{"id":"f478a89e-0f02-4ce8-81aa-04270f289266","arxiv_id":"2505.08213","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A visual-inertial fusion system with a wrist camera and miniature IMUs estimates dexterous hand joint angles within a few degrees without visible drift, using synthetic-only visual training.","lead":"This paper presents HandCept, a system that estimates the joint angles of a robotic hand by combining a wrist-mounted camera with tiny motion sensors. It aims to give dexterous robot hands accurate joint feedback without bulky encoders, which could make hand designs smaller, cheaper, and easier to control.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The uniformity variance in Table I (~2.9 degrees standard deviation) is the same order as the claimed 2-4 degree joint error, so the common-base IMU assumption and the 'no drift' claim are internally unsupported; no numeric error table backs the headline accuracy.","rationale":"I read the paper as proposing a visual-inertial proprioception module whose value rests on the quantitative claim of 2-4 degrees joint-angle error without drift. The strongest vulnerability is not the architecture, which is plausible, but the internal consistency of the supporting experiment. Table I is presented as validating both stability and uniformity, yet the largest reported variance (8.21 deg^2, about 2.9 degrees standard deviation) is not obviously 'very small' relative to the claimed error range. Since the common-base assumption directly feeds the inertial measurement model in Section V, an uncorrected per-chip discrepancy of about 3 degrees would make the claimed 2-4 degree error unattainable unless the visual stream absorbs the bias, which the paper does not demonstrate. This is a concrete, quantitative inconsistency, not a stylistic complaint. I agree with the reader's identification of the uniformity assumption as the weakest point, and I also note the absence of numeric error tables for Fig. 7C, but the variance table is the more specific and damning issue. The open-sourced Blender rendering pipeline and the miniature IMU hardware are genuine contributions and should be credited; however, they do not rescue the central accuracy/drift claim. Thus the reader's REJECT verdict should stand.","tokens_in":11289,"tokens_out":4935,"duration_ms":50964,"concrete_test":"Separate the two variance sources: compute per-IMU temporal Allan deviation over the two-hour static records, and compute cross-IMU quaternion residuals for each pose after best-fit common-base alignment. Then propagate the cross-IMU residual distribution through Eq. (22) and Algorithm 1 to obtain a joint-angle error bound. If that bound exceeds 2-4 degrees, the common-base assumption cannot support the accuracy claim and per-IMU calibration is required; if the large variance is temporal, the 'no drift' statement fails. In addition, rerun the Fig. 7C evaluation and report per-joint RMSE/MAE and standard deviations for IMU-only, visual-only, and fused estimates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that HandCept keeps joint-angle error between 2 and 4 degrees with no observable drift, and Section IV.A justifies the common IMU base frame by asserting that the magnetic field is uniform and the measured variance is 'very small.' Table I, however, reports angular variances up to 8.21 deg^2 (Pose #2), i.e., about 2.9 degrees standard deviation. If that variance is across the four IMUs, a common base frame leaves per-chip orientation residuals of the same magnitude as the headline joint-angle accuracy, and these residuals propagate through Eq. (22) into joint angles. If the variance is instead temporal drift during the two-hour static test, it directly contradicts 'without observable drift.' The paper does not report which interpretation is correct, nor does it give a numeric table of joint-angle errors for Fig. 7C, so the 2-4 degree claim cannot be checked against the reported uniformity data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HandCept, a visual-inertial fusion framework for estimating joint angles of dexterous robotic hands without joint encoders. A wrist-mounted RGB-D camera provides zero-shot 6D pose estimates for each link using a synthetic-data-trained FFB6D network, while multiple miniature 9-axis IMUs provide orientation estimates. The two modalities are fused in a latency-compensated Extended Kalman Filter, and joint angles are recovered from the fused rotation matrices by solving a kinematics-constrained equation. The authors claim joint-angle errors between 2 and 4 degrees without observable drift, outperform visual-only and inertial-only methods, and validate that the IMU system is stable and uniform enough to share a common base frame. They also open-source a Blender-based rendering pipeline for sim-to-real training.","tokens_in":11478,"tokens_out":4600,"duration_ms":49713,"significance":"If the central accuracy and drift-free claims were quantitatively established, HandCept would be a useful contribution to encoder-free proprioception for dexterous hands: the compact 12 mm × 15 mm IMU module, the open-source rendering pipeline, and the real-hardware demonstration with ArUco ground truth are all strengths. The paper also makes a falsifiable architectural claim, namely that a common base frame across IMUs