{"id":"db240bf1-fe18-4fa0-a0fc-08a4486c0d2d","arxiv_id":"2608.02571","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Adding rescue-relevance, observation-deficit, terrain, and travel terms to frontier ranking improves completion and victim recovery for a simulated quadruped in cluttered indoor search and rescue.","lead":"Search-and-rescue robots normally explore by expanding maps, not by hunting for victims. This paper adds a victim-relevance score to the usual frontier-selection algorithm and tests it on a simulated quadruped, reporting better completion and victim recovery in a cluttered indoor scenario.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The S2 advantage may be an artifact of tentative victim cues acting as near-ground-truth victim locations; without specifying the cue detector or giving baselines the same cue stream, the comparison is not yet fair.","rationale":"The reader flagged the same weakest assumption: R(f) depends on tentatively detected victim cues whose generation is underspecified, creating a possible privileged-information artifact. My stress-test agrees that this is the most load-bearing threat to the central claim. The numerical result in Table 1—20/20 vs. 19/20 and 2.00 vs. 1.95—is a small difference, and the absence of statistical tests compounds the issue, but the cue-pipeline concern is more fundamental: if V^⊔ is effectively ground truth, the entire comparison is unfair regardless of number of runs. However, the paper's description is not conclusive. The cues could be legitimate onboard detections that become available once a victim is in view, and using them for exploitation is a sensible rescue behavior; in that reading the method is not circular. Because the matter is empirically checkable and the existing CONDITIONAL verdict already calls for clarification, I do not recommend changing the verdict. The concrete test would settle whether the concern actually lands.","tokens_in":6775,"tokens_out":6125,"duration_ms":58890,"concrete_test":"Download the repository (github.com/ricardoGrando/go2_rescue_eval) and trace the source of the tentative victim topic: if it is /gazebo/model_states or a ground-truth subscriber, the concern lands. Then run two ablations in S2: (a) disable R(f) (w_R=0) and confirm whether the method degrades to baseline-level completion/recovery; (b) feed the same tentative-cue stream to a baseline (e.g., information gain plus a shortest-path-to-nearest-unconfirmed-cue behavior) without changing its exploration objective. If the baseline with cue access matches or beats SA's 20/20, the advantage comes from having the cues, not from the frontier prioritization. If SA still outperforms with a noisy cue detector (false positives and false negatives), the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that situation-aware frontier prioritization is beneficial in cluttered S2—rests on R(f) in Eq. (3), which sums Gaussian contributions from tentative victim observations V^⊔. Section 3.2.3 describes only a 'lightweight visual confirmation process' and does not report detection range, false-positive/false-negative rates, or whether the baseline planners receive the same tentative-cue stream. In the released Gazebo setup, if V^⊔ is populated by direct reads of victim proxy states (or by a perfect detector), then the proposed method's score contains near-ground-truth victim positions, while baselines optimize only geometry/information. The 20/20 completion and 2.00 victims/run in Table 1 would then demonstrate 'go toward known cue locations,' not generalizable rescue-aware frontier prioritization. This is load-bearing because the paper's stated contribution is precisely the balancing of rescue utility against exploration; if the rescue utility term is oracle-fed, the comparison is not evidence for that contribution. The concern is not that the math is wrong but that the experimental protocol does not establish that R(f) is computed from the same kind of uncertain, false-positive-prone evidence a real quadruped would have.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a situation-aware frontier prioritization method for single-robot quadruped search and rescue. The method preserves the classical frontier-exploration framework but adds terms for information gain, observation deficit, rescue relevance, terrain penalty, and travel cost to a frontier score J(f) defined in Eq. (1). Rescue relevance is computed in Eq. (3) from tentative victim observations. The method is evaluated in Gazebo with a Unitree Go2 in two indoor scenarios: an easy sanity-check scenario (S1) and a cluttered, branching scenario (S2), compared against nearest frontier, information gain, and risk-aware baselines. The main reported result is that in S2 the proposed method achieves the highest completion rate (20/20) and highest victim recovery (2.00 victims/run), suggesting that rescue-aware