PROCEEDINGS OF THE IEEE, VOL. 92, NO. 11, NOVEMBER 2004 On Activating Human Communications
With Pet-Type Robot AIBO MASAHIRO FUJITA, MEMBER, IEEE Invited Paper     In this paper we describe effects of human interactions with
a pet-type robot, especially with AIBO. First, we describe a design concept
for AIBO based on how to increase its “lifelike� appearance. By
introducing statistical results of marketing, and experiments involving human–robot
interactions using AIBO, we show that this pet-type robot activates human
emotions effectively. Furthermore, the experiments demonstrate that AIBO helps
in human-human communication. We discuss the phenomena of interaction with
AIBO, and attempt to explain why this happened.     Manuscript received June 20, 2003; revised May 27, 2004. II.  BACKGROUND: ROBOT ENTERTAINMENT III.  DESIGN CONCEPT FOR PET-TYPE ROBOT     A.  Configuration With High Degree of Freedom     B.  Multiple Motivations for Movement     C.  Nonrepeated Behavior Exhibition     D.  Behavior Control Architecture IV.  MARKET REACTIONS V.  SOME EXPERIMENTS USING AIBO     A.  Experiments in Pediatrics Hospital Ward     B.  Effect on Immune System     C.  Online Discussion Analysis VI.  DISCUSSION VII.  CUSTOMIZATION AND OPEN-ENDED SYSTEMS VIII.  SUMMARY AND FUTURE WORK ACKNOWLEDGMENT REFERENCES I.  INTRODUCTION     We are a couple whose children are all married. After all our children left our home and now there are only two of us, we seldom talked with each other. We had been spending a depressed life for a while. Recently we purchased AIBO. After we live with AIBO, we started talking with each other, and we became a cheerful couple.     We received similar reports from other users. Furthermore, there were some reports from institutions which said children with autism recovered after they had interacted with AIBO. Thus, AIBO seems to be a good partner with users, having a positive effect on their emotional state.     Recently, some research groups in Japan and in the United States started experiments using AIBO to show that AIBO is useful for mental therapy from a medical point of view [3][4][5][6][7][8]. In this paper we introduce these experiments and discuss what is essential for the mental benefits.     Before entering into the discussion, we start by explaining the development of AIBO. Why do we focus on entertainment applications rather than a “useful� robot? Why did we choose a pet-type application among many other entertainment applications such as remote-controlled robots? In the following section, we briefly explain the reasons.     Then we describe the design concept of our pet-type robot AIBO focusing on “how to maximize its lifelike appearance,� which is then followed by the implementation on how to maximize this effect. II.  BACKGROUND: ROBOT ENTERTAINMENT
Current technologies are not sufficiently mature to solve the reliability problem. The second approach is to reduce the size and weight of the robot; however, it now becomes difficult for a small robot to conduct substantial physical work.     This is the primary reason why we proposed robot entertainment as our target application. For entertainment purposes, a small-size robot can entertain a human in many different ways. Control exceptions and misrecognition can be allowed to occur in entertainment applications. For example, if a pet-type robot misrecognizes its owner, it would not pose a serious problem. On the other hand, if a security robot mistakes its owner for someone else, it is clearly serious.     Thus, by focusing on entertainment applications, the reliability problem caused by immature technologies becomes acceptable to users. This means that we can start to build a product using currently existing technologies. Of course, we not only use current technologies, but further advances are welcomed to enhance the features of the pet-type robot or to build new applications. Based on this concept, we have built a number of entertainment robots. Fig. 1 shows the first prototype of our entertainment robot as a pet-type robot, which is named MUTANT [9].     MUTANT was developed with the aim of studying pet-type robots. In order to develop Robot Entertainment, many styles of entertainment robots were built, including six legged, wheel based, and four legged. In terms of applications, we tried to build a remote controlled robot in parallel. For example, using a head-mounted display, we tested how it was interesting to control the robot from the robot's point of view. Although some technical problems remained, we concluded that it was problematic to our product plan if we would like to build a remote-controlled robot, but it was not a technical problem. Unfortunately, we did not feel much enthusiasm for this type of robot. However, we started feeling something new for the pet-type robot. There are many technical problems to be solved to achieve the pet-type robot. However, when we saw its motion with many degrees of freedom, we felt that the robot was a living thing.     We realized that a pet-type robot is a very good target as a first application for Robot Entertainment, mainly because it is quite a new product, it can generate big impact to the public, and it can be realized using state-of-the-art technology. There are many research and development issues that remain to build even better pet-type robots. III.  DESIGN CONCEPT FOR PET-TYPE ROBOT
