{"id":"cfef5f3a-359c-43a0-8458-a7b7643425fc","arxiv_id":"2412.18292","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MCoCoNav coordinates multiple zero-shot robots for object-goal navigation by combining multimodal chain-of-thought scoring with a shared semantic map.","lead":"This paper introduces MCoCoNav, a system that lets several home robots coordinate an object search using a shared map and a local vision-language model that reasons aloud before choosing where to go. It reports better search success than several prior multi-robot navigation methods on two indoor benchmarks, while using only a 9-billion-parameter model running locally rather than a cloud service.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 3 defines HFOVS as exp(τ_ES·ES + τ_JS·JS), which is always ≥ 1, so the 0.5 explore-vs-revisit threshold can never trigger; the central decision rule is mathematically vacuous as written.","rationale":"The reader's weakest assumption concerned calibration of the VLM probabilities and threshold sensitivity. That is a reasonable empirical worry, but the more fundamental problem is that Eq. 3 as written makes the 0.5 threshold unreachable. Since ES and JS are probabilities, exp(τ_ES·ES + τ_JS·JS) ≥ 1 always, so the 'revisit history node' branch can never execute. This is not a matter of external calibration or unknown scene distribution; it is an internal inconsistency in the formal definition of the planner. The entire history-score machinery and the ablations that credit it depend on the threshold being able to fire. Therefore the central claim that MCoCoNav outruns baselines via this explore-vs-revisit mechanism is not supported by the published mathematics. The public code could resolve the discrepancy, which is why the concrete test is to inspect it, but the manuscript itself is not self-consistent and requires major correction before the results can be trusted. Given the centrality of Eq. 3, I recommend REJECT rather than CONDITIONAL, with the clear path to resubmission being a corrected equation and re-validated experiments.","tokens_in":18314,"tokens_out":5017,"duration_ms":46183,"concrete_test":"Inspect the released implementation (github.com/FrankZxShen/MCoCoNav) for the actual HFOVS computation. Run a unit check with ES=0.0, JS=0.0 and ES=1.0, JS=1.0 under Eq. 3 to confirm that the output is always ≥ 1. If the code instead uses a sigmoid or exp(-(...)) form, verify which formula was used to produce Tables 1-3 and require the paper to state the correct equation; if the code matches Eq. 3, demonstrate whether HFOVS < 0.5 ever occurs in practice — if not, the exploration/revisit decision is vacuous and the reported results must be re-evaluated.","verdict_should_be":"REJECT","load_bearing_attack":"The central decision rule in Eq. 3 is internally inconsistent. ES and JS are probabilities of the VLM answering 'Yes', so ES, JS ∈ [0,1]. With τ_ES=2.0 and τ_JS=1.0 from Table 4, the weighted sum τ_ES·ES + τ_JS·JS lies in [0,3], and exp(...) therefore always lies in [1, e^3 ≈ 20.1]. The paper then states: 'If HFOVS is greater than or equal to 0.5, robot ri continues to select frontier points; otherwise, robot ri returns to history nodes.' Since HFOVS ≥ 1 > 0.5 under every possible input, the 'otherwise' branch is unreachable. This makes the history-return mechanism dead code, and the reported ablation for the Decision/History modules cannot be explained by the published equations. The issue is not a calibration or threshold-tuning matter; it is a hard mathematical contradiction that prevents the method from working as described. The code is publicly available, so the discrepancy is checkable, but as the paper stands the central algorithmic mechanism is not well-defined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes MCoCoNav, a decentralized multi-robot object-goal navigation system built around a local quantized 9B VLM (GLM-4V-9B). A Perception module obtains an exploration score ES as the VLM's probability of \"Yes\" for a scene view; a Judgment module obtains JS from the global top-down semantic map annotated with history nodes and frontiers; Eq. (3) combines these into a horizontal field-of-view score HFOVS, thresholded at 0.5 to choose between exploring frontier points and returning to history nodes; a Decision module selects one of A-D frontier points; a Logical Analysis module handles continuation and collision traps; FMM