{"id":"7336b1ea-ce2b-400d-bfb2-eafec0320c6c","arxiv_id":"2602.19108","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Fusing thermal and depth images lets a legged robot build a real-time radiation-field cost map and plan paths that avoid fire heat while still reaching its goal.","lead":"Fire is hard for robots to avoid because cameras see hot surfaces, not the heat in the air; this paper builds a live map of that heat and steers a robot around it. A Boston Dynamics Spot uses a thermal camera and depth sensors to construct a radiation-field cost map, then plans paths that stay cool while still reaching the goal.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 omits radiative fraction X used in the model and validation, so the experiments may not test the calibrated radiation field.","rationale":"The reader's weakest assumption focuses on the values of T0, γ, and X from literature and the fact that the model overpredicts measured flux by 68%. That is a reasonable concern about calibration, and it is supported by the paper's own calorimetry. However, the overprediction is conservative: if the model overestimates the hazard, the planned paths are further away than necessary, which still satisfies a safety claim. The more directly load-bearing issue is the internal inconsistency between the algorithmic description and the validation: Algorithm 1 omits the radiative fraction X that is present in Eq. (3) and used in the calibration. This means the experiments may not test the model that is described and calibrated. If the implementation indeed omits X, the safe distance is 1.58 m instead of 0.93 m, so the reported error is not the error of the implemented system. This does not necessarily invalidate the qualitative conclusion that the robot avoids the fire, but it makes the quantitative claims unreliable and suggests the method as described may not be the method as tested. I agree with the reader's conditional verdict, but I flag this specific implementation mismatch as the most concrete and checkable problem. The paper deserves credit for including an independent calorimetric check, but the missing link between the model equation and the code is a serious gap that should be resolved before the safety claims are accepted as quantitatively verified.","tokens_in":9896,"tokens_out":8526,"duration_ms":82049,"concrete_test":"Audit the source code used for the experiments to check whether the radiative fraction X is applied in the radiation update (Algorithm 1, line 24). Recompute the predicted flux at 0.45 m and the corresponding safe distance with and without X using the reported P = 78.5 kW. Then re-run the navigation experiments (or simulation) with the corrected implementation that includes X, and measure the minimum distance to the fire along the planned paths. If the paths change materially (e.g., the 'close pass' in Fig. 6i-l becomes longer), the reported results do not reflect the validated model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the robot avoids hazardous regions using a Stefan-Boltzmann inverse-square radiation field. However, there is a concrete inconsistency between the model and its implementation. In Section III-B, Eq. (3) defines the radiative flux as P(r) = P / (4πr²) · X, with X = 0.35. In the validation (§IV-B), this X is used to obtain 10.79 kW/m² at 0.45 m and a safe distance of 0.93 m. But Algorithm 1, line 24, computes T[i,j] = P / (4πd²) without multiplying by X. If the actual planner code matches Algorithm 1, the flux at 0.45 m would be 30.85 kW/m² and the safe distance would be 1.58 m — not the reported 0.93 m. This means the field actually used in the experiments is not the one that was calibrated, and the reported 0.22 m safe-distance error does not reflect the deployed system. The qualitative detour behavior is plausible, but the quantitative basis for 'thermally safe paths' is undermined because the experiments may validate a different (overly conservative) model. This is a correctness issue in the method description, not just an uncertainty in physical constants.