{"id":"060e5c2e-2f5a-44d3-b9dd-f7e05f2b927a","arxiv_id":"2501.17881","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A wireless localization method that recovers target positions by gradient descent through a differentiable ray-tracing simulator calibrated to the environment.","lead":"RayLoc turns wireless indoor localization into an inverse problem: instead of estimating a few signal parameters, it optimizes the position of a target inside a differentiable ray-tracing simulator until the simulated CSI matches the measured CSI. The approach unifies device-free and device-based localization and, if it holds up, could give sub-meter accuracy with a single access point.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eqs. 16 and 19 require per-path delays that a 20 MHz CSI cannot resolve; without a stated extraction procedure, the central experimental claim rests on an unverified step.","rationale":"I read the paper as making a specific, falsifiable claim: that by inverting a fully differentiable ray tracer and calibrating the scene, RayLoc can localize both devices and passive objects from real-world CSI more accurately than conventional baselines. The central mechanism requires a loss that compares simulated and measured propagation paths, and the paper's Eqs. 16 and 19 both require per-path delays from real-world measurements. The reader's weakest_assumption identifies exactly this gap, and I agree with that diagnosis. My stress-test sharpens it into an internal inconsistency: Eq. 16 sums per-path delays over an index set Ntot that is defined as the CSI matrix dimension rather than the number of traced paths, so even the notation does not support the intended computation. The 20 MHz bandwidth makes the problem concrete rather than merely notational, because individual paths in a room-scale environment cannot be resolved at 15 m delay resolution. This is not an external disagreement with the field; it is a missing link inside the paper's own loss construction. The concern is serious, because if τ_i cannot be obtained, the reported localization accuracy in Section 5.4 is not supported by the described method. However, it is also fixable in principle: the authors could clarify that τ_i are obtained by a specific super-resolution or model-fitting step, or they could reformulate the losses using frequency-domain complex CSI. Therefore, I do not move the verdict to rejection; the appropriate disposition remains conditional, matching the reader's verdict. The single concrete check that would settle the issue is to require a reproducible delay-extraction result from the measured CSI and to test whether localization survives without per-path delays.","tokens_in":20332,"tokens_out":3523,"duration_ms":42549,"concrete_test":"Ask the authors to specify and execute their delay-extraction pipeline on one measured 20 MHz CSI snapshot from the meeting-room dataset: apply the proposed procedure (or a standard super-resolution method such as MUSIC/ESPRIT) to obtain the {τ_i} used in Eq. 16, and compare the number and values of extracted delays with the path delays produced by the RT simulator. If the extracted delays do not correspond to the simulated multipath components, then Eqs. 16 and 19 cannot be evaluated as written. As a complementary check, rerun localization with L_p and L replaced by a frequency-domain complex-CSI MSE that does not require per-path delays; if the reported sub-meter errors do not persist, the claimed gains depend on unavailable per-path delay information.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for RayLoc's central claim is that the losses actually used for calibration and localization can be evaluated from the measured CSI. In Eq. 16, L_p sums τ_i 'obtained from the real-world measurements' over Ntot = Nt × Nr × Ns, i.e., over subcarriers and antennas, but a per-path delay τ_i is a property of an individual multipath component and is not directly available from a CSI matrix. The number of physical multipath delays is the number of traced paths Nw, not the CSI dimension Ntot, so the index set in Eq. 16 is dimensionally mismatched. Moreover, the system uses a 20 MHz bandwidth, giving a delay resolution of roughly c/BW ≈ 15 m, so individual paths inside a 3.53 m meeting room cannot be separated by an IFFT or simple peak picking. The paper never states the algorithm that converts the measured 128-subcarrier CSI into the τ_i used in Eqs. 16 and 19, nor does it release code or data. If the intended τ_i are aggregate per-subcarrier phase-slope delays, then the multipath information RayLoc claims to exploit is reduced to one or two resolvable components, which would undercut the central mechanism. Either way, the experimental support in Section 5.4 depends on an unstated and possibly infeasible delay-estimation step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RayLoc, a