provides local actions. Experiments on HM3D v0.2 and MP3D report SPL/SR improvements over Greedy, Cost-Utility, Random Sampling, Multi-SemExp, and Co-NavGPT, plus ablations and robot-number/semantic-accuracy studies.","tokens_in":18528,"tokens_out":6362,"duration_ms":56369,"significance":"If substantiated, the main result is notable: a local, quantized 9B VLM with a shared semantic map beats a remote GPT-3.5 centralized planner (Co-NavGPT) on HM3D v0.2 (SPL 0.387 vs 0.331; SR 0.716 vs 0.661), with code released and no task-specific training. The cross-image CoT prompting and map-as-communication-bridge design are interesting, and the modular ablation story is coherent. However, the published decision rule is mathematically vacuous as written, the reported numbers lack variance and statistical support, and the communication-cost claim is not measured; the contribution is therefore conditional on correcting and re-validating the central mechanism.","major_comments":[{"comment":"Because ES and JS are next-token probabilities in [0,1] and τ_ES=2.0, τ_JS=1.0, HFOVS = exp(τ_ES·ES + τ_JS·JS) always lies in [1, e^3] ≈ [1,20.1]. The decision rule \"If HFOVS ≥ 0.5 ... otherwise return to history nodes\" therefore always takes the frontier-exploration branch; the history-return branch and the Logical Analysis condition \"HFOVS < 0.5\" are unreachable. Since the Decision/History switching is the mechanism that the ablations in Table 3 are meant to validate, the paper as written does not specify the algorithm it evaluates. Please correct Eq. (3) or the threshold (or both) to match the code and report whether the history-return branch fires in any episode.","section":"Eq. (3), §Method (Global Map Exploration Judgment), Table 4"},{"comment":"τ_ES=2.0, τ_JS=1.0, δ=25, and the 0.5 threshold are fixed constants, but the reported HM3D v0.2 results are on the same validation split used to choose them; no sensitivity analysis or hold-out split is provided. Because raw VLM probabilities are not calibrated, the absolute value of ES/JS is prompt- and scene-dependent, so the claimed zero-shot gains may partly reflect fitted constants. Please report validation/test separation and a sweep over τ_ES, τ_JS, and the threshold.","section":"Table 4, §Experimental Setup"},{"comment":"All metrics are reported as single numbers with no error bars, number of seeds, or significance tests, yet the Table 1 caption states MCoCoNav \"significantly outperforms all baseline methods.\" Given the small margins over Random Sampling on MP3D and over Co-NavGPT on HM3D, please provide mean±std across multiple runs (at least 3 seeds) and/or paired tests per episode.","section":"Table 1, Table 3"},{"comment":"The paper asserts that \"physical time for communication between the robots is negligible\" without measuring it; the experiments run in a single Habitat process, so no inter-robot communication bandwidth or latency is modeled. The introduction and abstract claim reduced communication overhead, but the global map \"bridge\" is an architectural choice whose cost is never quantified. Please specify the communication model and provide measurements or a bound.","section":"§B.2 Computational Resources"},{"comment":"Co-NavGPT, the strongest centralized LLM baseline, is reported only on HM3D v0.2; the MP3D comparison therefore does not include the main competing method. The claim that MCoCoNav outperforms \"all baseline methods on all metrics\" across both datasets is not supported by the table as presented; please run the strongest baselines on MP3D or qualify the claim.","section":"Table 1"}],"minor_comments":[{"comment":"DT (Detection Trap) is used as a metric but is never defined in the experimental setup; please define it and explain how it is counted.","section":"Table 2 and §Metrics"},{"comment":"EXP is used as a denominator (line 21) but is never initialized or updated in the pseudocode, so the algorithm is underspecified for the first visit to a location.","section":"Appendix A.2, Algorithm 1"},{"comment":"The term \"temperature scaling\" is nonstandard here: taking exp of a probability-weighted sum does not implement softmax temperature scaling; please clarify the intended transformation.","section":"Eq. (3)"},{"comment":"The checkmark rows are ambiguous because the column alignment makes some rows identical in appearance; please use explicit module names or a legend for each ablation row.","section":"Table 3"},{"comment":"The success-distance criterion differs across single-robot baselines (Pixel-Nav uses 1.0m while others use 0.2m); please make the comparison protocol explicit.","section":"Table 6 and §B.3"},{"comment":"The failure-mode percentages are reported without denominators or error bars, making it hard to assess the reliability of the failure-mode analysis.","section":"Figures 5 and 6"},{"comment":"There are copyediting issues such as \"V oronav\" in the Related Work section and \"Jug\" in the Table 3 caption; please revise.