{"id":"57e8777d-5a87-4363-b576-a08e9b96e542","arxiv_id":"2505.10816","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A mmWave radar and a low-power Van Atta Array reflecting surface together localize hidden humans with about 9 cm median error and communicate via amplitude-modulated antenna switching.","lead":"mmMirror uses a special reflecting surface (a Van Atta Array IRS) and a standard 24 GHz radar to locate people around corners without them wearing any device. It reports median localization errors around 9 cm at 3 meters and also lets the radar talk to the reflector, cutting scan time by up to 75%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Localization accuracy is only evaluated for targets exactly on the four discrete IRS beam rays; off-beam targets at 3 m incur up to ~39 cm angular mismatch, so the centimeter-level claim is overgeneralized.","rationale":"The reader's weakest assumption centered on the 2D geometric model: same height and the dominant two-segment path via the IRS. That is related, but the more specific and more damaging issue is that the system measures only the range along a preconfigured IRS reflection angle; it has no means of determining the target's actual angular position. The localization equations use the nominal reflection angle α_i, so any target not exactly on one of the four discrete beams will be reported on that beam at the measured distance, producing an error equal to the perpendicular distance from the true position to the beam ray. With 15° beam spacing, this error reaches roughly 39 cm at 3 m, far exceeding the claimed centimeter-level accuracy. The experimental evaluation deliberately aligns targets with the beam angles, so the reported errors reflect range estimation errors only, not the system's true capability for arbitrary target positions. This is load-bearing because it undermines the central claim of device-free NLoS localization: a system that cannot localize a target between its discrete beams is not a general 2D localizer. The reader's height and multipath concerns are secondary and are at least disclosed in Section 5; the off-beam limitation is neither disclosed nor tested. The proposed experiment, placing targets at angular offsets between beams, would directly settle whether the system can maintain accuracy for non-aligned targets. I keep the verdict CONDITIONAL because the paper remains a plausible systems contribution, but the condition should explicitly require off-beam evaluation or a clear scope limitation in the claims.","tokens_in":16684,"tokens_out":8094,"duration_ms":86233,"concrete_test":"Place a target at 3 m from the IRS at an angle exactly midway between two adjacent reflection angles (e.g., 37.5° when beams are at 30° and 45°) and run the mmMirror localization procedure. Repeat for several off-beam angles (e.g., offsets of 3.75°, 7.5°, 11.25° from the nearest beam) at distances of 2 m and 3 m. If the median overall distance error exceeds the claimed 8.93 cm by a substantial margin (e.g., >20 cm), then the reported accuracy only holds for targets aligned with the discrete IRS beams, and the abstract's general 'localization error of 8.93 cm at a 3 m range' should be qualified accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 states that targets were placed 'aligned with the mmMirror beam angle,' and Section 2.4, Eqs. (4)-(5), compute the target position using the nominal IRS reflection angle α_i broadcast by the IRS. The radar obtains only the total path length D_RS + D_ST via 2D-MUSIC; it does not measure the target's angular offset relative to the IRS ray. Therefore, mmMirror is effectively a 1D range measurement along a known ray, not a 2D localization of arbitrary targets. The IRS prototype supports only four discrete reflection angles separated by 15° (Table 2). A target at 3 m from the IRS located midway between two adjacent beams (e.g., 37.5° when beams are 30° and 45°) lies about 39 cm away from the nearest ray, so the position reported by Eqs. (4)-(5) would be off by roughly that amount. The paper's statement that the design 'reduces the maximum angular mismatch error to 10.4 cm' is only consistent with a target distance near 0.8 m, not the 3 m evaluated in Section 4.4. No experiment places a target off the discrete beam angles. Thus the headline accuracy of 8.93 cm at 3 m is demonstrated only for targets exactly on a beam, leaving the general claim of centimeter-level device-free NLoS localization unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes mmMirror, a device-free NLoS localization system built around a 24 GHz FMCW radar and a PCB-based Van Atta Array (VAA) IRS. The IRS retro-reflects to identify itself to the radar, then switches to one of four discrete reflection angles (30 degrees, 45 degrees, 60 degrees, 75 degrees) to redirect radar energy around a corner. The radar estimates the total path length via 2D-MUSIC and combines it with the known IRS position and the broadcast reflection angle to compute the target location (Eqs. 4-5). The paper also contributes an OOK-based radar-to-IRS communication