acceptodds
Under review as a conference paper at ICLR 2027

STAMP: TASK-STRUCTURED TRANSFORMERS FOR STREAMING SEMG-TO-POSE

Abstract

Streaming hand-pose estimation from surface electromyography (sEMG) requires causal reconstruction from noisy, non-stationary signals. We propose STAMP, a task-structured Transformer that organizes electrode interactions, temporal evidence, pose evolution, and motion-dependent computation. SpatioTemporal Electrode Modeling (STEM) preserves the electrode axis and combines spatial softmax attention with temporally decayed causal linear attention, informed by sensing geometry. Affine Scan provides compact, parallelizable pose updates through gated propagation and correction, while history-conditioned Motion MoE adapts computation across timescales. On emg2pose, STAMP achieves the lowest cross-user mean angular and whole-hand landmark errors in every Regression and Tracking split among matched retrained implementations. Controlled replacements, including near-parameter-matched alternatives, support structured electrode modeling, historical expert selection, and complementary pose propagation and correction. At nearly matched capacity, STAMP has lower errors and steeper fitted scaling than vEMG2Pose over 32–158 training users, outperforming the 158-user baseline with only 95 users. Increasing capacity generally yields further gains within the measured grid. Dataset-specific retraining with interface-only adaptation also yields the lowest pooled errors on EgoEMG and NinaPro, supporting architectural portability across electrode layouts.

Primary area:
learning on time series and dynamical systems
Keywords:
surface electromyography (sEMG), continuous hand pose estimation, causal sequence modeling, streaming inference, structured Transformers, mixture-of-experts, cross-user generalization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FDiVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents STAMP, a task-structured Transformer for streaming hand-pose estimation from surface electromyography (sEMG) signals. It introduces three components: a STEM encoder for spatiotemporal electrode modeling, an Affine Scan decoder for pose propagation and correction, and a history-conditioned Motion MoE for adaptive computation. Experiments on emg2pose, EgoEMG, and NinaPro demonstrate improved pose estimation accuracy over retrained baselines. Additional ablations, scaling studies, and streaming evaluations support the effectiveness and practicality of the proposed architecture.

Critical strengths
  • The proposed Motion MoE leverages multi-scale motion histories for adaptive expert selection, improving the model's ability to capture different motion patterns and enhancing pose estimation accuracy.
  • The replacement studies are useful. Near-parameter-matched temporal attention and wider non-MoE controls, together with routing and decoder replacements, provide evidence beyond a simple capacity explanation.
  • I also appreciate the explicit distinction between native streaming causality, benchmark alignment, and the steady-state execution latency.
Critical weaknesses
  • My main concern is experimental robustness. The experiments do not sufficiently validate the effectiveness of Motion MoE across different motion regimes or STEM's robustness to electrode displacement. The main emg2pose comparison and cumulative ablations use a single seed, and the scaling experiments also lack repeated seeds and participant-subset sampling. This limits confidence in both the smaller performance differences and the fitted scaling slopes.
  • STAMP shows limited improvements on unseen users in NinaPro and the audited EgoEMG protocol, raising concerns about its generalization ability across different users and datasets.
  • The released-protocol EgoEMG results have split and model-selection limitations. The additional audited evaluation is valuable and should carry more weight in the main presentation. A unified comparison with the complete PI-EMG method would also strengthen the accuracy claims.
  • Long-horizon tracking exposes an important limitation. At 30 seconds, the original absolute-position configuration performs worse than the baseline.
Questions for authors
  1. To demonstrate the effectiveness of MotionMoE, the authors compare it with a parameter-matched wider backbone. Could the authors clarify the specific architecture of this wider backbone. Furthermore, the paper argues that a fixed computation path must compromise across static, transitional, slow, and fast motion regimes, motivating the proposed Motion MoE. However, the current experiments do not directly demonstrate whether Motion MoE alleviates this limitation. Could the authors provide a comparison of STAMP and the baselines across these different motion regimes to verify the benefits of motion-dependent expert selection?
  2. The paper motivates STEM by highlighting the importance of electrode spatial relationships and the sensitivity of sEMG signals to electrode placement shifts. Although the authors compare STEM with a temporal Transformer and evaluate circular positional encoding, the current experiments do not directly demonstrate robustness to electrode displacement. Could the authors evaluate STAMP under different electrode placement conditions or controlled sensor perturbations to verify whether the proposed spatial modeling improves robustness compared with the baselines?
  3. Although STAMP achieves strong performance on emg2pose, its improvements are less consistent across other datasets. In particular, STAMP underperforms NeuroPose on the User and User+Stage splits of NinaPro, while its performance gains become marginal under the audited EgoEMG protocol. Could the authors explain why the proposed method shows less consistent improvements in cross-user generalization across different datasets?
  4. While the motivation and mathematical formulations are reasonably clear, the correspondence between Figure 2 and the method description could be improved. Several architectural components, such as the multi-scale Patch MLP and the two decoder attention are only briefly described in the main text. The Method section relies heavily on mathematical formulations, many of which would be more appropriate for the appendix. This makes it difficult to follow the complete data flow directly from the method section and Figure 2. The authors should simplify the equations in the main text and focus on clearly explaining the inputs, outputs, functions, and roles of each module within the overall architecture, while moving detailed mathematical derivations to the appendix.
2 comments
AuthorsOpenReview ↗

