acceptodds
Under review as a conference paper at ICLR 2027

MARC: Asymmetric Memory Coupling for Myoelectric Hand Motion Decoding

Abstract

Streaming hand motion decoding from surface electromyography (EMG) benefits from motion history, but feeding predicted poses back into memory can also propagate their errors. We study how an explicit motion predictor should interact with persistent sensor memory. MARC combines joint-conditioned EMG encoding, observation-driven memory updates, and pose-conditioned motion prediction. Predictions contribute to the decoded pose without writing back into observation memory, giving the two state pathways an asymmetric dependency. Across emg2pose, MANUS, and NinaPro DB8, MARC improves mean reconstruction accuracy over reproduced TDS+LSTM and EMGFormer-L baselines. On emg2pose, it reduces angular error by 9.5–19.3% and fingertip error by 13.1–21.6% relative to TDS+LSTM, while improving displacement accuracy and motion direction. Matched feedback controls and complementary state interventions reveal a trade-off: restricting pose feedback reduces the output response to an injected pose error, but slows recovery of the unperturbed output after observation-memory erasure. Together, these results establish the practical value of MARC and identify prediction-to-memory feedback as a testable design choice in continuous myoelectric decoding.

Primary area:
learning on time series and dynamical systems
Keywords:
surface electromyography, hand motion decoding, streaming inference, recurrent memory, prediction feedback, structured sequence modeling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_EcUcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents MARC for continuous hand-pose decoding from surface EMG. Its central design separates observation-driven memory updates from pose-conditioned motion prediction: predicted poses contribute to the output but do not write back into observation memory. MARC also incorporates joint-conditioned encoding and auxiliary motion and geometry objectives. Experiments on emg2pose, MANUS, and NinaPro DB8 report improvements over reproduced TDS+LSTM and EMGFormer-L baselines. Additional feedback controls and state interventions examine pose-error propagation and recovery after observation-memory reset.

Critical strengths

The separation between observation memory and pose prediction is clearly specified. The matched observation-projection control is useful because it distinguishes the source of the added information from added parameter capacity.

The evaluation covers three recording settings and includes both pose and motion diagnostics. The paper also appropriately distinguishes recovery of a model’s unperturbed output from recovery of the ground-truth pose, and acknowledges the opposing effects of feedback under the two interventions.

Critical weaknesses

The experiments do not sufficiently isolate the contribution of the proposed memory organization. MARC changes the encoder, joint representation, prediction pathway, fusion mechanism, and training objectives together. The external baseline comparisons therefore establish the performance of the complete system, but do not establish which components account for its gains. Important missing controls include the same encoder and observation memory without the prediction pathway, a conventional recurrent decoder using the same encoded features, and ablations of the auxiliary objectives. Although the feedback controls partially address the central design choice, they use only one initialization and an 18-epoch budget (Appendix F), whereas the main emg2pose recipe allows up to 64 epochs and early stopping only after 24 epochs. Participant-bootstrap intervals do not establish robustness across training runs. These limitations leave uncertainty about the necessity and reproducibility of the proposed architectural contribution.

The benchmark evidence requires stronger validation. The reproduced EMGFormer-L obtains 14.788° and 15.500° on the unseen-movement and combined emg2pose splits, respectively, whereas its original paper reports 9.28° and 12.25° (EgoEMG, Appendix D.3, Table 11). These results are not necessarily directly comparable: causal restrictions and evaluation-protocol changes may explain the discrepancy. However, the paper does not sufficiently reconcile this substantial gap to establish that the adapted baseline remains competitive. In addition, Appendix C.4 explicitly states that test comparisons informed architecture development. Selecting checkpoints on validation data does not remove potential selection bias from test-informed architectural decisions. The scope of this test-set use and the availability of an untouched final evaluation need clarification.

Missing validation in a real online system. For a method motivated by streaming myoelectric decoding and practical use, I consider evaluation with a live EMG acquisition system necessary. Experiments on prerecorded datasets establish offline decoding accuracy, but do not establish whether the proposed system can maintain accurate and timely tracking during actual operation. The timing results in Appendix G exclude acquisition, host transfer, and visualization, and therefore do not measure end-to-end latency. Likewise, manually perturbing internal states does not establish robustness to disturbances encountered during live sensing. A convincing evaluation should demonstrate sustained online tracking with synchronized reference measurements, report end-to-end latency distributions and missed update deadlines, and compare MARC with the baselines and a simplified variant on identical hardware. Without this evidence, the practical benefit of the additional memory and prediction components remains insufficiently validated.

