Qualitative Research of Robot-Helping Behaviors in a Field Trial

Helping behavior, which has been studied in both social and developmental psychology since the late 1960s, is also called prosocial behavior or altruism.

Social psychology researches on helping behavior have clustered around work on interpersonal processes (such as the relationship between potential helpers and those who were helped) to answer why people assist others [11]. Latane & Darley [7] presented the following 5-step decision-making model that describes the process by which a person takes action to help: 1) recognizing a problem, 2) interpreting it as an emergency and identifying another's need, 3) deciding whether one has a responsibility to act, 4) recognizing whether one knows how to help, and 5) deciding whether to help.

In developmental psychology, the process of prosocial behaviors can be divided into three stages [12]. The first stage, which is attention to the demands of others, is influenced by the level of social cognitive development, the socialization experiences, the degree of the favor, and interest in others [13]. The second stage is motivational, which includes those who are aware of the demands of others and choose prosocial behavior. This is the prosocial judgment stage, which is influenced by the attribution of causality and empathy [14]. The third stage is the connection between intention and behavior, where prosocial judgment is enacted. This stage is also shaped by self-efficacy and the awareness of prosocial behavior [15]. In any field, helping behavior is a process.

On the other hand, studies have also focused on various factors of helpers, such as personalities, moods, and emotions in helping situations. Personality might exert more influence on prosocial behavior when the situational pressures for providing assistance are weak and its costs are high [8]. Among several prosocial personalities, empathy is considered the foundation for sustained prosociality in society [16, 17]. Bierhoff et al. [18] showed that it has the highest single correlation between prosocial behavior and personality tendencies.

During laboratory experiments, people often help a robot when it seeks assistance. Previous studies identified the factors that influence people to accept requests from robots. These factors, however, seem to vary depending on the situations. One is the relationship between a human and a robot. A pioneering study [19] argued that participants accept requests from a robot if a person with an authority was on the screen on the robot. Some type of relationship must exist between humans and robots in advance to expect that people will heed a robot's request [20]. However, Hüttenrauch and Eklundh [21] reported that about half of their participants, even those encountering a service robot for the first time, helped a robot that requested assistance.

Other factors from the robot's side, such as its form and attitude, seem related to human-helping behavior. Human perception and robot behavior depend on the physical and behavioral cues of robots (e.g., [1, 22, 23]). When comparing humanoid robots and video-displayed robots in collaborative tasks, humans more frequently respond to the instructions of the former: a physically co-present cohort [24]. Fischer et al. [25] found that it is possible to get people's attention and increase their affinity for a robot when it speaks rather than when it beeps. People helped a robot nearly 50% faster if they perceived it to be autonomous rather than being teleoperated [1]. An experiment measured robot animacy and found that an agreeable and intelligent robot maximized its animacy and caused hesitancy in users over the question of switching it off [22]. On the other hand, people's situational factors when helping others also affect such behavior toward robots. Hüttenrauch and Eklundh [21] experimentally reported that a willingness to help a robot fell sharply when they were busy or had other more pressing demands on their time.

Robots that make requests have also been studied in wild settings. A study described how some people accepted requests from a robot and provided directions on public roads [6]. This study also showed that people helped a robot without any prior relationship with it. Fallatah et al. [26] conducted field experiments in which a robot asked about items for sale at a university café and found that many people went beyond just responding to requests and actually anticipated its needs. Such responses were influenced by elements of microculture, including architecture, atmosphere, and the requested items.

Although these previous studies investigated people's helping behavior during requests from robots, they overlooked when, how, and who spontaneously assisted them in distress. The factors may be even different in the case of spontaneous-helping behavior.

In human-human communication, we often see people who aid others without being asked. For example, common prosocial behaviors include picking something up that a person dropped or asking, “can I help you?” of a person who seems to need it. The term spontaneous-helping behavior denotes such unprompted helping behaviors.

Children even help people achieve their goals without verbal requests or physical rewards. Such spontaneous prosocial (helping) behavior is usually described as indiscriminate because a child's prosocial behavior is robust across many aspects of its recipient. A recent study showed that children's helping behavior extends to robots as well. Martin et al. [27] conducted a laboratory experiment with 40 children who witnessed a robot that intentionally/unintentionally dropped an object beyond its reach. The children only returned it when they thought it was accidentally dropped. In other words, the children understood when the robot needed help, and in such situations they were willing to provide it.

