{"id":"0ce8137c-a6cb-42fc-8f21-6cdc7d76862c","arxiv_id":"2507.20437","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"EchoForce estimates grip force from ultrasound reflections off skin deformations on the wrist, achieving 9.08% user-dependent and 12.29% user-independent mean error in an 11-person study.","lead":"A wristband with a speaker and microphone estimates grip force by bouncing ultrasound off forearm skin and reading the deformation. It reports about 9% error for known users and 12% for new users, which could make continuous grip-strength monitoring practical outside the lab.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth interpolation could smooth the target and deflate RMSE; the headline error rates need an independent check against a higher-fidelity force reference.","rationale":"The reader's weakest_assumption identifies exactly the same issue: the ground-truth trace is linearly interpolated from a low-rate, partially frozen dynamometer display, and the authors themselves flag in Section 6.4 that 'this interpolation may have compromised the fidelity and accuracy of our ground truth.' That is the load-bearing soft spot because every reported error metric is RMSE against that reconstructed trace. A smoothed target makes the regression task look easier, and the paper's own limitation statement acknowledges the risk. I agree with the CONDITIONAL verdict and do not see reason to move it. Independent supporting evidence is real: the device was remounted between sessions, cross-orientation evaluation was performed, user-independent testing used leave-one-user-out, and design choices such as the 45-degree sensor angle were grounded in a pilot comparison. The interpolation concern is concrete and testable: replacing the camera-plus-interpolation reference with a streaming dynamometer would settle whether the headline numbers survive. I keep the verdict CONDITIONAL rather than ACCEPT because the paper reports no confidence intervals, no ground-truth reconstruction error, and no released artifacts; these are verification shortfalls rather than fatal flaws. Even a worst-case moderate bias would shift the headline from about 9% to perhaps 10-11%, which still supports the central claim at reduced strength. The proposed test directly measures the bias; the verdict should move to ACCEPT only after such a check confirms the bias is negligible.","tokens_in":14896,"tokens_out":3049,"duration_ms":29336,"concrete_test":"Re-run the full pipeline on a subset of sessions with a synchronized high-fidelity force reference (a streaming dynamometer or instrumented handle with load-cell output at >=100 Hz) and compare the reconstructed 100 Hz ground truth against that reference. Quantify the reconstruction RMSE separately on ramp and release transients; if it exceeds ~0.5 kg, recompute the user-dependent and user-independent error rates using the high-fidelity reference and check whether the 9.08% and 12.29% figures shift by more than 1-2% MVC.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is that EchoForce estimates grip force at 9-12% MVC error, comparable to EMG. Every reported number depends on the ground-truth force trace reconstructed from a 30 Hz camera reading of a ~15 Hz display, with gaps filled by linear interpolation and then upsampled to 100 Hz (Section 3.3). The participants were asked to ramp to a displayed target 'as quickly and accurately as possible' and then release, so the true force has rapid ramps and releases. Linear interpolation of a 15-30 Hz sampled, hold-indicator-flagged signal systematically low-passes the true force trace: it rounds off the shoulders of the ramps, blurs short peaks, and replaces fast release transients with straight lines. Because the loss and metric are RMSE against this smoothed target, missing those fast transients is penalized less, so the reported RMSE could be artificially low even if predictions are accurate only in the plateau region. This is a measurement-reference concern, not an attack on the acoustic sensing mechanism, which is externally referenced by the dynamometer. The plausible severity: a bias of only ~0.25-0.5 kg over a 2-3 kg error budget would move P1's 6.01% to near 7-8%, and the 9.08% headline toward 10-11%, weakening but not destroying the claim. The user-independent result at 12.3% is comparatively less affected because smoothing artifacts apply across users, so the 'comparable to EMG' claim is the most exposed part.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"EchoForce presents a wristband that uses active acoustic sensing (FMCW ultrasound) to estimate grip force from skin deformation at the wrist. The authors report a user study with 11 participants, across 15 sessions (5 per each of three wrist orientations) with device remounting, and evaluate four model configurations: user-dependent, user-independent (leave-one-user-out), fine-tuned, and cross-orientation. Headline results are 10.10% user-dependent error, 9.08% fine-tuned, 12.29% user-independent, and 12.94% cross-orientation error, expressed as RMSE divided by MVC. The paper claims these results are comparable to or better than EMG-based grip force estimation, and robust to remounting, orientation, and new users.","tokens_in":15063,"tokens_out":6470,"duration_ms":73893,"significance":"The paper makes a plausible and useful contribution: a low-cost, non-contact wristband for continuous grip force estimation with a careful evaluation protocol. The study design is a genuine strength—15 sessions with remounting, three orientations, leave-one-user-out testing, and cross-orientation evaluation are appropriate steps for demonstrating robustness. The hardware and signal-processing pipeline are described clearly enough to be reproduced. However, the headline accuracy