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"The Eq. (3) issue should be verified against the public code before acceptance: if the code contains a different formula, the authors need to fix the paper and re-run the ablations; if the code matches Eq. (3), the reported Decision/History ablation is not explained by the published mechanism. I also note that the \"zero-shot\" claim should be qualified because the thresholds are tuned on the benchmark, and the communication-cost claim is asserted rather than measured."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a realistically described multi-robot ObjectNav system that uses a local 9B VLM, a shared semantic map, and chain-of-thought scoring. The empirical gains over Co-NavGPT and the classical baselines are credible, and the appendix gives enough detail to reimplement. But there is a hard mathematical problem in the core decision rule that the authors need to fix before the paper says what it means.\n\nThe nub: Eq. 3 defines HFOVS as exp(τ_ES·ES + τ_JS·JS), where ES and JS are VLM \"Yes\" probabilities in [0,1]. With the τ values from Table 4, the exponent is in [0,3], so HFOVS is always at least 1. The paper then says: if HFOVS ≥ 0.5, explore frontiers; otherwise, revisit a history node. Since HFOVS ≥ 1 > 0.5 always, the revisit branch is unreachable. The history-score mechanism and the \"balance between exploration and return\" contribution are dead code under the published equations. The stress-test note is correct.\n\nWhat the paper does well: the cross-image multimodal CoT, feeding both the scene view and the annotated global map into the VLM, is a genuinely new combination. The three-module decomposition is clean, and the ablations show each module adds something. The authors are honest about detection failures, and they include a full prompt template and a public code link. The claim that a local quantized VLM can beat a remote GPT-3.5 planner in this setting is interesting and worth checking.\n\nThe soft spots beyond the equation are about evidence quality: there are no error bars or significance tests, the communication-cost advantage is overstated because map merging itself is communication, and the temperature/threshold values appear fixed on the same validation benchmarks. Those are secondary. The primary issue is that the published decision rule contradicts itself. It is possible the code implements something different – for example, the Judgment module's text may still influence the Decision VLM through the chain-of-thought, which would explain the ablation numbers – but the paper does not describe that, and as written the revisit logic cannot fire.\n\nWho this is for: anyone building VLM-driven multi-robot navigation will find the prompting scheme and system design useful. The paper deserves a serious referee, because the idea is good and the empirical setup is real, but it needs major revision: correct or clarify Eq. 3, show the actual code path for revisits, add multi-run statistics, and soften the communication claim. I would not cite it until those are resolved.","headline":"The reported system is plausible and the paper is well-specified, but Eq. 3 makes the history-return decision impossible as written, so the central contribution is currently not well-defined.","tokens_in":19084,"tokens_out":3101,"would_cite":false,"duration_ms":28119,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A local 9B vision-language model with a shared semantic map can beat a remote GPT-3.5 centralized planner in multi-robot object search.","keywords":["multi-robot semantic navigation","object goal navigation","zero-shot navigation","vision-language model","chain-of-thought prompting","semantic mapping","frontier-based exploration","decentralized planning"],"falsifier":"Re-run the two-robot HM3D v0.2 evaluation with the exact hyperparameters