{"id":"a7e4cb01-e2e3-4143-897d-7700beac6f7f","arxiv_id":"2507.05597","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Baton reconstructs missing Wi-Fi Doppler features using temporal and cross-link correlations, achieving a median tracking error of 0.46 m at a 20% communication duty cycle.","lead":"This paper presents Baton, a Wi-Fi tracking system that keeps locating people even when Wi-Fi links are silent most of the time. It fills in missing signal readings by borrowing information across time and across nearby links, and tests it in real indoor settings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 0.46 m median error at 20% CDC is measured under random post-hoc masking of already-computed PLCRs, not under bursty acquisition-level duty cycling; the paper's own continuous-loss tests show large degradation, so the practical CDC claim is not yet established.","rationale":"The reader's weakest-assumption is the neural network's reliance on simulated walking traces. That is a legitimate generalization risk, but the paper's real-hardware experiments across ten subjects and multiple environments provide direct evidence that the trained network transfers to the evaluated settings, at least under the reported missingness protocol. The more load-bearing weakness is the evaluation protocol for the key independent variable: the 20% CDC result is obtained by randomly deleting entries from a fully sampled PLCR matrix, whereas the motivating scenario is intermittent, bursty Wi-Fi traffic. This concern is internal to the paper: Section VI-D itself shows that clustered feature loss substantially degrades accuracy. Because the abstract's headline number is the paper's central quantitative claim, and because the paper does not validate that random masking is representative of acquisition-level duty cycling, this is the least secure condition for the central claim. A targeted bursty-duty-cycle experiment would settle whether the 0.46 m result is an artifact of the masking model or a genuine property of the STAP algorithm. This does not overturn the conditional verdict: the system is well-engineered and the core STAP idea is plausible, but the evidence as presented is not airtight for the practical scenario claimed.","tokens_in":26315,"tokens_out":13979,"duration_ms":172171,"concrete_test":"Re-run the Section VI-B evaluation without post-hoc random masking: configure each of the four Wi-Fi links with a bursty 20% duty cycle (e.g., 200 ms active / 800 ms idle, with both synchronized and unsynchronized link schedules), process the raw CSI through the same STFT pipeline to extract PLCRs only from packets available during active periods, and recompute the median tracking error over the same eight trace shapes and ten subjects. If the median error exceeds roughly 0.65 m or the error CDF shifts materially, the headline 0.46 m result is specific to random feature loss rather than to practical CDC-limited Wi-Fi communication.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline quantitative claim depends on how 'communication duty cycle' is operationalized. In Section VI-D the paper states that for the baseline 'the probability of any element in the PLCR matrix being missing is the same,' and 80% of PLCR values are removed from an already-computed feature matrix. This is a random Bernoulli mask over clean, densely sampled PLCR estimates, not a true 20% duty-cycled link in which no CSI exists during idle periods and PLCRs must be estimated from short, sparse bursts. Real intermittent IoT traffic produces bursty idle intervals; when all links are idle simultaneously, Pred.2 is undefined and the system must rely on stale Pred.1 and model-based Pred.3, while the STFT-based PLCR extraction itself becomes less reliable. The paper's own robustness test in Section VI-D shows that adding only 1 s, 2 s, or 3 s of continuous loss to the random-missing baseline raises mean error from 0.1764 m to 0.3024 m, 0.6702 m, and 0.7125 m respectively, demonstrating that the system is sensitive to the missingness pattern. Unless the 0.46 m result is reproduced under a bursty, acquisition-level 20% duty cycle, the central claim that Baton accurately tracks under practical 'severe Wi-Fi feature deficiencies' is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Baton, a device-free Wi-Fi tracking system designed to maintain tracking accuracy when many Wi-Fi features are missing, motivated by low communication duty cycles (CDCs) in IoT-style traffic. The system extracts path length change rates (PLCRs) from CSI and fills missing PLCR values by combining three prediction mechanisms: observation-based prediction (Pred.1) using the most recent observed value on the same link, proportionate prediction (Pred.2) using cross-link PLCR ratios, and model-based prediction (Pred.3) using a Fresnel-zone model fed by the current trajectory estimate. The predictions are integrated using a reliability-weighted scheme, and the tracking and prediction steps are iterated in an incremental loop (STAP). A