figures depend on a ground-truth force trace reconstructed by linear interpolation from a ~15 Hz display filmed at 30 Hz (Section 3.3), a limitation the authors acknowledge in Section 6.4. Because the loss and error metrics are computed against this reconstructed trace, the reported RMSE values may be optimistic. If the accuracy claims survive a higher-fidelity ground truth, the system would be a meaningful advance for wearable grip-force sensing.","major_comments":[{"comment":"The ground-truth force signal was reconstructed by filming a dynamometer display that updates at approximately 15 Hz, discarding frames with the hold indicator active, and filling gaps with linear interpolation upsampled to 100 Hz. Since the task instructed participants to ramp to target \"as quickly and accurately as possible\" and then release (Section 4.1), the true force contains rapid ramps and releases that linear interpolation would smooth. The RMSE-based error rates reported in Table 1 are therefore computed against a target that may be missing exactly those transients, which can artificially lower the reported error. The manuscript itself states in Section 6.4 that the interpolation \"may have compromised the fidelity and accuracy of our ground truth.\" This is a load-bearing measurement-reference issue, not a criticism of the acoustic sensing mechanism. The authors should either re-analyze a subset of data with a streaming dynamometer (e.g., the one cited in Section 6.4) or quantitatively bound the effect of interpolation, for instance by simulating the interpolation on a known high-rate signal and reporting the induced RMSE reduction.","section":"Section 3.3, Section 5.1, Table 1"},{"comment":"The claim that EchoForce is \"comparable with EMG\" relies on comparing error rates to Keir and Mogk's 11.4% (Section 6.1). However, the paper does not specify whether the RMSE/error rate is computed over the entire recording (including the one-second rest periods between trials, where force is near zero) or only over the active force-producing portions. If rest periods are included, the error over these near-zero intervals can dilute the RMSE, and the comparison to studies that may use different evaluation windows would be misleading. The authors should state the exact evaluation window and, if it includes rest, provide a sensitivity analysis with only active periods.","section":"Section 5.1, Section 6.1"}],"minor_comments":[{"comment":"The description of the input-output mapping is ambiguous: \"The ground truth labels from the past two seconds and the extracted moving window from the echo profile compose a single pair of label and input.\" It is unclear whether the model predicts the entire two-second force trace, the instantaneous force at the end of the window, or something else, given that the decoder is an average pooling layer followed by a fully connected layer. Please clarify with a precise statement of tensor shapes and how echo windows are aligned with force labels.","section":"Section 3.4"},{"comment":"The user-independent model is called a \"foundation model\" in the abstract, but it is trained on only ten users' data. This term is likely to be misread; consider using \"pre-trained user-independent model\" instead.","section":"Abstract and Section 5.1.2"},{"comment":"Error rates are reported to two decimal places; given the ground-truth interpolation uncertainty, one decimal place or confidence intervals would better reflect the actual precision.","section":"Table 1"},{"comment":"There is a typographical error in the sentence about future dynamometers: \"more suitable for precise, continuous monitoring of grip force over time [27]).\" has an extra closing parenthesis.","section":"Section 6.4"},{"comment":"The caption says a five-point moving-average filter was applied for clarity; please specify whether this filter was applied to predictions only or to the ground truth as well, since applying it to both would affect the visual comparison.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"This is a well-executed systems paper with a careful evaluation protocol, but the central accuracy numbers rest on a ground-truth signal reconstructed via linear interpolation from a low-rate display. The authors' own acknowledgment of this limitation in Section 6.4 is honest, but the load-bearing comparison to EMG and the headline percentages need to be supported by a quantitative assessment of the interpolation's effect or a re-analysis with a streaming dynamometer. The paper fits the venue well and the core idea is sound; the requested revision is feasible within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"EchoForce applies active acoustic sensing to a new target—continuous grip force estimation from skin deformation on the wrist—and the study design is genuinely careful. Fifteen sessions per participant with full remounting, three wrist orientations, leave-one-user-out testing, and fine-tuned transfer. The user-independent 12.3% error rate is the headline that matters, because that is where EMG and light-based approaches fall apart. The central claim—that reflected ultrasound tracks skin deformation enough to estimate grip force—is plausible and externally referenced to a dynamometer, so there is no equation-level circularity.\n\nThe weak spot is the ground truth. The dynamometer display updates at about 15 Hz, filmed at 30 Hz, with hold-mode frames discarded and gaps linearly interpolated and upsampled to 100 Hz. Participants ramped to target as quickly as possible and released, so the true force trace has fast ramps and releases that linear interpolation will smooth away. Since RMSE is measured against that smoothed trace, the reported errors are likely understated. The authors openly acknowledge this in Section 6.4. I don't think it kills the paper—a 0.25–0.5 kg bias would push the headline 9.1% toward 10–11%, which still beats the capacitive skin-deformation baseline at 27.9% and stays in EMG territory—but the \"comparable to EMG\" claim is the most exposed part and needs an independent check with a higher-fidelity force reference.