Abstract

Vision-language-action (VLA) models and diffusion-based visuomotor policies are bottlenecked by the cost and narrowness of paired vision–action data. This has motivated a search for cheaper, higher-bandwidth, or higher-information demonstration media. Electromyographic (EMG) and prosthetic teleoperation interfaces are an appealing candidate because, uniquely among human input devices, they carry a direct correlate of muscle activation and therefore of force and limb impedance — signals that motion-capture, VR controllers and leader-follower arms discard by construction.

We conducted a systematic search across arXiv, IEEE Xplore, CoRL/RSS/ICRA/IROS proceedings, PhysioNet, Hugging Face, Zenodo and IEEE DataPort, and assembled a corpus of approximately 120 primary sources. We report four principal findings.

First, the target dataset class is empty. No openly released dataset pairs visual observations with robot action logs generated through an EMG-driven or prosthetic teleoperation interface. Two independent 2026 efforts — ForceBand (Amazon FAR/UMD/JHU) and DexEMG (Sharpa/SJTU) — converged on the problem within four months of each other and neither released a corpus. The intersection of the three required properties is currently unoccupied.

Second, EMG is a poor pose channel and this is quantitatively established. In the only direct head-to-head we located with full metrics (M4Bench, N=50), an eye-tracker-plus-EMG interface was ~26% slower, carried a ~58× higher error rate and ~2.8× the NASA-TLX workload of mouse-and-keyboard or gamepad. Combined with an electromechanical delay floor of ~50 ms, a controller-delay usability ceiling of 100–125 ms, 7.6–20% accuracy loss under 2 cm electrode shift, a 3.8%→18% error increase under limb-position change and ~10–12 percentage points of cross-session degradation, the resulting action logs sit on the wrong side of every demonstration-quality axis identified in the imitation-learning literature.

Third, EMG is a promising force channel, and the correct reframing is EMG-as-annotation rather than EMG-as-interface. ForceBand (June 2026) uses a $300 8-channel band and a 15-minute calibration to label ordinary egocentric human video with fingertip force, then discards the force sensors. It reports ~1.9× better finger-level contact detection than vision-based force inference and 87% success on pick–squeeze–place with an 11-dimensional action space that includes a commanded grip force spanning 3.2–19.3 N, against 0/10 squeeze success for a binary-gripper baseline. This uses EMG's information advantage without paying its bandwidth penalty.

Fourth, the appropriate integration point is fine-tuning with action-space extension, not direct end-to-end training. Modern pretrained policies adapt to new tasks with 50–150 demonstrations and to new embodiments with fewer than 200 examples; LoRA matches full fine-tuning at 1.4% of parameters and one-eighth the compute. Any realistically collectible EMG corpus is orders of magnitude too small for pretraining and comfortably sufficient for post-training — provided the contribution is a new action dimension (force, stiffness) rather than more samples of an existing one.

We close with eleven falsifiable experiments, the most important of which — a matched-data-budget comparison of EMG-derived versus covariance-derived endpoint stiffness — has, to our knowledge, never been run.

---