wireless indoor localization system that frames localization as the inverse of a fully differentiable ray-tracing process. The system calibrates a high-fidelity background scene model from measured CSI, then uses gradient-based optimization with a Gaussian-smoothed loss to estimate target positions. The authors claim that this approach unifies device-free and device-based localization, and they report experimental results in four indoor environments showing lower localization errors than six CSI-based baselines, with ablations supporting the smoothing components.","tokens_in":20678,"tokens_out":4134,"duration_ms":43322,"significance":"If the pipeline is fully specified and reproducible, RayLoc is a noteworthy contribution: it introduces a physically interpretable alternative to fingerprinting and blind parameter estimation, unifies two previously separate localization paradigms, and reports consistent improvements over strong baselines across real-world scenes. The paper also identifies a genuine optimization difficulty--sparse gradients and local minima in the ray-tracing loss landscape--and proposes a concrete mitigation. However, the current manuscript omits several load-bearing implementation details, most importantly how per-path delays are extracted from measured 20 MHz CSI, and it does not release code or data, which limits independent verification.","major_comments":[{"comment":"The loss functions require per-path delays tau_i 'obtained from the real-world measurements', but the index set is Ntot = Nt * Nr * Ns, which counts subcarrier/antenna pairs, not individual multipath components. A single CSI matrix does not directly expose individual path delays; with the stated 20 MHz bandwidth the delay resolution is about 50 ns, i.e., roughly 15 m of path length, so individual multipath components in a room-scale environment cannot be separated by conventional IFFT or peak picking. The paper must specify the exact algorithm that maps the measured 128-subcarrier CSI to the tau_i used in Eqs. (16) and (19), or reformulate the losses using observable aggregate quantities. Without this step, the experimental results in Sec. 5.4 cannot be reproduced or fully interpreted.","section":"Sec. 3.3.2, Eq. (16), and Sec. 3.4.1, Eq. (19)"},{"comment":"The claim that 'discrete decisions, such as testing the validity of paths, are replaced with soft functions' is made in a single sentence and is never defined. Occlusion and visibility changes are exactly the points where the gradient of the simulated CSI with respect to an object position is needed, so a hand-waved smoothing technique leaves the central differentiability claim unsubstantiated. Please provide the concrete soft-visibility model and its derivative, or state the approximation used and its limits.","section":"Sec. 3.2.3"},{"comment":"The gradient formula in Eq. (24) does not follow directly from the smoothed loss in Eq. (23). In Eq. (23) the bias term B = exp(-alpha * L) is declared constant for gradient computation, but Eq. (24) appears to differentiate through L_{u_j} in the product g(u_j)L_{u_j}; moreover the left-hand side, derivative of L with respect to u, is not the derivative of the Monte Carlo estimate in Eq. (23) unless additional identifications are made between u and the sampled positions u_j. Since Eq. (24) is the update rule actually used for localization, a correct derivation or a precise statement of the stochastic estimator is needed.","section":"Sec. 3.4.2, Eqs. (22)-(24)"},{"comment":"The paper asserts, without proof or reference, that convolving the loss with a Gaussian kernel removes plateaus and local minima while preserving the global optimum. This property is load-bearing for the method's convergence claims, and the ablation results are only indirect evidence. The authors should either prove the claim under stated assumptions, or temper the statement and clarify that the improvement is empirical.","section":"Sec. 3.4.2 and Sec. 5.6"}],"minor_comments":[{"comment":"The legend lists 'UbiLocate' as a baseline, but the baseline selection in Sec. 5.2 describes SpotFi, M3, and CiFi; please clarify whether this is a typo or a different baseline.","section":"Fig. 10(b)"},{"comment":"The notation 'softmax(Bc)' with scalar Bc is unusual; softmax is conventionally defined for a vector. Please specify the intended mapping (possibly a sigmoid or a vectorized softmax) and define how Bc is expanded.","section":"Eq. (25)"},{"comment":"The CSI matrix H is described as having size Nt * Nr * Ns, but the displayed expression is indexed only by Nt and Nr with Ns frequency entries; the nesting and orientation should be clarified.","section":"Eq. (12)"},{"comment":"The text mentions 195 measurement locations for the meeting room and later says 'we collect CSI data entries from predefined 143 measurement points'; please clarify which number was