Wearable exoskeletons and prostheses would be far more useful if they could read a new user's muscle signals out of the box. Surface electromyography (sEMG) is the electrical activity of muscles recorded through skin electrodes. It changes so much from person to person, with electrode position, fat thickness and limb shape, that decoders trained on one small cohort fail on the next person. The usual fixes are to collect motion-capture labels from every new user or to train ever larger networks on cohorts of 8 to 12 people. Neither works well in practice. Motion capture is not available in a clinic or at home, and large networks overfit when only about ten subjects exist.
Motor neuroscience points to a different, more structured explanation of what varies between people. The nervous system drives muscles through a small number of coordinated groups, called muscle synergies or functional clusters, that share a common neural input. The timing of these drives during walking is thought to be largely common across people. If that is right, most of the between-subject variation sits in a per-person mixing matrix that maps a few shared drives onto the electrodes. That matrix can be estimated from a minute of unlabeled walking. We propose to build a lower-limb decoder around this factorization and test it directly. The decoder's shared core learns once, from a small cohort, how synergy activations predict upcoming hip, knee and ankle angles. A new user is handled by fitting only their nonnegative mixing weights, with no kinematic labels. The test of whether this works also answers a live question in motor control: across people, speeds and inclines, is it the synergy timing that stays fixed, or the synergy composition?
Core Claim
Write the multichannel sEMG envelope of subject $s$ at time $t$ as
$$x_s(t) \approx W_s\, h(t), \qquad W_s \ge 0,\; h(t) \ge 0,$$
where $x_s(t)$ is the vector of $m$ electrode envelopes, $W_s$ is an $m \times k$ subject-specific mixing matrix (synergy composition, volume conduction, electrode placement), and $h(t)$ is a $k$-dimensional vector of synergy activations whose dependence on gait phase and upcoming movement is shared across subjects. A shared decoder $g$ maps a causal history of $h$ to future joint angles $\hat{q}(t+\tau)$ at prediction horizon $\tau$.
The claim is that, under leave-one-subject-out testing on small cohorts, a model that re-estimates only $W_s$ from unlabeled data will match or beat abstract latent-disentangling networks. These networks have far more subject-dependent capacity than the $mk$ numbers in $W_s$. The advantage should be largest at ahead-of-time horizons and during locomotor transitions, which are exactly the regimes where inertial sensors alone cannot help.
The claim is falsifiable in two specific ways. First, if per-channel gain normalization (a diagonal $W_s$) does as well as a full synergy mixing, then synergy structure is not what carries transfer, and we will say so. Second, if $W_s$ must change across inclines or speeds within the same person for decoding to hold, then synergy composition is task-dependent and the "fixed composition, shared timing" picture is wrong for locomotion.
Relation to Prior Work
The closest work is SAFE-Net (Efficient sEMG-based Cross-Subject Joint Angle Estimation via Hierarchical Spiking Attentional Feature Decomposition Network; journal version sEMG-Based Joint Angle Estimation via Hierarchical Spiking Attentional Feature Decomposition Network). It splits a spiking-network latent into a "kinematic" stream and a "biological" stream, using a subject-identification loss and an orthogonality penalty, and evaluates cross-subject hip, knee and ankle estimation on SIAT-DB1 ($N=8$) and SIAT-DB2 ($N=12$). We differ in three ways. Our subject-specific part is an explicit, physiologically interpretable nonnegative mixing rather than an abstract latent. We fit it to the new subject at test time without labels rather than hoping the invariant stream transfers. And we evaluate on transitions and ahead-of-time prediction, which SAFE-Net's authors list as untested.
Two papers motivate the synergy structure. Common synaptic input, synergies and size principle by Hug et al. argues that motor neurons are organized into functional clusters defined by shared synaptic input, sometimes crossing muscle boundaries and sometimes splitting a single muscle. That paper is the physiological basis for $x = Wh$. The manifold structure of limb coordination in walking Drosophila shows that walking coordination lies on a continuous low-dimensional manifold organized by stance duration and speed. That result motivates conditioning the shared activation dynamics on phase and speed rather than on discrete gait modes.
M2M-Net (Mapping Motor Cortex Stimulation to Muscle Responses) supplies a design lesson. In that paper, decoding through synergies plus a residual beat direct per-muscle decoding (NRMSE 0.500 vs. 0.661 with raw inputs). However, it uses one subject and static evoked responses, so it says nothing about transfer. We borrow its synergy-plus-residual output structure.
Several papers constrain the design. Continuous Prediction of Wrist Joint Kinematics Using Surface Electromyography From the Perspective of Muscle Anatomy and Muscle Synergy Feature Extraction reports that plain NMF is fragile to noise and local optima, and that its anatomically arranged network beat both NMF and musculoskeletal models ($R^2$ 0.95 vs. 0.88) within subjects. We therefore anchor NMF to a cohort template rather than running it from scratch, and we keep a nonlinear decoder on top of the synergies. KinEMbed shows that contrastive alignment of sEMG to kinematics outperforms PCA and PLS on NinaPro DB8, but it requires labels and suffers false negatives on repeating movements. Our adaptation step needs no labels at all.