If helping behavior is a process, it can be assumed that its factors will vary depending on the stages of the process. One factor at a certain stage in the generation of helping behavior towards a robot is noticing and recognizing when it needs help.

Intuitively, abuse is bad; an abuser is not good. On the other hand, a person who helps another is a good person. However, in fact, in our field trial we observed that the same person abused a robot and subsequently helped it.

Since robots are generally new in most situations and contexts, they continue to attract the attention of many people. In their interactions with robots, most people act appropriately within the bonds of social interaction that have been forged over the years. Unfortunately, some exhibit abusive behaviors, including hitting and teasing a robot. Such conduct is called robot abuse. Robot abuse is much less likely in laboratory experiments, where participants are aware that they are being observed by experimenters. However, after robots were introduced into public spaces, robot abuse began to emerge. Salvini et al. [28] reported that when they demonstrated human-sized robots in public spaces, people closely approached them out of curiosity, and some young people kicked, beat, and hit the robots. They described how the nature of the abuses suffered by the robots resembled bullying behaviors more than vandalism. Brscic et al. [29] observed children who relentlessly disturbed a robot by directing abusive language at it and kicking/hitting it when it served as a shopping mall guide. Of course, vandalism has also been identified that greatly exceeds simple bullying of robots. A notorious episode occurred in America with the HitchBot robot, which was completely destroyed [30] after it had already safely traveled across several other countries [31]. Yamada et al. [10] described how a group of children, stimulated by other children's previous similar behavior, escalated their own behavior from mild to physical forms and finally to such severe abuse as repeatedly kicking, slapping while yelling, and hitting with a stick. Robot abuse includes a wide range of behaviors, from such verbal behavior as teasing and saying nasty things to such physical aggression as hitting, kicking, and vandalizing. Although vandalism may be the ultimate form of abuse, we must first examine abusive behavior from its mild form, because it often escalates from teasing.

The factors that cause such abuse are undoubtedly complex, although some studies have focused on uncovering why it occurs. Robot abuse occurs when a robot is unattended in wild settings [28, 29]. Keijsers and Bartneck [32] concluded that low mind attribution is one source for abusing robotic agents. On the other hand, Nomura et al. [33] interviewed children who abused a robot at shopping centers and reported that their actions were fueled by curiosity or enjoyment. These children perceived the robot as more human-like than a machine and perceived that it has emotions and experiences pain, all of which are attitudes that show that they abused the robot without any hostility.

A recent study investigated ways to suppress robot abuse. After acknowledging the difficulty of getting children to stop abusing, Brscic et al. [29] proposed a behavioral design for a robot that flees from children. Ku et al. [34] similarly showed that children who abused a turtle-shaped robot stopped their abusive behaviors when it crouched (hid) in its shell. However, even though these methods stopped the abuse, they also inhibited the role for which the robots were designed. Alternatively, Connolly et al. studied whether robots can induce prosocial behavior in human adults [35]. In their study, a confederate and a participant engaged in a collaborative task with three robots, and the former abused one of the robots. Their results showed that the study's participants were more likely to prosocially intervene when the other (bystander) robots expressed sadness in response to the abuse as opposed to just ignoring these events. Lucas et al. [36] found that when verbal abuse was directed at a larger robot, participants did not label such behavior as mistreatment, although they did when similar abuse was directed at a child-sized robot. Perhaps a smaller robot is safer from abuse. Even though various studies have suppressed robot abuse, further research is needed to ensure that robots can fulfill their roles without being abused.

The robot used in the research stood 120-cm high with a 40-cm diameter on a mobile platform [37]. It conducted direction-giving and leaflet-distributing services in an area where visitors often browsed around shops and other locations. The robot wore a sign on its chest that said "Information staff" to clearly identify its role to visitors. It was equipped with a printer and could take and distribute leaflets with its left hand.

Although the robot was developed to be autonomous, it used a semi-autonomous approach to speech recognition because the automatic speech recognition (ASR) struggled in the shopping mall. In this system, the operator simply input the language spoken by the user without supplying any further knowledge.