neural network trained on simulated walking traces provides initial trajectory estimates. The authors implement Baton on commercial Intel 5300 Wi-Fi devices and evaluate it in four environments, reporting a median tracking error of 0.46 m at a 20% communication duty cycle and large improvements over NNE-Tracking and WiTraj under random PLCR removal.","tokens_in":26573,"tokens_out":3715,"duration_ms":46453,"significance":"If the central claims are correct, Baton would be an important step toward practical ISAC: it would show that device-free tracking can tolerate losing most Wi-Fi measurements through a combination of temporal and cross-link compensation, and it offers a plug-and-play mechanism that could be layered on existing data-driven trackers. The work has clear strengths: it is implemented on real COTS hardware, evaluated in multiple physical environments including an NLoS scenario, compared against two relevant baselines, and the main evaluation is anchored to external ground truth, so the headline result is not circular. The self-correction observation in Fig. 19 and the careful treatment of the all-links-missing case are also valuable. However, the practical CDC claim is weakened by the evaluation protocol, which removes PLCR values randomly after feature extraction rather than emulating bursty acquisition-level duty cycling, and the evaluation statistics are thin for the strength of the claims. The simulated-training dependence also adds a sim-to-real risk.","major_comments":[{"comment":"The headline 0.46 m median error at 20% CDC is obtained by randomly removing 80% of the entries of an already-computed PLCR matrix, not by running the system under a bursty acquisition-level 20% duty cycle. In §VI-D the paper states that 'the probability of any element in the PLCR matrix being missing is the same,' which is a Bernoulli mask over clean, densely sampled PLCR estimates. Real intermittent IoT traffic produces idle intervals during which no CSI exists; in such intervals STFT-based PLCR extraction itself becomes unreliable, Pred.2 is undefined when all links are idle, and the system must rely on stale Pred.1 and model-based Pred.3. The paper's own continuous-loss robustness test in §VI-D shows that adding 1 s, 2 s, or 3 s of continuous loss to the random-missing baseline raises mean error from 0.1764 m to 0.3024 m, 0.6702 m, and 0.7125 m, respectively. Thus the central claim that Baton accurately tracks under practical severe Wi-Fi feature deficiencies is not yet established; the 0.46 m result needs to be reproduced under a bursty, acquisition-level duty-cycle protocol.","section":"§VI-D"},{"comment":"The tracking-stage neural network is trained exclusively on simulated walking traces generated with randomized step lengths, speeds, and turning ranges. The paper asserts that these traces are 'remarkably comparable' to real trajectories, but it provides no quantitative validation of the simulated trace distribution against the real walking data used in evaluation. This is load-bearing because the entire STAP loop depends on the network producing a reliable initial trajectory from the first Nf observation-based PLCR rows; any systematic difference between simulated and real human locomotion will bias the initial trajectory, the derived velocity, and every model-based Pred.3 value, and it is not clear that the self-correction in Fig. 19 can recover from such a bias. The authors should provide a sim-to-real analysis, for example by comparing trajectory statistics or by evaluating with a network trained on a held-out portion of real traces.","section":"§V-C, Fig. 9"},{"comment":"The quantitative evaluation is based on only five repetitions per trace shape, and the paper reports median or mean errors without confidence intervals or statistical significance tests. For example, the improvement percentages in Table I and the overall medians in Fig. 16a are point estimates; the non-monotonic ordering in Fig. 16a (0.47 m at 40%, 0.50 m at 30%, 0.46 m at 20%, 0.65 m at 10%) further suggests that the differences between conditions may be within the noise. Additionally, the key parameters Nf, Tw, and the quadratic reliability weight in Eq. (9) are chosen empirically, and no sensitivity analysis with error bars is reported for these choices. Given the strength of the 'first system' and 'reduces error by 79.19%' claims, the evaluation needs more repetitions or confidence intervals to support them.","section":"§VI-B, Table I, Fig. 16"},{"comment":"The STAP loop is a feedback cycle: the trajectory is produced by a neural network fed with predicted PLCRs, and the model-based PLCR predictions (Pred.3) are computed from that same trajectory. In the all-links-missing case, the final PLCR prediction is w·Pred.1 + (1-w)·Pred.3, so errors in the trajectory estimate are fed back into the PLCR predictions and then back into the next trajectory estimate. The paper's empirical self-correction result in Fig. 19 is encouraging, but it is not accompanied by any analysis of when this