used for the calibration experiment.","section":"Sec. 5.3"},{"comment":"There are several typographical issues, including 'T echnologies' in the author affiliation and 'V ary' in the introduction; a proofreading pass is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central idea is timely and the reported results are promising, but the missing specification of the delay-extraction step is a serious reproducibility barrier. I would encourage the editor to request that the authors either provide the complete algorithm, including the mapping from measured CSI to the loss terms, or release the code and data used in the experiments. The paper's fit with the signal-processing audience is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know about RayLoc: it reformulates indoor WiFi localization as an inverse differentiable ray-tracing problem, calibrating a scene model from CSI and then optimizing the target position by gradient descent. The experiments across four real rooms show consistent gains over six baselines, and the ablation makes a fair case that the Gaussian smoothing of the loss landscape is doing real work. If the central mechanism holds, this is a genuine step toward unifying device-free and device-based localization under one optimization framework.\n\nWhat is genuinely new: prior differentiable RT work [40] used the technique for radio network optimization, and [42] for scene calibration; RayLoc is the first to push it to target localization and to address the resulting sparse-gradient problem. That is a meaningful, if incremental, application. The paper does a decent job of motivating the problem with the loss-landscape plots, and the device-based/device-free unification is conceptually clean.\n\nThe soft spots are in the middle, not the edges. The load-bearing issue is the definition of τ_i in the loss functions (Eqs. 16 and 19). The paper says these are 'obtained from the real-world measurements,' but with a 20 MHz bandwidth the delay resolution is about 50 ns (≈15 m), so individual multipath components in a meeting room cannot be separated from the CSI. The index set in Eq. 16 (N_t·N_r·N_s) doesn't match the number of physical paths N_w, and the paper never states the extraction algorithm. Until that is specified, the reported localization accuracy rests on an unverified step. This is fixable—the authors likely have an implementation that works—but as written it is a real gap, not a cosmetic one.\n\nTwo smaller issues: the claim in the intro that no fully differentiable RT simulator with gradients w.r.t. scene geometries exists is contradicted by their own reference [40]; and the visibility-differentiation trick in Section 3.2.3 is hand-waved with a citation rather than explained. Neither sinks the paper, but both need attention.\n\nWho this is for: anyone working on wireless localization, digital twins, or differentiable simulation. The paper deserves a serious referee—it is not desk-reject material. But I would not accept it without the authors clarifying the delay extraction and ideally releasing code/data.","headline":"RayLoc makes a plausible, incremental advance by applying differentiable ray tracing to localization, but the unstated per-path delay extraction undermines the central loss functions as written.","tokens_in":21134,"tokens_out":3138,"would_cite":false,"duration_ms":33033,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By making ray tracing fully differentiable, RayLoc reframes wireless indoor localization as an inverse-rendering problem, recovering target positions from CSI through gradient descent on a calibrated digital scene model.","keywords":["wireless indoor localization","differentiable ray tracing","channel state information","inverse problem","device-free localization","device-based localization","scene calibration","Gaussian smoothing"],"falsifier":"Take CSI measurements with the same 20 MHz, 128-subcarrier setup in the meeting room, convert the frequency-domain CSI to a power-delay profile, and check whether distinct peaks appear at the per-path delays $\\tau_i$ that the calibrated ray tracer predicts. If the roughly 50 ns delay resolution cannot resolve those paths, then the delay terms in Eqs. (16) and (19) have no measurable input from the real CSI, undermining the gradient-based localization claim.","tokens_in":20154,"feed_emoji":"📡","tokens_out":9016,"duration_ms":83860,"temperature":0.7,"pith_summary":"RayLoc's central claim is that wireless indoor localization should be posed as the inverse of a fully differentiable ray-tracing simulation, not as blind estimation of a few propagation parameters from channel state information (CSI). Given a calibrated digital model of the room—geometry, materials, and antennas—the position of a target object or device is found by gradient descent on the mismatch between simulated and measured CSI. The