1 Introduction

The applications of artificial intelligence (AI) have reached into more and more corners of human life. AI can solve small-scale problems like winning a chess game, recognizing human languages, or detecting human faces, as well as large-scale problems such as detecting outbreaks of pandemics [1] or tracing human contact for preventing the spread of a pandemic [2]. To interact with AI systems, humans receive and provide information via human–computer interfaces such as smartphones and screens. The quality of the interfaces thus plays an important role in the communication between the AI system and human users. In fact, endowing the AI systems with humanoid features has shown to be effective for encouraging health-related behavior. A series of studies have shown that the humanoid AI system, or robots, were more effective for healthcare treatment delivery [3, 4], weight management [5, 6], and persuasion in health promotion [7,8,9], than non-humanoid media such as smartphones, computers, or screen-based avatars.

The present study explored the possible persuasive influence of robots in the promotion of pro-social behaviors to satisfy societal needs. Recycling was chosen as the target behavior because of its potential to reduce waste, conserve natural resources, and enhance environmental sustainability. More importantly, appropriate recycling does not require much effort. It is a simple act; however, the benefits are significant. Consistent with the view that endowing the AI systems with humanoid features is effective in behavioral change, anthropomorphizing recycling bins has been found to be effective in the promotion of recycling [10, 11].

Several theories and models have been proposed to explain human reactions to technology. This includes the theory of reasoned action [12], theory of planned behavior [13], technology acceptance model [14], combined technology acceptance model and theory of planned behavior [15], motivation model [16], model of personal computer utilization [17], social cognitive theory [18], and the diffusion of innovation theory [19]. Venkatesh et al. [20] integrated these theories and proposed the unified theory of acceptance and use of technology (UTAUT). According to this theory, the four key factors in individuals’ intention to use a technology are performance expectancy (i.e., perceived usefulness), effort expectancy (i.e., perceived ease of use), social influence (i.e., perceived appreciation by important social networks), and facilitating conditions (i.e., perceived acquisition cost). Gender, age, experience, and voluntariness of use were found to be crucial moderating factors. All of these factors combined could account for use intention with an explanatory power of 70% [20].

On the basis of the UTAUT, Heerink et al. [21] investigated older adults’ reactions to robotics technology. They proposed the Almere model, which was named after the city in which the data were collected. Regarding the intention to use robotics technology, the following additional constructs were identified: social presence and trust, as well as perceived sociability, enjoyment, and adaptability. The first three constructs are related to user perceptions of humanness in a device. Social interactions occur when the interface has enough cues to lead individuals to categorize it as a social agent [22,23,24]. Devices that enable individuals to develop trust [9] leads to users becoming more willing to use or to interact with them. Consequently, an anthropomorphic device has more power to shape human behavior than a non-anthropomorphic one [25,26,27].

From a psychological perspective, anthropomorphism might exert effects through empathy induction [28,29,30,31,32], as similarity is an important cue in empathy induction [33, 34]. As an anthropomorphic device looks similar to human users, users can easily develop empathetic feelings towards it. Therefore, users are more likely to respond to a request from an anthropomorphic rather than a non-anthropomorphic device.

The term empathy could be defined as “an other-oriented emotional response congruent with another’s perceived welfare” [35, p. 105]. Most studies have viewed empathy as a personal trait, termed dispositional empathy. The present study focuses on context-induced empathetic emotion: situational empathy [36]. Early studies on empathy focused on human empathy towards humans, while recent studies have shown that human empathy could be extended towards non-human agents. In the study of Costa et al. [37], for example, participants were instructed to listen to a story narrated by agents with different degrees of anthropomorphic features, and the listeners’ emotional responses increased as storytelling agents possessed more anthropomorphic features.

