Improving Control of Dexterous Hand Prostheses Using Adaptive Learning

At the time of writing, highly dexterous hand prostheses are being manufactured and, to some extent, marketed. This means wearable, implantable mechanical hands with many independently controllable degrees of freedom, e.g. finger flexion and thumb rotation. Still, control by the patient is an open issue, and the most promising way ahead is probably machine learning applied to surface electromyography (sEMG). Researchers have mainly concentrated so far on improving the accuracy of sEMG classification and/or regression; but in general, a finer control implies longer and harder training times. A more natural form of control might shorten the time a patient requires to learn how to use the prosthesis, but the machine training time will inevitably be longer. In this work we propose a general method to re-use past experience, in the form of models synthesised from previous users, to boost the adaptivity of the prosthesis and dramatically shorten the training time. Extensive tests on a database recorded from 10 healthy subjects in controlled and non-controlled conditions reveal that the method is highly effective.