scheme using two transmit antennas and envelope detection, a multi-radar frequency-division approach based on antenna switching frequencies, and an adaptive time-slot allocation algorithm to scan only areas of interest. Experiments in three indoor environments report median overall localization errors of 4.63 cm at 2 m, 6.8 cm at 2.5 m, 8.61 cm at 3 m, 11.28 cm at 3.5 m, and 11.83 cm at 4 m total radar-target distance for single targets, with somewhat worse errors for multiple targets. Communication bit error rates remain below 10^-1 at up to 2.0 m and 60 degrees. The adaptive scanning reduces average scan time from 7.15 s to 1.80 s for a single target.","tokens_in":16953,"tokens_out":7026,"duration_ms":61940,"significance":"If the reported results hold, mmMirror is one of the few working prototypes of device-free centimeter-level NLoS localization with commodity radar hardware. The experimental evaluation uses a real PCB IRS prototype and a COTS radar, and the localization metrics are measured rather than curve-fit to model parameters, which is a strength. The radar-IRS communication design is novel in that it reuses the FMCW signal as the carrier for OOK data, avoiding additional RF chains at the IRS. The adaptive time-slot algorithm is a sensible system contribution for reducing scan time. However, the significance is currently qualified by the fact that all localization experiments place targets exactly on the discrete IRS beam axes; the general 2D localization claim is not yet supported.","major_comments":[{"comment":"The localization method is fundamentally a 1D range measurement along a known IRS ray, not a 2D positioning of an arbitrary target. The radar measures only the total path length D_RS + D_ST via 2D-MUSIC; the target angle is taken to be the nominal IRS reflection angle alpha_i that the IRS broadcasts. Equations (4) and (5) then project D_ST along that fixed ray. The manuscript explicitly states in Section 4.2 that \"we selected four target positions aligned with the mmMirror beam angle,\" and no experiment places a target between beams. For a target at D_ST = 3 m midway between adjacent beams (offset 7.5 degrees), the lateral position error would be about 39 cm, which is an order of magnitude larger than the reported median 3 m error of 8.61 cm. Thus the headline claim of centimeter-level 2D NLoS localization is unsupported for the general case; the paper should either add an interpolation or secondary measurement to determine the target's angular offset, or restrict its claims to localization along predefined beams.","section":"Section 4.2 and Eqs. (4)-(5)"},{"comment":"The statement \"This reduces the maximum angular mismatch error to 10.4 cm\" is only consistent with a target at approximately 0.8 m from the IRS (0.8 m * tan(7.5 degrees) = 10.5 cm). At the distances evaluated in Section 4.4 (IRS-target distances of 1.0-3.0 m), the same 7.5-degree half-beam spacing produces angular mismatch errors of roughly 13-39 cm. The claim is therefore numerically misleading and should be corrected to state the range at which the 10.4 cm figure applies.","section":"Section 3.2"},{"comment":"The radar-to-IRS communication BER reaches about 10^-1 at a distance of 2.0 m and at an angle of 60 degrees. This is a high error rate for a control link that conveys the IRS reflection angle and the area-of-interest commands; a single bit error could cause the IRS to use the wrong beam direction or scan set. The paper does not explain how bit errors are handled (e.g., retransmissions, coding, or acknowledgment), nor whether the localization experiments in Section 4.4 were conducted with error-free communication. Without this information, the end-to-end reliability of mmMirror under realistic operating conditions is not established. The authors should add an error-control mechanism or quantify the impact of communication errors on localization.","section":"Section 3.3 and Figure 14"},{"comment":"The multi-radar demonstration uses antenna switching frequencies of 1 Hz and 2 Hz, which correspond to data rates of about 1 bit per second per radar; even at the maximum prototype switching frequency of 20 Hz, the data rate is only 10 bps with N_r=1. With an 11-bit packet, a single radar-to-IRS command takes more than 2 seconds at the 10 Hz setting used in the single-radar experiments. This communication latency is not included in the scanning-time reduction numbers in Section 4.5 (Table 3), so the claimed 75% latency reduction for single targets does not account for the full system response time. The authors should include the communication overhead in the latency analysis.","section":"Section 3.3 and Figure 7"}],"minor_comments":[{"comment":"The abstract says \"localization error of 8.93 cm at a 3 m range\" while Section 4.4 reports a median overall error of 8.61 cm at a total radar-target distance of 3 m. Please clarify the exact definition of \"range\" (total path vs. IRS-target distance) and reconcile the two numbers.","section":"Abstract and Section 4.4"},{"comment":"The angle conventions in these equations