The robot wandered around the area, approached visitors, offered a leaflet (Figure 1, left) that included a map of the area, and talked to them. Visitors freely interacted with the robot, which greeted the visitors and initiated its direction-giving service: “May I help you? I can give directions or recommend shops. Where do you want to go?” When visitors asked about specific locations, it provided detailed explanations: “Go straight along this corridor and down the escalator.” It also pointed at them or in specific directions (Figure 1, right). The robot's area of operation was recorded with several cameras in the ceiling and a microphone attached to the robot. During 13 days of recording, the robot operated for 72.25 hours and 8,493 visitors appeared around it.

The robot occasionally accidentally dropped its leaflet during its leaflet-distributing service (Figure 2). It was supposed to grab the leaflets when they emerged from its printer. However, since its arm manipulation lacked precision, it sometimes dropped them. From the recorded data, we identified 47 cases where the robot dropped the leaflets in contexts where visitors could have provided assistance.

This study was approved by the ethical review board of the researchers’ affiliations. All the interactions were in Japanese, and we translated the examples/cases into English.

We repeatedly observed and coded the video data of every case. The factors that emerged included the presence/absence of interaction prior to the helping behavior between the helper and the robot and the presence/absence of a helper. In addition, we categorized the interactions prior to the helping behavior into whether the interaction, if any, was abusive. If a helper was present, such situations were categorized by the type of pre-interaction.

Two independent coders coded each case for the following factors:

Interaction type prior to helping behavior: First, the coders determined whether each case qualified as an interaction. A case was classified as an interaction if it included any visitors who took any actions toward the robot (approaching, talking, abusing, etc.); otherwise it was coded as a no-interaction, even if the robot addressed a passerby who ignored it. Next among the interaction cases, the presence or absence of abuse was determined based on the following definitions [29]:

“Persistent offensive action, either verbal or non-verbal, or physical violence that violates the robot's role or its human-like nature.”

Thus, each case was classified based on the following three categories: an abusive interaction (at least one moment when the robot was mistreated), a non-abusive interaction, or no-interaction.

Helper: If the robot was helped by someone, the types of such visitors who picked up the dropped leaflets were sorted into the following categories: abusers (those who mistreated the robot), users (those who conventionally interacted with it), observers (people who did not interact, including those who just briefly watched the interactions) as well as those who just looked at the robot, i.e., passersby). If no one helped the robot, we labeled it as nobody.

The inter-rater agreement (the kappa coefficient) for the type of interaction was .97. We coded each interaction situation by the type of helper: .86 for abusive interactions, .94 for non-abusive interactions, and 1.00 for no-interaction situations.

Note that in our coding, every person who picked up leaflets tried to return them to the robot; if they acted out of their own self-interest, e.g., to acquire a leaflet, we did not categorize them as helpers. However, we found no such cases. Nor did we observe any other helping behaviors, e.g., efforts to stop the abuse, except for a mother who warned her son, “you'd better stop before you break it,” which failed to deter him; she made no further effort. We did not count her as a helper.

Table 1 shows the numbers of people in each category. Among the 47 analyzed cases, 29 had prior interactions with the helping behavior that were either abusive (8) or non-abusive (21). Among those 29 cases, in 23 the visitors were either only children or groups that included children. The remaining 6 cases of interacting visitors were exclusively comprised of adults or teenagers. Of the 8 cases with abusive interactions, 7 were children and one was a teenager.

Concerning the 8 cases where the robot was abused, the abusers helped the robot in 3 of them, users (other than the abuser) helped in 4, and in no cases did an observer help. Not a single person offered help in one case. Of the 21 cases of non-abusive interaction, the robot was helped in 9 by users, in one by observers, and nobody helped in 11. Of the 18 no-interaction cases, the robot was helped in 2 by observers, and nobody helped in 16. As shown in Table 1, the ratios with which the robot received helped were 88%, 48%, and 11% with abusive, non-abusive, and no-interaction.

Figure 3, which shows the TEM diagram of those who helped the robot, integrated the models of each category obtained from the procedure reported in Section 5.1.2.

