{"id":"fe71c3cb-a48d-4277-8980-686e3059994a","arxiv_id":"2504.14495","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RadarTrack derives a fourth-order kinematic equation from mmWave radar phase returns and takes the most common root to estimate ego-vehicle speed, achieving median MAE around 2 cm/s across UGV, UAV, and handheld platforms.","lead":"RadarTrack estimates a radar-equipped vehicle's own speed by tracking tiny phase changes in reflected signals, solving a math equation derived from motion physics instead of using neural networks. A smart generalist might read it because it promises accurate, low-power speed sensing for small robots, drones, and phones using a single cheap radar chip.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sub-doppler claim is not self-contained: static/dynamic segmentation and the angle α feeding Eq. (5) are computed from Doppler radial velocities that quantize to zero at 1.5 cm/s, so the phase-only low-speed result uses an unstated input.","rationale":"I read the paper's central claim as: a purely signal-processing, radar-only pipeline recovers ego-speed from phase changes, with the key differentiator being sub-doppler operation. The derivation of Eq. (5) is internally consistent after the sign correction in Eq. (4) noted by the reader, and the root-mode idea is plausible. The weakest point is not the algebra but the availability of its inputs below the Doppler resolution: α and the static/dynamic mask are Doppler-derived, exactly where Doppler is quantized. This is more pointed than the constant-velocity assumption: that assumption limits UAV/handheld accuracy (MAE 0.15–0.16 m/s) but does not threaten the UGV headline. The sub-doppler dependency threatens the UGV headline itself. The reader's static-majority concern is related, so I mark partial agreement. I would keep the verdict CONDITIONAL rather than REJECT because the authors could resolve the issue by releasing code/data or specifying a non-Doppler source for α; until then the claim is not fully supported. UNCHANGED in the verdict field means my read leaves the reader's CONDITIONAL recommendation in place.","tokens_in":12257,"tokens_out":10876,"duration_ms":99707,"concrete_test":"Re-run the Fig. 8(b) low-speed experiment (≈1.5 cm/s) with instrumentation that logs vri, \\hat α_{i,j}, the selected static set S, and the roots of Eq. (5). Then recompute the speed estimate twice: once with α from the Doppler-based Eq. (2) and once with α from Vicon ground truth. If the root-mode MAE changes by more than ~1 cm/s, or if Eq. (2) yields \\hat α values with error >5° in more than 10% of frames, the reported sub-doppler accuracy depends on an input the paper does not specify; if the two runs agree, the concern is resolved and the method is self-contained.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ego-speed can be recovered from phase alone below Doppler resolution, with static objects at oblique angles. The load-bearing gap is in the input chain, not the quartic algebra. Sec. III-C segments static/dynamic points using \\hat v_bi = vri / cos(θi − α) and estimates α by Eq. (2) from pairwise ratios vri/vrj. Both use PCD Doppler radial velocities vri. With the stated radial-speed resolution of 0.0496 m/s, the paper's own low-speed demonstration at ≈1.5 cm/s (Fig. 8(b)) is below one Doppler bin, so vri values are zero or ±1 bin; the pairwise ratios in Eq. (2) are undefined or noise-dominated, and Eq. (1) cannot separate static from dynamic reflectors. However, Eq. (5) is not phase-only: coefficients b, c, d depend on K(α) = r0 cos(θ−α), and K comes from the same α. The paper does not state how α is initialized or maintained during the 1.5 cm/s trial, and no Vicon/inertial feedback is described in the pipeline. If α is fixed to a known mounting direction or estimated from previous high-speed frames, that is an unstated input; if it is truly Doppler-estimated per frame, the sub-doppler result lacks a reliable α. In either case, the abstract's 'phase-based ... solely based on signal processing' claim is not established for the regime that motivates the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"RadarTrack proposes to estimate ego-vehicle speed from a single-chip mmWave radar using the phase variations of static reflectors. The pipeline first separates static from dynamic radar points using Doppler radial velocities and an estimated angle alpha (Sec. III-C), selects the corresponding range bins, and then, for each static point and chirp, derives a fourth-order polynomial in ego-speed from the kinematics of the range evolution and takes the mode of the polynomial roots across the frame (Sec. III-D). The system is implemented on a TI IWR1843 radar with a Jetson Nano and evaluated on a UGV, a UAV, and a handheld stick against Vicon ground truth, with comparisons to Doppler-only, IMU, MilliEgo, and Radarize baselines, plus resource and latency measurements.","tokens_in":12495,"tokens_out":5988,"duration_ms":48000,"significance":"If the claims hold, RadarTrack would be a valuable lightweight, DNN-free radar-only speed estimator for micro-robots, AR, and other resource-constrained platforms, and its closed-form analytic model is a refreshing alternative to learned odometry. The paper's strengths include a transparent kinematic derivation in Sec. III-D, a real hardware prototype with Vicon ground truth across three platforms, direct comparison against four baselines, and explicit reporting of power, CPU, memory, and latency. The central derivation is almost correct, but the printed Eq. (4) contains a sign error, the sub-Doppler claim currently relies on an unstated source for the angle alpha, and the evaluation narrative contains internally inconsistent error numbers. These issues are addressable, which is why I recommend major revision rather than rejection.","major_comments":[{"comment":"Eq. (4) prints r_t = sqrt(r_0^2 - 2 v_b t K(alpha) - v_b^2 t^2), but expanding r(t)^2 = (X_0 - v_b cos(alpha) t)^2 + (Y_0 - v_b sin(alpha) t)^2 gives a + v_b^2 t^2 term, not a minus term. The coefficients of Eq. (5) correspond to the corrected plus sign; with the printed minus sign, differentiating Eq. (4) would give different coefficients, including a v^2 coefficient of K^2 + t^2 c^2 Theta^2 instead of the stated K^2 - t^2 c^2 Theta^2. Please correct Eq. (4) and add a sentence showing the consistency check that leads to Eq. (5), since this is the load-bearing derivation of the paper.","section":"Sec. III-D, Eq. (4)"},{"comment":"The static/dynamic segmentation in Eq. (1) and the angle alpha in Eq. (2) both use the PCD radial velocities v_ri. With the stated radial-speed resolution of 0.0496 m/s, the paper's headline low-speed demonstration at about 1.5 cm/s (Fig. 8(b)) is below one Doppler bin, so v_ri values are quantized to zero or plus/minus one bin. At v_ri = 0, the ratio in Eq. (2) is undefined and Eq. (1) cannot separate static from dynamic points; at plus/minus one bin, the ratio is noise-dominated. The paper does not state how alpha is initialized or maintained during the 1.5 cm/s trial. If alpha is fixed to a known mounting direction or taken from previous high-speed frames, that is an unstated input; if it is estimated per-frame from Doppler, the sub-Doppler result does not have a reliable alpha. In either case, the abstract's claim of a phase-based method 'solely based on signal processing' is not established for the regime that motivates the paper. Please specify the source of alpha in each experiment and add an ablation that runs the estimator without Doppler-based alpha.","section":"Sec. III-C and Sec. III-D, Doppler dependence at sub-Doppler speeds"},{"comment":"The text says that RadarTrack achieves 'approx 5% of the base speed, on average', but the numbers reported in Sec. V-B and Fig. 6 are inconsistent with that claim. The UGV MAE of 0.02 m/s at a 0.25 m/s baseline is 8%, the UAV MAE of 0.15 m/s at a 0.21 m/s baseline is 71%, and the handheld MAE of 0.16 m/s at a 0.35 m/s baseline is 46%. The median MAE of about 2 cm/s in Fig. 6 also does not correspond to a single 'base speed'. Please report relative errors with a clearly defined denominator, or remove the '5%' claim.","section":"Sec. V-A, error summary"},{"comment":"Eq. (5) is a quartic, so for each static point and chirp it can produce up to four roots, and some roots may be complex or negative. The paper does not state how complex or nonphysical roots are filtered before the mode is taken, nor whether the mode is computed over all N x N_c x 4 roots or only over the real, positive subset. Without this rule, a single spurious repeated root could dominate the mode. Please specify the root-filtering procedure and report the distribution of root multiplicities across a representative frame.","section":"Sec. III-D, root-mode selection"}],"minor_comments":[{"comment":"The sentence describing the integration is imprecise: Eq. (4) follows by directly expanding r(t)^2 and then taking the square root, not by integrating Eq. (3) from t = 0 to t and from r_0 to r_t. Please rephrase.","section":"Sec. III-D, Eq. (4)"},{"comment":"The 'mode' is used for continuous quantities in Eqs. (1) and (2) and in the final root mode; please specify the bin width or kernel used to compute the mode in each case.","section":"Sec. III-C, Eq. (1)"},{"comment":"The