Liquid-metal-based soft multiaxial force sensor and deep learning-based signal processing for electronic skin

Abstract

Robots with rigid structures have been studied for a long time, but recently, interest in soft robots has been increasing. Humans understand objects with their excellent skin properties. Skin is the largest organ in the human body and has a variety of outstanding properties such as stretchability, self-healing ability, high mechanical toughness, and tactile sensing capability. Furthermore, skin can sense force, temperature, vibration, and so on. Human-like sensing capabilities will provide more opportunities for soft robots. Therefore, there are many attempts to create human-like sensing abilities for robots through fabricating artificial skin so-called electronic skin (e-skin). The skin can perceive different kinds of forces, such as normal, shear, and tangential/radial forces. However, all soft multiaxial force sensor with accurate force prediction for electronic skin is still challenging obstacles to be solve. This work introduces liquid-metal-based soft multiaxial force sensor and deep learning-based signal processing for electronic skin. Fused deposition modeling (FDM) is used to fabricate the master mold that is based on a water-soluble thermoplastic polymer filament (Polyvinyl alcohol, PVA). The PVA master mold is used to create the liquid-metal-based channels and Dragon skin 10 is used as an elastomer. The proposed multiaxial force sensor consists of two layers of liquid metal microchannels and each layer consists of multichannel. The sensor matrix is created with the perpendicular placement of two layers. Then, 3D dome structure is made by vacuum chamber. The crossing area of sensor matrix is placed at a hole and vacuum was applied to construct a dome structure. The dome structure is fabricated with Dragon skin 10 which is the same material with sensor elastomer. Finally, the multichannel sensor matrix is placed on the top of the dome structure. Thus, this structure enables the measurement of normal force and radial force applied to the sensor matrix. Furthermore, in virtue of the proposed structure, the direction and the angle of the force can be detected. Data acquisition system and data synchronization with dynamic time warping processes are done to characterize the sensor. The sensing range of the liquid-metal-based soft sensors can be adjusted by the thickness of the elastomer. Using the minimum thickness for the elastomer resulted with maximum 2 N sensing for the proposed sensor. Due to the 3D printing fabrication process, the sensor demonstrates transient response at location-to-location. Sensor characterization results showed that the transient response pattern is different in each location. Thus, contact localization of each location in the sensor matrix could be done by different transient response patterns. A dual-task convolutional neural network is used for applied force prediction and contact localization of the 5 different points of the proposed sensor. Finally, classification of 5 different points was achieved with the accuracy of 98% by applying preprocessed sensor data which was collected from 6 different channels of the proposed sensor.