feedback loop converges versus diverges; the non-uniform motion experiment in §VI-E shows instability at 20% CDC. A stability or error-propagation analysis, or at least a diagnostic of per-iteration error for failure cases, would materially strengthen the claim that the algorithm is robust rather than merely tuned to the tested traces.","section":"§V-B, Algorithm 1"}],"minor_comments":[{"comment":"The abstract reports a median tracking error of 0.46 m at 20% CDC, while Fig. 16a reports median errors of 0.47 m, 0.50 m, and 0.46 m at 40%, 30%, and 20% CDC respectively; the non-monotonic relationship should be explained or reconciled.","section":"Abstract/§VI-B"},{"comment":"In Eq. (5) the coefficient is written as kBC in both the second and third expressions; the third should presumably be kAC or another distinct symbol to match the text.","section":"§III-B, Eq. (5)"},{"comment":"The reliability weight w is defined for '0 < t < Tw' and 't ≥ Tw', but the units of t (time slots versus seconds) and the discrete-time nature of the weight are not specified; clarifying this would help reproducibility.","section":"§V-A.3, Eq. (9)"},{"comment":"The set of raw PLCR values in the PLCR matrix is said to be set to 0 when missing, but PLCR is a signed quantity and 0 is also a valid measured value; the use of 0 as a missing-value sentinel should be stated explicitly or replaced with a NaN/flag mechanism.","section":"§V-A.1"},{"comment":"The paper states that ten PLCR values are extracted per second and that the tracking result is updated ten times per second, but the transmitter sends at 1000 Hz; the relationship between these rates and the STFT window/hop length is not described, making it hard to reproduce the feature extraction.","section":"§VI-B"},{"comment":"No data or code release is mentioned, and the experimental protocol does not include a statement about randomized ordering or blinding; a reproducibility appendix or a link to the implementation would improve the manuscript.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper has already appeared in IEEE TMC, and the core algorithmic idea is interesting and potentially publishable. My main concern is that the evaluation protocol does not match the paper's practical claim: random masking of extracted PLCRs is not equivalent to a 20% communication duty cycle, and the paper's own continuous-loss results show substantial degradation. This is fixable with new experiments under bursty or scheduled idle patterns, along with a more rigorous statistical comparison and a sim-to-real sanity check. I would not reject the work, but I would require those additions before treating the headline claim as established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nBaton deserves a look, but read the headline with a grain of salt. The core idea is genuinely new: instead of spatial interpolation or assuming continuous links, it simultaneously tracks and predicts, using temporal continuity and cross-link proportional correlation of PLCRs to fill in missing features. The STAP loop (observation-based, proportionate, and model-based predictions weighted by reliability) is a reasonable and well-engineered response to the problem of intermittent Wi-Fi traffic. The paper gives real hardware experiments, multiple indoor environments including obstacles and NLoS, ten subjects, and comparisons against NNE-Tracking and WiTraj. The PLCR ratio stability observation is backed by both empirical data and a simple Fresnel-zone derivation. That is real work and worth taking seriously.\n\nThe soft spot is the operationalization of the headline 20% duty-cycle claim. Section VI-D reveals that missing PLCR values are randomly removed from an already-computed feature matrix; that is a Bernoulli mask over dense observations, not a true duty-cycled link where no CSI exists during idle periods. Real IoT traffic is bursty, with all links idle at once. The paper's own robustness tests show the system is sensitive to the missingness pattern: adding 1s, 2s, or 3s of continuous loss to the random-missing baseline raises mean error from 0.18 m to 0.30, 0.67, and 0.71 m. That is a large degradation, and it directly undercuts the practical reading of the 0.46 m result at 20% CDC.\n\nOther weaknesses are real but minor in comparison: no code or data release, only five repetitions per trace shape with no confidence intervals, and Nf plus reliability weights tuned empirically. The neural network for initial trajectory is trained purely on simulated walking traces; if the step-and-turn statistics mismatch the deployment environment, the initial velocity estimate is biased. The paper acknowledges several of these limitations in the parameter and robustness sections, which is honest.