{"id":"f26ce34c-997c-4535-a9b2-05dcdd20deed","arxiv_id":"2504.14862","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FERMI combines physics-based direct path attenuation with a neural network over line-of-sight surface points to predict Wi-Fi signal strength for unseen transmitter-receiver pairs from sparse data, and plans multi-robot data collection.","lead":"FERMI is a robotic mapping system that predicts Wi-Fi signal strength between any two points in a building using a mix of physics for direct paths and a neural network for reflections and shadows. A companion planner tells a team of robots where to go so the needed measurements are collected quickly without human help.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The generalization guarantee in Sec. IV-A is not established: the network learns pairwise surface-point interactions, and \"LOS points covered\" does not imply the required pair combinations are covered; Eq. (2)'s direct-only retransmitter weights also drop indirect illumination and incident phase.","rationale":"The reader's weakest assumption about the independence of Tx LOS points in Eq. (2) is a real and important concern, and I agree it directly threatens the physical fidelity of the model in reflective or shadowed scenes. However, the most load-bearing gap in the paper's central claim is slightly broader: the stated generalization condition, 'LOS points are covered,' is not sufficient because the learned network is a pairwise function and the training data supervises pairs, not individual points. The paper provides no coverage statistics, no pair-level generalization analysis, and no error bars, so the mechanism underlying the reported 5.59-11.29 dBm MAE results is not demonstrated. These issues are addressable and do not by themselves falsify the method, so the existing CONDITIONAL verdict is appropriate; no verdict change is needed.","tokens_in":15114,"tokens_out":8420,"duration_ms":87825,"concrete_test":"Using the Sionna ray tracer, construct a scene with a dominant two-bounce path: Tx illuminates wall A, wall A illuminates wall B in shadow, wall B reaches Rx, while Tx-B is NLOS and Rx-A is NLOS. Train FERMI on dense LOS-point data collected around this scene, then compare predictions for test Tx/Rx pairs where the two-bounce contribution is large. If FERMI's error on those pairs is substantially higher than on equal-coverage single-bounce pairs, the direct-only alpha_j in Eq. (2) is the cause. A stronger variant is to re-run with alpha_j replaced by the full complex incident field from Sionna; if MAE drops, the independence assumption is the load-bearing defect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. IV-A claims FERMI generalizes to unseen Tx-Rx pairs 'provided their corresponding LOS points are covered by those in the training data.' This condition is insufficient. F_Theta in Eq. (1) is a function of one Tx-LOS point and one Rx-LOS point; training supervises only the Cartesian product of LOS-point pairs generated by the collected Tx-Rx measurements. An unseen Tx/Rx can have LOS point sets that are individually well-covered while their pairwise combinations are unseen, and the paper gives no argument or measurement that the MLP/hash representation interpolates reliably across this pair space. The experimental section does not report LOS-point coverage rates, error bars, or a breakdown of error by pair-novelty, so the claimed mechanism is untested. Compounding this, Eq. (2) weights each Tx LOS point by the normalized direct-path energy alpha_j from the Tx only. Under the Huygens-Fresnel motivation, the retransmitting surface is driven by the total incident field, including indirect illumination; using a scalar energy weight also discards the phase of the direct path. Thus even complete LOS-point coverage would not recover multiple-bounce contributions whose first surface interaction is not directly illuminated, and the model is not guaranteed to produce the reported accuracy in heavily reflective or deeply shadowed environments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FERMI, a framework for building radio signal strength maps in large, occluded environments using sparse measurements collected by multiple robots. The approach decomposes propagation into three stages: a physics-based direct-path model from the transmitter to sampled line-of-sight (LOS) surface points, a neural network that models multipath propagation between pairs of LOS points associated with the transmitter and receiver, and a physics-based direct-path model from the receiver's LOS points to the receiver. A visibility-based scene partitioning method and a configuration-planning formulation are introduced to coordinate multi-robot data collection. The authors evaluate FERMI on five simulated scenes using Sionna ray tracing and