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fire-aware navigation framework for mobile robots. Depth and thermal images are registered into a temperature-annotated 3D point cloud; high-temperature points are clustered and fit with a hemisphere to localize the fire and estimate its radiative power via the Stefan–Boltzmann law. An inverse-square decay model is used to construct a 2D thermal radiation field with line-of-sight occlusion handling, which is fused into an occupancy grid and used to plan A* paths. The system is demonstrated on a Boston Dynamics Spot robot with a controlled propane fire, and a calorimetry experiment is used to validate the radiation magnitude.","tokens_in":10219,"tokens_out":7463,"duration_ms":65706,"significance":"The idea of extending surface-centric thermal maps to a navigable free-space radiation field is timely and practically relevant for fire response robotics. The paper includes real-robot experiments, a physical model with explicit assumptions, and an independent calorimetric check, which are strengths. However, the quantitative claims rest on several assumed constants (T0, γ, X) and the validation shows a 68% overprediction in flux. More importantly, there is a concrete mismatch between the model in Eq. (3) and the algorithm pseudocode in Algorithm 1. If the deployed planner follows Algorithm 1, the calibrated and validated field is not the field used during navigation. These issues need to be resolved before the central 'thermally safe paths' claim is fully supported. The qualitative detour behavior is plausible and useful, but the quantitative basis needs repair.","major_comments":[{"comment":"Algorithm 1 line 24 computes T[i,j] ← P/(4πd²), omitting the radiative fraction X=0.35 from Eq. (3). With P=78.5 kW and r=0.45 m, Eq. (3) gives 10.79 kW/m² and a safe distance of 0.93 m (matching §IV-B), whereas the Algorithm 1 expression gives 30.85 kW/m² and a safe distance of 1.58 m. The pseudocode and the validation therefore describe different radiation fields. Please clarify which expression is used in the deployed planner, and if the pseudocode is wrong, correct it. This is load-bearing because the claimed calibrated safety distance (0.93 m) does not correspond to the algorithm as written.","section":"§III-B, Algorithm 1, §IV-B"},{"comment":"There is a physical inconsistency in the definition of P. Table I labels P as 'Estimated radiative power of fire source,' but Eq. (3) multiplies P by X=0.35, described as the radiative fraction of total heat release. If P is already radiative output, X should be 1; if P is meant to be total heat release, Eq. (2) should be described as estimating heat release rate, not 'thermal radiation emitted from the fire surface.' This ambiguity changes the field by a factor of 1/X≈2.86 and affects every cost-map threshold.","section":"§III-B, Table I"},{"comment":"The validation shows the model overpredicts incident flux by 68% (10.79 vs 6.4 kW/m²) and safe distance by 0.22 m. Because the overprediction is conservative, it does not invalidate the avoidance demonstration, but the abstract and conclusions use the phrase 'thermally safe paths.' As the limitations section correctly notes, fire strength is assumed rather than estimated. I ask for either a margin/safety analysis demonstrating that the executed paths remain outside the measured hazard zones, or a more modest claim such as 'conservative avoidance of high-radiation regions.'","section":"§IV-B, §V"}],"minor_comments":[{"comment":"There is a typographical error: 'max(O[x, y],; min(...))' contains an extra semicolon after the comma.","section":"Eq. (7)"},{"comment":"The paper claims 'real-time' thermal radiation fields, but no timing or latency measurements are reported. Please provide runtime data or remove the 'real-time' wording.","section":"Abstract, §IV"},{"comment":"The sensor description is inconsistent: §III-A says depth is obtained from the stereo camera, while Algorithm 1 processes 'LiDAR scans.' Clarify which sensor(s) are used for occupancy and obstacle detection.","section":"§I, §III-A, Algorithm 1"},{"comment":"Minor typos: 'Fore one' should be 'For one' (Discussion), 'UA V' should be 'UAV' (Introduction), and 'scaler' should be 'scalar' (Eq. 6). Also, some figure references are repeated or ambiguous (e.g., 'In Fig. 7 we show...' appears twice in §IV-A).","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central contribution is potentially useful for fire-response robotics, but the load-bearing quantitative linkage between the model, the algorithm, and the validation is currently broken. The revisions requested are local: fix the Algorithm 1/Equation (3) mismatch, resolve the P/X interpretation, and temper the safety claim. With those changes, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the free-space thermal radiation field: instead of stopping at surface temperature maps or thermal SLAM, the authors take a fused thermal-depth point cloud, localize the fire by clustering hot points and fitting a hemisphere, and propagate a Stefan-Boltzmann inverse-square field into empty space for A* planning. That is a real extension of the prior work they cite, and they show it working on a Spot robot with a live propane fire. The calorimetric validation with water heating plus the fuel-consumption cross-check is more than most robotics papers do, and the limitations section is honest about single-source, radiative-only, assumed-temperature weaknesses. Credit is due for that.