Nerva Apply for a pilot

Make your egocentric data 12× more useful by wearing a wristband.

Nerva is a wristband your capture team wears while recording. Its 16 EMG channels read the muscles that move the fingers and turn them into per-finger force and 20-DoF hand pose, synced to your head camera. In published wrist-EMG research, robots trained on demos with per-finger force succeeded 12× more often on squeeze tasks: 90% versus 7.5% without it.

He et al., ForceBand (2026): 72 of 80 squeeze trials with force, 6 of 80 for the best policy without.

Apply for a pilot Limited spots in the first cohort.
Pilots launch in November. Get better data, fast.

Fingertip force on all five fingers0.92 Nerror, published
Hand pose, including fingers the camera can’t see20 DoFper hand, target
Grasp and release events, ahead of contact30–100 mslead, published
16-channel EMG, time-synced to your head camera2 kHzsample rate, target
Skin temperature beside the electrodes±0.2 °Caccuracy, target
Humidity at the skin, to track sweat and contact drift±2 %RHaccuracy, target
Dry electrodes, calibrated once per wearer15 mincalibration, published

Figure 1. Hand pose decoded from 16-channel wrist EMG against motion capture, for a wearer outside the training set.

Specifications

Every value is labelled by its basis. Target: a design goal for Nerva’s first units, which have not been measured yet. Published: results from other teams’ wrist-EMG hardware or the physiology literature. Measured: our own run of an open model on public data.

Parameter	Conditions	Value	Unit	Basis
Signal
EMG channels	Dry electrodes at the wrist	16	ch	Target
EMG sample rate		2000	Hz	Target
Hand pose
Joint angles	Per hand	20	DoF	Target
Joint angle error	Wrist EMG only, wearers not in training set	12.2	°	Published¹
Joint angle error	Starting pose given, 328 windows, 36 held-out wearers; holding the starting pose scores 13.3°	10.2	°	Measured²
Force
Fingertip force channels		5	ch	Target
Fingertip force error	Calibrated wearer, per finger	0.92	N	Published³
Force error relative to video	Against video-based estimators	< 0.5	×	Published³
Robot squeeze success, with force	80 trials, same human demos	90	%	Published³
Robot squeeze success, without force	Best force-free policy, 80 trials	7.5	%	Published³
Signal quality
Skin temperature accuracy	Beside the electrodes	±0.2	°C	Target
Skin humidity accuracy	Skin side of the band	±2	%RH	Target
Temperature and humidity rate	Logged with every episode	10	Hz	Target
Timing and setup
EMG onset before force	Electromechanical delay	30–100	ms	Published⁴
Calibration	Per wearer, against fingertip force sensors	15	min	Published³

Salter et al., emg2pose, NeurIPS 2024. Regression model, wearers outside the training set.
Our evaluation of the released vemg2pose tracking weights on emg2pose sessions from wearers outside the training set (Figure 1).
He et al., ForceBand: Learning Forceful Manipulation with sEMG, arXiv:2606.26093 (2026).
Physiology literature on electromechanical delay.

Timing

Muscle activity starts 30 to 100 ms before fingertip force builds. Grasp onset and release come out of the EMG stream as timestamped events, before contact is visible on video.

Figure 2. Idealized timing, not measured data.

Output

One folder per episode, every stream aligned to the first frame of your video. Delivered as LeRobot, RLDS or plain HDF5.

episode_000123/
  head_rgb.mp4          # your camera, unchanged
  emg                   float32 [T, 16]   2000 Hz
  imu                   float32 [T, 9]     200 Hz
  hand_pose             float32 [T, 20]   rad, joint angles
  fingertip_force       float32 [T, 5]    N
  skin_temp             float32 [T_aux]   10 Hz, °C
  skin_humidity         float32 [T_aux]   10 Hz, %RH
  events.json           # grasp_onset, release timestamps
  calibration.json      # per-wearer force model

Pilots

Pilots start November 2026.

We train a policy on your sessions with and without force, so your buyers see the difference.