The participants were 20 healthy Japanese adults, including 10 males (mean age = 21.9, SD = 2.3) and 10 females (mean age = 22.9, SD = 1.2). Of these, 17 participants had no experience of contact with humanoid robots until the experiment, and 11 participants had no knowledge of humanoid robots. Two participants had contact with infants in the past 5 years. All participants provided written informed consent approved by the Ethics Committee of Graduate School of Engineering, Osaka University.
A child-like android robot named Affetto  was set on a desk in front of a chair, as shown in Fig. 2. The robot had a head and upper body, which were covered with a cloth or gloves; only its face and left forearm were exposed. The joints of the robot were physically fixed to maintain a particular posture and its left hand was held out to the participants so that the joint movements did not affect its impressions on the participants. A partition was placed between the robot and participants so that the robot could be hidden during forearm exchanges and could be shown to the participants only during the experiment.
Four different types of left forearms—A, B, C, and D—with identical sizes and appearances were prepared, and they were set (one at a time) between the left elbow and left hand of the robot under each experimental condition. Figure 3 shows the structural overview of the left forearm. Overall, the forearms were cylindrical with a diameter of 40 mm and a length of 105 mm. The main material of the forearm was cured platinum silicone rubber (Dragon Skin FX-Pro, Smooth-On Inc.), and its center was supported by a metal rod with a diameter of 8 mm. The outer surface of the silicone rubber was wrapped with a thin (7 \(\upmu \)m) polyurethane film (Airwall UV, Kyowa Ltd.) to ensure constant surface friction along the forearm.
Different amounts of two types of additives were mixed with the silicone rubber to provide different touch sensations to the forearms. One of the additives was a plasticizer (Silicone Thinner, Smooth-On Inc.), which reduces hardness, whereas the other was a thickener (Slacker, Smooth-On Inc.), which increases viscosity while reducing hardness. The additive contents were adjusted so that the hardness decreased from forearm A to forearm DFootnote 1, and so that forearms C and D had higher viscosity than forearms A and B. The different percentages of the plasticizer and thickener used for preparing the forearms are listed in Table 1.
The semantic differential (SD) method  was used to measure the touch sensations and personality impressions. Two sets of a touch sensation questionnaire (TSQ) and personality impression questionnaire (PIQ), each with lists of several pairs of opposite Japanese adjectives, were provided to the participants. They were instructed to choose their responses on a seven-point scale between opposite adjectives, e.g., “Soft or Hard” for the TSQ and “Active or Passive” for the PIQ.
Table 2 summarizes the 19 adjective pairs used in the TSQ. Most of these pairs were selected from previous studies on touch sensations of artificial skins with eight adjective pairs  and touch sensations of robot hands with 10 adjective pairs . Table 3 summarizes the 46 adjective pairs used in the PIQ. Most of these pairs were derived or selected from previous studies on personality impressions of robot hands with 12 pairs , impressions of hug dolls with 12 adjective pairs , quantification of impressions of humanoid robots with 33 adjective pairs, and meta-analysis of SD adjective pairs for personality impressions . Thus, our adjective pairs were prepared such that we could investigate touch sensations and personality impressions thoroughly. These adjective pairs were translated into English by a professional translator for use in this manuscript.
The experimental procedures were divided into three sessions. In the first one, the participants were instructed on the manner in which the robot was to be evaluated. In the second one, they practiced evaluating their own forearms in the instructed manner. In the third one, they evaluated the robot by touching its forearm and then answering the questionnaires.
In the first instruction session, a movie describing the manner of touching the robot and the two questionnaires (TSQ and PIQ) were shown to the participants. The movie showed a demonstrator pinching one of the forearms of the robot with his thumb and forefinger, touching the forearm with the pads of his fingers, and holding the forearm with his hand several times. The participants were told to look at the forearm and touch it with their dominant hand as shown in the movie. In the second session, for training purposes, the participants were told to touch their own forearm with their dominant hand and answer the TSQ and PIQ. This session was conducted to check whether the participants had understood the instructions regarding the manner of touching the forearm and to allow the participants to get acquainted with the questionnaires. The third evaluation session was divided into four subsessions in which the participants evaluated each of the four forearms attached to the robot and the entire robot. In each subsession, the participants were instructed to touch the forearm of the robot for arbitrary time durations and then answer the TSQ. After answering it, they were told to touch the forearm again and then answer the PIQ. This subsession took approximately 10 min. Between subsessions, the robot was hidden from the participants by the partition, and the robot’s attached forearm was changed; then, the robot was shown to the participants again. The forearm replacement was completed in 1 min. The order in which each forearm was showed and the order of the adjective pairs in the questionnaires were shuffled for each participant. The entire series of sessions was completed in an hour.
An exploratory factor analysis was carried out to identify several underlying factors, each of which was statistically reflected by several observed variables or evaluation scores of the adjective pairs. The maximum-likelihood estimation and varimax rotation were chosen as the factor extraction and rotation methods, respectively. The number of factors chosen according to the scree test was investigated by the Bayesian information criterion (BIC) and the root mean square error of approximation (RMSEA).
Path analysis was carried out to determine significant causal relationships between the found factors. Here, we assumed a multivariate multiple regression model whose independent variables were the touch sensation factors and the dependent variables were the personality impression factors. The maximum-likelihood estimation was used for estimating the model parameters. The R language (version 3.2.2)  was used for the above analyses.