A prosthetic hand controlled by a wrist camera — no electrode patches, no conscious muscle flexing — can now pick up and put down everyday objects on its own.
Researchers trained a prosthetic hand control model using imitation learning: a human operator demonstrated grasps via a teleoperation rig, and the system learned to mimic those actions. The camera reads the hand's position relative to nearby objects and triggers the appropriate grip automatically. To release, the user simply moves the hand near a surface — no signal required. Training data came from a single participant handling a limited object set, yet the model generalized to new users and previously unseen items of varying weights.
Most current prosthetics lean on surface electromyography, which reads electrical signals from residual muscle tissue. That approach works, but it demands conscious effort on every grip, fatigues users, and degrades in accuracy with sweat or electrode shift. Cutting biosignals out of the loop entirely removes that cognitive and physical tax — which matters most for people who are already managing significant physical change.
The sample size here is small and lab conditions are controlled, so the gap between a research demo and a device someone wears at a construction site or a kitchen remains wide. Still, the direction is notable: vision-driven autonomy is advancing fast enough that the muscle-interface bottleneck in prosthetics is starting to look optional rather than inevitable.