are not fully defined; in particular, the offset (alpha_i - phi) and the sign of the y-term should be derived from the geometry in Figure 2 so that readers can reproduce the coordinate computation.","section":"Section 2.4, Eqs. (4)-(5)"},{"comment":"The FFT threshold for radar identification is described as \"the half of the first peak,\" which may not be robust when multiple radars have unequal received powers or when a radar is much closer than another. A calibration or normalization procedure would strengthen the evaluation.","section":"Section 3.3"},{"comment":"The constraint in line 23 (D_i >= (r_max_energy^4 / r_i^4) * D_max_energy) is introduced without derivation. Please justify the exponent 4 and explain how reflection energy maps to the required number of chirps.","section":"Section 4.5, Algorithm 3"},{"comment":"The assumption that the radar and IRS are at the same height is stated as a limitation, but the paper does not quantify the sensitivity of localization accuracy to height mismatch. A short robustness experiment or simulation would help the reader judge deployability.","section":"Section 5"},{"comment":"The CDFs aggregate over all target distances; per-distance CDFs would be more informative since the error clearly grows with range.","section":"Figure 17"},{"comment":"There are minor typos: \"withe\" in the Algorithm 1 explanation (Section 3.4) and the broken \"˝len(T)\" symbol in Algorithm 3 should be fixed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems paper with a real prototype, and the measured localization results along the beam axes are credible. The main issue is that the title and abstract claim general 2D centimeter-level localization, whereas the evaluation only demonstrates localization for targets lying on the discrete IRS beam rays; the manuscript should either add an off-beam localization mechanism (e.g., beam interpolation or angle estimation at the IRS) or be reframed as localization along predefined beams. I would also ask the authors to add a communication error control analysis and to include the communication latency in the scanning-time comparison. The angular-mismatch claim in Section 3.2 should be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take on mmMirror. It's a real systems paper: a PCB IRS, commodity 24 GHz FMCW radar, and experiments in three environments. What's new is the package: a reconfigurable Van Atta array used for NLoS localization, plus two-TX amplitude encoding that lets the radar talk to the IRS, plus adaptive time-slot allocation. The encoding idea is the cleverest bit, and the paper credits Millimetro for the OOK IRS-to-radar direction. The hardware description is clear enough that someone could reproduce the board. Measured median errors around 8.6 cm at 3 m total path are believable for on-beam targets, and the CDFs across environments are a good-faith evaluation. The latency reduction numbers (7.15 s down to 1.8 s single target) are concrete and useful.\n\nNow the soft spots. The stress-test note is right: localization is only evaluated for targets placed on the discrete beam rays. Section 4.2 says exactly that. Equations (4)-(5) plug in the nominal IRS reflection angle alpha_i broadcast by the IRS; the radar only measures total path length D_RS + D_ST. So for a target off that ray—say halfway between two of the four beams, which are 15 degrees apart—the reported position will be off by roughly half the angular separation times range: about 39 cm at 3 m. The paper says the design reduces maximum angular mismatch error to 10.4 cm, which only holds near 0.8 m. No off-beam experiment is reported. So the headline claim of centimeter-level NLoS localization, as stated in the abstract, overgeneralizes. It is centimeter-level for targets sitting on known rays, which limits it as a true 2D localization system. That is a scope problem, not an invalidation of the prototype.\n\nOther softer spots: BER around 10^-1 at 2 m and 60 degrees is marginal, though for a low-rate control channel it might be tolerable. The multi-radar demo at 1-2 Hz switching is unrealistic for the stated collision scenario, and the 64 Hz MCU sampling caps scalability. No code or data are released, which hurts reproducibility. The paper also does not benchmark against Mosaic, SuperSight, or Metasight in a common indoor setup, so we do not know how it stacks up as a system. The same-height, same-plane geometric assumption is disclosed but load-bearing.