We found that children and a teenager spontaneously helped the robot, even though they previously abused it. That is, we observed that not only children and a teenager who engaged in friendly communication with the robot but also those who yelled at it performed helping behaviors like picking up leaflets dropped by the robot. This counterintuitive phenomenon is confounding. Nevertheless, the target children and a teenager who offered help only engaged in relatively mild abuse. Yamada et al. [10] proposed a four-stage model of robot abuse: approach, mild abuse, physical abuse, and escalation. The target children and a teenager in our study only reached the second stage of this model, i.e., mild abuse; other children escalated to the physical abuse stage. Future work must study whether children who escalate their abuse also engage in helping behavior.

Keijsers and Bartneck [32], in a study with adults, concluded that one cause of bullying is a low estimate of the robot's mentality. In our observations, the robots were sometimes slow to respond or failed to appropriately satisfy the user's intentions. This situation might have fueled the abuse. On the other hand, Nomura et al.’s work [33] with children cited curiosity and interest as reasons. They reported that children who abused the robots said they enjoyed the reactions of the robots. Even though they abused them, they did not intend any harm. Again, children and a teenager abused the robot in our study. Since ours is an observational study, we haven't identified the reason for their abuse and helping. However, from video observations, we inferred that they were curious about the robot, teased it and enjoyed it. This is the same reason provided by the children for their abuse in Nomura et al. [33]. Since we believe they continued to enjoy their interaction with the robot, their mild abuse and helping behaviors are not necessarily contradictory. Perhaps the abuse and helping behavior are caused by mind-attribution, in which the children and a teenager who were curious about the robot see others abuse it and suppose that they can abuse it, too. Perhaps they abuse it slightly more than the others; since it dropped a flyer, they help it because nobody else was around.

The ratio of the robot being helped increased when the interaction was abusive (Table 1: abusive: 88%, non-abusive: 48%, no-interaction: 11%). Our TEM analysis highlighted the importance of attention being paid to the robot in every category of the cases. In the five-step decision-making model of helping behavior [7], the process by which a person reaches a helping behavior is divided into the following five categories: (1) recognizing that something serious is happening; (2) recognizing a crisis; (3) recognizing a responsibility to help; (4) recognizing that you know how to help; (5) choosing to help. In other words, helping behavior starts by paying attention to the person who is being helped. Robot abuse might gather more attention from others to the robot than normal interaction and ironically increase the chance that a robot will be helped.

People frequently helped the robot when it dropped a leaflet. If a person had dropped a leaflet (e.g., discount coupons), little help from strangers seems likely, although it depends on the situation. People would probably recognize that the person can easily pick it up herself. In other words, the lack of a crisis situation at stage 2 of the decision-making model of helping behavior [7] (described above) discouraged helping behaviors. In contrast, since the visitors in our studies probably felt that a robot might struggle to retrieve a dropped leaflet, they recognized that the situation is critical and provided help.

As discussed in Section 6.1, getting attention is one critical factor for the occurrence of behavior that helps robots. Note also that help was often offered by children. Our results match previous literature where children provided helping behaviors rather indiscriminately to diverse people (e.g., [43]) as well as robots [27] when they noticed such needs. In our study, children spontaneously helped the robot when they noticed its failure.

Moreover, the absence of others around the robot encouraged helping behaviors. As noted above, the presence of others diffuses responsibility and inhibits helping behavior [44]. Since the need for help is apparent from a situation in which the robot dropped its leaflet, we infer that the children who helped recognized that such assistance is their responsibility because no one else was present. In fact, this interpretation was consistent with the data. We assume that the children who helped the robot recognized their responsibility to assist it because no other person was present at stage 3 of the decision-making model [7]. This is consistent with the concentration of responsibility identified by Staub as one helping determinant [45].

Finally, a relationship built upon previous interaction with the robot might also contribute, as reported by previous studies [20, 46]. Children become familiar with robots through prior interaction with them. Perhaps children would be more motivated to assist robots if they interacted more with them, regardless whether the interaction was abusive.