Continuous Prediction of Lower-Limb Kinematics From Multi-Modal Biomedical Signals (KinPreNet) establishes why sEMG is worth the trouble. Muscle activity leads joint motion by an electromechanical delay of roughly 27–108 ms, which allows forecasting. KinPreNet itself has no cross-subject scheme. Finally, PhASER treats signal phase as nuisance drift for classification. That is a warning for our setting, because in forecasting the timing between muscle activation and movement is the signal itself. Our model therefore normalizes amplitude through $W_s$ and never discards timing.
Research Program
The model has four parts. The first is a fixed preprocessing step that turns raw sEMG into rectified, low-pass envelopes, using causal filters only. The second is a cohort template $W_0$ with $k$ synergies, learned by NMF on pooled training subjects. The third is a causal temporal convolutional decoder $g$ that maps the last few hundred milliseconds of $h$, plus optionally IMU signals, to joint angles at horizons $\tau \in \{0, 50, 100\}$ ms. The fourth, following M2M-Net, is a small residual pathway that takes raw envelopes, so the model can use fine detail the synergies miss. The residual pathway is penalized so that it carries little of the prediction.
Training alternates between two steps. In one, the shared decoder is fit on activations obtained by projecting each training subject's data onto their own $W_s$. In the other, each $W_s$ is refit with a term pulling it toward $W_0$.
Adapting to a new subject uses only 30–120 seconds of unlabeled walking. We run NMF initialized at $W_0$. We then remove NMF's permutation and scale ambiguity by matching each recovered synergy to a template synergy through its gait-phase-averaged activation profile, using Hungarian matching. Gait phase here comes from a heel-strike detector on the shank IMU, so no kinematic labels are involved. After this step, decoding is a nonnegative projection followed by the frozen decoder, which is cheap enough for embedded hardware.
We also build a learned variant in which $W_s$ is fit by minimizing reconstruction error of $x_s$ through the frozen activation model. This checks that the gains do not depend on the particular NMF recipe.
The analysis part asks what moves between people and between conditions. Within each subject, we fit $W$ separately for each incline or speed in SIAT-DB1/DB2 and compare the similarity of synergy compositions with the similarity of activation profiles. This bears directly on the open question of hardwired versus task-assembled synergies.
On the Dimitrov dataset, the 64-channel grid covers a single muscle, the fibularis longus. There, $k$ components correspond to candidate sub-muscle functional clusters in Hug's sense. We ask whether their relative timing reorganizes between level walking, stairs and ramps, which is an open problem the research file lists for locomotion.
Data
We use three sources. The first is SIAT-DB1 (5 muscles, $N=8$, three inclines) and SIAT-DB2 (8 muscles, $N=12$, three speeds), which allow direct comparison with SAFE-Net on the same splits. We will request these from the authors; the file does not state that they are public. If they are unavailable, the KinPreNet level-walking data (9 muscles, $N=10$, knee angle) serves as the multi-muscle fallback.
The second is the open dataset High-density EMG, IMU, kinetic, and kinematic open-source data for comprehensive locomotion activities ($N=10$). It provides 64-channel HD-sEMG with OpenSim joint angles across sit-to-stand-to-walk, stairs, 5° and 15° ramps and sidestepping. This is the only source in the file with transitions.
The third addresses a gap none of these sources fill: multi-muscle EMG during transitions. For this we budget a modest new collection of about 12 healthy adults with 8 bipolar lower-limb muscles, shank and thigh IMUs and motion capture, using the Dimitrov task protocol.
Baselines
The naive reference is a causal temporal convolutional network trained on pooled envelopes, with per-subject z-scoring of each channel computed from the same unlabeled calibration minute our method uses. This is the "do nothing clever" decoder. It also tests the simplest form of subject adaptation, which is a per-channel gain.
The strongest existing cross-subject method for lower-limb sEMG is SAFE-Net. On SIAT-DB1 it reports a 27.3% reduction in average RMSE versus Informer and 59.8% versus Spikformer at slope 0, and 27–28% average reduction for hip-knee-ankle. The file gives no absolute RMSE values, and for DB2 it reports no numerical result found on Lacuna (results appear only in box plots). We will therefore reimplement SAFE-Net with its stated composite loss. We will also rerun Informer, so that the reported relative gap can be reproduced as a sanity check, on identical leave-one-subject-out splits, causal windows and envelope inputs. Every method receives the same hyperparameter search budget, set by a fixed number of trials using inner-fold validation on training subjects only. Beating SAFE-Net means lower mean leave-one-subject-out RMSE over hip, knee and ankle, with a paired test across held-out subjects. The paired test matters because with $N=8$–12 a small average gain can easily be noise.
For forecasting, the reference is a KinPreNet-style two-stage LSTM fusing sEMG and IMU. Its paper reports 3.98° knee RMSE and Pearson R of 0.88 at 100 ms on level walking, within its own cohort. We run it under our cross-subject protocol, where no number has been reported. We also include an IMU-only forecaster with the same architecture to measure what sEMG adds.
As a ceiling, we fine-tune the pooled network on labeled data from the held-out subject, at label budgets of 30 seconds to several minutes. This shows how much of the labeled-calibration gap our label-free method closes.