\n\nVerdict: this deserves a serious referee. It is a useful systems contribution with honest discussion of its own limitations. The authors should be asked to either evaluate off-beam targets or downgrade the claim to \"along predefined beams,\" and to release artifacts. I would bring it to a reading group for the communication design alone, but I would not build on the localization numbers until the beam-angle question is answered.","headline":"Solid prototype paper with a genuinely new integrated VAA-IRS design, but the centimeter-level claim is only demonstrated for targets sitting on one of four discrete beam angles.","tokens_in":17537,"tokens_out":2135,"would_cite":false,"duration_ms":22711,"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":"mmMirror claims that a single 24 GHz FMCW radar and a Van Atta-array IRS can localize an untagged human around an indoor corner with median errors of 4.63 cm at 2 m, 8.61 cm at 3 m, and 11.83 cm at 4 m total path length.","keywords":["mmWave radar","NLoS localization","Van Atta array","intelligent reflecting surface","FMCW","device-free sensing","indoor localization","human-robot collision avoidance"],"falsifier":"Run the same L-corridor experiment with the radar raised or lowered by 10 cm relative to the IRS while the target stays at 3 m total path length: if the median error remains near 8.6 cm, the coplanar assumption is not doing the work; if it jumps to tens of centimeters, the geometry's load-bearing premise is confirmed. A second check is to place a large metal cabinet near the IRS and see whether 2D-MUSIC still selects the radar-IRS-target path rather than a multipath return.","tokens_in":16466,"feed_emoji":"📡","tokens_out":8923,"duration_ms":81787,"temperature":0.7,"pith_summary":"mmMirror sets out to establish that a person hidden around an indoor corner can be localized without carrying any device, using one off-the-shelf 24 GHz frequency-modulated continuous-wave (FMCW) radar and a low-power reflecting surface built from a Van Atta array. The paper reports median overall distance errors of 4.63 cm at 2 m, 8.61 cm at 3 m, and 11.83 cm at 4 m total radar-to-target path length in L-shaped corridors, and says the system also handles multiple radars, multiple targets in distinct beams, and a 42-75% reduction in scanning time via adaptive time-slot allocation. The reason this matters is that collision avoidance for robots sharing factories and hospitals with people needs sensing that works without line of sight, without bright light, without cameras that raise privacy concerns, and without the person wearing a tag. If the reported accuracy holds, a robot that can see the reflector at a corner can know where a human is on the other side before the human appears.","feed_headline":"Smart reflector lets radar see around corners to ~9 cm","feed_subtitle":"A 24 GHz radar and a Van Atta-array surface track people hidden around corners, no tag needed.","key_machinery":"The load-bearing piece of hardware is a non-retro-reflective Van Atta array: paired patch antennas joined by switched microstrip transmission lines whose lengths set a phase gradient, so the surface reflects an incoming 24 GHz wave at one of four discrete angles (30°, 45°, 60°, 75°) instead of straight back at the radar. With the switches off it reverts to true retro-reflection, which lets the radar locate the IRS and receive OOK-modulated identity and angle bits, and with the switches on it steers energy into the NLoS region. The load-bearing math is the two-segment geometry of Eqs. (2)-(5): given the radar-IRS position from range and angle of arrival, the IRS-target distance from the 2D-MUSIC total path length minus $D_{RS}$, and the broadcast reflection angle $\\alpha_i$, the target position follows directly. The paper's communication and scheduling layers, amplitude-modulated antenna encoding, envelope detection, and greedy area-of-interest time-slot allocation, exist to make that geometry usable in real time, but the VAA steering and the two-segment trigonometry are what carry the localization claim.","core_discovery":"The paper's central claim is that a Van Atta array, normally a passive retro-reflector that sends energy back to its source, can be turned into a controllable non-line-of-sight (NLoS) mirror by switching the lengths of transmission lines between paired antennas, and that this mirror plus a single FMCW radar is enough for device-free localization around a corner. The radar first finds the IRS from its retro-reflected signal and decodes the IRS's identity and current reflection angle via on-off keying (OOK) modulation. It then applies the 2D-MUSIC algorithm to the two-segment reflected path to estimate the total radar-IRS-target length, subtracts the known radar-IRS distance, and uses the trigonometric relations in Eqs. (2)-(5) with the decoded reflection angle to produce target coordinates. The reported outcome is centimeter-level localization for a human moving at 0.5-1 m/s in three indoor environments, with median overall errors of 4.63 cm at 2 m, 8.61 cm at 3 m, and 11.83 cm at 4 m total path length; the abstract cites 8.93 cm at 3 m.","pith_inferences":["If the reported error is dominated by the 15° angular quantization and 2D-MUSIC resolution rather than noise, adding more switched line lengths or a 3D Van Atta array should push accuracy toward the radar's range resolution without changing the architecture.","The 6-bit OOK packet and BER below $10^{-1}$ at 2 m leave room for the radar-IRS link to carry extra data, such as which IRS is active or a sensor reading, at negligible cost.","The reported 183.9 µW average power and 566-day battery estimate suggest the IRS could be deployed