Although we observed helping behaviors fueled by the robot's failures, i.e., dropping leaflets, we did not observe any helping behaviors toward another apparent problem: robot abuse. In terms of the decision-making model, a person must recognize how she can help at the fourth stage [7]. Since helping behavior in response to robot abuse is more complex than helping a robot that just dropped a leaflet, perhaps the children did not know how to respond to abuse. In fact, helping behavior is less likely to occur in human bullying [47]. Stepping in against abuse and abusers is very risky, since abusers might protest or “fight back,” fueling escalation of the conflict. The potential costs incurred by helping behaviors (e.g., physical danger, money, time, social aspects, etc.) are inhibitory factors (e.g., [45]). In addition, people judge whether the person who seeks help is worthy of such efforts (e.g., [45]). Perhaps the robot was not deemed sufficiently worthy of assistance even during its conflicts with others. In other words, one result of low mental attribution to robots is that they receive less help.

Another work that focused on group dynamics concluded that participants are more likely to intervene during robot abuse when a bystander robot expresses sadness [32]. Unfortunately, our study lacked such group dynamics, i.e., no third-party members who confirmed that the abused robot needed help. If another robot were present, and if it expressed sadness over robot abuse, or if the abused robot actively sought help from a nearby third party, some people might have intervened.

Our study's results show that a robot is more likely to be helped if any type of interaction had occurred with it. In other words, designing a robot that attracts interactions with others may lead to a robot that elicits spontaneous-helping behavior from people. Although robots that require human help will increase the burden on humans, robots that can leverage human assistance might have the potential to accomplish much more than those that do not [1]. Such a scenario might eventually increase human convenience.

Our study raises the following question: is robot abuse simply immoral behavior? We revealed that abusive behavior, at least mild abuse by children, is a manifestation of their curiosity and does not necessarily signal hostility to robots. The act of abusing itself is not ethical; abuse should not occur. About half of the children who approach robots in wild settings engage in mild abuse [10]. As more and more robots are introduced into human societies, more robots will be placed in wild settings without human supervision. What this study shows is that we must redirect the spark of interest in robots from abuse to assistance. We envision a new direction upon which to base the design of robots that can lead from mild abuse to helping, as well as a direction that reduces robot abuse.

Nevertheless, finding ways to suppress robot abuse remains critical. From the perspective of bystanders, witnessing abuse is uncomfortable (e.g., [48]). Abuse by one person invites similar treatment from others [10]. If a robot cannot cope with abusive behaviors, it might fall into a chain of abusive behaviors from many people. Robots must be designed so that they can more effectively “defend” themselves against mild-level abusive behaviors. Children who slightly abuse a robot are probably just curious about it, not bullies who dislike robots. Helpful techniques are available, e.g., the bystander robot effect [35]. In future societies when robots inevitably play even more important roles, they will be introduced into our lives and group interactions between people and robots will occur. The effectiveness of such methods based on group dynamics must be tested to determine whether they can indeed reduce robot abuse. We also hope that social cognition toward robots will eventually become more enlightened.

Since all the data in this study were observational, we were unable to examine any of the motivations of the helpers. Future studies must interview those who have been identified as helpers to clarify the causes of their abuse and/or their motivation for helping behaviors. Since our study's dataset was very small, our result must be confirmed quantitatively with a larger dataset. In addition, the helping behavior modeled in our study is "picking up flyers dropped by robots," which is obviously only one of many such potential helping behaviors that can occur in HRI. Different tasks and environments may generate different helping behaviors or perhaps none at all. In fact, in another study, we observed situations where no helping behavior occurred against abuse. In the future, we must verify whether other forms of helping behavior also occur for robots.

Cultural differences also influence helping behaviors toward robots, including those reported in HRI, prosocial, and altruistic behaviors. For example, Korte & Ayvalioglu [49] found that people from rural areas are more prosocial than urban dwellers. Levine et al. [50] conducted a worldwide field experiment on assisting behavior in 23 large cities and found a more than two-fold difference in the occurrence of such behavior among them. The differences in the motivation for altruism have also been studied. Eisenberg and Mussen [51] cite prosocial moral reasoning as a cognitive aspect of altruism, identifying five stages of this development associated with age as well as cultural differences in its development. Our study was conducted in a city shopping center in Japan and reflects urban and/or Japanese culture. We must verify whether similar behavior occurs in other cultures and/or rural areas. Further research on spontaneous human-helping behavior toward robots is desirable.