Differential design in homogenous sensors for classification and decoupling kinesthetic information through machine learning

Abstract

To accurately reflect the movement of humans or robots, multi-sensor integration strategy is necessary to decouple complex deformations caused by motion in the wearable artificial kinesthetic perception system. The multi-sensor integration strategy could collect multi-dimension information, making up for the deficiency of robustness and accuracy of single sensor in complex motion scenes and extending the sensing capability of the system. However, the inevitable inconsistency between multiple sensors causes difficulties in fabrication, integration, and perception, limiting the development of artificial kinesthetic perception systems toward the direction of compact integration, large-area sensing, and high-resolution decoupling. Here, we propose a strategy to build an artificial kinesthetic perception system based on the differential design of homogeneous sensors. The strategy aims at guiding system design to avoid the inconsistency in multi-sensor integration by revealing the explicit relationship among structure, signal, and movements from the perspective of the physics model. According to the strategy, we presented a simple fabrication method of the kinesthetic perception prototype. The prototype with two homogenous sensors (0.019 56 residual on average) shows three differential signal modes to three deformations. With the help of machine learning, it realized the decoupling task to 25 kinds of complex deformations. The accuracy remains at 95% even though the decoupling resolution is up to 0.2 mm. With more than one prototype arrayed, complex deformation composed with more kinds of basic deformation (shear and twist) could be further decoupled. We believe that the strategy described in this paper will contribute to the development of a compact and programmable kinesthetic perception system.