People who have lost one hand can control a prosthesis through electromyography (EMG): electrodes on the skin of the remaining forearm, the "stump," pick up the electrical activity of muscles the person tries to move. Turning that signal into continuous joint angles requires a regression model, and a regression model needs examples of the target. Since there is no hand on the amputated side to measure, the field uses mirrored bilateral training. The person moves the intact hand while imagining the same movement with the missing one, and a data glove on the intact hand supplies the labels. This works well enough for people with long stumps. For people with short stumps it largely fails: in EMG-based simultaneous and proportional estimation of wrist/hand kinematics in uni-lateral trans-radial amputees, the variance explained ($R^2$) was 72.0% in intact controls, 62.5% for long stumps, and only 19.7% for stumps shorter than 10 cm.
We propose to also record EMG from the intact forearm during mirrored training. This is a small change to the recording setup. The intact arm then provides a clean, same-person EMG-to-kinematics mapping, measured on a real hand with real ground truth. It also provides a time-locked "partner" signal for every stump sample. The usual cross-domain problem in this area is adapting a model trained on able-bodied data to an amputee's stump. With both arms recorded, that problem becomes a paired translation problem: learn to map stump EMG into the representation of the patient's own intact-arm EMG, then decode with a frozen intact-arm decoder. Our claim is that the intact-arm EMG target is denser, causally cleaner, and more data-efficient than the glove trajectory alone, and that this matters most where current methods break down, in short stumps and with small calibration budgets.
Core Claim
Mirrored bilateral training is normally treated as a label-acquisition trick. The kinematics come from one side and the EMG from the other. That pairing is noisy in known ways: inter-limb latency, asymmetric effort, and phantom-limb sensations mean the intact hand only approximates the intended motion, as discussed in ReMAP. We argue the pairing is more useful at the level of EMG itself. Both forearms are driven by the same bilateral motor intent, so intact-side EMG gives a 16-dimensional supervisory signal at every time step, against 3 to 5 joint angles from the glove. A decoder trained on the intact side, either on the patient alone or on able-bodied populations, never sees the stump domain shift.
The objective trains a small translator $T$ that maps stump EMG features $x_s$ toward intact-side EMG $x_i$. Two modules are held frozen: $E$, an encoder of intact-side EMG, and $g$, the regression head that maps that encoding to joint angles $y$:
$$\mathcal{L}(T,\tau) = \big\| g\big(E(T(x_s^{t}))\big) - y^{t} \big\|^2 + \lambda \,\big\| E(T(x_s^{t})) - E(x_i^{t+\tau}) \big\|^2 .$$
Here $\lambda$ weights the representation-matching term, and $\tau$ is a per-session inter-limb lag. We estimate $\tau$ jointly instead of assuming the two arms are perfectly synchronous. Only $T$ and $\tau$ are fit on the patient's stump data.
Relation to Prior Work
Three lines of work motivate or supply parts of this design.
Methods that align EMG with kinematics. KinEMbed pre-trains an EMG encoder contrastively against joint angles on NinaPro DB8 and then regresses with a temporal convolutional network. It reaches $R^2 = 0.769$ on able-bodied subjects but only 0.568 on two amputees. It also reports a false-negative problem: time points with nearly identical poses get pushed apart. CPEP similarly anchors EMG to a frozen pose encoder, but it evaluates only gesture classification. Both align EMG to kinematics. We align stump EMG to intact EMG with a regression-style matching loss, which has no negatives at all and so sidesteps the false-negative problem entirely.
Model reprogramming. ReMAP supplies the core method. It freezes an able-bodied gait model and trains a three-layer adapter on the inputs, reaching $R^2$ 0.86–0.87 with 10% of patient data, against about 0.74 for fine-tuning. However, ReMAP finds its targets through nearest-neighbour retrieval and network inversion over other people's movements, because it has no paired target. Mirrored recording gives us the paired target directly, from the same person, at the same moment.
Explicit factorization of subject identity. SAFE-Net separates movement-intent features from subject-identity features, but only for healthy treadmill walking. We test whether a simple paired adapter does better than explicit factorization under the much larger stump-versus-intact shift.
Constraints on the design. Two results limit what we should expect. The cross-domain few-shot study Experiments in Cross-domain Few-shot Learning found that the deepest pre-trained layers transfer worst to distant domains. We therefore let the translator target an intermediate layer of $E$ and treat the choice of layer as an experimental variable. The unsupervised autoencoder of Online mapping of EMG signals into kinematics by autoencoding shows that low-dimensional muscle-synergy structure can be found without labels. This means some of the benefit of intact EMG might be reproducible with no second arm at all, so we test for that.
Research Program
Stage one: simulation on existing data. We use the nine able-bodied NinaPro DB8 subjects. For each, we hold out a subset of channels to stand in for a truncated stump, using channels with deliberately reduced signal amplitude and added crosstalk. We then train $T$ to map the degraded subset onto the full-array representation. This is a crude stand-in for a real stump, and it is useful only for ruling out failure early. If paired translation does not beat direct regression on the degraded subset, the idea is unlikely to survive real anatomy.