paper argues this matters because conventional CSI methods discard multipath and scene details as noise, while RayLoc turns them into localization signal and unifies device-free and device-based localization in one pipeline. Experiments across four indoor scenes with a single access point show average error reductions of 0.44 m to over 1.17 m compared with six baselines.","feed_headline":"Reverse ray tracing cuts WiFi localization error by up to 1.17 m","feed_subtitle":"RayLoc turns room reflections into position information, locating objects and devices from WiFi CSI in one framework.","key_machinery":"The central object is a fully differentiable ray-tracing simulator built on shooting and bouncing rays, with target positions, triangle-mesh geometry, material conductivity and permittivity, antenna patterns, and transceiver locations all as trainable parameters. Because the path from scene to CSI is differentiable, automatic differentiation supplies $\\partial L/\\partial I$ for the localization loss, and the paper smooths that loss by convolving it with a 2D Gaussian kernel and weighting Monte Carlo samples by a bias toward low-loss positions. The CSI ratio of two antennas cancels the random phase offset in real measurements, SMAPE is used for background calibration, and MSE localizes the target.","core_discovery":"The paper claims that the radio channel itself can be rendered like an image: a shooting-and-bouncing-ray simulator maps a fully parameterized scene to CSI, and localization is the inverse problem of recovering the target position that generated the observed CSI. RayLoc makes every scene parameter trainable, first calibrating background geometry and material electromagnetic properties from measured CSI, then optimizing the target position while a Gaussian-kernel-smoothed loss prevents the optimizer from stalling on plateaus and local minima. With a single three-antenna access point at 5 GHz and 20 MHz bandwidth, the system localizes passive objects (device-free) and active devices (device-based) in a meeting room, laboratory, classroom, and corridor, under both line-of-sight and non-line-of-sight conditions, reducing average localization errors by 0.44 m to over 1.17 m relative to six CSI-based baselines.","pith_inferences":["At the reported 20 MHz bandwidth the delay resolution is about 50 ns, or roughly 15 m of path length, so individual indoor multipath components cannot be separated from the measured CSI; making the per-path-delay losses in Eqs. (16) and (19) operational would require a super-resolution or parametric delay-estimation step that the paper does not describe.","If the inverse-RT formulation is sound, the same differentiable scene graph could be used to estimate other hidden scene properties, such as material parameters or reflector maps, not just positions.","The calibration dataset is the practical bottleneck; replacing the 143-195 measurement points with a one-shot depth-sensor scan, as the paper itself suggests as future work, would be the natural test of whether RayLoc scales beyond curated scenes.","A reader could test the method's generality by re-running the optimization with the Gaussian smoothing removed but wider bandwidth, to separate the contribution of loss smoothing from the contribution of resolvable delays."],"forward_implications":["Localization accuracy should improve as the environment becomes more cluttered, because each reflector is treated as information; the experiments show RayLoc's advantage grows with the number of background objects.","One formulation covers both device-free and device-based localization, since the target position is simply one entry in the scene parameter vector.","The method runs on legacy WiFi hardware: a single three-antenna AP using only 20 MHz of bandwidth is sufficient, lowering deployment cost.","Scene calibration is the enabling step: material properties and object positions converge from measured CSI in roughly 150 iterations, and localization quality depends on this calibration.","Removing the Gaussian smoothing collapses device-free median error to above 1.75 m, indicating that the smoothing is what makes gradient descent tractable."],"supporting_citations":[{"why":"Supplies the differentiable ray-tracing formulation and SMAPE loss normalization that RayLoc adapts for scene calibration.","marker":"[28]"},{"why":"The differentiable ray-tracing simulator that RayLoc extends; cited as lacking differentiability with respect to target positions.","marker":"[27]"},{"why":"Device-based baseline that uses super-resolution angle-of-arrival estimation; RayLoc is compared against it in all scenes.","marker":"[14]"},{"why":"Single-access-point multipath-assisted device-based baseline; one of the six systems RayLoc outperforms.","marker":"[20]"},{"why":"Device-free passive human tracking baseline; the comparison target in