Tan et al. [38] manipulated perceived anthropomorphism in robotic bins to test the hypothesis that anthropomorphic devices could promote recycling through empathy induction. They found that the greater the perceived anthropomorphism, the greater the participants’ empathy towards the robotic bin, and thus, inclination to use it for recycling. However, simply measuring use intention may not necessarily be predictive of the actual use of the bin. Furthermore, participants did not interact with the robot; instead, they watched a video of individuals interacting with it. This might not have reflected actual use behavior, given that studies have found that humans were more empathetic toward a physical robot than a simulated robot seen on a screen [32]. To genuinely understand a robot’s persuasive influence in the promotion of recycling, empirical data on authentic human–robot interactions are required.

In the present study, a robot or tablet computer was placed next to waste disposal bins to be used as a persuader to educate users on appropriate recycling in order to induce authentic human–robot interactions and measure actual use behavior. We hypothesized that because of the robot’s humanoid appearance, it can evoke empathy and thus increase use intention and appropriate recycling behavior.

A challenge for the design of this experiment was choosing an appropriate control device. A tablet computer was used because robots and tablet computers can produce the same amount of visual and auditory information. It would thus be reasonable to conclude that differences in participants’ responses to the robot and tablet computer are attributed to the anthropomorphic features of the robot instead of the content of the conveyed messages.

2 Methods

2.1 Participants

The experimental protocols were approved by the Research Ethics Committee for Human Subject Protection at National Chiao Tung University. All participants gave their written consent before participating. Two experiments with identical purposes but slightly different protocols were conducted approximately one month apart. There were a total of 77 participants (39 in Experiment 1; 38 in Experiment 2) for both experiments (32 of whom were females) of ages 20–36 years (M = 22, SD = 3.04).

2.2 Procedure

2.2.1 Experiment 1

Thirty-nine participants were divided into four groups of 10, 12, 9, and 8. Participants of the same group attended the experiment at the same time, and were informed that the study aimed to assess the effects of their knowledge of artificial intelligence (AI) and their attitudes toward it. The task is laid out in six steps (Table 1):

  • Step 1. Participants of the same group sat in the same room and watched the following TED Talks: Three Principles for Creating Safer AI by Russell [39], How AI Can Save Our Humanity by Lee [40], and How Humanistic AI Can Enhance Our Memory, Work and Social Lives by Gruber [41]. Participants were told that they would be called to another room to complete a questionnaire at any time while the videos were being played. They were told that the study aimed to examine the effects of their knowledge on their attitudes toward AI and the information received was manipulated through the duration of the video they had watched. There were three experimenters. Experimenter A was responsible for playing the videos and explaining the procedure in the video room.

  • Step 2. Approximately 10 min after the video began, a randomly chosen participant was called out of the video room and directed to the questionnaire room to complete the questionnaire.

  • Step 3. On the way to the questionnaire room, Experimenter B would thank them, give them a cup of water, and then guide them into the questionnaire room.

  • Step 4. As the participant proceeded, they would see Experimenter C, who showed them a box with packs of raffle tickets. The participant was asked to draw a pack from the box and unwrap it to see whether they had won a prize.

  • Step 5. After being told that they had not won a prize, they were asked to dispose of all the items in the waiting area outside the questionnaire room. This waiting area was the recycling location, which was equipped with an electronic instructor and four bins labeled “Paper,” “Soft Plastic,” “Hard Plastic,” and “Garbage.” The electronic instructor was either a Zenbo robot (Fig. 1a) or a tablet computer (Fig. 1b). The Zenbo robot, designed and manufactured by ASUSTeK Computer Inc., is 37 cm long, 37 cm wide, and 62 cm tall with a 10.1-inch liquid–crystal display screen. The tablet computer, an iPad by Apple Inc., has a 9.7-inch screen. It rested on a cuboid-shaped chair that was approximately 90 cm long, 60 cm wide, and 60 cm high. The bins were empty, and an empty cup and a plastic bag lay near the bins.

