Machine-Learning-Based Fine Tuning of Input Signals for Mechano-Tactile Display

Deducing the input signal for a tactile display to present the target surface (i.e., solving the inverse problem for tactile displays) is challenging. We proposed the encoding and presentation (EP) method in our prior work, where we encoded the target surface by scanning it using an array of piezoelectric devices (encoding) and then drove the piezoelectric devices using the obtained signals to display the surface (presentation). The EP method reproduced the target texture with an accuracy of over 80% for the five samples tested, which we refer to as replicability. Machine learning is a promising method for solving inverse problems. In this study, we designed a neural network to connect the subjective evaluation of tactile sensation and the input signals to a display; these signals are described as time-domain waveforms. First, participants were asked to touch the surface presented by the mechano-tactile display based on the encoded data from the EP method. Then, the participants recorded the similarity of the surface compared to five material samples, which were used as the input. The encoded data for the material samples were used as the output to create a dataset of 500 vectors. By training a multilayer perceptron with the dataset, we deduced new inputs for the display. The results indicate that using machine learning for fine tuning leads to significantly better accuracy in deducing the input compared to that achieved using the EP method alone. The proposed method is therefore considered a good solution for the inverse problem for tactile displays.