Questions for authors

Can you report matched controls that remove the prediction pathway or replace the specialized decoder with a conventional recurrent decoder, together with auxiliary-loss ablations? Do the current memory-routing rankings persist across multiple seeds and the main training schedule?

What explains the large EMGFormer-L discrepancy on the unseen-movement and combined splits? Can you separate the effects of causalization, evaluation windows/alignment, and training objectives, and provide the exact baseline configurations?

Which architectural decisions were informed by test comparisons, and which datasets or splits were consulted? Were any final results obtained on evaluation data that had not been used during architecture development?

Can the authors include more recent and methodologically relevant baselines, particularly those incorporating explicit motion prediction or structured memory, under the same causal evaluation protocol?

5 comments
AuthorsOpenReview ↗

Response to Reviewer EcUc (1/2): baseline audit and evaluation scope

Thank you for identifying the discrepancy with the published EMGFormer-L results. We independently checked the public reference and our evaluation. The audit explains much of the discrepancy, while the matched causal comparisons establish MARC's gains in pose and motion reconstruction.

Q2. What explains the baseline discrepancy?

The authors' correction withdraws the earlier scores: 16,831 training records had entered the unseen-movement evaluation. The checkpoint weights are unchanged. The emg2pose participant-macro angular MAE (°) is:

Author EMGFormer-L referenceUnseen movementsUsers and movements unseen
Earlier result, withdrawn9.28012.250
Corrected result13.71614.845

We independently recomputed these macro scores from released aggregates. Our own index contains 6,167 records, all test-labeled, with the corrected group counts. The apparent gap between our causal reproduction and the author reference falls from 5.508° to 1.072° for unseen movements and from 3.250° to 0.655° for the combined setting. For identical author predictions, macro versus micro aggregation changes MAE by only 0.056–0.081°.

This correction narrows the discrepancy; it does not make the protocols identical. Noncausal context, windows/alignment, normalization and training still differ, and their individual effects have not been isolated. Adapted architectures and training recipes are specified in Appendix B, pp. 13–14; the reference audit is in Table 22, p. 28.

MARC's gains come from the separate matched causal comparison. All three models below receive causal EMG inputs and use common scored timestamps and participant aggregation. emg2pose angular MAE (°), mean ± SD over three training runs:

ModelUnseen usersUnseen movementsUsers and movements unseen
TDS+LSTM12.861 ± 0.05216.380 ± 0.05016.519 ± 0.067
EMGFormer-L12.272 ± 0.00514.788 ± 0.33215.500 ± 0.098
MARC11.645 ± 0.04413.217 ± 0.24314.527 ± 0.104

MARC reduces MAE by 9.5–19.3% versus TDS+LSTM and 5.1–10.6% versus causal EMGFormer-L. Relative to EMGFormer-L, gains cover 19/20 unseen users, 157/158 participants on unseen movements and 19/20 in the combined setting (Table 2/Figure 3, p. 6).

The benefit also appears in hand geometry and motion. On unseen emg2pose movements, the same models give:

ModelFingertip distance (nominal mm) ↓100 ms displacement MAE (°) ↓100 ms direction cosine ↑
Causal EMGFormer-L34.148 ± 0.6775.631 ± 0.0610.512 ± 0.015
MARC29.769 ± 0.6765.018 ± 0.0340.589 ± 0.013

This corresponds to 12.8% lower fingertip error and 10.9% lower displacement error, alongside better direction consistency. These are reconstruction metrics, with motion measured between outputs 100 ms apart (Figure 4, p. 7; Table 11, p. 18; Figure 7, p. 19). The gains are therefore visible beyond aggregate angular error.

Q3. Checkpoint selection and development scope.

Validation selected run checkpoints. Earlier test comparisons informed family/capacity choices across emg2pose, MANUS and DB8, plus emg2pose state/feedback candidates, MANUS protocol variants and DB8 objectives. Final results and the added controls use these same, previously consulted test splits (Sec. 4.1, p. 5; Appendix C.4, p. 16).