Motion-Dependent Performance, Sensor Perturbation Tests, and Cross-User Generalization

Thank you for the suggestions. We have added three-seed evaluations and will revise the presentation as follows. Unless stated otherwise, results are mean angular errors; ± denotes the sample SD across seeds 42, 43, and 44.

1. Wider backbone and motion regimes. The wider control retains electrode-wise patches, spatial softmax attention, causal temporal attention, feed-forward layers, dual-query pooling, and the two-stage Affine Scan. It removes all three Motion MoE branches and widens the remaining encoder, giving 6.056M parameters versus STAMP's 6.059M under the same training and evaluation protocol.

On emg2pose, training-derived thresholds define low-motion, slow, fast, and transition regimes. At 50 Hz, STAMP outperforms the evaluated EMGFormer-M, NeuroPose, and single-task vEMG2Pose baselines in all four regimes across the three-seed evaluation. Fast-motion errors are 12.049° versus 13.286° for EMGFormer-M; transition errors are 13.109° versus 14.390°. These results establish full-model performance across regimes; the capacity-matched control addresses the contribution beyond model size.

2. Controlled sensor perturbations. Three-seed emg2pose tests cover circular channel shifts of ±1/±2, zeroing channel 0 or 8, and alternating gains of 0.5/1.5. STAMP has lower absolute error than the evaluated baselines under all shifts and channel-zeroing conditions. Under gain perturbation, STAMP/EMGFormer-M/vEMG2Pose obtain 17.678°/15.766°/16.831°.

On EgoEMG, we also trained STAMP and EMGFormer-M with identical channel-shift augmentation. Averaging left/right one-channel shifts over the first five seconds and 20 finger angles, errors change from 16.359°/18.163° to 15.535°/15.353°. Both improve for all 32 paired users. We will report absolute errors and degradation from clean inputs, and distinguish these synthetic perturbations from physical re-donning.

3. Cross-user results and EgoEMG. On NinaPro, STAMP improves Stage performance to 16.089° versus EMGFormer-M's 17.759°, but trails NeuroPose by 0.350° on User and 1.060° on User+Stage. Thus, the overall benefit comes mainly from unseen gestures performed by seen users.

For EgoEMG, the restored strict-v1 split with independent validation-based checkpoint selection will become the main result; released-protocol results will move to the appendix. The three-seed overall regression errors are 15.533±0.212° for STAMP and 14.991±0.271° for EMGFormer-M. STAMP performs better in User and User+Stage but worse in Stage; 26 of the 32 scored users contribute only Stage windows, giving this condition substantial weight in the user-balanced mean.

Low motion helps explain this pattern. In EgoEMG Stage, static or very-low-speed samples account for 61.3% under the matched 20-finger-angle analysis. STAMP trails EMGFormer-M by 1.105° in this group but leads by 0.678° in fast motion. Tracking shows a related effect: holding the provided initial pose already gives 9.164° overall, close to the trained models' 8.976°–9.171°. Splitting each user's windows equally by ground-truth displacement from the initial pose, STAMP/vEMG2Pose errors are 5.969°/5.543° in the lower-displacement group and 12.108°/12.300° in the higher-displacement group. The ranking reverses in all three seeds. Their contributions to the original user-balanced gap are +0.209° and −0.097°, yielding +0.112° overall. We will report these conditions separately and describe cross-dataset retraining as evidence of architectural portability.