Table 1 Steps related to participants’ tasks
Full size table
Fig. 1
figure 1

Experimental setup at the recycling location for the a robot group and the b tablet computer group. In addition to the electronic instructor, there are four bins labeled (right to left) Paper, Soft Plastic, Hard Plastic, and Garbage

Full size image

The raffle pack contained seven items (Fig. 2a): the outer bag (paper), a paper raffle ticket (paper), a plastic bag (soft plastic), a sheet of cellophane (soft plastic), a plastic toy shell (hard plastic), a strand of wool (garbage), and a piece of aluminum foil (garbage).

Fig. 2
figure 2

a The raffle pack given to the participants in Experiment 1. The gift-type bags that were used in Experiment 1 were no longer available during data collection in Experiment 2; thus, they were replaced with b trash-type bags in Experiment 2. The participants were told to unwrap the pack to obtain the enclosed raffle ticket and then to dispose of all the items in the pack

Full size image

The electronic instructor played the audio instruction repeatedly in Mandarin Chinese. The English translation is provided below:

Hello! Dear everyone, please sort the trash and help with recycling. Envelopes and paper tickets should be placed in the Paper bin, soft plastic materials should be placed in the Soft Plastic bin, hard plastic materials should be placed in the Hard Plastic bin, and other materials should be placed in the Garbage bin. With your help, we can make the earth a better place.

Each sentence of the instruction was followed by a short silence. While the electronic instructor was playing the sentence, a visual instruction was displayed (Fig. 3a, b). During the silent period, human-like expressions were shown on the robot instructor screen (Fig. 3c), and the tablet computer instructor screen was blank. The head of the robot instructor would turn toward the direction of the bin that was being mentioned in the instruction.

  • Step 6. After the participant had disposed of all the items, they entered the questionnaire room to complete the questionnaire. Experimenter C recorded the items in each bin and emptied the bins. After all the items had been recorded and the bins had been emptied, Experimenter C sent a message to Experimenter A to direct another participant from the video room to go through the aforementioned procedure.

Fig. 3
figure 3

a The visual instructions for recycling the paper (left), soft plastic (middle), and hard plastic (right) items. b The image shown on the screen during the playing of the audio instructions and c the inter-sentence silence. Due to copyright issues, some of the graphic elements shown in the figure are schematic illustrations, instead of the actual stimuli, used in the experiments

Full size image

It took approximately one hour for participants of the same group to finish the experimental procedure. For the four consecutive groups of 10, 12, 9, and 8 participants who attended the experiment on the same day, the electronic instructors were, respectively, the tablet computer, robot, robot, and tablet computer.

The questionnaire (see Appendix for the English translation) comprised three sections that were edited to be the same format. In accordance with the method of Batson et al. [35] and Seo et al. [32], the questions in the empathy section required participants to state how they would feel if the device was destroyed by someone. The more negative they felt, the more empathetic they were towards the device (i.e., the electronic instructor). The second section of the questionnaire measured the perceived anthropomorphic level of the electronic instructor. The questions were those used by Bartneck et al. [10]. The last section, which contained only one question, measured the participant’s intention to follow the electronic instructor’s directions if they saw a similar recycling facility in a public space in the future.

2.2.2 Experiment 2

Thirty-eight participants were divided into four groups of 10, 9, 9, and 10. The procedure was identical to Experiment 1 except for the differences described below.

2.2.2.1 Instructors in the Control Condition

In Experiment 1, the appearance of the tablet computer was different from the robot; thus, the effects on behavior might have been confounded by low-level features. Ideally, the differences between the robot group and the tablet computer group would be the result of participants’ perceptions rather than the physical properties of the instructor. However, if the low-level features of the tablet computer instructor were identical to the robot instructor, then the tablet computer would be perceived as a robot. Since it is impossible to equalize all the low-level features for the robot and tablet computer groups, we redesigned the features of the tablet computer In Experiment 2 to be more similar to the robot. If the physical differences between the robot and the tablet computer in Experiment 1 constituted confounding factors that stimulated different behaviors, we can expect more similarities in behavioral responses to the two types of instructors in Experiment 2 than in Experiment 1. The tablet computer instructor in Experiment 2 was manufactured by U-Best Digital Technology. Its 10.1-inch screen is the same size as that of the Zenbo robot. The tablet was erected on a white box to mimic the robot (Fig. 4).