frontier selection is beneficial when frontier choice is nontrivial.","tokens_in":7084,"tokens_out":3752,"duration_ms":35530,"significance":"If validated, the paper addresses a real gap between generic exploration and rescue-oriented decision-making for quadrupedal SAR, and the open-source repository plus shared execution layer are useful contributions. The evaluation design, with two scenarios of differing difficulty and the same low-level controller across methods, is a reasonable starting point. However, the headline result rests on a very small performance margin and on an insufficiently specified victim-cue pipeline; the current evidence does not yet establish that the proposed method is superior to the risk-aware baseline, nor that the rescue-relevance term is computed from realistic, non-privileged sensor information. Properly addressing these points would make the contribution meaningful.","major_comments":[{"comment":"The central S2 claim is supported by a difference of one run in completion rate (20/20 vs. 19/20 for Risk-aware Exploration) and 0.05 victims/run (2.00 vs. 1.95). The paper reports no confidence intervals, statistical tests, or error bars on mission time or path length. With n=20, 20/20 vs. 19/20 is not statistically distinguishable (e.g., Fisher's exact test p is about 1.0). The conclusion that the proposed method achieves the 'highest' completion rate and victim recovery is therefore not supported by the reported evidence. Please report effect sizes, confidence intervals, and appropriate tests, or explicitly qualify the claim as a non-significant trend.","section":"Section 4, Table 1"},{"comment":"The rescue-relevance term R(f) depends on the set of tentative victim observations V^⊔, but the paper never specifies how these observations are generated. It only says a 'lightweight visual confirmation process' uses 'visual cues corresponding to victim proxies.' No detection range, false-positive/false-negative rates, or sensor model are given, and it is not stated whether the baseline methods receive the same tentative-cue stream. If V^⊔ is populated directly from ground-truth victim locations in simulation, the proposed method is given privileged information, and the S2 result demonstrates moving toward known cue locations rather than rescue-aware exploration under uncertainty. Please specify the cue-detection pipeline, quantify its reliability, and ensure baselines either receive the same cues or an ablation without R(f) is reported.","section":"Section 3.2.3 and Eq. (3)"},{"comment":"The weights w_I, w_O, w_R, w_T, w_D and the kernel width sigma_R in Eq. (3) are never reported, nor is the tuning procedure. Without these values, the method cannot be reproduced, and it is impossible to assess whether the results are sensitive to hand-tuned parameters or overfit to the two scenarios. Please provide the parameter values, the selection procedure, and a sensitivity analysis over at least the key weights and sigma_R.","section":"Section 3.3, Eq. (1)"}],"minor_comments":[{"comment":"The figure numbering appears inconsistent: the bar chart labeled 'Figure 3. Completion rate and victim recovery in S2' is referenced in the text as 'Fig. 2,' and the path/mission-time chart is referenced as 'Fig. 3.' Please renumber or correct the in-text references.","section":"Section 4/Figures"},{"comment":"In S2, the first-victim time of the proposed method (349.4 s) is worse than Nearest Frontier (326.4 s) and Risk-aware Exploration (343.7 s). Since rescue speed is a relevant metric, this should be discussed rather than omitted from the interpretation.","section":"Section 4, Table 1"},{"comment":"The random starting-pose perturbation is described only qualitatively. Reporting the distribution of initial poses and the random seed policy would improve reproducibility.","section":"Section 3.5"},{"comment":"The abstract contains a grammatical error: 'whereas in a complex scenario is different.' Please revise.","section":"Abstract"},{"comment":"The notation V^⊔ is used in Eq. (3) but is not explicitly defined where the equation appears; define it in Section 3.3 for self-containment.","section":"Section 3.2.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not unsalvageable. The main risk is that the victim-cue term is oracle-fed in simulation; if so, the comparison is unfair and the central claim is misleading. I would ask the authors for a much more detailed description of the perception pipeline and, if necessary, rerun the S2 experiments with a realistic detector or a no-cue ablation. The statistical weakness in Table 1 also needs to be addressed head-on. I would not reject outright because the framework is sensible and the shared execution layer is methodologically sound, but the current evidence is insufficient for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: it's a plausible small extension of frontier exploration with a rescue-relevance term, but the evaluation doesn't yet back the headline claim. The S2 advantage could come from the method being given near-ground-truth victim cues rather than from smarter frontier scoring.