather more evidence and move more efficiently, regardless of whether it sticks to one hypothesis. Without counts of how often the agent switches from explore to pursue before and after the modules, or how many times it abandons a near-miss target, the attribution stays indirect. The paper does not appear to introduce new failure modes, which is good, but the central claim would be stronger with those episode-level diagnostics. This paper is aimed at robotics researchers who want practical improvements to zero-shot navigation without retraining models. Anyone working on home robots or warehouse pickers could get value from the executive design and the ablation results. It is solid enough to deserve a serious referee, especially since it includes both simulation benchmarks and real deployment. I would recommend sending it to peer review. The experiments are extensive enough to warrant detailed feedback, though the reviewers will likely push for clearer evidence on the mechanism.","headline":"ConsistNav adds a practical executive layer that boosts zero-shot navigation performance, though the link to closing the specific action consistency gap could use more direct diagnostics.","tokens_in":2423,"tokens_out":458,"would_cite":false,"duration_ms":46851,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"Navigation executive with finite-state commitment and persistent memory operates in applied robotics domain with no RS overlap","alignment":"orthogonal","rationale":"Paper introduces Finite-State Executive Controller (states SEARCH/SUSPECT/APPROACH/VERIFY etc.), Persistent Candidate Memory, and Stability-Aware Action Control to enforce cross-step semantic commitment and close oscillation/abandonment failures. These structures address practical evidence-to-action consistency in zero-shot ObjectNav but invoke none of the RS core machinery (J-cost, φ-ladder, 8-tick periodicity, ratio-symmetric forcing, or parameter-free constant derivations). RS modules such as ArrowOfTime, Atomicity, and Breath1024 contain related notions of temporal ordering, serialization, and periodic micro-structure, yet the paper neither cites nor parallels any specific RS theorem and remains confined to cs.RO engineering. Hence orthogonal.","tokens_in":49056,"confidence":"high","tokens_out":191,"duration_ms":22619,"cache_read_input_tokens":32896,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A semantic executive with three coordinated modules closes the action consistency gap by enforcing persistent commitment to target pursuit in zero-shot object navigation.","keywords":["zero-shot object navigation","action consistency gap","semantic executive","persistent memory","robot navigation","embodied AI","finite-state control","training-free navigation"],"falsifier":"Compare oscillation frequency and abandonment rate between the baseline and ConsistNav in identical MP3D episodes and check whether the executive modules produce a clear drop in switches between exploration and pursuit while success rate rises.","tokens_in":2649,"feed_emoji":"🤖","tokens_out":780,"duration_ms":32584,"temperature":0.7,"pith_summary":"The paper points out that even when zero-shot methods spot a plausible target, agents often switch back and forth between exploring and pursuing or quit near the goal because semantic evidence gets reinterpreted fresh at every step. ConsistNav adds a training-free semantic executive layer on top of existing detectors and planners to stage pursuit in guarded phases, keep stable object hypotheses across frames, and block wasteful actions such as spinning in place. A reader would care because this keeps the agent from abandoning a found object and raises success without any retraining or changes to the underlying perception and planning code. Experiments on HM3D and MP3D show the approach reaches state-of-the-art numbers among compared zero-shot methods and lifts success rate by 11.4 percent and SPL by 7.9 percent over a controlled baseline on MP3D. Ablations and real-robot tests confirm the executive modules are what drive the gains.","feed_headline":"Semantic executive closes action consistency gap in robot navigation","feed_subtitle":"Three modules enforce persistent target pursuit, raising success rate 11.4 percent on MP3D without retraining detectors or planners.","key_machinery":"Semantic executive, a training-free coordinator that decides when semantic evidence should drive navigation and when it should be suppressed or revisited through its three modules.","core_discovery":"The paper claims that the action consistency gap—repeated reinterpretation of semantic evidence without persistent commitment across the episode—explains why agents oscillate or abandon targets near success, and that this gap can be closed by a semantic executive composed of a Finite-State Executive Controller that stages guarded pursuit phases, a Persistent Candidate Memory that accumulates cross-frame target evidence into stable hypotheses, and Stability-Aware Action Control that suppresses rotational stagnation and unverified stopping, all without modifying the detector or low-level planner.","pith_inferences":["The same executive structure could be tested on other embodied tasks where agents must commit to a detected goal over time, such as object manipulation sequences.","Because the modules act after detection, they might combine with newer open-vocabulary detectors without retraining the consistency logic.","If the gap is truly central, similar executive controls could be added to language-guided exploration methods to reduce backtracking.","The approach leaves open whether the same gains appear when the underlying planner itself is also improved."],"forward_implications":["Agents maintain stable object hypotheses across multiple frames instead of reinterpreting evidence at each step.","Pursuit is staged through guarded semantic phases that prevent premature abandonment of detected targets.","Rotational stagnation and ineffective pursuit actions are suppressed while still allowing verified stopping.","The same detector and planner can be used with higher reliability simply by adding the executive layer.","The method transfers to real-world robot deployments without additional training."],"fun_headline_variants":["Semantic executive closes action consistency gap in ObjectNav","Coordinated modules enforce persistent target pursuit","Stability-aware control suppresses rotational stagnation","Persistent memory accumulates cross-frame target evidence"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The action consistency gap is the dominant failure mode in current zero-shot object navigation and the three executive modules can close it without creating new exploration failures or requiring detector or planner changes.","fun_headline_variants_meta":{"raw":{"variants":["Semantic executive closes action consistency gap in ObjectNav","Coordinated modules enforce persistent target pursuit","Stability-aware control suppresses rotational stagnation","Persistent memory accumulates cross-frame target evidence"]},"model":"grok-4.3","cost_usd":0.011848,"raw_usage":{"total_tokens":5120,"prompt_tokens":708,"num_sources_used":0,"completion_tokens":50,"cost_in_usd_ticks":118478000,"prompt_tokens_details":{"text_tokens":708,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":4362,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":708,"tokens_out":50,"duration_ms":72893,"temperature":1.0,"reasoning_tokens":4362,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-19T18:02:15.884466+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Compare oscillation frequency and abandonment rate between the baseline and ConsistNav in identical MP3D episodes and check whether the executive modules produce a clear drop in switches between exploration and pursuit while success rate rises.","supporting_citations":[],"review_version":2}