\n\nOther soft spots: 11 young, mostly male, right-handed participants from one institution, MVCs below population norms, no confidence intervals, no released code or data, and no dynamic wrist motion. These are addressable, and the authors flag most of them.\n\nWhat the paper does well: it credits prior work, does not overclaim first-principles novelty, and the limitation section is candid. The citation pattern is fair, including the closest competitors. The audience is the wearable-HCI and acoustic-sensing community.\n\nRecommendation: send it to peer review. The measurement-reference concern is serious enough to require revision, but the mechanism, study design, and honest reporting make it worth referee time. I would also want per-participant confidence intervals and, ideally, released data.","headline":"A well-built acoustic wristband for grip force with an honest limitations section, but the headline numbers rest on a ground truth that linear interpolation likely smoothed—worth refereeing, needs an independent force reference.","tokens_in":15714,"tokens_out":2854,"would_cite":true,"duration_ms":30217,"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":"EchoForce shows that a wristband can estimate continuous grip force from reflected ultrasound with EMG-level accuracy, including for new users and after remounting.","keywords":["grip force estimation","active acoustic sensing","skin deformation","wearable wristband","FMCW ultrasound","continuous force regression","user-independent model","health monitoring"],"falsifier":"Re-run the user study with a dynamometer that streams force data directly at 100 Hz or higher and recompute the three headline error rates; if the RMSE increases materially, the linear interpolation had smoothed away quick force transients and the reported percentages overstate accuracy.","tokens_in":14579,"feed_emoji":"💪","tokens_out":7726,"duration_ms":75701,"temperature":0.7,"pith_summary":"This paper proposes EchoForce, a wristband that measures continuous grip force by emitting inaudible ultrasound toward the forearm and analyzing the echoes reflected from the skin. The claim is that the subtle skin deformations caused by the finger flexor muscles carry enough information to estimate grip force with accuracy comparable to electromyography (EMG), but without the cost, setup expertise, or per-session recalibration EMG requires. In a study with 11 participants, the authors report a fine-tuned user-dependent mean error rate of 9.08% of each participant's maximum voluntary contraction, a user-dependent error of 10.10%, and a user-independent error of 12.29%, with accuracy maintained when the device is remounted and when the wrist is supinated, neutral, or pronated. If correct, this would make continuous grip force measurement practical for health monitoring, rehabilitation, and force-based interaction on ordinary wrist-worn devices.","feed_headline":"Wristband hears grip force from skin echoes at ~9% error","feed_subtitle":"Reflected ultrasound estimates grip force with EMG-level accuracy, even for new users and after remounting.","key_machinery":"The load-bearing mechanism is active acoustic sensing of skin deformation. The speaker emits frequency-modulated continuous-wave (FMCW) ultrasound in the 20–29 kHz range, and a microphone placed at 45 degrees to the skin captures reflected signals; the shape of the skin surface alters reflection delay, angle, and frequency. Consecutive FMCW frames are cross-correlated with the transmitted signal and subtracted frame-to-frame to produce a differential echo profile that isolates movement of the skin surface from static reflections. A 320-by-78-pixel moving window, representing two seconds of data, is fed into a FastViT encoder followed by pooling, dropout, and a fully connected layer, which regresses the window to grip force in kilograms. The 45-degree sensor angle, the wristband placement over the flexor tendons, and the differential echo profile are the design choices that carry the sensing claim.","core_discovery":"The central discovery is that active acoustic sensing can serve as a non-contact proxy for grip force: the wristband's speaker emits 20–29 kHz FMCW sweeps, and the microphone captures reflections whose timing, angle, and frequency shift change as the flexor tendons and overlying skin deform under tension. From these reflections the system builds a differential echo profile, and a vision-transformer regression model maps two-second windows of that profile to force in kilograms. Evaluated on 11 participants, the system reached a mean error rate of 10.10% (2.56 kg) in user-dependent testing, 9.08% (2.31 kg) after fine-tuning a user-independent foundation model with the participant's own data, and 12.29% (3.11 kg) in leave-one-user-out testing; error rates are reported relative to each participant's maximum voluntary contraction. The authors interpret this as evidence that the approach generalizes across remounting sessions, wrist orientations, and new users, thereby addressing a known weakness of EMG-based and optical grip-force sensors.","pith_inferences":["One consequence the authors leave implicit is that the same acoustic principle might estimate other fine-grained muscle outputs, not just grip force, since any flexor-driven skin deformation would modulate the echo profile.","A testable extension is to replace the video-tracked dynamometer with a dynamometer that streams ground truth