Shared Autonomy, Human–Robot Collaboration and Robotic Hands

Surface electromyography (sEMG) measured on the forearm reveals how a person intends to grasp. We regress these signals into continuous commands for multi-fingered robotic hands while avoiding the tedious instant-by-instant labelling that supervised learning requires: non-negative matrix factorization combined with deep neural networks in a self-supervised scheme, minimally supervised networks trained with a soft-DTW loss, weakly supervised regression, and temporal convolutional networks that run on a low-power microcontroller.

Kinesthetic teaching lets an operator program a collaborative robot by guiding it by hand, but it records only the path. We add a second channel: a wearable sEMG interface estimates how much the operator stiffens the hand (muscle co-contraction) without interfering with voluntary movements, and vibrotactile feedback makes this level perceivable. Stiffening the hand can request a more accurate trajectory, learned through constrained smoothing B-splines, or set the compliance of the robot point by point, in both kinesthetic and teleoperated programming by demonstration.

Muscle signals also drive devices that physically support the human body. We developed a lightweight soft exosuit actuated by twisted string actuators that assists the elbow in single- and dual-arm tasks, and human-in-the-loop controllers that let an assistive robot either compensate muscular effort or generate it for strength training. For safe physical interaction we designed variable-stiffness actuators, including an anthropomorphic elbow joint, and a twisted-string haptic interface for teleoperation with force feedback.