Study on Material Recognition of Artificial Finger for Pneumatic Robot Arm

Transcription

1 Study on Material Recognition of Artificial Finger for Pneumatic Robot Arm Mitsuru TOYAMA, Osamu OYAMA and Kazuya TAMURA Department of Mechanical Engineering Informatics Meiji University Higashimita, Tama-ku, Kawasaki-shi, Kanagawa-ken, Japan ( ABSTRACT We have reported that a pneumatic finger [1] can work in the field of human tasks, such as paper handling. In this field, to know the "Human Sensitivity" is becoming increasingly important. There are manyfactors which influence human sensibility, and we measured the physical information of material affect on human recognizing by using artificial nger[2] having force and temperature sensors. Also, we constructed a Neural Network [NN]system which represents fi a human feeling and then these physical information are applied as the inputs of thenn system. After the learning against many materials, the outputs of the NN can have images like a human. KEY WORDS SD Method, Human recognition, Artificial finger, Neural Network [NN] 1. Introduction It is very interesting to know a human's tactile sensing information. The information can be used in the elds of pneumatic robots. From a physiological field, fi it has been pointed out that the skins of human have various acceptance devices and the roles of a sensor. The various tactile sensing functions feel a thermal conductivity, a surface state, hardness, etc, of material. The human can presumes the material from those results As the first step of the realization of human sensitivity system, this research treats a mentality tactile sense of human. In the tests, the both of the mental and physical data are measured and are used to construct the human sensitivity propriety of model, and then, the model is verified by using Neural Network [NN]. 2. Mental consultation experiment Here, the SD (Semantic Differential) method is used as mental analysis examination to acquire sensory test of human. In the test, 11 kinds of adjective pairs are Fluid Power. Fifth JFPS International Symposium (c) 2002 JFPS. ISBN

2 selected about the sense of finger, and the 15 kinds of objects are prepared to touch. Moreover, in order to prevent the error by individual difference, the sensory analysis is performed after the testers learned enough and they became to be able to distinguish the materials. The normalized (they are made to average 0 and standard deviation 1) multivariate sensitivity data obtained from 20 testers are shown in Tablel. Furthermore, the main factor solution is calculated by the power method. Factor loadings (the grades of correlation of the 1st and 2nd factor to an adjective pair) are shown in Table 2. Table 2 The amount of load factors to an adjective pair Table 1 Multivariate sensitivity date Each factor is analyzed from the result of Table 2.The correlations between the meanings of adj ective pairs and the values of each factor are examined, and the 1 st factor is considered to be the parameter which shows the feeling image obtained from the state on the surface of a object. The vibration and friction of a finger may have concerned to the 1st factor. The 2nd factor is an image which human feels from the thermal internal characteristic of object. As the 3 rd factor is so small although it is not shown in the table, we take account only these two factors in this study. Table 3 The amount of load factors to adjective paper The next, the correlations between each object and factor obtained by the same technique are shown in Table 3. From each factor based on this result, it is thought

3 that the object which has high value of the 1st factor is the smooth surface, and that of the 2nd factor has a small thermal conductance and the changes of temperature of human's finger will be small. This result is taken as the physical human image to each object. And they are used as supervising signals of NN construction as later shown. Moreover, the physical parameters which human probably feels are measured. This research prepares "friction","heat", and "vibration" as the physical parameters, and verifies the validity of these parameters. 3. Experiment of physical parameters Figure 2 Waveforms of frictional force change (cherry tree) 3.1 Measurement of the amount of friction The frictional characteristic measurement system used by this research is shown in Figure 1. The friction is measured with the artificial finger made from silicone rubber. The artificial finger touches by constant force at the object. And the artificial finger skims the surface of object. The finger is driven by the DC torque motor. The six load cells are stuck under the object, and they measure the force that the artificial finger gave the object. (In this research the tangential direction is P, the normal direction is F) When the object of the cherry tree is used, the example of the result is shown in Figure 2. The moment when the tangential force suddenly degreases ('h) is the moment when the finger begin to slide. The ratio of P-F at this time is used as the friction parameter. These parameters about each object are normalized and are shown in Table 4. Table 4 The friction parameter of each object 3.2 Measurement of temperature change Both the thermocouple and the silicone rubber heater were built inside of the artificial finger (Figure 3). The thermocouple is located at separated portion from a heater and measures the temperature of artificial finger. The temperature of silicone rubber heater is kept at 40 Ž. While the artificial finger contacts the object for 5sec, the temperature change of the finger is measured. The Figure 1 The frictional characteristic measurement system amount of heat change between initial value and final one is used as the thermal parameter. The parameters are normalized and are shown in Table 5.