Fig. 4
figure 4

Electronic instructors in Experiment 2. On the left is the tablet computer device that mimicked the robot, and on the right is the Zenbo robot. The devices are placed together for the purpose of demonstration. In the experimental setting, the participants could see only the robot and four bins or the tablet computer and four bins. Due to copyright issues, some of the graphic elements shown in the figure are schematic illustrations, instead of the actual stimuli, used in the experiments

Full size image
2.2.2.2 Instructor Order

The electronic instructor order was robot, tablet computer, tablet computer, and robot for the four consecutive groups of 10, 9, 9, 10. This was the reverse of Experiment 1. Order-based effects were expected to be canceled if the data from the two experiments were collapsed for analysis. For Experiment 1, all four groups of participants were scheduled on the same day. Experiment 2 was scheduled for two consecutive days, with two groups each day.

2.2.2.3 Step 3 Omitted

The participants were not given cups with water. The reason for giving each participant a cup of water was to test their ability to dispose of the cup appropriately. However, some participants put the cups in their bags, as they did not deem the cup an unwanted item to be recycled.

2.2.2.4 Trash Appearance

The gift-type bags in Experiment 1 were no longer available during data collection in Experiment 2; thus, trash-type bags were used (Fig. 2b). Other trash items were the same as those in Experiment 1.

2.2.2.5 Trash on the Floor

There was no leftover trash (cups or plastic bags) on the floor in Experiment 2 because only one of the 39 participants in Experiment 1 picked it up. It is possible that the intention to pick up unknown items in public spaces might involve social or psychological factors that are irrelevant to the motivation to recycle. All the other aspects of the experimental procedure were identical to those in Experiment 1.

The aforementioned differences between Experiments 1 and 2 could have potentially affected the data. Nevertheless, in the absence of interactions between the experiment factor and interest factor, i.e., instructor type (robot or tablet), any differences between Experiments 1 and 2 were expected to be trivial and ignorable.

3 Results

3.1 Trash-Sorting Accuracy

Trash-sorting accuracy was based on the disposal of seven items: the outer bag (paper), a paper raffle ticket (paper), a plastic bag (soft plastic), a sheet of cellophane (soft plastic), a plastic toy shell (hard plastic), a strand of wool (garbage), and a piece of aluminum foil (garbage).

Trash-sorting accuracy was calculated as the percentage of items placed in the correct bin. The mean trash sorting accuracy values (Fig. 5) in the robot and tablet groups were 80% and 66% respectively in Experiment 1, and 89% and 74% respectively in Experiment 2. Two-way analysis of variance (ANOVA) was performed with the two factors being experiment (Experiment 1 or 2) and instructor type (robot or tablet computer). Instructor type had a significant effect [F (1, 73) = 7.3, p = 0.009, η2 = 0.09]. However, experiment [F (1, 73) = 2.35, p = 0.13, η2 = 0.03] and the interaction between experiment and instructor type had no significant effect [F (1, 73) = 0.009, p = 0.923, η2 < 0.001].

Fig. 5
figure 5

Mean trash-sorting accuracy in the robot and tablet computer conditions, plotted separately for Experiments 1 and 2. Error bars indicate 95% confidence intervals

Full size image

The results revealed that participants instructed by the robot sorted the trash more accurately than those instructed by the tablet computer. The pattern was consistent across the two experiments because of the minimal interaction between experiment and instructor type.

