Chapter 6
Dataset Audit
3 sections · about 4 minutes
6.1 The verdict
No openly released dataset pairs (a) visual observations, (b) robot action logs, and (c) EMG- or prosthesis-driven teleoperation provenance. The intersection is empty as of August 2026. The landscape decomposes into three non-overlapping silos.
Silo 1 — EMG corpora. Large and mature, with hand pose, gesture class or finger force as ground truth. emg2pose: 193 participants, 370 hours, 25,253 files, 2 kHz, CC-BY-NC-SA-4.0. emg2qwerty: 108 users, 346.4 hours, 5,262,671 keystrokes, 32 channels. EMG-EPN-612: 612 subjects. Ninapro DB1–DB10 (DB3, DB7, DB8 and DB10 include amputee participants). Hyser: 256-channel HD-sEMG, 142.8 GB. GRABMyo: 43 subjects across 3 days. Of roughly twenty corpora, exactly two contain a vision modality — MeganePro / Ninapro DB10 (12-channel sEMG at 1926 Hz + eye tracking + 1920×1080 scene video, 45 participants including 15 transradial amputees) and putEMG (24-channel at 5120 Hz + RGB + depth + isometric finger force, 44 subjects). Zero contain robot action logs.
Silo 2 — Robot manipulation corpora. Enormous, some with force. RH20T is the structural template: 110k+ contact-rich sequences, 147 tasks, 7 robot configurations, 6-DoF force/torque at 100 Hz on all configurations, fingertip tactile at 200 Hz on one, collected via haptic device and pedal. AgiBot World: 1,001,552 trajectories / 2,976.4 hours. DROID: 76k / 350 hours. Open X-Embodiment: 1M+ trajectories, 22 embodiments — note that force/torque is not an RLDS standard field and appears only in specific constituent datasets; we could not obtain a defensible fraction and recommend citing no percentage. Zero contain any biosignal.
Silo 3 — Human hand corpora. GRAB, DexYCB, ARCTIC and ContactPose provide contact geometry, not measured force. Only PressureVision++, Feel the Force and the OSMO glove provide instrumented pressure. EgoDex provides 829 hours of egocentric hand pose with no force at all. Zero contain EMG; zero contain robot actions.
6.2 The near misses, characterised precisely
ForceBand — He, Wang, Kuang, Ghosh, Malik, Fermüller, Wu, Mao, Liu, Qi & Aloimonos (Amazon FAR / UMD / JHU), 24 June 2026 [arXiv:2606.26093] [C]. The only work that puts sEMG, egocentric vision and robot manipulation in one paper. 10 hours of synchronised egocentric video + 8-channel sEMG + IMU + fingertip forces from 4 subjects at 250 Hz; separately, 15 robot demonstrations per object across 9 objects on a UR-5 with Robotiq gripper. Fails the target class on two counts: the sEMG+video corpus is human demonstration, not robot teleoperation (sEMG is converted to force labels post hoc via a learned EMG2Force model), and the robot action logs are a separate set without EMG. The project page states the dataset is "Coming Soon"; it was not downloadable at the time of this review.
DexEMG — Zhao, Li, Wang & Zhang (Sharpa / SJTU), March 2026 [arXiv:2603.05861] [C]. Genuine EMG-driven teleoperation producing real 22-DoF robot action logs on a Sharpa Wave hand, using an emg2pose-style encoder (2× 1-D conv → 2× TDS stages → LSTM) predicting joint velocities, then kinematic retargeting by L2 keypoint optimisation. Grasping success 76.0% trained / 66.0% unseen objects / 56.0% novel environments; pose MAE 0.09 rad grasping, 0.15 rad in-hand rotation. Fails the target class because nothing is released, no vision-observation stream is part of the pipeline (retargeting is purely kinematic), and the paper explicitly states it "does not train reinforcement learning policies." It also states that DexEMG "currently requires individual calibration for new users," and does not report control rate or latency.
HannesImitation — Alessi, Vasile, Ceola, Pasquale, Boccardo & Natale (IIT), 2025 [arXiv:2508.00491] [C]. The closest thing to prosthesis-sourced robot training data that actually exists. The Hannes prosthetic hand (3 DoF) with a palm-embedded RGB camera; 450 demonstrations across 15 YCB objects collected by teleoperating the prosthesis via keyboard; a reduced-U-Net diffusion policy running at ~35 Hz. Success: table grasp 80.6%, shelf grasp 68%, human-to-prosthesis handover 89.3%, overall 79.3% across 450 trials, unseen objects 76%, beating a visual-servoing baseline by 13.8 percentage points. Fails the target class because the teleoperation is by keyboard, not EMG, and the prosthesis is not mounted on a robot arm — but it is the existence proof that a prosthetic hand can serve as a data-collection end-effector for a modern visuomotor policy.
ABB Robotics × PSYONIC, June 2026 [D]. PSYONIC's myoelectric Ability Hand reportedly supplying "movement, contact and grip force" data from real prosthetic use into ABB GoFa dexterity work. This is a press release with no hours, no demonstration counts, no benchmark and no paper. Cite only as industry signal. It is nonetheless the clearest indication that the idea has commercial traction.
6.3 Why the gap is structural rather than an artifact of search
The two research communities have incompatible ground-truth conventions. The EMG community's target variable is hand pose, gesture class or finger force, collected under constrained laboratory protocols with mocap or force-sensor supervision and no robot. The robot-learning community's teleoperation stack — VR controllers, haptic devices, leader-follower arms, exoskeletons — is chosen specifically because it delivers clean, high-bandwidth pose signals that EMG cannot match.
The exoskeleton line makes this explicit. DEXOP (MIT), DexUMI, DexEXO, ACE, AnyTeleop, Open-TeleVision, HOMIE and TWIST2 all record joint positions via mechanical or optical encoders, and not one of them records muscle activity — despite several of them placing hardware directly over the forearm musculature. The omission is deliberate: mechanical sensing is more accurate, needs no calibration, and does not drift.
That two independent groups arrived at the problem in 2026 and neither shipped a corpus is itself the strongest evidence that the class is nascent. This is a citable gap, not a search failure.