Stage two: the decisive dataset. No dataset in our file records EMG on both forearms during mirrored movements. We will therefore collect one. The protocol follows the mirrored 3-DoF wrist design of EMG-based simultaneous and proportional estimation… and adds the 5-degree-of-actuation hand tasks of DB8. Each participant wears 16 EMG channels on each forearm and a CyberGlove on the intact hand, and is recorded over three sessions on separate days. We target roughly a dozen transradial amputees, stratified by stump length so that both long and short stumps are well represented, plus matched able-bodied controls.
Choice of decoder. We compare two ways of building $E$ and $g$. One is trained on the patient's own intact arm, which gives true ground truth after mirroring the electrode layout. The other is trained on the pooled able-bodied DB8 population. Comparing them tells us whether the bridge works because it is personal or because it is a well-trained prior.
Baselines
Naive within-subject regression. This is the reference point a clinic would actually use: per-DoF MLPs on time-domain and autoregressive features, trained directly from stump EMG to glove kinematics, as in EMG-based simultaneous and proportional estimation…. That paper reports $R^2$ of 62.5% for long stumps and 19.7% for short stumps. We re-implement it on our data with the same features and the same calibration minutes.
KinEMbed. This is the strongest existing continuous decoder. On NinaPro DB8 it reports an overall $R^2$ of 0.732 and 0.568 on amputees, with CEBRA at 0.570. We run it with its contrastive pre-training on the same patient's stump EMG and glove data, its temporal convolutional regressor, and its train-on-two-sessions, test-on-the-third protocol. Its hyperparameter search budget matches ours. Beating it means a higher session-3 $R^2$ that is statistically reliable across amputees in a paired test, and especially within the short-stump group.
Fine-tuning and ReMAP-style input adaptation. Here we take the same frozen able-bodied decoder and either fine-tune all of it on stump-to-glove data or train an input adapter using only the kinematic term ($\lambda = 0$). The $\lambda = 0$ adapter is the key ablation. It is identical to our method except that it never sees intact-side EMG, so any gain from our full method over it isolates the contribution of the bilateral EMG pairing.
Incremental ridge regression. For cross-day robustness we include the incremental ridge regression with random Fourier features from Stable myoelectric control…. It restored Day-2 error from about 1.0 to 0.3–0.6 normalized mean squared error using 12-second updates. We give it the same short recalibration bursts that our translator gets.
All methods receive identical recordings, calibration budgets of 1, 3, and 10 minutes, and equal tuning budgets.
Why not just…?
Why not just regress stump EMG onto glove kinematics with a better model? This is the real threat. The information available from the stump is the same in both cases, and if the stump simply lacks the information, no target will recover it. The proposal is justified only if the denser intact-EMG target makes better use of scarce data. This is the kind of benefit ReMAP showed when it beat direct mapping (about 0.71) with frozen priors. Our calibration-budget curves settle the question. If the advantage over KinEMbed and over the $\lambda = 0$ adapter disappears at every budget, the idea is wrong, and we will report it that way.
Why not just apply contrastive alignment, KinEMbed-style, to mirrored data? That is the obvious combination of these two directions. It inherits the false-negative problem that KinEMbed's own soft variant did not fix. It also aligns to a 5-dimensional pose space that throws away the muscle-level structure the intact arm provides. KinEMbed is included as a full baseline, so the comparison is made directly.
Why not just fix the inter-limb lag by cross-correlation and keep direct regression? Lag correction is cheap and may recover part of the gain. We include a cross-correlation-shifted direct-regression baseline and report how much of our improvement survives it.
Why not just use an unsupervised autoencoder and let the user adapt? The autoencoder approach of Online mapping… was validated only on seven able-bodied users. KinEMbed reports that unanchored autoencoders decode kinematics worse. We include an autoencoder-initialized translator so we can separate the value of unsupervised structure from the value of paired intact EMG.
Decisive Checks
The primary outcome is session-3 $R^2$ per degree of freedom, stratified by stump length. The proposal succeeds if the full method beats both KinEMbed and the $\lambda = 0$ adapter for short-stump amputees at small calibration budgets. It fails if the $\lambda = 0$ adapter matches it, because that would mean the intact EMG adds nothing.
Two secondary checks follow. First, the learned lag $\tau$ should be stable within a person across sessions; if it is unstable, it is absorbing noise rather than physiology. Second, we test whether gains concentrate on pronation and supination, the rotational degrees of freedom where masking by deep muscles is worst.
Finally, a closed-loop Fitts' law target-acquisition task with a subset of amputees checks whether offline $R^2$ gains carry over to real-time control. That task reports throughput and path efficiency, the metrics used by Online mapping….
References
- EMG-based simultaneous and proportional estimation of wrist/hand kinematics in uni-lateral trans-radial amputees
- ReMAP: Neural Model Reprogramming with Network Inversion and Retrieval-Augmented Mapping for Adaptive Motion Forecasting
- KinEMbed: Decoding Kinematics from Electromyography via Cross-Modal Contrastive Learning
- CPEP: Contrastive Pose-EMG Pre-training Enhances Gesture Generalization on EMG signals
- Efficient sEMG-based Cross-Subject Joint Angle Estimation via Hierarchical Spiking Attentional Feature Decomposition Network
- Experiments in Cross-domain Few-shot Learning for Image Classification: Extended Abstract
- Online mapping of EMG signals into kinematics by autoencoding
- Stable myoelectric control of a hand prosthesis using non-linear incremental learning