of Table 4 on scenes never used for prompt or threshold selection, and compare SPL and success rate against Co-NavGPT and Random Sampling; if the margin collapses to within episode-level noise, the zero-shot transfer claim fails. A second check is to record the distributions of 'Yes' token probabilities in successful versus failed episodes: if the two distributions do not separate around 0.5, the score is not behaving as an exploration-value estimate.","tokens_in":18077,"feed_emoji":"🤖","tokens_out":8078,"duration_ms":63681,"temperature":0.7,"pith_summary":"The paper proposes MCoCoNav, a decentralized planning framework in which several robots searching for a target object in an unseen home share a single top-down semantic map and use a locally deployed vision-language model to score how promising each view and each candidate frontier is. The central claim is that this modular, zero-shot design beats both a centralized GPT-3.5 planner (Co-NavGPT) and trained exploration baselines on the HM3D v0.2 and MP3D benchmarks, while keeping communication cost low because the shared map is the only message exchanged. On the two-robot HM3D v0.2 evaluation with predicted segmentation, MCoCoNav reports a success-weighted path length (SPL) of 0.387 and a success rate of 0.716, compared with Co-NavGPT's 0.331 and 0.661. If the claim holds, it means a 9-billion-parameter quantized local model can replace a remote large model for multi-robot object search without losing performance.","feed_headline":"Local 9B VLM beats remote GPT-3.5 in multi-robot search","feed_subtitle":"Shared semantic map plus chain-of-thought scoring lifts HM3D SPL from 0.331 to 0.387 over Co-NavGPT.","key_machinery":"The load-bearing object is the cross-image multimodal chain-of-thought: three VLM calls (Perception, Judgment, Decision) that consume both the robot's first-person image and the annotated shared semantic map, with each call's text output becoming part of the next prompt. The navigation decision is carried by the VLM's token probability of 'Yes' (Eqs. 1 and 2), temperature-scaled with $\\tau_{ES}=2.0$ and $\\tau_{JS}=1.0$ and combined into the horizontal-field-of-view score of Eq. 3, whose 0.5 threshold routes each robot either to explore a frontier or to revisit a history node. A history-score algorithm accumulates these scores into 360 directional bins and normalizes by exploration count, so the global semantic map acts simultaneously as the communication channel, the scoreboard, and the source of the visual prompt for the Decision VLM.","core_discovery":"The paper's discovery claim is that the next-token 'Yes' probability of a local vision-language model, carried through a chain of three prompts, can act as a stable exploration-value score for multi-robot navigation. The Perception module turns a first-person RGB view into a spatial-relationship description; the Judgment module reads an annotated global semantic map and produces a judgment score; the two are temperature-scaled and combined into a horizontal-field-of-view score $HFOVS=\\exp(\\tau_{ES}ES+\\tau_{JS}JS)$ (Eq. 3), and a 0.5 threshold decides whether the robot continues to frontier points or returns to the highest-scoring history node. The Decision module then selects among labeled frontier points by normalized token probabilities over 'A', 'B', 'C', 'D'. With two robots and predicted segmentation, this procedure reports SPL 0.387 and SR 0.716 on HM3D v0.2 against Co-NavGPT's 0.331 and 0.661, and SPL 0.334 and SR 0.568 on MP3D. The paper argues that this specific combination of perception, judgment, and decision scores is what makes zero-shot local multi-robot semantic navigation work.","pith_inferences":["Editorial extension: Replacing the VLM's 'Yes' probabilities with a fixed or random score while keeping the map, history-node algorithm, and frontier selection would test whether the gains come from VLM semantics or from the cost-geometry of the shared map; the paper's ablations remove whole modules but never replace the score with a non-semantic baseline.","Editorial extension: The 0.5 threshold and temperature scales are constants tuned on the same benchmark family, so the system's zero-shot claim would be stronger if the threshold were shown to be stable across scene types, for instance from bedroom-dominated episodes to kitchen-dominated episodes.","Editorial extension: The cross-image chain-of-thought pattern, one