Introduction and Scope

1.1 Motivation

The dominant constraint on generalist manipulation policies is not architecture but data. Ken Goldberg's estimate places the corpora used to train frontier vision-language models at roughly 100,000 human-years of content, against about one year for the largest teleoperated robot dataset [Goldberg, Science Robotics, Aug 2025]. Purpose-built robot corpora have grown fast — AgiBot World reports 1,001,552 trajectories totalling 2,976.4 hours from 100+ robots in a 4,000 m² facility [arXiv:2503.06669]; DROID required 50 collectors across 13 institutions over twelve months to produce 350 hours [droid-dataset.github.io] — but remain small by any comparison to language or vision.

This has produced a proliferation of alternative demonstration media: handheld grippers (UMI), wearable exoskeletons (DexUMI, DEXOP), egocentric video (Ego4D, EgoDex), simulation, and hand-matched gloves (Sunday Robotics). Each trades some combination of cost, throughput, fidelity and action-space alignment.

EMG occupies an unusual position in this space. Every other interface measures kinematics — where the hand is, how it moves. EMG measures activation — the neural drive that produces force. This is a categorically different signal, and it is the only cheap, wearable channel that observes two quantities the rest of the field cannot: grip force and limb impedance via co-contraction. Since co-contraction changes stiffness without changing net torque, it is invisible to any pose-based sensor. The question this review addresses is whether that informational advantage survives the practical costs of the medium.

The question is timely for three reasons. Cross-user generic sEMG decoding became credible in 2025 with a Nature paper reporting >90% held-out-participant accuracy with no per-user calibration across 11,236 participants [Kaifosh, Reardon & CTRL-labs at Reality Labs, Nature 645(8081):702–711, DOI]. Force-conditioned imitation learning matured rapidly through 2025–26 with a dozen credible systems. And VLA post-training data requirements fell to the 50–200 demonstration range, bringing them within reach of any lab that can run a 15-minute calibration.

1.2 Research questions

• RQ1 (medium quality). How do action logs produced through EMG/prosthetic teleoperation compare, on measurable dimensions, to those produced by VR controllers, leader-follower arms, space-mouse and kinesthetic teaching?
• RQ2 (artifact propagation). How do EMG-specific artifacts — electromechanical delay, decoder jitter, electrode shift, limb-position dependence, session nonstationarity — propagate into the smoothness, precision and consistency of the resulting robot trajectories?
• RQ3 (dataset availability). Do open datasets exist with the required tripartite structure (vision + robot actions + biosignal provenance)?
• RQ4 (performance impact). Does training on prosthetic-sourced data deliver measurable gains in task success or sample efficiency, and does EMG capture force modulation, variable stiffness and contact adaptation better than rigid controllers?
• RQ5 (scalability and integration). Is collection scalable, and is the resulting distribution better used for direct end-to-end IL training or for fine-tuning a pretrained VLA?

1.3 Contributions

1. A dataset audit establishing that the target class is empty, with the three closest near-misses characterised precisely.
2. A propagation analysis mapping EMG artifact magnitudes onto the formal demonstration-quality axes of Belkhale, Cui & Sadigh (2023).
3. A comparative table of teleoperation media with sourced numbers, including the only direct EMG-versus-conventional benchmark we could locate.
4. A three-way disambiguation of the "force helps" literature into force-as-observation, force-as-action and force-as-relabelling, with evidence that the third dominates.
5. A taxonomy of four candidate architectures, three viable and one not, with the reasoning made explicit.
6. Eleven falsifiable experiments, several of which are cheap.

---