We hope the independently checked correction, explicit configurations and matched results address the baseline discrepancy that motivated your concern. We would appreciate your reassessment of baseline validity on this evidence.

AuthorsOpenReview ↗

Response to Reviewer EcUc (2/2): component controls and continuous operation

Thank you for requesting complete-training controls, loss ablations and a recent structured-memory comparison. These experiments are complete.

Q1 and Q4. Prediction-free, LSTM and Gated DeltaNet controls.

Each control used three runs and the main stopping rule (minimum 24 epochs, maximum 64, patience 20). Two original write controls continued training; nine other groups began from scratch. LSTM and Gated DeltaNet retain the prediction-free encoder/readout architectures and objective.

On emg2pose, prediction-free has lower mean MAE than the full architecture under the same non-future objective on two of three settings. Relative to prediction-free, the three setting means shift by −0.032 to +0.060° for LSTM and −0.007 to +0.257° for Gated DeltaNet (Table 15, p. 25). Direct state interventions complement these reconstruction controls by isolating what pose feedback changes.

Feedback has a measurable effect on error propagation and memory-loss response.

The following emg2pose results measure deviation from each model's own clean output 200 ms after intervention. Values are pose-write minus MARC, mean ± SD of three paired runs (°): positive favors MARC; negative favors pose-write.

InterventionUnseen usersUnseen movementsUsers and movements unseen
Stored pose perturbed by ±5°+1.303 ± 0.280+1.259 ± 0.184+1.310 ± 0.183
Observation memory alone erased−1.161 ± 0.230−1.761 ± 0.238−1.840 ± 0.113

All nine paired run–setting differences have the corresponding sign for each intervention. Restricting pose writes reduces the propagation of injected pose errors; allowing writes reduces clean-trajectory deviation after observation-memory loss (Figure 5, p. 7). This comparison measures trajectory deviation, not a recovery-rate advantage.

Q1. Loss ablations identify angular and spatial benefits on unseen movements.

Each row removes one loss while retaining the full inference architecture and other coefficients. emg2pose, unseen movements, mean ± SD over three runs; fingertip distance uses identical generic-hand forward kinematics for predictions and targets:

Training objectiveAngular MAE (°)Fingertip distance (mm)
Full MARC13.217 ± 0.24329.769 ± 0.676
Without observation loss13.389 ± 0.04130.096 ± 0.030
Without displacement loss13.551 ± 0.18630.634 ± 0.648
Without future loss13.324 ± 0.61430.103 ± 1.693
Without geometry loss13.249 ± 0.25330.836 ± 0.755

All four removals increase both metric means in this setting; their angular effects elsewhere vary. Across all three settings, removing geometry supervision increases mean fingertip error by 0.85–1.07 mm, with increases in all nine paired comparisons (Figure 6, p. 8; Table 14, p. 24).

Continuous operation and measured execution time.

Across 845 MANUS trials, maintaining state reduces mean error relative to clearing the complete streaming state every five seconds. Angular MAE (°); reduction is reset minus continuous, mean ± SD of three paired runs:

Test settingFive-second resetContinuousReduction
Unseen users10.85010.6840.166 ± 0.015
Unseen gestures14.33614.0800.256 ± 0.065
Users and gestures unseen14.83714.8130.024 ± 0.030

This is prerecorded replay (Table 21, p. 28). On one RTX 5090, FP32/batch one, MARC's input-transfer-inclusive p95 update time is 5.703 ms, measured with 20 ms input packets and 5.88 s emg2pose context (Table 23, p. 29). Acquisition hardware was unavailable; synchronized live accuracy, sensor-to-display latency and live deadlines remain untested.

We would appreciate your reassessment of the training-budget and missing-control concerns in light of the completed comparisons, consistent spatial-supervision evidence and directly measured feedback effects.

AuthorsOpenReview ↗

Response to Reviewer EcUc: Follow-up on the baseline comparison and completed controls

Thank you again for your careful review. Your questions prompted us to examine the baseline reference more closely and strengthen the controlled experiments. We would particularly appreciate your view on the following evidence addressing the baseline comparison and training budget.