the device-free experiments.","marker":"[31]"},{"why":"CSI-ratio-based device-free motion tracking baseline; compared in the device-free experiments.","marker":"[49]"},{"why":"CNN fingerprint-based device-free localization baseline; compared in the device-free experiments.","marker":"[50]"},{"why":"Material electromagnetic property values used to initialize scene parameters before calibration.","marker":"[32]"},{"why":"Source of the CSI-ratio method used to cancel the random phase offset in measured CSI.","marker":"[43]"},{"why":"Tool used to capture the 20 MHz, 128-subcarrier CSI from commercial WiFi hardware in the experiments.","marker":"[48]"}],"fun_headline_variants":["Differentiable ray-tracing turns CSI into precise indoor positions","RayLoc: Backprop through ray-tracing for wireless localization","CSI rendered like an image, then inverted for sub-meter accuracy","WiFi localization via fully differentiable ray-tracing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the per-path delays $\\tau_i$ used in the background-calibration and localization losses can be extracted from real CSI, even though the 20 MHz bandwidth resolves delays only to about 50 ns (roughly 15 m of path length), and the paper does not explain how individual room reflections are separated.","fun_headline_variants_meta":{"raw":{"variants":["Differentiable ray-tracing turns CSI into precise indoor positions","RayLoc: Backprop through ray-tracing for wireless localization","CSI rendered like an image, then inverted for sub-meter accuracy","WiFi localization via fully differentiable ray-tracing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000849,"raw_usage":{"total_tokens":3673,"prompt_tokens":907,"completion_tokens":2766,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":2699}},"tokens_in":523,"tokens_out":2766,"duration_ms":18525,"temperature":1.0,"reasoning_tokens":2699,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:45:40.093717+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take CSI measurements with the same 20 MHz, 128-subcarrier setup in the meeting room, convert the frequency-domain CSI to a power-delay profile, and check whether distinct peaks appear at the per-path delays $\\tau_i$ that the calibrated ray tracer predicts. If the roughly 50 ns delay resolution cannot resolve those paths, then the delay terms in Eqs. (16) and (19) have no measurable input from the real CSI, undermining the gradient-based localization claim.","supporting_citations":[{"cited_title":"WiTraj: Robust Indoor Motion Tracking With WiFi Signals,","cited_arxiv_id":null,"evidence_quote":"CSI-ratio-based device-free motion tracking baseline; compared in the device-free experiments."},{"cited_title":"Learning Radio Environments by Differentiable Ray Tracing,","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable ray-tracing formulation and SMAPE loss normalization that RayLoc adapts for scene calibration."},{"cited_title":"Sionna RT: Differentiable Ray Tracing for Radio Propagation Modeling,","cited_arxiv_id":null,"evidence_quote":"The differentiable ray-tracing simulator that RayLoc extends; cited as lacking differentiability with respect to target positions."},{"cited_title":"SpotFi: Decimeter Level Localization using WiFi,","cited_arxiv_id":null,"evidence_quote":"Device-based baseline that uses super-resolution angle-of-arrival estimation; RayLoc is compared against it in all scenes."},{"cited_title":"M 3: Multipath Assisted Wi-Fi Localization With a Single Access Point,","cited_arxiv_id":null,"evidence_quote":"Single-access-point multipath-assisted device-based baseline; one of the six systems RayLoc outperforms."},{"cited_title":"Widar2.0: Passive Human Tracking with A Single Wi-Fi Link,","cited_arxiv_id":null,"evidence_quote":"Device-free passive human tracking baseline; the comparison target in the device-free experiments."},{"cited_title":"DSCP: Depthwise Separable Convolution-Based Passive Indoor Localization Using CSI Fingerprint,","cited_arxiv_id":null,"evidence_quote":"CNN fingerprint-based device-free localization baseline; compared in the device-free experiments."},{"cited_title":"Effects of Building Materials and Structures on Ra- diowave Propagation Above About 100 MHz,","cited_arxiv_id":null,"evidence_quote":"Material electromagnetic property values used to initialize scene parameters before calibration."},{"cited_title":"FarSense: Pushing the Range Limit of WiFi-based Respiration Sensing With CSI Ratio of Two Antennas,","cited_arxiv_id":null,"evidence_quote":"Source of the CSI-ratio method used to cancel the random phase offset in measured CSI."},{"cited_title":"AX-CSI: En- abling CSI Extraction on Commercial 802.11 Ax Wi-Fi Platforms,","cited_arxiv_id":null,"evidence_quote":"Tool used to capture the 20 MHz, 128-subcarrier CSI from commercial WiFi hardware in the experiments."}],"review_version":1}