in a real corridor environment with three robots, reporting lower mean absolute error than PropEM-L, Gaussian processes, and NeRF2 baselines, as well as improved scalability of data collection planning compared with a prior greedy method.","tokens_in":15393,"tokens_out":4784,"duration_ms":47667,"significance":"If the claims hold, FERMI would be a practically valuable contribution: it targets a real limitation of existing radio mapping methods, namely prediction for transmitter-receiver pairs not seen in training, and it combines this with an autonomous multi-robot collection strategy. The paper's strengths include open-sourced code, a real-world validation with a three-robot system, a clear problem formulation for data collection planning, and an explicit decomposition of propagation that is more physically motivated than fully black-box baselines. However, the central generalization claim is not established as stated, and several parameters and baseline adaptations are left underspecified; these issues are load-bearing for the paper's main contribution and need to be addressed before the results can be fully credited.","major_comments":[{"comment":"The paper's central claim that the network generalizes to unseen Tx-Rx pairs 'provided their corresponding LOS points are covered by those in the training data' is not supported. The network input in Eq. (1) is a single pair of LOS points (P_Tj, P_Rwksi), and training supervises only the aggregate rendered signal for collected Tx-Rx pairs. Coverage of the individual LOS point sets does not imply coverage of the Cartesian product of Tx-LOS and Rx-LOS point pairs that is needed for a novel pair. Please provide either a proof of an interpolation or Lipschitz property of the hash-encoded MLP over this pair space, or experiments that quantify performance as a function of pair-novelty (e.g., error broken down by the fraction of LOS-point pairs seen in training). Without this, the generalization mechanism asserted in the abstract and Section IV-A remains untested.","section":"Section IV-A, Generalization"},{"comment":"In Eq. (2), the weight alpha_j is defined as the normalized energy of the direct path from the transmitter to the LOS point P_Tj. Under the stated Huygens-Fresnel motivation, each surface point should be driven by the total incident field, including indirect illumination, and the phase of that field matters. Using scalar direct-path energy discards phase and assigns zero weight to surface points that are not directly visible from the transmitter, even if they are strongly illuminated by multipath. Consequently, the model cannot represent multipath contributions whose first surface interaction is with an NLOS point from the transmitter. The authors should quantify this approximation error (for example, by comparing against full ray tracing in deliberately shadowed scenes) or explain why such contributions are negligible in the tested environments.","section":"Section IV-B, Eq. (2)"},{"comment":"All quantitative results appear to be based on single runs without error bars or repeated-seed statistics. The reported MAE differences (for example, 5.59 dBm versus 9.12 dBm for the Room scene in Table I) are hard to interpret without variance estimates. In addition, the adaptation of the NeRF2 baseline to the multi-transmitter training and test protocol is not described; since NeRF2 in [53] models a fixed transmitter, the comparison may disadvantage the baseline. Please describe the baseline adaptations and report repeated-run statistics (mean and standard deviation) for the main tables and figures.","section":"Section VI-A, Tables I-III and Figs. 8-9"},{"comment":"The LOS compensation in Eq. (6) introduces alpha_los and Rref without specifying how they are computed. The text says alpha_los is a distance-related attenuation factor and equals zero under NLOS conditions, but no formula or calibration procedure is given, even though Fig. 9 shows this term is important for accuracy. The notation also mixes the complex quantity RRx with the magnitude |RRx| and the presumably real Rref. Please provide the exact definition of alpha_los, the normalization of Rref, and clarify the units and whether Eq. (6) operates on complex values or real magnitudes.","section":"Section IV-B, Eq. (6)"}],"minor_comments":[{"comment":"Please clarify the relationship between RRx, |RRx|, and Rref in Eq. (6); it is unclear whether the equation combines complex amplitudes and then takes a magnitude, or whether it operates directly on dBm values.","section":"Section IV-B, Eq. (6)"},{"comment":"The scene partitioning depends on a distance threshold D, but no value is reported; please state the threshold used in the experiments, along with the ray sampling count and hash-encoding resolution for the LOS point generation.","section":"Section