The EMGFormer authors' correction substantially changes the reference underlying the discrepancy raised in your review:

Author EMGFormer-L reference: angular MAE (°)Unseen movementsUsers and movements unseen
Earlier, withdrawn9.28012.250
Corrected13.71614.845

The author correction and our independent audit are documented in Table 22, p. 28.

Separately, our matched causal comparison shows that MARC's gains extend beyond angular accuracy. On unseen emg2pose movements, the participant-macro results are:

Reconstruction metricCausal EMGFormer-LMARC
Angular MAE (°) ↓14.788 ± 0.33213.217 ± 0.243
Fingertip error (nominal mm) ↓34.148 ± 0.67729.769 ± 0.676
100 ms displacement MAE (°) ↓5.631 ± 0.0615.018 ± 0.034

These are 10.6%, 12.8% and 10.9% reductions, respectively, with values reported as mean ± SD over three runs. The comparison therefore supports improvements in pose accuracy, spatial reconstruction and motion reconstruction within this setting (Table 2, p. 6; Figure 4, p. 7; Table 11, p. 18).

The requested prediction-free, same-encoder LSTM, Gated DeltaNet and individual-loss controls have also completed three runs under the main stopping rule (Tables 14–15, pp. 24–25). One consistent result is that removing geometry supervision increases mean fingertip error by 0.849–1.067 mm across the three test settings, with increases in all nine paired run–setting comparisons. This directly supports the spatial benefit of that training objective.

Do the corrected reference and completed controls address your concerns about baseline validity and training adequacy? If a specific comparison or protocol detail remains important to your assessment, we would be glad to clarify it. Thank you again for helping us strengthen the experimental evidence and its presentation.

Reviewer_DAm4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MARC, Asymmetric Memory Coupling, for continuous hand pose estimation from sEMG signals. The model combines a joint-conditioned EMG encoder, observation memory, and motion prediction, while excluding direct pose feedback from observation memory updates to limit error propagation. Experiments on three public datasets show lower errors than TDS+LSTM and EMGFormer-L across the evaluated settings, including improved fingertip accuracy and motion reconstruction. The ablations also show a trade-off: pose feedback increases error propagation but helps recovery toward the original output trajectory after observation memory loss.

Critical strengths

MARC separates observation memory updates from motion prediction to examine how previous pose estimates affect prediction accuracy and stability. Controlled experiments compare memory update strategies and show how pose feedback affects both error propagation and recovery after memory loss, providing useful insights into memory design for sequential pose estimation.

Critical weaknesses
  1. Artificial pose perturbations may not reflect actual prediction errors. Previous estimates can still provide useful temporal information, but the paper does not sufficiently examine how their benefit depends on error magnitude and characteristics.

  2. Both MARC and Pose Feedback use historical EMG information. The performance gap alone does not sufficiently explain why excluding direct pose feedback improves performance when observation memory already contains past EMG information.

Questions for authors
  1. Could the authors provide component ablations to separate the contribution of Asymmetric Memory Coupling from those of the encoder, motion prediction, and fusion modules?

  2. Beyond fixed-magnitude artificial perturbations, does MARC outperform Pose Feedback across different magnitudes of naturally occurring estimation errors? Could the authors provide results grouped by error magnitude?

  3. Given their shared use of historical EMG information, could adjusting how pose estimates enter observation memory achieve comparable accuracy and robustness without fully excluding direct pose feedback?

3 comments
AuthorsOpenReview ↗

Response to DAm4: isolating prediction-to-memory feedback

Thank you for asking us to separate memory access from overall accuracy. Our new emg2pose controls connect pose feedback to both error propagation and context after memory loss. We test this through matched input sources, learned gains and natural-error strata.