\n\nThe soft spots are just as real. The model is a toy: fixed T0=1473.15 K, an empirically chosen gamma=0.4, and a radiative fraction X=0.35 lifted from pool-fire literature. Their own calorimetry shows the model overpredicts incident flux by about 68% (10.79 vs 6.4 kW/m2) and the safe distance by 0.22 m. That overprediction may be conservative, but it is not a validated safety metric. More concerning is a genuine inconsistency between Eq. (3), which includes X, and Algorithm 1, which computes T[i,j]=P/(4*pi*d^2) with no X. If the implementation matched the algorithm as printed, the actual radiation at 0.45 m would be 30.85 kW/m2, not 10.79, and the quoted 0.93 m safe distance would instead be 1.58 m. The paper does not explain this mismatch. It could be that the intensity P passed into the algorithm is already the X-scaled power, but as written it is a bug in the description, and the reader cannot tell what was actually run. The navigation results are also thin: single runs, no error bars, no baselines, no released code or data.\n\nAll that said, the qualitative behavior in the figures is plausible, and the central concept does not collapse. The paper is a proof of concept, not a finished deployment. It deserves a serious referee because the representation is useful and the validation approach is a good template, even if the numbers need tightening. My recommendation: send it out, but ask the authors to fix the algorithm/model inconsistency, report the calibration error honestly in the navigation section, and release the code and data so the claims can be checked.","headline":"A worthwhile proof-of-concept for fire-aware navigation that extends thermal maps into free space, but the quantitative claims need cleanup: hand-tuned constants, a 68% flux overprediction, and an inconsistency between the model equation and the algorithm pseudocode.","tokens_in":748,"tokens_out":1586,"would_cite":true,"duration_ms":38012,"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":"A mobile robot can navigate around fires safely by building a real-time thermal radiation field from fused depth and thermal images and letting an A* planner trade distance against heat.","keywords":["thermal radiation field","fire-aware navigation","Stefan-Boltzmann law","thermal-depth fusion","A* path planning","quadruped robot navigation","line-of-sight heat propagation","costmap safety margins"],"falsifier":"Set up a heat-flux or thermocouple array at the planner-computed danger boundary (e.g., 0.71 m and 0.93 m from the flame used in the paper) and at several larger radii. If the measured flux at any cell the planner marks safe exceeds the q_danger threshold, or if the decay with distance diverges visibly from 1/r^2 (e.g., because convection dominates), the central claim that the radiation field predicts safe navigation fails.","tokens_in":9773,"feed_emoji":"🔥","tokens_out":7164,"duration_ms":63831,"temperature":0.7,"pith_summary":"Fire is usually mapped as surface temperatures, but the hazard a robot faces lives in the empty space between it and the flame. This paper tries to establish that a robot can construct that missing free-space hazard in real time: fused depth and thermal images yield a temperature-annotated 3D point cloud, a fitted hemisphere localizes the fire, and a Stefan–Boltzmann inverse-square model spreads the flame's radiative power through the environment, gated by line of sight. The resulting thermal radiation field plugs directly into an A* cost map, so the planner can avoid hot regions while still reaching goals. Experiments on a quadruped with a controlled propane fire show paths that widen or tighten with a safety margin and that exploit wall occlusion as a heat shield. A calorimetric check shows the model overpredicts incident heat by about 68% and pushes the estimated safe distance out by 0.22 m—conservative for this test fire, but a sign that the constants carry the quantitative weight.","feed_headline":"Robot dodges fire by mapping radiant heat in empty space","feed_subtitle":"Thermal and depth cameras feed an A* planner that avoids predicted hot zones and still reaches the goal.","key_machinery":"The mechanism that carries the argument is a Stefan–Boltzmann inverse-square radiation field: a scalar map T[x,y] = P/(4πr^2)X defined at every free-space cell with line of sight to a fitted hemisphere fire source. The field is generated from registered depth–thermal clouds (a cluster of points above 100 °C for fire detection, a hemisphere