\n\nOverall, the central idea is sound and the engineering is credible, but the evaluation does not yet establish the headline claim under realistic intermittent traffic. A reader working on ISAC or device-free tracking should know Baton and can reuse its compensation logic, but should re-test it under bursty acquisition-level duty cycling before believing the 0.46 m number. This deserves serious referee engagement, ideally with requests for code/data and those bursty experiments.","headline":"Genuinely new compensation idea, but the headline 20% duty-cycle result is measured under random feature dropout, not bursty traffic; the paper's own continuous-loss tests show the practical claim is not yet nailed down.","tokens_in":27117,"tokens_out":2534,"would_cite":true,"duration_ms":27635,"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":"This paper claims that Wi-Fi device-free tracking can keep working when 80% of signal features are missing, with a median tracking error of 0.46 m at a 20% communication duty cycle.","keywords":["Wi-Fi sensing","device-free tracking","channel state information","path length change rate","feature compensation","communication duty cycle","integrated sensing and communication","STAP"],"falsifier":"Retrain the tracking network on simulated traces with a very different motion model (e.g., constant speed, no turns) and evaluate on turn-heavy walks at CDC=20%; if median error rises much above 0.46 m, the simulated-training statistics are load-bearing. Alternatively, measure PLCR ratios between two links during rapid back-and-forth pivots and check whether the half-second-ratio stability assumption breaks.","tokens_in":26091,"feed_emoji":"📶","tokens_out":6848,"duration_ms":71016,"temperature":0.7,"pith_summary":"This paper argues that device-free Wi-Fi tracking can be made practical even when transceivers communicate only intermittently, so that a large share of sensing features is simply absent. Baton claims to be the first system to track a person accurately under such severe feature deficiencies, reporting a median tracking error of 0.46 m at a 20% communication duty cycle, compared with multi-meter errors for existing systems in the same conditions. The argument rests on two observed regularities in the path length change rate (PLCR) of reflected signals: values in one link change smoothly over short time intervals, and the ratio between two links stays roughly constant. Baton's STAP algorithm fills the missing entries by combining these regularities with a model-based prediction, then iteratively re-tracks the trajectory so that early errors shrink as the loop proceeds.","feed_headline":"Wi-Fi tracking works with 80% of signals missing","feed_subtitle":"Baton's STAP loop fills in missing link data, hitting 0.46 m median error at a 20% duty cycle.","key_machinery":"The load-bearing object is the PLCR (path length change rate), the derivative of the reflected-path length, equal up to a sign to the Wi-Fi Doppler frequency shift. Two observations carry the argument: within a half-second window a link's PLCR stays nearly constant, and across two links the PLCR ratio remains stable because both scale with the normal component of the human velocity. STAP converts these into three predictors and fuses them: Pred.1 repeats the most recent observed PLCR, Pred.2 multiplies a known link's value by the stable cross-link ratio, and Pred.3 computes the PLCR from the Fresnel-zone geometry of the current position and velocity. A reliability matrix assigns a quadratically decaying weight to Pred.1 over time, and a BiLSTM neural network, trained on simulated walking traces, maps the first rows of PLCRs into the initial trajectory that seeds the tracking loop.","core_discovery":"On its own terms, the paper claims that missing Wi-Fi features do not have to be recovered from raw packets; they can be re-created from the structure of the features that are present. The central discovery is that the PLCR matrix has two usable dimensions of correlation—temporal continuity within each link and a stable proportional relationship across links—and that these can be exploited simultaneously. The paper packages this into the STAP algorithm, which alternates between predicting missing PLCRs and updating the trajectory, weighting the three prediction sources by a reliability matrix that decays trust in stale observations. The empirical payoff is a median tracking error of 0.46 m at CDC=20%, a 79.19% error reduction over the strongest prior work it compares against.","pith_inferences":["If the cross-link PLCR ratio stability holds beyond the tested layouts, the same temporal-and-proportional compensation could be applied to other CSI-derived quantities or other RF bands, since the derivation only assumes path geometry and human inertia.","A sharper test of the simulated-training assumption would be to retrain the network on measured walking traces from one environment and deploy in another; the paper's reported error is likely sensitive to the step-and-turn distribution.","The paper's own robustness result—unstable tracking at CDC 20% under sudden speed changes—suggests the compensation works best for walkers whose motion obeys smooth inertia; extending STAP to erratic motion would require a stronger model for velocity changes."],"forward_implications":["A