is justified by field uniformity. However, the significance is currently limited by the absence of any numeric joint-angle error evaluation, by internal inconsistencies in Table I that bear directly on the drift and uniformity assumptions, and by unreported EKF noise parameters. As submitted, the headline 2–4 degree accuracy cannot be checked or reproduced from the reported evidence.","major_comments":[{"comment":"The paper's central claim of 2–4 degree joint-angle error without observable drift has no quantitative support. Figure 7C is a qualitative time-series plot with no error bars, no per-joint numerical errors, no peak errors, no drift statistics, and no stated number of trials. The abstract and conclusion make a specific accuracy claim, but the reader cannot verify it from any table or error metric. Please report per-joint MAE/RMSE, peak error, drift rate, and standard deviations for joints 1–6, with the number of trials.","section":"Section VI.B, Fig. 7C"},{"comment":"Table I reports first-order drift in units of 10^-4 deg/s, with entries up to 36×10^-4 deg/s (Pose #2, Yaw). Over a two-hour static test, a constant drift at that rate accumulates to roughly 26 degrees, which is incompatible with the statements in Section VI.A that drift is 'minimal' and with the abstract's claim of 'without observable drift.' Please clarify whether these values are temporal slopes after Kalman filtering and report the accumulated orientation error over the full two-hour window.","section":"Table I, Section VI.A"},{"comment":"The uniformity claim rests on the assertion that the angular variance in Table I is 'very small,' but variances up to 8.21 deg^2 correspond to a standard deviation of about 2.9 degrees, which is the same order as the claimed 2–4 degree joint-angle accuracy. If the variance is across the four IMU chips, the common-base assumption leaves per-chip orientation residuals that propagate through Eq. (22) into joint angles; if the variance is temporal, it contradicts the no-drift claim. The paper must state which interpretation applies and provide per-chip residuals.","section":"Table I, Section IV.A"},{"comment":"The EKF fusion depends on process noise covariances Q_q and Q_b and measurement noise covariances R_cam and R_imu, as well as on the settings of the IMU orientation Kalman filter, but none of these values are reported. Without these parameters the fusion result is not reproducible and the claimed robustness cannot be assessed. Please provide the numerical values or a sensitivity analysis showing that the headline accuracy does not depend on fine-tuning them.","section":"Section V.A, Eqs. (7), (14), (15)"}],"minor_comments":[{"comment":"There are typos that should be corrected: 'quanlified' should be 'quantified', 'Kinematics Constrains' should be 'Kinematics Constraints', and 'Appilying' should be 'Applying.'","section":"Section IV.B"},{"comment":"The notation for homogeneous transformations is ambiguous: 'aTb' is defined as the transformation from frame a to frame b, but Eq. (1) then writes products such as 'IMU_i \\hat{T}_ee = IMU_i \\hat{T}_{IMU_b} ...' without making the source and target frames explicit in the superscripts. Please clarify the convention and verify the chain of products.","section":"Eq. (1)"},{"comment":"The accuracy claim is stated inconsistently: the abstract says 'errors between 2 and 4 degrees,' while the introduction says 'errors up to 2 degrees.' Please harmonize these statements.","section":"Abstract and Section VII"},{"comment":"Figure 7C lacks axis labels and a clear legend identifying the ground truth, visual-only, inertial-only, and HandCept traces; adding these would make the comparative claim easier to interpret.","section":"Fig. 7C"},{"comment":"The latency-compensation equations contain undefined or inconsistent indices and superscripts (e.g., \\tilde{H}^T_t and the treatment of \\tilde{P}_{t+1|t+k}); please rewrite this section with consistent subscripts and define every symbol.","section":"Section V.B"}],"recommendation":"major_revision","confidential_remarks":"The circularity concern raised in the review is not, in my reading, valid: the joint angles are obtained from measured poses, not from parameters fitted to the target outputs. The decisive issue is purely quantitative: the central 2–4 degree, no-drift claim is unsupported by any numerical joint-angle evaluation, and Table I appears to contradict the drift and uniformity assumptions on their face. If the authors cannot produce per-joint error statistics and reconcile Table I with the no-drift claim, I would recommend rejection; otherwise a revised manuscript with those additions could be reassessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"HandCept has a real system and a real open-source artifact, but the headline accuracy claim is not backed by the paper's own experiments. The joint-angle error figure (Fig. 7C) is qualitative; there is no numeric RMSE table, no per-joint errors, no error bars, and no comparison to existing hand proprioception methods. That alone would make the 2-4 degree claim unverifiable. Worse, Table I undercuts the 'no observable drift' claim: a first-order drift of 36e-4 deg/s is 0.0036 deg/s; over the two-hour static test that is about 26 degrees of uncompensated linear drift. The angular variances up to 8.21 deg^2 (about 2.9 deg std) are the same order as the claimed joint accuracy, so the 'very small' variance used to justify the common IMU base frame is not obviously consistent with a 2-4 degree system. The stress-test note is on target; if anything it understates the drift issue.