4. Method presentation. Figure 2 will show EMG → dual-scale electrode-wise Patch MLP → STEM → dual-query pooling → two-stage Affine Scan, with both decoder attentions, all three MoE locations, temporal strides, and cached states. We will retain the core update equations in the main text and move detailed derivations to the appendix. Regression and Tracking initialization will be labeled explicitly.

Additional robustness and streaming results. Three seeds cover the main comparisons, new STEM controls, and routing sensitivity. On emg2pose, STAMP/EMGFormer-M obtain 12.762±0.228°/14.199±0.169° in Regression; STAMP/single-task vEMG2Pose obtain 9.774±0.050°/10.826±0.013° in Tracking. Original cumulative ablations and scaling curves remain single-seed; participant-subset sampling was not repeated, so scaling trends will be treated as exploratory. A unified comparison with full PI-EMG is not included in the current results.

We recommend Absolute + MOD-295, applied to the trained absolute-position model without additional fine-tuning. With the same checkpoint per task and seed, Tracking errors at 5 s/30 s are 9.810±0.048°/12.997±1.147° under native scoring and 9.838±0.046°/13.034±1.144° under past-only scoring. The original MOD-295 engine measured a steady-state P99 of 1.216 ms on an RTX 5060 Laptop; this is the existing engine measurement, not a new timing run for these checkpoints.

AuthorsOpenReview ↗

imited Sensitivity of Aggregate EgoEMG Tracking Scores to Dynamic Performance

We provide an additional analysis to clarify why the aggregate EgoEMG tracking scores are close and how they reflect dynamic tracking performance.

1. A strong initial-pose baseline

On the restored strict-v1 split, holding the supplied ground-truth initial pose throughout each five-second window yields 9.164° joint-angle MAE. The three-seed means of the four trained methods range from 8.976° to 9.171°. Thus, a competitive aggregate score can be obtained without estimating subsequent motion from EMG.

2. Motion stratification reveals different rankings

For each user, we divided test windows into lower- and higher-displacement halves according to the ground-truth pose deviation from the initial pose. The groups contain 832 and 843 windows, respectively, and both cover the same 32 users. The table reports joint-angle MAE, averaged over three seeds for the trained methods.

MethodLower displacementHigher displacement
Hold initial pose5.460°12.749°
STAMP5.969°12.108°
EMGFormer-M5.492°12.731°
NeuroPose5.494°12.698°
vEMG2Pose (velocity integration)5.543°12.300°

Holding the initial pose has the lowest error in the lower-displacement group, whereas STAMP has the lowest error in the higher-displacement group. Based on each user's three-seed mean, vEMG2Pose performs better than STAMP for all 32 users in the lower-displacement group; STAMP performs better for 25/32 users in the higher-displacement group. The group-level ranking reversal between these two methods occurs in all three seeds.

Under the paper's original equal-user aggregation, the lower- and higher-displacement groups contribute +0.209° and −0.097°, respectively, to the overall STAMP-minus-vEMG2Pose difference of +0.112°. The aggregate comparison therefore combines a penalty at lower displacement with a gain at higher displacement.

3. Implication for evaluation

These results suggest that the aggregate EgoEMG tracking metric has limited sensitivity to differences in dynamic tracking performance: preserving the initial pose is already competitive, and opposing effects across motion regimes partly cancel. We will report the initial-pose baseline and motion-stratified results alongside the aggregate score to make these differences explicit.

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

STAMP estimates hand pose from wrist sEMG using an electrode-aware STEM encoder, a two-stage Affine Scan decoder, and expert routing conditioned on recent signals. The affine formulation allows parallel training and causal execution at inference. The authors evaluate the model on emg2pose, retrain it across datasets, and include parameter-matched comparisons, ablations, and longer streaming tests. On five-second emg2pose Regression, the reported angle error falls from 13.99 degrees for the strongest listed retrained comparator to 12.48 degrees. My reservations concern whether the gains survive retraining, what the dataset protocols establish about generalization, and how the model performs during sustained streaming.