smart home can deploy Wi-Fi sensing without dedicating continuous packet streams to it: Baton's reported performance at 20% duty cycle means sensing traffic can share airtime with normal IoT communication.","Because the STAP loop corrects its own earlier errors, the final trajectory is more accurate than the initial network output, so the system does not need a perfect first guess.","The compensation method is modular: Baton can be placed on top of existing trackers such as NNE-Tracking and HyperTracking, extending them to low-duty-cycle and NLoS conditions.","Tracking degrades gracefully rather than collapsing as data disappears: median errors stay under 0.65 m down to a 10% duty cycle, and using two receivers instead of four only raises the error to 0.59 m."],"supporting_citations":[{"why":"Supplies the PLCR feature extraction from CSI (STFT) and the velocity-monitoring model that Baton builds on.","marker":"[22]"},{"why":"Establishes the single-link tracking model and the assumption that reflection angles stay constant over short intervals, used in the ratio derivation.","marker":"[40]"},{"why":"Provides the state-of-the-art neural-network tracking baseline that Baton compares against and improves by 79.19% in severe feature deficiency.","marker":"[43]"},{"why":"Provides the model-based DFS tracking baseline (WiTraj) used as a second comparison in the evaluation.","marker":"[57]"},{"why":"Supplies the hyperbolic NLoS model that Baton adapts to extend compensation to non-line-of-sight scenarios.","marker":"[47]"},{"why":"Gives the conjugate-multiplication method for removing initial phase offset from raw CSI before PLCR extraction.","marker":"[56]"}],"fun_headline_variants":["Wi-Fi tracking survives 80% signal loss","Baton's STAP cuts sparse Wi-Fi tracking error 79%","Fill in Wi-Fi gaps: Baton tracks anyway","Sparse Wi-Fi? Baton reconstructs missing features","Track at 20% Wi-Fi duty cycle with Baton"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything downstream depends on the simulated-walking training set: if the simulator's step lengths, speeds, and turning statistics do not match real locomotion, the initial trajectory is biased and the self-correction loop cannot fully recover, and the paper's own robustness test already shows instability at CDC 20% when people suddenly accelerate or brake.","fun_headline_variants_meta":{"raw":{"variants":["Wi-Fi tracking survives 80% signal loss","Baton's STAP cuts sparse Wi-Fi tracking error 79%","Fill in Wi-Fi gaps: Baton tracks anyway","Sparse Wi-Fi? Baton reconstructs missing features","Track at 20% Wi-Fi duty cycle with Baton"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000607,"raw_usage":{"total_tokens":2824,"prompt_tokens":937,"completion_tokens":1887,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1805}},"tokens_in":553,"tokens_out":1887,"duration_ms":19777,"temperature":1.0,"reasoning_tokens":1805,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:21:59.445440+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the tracking network on simulated traces with a very different motion model (e.g., constant speed, no turns) and evaluate on turn-heavy walks at CDC=20%; if median error rises much above 0.46 m, the simulated-training statistics are load-bearing. Alternatively, measure PLCR ratios between two links during rapid back-and-forth pivots and check whether the half-second-ratio stability assumption breaks.","supporting_citations":[{"cited_title":"Widar: Decimeter- level passive tracking via velocity monitoring with commodity wi-fi,","cited_arxiv_id":null,"evidence_quote":"Supplies the PLCR feature extraction from CSI (STFT) and the velocity-monitoring model that Baton builds on."},{"cited_title":"Widar2. 0: Passive human tracking with a single wi-fi link,","cited_arxiv_id":null,"evidence_quote":"Establishes the single-link tracking model and the assumption that reflection angles stay constant over short intervals, used in the ratio derivation."},{"cited_title":"Nne-tracking: A neural network enhanced framework for device-free wi-fi tracking,","cited_arxiv_id":null,"evidence_quote":"Provides the state-of-the-art neural-network tracking baseline that Baton compares against and improves by 79.19% in severe feature deficiency."},{"cited_title":"Witraj: Robust indoor motion tracking with wifi signals,","cited_arxiv_id":null,"evidence_quote":"Provides the model-based DFS tracking baseline (WiTraj) used as a second comparison in the evaluation."},{"cited_title":"Hypertracking: Exploring the hyperbolic model for non-line-of-sight device-free wi-fi tracking,","cited_arxiv_id":null,"evidence_quote":"Supplies the hyperbolic NLoS model that Baton adapts to extend compensation to non-line-of-sight scenarios."},{"cited_title":"Wiborder: Precise wi-fi based boundary sensing via through-wall dis- crimination,","cited_arxiv_id":null,"evidence_quote":"Gives the conjugate-multiplication method for removing initial phase offset from raw CSI before PLCR extraction."}],"review_version":1}