3.2 Anthropomorphism

The questions to measure anthropomorphism were printed on the back of the page, and three participants (two in the robot group and one in the tablet computer group) in Experiment 1 did not turn over the questionnaire; consequently, they did not provide an anthropomorphism rating. The analysis was based on the data from the remaining participants.

There were five questions for the measurement of anthropomorphism. The Cronbach’s alpha for these questions was 0.90. Given the high reliability of these questions, the mean was used as the anthropomorphism index for further analysis.

The mean perceived anthropomorphism ratings in the robot and tablet instructors were 3.06 and 2.26 respectively (max = 7) in Experiment 1, and 2.87 and 2.12 respectively (max = 7) in Experiment 2 (Fig. 6). The ANOVA results revealed that instructor type had a significant effect [F (1, 70) = 6.78, p = 0.01, η2 = 0.09]. Experiment [F (1, 70) = 0.31, p = 0.58, η2 = 0.004] and the interaction between experiment and instructor type had no significant effect [F (1, 70) = 0.009, p = 0.925, η2 < 0.001].

Fig. 6
figure 6

Mean perceived anthropomorphism ratings in the robot and tablet computer conditions, plotted separately for Experiments 1 and 2. Error bars indicate 95% confidence intervals

Full size image

The results indicated that the robot was perceived to be more anthropomorphic than the tablet computer. The lack of significant interaction suggests that this pattern was consistent across the two experiments.

3.3 Empathy

There were 19 questions for the measurement of empathy. The Cronbach’s alpha was 0.93. Given the high reliability of the questions, the mean was used as the empathy index for further analysis.

The mean empathy ratings in the robot and tablet groups were 4.30 and 3.72 respectively (max = 7) in Experiment 1, and 4.15 and 3.66 respectively (max = 7) in Experiment 2 (Fig. 7). The ANOVA results revealed that instructor type had a significant effect [F (1, 73) = 4.49, p = 0.04, η2 = 0.06]. However, experiment [F (1, 73) = 0.17, p = 0.68, η2 = 0.002] and the interaction of experiment and instructor type did not have a significant effect [F (1, 73) = 0.03, p = 0.87, η2 < 0.001].

Fig. 7
figure 7

Mean empathy ratings in the robot and tablet computer conditions, plotted separately for Experiments 1 and 2. Error bars indicate 95% confidence intervals

Full size image

The results indicated that the robot instructor evoked a higher level of empathy than the tablet computer instructor. The lack of significant interaction suggests that this pattern was consistent across the two experiments.

Tan et al. [38] found that the effect of anthropomorphism on use intention was mediated by empathy. If this was the case for the data in the present study, the rating values on empathy and anthropomorphism would be correlated. We collapsed the data of the two instructor types of the two experiments, and found a significant positive correlation [Pearson’s r = 0.31, t (72) = 2.74, p = 0.008] between the two variables in the present study. As some methodological differences between the two experiments might have affected the correlation between perceived anthropomorphism and empathy, we conducted a partial correlational analysis with experiment as the covariate, and the correlation was still significant [Partial r = 0.31, t (71) = 2.69, p = 0.009]. This is an index of purely the association between anthropomorphism and empathy after removing possible confounding factors between Experiments 1 and 2 (the appearance of the tablet computer, the appearance of trash, etc.).

3.4 Use Intention

The last question on the questionnaire measured participants’ willingness to follow the instructions of a similar device in the future. Following the instructor’s directions was considered use behavior in the sense that the electronic instructor provided advice for action. Unfortunately, the last question was also printed on the reverse side of the questionnaire. Three participants Experiment 1 (two in the robot group and one in the tablet computer group) did not turn over the questionnaire. The analysis was thus based on the remaining participants.