footprint for area A), multiplied by the empirical factor γ = 0.4, and blended into occupancy as O = max(geometric, min(T/q_danger, 1)). The named danger threshold q_danger = 2.5/max(0.1, ϕ) converts heat flux into an impassable boundary; the A* cost C = 1 + Oβ then performs the route search. This lets one physical formula turn sparse surface temperatures","core_discovery":"The paper's central claim is that fire, for navigation purposes, is best represented not as a surface-temperature overlay but as a continuous radiative heat-flux field in free space, recoverable in real time from onboard sensors. The method fuses thermal and depth images into a 3D point cloud whose hottest cluster is treated as the fire source; a hemisphere fitted to that cluster supplies the emitting area A, and the Stefan–Boltzmann law P = σAT0^4γ with fixed T0 = 1,473.15 K and empirical γ = 0.4 yields the fire's radiative power. That power is spread over the environment by an inverse-square law with radiative fraction X = 0.35, gated by line of sight to the fire, and written into a 2D occ","pith_inferences":["A testable extension would estimate the fire's radiative power online from measured fuel mass-flow rate, replacing the fixed T0 and γ and likely removing the 68% overprediction.","Because smoke and hot combustion gases also radiate and convect heat without line of sight to the flame, the safe-corridor-behind-a-wall behavior may not transfer to real smoke-filled fires; a heat-flux probe behind an occluding wall would test whether the line-of-sight assumption is conservative or unsafe.","The same representation should scale to multiple fires by summing P(r) contributions from independently detected clusters, provided line-of-sight occlusion is tested per source.","The 192.5 °C ceiling of the thermal sensor likely truncates the fire cluster and biases the hemisphere fit, so repeating with a higher-dynamic-range camera is a cheap way to test whether the power overestimate comes from the sensor or the physical constants."],"forward_implications":["If correct, any cost-based planner can become fire-aware simply by adding a thermal radiation layer computed from a depth camera and a low-cost thermal camera.","The adjustable safety margin ϕ gives operators direct, real-time control over the efficiency–safety trade-off, as shown by the wide and narrow detours around the same fire.","The line-of-sight treatment of radiation means walls and tall obstacles create genuinely cooler corridors; the wall-occlusion experiment confirms the planner exploits this.","The calorimetric comparison implies that safe-distance estimates are conservative for the tested propane device—the model predicts 0.93 m clearance where measured flux would allow 0.71 m—so the system errs on the side of caution at the cost of extra detour.","The method is not surface-centric: it extrapolates from the visible flame to empty space, which is exactly the region a thermal camera alone cannot measure."],"fun_headline_variants":["Robot maps radiant heat to avoid fire zones","Thermal radiation fields steer robots from fire","Spot robot uses Stefan-Boltzmann law to dodge flames","Real-time heat mapping guides robots around fire"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The quantitative safety of every planned path rests on assumed fire constants—flame temperature T0 = 1,473.15 K, correction factor γ = 0.4, and radiative fraction X = 0.35—and the paper's own calorimetry shows the model overpredicts incident heat by 68%, so the computed 'safe' distance is only as trustworthy as those constants.","fun_headline_variants_meta":{"raw":{"variants":["Robot maps radiant heat to avoid fire zones","Thermal radiation fields steer robots from fire","Spot robot uses Stefan-Boltzmann law to dodge flames","Real-time heat mapping guides robots around fire"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1031,"prompt_tokens":717,"completion_tokens":314,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":255}},"tokens_in":461,"tokens_out":314,"duration_ms":3933,"temperature":1.0,"reasoning_tokens":255,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T21:44:03.980325+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Set up a heat-flux or thermocouple array at the planner-computed danger boundary (e.g., 0.71 m and 0.93 m from the flame used in the paper) and at several larger radii. If the measured flux at any cell the planner marks safe exceeds the q_danger threshold, or if the decay with distance diverges visibly from 1/r^2 (e.g., because convection dominates), the central claim that the radiation field predicts safe navigation fails.","supporting_citations":[],"review_version":1}