1. Memory access controls error propagation and retained context.

The matched controls add equal-capacity, zero-initialized projections, fed either the current EMG-derived pose observation or the previous predicted pose. Both retain EMG history; the changed dependency is whether a predicted pose can directly write observation memory. After a ±5° stored-pose perturbation, deviation from each model's clean output at +200 ms is:

Model / deviation (°; three-run mean ± SD)Unseen usersUnseen movementsBoth unseen
MARC1.318 ± 0.4411.092 ± 0.2851.079 ± 0.312
Observation-write1.391 ± 0.5311.127 ± 0.4851.119 ± 0.477
Pose-write2.622 ± 0.1862.351 ± 0.1822.389 ± 0.174

Observation-write remains close to MARC, supporting an effect of the admitted information rather than added capacity alone. Ground-truth comparisons also show a cost of admitting pose errors. Below, values are fixed pose-write minus MARC, paired-run mean ± SD:

Induced-error diagnosticUnseen usersUnseen movementsBoth unseen
±5° pose injection: excess GT MAE after 200 ms (°)+0.197 ± 0.053+0.117 ± 0.066+0.125 ± 0.051
200 ms EMG outage: excess-error AUC (°·s)+0.111 ± 0.382+0.264 ± 0.237+0.155 ± 0.131

The first row compares the perturbation-induced increase over each model's clean-run ground-truth MAE; AUC integrates interrupted-minus-clean ground-truth error. Both mean contrasts favor MARC across settings, while absolute recovery MAE remains mixed (Table 13, p. 22; Table 20, p. 28). Conversely, pose-write reduces clean-output deviation after memory erasure by 1.16–1.84°, evidence that pose history supplies context when observation memory is lost (Figure 5, p. 7).

The matched controls support an information-routing explanation. Prediction-free and fusion-free variants can match or improve clean MAE (Table 15 and Appendix G.1, pp. 23–25), distinguishing this feedback effect from overall reconstruction accuracy. An encoder-only replacement remains untested.

2. Does the effect also appear across natural-error levels?

We compared the same 768 windows, grouped by a fixed TDS+LSTM reference with cutoffs of 10.487° and 15.150° fitted outside test. Group-mean own-anchor error increases from low to middle to high in all 27 model–setting–run combinations (Table 17, p. 26). With EMG continuing normally, angular MAE at 1.6 s after the anchor gives the following pose-write minus MARC differences (°; paired-run mean ± SD; positive favors MARC):

Reference-error groupUnseen usersUnseen movementsBoth unseen
Low+0.427 ± 0.504−0.538 ± 0.894+0.118 ± 0.461
Middle−0.127 ± 0.609−0.036 ± 0.915−0.296 ± 0.260
High+0.211 ± 0.424−0.065 ± 0.442+0.019 ± 0.275

MARC has lower mean error in the low/high groups for unseen users, but the pattern varies across settings and is not monotonic in initial error. Thus natural error magnitude alone does not determine the benefit of pose feedback; the controlled interventions isolate its propagation pathway (Table 18, p. 27).

3. Does learning the write strength remove the feedback effect?

Following your suggestion, we learned a joint-wise gain on the same pose-write projection, initialized at one and bounded in (0,2). Values below are learned-write minus MARC in clean-output deviation at +200 ms (degrees; paired-run mean ± SD):

State interventionUnseen usersUnseen movementsBoth unseen
±5° pose error+1.380 ± 0.355+1.304 ± 0.225+1.293 ± 0.237
Observation-memory erasure−1.383 ± 0.406−1.933 ± 0.537−2.007 ± 0.438

The opposing responses persist when write strength is learned, so the pattern is not specific to a fixed gain. Learned feedback is also a useful alternative: full-test mean MAE changes span approximately −0.08° to +0.16° relative to MARC; during a 200 ms EMG outage it lowers mean excess-error AUC by 0.023–0.163°·s, with setting-dependent absolute recovery error (Table 15, p. 25; Figure 5, p. 7; Tables 24–25, p. 30).

Across fixed and learned gains, restricting pose-to-memory feedback reduces sensitivity to injected pose errors. The matched-source controls support an information-routing explanation, while the natural-error analysis clarifies its relation to reconstruction accuracy. We hope these additions address your three questions and would appreciate your reassessment of the mechanism evidence.

Thank you for the additional experiments and detailed response. The new results clarify how pose feedback affects error propagation and recovery after memory loss. These additions have addressed most of my concerns, and I will raise my score to 3.

AuthorsOpenReview ↗

Response to Reviewer DAm4: Thank you for your thoughtful feedback

Thank you for your encouraging reassessment and for recognizing the value of our additional experiments.