Critical strengths
  1. The design gives electrode structure and recent signal history explicit roles in hand-pose estimation. Recurrence operates in pose space, while recent signals determine expert selection. I find this combination worth investigating. The five-second results support it, and the parameter-matched controls help distinguish architectural benefits from gains due to model size.

  2. The authors test causality and stateful execution directly. They also report long-duration results and problems in the released EgoEMG protocol, then provide an audited evaluation. These corrective experiments make the limitations easier to assess.

Critical weaknesses
  1. Training variability. I cannot tell whether the smaller gains would persist after retraining. The main emg2pose experiments, scaling runs, and audited EgoEMG comparisons use one training seed. Standard deviations across users describe a different source of variability. The 0.11-degree pooled Regression margin in the audited EgoEMG evaluation is particularly difficult to interpret without uncertainty across training runs.

  2. Attribution of architectural gains. The parameter-matched replacements help assess the architecture, but the cumulative ablations also change model size. It remains less clear how much spatial attention, temporal decay, and multiscale patching each contribute within STEM.

  3. Implementation details. Reproducing these comparisons requires more information. Although the appendices provide encoder dimensions and optimizer settings, I could not locate complete expert architectures, router normalization, dropout probabilities, or full training schedules.

  4. Interpretation of generalization. Each model is retrained in the cross-dataset experiments, so these results support architectural portability; they do not test transfer of one trained checkpoint. The released EgoEMG protocol also has documented overlap and model-selection problems, which make the audited results more informative.

  5. Continuous streaming. The default absolute-position model underperforms vEMG2Pose on 30-second Tracking, whereas RoPE improves the result. Latency is measured for MOD-295 instead. It remains unclear what sustained accuracy and processing time a single recommended configuration achieves.

Questions for authors
  1. Which comparisons, if any, were repeated with different training seeds? Do those runs support the reported advantages, particularly the small margin in the audited EgoEMG evaluation?

  2. Does NinaPro use the same eight held-out movement classes for validation and Stage testing? Please identify the classes and explain how validation informed model selection.

  3. Which positional variant do you recommend for continuous streaming? What existing accuracy and latency measurements apply to that same configuration?

  4. How did you choose the routing-history lengths and active-expert counts? Is there existing evidence that the results are stable under other choices?

  5. Where are the complete expert architectures, router normalization, dropout probabilities, and training schedules specified? Please point to the relevant configurations or clarify the missing settings.

  6. Do existing experiments isolate spatial attention, temporal decay, or multiscale patching within STEM, beyond the cumulative additions and parameter-matched replacements?

2 comments
AuthorsOpenReview ↗

Repeated Training, Component Controls, and Continuous Streaming Evaluation

Thank you for the questions. We added independent training runs and targeted controls. All ± values below are sample SDs across seeds 42, 43, and 44.

1. Repeated runs and EgoEMG. Three-seed experiments now cover the emg2pose main comparisons, all four methods on NinaPro and audited EgoEMG, STEM controls, and routing sensitivity. On emg2pose, STAMP reduces regression error by 1.438° against EMGFormer-M and tracking error by 1.052° against single-task vEMG2Pose. Paired seed/user resampling gives descriptive 95% intervals of [−1.706, −1.217]° and [−1.126, −0.978]° for STAMP minus baseline.

The earlier 0.11° EgoEMG regression advantage did not recur in any of the three seeds. On strict-v1, STAMP/EMGFormer-M achieve 15.533±0.212°/14.991±0.271°; tracking is 9.088±0.048°/8.976±0.018° for STAMP/current velocity-integrating vEMG2Pose. The paired overall gaps are +0.542° [0.208, 0.841]° and +0.112° [0.037, 0.192]°, respectively. We will make these validation-selected results the main EgoEMG evaluation.

