The Teleoperation Medium: EMG Artifacts and Their Propagation

4.1 The physical chain

An EMG teleoperation system is a cascade, and every stage adds delay, distortion, or both:

$\text{intent} \to \text{motor unit recruitment} \to \text{sEMG} \to \text{windowing} \to \text{decoder} \to \text{mapping} \to \text{robot controller} \to a_t$

Compare this to a leader-follower arm (GELLO, ALOHA), where the chain is: intent → arm motion → joint encoder → joint target. One measurement, one transform, no learned component, no statistical nonstationarity. The comparison is not close, and every quantity in the rest of §4 exists only in the EMG chain.

4.2 Latency

The physiological floor is electromechanical delay. Begovic, Zhou, Li, Wang & Zheng (2014), Frontiers in Physiology 5:494 [DOI] [A] measured total EMG-to-force EMD in quadriceps at 49.73 ± 6.99 ms, decomposed as EMG→mechanomyogram 20.5 ± 4.73 ms (41.1%) and MMG→force 28.63 ± 6.31 ms (57.1%). Literature ranges collated therein span 37.8–56.5 ms for voluntary contraction. Electrically stimulated EMD is an order of magnitude lower (2.2–8.5 ms), which confirms that most of the delay is excitation-contraction coupling rather than instrumentation.

The usability ceiling has an exact citation, which is worth stating because it is usually cited second-hand. Farrell & Weir (2007), IEEE TNSRE 15(1):111–118 [DOI] [A] ran 20 subjects on a Box-and-Block Test with the PHABS testbed across seven delay levels from ~0 to 300 ms and two prehensor speeds, finding an optimal controller delay of 100 ms for fast prehensors and 125 ms for slower ones, with linear degradation beyond.

So the budget is: ~50 ms is spent before any engineering choice is made, and the total must stay under ~100–125 ms. Windowing typically consumes 80–250 ms (Yang et al. use 80 ms windows with 40 ms stride; Yang, Shibata, Weber & Erickson use 250 ms). Modern decoders are fast enough not to be the problem — Lin, Zhang & Zhao (2025), Scientific Reports [DOI] [A] report a parallel efficient transformer running inference in ~25 ms on a Raspberry Pi 4B. But the aggregate leaves very little headroom.

Why this matters for data, not just for usability. A demonstration recorded through a channel with 100 ms of variable delay contains a systematic temporal misalignment between what the operator saw and what the robot did. In an action-chunked policy predicting $a_{t:t+k}$ from $o_t$, this misalignment is baked into the supervision target. The policy learns to act on stale observations because the data says that is what the expert did.

4.3 Electrode shift

Tanaka, Nambu & Wada (2025), Sensors 25(13):4119 [DOI] [A] both measure and consolidate:

Source	Shift	Degradation
Tanaka 2025 (own)	2 cm perpendicular	−7.6%; −1.0% with sliding-window normalisation
Young, Hargrove & Kuiken	2 cm	−15% (7-class)
Gao et al.	0 to ±2 cm	−20% (6-class)
Côté-Allard et al.	—	93.58% → 75.50% (−18.08%, 11-class)
Ameri et al.	—	6% → 7% error (−1%, CNN, 8-class)

(The Young, Gao and Côté-Allard figures are cited second-hand through Tanaka 2025; the primary PubMed records were CAPTCHA-blocked during this review. The Ameri outlier suggests CNN-based decoders are substantially more shift-robust, which is a live research direction.)

4.4 Limb position

Fougner, Scheme, Chan, Englehart & Stavdahl (2011), IEEE TNSRE 19(6) [A] report that average classification error rises from 3.8% to 18% — a 4.7× increase — when limb position changes. Mitigations exist and work: training across multiple positions gives 5.7%, and adding accelerometers gives 5.0%.

This is arguably the single most damaging artifact for manipulation data collection, because manipulation is limb-position variation. A dataset collected in one arm configuration and evaluated in another is exactly the failure mode.