\n\nWhat is genuinely new: the compact 12x15 mm IMU module with serial/parallel I2C expansion, the open-sourced Blender rendering pipeline for zero-shot sim-to-real training, and the specific integration of wrist RGB-D + multiple IMUs + a delayed-state EKF for dexterous hand proprioception. The hardware and software artifacts are real and useful to the community. The self-citation to the DexCo hand is appropriate as a testbed. The kinematic-constraint section is under-specified (Eq. 2 is skeletal), and the latency-compensation derivation is hard to follow, but those are fixable. The euler-angle uniqueness algorithm is a nice practical detail.\n\nOverall, this is a promising system paper that needs a proper quantitative evaluation. I would not accept it in current form, but I would send it to peer review rather than desk-reject, because the contribution is concrete and the authors have released code/data. The right response is major revision with a real error table, a clear statement of what the drift and variance numbers actually mean, and a comparison baseline.","headline":"Promising hardware and pipeline, but the headline accuracy claim is unsupported by the paper's own experiments and appears contradicted by its drift table.","tokens_in":12062,"tokens_out":2729,"would_cite":false,"duration_ms":26831,"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":"A wrist-mounted RGB-D camera fused with miniature inertial sensors can keep a dexterous hand's joint angles accurate to within 2–4 degrees without drift.","keywords":["proprioception","dexterous hand","visual-inertial fusion","Extended Kalman Filter","joint angle estimation","zero-shot sim-to-real","9-axis IMU","RGB-D pose estimation"],"falsifier":"Repeat the system's uniformity test with four motion sensors mounted near the actuation motors of the hand and power the motors during measurement; if the between-sensor orientation variance grows beyond the roughly 3-degree standard deviation already seen in the paper's Table I, or if joint-angle error exceeds 4 degrees on any axis, then the common-calibration-frame assumption is not reliable enough for the claimed accuracy.","tokens_in":11069,"feed_emoji":"🖐️","tokens_out":8517,"duration_ms":78741,"temperature":0.7,"pith_summary":"HandCept is presented as a visual-inertial proprioception framework for dexterous robotic hands, giving them a sense of their own joint configuration without joint encoders. The paper claims that fusing a wrist-mounted RGB-D camera with miniature 9-axis inertial sensors through a latency-compensated Extended Kalman Filter yields joint-angle estimates accurate to between 2 and 4 degrees, with no observable drift. This matters because encoder-free and tendon-driven hands currently lack reliable, generalizable proprioception, which blocks closed-loop manipulation. The paper also claims that a uniform magnetic field across the hand's sensors lets all inertial modules share one calibration frame, and that a purely synthetic-image training pipeline transfers zero-shot to real images.","feed_headline":"Wrist camera plus tiny sensors measure hand joints to a few degrees","feed_subtitle":"No joint encoders needed: fusing RGB-D vision with inertial sensing keeps dexterous-hand joint angles drift-free.","key_machinery":"The load-bearing mechanism is a latency-free Extended Kalman Filter whose state is each link's unit quaternion plus a gyro bias term. IMU data drives a real-time random-walk prediction; when a delayed camera pose estimate arrives, the filter rewinds and re-propagates the state across the latency window, so slow visual corrections reach the current estimate without stalling the output. Two supporting pieces carry the accuracy: a kinematic-constraint step that removes spurious rotations from raw visual estimates using the known revolute-joint chain, and a continuity-based Euler-angle solver that chooses the configuration-space solution closest to the previous timestep to avoid dual solutions.","core_discovery":"The central discovery claimed is that proprioception for a rigid-link dexterous hand can be recovered from sensing mounted on the wrist and links rather than from joint hardware: per-link orientations come from small 9-axis IMUs, per-link full poses come from an RGB-D camera, and the two are combined in a modified extended Kalman filter that retroactively corrects the inertial trajectory whenever a delayed visual measurement arrives. On a dexterous hand with no onboard joint estimation system, the fused estimates stay within 2 to 4 degrees of ground truth during random teleoperated motion, while either modality alone is less accurate or drifts over time. The paper further claims that the IMU modules are small and stable enough that a common base frame across all modules is justified, which makes multi-sensor calibration simple.","pith_inferences":["The paper's own uniformity data show between-chip angular variances up to 8.21 square degrees, roughly 3 degrees of standard deviation; in a real hand with motors and wiring nearby, that floor may set the