Low-motion windows help explain the close tracking scores: simply holding the supplied initial pose gives 9.164°, versus 8.976–9.171° across the four trained methods. Splitting each user's windows into lower/higher displacement halves by ground-truth deviation from the initial pose, both covering the same 32 users, gives STAMP/vEMG2Pose errors of 5.969°/5.543° and 12.108°/12.300°, respectively. The ordering reverses in every seed. Under the original user-equal metric, these groups contribute +0.209° and −0.097°, yielding the overall +0.112° gap. We will report this motion-dependent behavior and distinguish training-seed SD from cross-user SD. Existing cumulative ablations and scaling curves remain single-seed; we will label their coverage and treat scaling trends as exploratory.

2. NinaPro classes and selection. Validation and Stage-related tests use the same held-out action IDs: 1, 2, 4, 15, 23, 28, 30, and 38. The participant split assigns 28/4/8 users to disjoint training/validation/test groups. These actions are excluded from training supervision but used for validation. Checkpoints are selected by masked angular L1 on validation User and User+Stage samples, with early-stopping patience 30. We will state this scope explicitly.

3. Recommended streaming configuration. We recommend Absolute + MOD-295, applying t mod 295 to the trained absolute-position model without additional fine-tuning. On emg2pose, with a fixed checkpoint per task and seed, tracking errors at 5 s are 9.810±0.048° (native) and 9.838±0.046° (past-only); at 30 s, they are 12.997±1.147° and 13.034±1.144°. The original MOD-295 engine measured 1.216 ms steady-state P99 on an RTX 5060 Laptop. This is the existing latency measurement, not a new timing result for these checkpoints. We will state checkpoint identities and timing provenance together.

4. Routing choices and sensitivity. History half-lives of 0.25/0.5/1/2 s and top-8 routing were predefined. The default gives 12.762±0.228°; halved/doubled histories give 12.866±0.170°/12.682±0.123°, and top-4/top-16 give 12.787±0.204°/12.541±0.147°. All paired difference intervals include zero. We will report the grid without presenting the default as a test-set optimum.

5. Implementation and training. The emg2pose model has 32 independent experts at each of three locations. Each expert is 64→128→SiLU→Dropout(0)→64, with zero-initialized output weights and bias. The router has width 32, four heads, FFN width 64, and pre-attention, pre-FFN, and output LayerNorm. Temperature is 1; softmax weights are restricted to the top eight and renormalized. STEM/pooling dropout is 0.1; decoder/expert/linear-router dropout is 0. Training uses FP32, eight-GPU DDP, global batch 64, Adam, up to 100 epochs, patience 30, and learning rates 1e-4 for the backbone and 1e-3 for experts/router. There is no dense warmup, top-k annealing, or learning-rate scheduler. We will consolidate the dataset-specific configurations in the appendix.

6. STEM attribution. Replacing spatial MHA with an electrode-wise MLP raises error from 12.762±0.228° to 14.850±0.122°. Removing decay or using repeated short-/long-scale branches gives 12.797±0.023°, 12.721±0.309°, and 12.763±0.120°; their paired difference intervals include zero. Repeated-scale controls retain both branches and approximately match parameters by adjusting widths. We will emphasize electrode interaction and describe decay/multiscale patching as design choices without claiming separate accuracy gains.

We also agree that separately retrained models demonstrate architectural portability. We will distinguish this from transferring one checkpoint across datasets.

AuthorsOpenReview ↗

imited Sensitivity of Aggregate EgoEMG Tracking Scores to Dynamic Performance

We provide an additional analysis to clarify why the aggregate EgoEMG tracking scores are close and how they reflect dynamic tracking performance.

1. A strong initial-pose baseline

On the restored strict-v1 split, holding the supplied ground-truth initial pose throughout each five-second window yields 9.164° joint-angle MAE. The three-seed means of the four trained methods range from 8.976° to 9.171°. Thus, a competitive aggregate score can be obtained without estimating subsequent motion from EMG.

2. Motion stratification reveals different rankings

For each user, we divided test windows into lower- and higher-displacement halves according to the ground-truth pose deviation from the initial pose. The groups contain 832 and 843 windows, respectively, and both cover the same 32 users. The table reports joint-angle MAE, averaged over three seeds for the trained methods.