Your questions prompted us to examine pose feedback more carefully and distinguish error propagation from recovery after memory loss. Your suggestions substantially strengthened both the experimental evidence and the clarity of the paper.

We sincerely appreciate the time and care you devoted to reviewing our work and engaging with our rebuttal.

Reviewer_2VNgOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work focuses on motion decoding from EMG and studies how the memory of motion history should interact with motion prediction, since feeding predicted poses back into memory can propagate errors. The authors propose MARC, a motion decoder with asymmetric memory coupling: EMG signals update an observation memory, and a motion predictor uses this memory to predict the pose, but the predictions do not feed back into the observation memory. MARC is evaluated on the emg2pose, MANUS, and NinaPro DB8 datasets, examining whether MARC improves accuracy across recording settings, whether its gains extend to hand geometry and movement, and how giving predicted poses access to observation memory affects error propagation and recovery. The results show that MARC outperforms the TDS+LSTM and EMGFormer-L baselines, and reveal a trade-off between reducing the effect of injected pose errors and recovery speed.

Critical strengths

Thorough evaluation: MARC consistently outperforms the TDS+LSTM and EMGFormer-L baselines on three datasets, including unseen users and movements. The evaluation also covers spatial and temporal reconstruction (e.g., fingertip errors, displacement, and motion direction), ablations on memory access, and controlled experiments on error propagation and recovery. This shows that MARC performs well and generalizes, and it supports the paper’s claims.

Critical weaknesses

Flow and structure: The formal definitions of functions and variables, including the loss functions, are provided only in the appendix and not in the main body. This makes the methods section especially hard to follow without constantly referring to the appendix. A self-contained methods section is important for readers to understand and assess the design.

Questions for authors
  • Have the authors performed an ablation study on the loss terms? It would help to understand how much each term contributes to the final performance.

Regarding the weaknesses section above, moving the key definitions of functions, variables, and loss functions from the appendix into the main body would make the methods section self-contained and easier to follow.

3 comments
AuthorsOpenReview ↗

Response to 2VNg: completed loss ablations and a self-contained method

Thank you for recognizing the cross-dataset evaluation and for requesting loss ablations and clearer method definitions. We have completed both revisions. The new results show how the objectives contribute to angular and spatial reconstruction.

1. The requested loss ablations are now complete.

Each control retains the full inference architecture and removes one loss, with all remaining coefficients fixed. We first show unseen emg2pose movements, where the angular and spatial effects can be compared directly. These are complete-test participant-macro means ± SD over three runs; lower is better. Fingertip distance uses the same generic-hand forward kinematics for predicted and target poses.

Training objectiveAngular MAE (°)Fingertip error (nominal mm)
Full MARC13.217 ± 0.24329.769 ± 0.676
Without observation loss13.389 ± 0.04130.096 ± 0.030
Without displacement loss13.551 ± 0.18630.634 ± 0.648
Without fingertip loss13.249 ± 0.25330.836 ± 0.755
Without future loss13.324 ± 0.61430.103 ± 1.693

On this setting, each removal raises both mean errors. The effects differ: displacement supervision reduces mean angular MAE by 0.334°, whereas geometry supervision produces the largest mean fingertip improvement, 1.067 mm.

The geometry result also extends across all three test settings:

Fingertip error (nominal mm)Unseen usersUnseen movementsUnseen users and movements
Full MARC24.262 ± 0.30829.769 ± 0.67633.085 ± 0.442
Without fingertip loss25.111 ± 0.12830.836 ± 0.75534.041 ± 0.116
Increase after removal+0.849+1.067+0.956

All nine paired run–setting comparisons worsen when geometry supervision is removed. Its removal also raises mean angular MAE by 0.204°, 0.032° and 0.117° in the respective settings. Displacement and future supervision lower mean fingertip error in all three settings, although angular effects of these and observation supervision vary by setting. The objectives contribute differently: geometry supervision yields a consistent spatial benefit that angular MAE alone understates. Figure 6, p. 8 and Table 14, p. 24 give all removals and settings.

