Conclusions

RQ1 (medium quality). EMG-sourced action logs are measurably inferior to those from every conventional interface on every dimension the imitation-learning literature identifies as mattering. The single direct comparison found 26% slower completion, a 58× error rate, and 2.8× workload, against baselines that themselves underperform leader-follower rigs by roughly 30 points.

RQ2 (artifact propagation). The propagation chain is well characterised and every link is damaging: ~50 ms of irreducible electromechanical delay against a 100–125 ms budget; 7.6–20% accuracy loss per 2 cm of electrode shift with 1.16 cm re-introduced at each donning; a 4.7× error increase under limb-position change; 10–12 points of cross-session degradation. The most under-appreciated mechanism is that closed-loop operator compensation converts decoder error into action inconsistency rather than task failure, which hides the damage from teleoperation metrics while preserving it exactly where behaviour cloning is most sensitive.

RQ3 (dataset availability). The target class is empty. Two 2026 efforts converged on it and neither released a corpus. The closest released artifacts are MeganePro/Ninapro DB10 (EMG + scene video + gaze, 15 amputees, no robot) and RH20T (vision + actions + 6-DoF force, no biosignal).

RQ4 (performance impact). No study has trained a visuomotor policy on EMG-collected demonstrations and compared it against a conventional control. What is established: EMG-derived stiffness improves contact-rich teleoperation over constant impedance by 5–23 points and halves peak contact force under delay; EMG-derived force labels on human video enable an 11-dimensional action space that reaches 87% on pick–squeeze–place where binary-gripper baselines score zero. But the stiffness comparison has never been run against the free alternatives — demonstration covariance or vision — and a 2026 result achieves the same objective from wrist-camera RGB alone.

RQ5 (scalability and integration). Collection through EMG teleoperation is not scalable: 30–45 minutes of per-session overhead, fatigue-bounded sessions, and multi-day drift that converts single-operator corpora into mixed-quality ones. EMG as annotation inherits egocentric video's throughput and is scalable. The integration point is fine-tuning, not pretraining — modern policies adapt with 50–200 examples — and specifically fine-tuning that extends the action space rather than adding samples to an existing one.

The one-paragraph version

EMG's information advantage is real and unique: it is the only cheap wearable channel that observes grip force and co-contraction, and co-contraction is invisible to every pose sensor by construction. Its bandwidth disadvantage is equally real and is not closing at the rate that would matter — 12.2° cross-user hand-pose error is two orders of magnitude worse than a leader-follower encoder. The productive move, therefore, is to use EMG where its information advantage applies and its bandwidth disadvantage does not: out of the control loop, as an offline annotator of force on demonstrations collected by faster means, feeding a force-extended action space in a pretrained policy's post-training set. Exactly one paper does this, it is four months old and unrefereed, and the two cheapest experiments that would validate or kill the idea — a matched-budget comparison against covariance-derived stiffness, and a FELT-style test of whether vision predicts grip force anyway — have not been run.

---