MethodLower displacementHigher displacement
Hold initial pose5.460°12.749°
STAMP5.969°12.108°
EMGFormer-M5.492°12.731°
NeuroPose5.494°12.698°
vEMG2Pose (velocity integration)5.543°12.300°

Holding the initial pose has the lowest error in the lower-displacement group, whereas STAMP has the lowest error in the higher-displacement group. Based on each user's three-seed mean, vEMG2Pose performs better than STAMP for all 32 users in the lower-displacement group; STAMP performs better for 25/32 users in the higher-displacement group. The group-level ranking reversal between these two methods occurs in all three seeds.

Under the paper's original equal-user aggregation, the lower- and higher-displacement groups contribute +0.209° and −0.097°, respectively, to the overall STAMP-minus-vEMG2Pose difference of +0.112°. The aggregate comparison therefore combines a penalty at lower displacement with a gain at higher displacement.

3. Implication for evaluation

These results suggest that the aggregate EgoEMG tracking metric has limited sensitivity to differences in dynamic tracking performance: preserving the initial pose is already competitive, and opposing effects across motion regimes partly cancel. We will report the initial-pose baseline and motion-stratified results alongside the aggregate score to make these differences explicit.

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

STAMP is a task-structured architecture for streaming sEMG-to-hand-pose estimation. STEM preserves electrode structure and combines spatial softmax attention with decayed causal linear attention. Affine Scan maintains pose through gated propagation and absolute correction, while Motion MoE routes computation using causal history summaries. The paper evaluates short-window accuracy, architectural replacements, data and capacity scaling, long-horizon behavior, native causality, and deployment latency. The strongest evidence concerns matched short-window emg2pose comparisons; other datasets and deployment configurations require more qualified interpretation.

Critical strengths

The task decomposition is coherent and the affine update is clearly specified. Near-parameter-matched controls against a temporal Transformer and a wider non-MoE backbone go beyond a simple larger-model comparison. On emg2pose Regression, STAMP reports 12.48 degrees and 16.46 mm versus 13.99 degrees and 18.79 mm for EMGFormer-M. Tracking improves from 10.86 degrees and 14.68 mm for vEMG2Pose to 9.80 degrees and 13.10 mm.

Critical weaknesses

The main emg2pose models and cumulative ablations use one training seed, and the scaling grid is also single-seed. Cross-user population standard deviations describe heterogeneity across people, not optimizer variability or uncertainty in a paired model difference. Several larger short-window improvements look promising, but the smaller portability gains especially need repeated training and paired-user analysis.

Questions for authors
  1. Can the principal emg2pose comparisons and audited EgoEMG results be repeated across training seeds with paired user-level uncertainty?
  2. Which exact checkpoint and position treatment is recommended for deployment, and can it be evaluated jointly at five and thirty seconds under native and past-only scoring?
2 comments
AuthorsOpenReview ↗

Three-Seed Comparisons and Streaming Evaluation with a Single Configuration

Thank you for your comments. We have completed independent retraining and a joint evaluation of the recommended streaming configuration.

1. Training variability and paired comparisons

We repeated the main comparisons with seeds 42, 43, and 44. Values below are mean ± sample SD across training seeds.

emg2pose taskSTAMP angle / landmark errorComparator angle / landmark error
Regression12.762 ± 0.228° / 16.810 ± 0.330 mmEMGFormer-M: 14.199 ± 0.169° / 19.007 ± 0.189 mm
Tracking9.774 ± 0.050° / 12.986 ± 0.144 mmSingle-task vEMG2Pose: 10.826 ± 0.013° / 14.677 ± 0.064 mm

With 10,000 paired seed-and-user resamples, the angular-error differences (STAMP minus comparator) are −1.438° for Regression, with a descriptive 95% interval of [−1.706, −1.217]°, and −1.052° for Tracking, with [−1.126, −0.978]°. Averaging each user's difference across the three seeds, STAMP has lower error for 175/178 users in Regression and 178/178 in Tracking; paired sign tests give p < 0.001 for both after Holm correction across 12 comparisons.

For EgoEMG, all four methods were retrained on the restored strict-v1 split, using independent validation for checkpoint selection. Across three seeds, STAMP has higher overall angular error than EMGFormer-M in Regression by 0.542° [0.208, 0.841]°, and than the current velocity-integration vEMG2Pose in Tracking by 0.112° [0.037, 0.192]°. These paired intervals preserve the 26 seen-user and 6 unseen-user composition. We will make these audited results the main EgoEMG comparison and update the claims accordingly.