caption says that negative speed represents the ego-vehicle approaching a target, but the vertical axis in both panels is labeled with positive velocity values from 0 to 0.4 m/s. Please reconcile the sign convention with the plotted data.","section":"Fig. 8 caption"},{"comment":"The statement that RadarTrack 'utilizes only 84% of CPU' is surprising for a method presented as lightweight; please clarify whether this is single-core utilization, how it is measured, and whether the comparison with Radarize and MilliEgo was made on the same hardware with the same measurement period.","section":"Sec. V-E, resource consumption"},{"comment":"At 10 frames per second the frame time is 100 ms, yet the reported latency is about 0.29 s. Please clarify whether this includes the 150 ms frame accumulation, the processing time, or both, and how the baseline latencies of 3 s and 5 s were measured.","section":"Sec. V-E, latency"},{"comment":"Please report the number of trials, standard deviations, and trajectory lengths for each reflector configuration, since the MAE differences in Table II are small and may not be statistically significant.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a robotics or mobile sensing venue. The main risk is that the 'phase-only' claim is overstated because the front end of the pipeline is Doppler-dependent; if the authors can clarify how alpha is obtained in the sub-Doppler regime and validate the low-speed result with an independent alpha source, the contribution could be acceptable. The related-work positioning should also engage more carefully with Ding et al. [23], which already addresses ego-motion estimation with single-chip mmWave radar in dynamic scenes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is genuinely new, but the sub-doppler claim isn't self-contained. Deriving a fourth-order polynomial in ego-speed from the phase history of an oblique static point, then taking the mode of roots across chirps, is a neat contribution and a real extension over [17]'s boresight-only restriction. The hardware work on UGV/UAV/handheld with a COTS radar is also useful, and the latency and power numbers are believable.\n\nThe load-bearing problem is upstream of the quartic. The static/dynamic segmentation (Eq. 1) and the angle alpha (Eq. 2) both use radial velocities vri from the Doppler FFT. With the stated resolution of 0.0496 m/s, the paper's own 1.5 cm/s demonstration is below one bin, so those vri values are zero or a single bin. Pairwise ratios are then undefined or noise-dominated, and alpha cannot be recovered from them. The paper never says how alpha is initialized or maintained in that regime. It might be held at a known mounting angle, or estimated from earlier higher-speed frames, but that would be an unstated input and would break the 'phase-only' claim. This is a gap the authors need to close, not a minor omission.\n\nSmaller issues: Eq. (4) has a sign typo (the v_b^2 t^2 term should be plus, and Eq. (5)'s coefficients match the corrected version), which is repairable but sloppy. The 'about 5% average error' statement conflicts with the reported UAV and handheld MAEs around 0.15–0.16 m/s at baselines near 0.2–0.35 m/s; the 2 cm/s median in Fig. 6 likely just reflects a UGV-heavy dataset. No code or data is released, and the baseline training for Radarize/MilliEgo is underspecified. The two-dynamic-object limitation is acknowledged in the text, but the abstract's claim about handling dynamic environments is stronger than the evidence.\n\nI'd engage with this: the core idea deserves referee time and the flaws are fixable. I'd ask for major revision, a clear statement of how alpha is obtained in the sub-doppler regime, and artifacts before it's publishable. As is, I wouldn't cite it, and I'd keep it out of the reading group until the alpha question is answered.","headline":"The quartic phase-based speed estimator is a real idea, but the sub-doppler claim isn't self-contained: the pipeline still depends on Doppler-derived quantities that quantize to zero at the exact speeds the method targets.","tokens_in":13113,"tokens_out":4577,"would_cite":false,"duration_ms":41589,"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":"RadarTrack recovers ego-vehicle speed from mmWave radar phase changes alone, solving a fourth-order kinematic equation and reaching about 2 cm/s median error without any neural network.","keywords":["ego-speed estimation","mmWave radar","phase-based odometry","sub-Doppler resolution","static-dynamic segmentation","real-time embedded systems","kinematic polynomial roots","single-chip FMCW radar"],"falsifier":"Mount the