V-C"},{"comment":"For the Gaussian process baseline, the choice of kernel and hyperparameter settings is not given; please provide these details for reproducibility.","section":"Section VI-A"},{"comment":"Figure 6 is described as illustrating all five simulation scenes, but only one scene appears to be shown; please show all scenes or adjust the caption.","section":"Section VI-A"},{"comment":"There are minor typographical and stylistic issues throughout, such as inconsistent capitalization in the reference to 'Ego-Swarm' [55]; a careful proofread is recommended.","section":"References and text"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the real-world validation is encouraging, but the theoretical generalization argument needs substantial rework or experimental backing. The undefined alpha_los in Eq. (6) and the underspecified baseline adaptation are fixable, but the pair-coverage issue goes to the heart of the claimed contribution and should be the focus of the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on FERMI. The central idea is a good one: split radio propagation into two explicit physics stages (Tx to LOS points, Rx LOS points to Rx) and one learned stage (LOS-point to LOS-point). That decomposition makes the network input agnostic to absolute Tx/Rx positions, so a handful of collected measurements can inform many LOS-point pairs—this is why the method does well with sparse data. The planning half (visibility-based region partitioning, then set-cover plus TSP for robot configurations) is standard but effective, and the real-world demo with three robots in a corridor is a respectable proof of concept.\n\nThe soft spots are mostly about overclaiming and insufficient evidence. The claim in Sec IV-A that generalization follows from LOS-point coverage is too strong. The network maps a pair of LOS points, so an unseen Tx/Rx pair with individually covered LOS points can still have unseen pair combinations; no argument or experiment shows the network interpolates reliably across that pair space. The stress-test is right to push on this. Similarly, the retransmitter weight alpha_j in Eq (2) only includes direct-path energy from the Tx, so a LOS point that is not directly illuminated but is energized by indirect bounces gets zero weight. That is a real limitation in reflective or deeply shadowed scenes, and it undercuts the Huygens-Fresnel motivation.\n\nThe experiments themselves are encouraging but reported as single runs—no error bars in Tables I-III or Figures 8-9. The NeRF2 baseline adaptation to multiple training Txs is not described, which makes the comparison hard to interpret. Eq (6) introduces alpha_los without saying how it is computed, and the Section V coverage guarantee is stronger than the diagonal-exclusion argument supports. None of these are fatal; they are all addressable in revision.\n\nMy overall read: this is a solid engineering contribution with a genuinely novel propagation-model decomposition and a real system demonstration. The generalization story needs to be honestly restated—as an empirical property of the representation, not a formal guarantee—and the evaluation needs error bars and a described baseline. I would send it to peer review with that expectation. The code is open-sourced, though without a commit hash or dataset, so reproducibility is currently partial.","headline":"FERMI's LOS-point decomposition is a genuinely useful idea with solid real-world results, but the paper overclaims its generalization guarantee and needs error bars and a described baseline.","tokens_in":15937,"tokens_out":2768,"would_cite":true,"duration_ms":25220,"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 hybrid physics-plus-neural model predicts radio signal strength between unseen transmitter-receiver pairs from sparse data.","keywords":["radio mapping","hybrid propagation model","multipath propagation","neural radiance fields","multi-robot data collection","visibility-based partitioning","signal strength prediction"],"falsifier":"Place a transmitter in a corner of a highly reflective room, train FERMI on dense coverage of that transmitter's LOS points, then predict signal strength at a receiver in a deep shadow niche and compare against ray-tracing ground truth that includes multi-bounce paths; a systematic large error in those receiver positions would falsify the independent-retransmitter assumption.","tokens_in":14935,"feed_emoji":"📡","tokens_out":7815,"duration_ms":62736,"temperature":0.7,"pith_summary":"FERMI claims that radio mapping in obstacle-rich environments can be made accurate and data-efficient by splitting propagation into two explicit physical stages and one learned stage. The direct path from a transmitter to its line-of-sight (LOS) surface points, and the direct path from a receiver's LOS surface points