Evaluation of simple microphone-based mechanomyography (MMG) probe sets for hand stiffness classification

We describe simple to build mechanomyography sensors, with one or two channels, based on electret microphones. We evaluate their application as a source of information about the operator’s hand stiffness, which can be used for changing a robot’s gripper stiffness during teleoperation. We explain a data acquisition procedure for further employment of a machine-learning. Finally, we present the results of three experiments and various machine learning algorithms. support vector classification, random forests, and neural-network architectures (fullyconnected articial neural networks, recurrent, convolutional) were compared in two experiments. In first and second, two probes were used with a single participant, with probes displaced during learning and testing to evaluate the influence of probe placement on classifcation. In the third experiment, a dataset was collected using two probes and seven participants. As a result of the singleprobe tests, we achieved a (binary) classification accuracy of 94 % during the multi-probe tests, large crossparticipant differences in classifcation accuracy were noted, even when normalizing per-participant.