The close Tracking scores also reflect limited motion in many windows: simply holding the supplied initial pose gives 9.164°, compared with 8.976–9.171° for the four trained methods. Splitting each user's windows into equal lower- and higher-displacement groups, based on ground-truth deviation from the initial pose, gives:

EgoEMG Tracking groupHold initial poseSTAMPvEMG2Pose
Lower displacement, 832 windows5.460°5.969°5.543°
Higher displacement, 843 windows12.749°12.108°12.300°

Both groups contain the same 32 users. vEMG2Pose performs better for all 32 users in the lower-displacement group; STAMP performs better for 25/32 in the higher-displacement group, using each user's three-seed mean. The ranking reversal appears in all three seeds. Under the paper's equal-user aggregation, the two groups contribute +0.209° and −0.097°, respectively, to the overall +0.112° difference. The lower-displacement penalty and higher-displacement gain partly cancel in the overall score.

The new STEM controls and routing sensitivity study also use three seeds. The original cumulative ablations and scaling curves remain single-seed; we will label their coverage explicitly and treat scaling trends as exploratory.

2. One streaming configuration at 5 and 30 seconds

We recommend Absolute + MOD-295: the trained absolute-position model uses position index t mod 295, with no additional positional fine-tuning. On emg2pose, we fixed one checkpoint per task and seed across both horizons and both scoring modes. Angular errors are:

Task / horizonNativePast-only
Regression / 5 s12.787 ± 0.232°12.879 ± 0.228°
Regression / 30 s12.226 ± 0.212°12.274 ± 0.207°
Tracking / 5 s9.810 ± 0.048°9.838 ± 0.046°
Tracking / 30 s12.997 ± 1.147°13.034 ± 1.144°

The original MOD-295 deployment engine measured 1.216 ms steady-state P99 on an RTX 5060 Laptop. The recommendation retains that execution path; this timing is from the original engine evaluation, while the table reports the new fixed-checkpoint accuracy evaluation. We will list the checkpoint identifiers, scoring settings, and timing provenance together in the revision.

AuthorsOpenReview ↗

Limited Sensitivity of Aggregate EgoEMG Tracking Scores to Dynamic Performance

We provide an additional analysis to clarify why the aggregate EgoEMG tracking scores are close and how they reflect dynamic tracking performance.

1. A strong initial-pose baseline

On the restored strict-v1 split, holding the supplied ground-truth initial pose throughout each five-second window yields 9.164° joint-angle MAE. The three-seed means of the four trained methods range from 8.976° to 9.171°. Thus, a competitive aggregate score can be obtained without estimating subsequent motion from EMG.

2. Motion stratification reveals different rankings

For each user, we divided test windows into lower- and higher-displacement halves according to the ground-truth pose deviation from the initial pose. The groups contain 832 and 843 windows, respectively, and both cover the same 32 users. The table reports joint-angle MAE, averaged over three seeds for the trained methods.

MethodLower displacementHigher displacement
Hold initial pose5.460°12.749°
STAMP5.969°12.108°
EMGFormer-M5.492°12.731°
NeuroPose5.494°12.698°
vEMG2Pose (velocity integration)5.543°12.300°

Holding the initial pose has the lowest error in the lower-displacement group, whereas STAMP has the lowest error in the higher-displacement group. Based on each user's three-seed mean, vEMG2Pose performs better than STAMP for all 32 users in the lower-displacement group; STAMP performs better for 25/32 users in the higher-displacement group. The group-level ranking reversal between these two methods occurs in all three seeds.

Under the paper's original equal-user aggregation, the lower- and higher-displacement groups contribute +0.209° and −0.097°, respectively, to the overall STAMP-minus-vEMG2Pose difference of +0.112°. The aggregate comparison therefore combines a penalty at lower displacement with a gain at higher displacement.

3. Implication for evaluation

These results suggest that the aggregate EgoEMG tracking metric has limited sensitivity to differences in dynamic tracking performance: preserving the initial pose is already competitive, and opposing effects across motion regimes partly cancel. We will report the initial-pose baseline and motion-stratified results alongside the aggregate score to make these differences explicit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.