A computational framework that is extremely fast to learn without a labyrinth of coding is known as physical reservoir computing. This computation inputs some action into a real physical system and obtains an output from the observed nonlinear dynamics of the system. It has been suggested that this output can be applied to machine learning to achieve high accuracy in time series data prediction and pattern recognition. Therefore, verification of the materials and structural factors required for highly practical physical reservoir computing is underway from a materials engineering perspective. In particular, it is important to clarify how the different physical properties of materials affect the time series data extracted from the reservoir, which responds to physical actions, and how this positively impacts learning efficiency. In this study, we developed two piezoelectric sensing end-effectors, Gel Biter and Gel Toucher, which consist of two groups of 3D structured polymeric materials inspired by the human mouth and fingers. Then, we discuss the accuracy of physical reservoir computing and the necessary elemental verification in non-destructive tactile discrimination between biting and touching. As a result, we found that both achieved a tactile discrimination accuracy of over 80%, and that the integration of different nonlinear responses, in other words, a multidimensional capture, is useful for accurate object recognition by selecting and combining the reservoirs to be used to improve accuracy.
抄録全体を表示