\n\nThe genuinely new bit is the score in Eq. 1: adding observation deficit and a Gaussian rescue-relevance term around tentative victim detections to the usual info-gain/terrain/travel-cost mix. That's a legitimate extension within the established frontier program, not a new paradigm, and the paper is honest that in the easy scenario S1 the proposed method is worse than pure information gain. Credit where due: the shared local controller across methods is the right way to isolate frontier-selection effects, and the release of code and Gazebo setups supports reproducibility.\n\nThe soft spots are real. The rescue-relevance term R(f) depends entirely on the 'tentative victim observations' V^⊔, but Sec. 3.2.3 never specifies the detector: no range, no false-positive rate, no false-negative behavior. If those cues are effectively victim ground-truth positions, then the method is told where the victims are while the baselines are not, and the 20/20 in S2 demonstrates 'go to the cue' rather than balanced rescue-aware exploration. The baselines need to receive the same cue stream, even if they don't use it in scoring. Second, the weights in Eq. 1 are never reported. Without them the method can't be reproduced, and the whole evaluation is a single point in weight space. An ablation or sensitivity study is needed. Third, the statistical basis is thin: 20 runs, no error bars or tests. The SA vs Risk-aware differences (20 vs 19 completions, 2.00 vs 1.95 victims) are within noise; the big contrast is against Information Gain, which is also the baseline most disadvantaged if it lacks the cues.\n\nThe easy-scenario result is a useful check: the method's advantage doesn't generalize trivially, which the authors acknowledge. That's to their credit, but it also weakens the strength of the S2 claim.\n\nBottom line: this is a serious, clearly-written paper that deserves a proper peer review. The core idea is reasonable and the flaws are fixable with more experiments, published weights, and an honest description of the victim-cue pipeline. I wouldn't cite its results as established until that revision is done, but I'd send it to a competent referee.","headline":"A plausible small extension to frontier exploration for rescue, but the evaluation doesn't yet support the central claim because the rescue-cue term may be oracle-fed and the score weights are unreported.","tokens_in":7554,"tokens_out":3003,"would_cite":false,"duration_ms":27711,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that a frontier-prioritization score which blends information gain, observation deficit, rescue relevance, terrain penalty, and travel cost lets a quadruped robot find more victims in cluttered rescue environments than cla","keywords":["quadruped robots","search and rescue","frontier exploration","rescue relevance","victim recovery","autonomous exploration","multi-objective frontier scoring"],"falsifier":"A controlled experiment in which the same tentative victim cues are provided to all baselines (for example, by adding the rescue-relevance term to nearest, information-gain, and risk-aware frontier scoring) would show whether the advantage comes from the cues themselves or from the specific way the proposed method combines them. If the baselines with the same cues match the proposed method's completion rate and victim recovery, then the claim that the ranking mechanism is responsible would be falsified.","tokens_in":6678,"feed_emoji":"🚑","tokens_out":3152,"duration_ms":64321,"temperature":0.7,"pith_summary":"The paper proposes a rescue-aware frontier-prioritization method for a single quadruped robot searching for victims in unknown indoor environments. The method keeps the classical frontier-exploration structure but scores each frontier with five terms: information gain, observation deficit, rescue relevance, terrain penalty, and travel cost. The rescue-relevance term biases the robot toward frontiers near tentative victim cues, while the other terms preserve exploration and safe motion. Tested in two simulated rescue scenes of different difficulty, the method matches baselines in the easy scene and achieves the highest completion rate (20/20) and highest mean victim recovery (2.00 per run) in the complex scene. The authors conclude that rescue-aware prioritization pays off exactly when frontier choice is ambiguous and generic exploration criteria are not enough.","feed_headline":"Rescue-aware frontier ranking finds more victims in cluttered scenes","feed_subtitle":"A five-term frontier score beats classical exploration on completion and victim recovery in the hard scenario.","key_machinery":"The central mechanism is the additive frontier score