at high rate; this would show whether linear interpolation of the 15 Hz display inflated the reported accuracy.","The authors evaluate static postures only; a direct next experiment is to measure grip force during walking or arm motion to see whether the learned echo features survive real-world artifacts.","Because the sensor bracket was tuned to a 45-degree angle and a specific wrist position, off-the-shelf smartwatches with different speaker and microphone layouts may need hardware adjustment before achieving the same accuracy."],"forward_implications":["Grip force can be tracked continuously from a wrist-worn device without obstructing the hand, because the sensing is non-contact and requires only a speaker and microphone.","A user-independent model removes the need for per-user calibration, so a wristband could work out of the box for new users at the reported 12.29% mean error rate.","Fine-tuning with a small amount of the user's own data improves mean error to 9.08%, suggesting accuracy can improve as more user data is collected.","The same skin-deformation signal could support force-based input for smartwatches and augmented reality, and grip-force monitoring for rehabilitation and aging-related health assessment."],"supporting_citations":[{"why":"Supplies the EMG-based error-rate baseline of 11.4% that the paper uses to claim EMG-comparable performance.","marker":"[17]"},{"why":"Is the Camry dynamometer whose display is video-recorded to produce the ground-truth force trace.","marker":"[3]"},{"why":"Supports the reliability and criterion validity of the low-cost Camry dynamometer used for ground truth.","marker":"[46]"},{"why":"Provides the C-FMCW signal-processing method that produces the cross-correlated echo profiles.","marker":"[48]"},{"why":"Is the prior wrist-topography capacitive sensing work whose 27.9% error motivates the acoustic skin-deformation approach.","marker":"[40]"},{"why":"Is the light-reflectivity grip-force system used as the comparison point that needs fine-tuning to reach low error.","marker":"[51]"},{"why":"Demonstrates the wristband active acoustic sensing platform and echo-profile pipeline that EchoForce adapts.","marker":"[21]"},{"why":"Is the FastViT model used as the encoder for regression from echo-profile windows.","marker":"[47]"},{"why":"Is the EMG study whose static-calibration user-independent error of 6.1 kg EchoForce compares against in kilograms.","marker":"[14]"}],"fun_headline_variants":["Wristband's ultrasound listens to skin to gauge grip force","EchoForce: Acoustic wristband estimates grip force within 12% for new users","Active acoustics turn skin echoes into grip strength readings","Non-contact grip force monitor uses sonar-like skin reflections","Skin echoes reveal grip force: wristband hits 9% error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ground-truth force trace is trustworthy, even though it was reconstructed from a 30 Hz video of a 15 Hz dynamometer display, with dropped frames filled by linear interpolation and upsampled to 100 Hz, a method the paper concedes may have compromised fidelity.","fun_headline_variants_meta":{"raw":{"variants":["Wristband's ultrasound listens to skin to gauge grip force","EchoForce: Acoustic wristband estimates grip force within 12% for new users","Active acoustics turn skin echoes into grip strength readings","Non-contact grip force monitor uses sonar-like skin reflections","Skin echoes reveal grip force: wristband hits 9% error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1542,"prompt_tokens":925,"completion_tokens":617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":527}},"tokens_in":541,"tokens_out":617,"duration_ms":6692,"temperature":1.0,"reasoning_tokens":527,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:34:14.157927+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the user study with a dynamometer that streams force data directly at 100 Hz or higher and recompute the three headline error rates; if the RMSE increases materially, the linear interpolation had smoothed away quick force transients and the reported percentages overstate accuracy.","supporting_citations":[{"cited_title":"Keir, , and Jeremy P.M","cited_arxiv_id":null,"evidence_quote":"Supplies the EMG-based error-rate baseline of 11.4% that the paper uses to claim EMG-comparable performance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the Camry dynamometer whose display is video-recorded to produce the ground-truth force trace."},{"cited_title":"Ruiz, Tamás Csányi, and Francisco B","cited_arxiv_id":null,"evidence_quote":"Supports the reliability and criterion validity of the low-cost Camry dynamometer used for ground truth."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the C-FMCW signal-processing method that produces the cross-correlated echo profiles."},{"cited_title":"Nielsen, Steffen Holmgaard, Ning Jiang, Kevin B","cited_arxiv_id":null,"evidence_quote":"Is the prior wrist-topography capacitive sensing work whose 27.9% error motivates the acoustic skin-deformation approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the light-reflectivity grip-force system used as the comparison point that needs fine-tuning to reach low error."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates the wristband active acoustic sensing platform and echo-profile pipeline that EchoForce adapts."},{"cited_title":"FastViT: A Fast Hybrid Vision Transformer using Structural Reparameterization","cited_arxiv_id":"2303.14189","evidence_quote":"Is the FastViT model used as the encoder for regression from echo-profile windows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the EMG study whose static-calibration user-independent error of 6.1 kg EchoForce compares against in kilograms."}],"review_version":1}