The complete decoder's gains also extend to MANUS and DB8, which do not use the geometry loss. All their test settings are shown below, using the lower-mean reproduced baseline in each row (three-run mean ± SD; MANUS in degrees, DB8 in its released calibrated sensor scale; Table 2, p. 6):

Dataset / test settingStronger reproduced baselineMARC
MANUS: unseen usersEMGFormer-L: 10.987 ± 0.15710.684 ± 0.454
MANUS: unseen gesturesTDS+LSTM: 14.499 ± 0.05214.080 ± 0.685
MANUS: users and gestures unseenEMGFormer-L: 15.239 ± 0.21714.813 ± 0.392
DB8: third acquisitionEMGFormer-L: 7.179 ± 0.0346.593 ± 0.095

These whole-model comparisons complement the emg2pose objective ablations: the former establish breadth across recording settings, and the latter identify the role of each supervision signal.

2. The main method now defines the state, update and objective.

Section 3 of the revised manuscript, pp. 3–4, introduces each quantity when it is first used. Observation memory hth_t is updated from its previous value and arrived EMG. The motion branch additionally reads the preceding predicted pose, but its temporary state htph_t^p does not overwrite observation memory. The output is

q^t=ot+gt⊙(pt−ot)+rt,\widehat q_t=o_t+g_t\odot(p_t-o_t)+r_t,

where oto_t is the observation estimate, ptp_t the motion prediction, gtg_t the joint-wise gate and rtr_t the bounded residual. This makes the central distinction explicit: pose history contributes to the output without entering the observation-memory update.

Equations 1–4 now define the state updates, output composition and complete training objective, including coefficients and normalization. Tables 3–4, pp. 11–13 provide tensor dimensions and the streaming algorithm. These definitions let readers follow the information flow and understand each loss before consulting implementation details.

The completed ablations explain the objectives' complementary roles, and the method now makes the state dependencies and training objective explicit. We hope these additions resolve your two concerns and strengthen your assessment of the paper.

I thank the authors for their detailed responses, which addressed my questions. I maintain my positive assessment of the paper.

AuthorsOpenReview ↗

Response to 2VNg: Thank you for your constructive feedback and continued positive assessment

Thank you for maintaining your positive assessment and for confirming that our responses addressed your questions.

Your constructive suggestions helped us make the method section more self-contained and strengthen the analysis of the training objectives through additional ablations.

We sincerely appreciate your careful review and thoughtful engagement throughout the discussion, which have improved both the clarity of the paper and its experimental support.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Official Comment by Authors

Dear Area Chair,

Thank you and all the reviewers for your time and thoughtful feedback. Your suggestions helped sharpen the paper’s central question and strengthen its experimental evidence.

Our paper studies a concrete design question in streaming inference: how should predicted poses interact with persistent sensor memory? MARC lets pose-conditioned prediction assist decoding without writing predicted poses into observation memory.

The revised paper offers both broad reconstruction gains and controlled evidence about memory design. MARC achieves lower mean reconstruction error than both reproduced baselines across all seven settings from three datasets.

On emg2pose, angular error decreases by 9.5–19.3% relative to TDS+LSTM and 5.1–10.6% relative to causal EMGFormer-L. After averaging training runs, MARC improves participant MAE over the latter for 19/20 participants in the unseen-user setting, 157/158 in the unseen-movement setting, and 19/20 when both are unseen (Table 2; Figure 3). New controlled interventions clarify how pose feedback propagates injected errors and supplies context after observation-memory loss, under both fixed and learned feedback (Figure 5). We also completed prediction-free and recurrent-operator controls, loss ablations, and natural-error analyses, and clarified the method and baseline configurations (Tables 14–15, 22; Appendix B).

Two reviewers have since explicitly acknowledged the value of these additions:

  • Reviewer DAm4 stated that the additional experiments had “addressed most of my concerns”.
  • Reviewer 2VNg confirmed that our detailed responses “addressed my questions.”

As of this note, Reviewer EcUc has not participated in the rebuttal discussion or responded to our detailed replies and follow-up messages. We would appreciate your consideration of their original concerns alongside the completed controls, baseline audit, and detailed clarifications now available in our responses.

We respectfully ask you to consider the paper for acceptance on the strength of its reconstruction gains, controlled findings on memory feedback, and the evidence strengthened during rebuttal. Thank you again for your careful consideration.

Best regards, The authors of Paper 60875

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.