J(f) = w_I I(f) + w_O O(f) + w_R R(f) − w_T T(f) − w_D D(f), where I is information gain, O is observation deficit, R is rescue relevance, T is terrain penalty, and D is travel cost. Rescue relevance R(f) is a sum over tentative victim observations of confidence-weighted Gaussian functions of distance, so frontiers near likely victim cues are promoted without abandoning exploration.","core_discovery":"In the complex scenario, the proposed Situation-aware Frontier Prioritization reaches a 20/20 completion rate and a mean of 2.00 confirmed victims per run, with a mission time of 373.5 s and path length of 57.45 m, outperforming nearest frontier, information gain, and risk-aware exploration on the rescue metrics while remaining competitive on travel effort. In the easy scenario, all methods perform well, and the information-gain baseline actually leads. The paper reads this pattern as evidence that rescue-aware prioritization matters when frontier choice is nontrivial and rescue utility must be traded against generic exploration.","pith_inferences":["If the tentative victim cues come from a real visual detector with nonzero false-positive rate, the Gaussian rescue-relevance term may steer the robot toward spurious locations; the paper does not quantify detector reliability, so a robustness test with noisy cues would clarify whether the advantage persists.","The Gaussian spatial prior is a simple parametric model; a natural extension is to learn the rescue-relevance prior from environment features (layout, occlusion, terrain) instead of relying only on current tentative detections.","The evaluation is entirely simulated with a single robot; moving to real hardware or multi-robot coordination would likely require recalibrating weights and the observation-deficit term to account for odometry drift and perceptual noise.","The result suggests that in cluttered search-and-rescue environments, the exploration policy should not treat all frontiers as equal: a small bias toward rescue-relevant regions can yield large gains in mission success, which is a testable design principle for field deployment."],"forward_implications":["The method improves mission reliability and victim recovery without requiring longer travel, since it achieves the best rescue metrics in the complex scenario without the longest path length.","The benefit is conditional on scenario difficulty: in an easy scene with a single victim and little clutter, simple information-gain exploration is sufficient, suggesting rescue-aware ranking is most valuable when frontiers are ambiguous.","Because the method preserves the frontier-exploration framework, it can be grafted onto existing frontier-based systems by replacing only the ranking function, leaving low-level control and mapping unchanged.","The same scoring structure could be extended to other task-aware exploration objectives beyond victim search, such as locating specific objects or monitoring regions of interest."],"fun_headline_variants":["Situation-aware frontiers find all victims in cluttered rescue scenes","Rescue ranking beats classical exploration in hard quadruped search","Quadruped rescue: new frontier score recovers more victims","In cluttered scenes, situation-aware frontiers beat baselines on victim recovery"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The rescue-relevance term R(f) depends on tentative victim observations produced by a visual confirmation process whose detection reliability is not reported; if these cues essentially mark true victim locations, the proposed method receives privileged information that the baselines do not, undermining the comparison's fairness.","fun_headline_variants_meta":{"raw":{"variants":["Situation-aware frontiers find all victims in cluttered rescue scenes","Rescue ranking beats classical exploration in hard quadruped search","Quadruped rescue: new frontier score recovers more victims","In cluttered scenes, situation-aware frontiers beat baselines on victim recovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1128,"prompt_tokens":710,"completion_tokens":418,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":346}},"tokens_in":454,"tokens_out":418,"duration_ms":4426,"temperature":1.0,"reasoning_tokens":346,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T04:27:28.235860+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment in which the same tentative victim cues are provided to all baselines (for example, by adding the rescue-relevance term to nearest, information-gain, and risk-aware frontier scoring) would show whether the advantage comes from the cues themselves or from the specific way the proposed method combines them. If the baselines with the same cues match the proposed method's completion rate and victim recovery, then the claim that the ranking mechanism is responsible would be falsified.","supporting_citations":[],"review_version":1}