to the receiver, are computed with classical distance-based attenuation; only the multipath exchange between the two LOS point sets is learned by a neural network. Because the network's inputs are LOS-point coordinates rather than absolute transmitter or receiver positions, it is not tied to any specific pair and can predict signal strength for transmitter-receiver combinations never seen in training, provided their LOS points appear in the training data. The paper reports the lowest mean absolute error among tested methods in all five simulated scenes (5.59-11.29 dBm) and in a real corridor (6.24 dBm), using roughly 1-2 m sparse sampling and three robots.","feed_headline":"Hybrid model predicts Wi-Fi strength at unseen spots","feed_subtitle":"Physics handles direct paths; a neural net learns multipath, so sparse robot data suffices.","key_machinery":"The central object is the multipath propagation network $F_\\Theta: (P_{T_j}, P_{R_{w_k s_i}}) \\rightarrow (\\delta(P_R), S(P_T, P_R))$, built from two MLPs with hash-encoded inputs, where $\\delta$ is a material-related attenuation coefficient at a receiver LOS point and $S = a\\,e^{j\\theta}$ is the complex signal retransmitted from a transmitter LOS point to a receiver LOS point. Each transmitter LOS point is treated as an independent retransmitter weighted by $\\alpha_j$, the normalized direct-path energy (Eq. 2); receiver LOS point signals are integrated along rays with transmittance weights $L = \\exp(-\\sum \\sigma \\delta)$ (Eqs. 3-4), summed over directions with receiver gain $G(\\omega)$ (Eq. 5), and combined with the LOS direct-path term (Eq. 6). The same LOS-point representation also drives the data-collection planner: visibility-based scene partitioning groups mutually visible surfaces so that placing robots across region pairs covers all mutually invisible surface-point pairs needed for multipath training.","core_discovery":"On its own terms, the paper's central claim is that the received signal strength for an arbitrary transmitter-receiver pair can be rendered from surface-level interactions rather than learned as a function of the two absolute positions. The transmitter's LOS points act as independent retransmitters, each carrying a weight equal to the normalized direct-path energy from the transmitter; the receiver-side signal is accumulated from its LOS points through a volume-rendering equation with learned attenuation coefficients, and under LOS conditions a physics-computed direct component is added. Trained only on sparsely sampled robot-collected measurements, this construction yields accurate predictions for unseen Tx-Rx pairs because the network's inputs are LOS point coordinates, which are largely shared between training and test pairs, and because each measured pair contributes thousands of LOS-point pairs to training.","pith_inferences":["The same LOS-point decomposition could plausibly transfer to other wave fields with Huygens-like secondary sources, such as acoustic impulse responses, although the paper only demonstrates Wi-Fi.","The visibility-based region partitioning is a reusable primitive for any pairwise field measurement where multipath interactions matter, not only radio mapping.","Because the model is trained per scene, a natural next step the paper does not take is to test whether features learned in one building warm-start or transfer to a new building, potentially removing the retraining requirement.","The planner's greedy set-cover step could be replaced by an uncertainty-guided objective that prioritizes LOS point pairs with the largest predicted variance, making collection even more data-efficient."],"forward_implications":["A sparse radio map can predict signal strength for arbitrary points in the scene, not just the measured transmitter-receiver combinations, which is the property needed for communication-aware planning.","The autonomous collection method removes the manual Tx-Rx deployment bottleneck; a three-robot team collected around 40,000 valid measurements in a 94 m by 64 m corridor within one hour.","The planning solution scales to larger teams: using 9 robots cuts the number of configuration transitions from 88 (three robots) to 10, and planning time stays under six seconds for 15 robots.","Because each collected data point trains thousands of LOS-point pairs, the model degrades only mildly when the training set is sparsified, an effect the paper demonstrates in its data-reduction experiments."],"supporting_citations":[{"why":"Supplies the distance-related path attenuation model used for the direct-path stages from the transmitter to its LOS points and from the receiver's LOS points to the receiver.","marker":"[2]"},{"why":"Provides the volume-rendering transmittance formulation adapted here to accumulate complex signals from receiver LOS points