We create a new tactile recording system with which we develop a three-axis fingernail-color sensor that can measure a three-dimensional force applied to fingertips by observing the change of the fingernail’s color. Since the color change is complicated, the relationships between images and three-dimensional forces were assessed using convolution neural network (CNN) models. The success of this method depends on the input data size because the CNN model learning requires big data. Thus, to efficiently obtain big data, we developed a novel measuring device, which was composed of an electronic scale and a load cell, to obtain fingernail images with 0$^\circ$ to 360$^\circ$ directional tangential force. We performed a series of evaluation experiments to obtain movies of the color changes caused by the three-axis forces and created a data set for the CNN models by transforming the movies to still images. Although we produced a generalized CNN model that can evaluate the images of any person’s fingernails, its root means square error (RMSE) exceeded both the whole and individual models, and the individual models showed the smallest RMSE. Therefore, we adopted the individual models, which precisely evaluated the tangential-force direction of the test data in an $F_x$-$F_y$ plane within around $\pm$2.5$^\circ$ error at the peak points of the applied force. Although the fingernail-color sensor possessed almost the same level of accuracy as previous sensors for normal-force tests, the present fingernail-color sensor acts as the best tangential sensor because the RMSE obtained from tangential-force tests was around 1/3 that of previous studies.

Stochastic resonance (SR) is one of the basic principles intrinsically possessed by any living thing to highly adapt to complicated environments including various disturbances. Although a noise is inevitably mixed by contact with an object and a sensor's movement on it in tactile sensing, a human being can evaluate the several micrometers of unevenness on the object surface by means of SR. We intend to apply SR to tactile sensing and to develop a tactile sensing system capable of measuring an object surface with high precision in not only a controlled environment like a precision measurement room but also in a living environment. First, we investigate the SR characteristic possessed by a Hodgkin–Huxley model capable of emulating squid's neuron activities. According to the simulation results, we develop a new electronic circuit capable of generating the SR. We perform the object surface scan using a linear stage equipped with a tactile sensor and the circuit. A series of object surface scanning tests is repeated while changing the intensity of applied noise, and the signal to noise ratio (SNR, hereafter) is calculated from the obtained measurement data to check the effect of the SR. In the experiment, striped textures with a height of δ = 5 ~ 30 μm are used as specimens. The SNR changes depending on the noise intensity, and the local maximum appears under a proper noise. It is found that the sensing accuracy is improved according to the aforementioned SR theory. Therefore, SR, which is usually applied to noisy environments, is effective for a tactile sensing system.

The degree to which a binary tactile (or visual) image matches the original object is limited by the resolution of the sensor array. Given this fundamental limitation it is still possible to minimize the error in the image formed by the interconnection of the centers of activated sensors along the object's edge. This is achieved by a suitable choice of the physical size of each sensor within the limits of the pixel size. An empirical investigation shows that normally a sensor area of about 50% of the square of the resolution yields an optimal result.

Traditionally, tactile sensors have been designed using compliant, rubber-like materials; when such a sensitized gripper grasps or otherwise manipulates an object, the normal strain deformation in the compliant material is sampled. The resulting information can be used to deduce simple local geometry of the contact, but the transduction process does not typically permit use of the individual strains in determining large-scale properties of the object (e.g., the inertia). Measurements of inertial parameters of grasped objects require accurate, low-hysteresis transduction that few tactile sensors currently provide.

An alternative is to work from the task specification, and determine the tactile information that is necessary to accomplish the task. Here, we consider how to sense the length and mass of a uniform object that is gripped in a gravitational field, and show the design and assessment of a new kind of tactile sensor that is based on the theory of the deformation of thin plates. Features of this design include its potentially rugged realization, and its high-accuracy measurement that is more typical of force sensors than of tactile sensors.

To evaluate our three-axis tactile sensor developed in preceding papers, a tactile sensor is mounted on a robotic finger with 3-degrees of freedom. We develop a dual computer system that possesses two computers to enhance processing speed: one is for tactile information processing and the other controls the robotic finger; these computers are connected to a local area network. Three kinds of experiments are performed to evaluate the robotic finger's basic abilities required for dexterous hands. First, the robotic hand touches and scans flat specimens to evaluate their surface condition. Second, it detects objects with parallelepiped and cylindrical contours. Finally, it manipulates a parallelepiped object put on a table by sliding it. Since the present robotic hand performed the above three tasks, we conclude that it is applicable to the dexterous hand in subsequent studies.

Tactile sensing is advantageous for the acquisition of local, proximal information such as the contact condition between a finger and an object. This type of sensing, however, is not suited for recognizing an entire object that is easily recognized by vision. The objective of this paper is to ease the limitations experienced in tactile sensing by using both a neural model based on the human tactile sensation and a tactile-oriented associative memory model to enable a robot to recognize object contours. In the model, first the direction vectors belonging to segments of the object contour are obtained from a filtered tactile pattern of the simulated neurons' excitation. Second, the vectors are quantized by the chain-symbolizing method and stored for use in a memory matrix that accumulates matrix-products between the vector and its transposition. In the recalling process, complete vectors are remembered even if some input vector elements are missing. In the experiments, a robotic manipulator equipped with a tactile sensor traces five types of contours, these being a circle, a square, a triangle, a star, and a hexagon. After the robot recalls the complete contours, it is able to recognize a complete contour by just touching even a part of a contour.

This paper describes a micro-optical three-axis tactile sensor capable of sensing not only normal force, but also shearing force. The normal force was detected from the integrated gray-scale values of bright pixels emitted from the contact area of conical feelers. The conical feelers were formed on a rubber sheet surface that maintains contact with an optical waveguide plate. The shearing force was detected from horizontal displacement of the conical feeler. In the experiments, a precise multi-axial loading machine was developed to measure sensing characteristics of the present sensor. Results show that the normal force was specified uniquely under combined force conditions and that the shearing force was specified by modifying the relationship between the shearing force and the horizontal displacement on the basis of normal force. We formulated a set of expressions to derive the normal force and the shearing force by taking into account this modification. Furthermore, calibration coefficients were identified for transforming the integration of gray-scale values into the normal force and for transforming the horizontal displacement into the shearing force. This result suggests that the expressions can estimate the normal force and the shearing force in wide-load regions.

This paper describes precision enhancement of an optical three-axis tactile sensor capable of detecting both normal force and tangential force. The sensor's single cell consists of a columnar feeler and 2-by-2 conical feelers. We have derived equations to precisely estimate the three-axis force from the area-sum and area-difference of the conical feelers' contact areas by taking into account wrench-length shrinkage caused by a vertical force. To evaluate the equations and determine constants included in the equations, we performed a series of calibration experiments using a manipulator-mounted tactile sensor and a combined load-testing machine. Subsequently. to evaluate the tactile sensor's practicality. it was mounted on the end of a robotic manipulator which rubbed flat specimens such as brass plates with step-heights of δ=0.05, 0.1, 0.2 mm and a brass plate with no step-height. We showed from the experimental data that the optical three-axis tactile sensor can detect not only the step-heights but also the distribution of the coefficient of friction, and that the sensor can detect fine plate inclination with accuracy to about ±0.4°.