The mean use intention ratings in the robot and tablet computer groups were 6.53 and 6.47 respectively (max = 7) in Experiment 1, and 5.80 and 5.94 respectively (max = 7) in Experiment 2 (Fig. 8). The ANOVA results revealed that experiment had a significant effect [F (1, 70) = 5.81, p = 0.02, η2 = 0.08]. This was attributed to the higher ratings in Experiment 1 than in Experiment 2. Instructor type [F (1, 70) = 0.03, p = 0.87, η2 < 0.001] and the interaction between experiment and instructor type did not have a significant effect [F (1, 70) = 0.15, p = 0.70, η2 = 0.002]. Thus, use intention was higher in Experiment 1 than in Experiment 2, regardless of instructor type. An explanation is provided in the General Discussion section.

Fig. 8
figure 8

Mean ratings, in the robot and tablet computer conditions, for the intention to follow the instructor’s instructions upon seeing a similar device in the future, plotted separately for Experiments 1 and 2. Error bars indicate 95% confidence intervals

Full size image

4 General Discussion

4.1 Summary of Results

The study tested the effectiveness of a robot serving as a persuader to educate and encourage recycling. A scenario was created to allow individuals to physically interact with a robot. The results suggest that a robot is more effective than a tablet computer in educating and persuading individuals to sort trash appropriately.

4.2 Why were the Robot’s Instructions Followed?

Why did the participants recycle the items more accurately when interacting with the robot than with the tablet computer? The data suggest that the robot was perceived to be more anthropomorphic than the tablet computer. The anthropomorphic features might have induced higher empathy, as evidenced by the higher ratings of empathy for the robot than for the tablet computer, as well as the positive correlation between the anthropomorphism and empathy ratings.

These findings are consistent with the view that individuals tend to empathize with targets that are similar to them [35]. Interestingly, this empathy effect could be extended to non-human agents. According to the computers-are-social-actors (CASA) perspective [22, 24], individuals tend to apply the same social heuristics to humans and non-human devices. Empathy has been demonstrated to have a powerful effect on persuasion [42, 43]; therefore, the anthropomorphizing of a device is an effective strategy.

Compared to the study of Tan et al., who had tested the effectiveness of robotics technology in the promotion of appropriate recycling [38], the results of the present study could be more useful for two reasons. First, Tan et al. measured only self-reported intentions as the major dependent variable. The present study measured actual human–robot interactions, which should offer greater validity for predicting human behavior outside laboratory settings. Second, to confirm the persuasive effectiveness of robots for promoting recycling, a non-robot condition is required to serve as a control. This was applied in the present study; however, it was not in the study by Tan et al.

4.3 Unexpected Findings

To increase the sample size and examine the generalizability of the results to other control manipulations, two experiments were conducted in the current study. There was no significant experiment and instructor type interaction in the measurements of interest; therefore, the effects of instructor type in the two experiments were consistent.

Nevertheless, experiment had a significant effect on the participants’ use intention ratings. Generally, participants in Experiment 1 were more willing than those in Experiment 2 to follow the instructor’s directions to recycle. It is possible that this inter-experiment discrepancy was caused by the raffle pack appearance. A gift-type pack was used in Experiment 1, and a trash-type pack was used in Experiment 2. It is possible that the gift-type pack materials were more conducive to reuse and, thus, recycling than the trash-type pack materials in Experiment 2. Even if this was the case, it likely does not affect the conclusion, because the lack of interaction between experiment and device indicated that the benefits of using a robot instead of a tablet computer remained constant regardless of trash appearance.

The dissociative results for trash-sorting accuracy and use intention suggest that subjective intentions are not always predictive of actual actions. According to the major human–computer interactions theories, such as the technology acceptance model [14] and UTAUT [20], intention is a strong predictor of use behavior. However, social desirability is a possible factor when use behavior is associated with social expectations, such as recycling [44, 45]. Individuals have reported that they would practice appropriate recycling because they know that they should; however, this does not always occur in the absence of monitoring. Future research on human–machine interactions should incorporate actual use behavior measurements.

