Robots see the hand. They don't feel the grip.
Vision-trained robots plateau at the last inch of fine motor control, because a camera never sees the force or the motor command behind a movement.
Manipulation policies trained on video and sim look great in benchmarks and break in the real world. The reason is simple: the camera never observed force, the model never learned tension, and the simulator never produced the rich, noisy, person-specific motor data a body actually generates. The gap shows up in the last inch — picking a delicate object, applying just enough pressure, knowing when contact has occurred.
We capture the neuromuscular signal and applied force directly at the wrist, while a human performs the task. The result is paired data: the action and the motor intent that produced it — the part of the demonstration that sim can't produce.
- Force-aware telemetry. EMG amplitude as a proxy for applied force, time-aligned with motion.
- Sim-to-real bridge. Real-world neuromuscular data to complement synthetic demonstrations.
- Imitation-learning ground truth. Wrist-EMG + IMU + (Pro) PPG streams during human task execution.
- Pre-motion intent. The command at the wrist forms before the hand finishes moving — useful for teleop loops.