Stimulation in Training Data to Improve Proportional Motor Control in a Dexterous Sensorimotor Task

While electrocutaneous stimulation provides a lowcost approach to sensory feedback, it produces signal artifact which corrupts sEMG and can create errors in sEMG motor control. Eight volunteers with intact limbs controlled a 5-degree-of-freedom virtual hand using surface electromyography. The hand was trained by mimicking flexion and extension of each degree-of-freedom five times, holding each position for three seconds. Participants received sensory feedback in the form of electrocutaneous stimulation: 100 microsecond pulses at 100 Hz with amplitude linearly proportional to the force experienced by the index finger while pressing a virtual block. We compared ability to discriminate between block compliance when using training data with and without stimulation artifact. Though not statistically significant, randomly occurring stimulation during training improved ability to discriminate more compliant objects from 76 to 82 % and 72 to 85 % correct for two proportional control algorithms-a modified Kalman filter and a convolutional neural network. This approach requires no extra computational resources while providing improved motor control for bidirectional prostheses.