untethered at corners, making the system practical for retrofit rather than wired installation.","The coplanar assumption predicts a specific failure mode: vertical misalignment between radar and IRS should inflate error mainly along the y-axis, so measuring error as a function of height offset would give a direct, quantitative test of the model."],"forward_implications":["A robot with line of sight to a corner-mounted IRS can detect and localize an untagged human on the other side before line of sight exists, with median errors of 5-12 cm over 2-4 m total paths.","The IRS and the radar exchange all needed information through the same FMCW waveform, so deployment requires no extra wireless link between them.","Adaptive time-slot allocation lowers average full-scan time from 7.15 s to 1.80 s for one target and 4.13 s for multiple targets, which the paper says is fast enough for its 6.25 s robot-stop collision scenario.","Multiple radars can share one IRS by choosing distinct antenna-switching frequencies, and the IRS supports several targets as long as no two occupy the same reflection angle.","Because the target carries nothing and the radar only sees reflected energy, the same hardware offers a privacy-preserving alternative to cameras in factories, warehouses, and healthcare spaces."],"supporting_citations":[{"why":"Supplies the retro-reflection principle the IRS is built on.","marker":"[35]"},{"why":"Provides the non-retro-reflective pairing and phase-delay method used to steer the reflected beam.","marker":"[3]"},{"why":"Prior VAA retro-reflective localization system that mmMirror extends from LoS tagged targets to NLoS device-free sensing.","marker":"[37]"},{"why":"The commodity 24 GHz FMCW radar used as the prototype's sensing hardware.","marker":"[12]"},{"why":"Recent around-the-corner mmWave imaging baseline that needs heavy compute, contrasted with mmMirror's real-time localization.","marker":"[14]"},{"why":"Supplies the human-robot collision scenario and reaction-time budget used to size the system's latency targets.","marker":"[36]"},{"why":"RF switches that select transmission lines and perform OOK modulation in the IRS prototype.","marker":"[13]"},{"why":"Envelope detector used to demodulate radar-to-IRS amplitude-modulated commands.","marker":"[11]"}],"fun_headline_variants":["Radar sees around corners to 9 cm with smart mirror","Passive mirror gives radar NLoS vision: 9 cm accuracy","Van Atta mirror boosts 24 GHz radar to 9 cm NLoS","No-tag radar + mirror: corner localization to 9 cm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The equations assume the radar, the reflector, and the target all sit in one horizontal plane with the radar and reflector at the same height, and that the strongest path the radar's angle-and-range estimator picks out is exactly the two-segment path that bounces off the reflector; a height mismatch or a competing strong reflector would distort the computed position.","fun_headline_variants_meta":{"raw":{"variants":["Radar sees around corners to 9 cm with smart mirror","Passive mirror gives radar NLoS vision: 9 cm accuracy","Van Atta mirror boosts 24 GHz radar to 9 cm NLoS","No-tag radar + mirror: corner localization to 9 cm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00045,"raw_usage":{"total_tokens":2309,"prompt_tokens":1027,"completion_tokens":1282,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":1206}},"tokens_in":643,"tokens_out":1282,"duration_ms":11663,"temperature":1.0,"reasoning_tokens":1206,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:02:14.207892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same L-corridor experiment with the radar raised or lowered by 10 cm relative to the IRS while the target stays at 3 m total path length: if the median error remains near 8.6 cm, the coplanar assumption is not doing the work; if it jumps to tens of centimeters, the geometry's load-bearing premise is confirmed. A second check is to place a large metal cabinet near the IRS and see whether 2D-MUSIC still selects the radar-IRS-target path rather than a multipath return.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the retro-reflection principle the IRS is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the non-retro-reflective pairing and phase-delay method used to steer the reflected beam."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior VAA retro-reflective localization system that mmMirror extends from LoS tagged targets to NLoS device-free sensing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The commodity 24 GHz FMCW radar used as the prototype's sensing hardware."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the human-robot collision scenario and reaction-time budget used to size the system's latency targets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RF switches that select transmission lines and perform OOK modulation in the IRS prototype."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Envelope detector used to demodulate radar-to-IRS amplitude-modulated commands."}],"review_version":1}