Three ablations isolate the key ingredient. In the first, adaptation is restricted to a diagonal $W_s$, which tests whether gain alone suffices. In the second, $W_s$ is replaced by an unconstrained-sign linear map of the same rank, which tests whether nonnegative synergy structure matters or any low-rank map would do. In the third, $W_s$ is frozen at $W_0$ with no adaptation. We also sweep $k$ from 2 to 8.
Why not just…?
Why not just normalize each channel's amplitude and train a good network? This is the most serious threat. Much between-subject variation is plausibly gain from fat thickness and electrode impedance, and a diagonal $W_s$ captures exactly that. Our full model contains it as a special case, so the diagonal-gain ablation settles the question directly. If diagonal adaptation matches full mixing on SIAT-DB2 and on our transition data, the synergy story adds nothing for decoding, and we would report that. The scientific analysis of synergy composition would still stand on its own. We expect full mixing to help mainly where electrode placement shifts which muscles each electrode picks up. That situation is most visible on the 64-channel grid, where a positional shift changes the mixing rather than the gain.
Why not just use IMUs, which do not drift with sweat or impedance? For estimating the current joint angle during steady walking, IMUs may be enough. Indeed, SIAT-DB1's labels themselves come from IMUs, which is one reason we do not consider instantaneous angle estimation on that dataset decisive. The case for sEMG rests on forecasting ahead of movement. KinPreNet's 27–108 ms lead window is reported within its own cohort, and its 15.95° figure for traditional SVM forecasting is not an IMU-only baseline, so the file gives no direct cross-subject comparison. The IMU-only forecaster, evaluated separately on transition onsets, will show whether sEMG earns its keep. If the gap closes at every horizon, the practical argument for our method fails, and we would say so.
Why not just fine-tune on a short labeled calibration from the new user? When motion capture is available, this works, and it is our ceiling. But labeled joint angles require optical capture or carefully calibrated IMU chains that clinical and home users lack. The fair comparison is the label budget at which fine-tuning overtakes label-free synergy adaptation. If that budget is only a few seconds, our contribution shrinks to a convenience.
Why not a generic domain-generalization method such as PhASER or adversarial adaptation? PhASER is built for classification, and its authors leave regression untested. It also regularizes away phase, which in forecasting carries the activation-to-movement lead. Adversarial adaptation for sEMG has been shown for inter-session gesture classification (Surface EMG-Based Inter-Session Gesture Recognition Enhanced by Deep Domain Adaptation), but not for continuous lower-limb regression. As a secondary baseline, we will run PhASER's magnitude path with Hilbert augmentation using the public code (payalmohapatra/PhASER), with a regression head. The specific prediction is that it degrades more than our model at the 100 ms horizon.
Decisive Experiments
The main experiment is leave-one-subject-out on SIAT-DB1/DB2, or on the KinPreNet data if needed. We report per-joint RMSE and $R^2$ at horizons of 0, 50 and 100 ms, against every baseline above, as a function of how many seconds of unlabeled calibration are available. The result that would support the claim is that full nonnegative mixing adaptation beats both SAFE-Net and diagonal gain at 100 ms with a paired test across held-out subjects, using 60 seconds or less of unlabeled data.
The transition experiment uses the Dimitrov dataset and our new collection. We train on level walking and ramps and test on stairs and sit-to-stand-to-walk for held-out subjects, scoring RMSE separately within ±500 ms of transition onsets. This tests whether shared activation dynamics, conditioned on a continuous phase-speed manifold, extrapolate to non-cyclic movement better than the baselines. This regime is where the field currently has no reported cross-subject result on Lacuna.
The scientific check fits synergy composition and activation separately per condition within subjects. If composition is stable across speeds and inclines while activation timing tracks the gait manifold, the factorization is supported for both decoding and physiology. If composition must be refit per condition to keep decoding error low, synergy composition is task-assembled in locomotion. In that case the method would need condition-dependent mixing, which we would report as the main finding.
References
- Efficient sEMG-based Cross-Subject Joint Angle Estimation via Hierarchical Spiking Attentional Feature Decomposition Network
- sEMG-Based Joint Angle Estimation via Hierarchical Spiking Attentional Feature Decomposition Network
- Common Synaptic Input, Synergies, and Size Principle: Control of Spinal Motor Neurons for Movement Generation
- The manifold structure of limb coordination in walking Drosophila
- Mapping motor cortex stimulation to muscle responses: a deep neural network modeling approach
- Continuous Prediction of Wrist Joint Kinematics Using Surface Electromyography From the Perspective of Muscle Anatomy and Muscle Synergy Feature Extraction
- KinEMbed: Decoding Kinematics from Electromyography via Cross-Modal Contrastive Learning
- Continuous Prediction of Lower-Limb Kinematics From Multi-Modal Biomedical Signals
- Phase-driven Generalizable Representation Learning for Nonstationary Time Series Classification
- High-density EMG, IMU, kinetic, and kinematic open-source data for comprehensive locomotion activities
- Surface EMG-Based Inter-Session Gesture Recognition Enhanced by Deep Domain Adaptation