4.5 Cross-session nonstationarity

Rehman et al. (2018), Sensors 18(8):2497 [DOI] [A] ran 7 subjects over 15 consecutive days × 2 sessions. Within-session accuracy was excellent (CNN 97.60 ± 1.99%, stacked sparse autoencoder 98.12 ± 1.07%, LDA ~86–87%); between-day error rose to ~11–15%, a drop of roughly 10–12 percentage points for the deep models. Deep models degraded less than LDA, which is the consistent finding across this literature.

Odeyemi & Zhang (2026) [arXiv:2607.27568] [C] report cross-session macro-F1 of 0.688 for a montage-agnostic encoder versus 0.540 for per-user LDA, and find that feature-statistic alignment "recovers about what a single labelled calibration repetition would," while batch-norm re-estimation was actively harmful.

4.6 Donning, doffing and skin impedance

Hwang, Hahne & Müller (2017), PLOS ONE [DOI] [A] measured mean electrode shift on re-donning at 1.16 ± 0.34 cm — i.e. every session re-introduces roughly half the shift magnitude that costs 7.6–20% accuracy in §4.3. They report significant online degradation in completion rate and time after donning/doffing (p<0.05), with path efficiency in an example subject falling 62.16% → 46.29%.

Their most interesting finding is a dissociation: arm position hurt offline $R^2$ substantially but online completion rates stayed above 90%, because closed-loop visual feedback lets the operator compensate. This matters enormously for our question and cuts against EMG as a data source specifically. The operator's compensation is invisible in the recorded action log — the log shows a trajectory that reached the goal, not the corrective effort that got it there. Closed-loop human compensation converts decoder error into action inconsistency rather than task failure, which is precisely the transformation that damages behaviour cloning while leaving teleoperation success rates looking acceptable.

Sousa, Noites, Vilarinho & Santos (2023), Sensors 23(20):8582 [DOI] [A] measured electrode–skin impedance every 5 minutes over 50 minutes: mean impedance falls from 0.56 kΩ at minute 5 to 0.49 kΩ at minute 10 (−12–13%), stabilising after minute 15 (only 2.9% variation from 15 to 50 minutes). Operationally: any EMG collection protocol needs a ~15-minute warm-up before calibration, and data collected in the first fifteen minutes is drawn from a different distribution than the rest.

4.7 Decoding: what is achievable

Two regimes must be distinguished.

Classification (pattern recognition) → discrete modes. This is what nearly all deployed EMG robot control uses. Yang, Shibata, Weber & Erickson (2025), npj Robotics [DOI] [A] achieve offline 99.6% (SD 0.8) and real-time 95.6% mean accuracy on 10 gestures with a 64-electrode HD-EMG array on a Stretch RE2. Wang et al. (2025), Biomimetics 10(7):464 [A] report >90% online recognition of 12 gestures under fatigue and electrode shift with an 8-channel gForcePro band at 200 Hz.

Note what these systems produce as an action: a discrete mode selection, mapped through a finite state machine to axis-wise velocity (Wang et al.: base translational 0.02 m/s, rotational 0.08 rad/s, at 50 Hz). This is not proportional control, and the resulting trajectories are piecewise-constant velocity segments — a highly non-human, non-smooth action distribution.

Regression → continuous proportional control. Harder and less accurate. Ameri, Akhaee, Scheme & Englehart (2018), PLOS ONE [A] report a negative result worth citing: a CNN on raw EMG achieved 91.61 ± 0.39% versus 90.63 ± 0.31% for a well-tuned SVM with hand-engineered features (p = 0.059, not significant), and in a Fitts'-law closed-loop test there was no significant difference on any control metric. Deep learning did not obviously beat classical methods in the loop.