practical accuracy limit, so the 2-degree headline likely assumes a benign magnetic environment.","The same sensor stack could transfer to prosthetic hands or teleoperation gloves, where encoders are hard to fit; the main required change would be retraining the visual estimator on the new link shapes.","Because the paper notes that real images are cluttered and noisy while no noise was added during synthetic training, the zero-shot result is likely to degrade under occlusion or strong lighting changes; injecting synthetic noise or fine-tuning on a small set of real images is a direct test.","A natural extension noted in the paper is adding torque, compliance, and tactile signals; the same EKF state could be expanded from geometry to full interaction states."],"forward_implications":["A dexterous hand built without joint encoders can still close the loop on manipulation, provided it carries a wrist RGB-D camera and miniature inertial sensors.","Because the IMU stream updates at 200 Hz or more, the latency-compensated filter can produce real-time joint feedback while the slower camera refines the history.","If the common-base-frame calibration holds across a hand, adding more fingers or links does not require per-module alignment, lowering the cost of scaling to five-fingered hands.","The fully synthetic training pipeline implies that new hand geometries can get visual pose estimators without collecting real ground-truth poses.","Sustained errors of 2 to 4 degrees over long random motion imply the estimate is stable enough to serve as feedback for higher-level manipulation policies."],"supporting_citations":[{"why":"Supplies the dexterous hand platform with no onboard joint estimation that all experiments run on.","marker":"[8]"},{"why":"Provides the instance-level 6D pose estimation network that forms the visual branch's core.","marker":"[29]"},{"why":"Supplies the Euler-angle extraction formulas used to convert link rotations into joint angles.","marker":"[30]"},{"why":"Represents prior visual-inertial human hand pose estimation that the paper argues is not precise enough for manipulation.","marker":"[28]"},{"why":"Documents why tendon-driven indirect joint sensing is limited, motivating the search for an alternative proprioception method.","marker":"[12]"}],"fun_headline_variants":["Visual-inertial fusion tracks hand joints to a few degrees","Wrist cam and IMUs replace encoders with 2-4 deg accuracy","No joint sensors: fusion keeps hand pose drift-free","HandCept: fusion achieves 2-4 deg joint accuracy without encoders","Accurate hand proprioception from wrist camera and IMUs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The magnetic field around the hand is uniform enough that all of the hand's tiny inertial motion sensors can share a single calibration frame; if motors, wiring, or nearby metal distort that field, the relative biases that appear in the sensor orientations will show up directly as joint-angle errors.","fun_headline_variants_meta":{"raw":{"variants":["Visual-inertial fusion tracks hand joints to a few degrees","Wrist cam and IMUs replace encoders with 2-4 deg accuracy","No joint sensors: fusion keeps hand pose drift-free","HandCept: fusion achieves 2-4 deg joint accuracy without encoders","Accurate hand proprioception from wrist camera and IMUs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1532,"prompt_tokens":960,"completion_tokens":572,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":481}},"tokens_in":576,"tokens_out":572,"duration_ms":5162,"temperature":1.0,"reasoning_tokens":481,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:00:34.149190+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the system's uniformity test with four motion sensors mounted near the actuation motors of the hand and power the motors during measurement; if the between-sensor orientation variance grows beyond the roughly 3-degree standard deviation already seen in the paper's Table I, or if joint-angle error exceeds 4 degrees on any axis, then the common-calibration-frame assumption is not reliable enough for the claimed accuracy.","supporting_citations":[{"cited_title":"A dexterous and compliant (dexco) hand based on soft hydraulic actuation for human inspired fine in-hand manipulation,","cited_arxiv_id":null,"evidence_quote":"Supplies the dexterous hand platform with no onboard joint estimation that all experiments run on."},{"cited_title":"Ffb6d: A full flow bidirectional fusion network for 6d pose estimation,","cited_arxiv_id":null,"evidence_quote":"Provides the instance-level 6D pose estimation network that forms the visual branch's core."},{"cited_title":"Computing euler angles from a rotation matrix,","cited_arxiv_id":null,"evidence_quote":"Supplies the Euler-angle extraction formulas used to convert link rotations into joint angles."},{"cited_title":"Visual–inertial fusion-based human pose estimation: A review,","cited_arxiv_id":null,"evidence_quote":"Represents prior visual-inertial human hand pose estimation that the paper argues is not precise enough for manipulation."},{"cited_title":"A com- pliant, underactuated hand for robust manipulation,","cited_arxiv_id":null,"evidence_quote":"Documents why tendon-driven indirect joint sensing is limited, motivating the search for an alternative proprioception method."}],"review_version":1}