{"id":"5da07296-f33e-4499-ba93-203743949d34","arxiv_id":"2509.03451","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"A phone and watch working together can track shoulder, elbow, and wrist positions to 11 cm median error using UWB ranging plus IMU data and no user-specific training.","lead":"SmartPoser estimates a person's arm pose in 3D using only a smartphone in the pocket and a smartwatch on the wrist, combining UWB radio distance with motion sensor data. Its median error of about 11 cm makes camera-free arm tracking practical for fitness, rehabilitation, and mobile AR applications.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy claim depends on unstated arm-span input; sensitivity to arm-span errors not evaluated.","rationale":"The reader's weakest assumption identifies the arm-span and calibration sensitivity as the primary gap. I agree, focusing specifically on arm span because it directly scales the final output and is completely unstated in the paper: no measurement method, no default, no robustness analysis. The central claim of 11.0 cm is a median after applying the correct arm span, but a deployed system would require the user to input this value without guidance, and any error would directly affect the reported accuracy. This is not a fatal flaw in the method but a missing specification that makes the headline result conditional. The reader's CONDITIONAL verdict already captures this; therefore no change is needed. The open-source release of code and data makes the proposed perturbation test straightforward and would settle whether the concern is practically significant.","tokens_in":16762,"tokens_out":5344,"duration_ms":59667,"concrete_test":"Using the open-source model and dataset, recompute the median elbow/wrist error while perturbing the per-participant arm span values used for scaling by −20%, −10%, +10%, and +20%, and also test using arm span estimated from the T-pose UWB distance (if such an estimator is added). If a ±10% arm span perturbation changes the median error by more than ~20% (e.g., to beyond 15 cm), the headline accuracy is sensitive to an unspecified input and the claim should be qualified accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reported 11.0 cm median error is obtained after scaling model outputs from normalized units to real-world centimeters using the user's arm span (Sec. 3.3, 3.4). The paper never states how arm span is measured or provided to the system in practice, nor does it analyze how errors in this value propagate. Because both the UWB correction network's input and the pose estimator's output are normalized by arm span, an inaccurate arm span changes the input features and the output scaling simultaneously. For a typical adult with arm span ~170 cm, the normalized median error is ~6.5% of arm span. If a user over- or under-estimates arm span by 10% (17 cm), the reported error could plausibly double or become negligible in absolute terms, yet no error bars or sensitivity analysis cover this. The limitations section (Sec. 8) discusses the calibration T-pose but is silent on arm span acquisition and accuracy. Without a specified method for obtaining arm span or a robustness evaluation, the central claim of 'median positional error of 11.0 cm' is not a fully reproducible end-to-end result from just the phone and watch; it assumes an external, unquantified human input.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"SmartPoser proposes a software-only arm-pose tracking system using an off-the-shelf smartphone and smartwatch. It fuses UWB distance measurements between the two devices with their IMU orientations and accelerations, feeding a two-stage bidirectional-LSTM pipeline: a UWB correction network that cleans noisy body-occluded range measurements, followed by a pose estimator that outputs shoulder, elbow, and wrist positions in a normalized arm-span coordinate frame. The system is trained end-to-end on Kinect-derived ground truth from 10 participants and evaluated with leave-one-participant-out cross-validation. The headline result is a median elbow/wrist positional error of 11.0 cm (SD=0.9 cm), with an ablation showing UWB reduces error by 20--25% relative to IMU-only input. The authors also report a real-time 25 Hz implementation and release the dataset, models, and code.","tokens_in":17040,"tokens_out":3532,"duration_ms":36896,"significance":"If the 11.0 cm median error is reproducible in the claimed deployment conditions, this is a meaningful advance: it would be the first demonstration of arm-pose tracking at this accuracy using only a phone and a watch in normal wearing positions, with no per-user training data and no external infrastructure. The paper's strengths include a real user study with 280k frames, leave-one-participant-out evaluation, a clean UWB/IMU ablation, an on-device CoreML latency measurement, and open-sourced data and models. These are concrete and valuable contributions beyond the headline number. However, the absolute error depends on an external input—the user's arm span—whose acquisition and sensitivity are not described, and the comparative claims against prior systems rest on incompatible datasets. Both issues are addressable but need attention before the central claim is fully