radar on a linear stage that moves at a constant 2 cm/s and carries a high-rate encoder as ground truth, with a single static reflector at 45 degrees; if the modal fourth-order root deviates systematically from the encoder speed once pitch oscillation is added, the constant-velocity phase model is disconfirmed.","tokens_in":11977,"feed_emoji":"📡","tokens_out":7746,"duration_ms":64393,"temperature":0.7,"pith_summary":"RadarTrack claims that an ego vehicle's speed can be recovered from the phase of single-chip mmWave radar returns alone, without Doppler bins, IMUs, or neural networks. The paper derives a fourth-order polynomial in ego-speed from the round-trip geometry of a static reflector, then takes the most frequent root across chirps and static points as the speed estimate. Because the estimate uses phase rather than the Doppler-FFT bin, it resolves speeds below the radar's Doppler resolution and works for static objects seen at oblique angles. If the claim holds, cheap, low-latency ego-speed estimation becomes available for micro-robots, drones, and handheld devices in cluttered indoor scenes.","feed_headline":"Phase-only mmWave radar speed hits ~2 cm/s median error","feed_subtitle":"No neural network, no IMU: RadarTrack solves a kinematic equation from radar phase, beating Doppler at slow speeds.","key_machinery":"The carrying object is the fourth-order kinematic phase equation, $a(t)v_b^4 + b(t,\\alpha)v_b^3 + c(t,\\alpha)v_b^2 + d(t,\\alpha)v_b + e(t) = 0$, derived from the time-dependent range to a static reflector with coefficients $a(t)=t^2$, $b(t,\\alpha)=-2K(\\alpha)t$, and so on, where $K(\\alpha)=Y_0\\cos\\alpha + X_0\\sin\\alpha$. It converts chirp-level phase changes into candidate speeds, and the mode of the candidate roots over static points and chirps is the advertised ego-speed. The supporting mechanism is a Doppler radial-velocity classifier that labels static versus dynamic points, plus phase unwrapping that keeps $\\Delta\\Phi < 2\\pi$, bounding the unambiguous speed at about 7.75 m/s.","core_discovery":"The paper's central claim is that the phase trace of static reflectors carries enough information to determine ego-speed exactly: differentiating $r(t)^2 = X(t)^2 + Y(t)^2$ under constant ego-velocity produces a fourth-order equation in $v_b$ whose coefficients involve the measured phase derivative $\\Theta(t) = d\\Phi/dt$ and the heading offset $\\alpha$. Each static point and each chirp yields four candidate roots; since the physical speed is unique, the modal root across a frame is the estimate. Static points are selected by a mode-based threshold on Doppler radial speeds, under the assumption that static reflectors dominate the field of view. The paper reports about 2 cm/s median absolute error on a ground vehicle, up to 4x improvement over Doppler-based estimation, and larger errors on the drone and handheld setups, where the constant-velocity assumption is violated by vibration and rotation.","pith_inferences":["Editorial inference: the fourth-order polynomial models only the scalar speed magnitude; a two-frame or multi-static-point version of the same phase geometry could also estimate the heading direction $\\alpha$ over time, turning the method into lightweight odometry rather than just speedometry.","Editorial inference: the pattern of errors across platforms suggests the method's effective envelope is rigid translation with slow heading change; adding gyroscope-based rotation compensation should recover much of the 0.15-0.16 m/s error seen in handheld and drone runs.","Editorial inference: because phase-based speed is absolute and drift-free, it could serve as an in-situ calibration signal for IMU bias or as a regularizer for visual-inertial odometry in dark and dusty conditions where cameras fail.","Editorial inference: a controlled test with sinusoidal speed profiles would separate the constant-velocity assumption from the mode-voting machinery; the model predicts the modal root tracks the mean speed while the root spread grows with acceleration."],"forward_implications":["Robots moving slower than the Doppler resolution (here 0.0496 m/s) can still get a continuous speed estimate, because phase tracking is not quantized to Doppler bins.","Static reflectors no longer need to sit near the radar's boresight; oblique reflectors contribute through the heading angle $\\alpha$ estimated from pairwise radial-velocity ratios.","The approach is real-time on a low-power embedded board, with roughly 0.29 s latency and about 2.25 W power, compared with seconds and higher power for DNN baselines.","In dynamic scenes, accuracy degrades gracefully with the number