Page 13 - Read Online

Jin et al. Soft Sci 2023;3:8 https://dx.doi.org/10.20517/ss.2022.34 Page 11 of 26

overlaid with another layer of humidity sensors, which achieves excellent performance in intelligent
[51]
prostheses and peripheral nervous system interface . Hua et al. presented a highly stretchable electronic
skin matrix combined with a stacked and distributed layout, which can measure outside stimuli of seven
[122]
types [Figure 6D] . Hybrid sensors are an extension of centralized and distributed multi-modal sensors,
and both the signal crosstalk and abundant data acquirement should be carefully considered. In the future,
it is an essential topic for multi-modal sensors to achieve more tactile information with low interference in
limited space.

APPLIED TACTILE SENSORS IN ROBOTIC APPLICATIONS
Recognition of object properties
Exploring object properties is one of the most significant tasks for robotics. It is necessary for robots to
gather information on the interacted objects, which can not only provide a clear view of the surrounding
environment but also assist decision-making in high-level tasks such as manipulation controlling or
human-machine interaction . For the past two decades, image or acoustic features obtained by vision or
[17]
[123]
audio devices have been widely used in robotic object recognition . However, the light and sound may be
obstructed and fluctuate during manipulation , and it is difficult to extract tactile features without
[3]
contacting the target object. Figure 7A shows a robotic hand detecting object properties with tactile sensors.
As for the complement of vision and audio information, tactile features can significantly enrich the
description of the object characteristics, which can be commonly divided into three categories: external local
properties, global properties, and internal properties. This section summarizes the principle and advanced
algorithms to extract object properties in recent years, and Table 2 summarizes the processing methods and
practical applications of some relative works.

External local properties recognition
External local properties can directly describe the characteristics of the contacted region between the tactile
sensor and the target object, which can hardly be measured by vision and audio sensors. Specifically, if the
object is entirely homogeneous or its size is less than a sensing unit, the local properties can be similarly
equivalent to global properties. Figure 7B illustrates the commonly used exploring actions and response
signals to detect external properties including surface texture, stiffness, thermal conductivity, and the
attached chemical substance.

Surface texture mainly represents the object superficial characteristics, such as roughness, friction
[7]
coefficient, and micromorphology structures , and it has been used in sliding monitoring, textile
recognition, or even Braille reading [124,143,144] . In general, tactile sensors measuring dynamic force and
vibration are used to slide alongside the object contour, and texture features can be extracted by frequency
spectrum analysis. After that, researchers can classify texture categories with advanced machine learning
methods such as support vector machine (SVM), Bayes exploration, or long short term memory
[90]
(LSTM) [66,145] . Besides, tri-axis force sensors can estimate the surface friction coefficient , and tactile-visual-
based sensors are able to extract the surface micro-structure patterns . Stiffness (or hardness, elasticity) is
[42]
one of the most significant properties for estimating the mechanical impedance of the object. It can help
robots to adjust grasping force and prevent destruction to the interacted target , and some can even assist
[146]
[147]
interaction activities, as seen in health monitoring . For precise value measurement, it is commonly
extracted through the comprehensive analysis of force and deformation signals by static pressing . For
[127]
object classification, some use advanced algorithms such as k-Nearest Neighbor (k-NN), LSTM, and so
on [43,126] .