supported.","major_comments":[{"comment":"The reported 11.0 cm median error is not an end-to-end system result as stated because the model's input features and output coordinates are both scaled by the user's arm span, yet the paper never specifies how arm span is obtained in practice or how sensitive the result is to that input. Section 3.3 says UWB distances are 'scaled by the wearer's arm span' and Section 3.4 says joint locations are 'normalized to the user's arm span' and then reprojected into real-world units by multiplying by arm span. A 10% error in arm span will propagate directly into the output scale and also alter the UWB-correction input features, so the measured 11.0 cm could change by several centimeters. Section 8 lists calibration and pocket placement as limitations but is silent on arm-span acquisition. The authors should state how arm span is measured or supplied (e.g., user entry, height-based estimate, or a","section":"Sec. 3.3, 3.4, and 8"},{"comment":"The comparison to ArmTrak, IMUPoser, and MUSE is made across datasets with different tasks, protocols, and evaluation metrics. The paper acknowledges 'these systems were evaluated on different datasets' but then asserts the comparison is applicable. This does not invalidate the absolute 11.0 cm measurement, but it does not support the stronger framing in the introduction that SmartPoser achieves 'the best tracking accuracy among systems that do not require special user instrumentation or external infrastructure.' A direct comparison on a shared benchmark or a reimplementation is needed, or the claims should be softened to 'comparable in magnitude to previously reported results under different evaluation conditions.'","section":"Sec. 5.5"}],"minor_comments":[{"comment":"The sentence 'our system achieved a real-time (25 Hz) median error of 11.0 cm for the shoulder and wrist' appears to be a typo: the headline metric is for elbow and wrist, and the next clause says '13.3 cm for just the wrist.' Please correct.","section":"Sec. 5.5"},{"comment":"The 'real-time' description should be qualified by the 200 ms output delay from using the n-5th frame. This is disclosed in Section 3.4 but not in the results summary or comparison section; a reader could interpret '25 Hz real-time' as zero-latency. Please state this latency explicitly wherever real-time performance is claimed.","section":"Sec. 3.4, 5.5"},{"comment":"The text says the scatter plot uses 'all 131k frames of study data from our gross arm pose procedure,' while Section 4.3 reports 129K frames for gross motions and Figure 7's caption says 280k datapoints. These numbers should be reconciled.","section":"Sec. 5.3 / Fig. 7"},{"comment":"The Kinect pairing procedure is described as 'paired it with the most recent Kinect frame' but the Kinect frame rate is not stated. Please report the Kinect capture rate so readers can assess synchronization latency, particularly for fast arm motions.","section":"Sec. 4.3"},{"comment":"The caption of Figure 7 states R^2 = 0.72 for the corrected-UWB vs Kinect-distance fit, but the text in Section 5.3 does not interpret this value in context. It would help to state what fraction of variance remains unexplained and how that relates to the residual 8.6 cm error.","section":"Sec. 5.3"}],"recommendation":"major_revision","confidential_remarks":"The central empirical result appears credible as measured, and the open-source artifact is a plus. The main concern is that the headline error is not fully pinned down without the arm-span input and its sensitivity; this is a fixable reproducibility issue rather than a reason to reject. The cross-dataset comparison is a second concern that the authors should address either by softening claims or by providing a matched evaluation. I would be comfortable with acceptance after these points are resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the engineering is sound and the central quantitative claim is probably right, but the packaging overreaches in one specific and checkable way—the paper calls itself the first to fuse UWB and IMU for wearable pose tracking while its own related work cites UWB+IMU pose trackers (refs 9, 15, 63). What is genuinely new is doing that fusion on an off-the-shelf phone and watch, and that is a real contribution. The 11.0 cm median error, measured with leave-one-participant-out over ten people, with an ablation showing a 20–25% benefit from adding UWB, is credible. The model is small (46.4k params), runs on-device at 0.99 ms/inference, and they open-source the data, code, and models. That is the right way to do a systems paper, and I believe the number.\n\nThe soft spot that needs the most attention is arm span. The system normalizes UWB distance and pose outputs by the user's arm span, then rescales to centimeters, but the paper never says how arm span is obtained. If a user enters it wrong by 10%, the reported error would scale roughly proportionally, and there is no sensitivity analysis or error bar. That makes the headline 11.0 cm an end-to-end result only under an unstated assumption. It is a limitation, not a fatal one, but it should be fixed before publication: either specify the measurement method or vary arm span in the analysis and show the error's sensitivity. The comparison to previous work across incompatible datasets is also weak, though they acknowledge it in Section 5.5. Two smaller points: the \"real-time\" figure includes a 200 ms prediction delay, which is fine but should stay visible, and the calibration T-pose requirement is acknowledged in Section 8 but its accuracy sensitivity is not analyzed.