Encord releases EMG-1 dataset pairing muscle signals with egocentric video

Encord has released EMG-1, a dataset pairing wrist-muscle measurements with synchronised first-person video for physical AI research. Developed with Wearable Devices, it combines 307 hours of field recordings with a separate 10.4-hour laboratory collection using optical motion capture. The release studies how muscle signals can help reconstruct finger movement during object handling.

The field set contains 3,804 clips from seven collection stations, using three-channel Mudra Pro wristbands with head and wrist cameras, depth and inertial measurements. The laboratory set contains 643 takes from ten participants, using eight-channel Mudra Ultimate sensing on the left wrist and independent motion-capture measurements.

A combined EMG-and-camera model achieved 15.3 mm median fingertip error, approximately 25% below Encord’s camera/held-pose reference. The reported evaluation covers 700 frames from nine takes and three participants. It supplies both methods with motion-capture palm and finger-base positions and uses half a second of future EMG in the fused model, making the result an offline finger-articulation evaluation.

Encord grades recording quality and identifies missing or degraded signals. Approximately 200 field clip-hours contain usable bilateral EMG. Dataset access is offered through registration, with raw signals, video and quality metadata; evaluated fused-pose outputs are planned for later release.

The dataset accompanies the launch of Encord Data Labs, a research division covering data capture, annotation and evaluation. Its initial work also includes Encord-REACH, a hand-tracking model for gloved hands, occlusion and changing viewpoints, connecting sensing research with the preparation of human demonstrations for robot learning.