along rays.","marker":"[29]"},{"why":"The neural radiance-field radio model whose two-MLP architecture inspires FERMI and which serves as a comparison method in the experiments.","marker":"[53]"},{"why":"Supplies the multiresolution hash encoding used to encode LOS-point positions as inputs to the network.","marker":"[32]"},{"why":"Supplies the ray-tracing engine used to generate the simulated ground-truth signal data in the five test scenes.","marker":"[16]"},{"why":"The prior multi-robot data-collection planner that FERMI's configuration-sequence approach is compared against on transition count, transition cost, and computation time.","marker":"[19]"},{"why":"The scene-encoding benchmark method used for comparison; the paper reports lower MAE with FERMI in both simulation and real-world tests.","marker":"[8]"},{"why":"The Gaussian-process baseline for radio mapping; the paper reports large MAE reductions over it.","marker":"[11]"}],"fun_headline_variants":["Sparse robot data, hybrid model: radio maps that generalize","Physics-AI combo maps Wi-Fi in complex, cluttered scenes","Robot fleets gather sparse data, FERMI maps radio signals","Unseen spots covered: hybrid model from sparse measurements"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The key assumption is that each surface point near the transmitter re-radiates based only on the unoccluded direct path from the transmitter, so a surface point lit indirectly by other reflected signals is treated as if that indirect illumination did not exist; if this fails in highly reflective or deeply shadowed scenes, the claimed generalization collapses.","fun_headline_variants_meta":{"raw":{"variants":["Sparse robot data, hybrid model: radio maps that generalize","Physics-AI combo maps Wi-Fi in complex, cluttered scenes","Robot fleets gather sparse data, FERMI maps radio signals","Unseen spots covered: hybrid model from sparse measurements"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000601,"raw_usage":{"total_tokens":2805,"prompt_tokens":938,"completion_tokens":1867,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":1796}},"tokens_in":554,"tokens_out":1867,"duration_ms":12679,"temperature":1.0,"reasoning_tokens":1796,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:39:09.138149+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place a transmitter in a corner of a highly reflective room, train FERMI on dense coverage of that transmitter's LOS points, then predict signal strength at a receiver in a deep shadow niche and compare against ray-tracing ground truth that includes multi-bounce paths; a systematic large error in those receiver positions would falsify the independent-retransmitter assumption.","supporting_citations":[{"cited_title":"Radar: An in-building rf-based user location and tracking sys- tem","cited_arxiv_id":null,"evidence_quote":"Supplies the distance-related path attenuation model used for the direct-path stages from the transmitter to its LOS points and from the receiver's LOS points to the receiver."},{"cited_title":"Optical models for direct volume render- ing","cited_arxiv_id":null,"evidence_quote":"Provides the volume-rendering transmittance formulation adapted here to accumulate complex signals from receiver LOS points along rays."},{"cited_title":"Nerf2: Neural radio-frequency radiance fields","cited_arxiv_id":null,"evidence_quote":"The neural radiance-field radio model whose two-MLP architecture inspires FERMI and which serves as a comparison method in the experiments."},{"cited_title":"Instant neural graphics primitives with a multiresolution hash encoding","cited_arxiv_id":null,"evidence_quote":"Supplies the multiresolution hash encoding used to encode LOS-point positions as inputs to the network."},{"cited_title":"Sionna: An open-source library for next-generation physical layer research","cited_arxiv_id":null,"evidence_quote":"Supplies the ray-tracing engine used to generate the simulated ground-truth signal data in the five test scenes."},{"cited_title":"Con- structing radio signal strength maps with multiple robots","cited_arxiv_id":null,"evidence_quote":"The prior multi-robot data-collection planner that FERMI's configuration-sequence approach is compared against on transition count, transition cost, and computation time."},{"cited_title":"PropEM-L: Radio Propagation Environment Modeling and Learning for Communication-Aware Multi-Robot Exploration","cited_arxiv_id":null,"evidence_quote":"The scene-encoding benchmark method used for comparison; the paper reports lower MAE with FERMI in both simulation and real-world tests."},{"cited_title":"Online methods for radio signal mapping with mobile robots","cited_arxiv_id":null,"evidence_quote":"The Gaussian-process baseline for radio mapping; the paper reports large MAE reductions over it."}],"review_version":1}