\n\nThe novelty overclaim and the arm-span gap are both addressable in revision. The underlying engineering and evaluation are sound, and the open-source release makes the result independently checkable. I'd send this to a serious referee; it deserves a full review, not a desk reject. I'd also bring it to a reading group if your group cares about on-device sensing, UWB, or practical pose tracking.","headline":"Solid commodity-device pose tracking with a credible 11-cm claim, but the 'first UWB+IMU' framing overreaches and the headline number depends on an unstated arm-span input.","tokens_in":17514,"tokens_out":2013,"would_cite":true,"duration_ms":22371,"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":"SmartPoser claims that an off-the-shelf smartphone and smartwatch, fused through ultra-wideband distance and inertial data, can track elbow and wrist positions with a median error of 11.0 cm in real time, with no user training data.","keywords":["arm pose estimation","ultra-wideband ranging","IMU fusion","smartwatch sensing","smartphone sensing","wearable motion tracking","real-time pose estimation"],"falsifier":"Run SmartPoser on participants whose arm span is deliberately mis-entered by ±10% and ±20%, and who rotate the phone in the pocket after calibration, then compare median elbow/wrist error against marker-based ground truth. If the 11.0 cm figure holds only with an exact arm span and an undisturbed calibration, the accuracy claim overstates robustness to inputs a real user controls.","tokens_in":16668,"feed_emoji":"⌚","tokens_out":14399,"duration_ms":113439,"temperature":0.7,"pith_summary":"Arm pose tracking typically needs cameras, body-worn sensor suits, or instrumented rooms — none of which match what consumers own. SmartPoser claims that two devices people already carry — a smartphone in the front pocket and a smartwatch on the wrist — can do it together. The trick is pairing a measurement that drifts (inertial orientation and acceleration from each device's IMU) with one that does not (ultra-wideband radio measures the absolute distance between the two devices about five times per second). Two small recurrent networks clean up the noisy distance signal and convert the combined stream into 3D elbow and wrist positions, normalized to the wearer's arm span. In a 10-person study with leave-one-participant-out evaluation, the system reports a median error of 11.0 cm averaged over elbow and wrist — the paper argues this puts commodity phone-and-watch tracking on par with far more instrumented systems.","feed_headline":"11 cm arm tracking from phone and watch alone","feed_subtitle":"Fusing UWB distance with inertial data estimates elbow and wrist joints without cameras or per-user training data.","key_machinery":"Two complementary measurements carry the argument: UWB ranging (~5 Hz) gives an absolute, drift-free distance between phone and watch; each device's IMU (~25 Hz) gives orientation and acceleration. Two small bidirectional LSTM networks do the computation — 46.4k parameters total. The UWB correction network cleans the raw distance stream using IMU context, cutting mean distance error from 21.0 cm to 8.6 cm. The pose estimator then converts corrected distance plus IMU features into 3D shoulder, elbow, and wrist positions normalized to arm span, reading the n−5th frame (200 ms behind real time) as its output. Side-by-side devices plus a T-pose align the two reference frames.","core_discovery":"SmartPoser's central claim is that the absolute distance between a smartphone and a smartwatch — measured with UWB time-of-flight ranging — is the missing signal that makes arm pose estimation practical on consumer hardware. Inertial data drifts and is relative; UWB distance is absolute and stable, so fusing the two gives the pose estimator an anchor. A UWB correction network reduces raw distance error from a mean of 21.0 cm to 8.6 cm, and a pose estimator outputs shoulder, elbow, and wrist positions normalized to the user's arm span, then reprojected to centimeters by that span. With leave-one-participant-out cross-validation against Kinect-derived ground truth, the system reports median er","pith_inferences":["Because every reported coordinate is normalized to arm span and then scaled back, an error in the user's arm-span estimate propagates roughly linearly into absolute position; the paper never measures this sensitivity, so everyday accuracy could deviate from 11.0 cm whenever a user's span is misentered.","The paper's own heatmaps show the worst errors where the UWB signal is blocked by the body (above the shoulder, at the right thigh); a second UWB anchor — e.g., earbuds — or a stronger correction model would likely attack the worst cases more than the median.","The same corrector-plus-regressor recipe could be retrained for other body parts: a pocket phone and a UWB-equipped shoe might track leg pose, or a UWB earbud could replace the phone as the absolute reference for the watch.","The