VLM describing a local view, another reading an annotated global map, and a third choosing among labeled candidates, could transfer to other multi-agent tasks such as search-and-rescue exploration where shared map scores and local language models matter."],"forward_implications":["Multi-robot object-goal navigation in unseen homes can be done entirely from local inference, without task-specific reinforcement learning and without remote foundation-model calls.","Communication cost drops to the cost of maintaining one shared semantic map, because the map carries all the information robots exchange.","Adding more robots continues to raise success rate and SPL, but with diminishing returns and rising detection traps, so two robots give the best efficiency trade-off on HM3D v0.2.","With ground-truth semantic segmentation, the reported success rate for two robots on HM3D v0.2 reaches 0.872, implying that detector and segmentation quality, rather than VLM reasoning, is the current bottleneck.","Failure analysis attributes most remaining failures to detection rather than planning or exploration, so better object detection and 3D scan quality would improve the method more than further prompt engineering."],"supporting_citations":[{"why":"Provides the HM3D v0.2 dataset and its 1000 validation episodes on which the main two-robot results are measured.","marker":"Ramakrishnan et al. 2021"},{"why":"Supplies the Matterport3D scenes used for the MP3D evaluation, testing cross-dataset transfer.","marker":"Chang et al. 2017"},{"why":"Defines the Habitat simulator, including the observations, discrete action space, and success conditions used in all experiments.","marker":"Savva et al. 2019"},{"why":"Introduces Co-NavGPT, the centralized remote-GPT-3.5 baseline that MCoCoNav claims to outperform.","marker":"Yu, Kasaei, and Cao 2023a"},{"why":"Provides SemExp, the trained semantic-exploration approach extended into the Multi-SemExp multi-robot baseline.","marker":"Chaplot et al. 2020"},{"why":"Provides the GLM-4V-9B model that is locally deployed, quantized, and used for every VLM score and decision in MCoCoNav.","marker":"GLM et al. 2024"},{"why":"YOLOv10 is the detection model that converts visual object labels and probabilities into natural-language input for the Perception module.","marker":"Wang et al. 2024"},{"why":"RedNet performs the semantic segmentation used to build the semantic maps that the Judgment and Decision modules read.","marker":"Jiang et al. 2018"},{"why":"The Fast Marching Method is the local planning policy that turns each long-term goal into low-level robot actions.","marker":"Sethian 1999"}],"fun_headline_variants":["Multi-robot nav boosted by chain-of-thought scoring","Local VLM scores guide multi-robot navigation","Chain-of-thought collab improves multi-robot search","MCoCoNav: shared map and CoT scores for robots"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a VLM's next-token probability of 'Yes' is a stable, well-calibrated measure of exploration value across different scenes, so the fixed 0.5 threshold and the temperature scales $\\tau_{ES}=2.0$, $\\tau_{JS}=1.0$ transfer to new environments without per-scene retuning.","fun_headline_variants_meta":{"raw":{"variants":["Multi-robot nav boosted by chain-of-thought scoring","Local VLM scores guide multi-robot navigation","Chain-of-thought collab improves multi-robot search","MCoCoNav: shared map and CoT scores for robots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000483,"raw_usage":{"total_tokens":2422,"prompt_tokens":1021,"completion_tokens":1401,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":1334}},"tokens_in":637,"tokens_out":1401,"duration_ms":9689,"temperature":1.0,"reasoning_tokens":1334,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:49:12.954182+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the two-robot HM3D v0.2 evaluation with the exact hyperparameters of Table 4 on scenes never used for prompt or threshold selection, and compare SPL and success rate against Co-NavGPT and Random Sampling; if the margin collapses to within episode-level noise, the zero-shot transfer claim fails. A second check is to record the distributions of 'Yes' token probabilities in successful versus failed episodes: if the two distributions do not separate around 0.5, the score is not behaving as an exploration-value estimate.","supporting_citations":[],"review_version":1}