The state of the art on continuous hand-pose regression comes from Meta's emg2pose benchmark [Salter et al., NeurIPS 2024 D&B, arXiv:2412.02725] [A] — 193 participants, 370 hours, 80M labelled frames at 60 Hz, 16-channel sEMG at 2 kHz with 26-camera mocap ground truth:

Condition	vemg2pose	NeuroPose	SensingDynamics
Regression, held-out users	12.2°	13.2°	15.5°
Regression, held-out stages	15.2°	17.2°	18.8°
Regression, held-out users + stages	15.8°	17.5°	18.7°
Tracking, held-out users	7.7°	—	—

Read the top-left cell carefully: 12.2° mean joint-angle error for a new user. For comparison, UMI's SLAM tracking achieves 6.1 mm position and 3.5° rotation absolute trajectory error, and GELLO's leader-follower encoders are essentially exact. A 12° joint error propagated through a hand kinematic chain is centimetres of fingertip error — larger than the clearance of most insertion tasks in the manipulation literature.

The 2025 scaling result. Kaifosh, Reardon & CTRL-labs at Reality Labs, Nature 645(8081):702–711 (23 July 2025) [A] is the most important EMG paper of the period. A 16-channel bipolar wristband at 2 kHz (2.46 µVrms noise), with 162 participants for wrist control, 4,900 for discrete gestures, 6,627 for handwriting. Cross-user, with no per-user calibration: >90% held-out-participant classification for handwriting and gesture detection, <13° s⁻¹ wrist-angle velocity error; closed-loop 0.66 target acquisitions/s, 0.88 gesture detections/s, 20.9 words per minute. Twenty minutes of personalisation cut handwriting character error rate by 16% relative to the largest generic model.

The contrast with emg2qwerty (2024), where the generic-model character error rate was 55.38% versus 15.38% personalised-from-scratch, shows how much the cross-user gap closed with scale in roughly one year.

But note the boundary condition: there is no robot in the Nature paper. It is cursor, gesture and handwriting. The demonstrated capability is discrete or low-dimensional continuous control, calibration-free. Extrapolating it to 6-DoF end-effector pose plus a multi-fingered hand is not supported by anything in that paper.

4.8 Propagation: mapping artifacts onto the quality axes

We can now state the argument compactly. Model the recorded action as

$a_t^{\text{rec}} = f_\phi\big( \mathrm{EMG}(t - \delta_t) \big) + \eta_t$

where $\delta_t$ is a time-varying latency (§4.2), $\phi$ is a decoder whose calibration drifts within and across sessions (§4.5–4.6), and $\eta_t$ is decoder noise with variance that depends on limb configuration (§4.4) and electrode registration (§4.3).

Each term maps onto a known failure mode:

• Variable $\delta_t$ produces observation–action misalignment. Under the Belkhale decomposition this appears as action divergence, since the same observation is paired with actions issued in response to a different, earlier observation.
• Configuration-dependent $\eta_t$ means action noise is state-correlated. This is the worst case: it is neither system noise (which teaches recovery) nor i.i.d. policy noise (which averages out) but a structured error that a sufficiently expressive policy will learn.
• Session-to-session drift in $\phi$ means that a corpus collected over multiple days is a mixed-operator dataset even with a single operator. This is the robomimic MH condition, arrived at by a different route, and robomimic tells us what it costs: BC-RNN Transport 72% → 42%.
• Discrete-mode decoding produces piecewise-constant velocity commands with abrupt transitions — the signature of low SPARC and high jerk, which RINSE shows is a usable negative quality filter.
• Closed-loop operator compensation (§4.6) converts all of the above from task failure into action inconsistency, hiding the damage from teleoperation-level metrics while preserving it in the training data.

The last point deserves emphasis because it is a genuine methodological trap. Every EMG teleoperation paper we surveyed reports task-level metrics — success rate, completion time, questionnaire scores. None reports SPARC, jerk, action variance, or downstream policy success on the collected data. The field has no measurement of the thing this review is about.

Hypothesis H1 (testable, cheap): for matched tasks and operators, action logs collected via EMG interfaces will show significantly worse SPARC and higher within-state action variance than logs from leader-follower or kinesthetic collection, and BC policies trained on them will underperform by a margin exceeding the teleoperation-level success-rate gap. Nobody has run this.

---