of moving reflectors, from 0.018 m/s MAE with three static objects to 0.054 m/s with two dynamic objects, and recovers when a wall enters view.","The speed estimate is independent of IMU integration, so it avoids drift accumulation at constant and low speeds where accelerometers struggle."],"supporting_citations":[{"why":"Provides the FMCW phase-to-range relation and the Doppler-resolution limit that defines the sub-Doppler regime RadarTrack targets.","marker":"[9]"},{"why":"The earlier phase-based ego-speed proof of concept that RadarTrack extends beyond a single boresight static object.","marker":"[17]"},{"why":"Radarize, the DNN Doppler-heatmap baseline, whose latency and accuracy are the main comparisons.","marker":"[16]"},{"why":"MilliEgo, the multi-modal DNN baseline, whose ego-motion output is adapted to speed and used as a comparison.","marker":"[11]"},{"why":"A dynamic-object mmWave ego-motion method that motivates the static/dynamic segmentation and is contrasted with RadarTrack's phase pipeline.","marker":"[23]"},{"why":"Supplies the phase-sensitivity argument that tiny phase changes correspond to sub-micrometer range changes, justifying chirp-level phase tracking.","marker":"[31]"},{"why":"Representative point-cloud-alignment ego-motion approach that RadarTrack positions itself against as a phase-based alternative.","marker":"[15]"}],"fun_headline_variants":["RadarTrack: phase-only speed from mmWave, 2 cm/s error","No AI, no IMU: mmWave phase nails ego-speed to 2 cm/s","Phase-based mmWave ego-speed: beats Doppler 4x, no DNN","Single-chip radar: phase math gives 2 cm/s vehicle speed","RadarTrack: phase-only mmWave speed, 2 cm/s median error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ego-vehicle's velocity magnitude and direction stay constant for the entire 150 ms frame, and that static reflectors outnumber dynamic ones in view; if a platform vibrates, pitches, or sees only moving objects, the kinematic equation and the mode vote lose their foundation.","fun_headline_variants_meta":{"raw":{"variants":["RadarTrack: phase-only speed from mmWave, 2 cm/s error","No AI, no IMU: mmWave phase nails ego-speed to 2 cm/s","Phase-based mmWave ego-speed: beats Doppler 4x, no DNN","Single-chip radar: phase math gives 2 cm/s vehicle speed","RadarTrack: phase-only mmWave speed, 2 cm/s median error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1407,"prompt_tokens":895,"completion_tokens":512,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":407}},"tokens_in":511,"tokens_out":512,"duration_ms":4075,"temperature":1.0,"reasoning_tokens":407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:47:32.377342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Mount the radar on a linear stage that moves at a constant 2 cm/s and carries a high-rate encoder as ground truth, with a single static reflector at 45 degrees; if the modal fourth-order root deviates systematically from the encoder speed once pitch oscillation is added, the constant-velocity phase model is disconfirmed.","supporting_citations":[{"cited_title":"Poster: Dynamic ego-velocity estimation using moving mmwave radar: A phase- based approach,","cited_arxiv_id":null,"evidence_quote":"The earlier phase-based ego-speed proof of concept that RadarTrack extends beyond a single boresight static object."},{"cited_title":"Radarize: Enhancing radar slam with generalizable doppler-based odometry,","cited_arxiv_id":null,"evidence_quote":"Radarize, the DNN Doppler-heatmap baseline, whose latency and accuracy are the main comparisons."},{"cited_title":"milliego: single-chip mmwave radar aided egomotion estimation via deep sensor fusion,","cited_arxiv_id":null,"evidence_quote":"MilliEgo, the multi-modal DNN baseline, whose ego-motion output is adapted to speed and used as a comparison."},{"cited_title":"Push the limit of single-chip mmwave radar-based egomotion estimation with moving objects in fov,","cited_arxiv_id":null,"evidence_quote":"A dynamic-object mmWave ego-motion method that motivates the static/dynamic segmentation and is contrasted with RadarTrack's phase pipeline."},{"cited_title":"mmspy: Spying phone calls using mmwave radars,","cited_arxiv_id":null,"evidence_quote":"Supplies the phase-sensitivity argument that tiny phase changes correspond to sub-micrometer range changes, justifying chirp-level phase tracking."},{"cited_title":"Precise ego-motion estimation with millimeter-wave radar under diverse and challenging conditions,","cited_arxiv_id":null,"evidence_quote":"Representative point-cloud-alignment ego-motion approach that RadarTrack positions itself against as a phase-based alternative."}],"review_version":1}