fixed 200 ms output lag sets a floor on interactivity; gesture-input and gaming applications that need immediate response would have to trade accuracy for a shorter context window or a shallower network."],"forward_implications":["Arm tracking becomes a software feature: UWB-capable phone-and-watch pairs could run pose estimation as an app update, with no cameras, body suits, or external infrastructure.","Fitness, rehabilitation, life-logging, occupational-safety, and context-aware assistant applications gain a continuous, real-time arm-position stream in mobile settings where prior systems required static or heavily instrumented setups.","The reported accuracy (8.7 cm elbow, 13.3 cm wrist, 11.0 cm pooled median) places commodity phone-and-watch tracking in the same range as far more instrumented systems, including ArmTrak, MUSE, and IMUPoser.","The approach is claimed to keep working while the user moves — a capability ArmTrak lacked — and the same pipeline is expected to extend to other phone placements such as rear pockets and jacket pockets, and to other UWB devices like earbuds or smart glasses."],"supporting_citations":[{"why":"ArmTrak, the single-IMU arm tracker that is SmartPoser's primary comparison baseline and the prior system it must beat.","marker":"[42]"},{"why":"IMUPoser, the closest prior work using commodity phones, watches, and earbuds; supplies the T-pose calibration procedure and a non-instrumented comparison point.","marker":"[32]"},{"why":"TransPose, source of the bidirectional RNN design, the n−5th-frame prediction trick, and the acceleration scaling convention.","marker":"[59]"},{"why":"Deep Inertial Poser, the architecture and windowed training approach the pose estimator is built on.","marker":"[16]"},{"why":"Physical Inertial Poser, another architecture and training source that also establishes the acceleration-scaling-by-30 convention.","marker":"[58]"},{"why":"MUSE, the wrist-tracking baseline used for comparison when the body moves.","marker":"[41]"},{"why":"The platform UWB ranging API that supplies the absolute phone-to-watch distance measurements on commodity devices.","marker":"[2]"},{"why":"The platform IMU API that supplies the phone and watch orientation and acceleration streams used at the 25 Hz frame rate.","marker":"[1]"}],"fun_headline_variants":["Phone and watch track arm pose with 11 cm accuracy","No cameras, just phone and watch: arm pose within 11 cm","UWB + IMU: arm tracking from pocket and wrist","Smartphone and smartwatch pin elbow and wrist without training","Arm pose to 11 cm via phone-watch UWB and IMU"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The 11.0 cm median assumes the user's arm span is known and the two-step calibration (devices held side by side, then a T-pose) was performed correctly; the paper does not quantify how much error grows when the arm span is wrong or the phone shifts or rotates in the pocket afterward.","fun_headline_variants_meta":{"raw":{"variants":["Phone and watch track arm pose with 11 cm accuracy","No cameras, just phone and watch: arm pose within 11 cm","UWB + IMU: arm tracking from pocket and wrist","Smartphone and smartwatch pin elbow and wrist without training","Arm pose to 11 cm via phone-watch UWB and IMU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000695,"raw_usage":{"total_tokens":2957,"prompt_tokens":700,"completion_tokens":2257,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":2178}},"tokens_in":444,"tokens_out":2257,"duration_ms":14393,"temperature":1.0,"reasoning_tokens":2178,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:52:26.636108+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SmartPoser on participants whose arm span is deliberately mis-entered by ±10% and ±20%, and who rotate the phone in the pocket after calibration, then compare median elbow/wrist error against marker-based ground truth. If the 11.0 cm figure holds only with an exact arm span and an undisturbed calibration, the accuracy claim overstates robustness to inputs a real user controls.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ArmTrak, the single-IMU arm tracker that is SmartPoser's primary comparison baseline and the prior system it must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"IMUPoser, the closest prior work using commodity phones, watches, and earbuds; supplies the T-pose calibration procedure and a non-instrumented comparison point."},{"cited_title":"Black, Otmar Hilliges, and Gerard Pons-Moll","cited_arxiv_id":null,"evidence_quote":"Deep Inertial Poser, the architecture and windowed training approach the pose estimator is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Physical Inertial Poser, another architecture and training source that also establishes the acceleration-scaling-by-30 convention."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MUSE, the wrist-tracking baseline used for comparison when the body moves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The platform UWB ranging API that supplies the absolute phone-to-watch distance measurements on commodity devices."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The platform IMU API that supplies the phone and watch orientation and acceleration streams used at the 25 Hz frame rate."}],"review_version":1}