4 Figure 3 Measurement system of temperature change Figure 4 Synthetic rubber's spectrum characteristic Table 5 The thermal parameter of each object The waveform concerns the surface state of the material. If surface of object is rough, the waveform will become oscillatory and the sectional area surrounded by the waveform and the horizontal axis is used as the surface parameter of object. The parameters are normalized and are shown in Table 6. Table 6 The surface parameter of each sample 3.3 Measurement of spectrum of vibration A small condenser microphone is attached in the artificial finger and measures the vibration when the artificial finger skims the surface of object. The Fast Fourier Transformation is performed to the output signal from the microphone. And the amplitude spectrum is induced. The spectrum of the synthetic rubber is shown in Figure 4 as an example. 4. Construction of a Neural Network [ N N] By the above experiment, the physical parameter of each object has been obtained. Human use pell-mell arrangement of these physical parameters to recognize material, the network system having the same function is constituted and the realization of human sensitivity is tried. This research use the back propagation type

5 Neural Network [NN]. The fundamental composition is shown in Figure 5. The input signals to the NN are three parameters as mentioned. The supervising signal in an output layer is the correlation between each object and factor in Table3. The composition element of a network is assumed to a Table 7. Table 7 The composition element of the network Figure 6 The error in the number of times of learning 4.2 Improvement of a network In order to improve the accuracy of NN more, the Input layer Middle layer Output layer optimal of composition element of the NN is examined. The final optimal composition elements are shown in Table8. Table 8 The composition element of the network Figure 5 The fundamental composition figure 4.1 Learning result The physical parameters of the object are put into the NN and the NN learns until the outputs agree with each supervising signal. When the NN is repeating the learning, the process of reduction of square root error to the supervising signal is shown in Figure6. From the figure, when the number of times of learning exceeds 10000, the error between the NN output and supervising becomes almost saturate. The output result after learning of the NN is shown in Table 9. They agree with original human sensitivity data (Table 3) well. The error is the sum of the square of the difference of each factor.

6 Table 9 Learning result 5. Conclusion If pneumatic robot arm has tactile sense of human, it can operate like human. In this research tactile sense information of human considered "friction", "heat", and "vibration" by mental analysis experiment, and artificial tactile sense of the human constructed by the NN. The parameters obtained in the experiment with an artificial finger were made into the input signal of NN. Also the supervising signal of NN is deduced from mental analysis experiment. Learning and normalization of result became to be good accuracy. Therefore, "Human sensitivity" has been expressed. Furthermore, we will attempt to realize the active motion of pneumatic finger with ability of recognition like human Generalizing result When an unknown object is given to the NN which have completed the learning using other objects, the comparison of the value of sensitivity data between the output of NN and the original one of human is shown in the Table9. They agree well. The human material recognition can be realized with the artificial finger and the NN. The error may be large about wooden objects such as "Cork" or "Cedar", because human can't distinguish them well, either. 6. References 1. Osamu OYAMA, Paper Handling Robot Using Pneumatic Actuators, Third JHPS International symposium on Fluid Power, PP , Hiromichi YAMADA, Study on Feeling Evaluation of Touch Sense, BPES'99 SINCE, Table 10 Generalizing result 1