4.4 Limitations and Future Directions

The present study used two types of tablet computers for the control conditions. It could be argued that a robot and a tablet computer are different in many factors besides perceived anthropomorphism; for example, one might argue that the benefit of using a robot might have been due to its novelty. As robots are still rare, participants paid more attention and felt more curious about the robot than the tablet computer, which may cause participants to be more willing to follow the robot’s instructions. The potential effect of novelty is indeed very difficult to rule out for studies that employ screen agents as a comparison to manifest the effect of embodied social robots, such as the present study as well as multiple previous works [5, 7, 46]. Future studies are needed to assess if this persuasion maintains through time, even after the user’s curiosity recedes. Nevertheless, our findings are still valuable from a practical perspective. Tablet computers are widely used because they provide rich visual and auditory information at a low cost. Therefore, it is reasonable to treat the tablet computer as a baseline condition to investigate the effectiveness of any advanced technology-assisted persuasion technique.

One potential problem of the present study was our measurement of perceived anthropomorphism. We included the phrase “please circle the number that best describes your perceptions of anthropomorphism,” which revealed to the participants what we were measuring. Therefore, the rating values provided by the participants might have convoluted participants’ susceptibility to experimental suggestion instead of the true perceived anthropomorphism. Although this may have biased our data, the bias should be minor. If the participants were indeed biased by our instructions, the value of internal reliability among the items to measure anthropomorphism should be substantially higher than other studies using the same measurement tool. In a study conducted by Bartneck et al. [46], who used the identical items to measure anthropomorphism (they called it “human-likeness” in their study), the value of Cronbach’s alpha ranged from 0.86 to 0.93, and the corresponding value was 0.9 in the present study. Therefore, the potential bias in our data should not be a major problem. Furthermore, the data on perceived anthropomorphism is for testing the hypothesis that robots can induce higher anthropomorphism than tablet computers, and so in the current study, the focus is on the difference between the perceived anthropomorphism induced by the two devices. As the wording was the same for both questionnaires designed for the robot and for the tablet computer, the biasing effect should be canceled when looking at the difference of the perceived anthropomorphism of the two devices. Nevertheless, future research is required to tease apart the effect of device anthropomorphism on human behavior.

4.5 Implications of Our Findings

Since the industrial revolution, new technologies have emerged and have continually evolved to improve daily life. However, some industries have had negative consequences, such as greenhouse gas emissions and increased global consumption that is threatening environmental sustainability. Individuals are generally aware of climate change and sustainability issues; however, they are less willing to take action to address these issues [47]. Progress towards a sustainable future requires an increase in recycling rate. This should be the focus of interaction design [48, 49]. The present study has provided evidence on the positive effects of an empathy-evoking design on recycling behavior, even in the absence of self-reported use intention. These findings could provide insights for reciprocal human–robot relationships in the future. The use of robots for behavior change, especially sustainability, is a new trend.

It should also be noted that strict autonomy is not the only method for robot application. Functions of the robot could be achieved through partial tele-operation between a human operator with one, or even multiple robots. As human intelligence and artificial intelligence excel in different aspects, many applications can be achieved by collaborations between humans and robots, which is exemplified through the so-called "human–robot cloud" approach [50]. In a “human–robot collaboration” scenario, the quality of the working alliance between the human and the robot is essential [5, 6], and adding anthropomorphic elements to robots might help improve this quality, as suggested by the results of the present study.

Future research is required to explore more factors to motivate human–robot interactions. For example, beyond having humanoid appearances, robots possessing the “grounded situation model” [51] of the environment may facilitate human–robot communications to a larger degree. To be specific, robots should be able to monitor the physical information of the environment and the human partner through computer vision that has existed since twenty years ago [52]. Robots should also possess mental aspects of a situation, such as “beliefs of others,” or share memories with the human partner [53]. In addition, the same robotic system may need to be embodied into different forms due to requirements of different applications; how to maintain the same level of partnership with the human partner is also worth investigating [54].