    Regarding the first factor, we considered that it is important to have a high number of degrees of freedom to obtain a lifelike appearance. There are many animal-like toys with motion but possessing low degree of freedom. However, many people feel a difference between these toys and AIBO, as the high-degree-of-freedom configuration makes Sony's robot different from other toys.     Regarding the second factor, the meaning of motivation has a similar meaning as in the motivation of animal behaviors [10]. It is alternatively referred to as drives, emotions [11], or instinct. Naturally, the instinct and emotion model for this pet-type robot processes information in a style similar to the mammalian brain and takes account of biological behavior. Eventually, we moved toward designing the instinct and emotion model based on ethological studies [11], [12], but in the beginning we focused on increasing complexity using instinct and emotions. For example, using emotions, we can simply design “laughing behaviorsâ€? that depend on the level of joy. In addition, we introduce some other mechanisms to maximize the complexity of movement and behaviors, which are described later.     The two factors explained above are mainly solutions to increase the number of manifested behaviors. However, maximizing the complexity of responses and movements does not mean only increasing the number of behaviors. The third factor provides a solution to the problem of how to realize nonrepeated behavior exhibition. Note that the instinct and emotions can also work to realize nonrepeated actions because, using the instinct and emotion internal variables, a robot can respond differently to the same external stimuli. We describe this more in details later. We further introduce learning and development of behaviors through interactions with humans and the environment, which are also effective in realizing nonrepeated behavior.     A.  Configuration With High Degree of Freedom     Fig. 2 shows the cosmetic design of AIBO ERS-110. In several trials of cosmetic design, we concluded that it must not try to mimic a “real animalâ€? too much. It is better that the cosmetic design gives a “mechanicalâ€? impression, but once it moves, with its smooth and complex motions, a user feels that it is alive. Thus, the expectation of the user is well managed.    Â
In order to realize such impressions, we use almost the same configuration as MUTANT for AIBO ERS-110, except that it now has two degrees of freedom for the tail and one degree of freedom for the mouth. In addition, each toe passively moves so that when AIBO walks the toes show smooth movement. Each ear also moves passively so that when AIBO shakes its head, the ear follows the head naturally. Thus, there are several careful design considerations so that AIBO's motions give a “lifelike� impression to a user.     B.  Multiple Motivations for Movement     We basically use three different motivations that can be interpreted as being distributed in three different domains. The first domain is time. We consider that a robot must respond very quickly to some stimuli such as a loud sound. On the other hand, a robot must also behave after deliberation. Thus, there are some motivations that lie at widely separated points on the time axis. Some behaviors must respond in a very short time, but some behaviors can respond slowly and with deliberation.     The second domain is “internal or external� domain. Like an animal, a robot has (artificial) instincts and emotions. In AIBO ERS-110's case, there are four instincts, which are “affection,� “investigation,� “exercise,� and “appetite.� AIBO's value of affection (representing the desire for obtaining affection) increases when there is less human interaction and decreases when there is more. If the affection value becomes large, then AIBO begins to search out a human to interact with. The investigation value increases when there is no stimulus and decreases when there are stimuli. If the value of investigation becomes large, then AIBO starts to search for stimuli by walking. The exercise value increases when AIBO does not move for a while and decreases when AIBO moves a lot. If the value of exercise becomes large, then AIBO starts to walk and dance. The appetite value increases when the battery discharges. When this value becomes large, AIBO asks a user to charge it using its motion and sound.     For additional internal motivation, we implemented six artificial emotions, derived from Ekman's six basic emotions [13]. These are joy, sadness, anger, surprise, fear, and disgust. As we described above for artificial instincts, these artificial emotions also increase and decrease depending on the increase and decrease of the artificial instincts. If the value of emotions becomes large, corresponding emotion expression behaviors are displayed.     In addition, the behavior selection mechanism is influenced by the emotions, so that the tendency of the behaviors to be exhibited properly changes depending on the robot's emotional status, which will be explained later.     The third domain for the motivations is “body parts of a robot.� Here, body parts are such things as the head, tail, and legs. For example, assume that the head part has its own motivations of behaviors. Then, the number of the behaviors exhibited by the robot becomes a multiple of the number of behaviors for each part. For example, if the head part has ![[${N}_{h}$]](http://mathfigs.ieeexplore.ieee.org/iel5/5/29676/1347460/1182395.gif){kind=small} behaviors, the tail part has ![[${N}_{t}$]](http://mathfigs.ieeexplore.ieee.org/iel5/5/29676/1347460/1182396.gif){kind=small} behaviors, and the legs part has ![[${N}_{l}$]](http://mathfigs.ieeexplore.ieee.org/iel5/5/29676/1347460/1182397.gif){kind=small} behaviors, then the total number of behaviors that can be
generated by the entire robot body becomes ![[${N}_{h} \ast {N}_{t} \ast {N}_{l}$]](http://mathfigs.ieeexplore.ieee.org/iel5/5/29676/1347460/1182398.gif){kind=small} , where the number of the designed behaviors is only ![[$({N}_{h} +{N}_{t} +{N}_{l})$]](http://mathfigs.ieeexplore.ieee.org/iel5/5/29676/1347460/1182399.gif){kind=small} . If we design the behaviors
only for the entire robot body, the number of generated behaviors becomes , which is of course the
same as the number of the designed behaviors.    C.  Nonrepeated Behavior Exhibition    Â
Even if we provide the two features described above for a “lifelikeâ€? appearance, when a user once realizes that the robot exhibits repeated behavior, they suddenly lose this impression. In order to avoid exhibiting repeated behavior, we introduce several mechanisms in our behavior control architecture. These are: 1) artificial emotions and instincts; 2) probabilistic state machine for behavior generation; 3) reinforcement leaning for the probabilistic state machine; and 4) development through interactions.     As we described above, the instincts and emotions generate varied responses to stimuli. For example, if a robot is in its normal state, and there is a hand presentation in front of the robot, then the robot displays the response for the “give me a pawâ€? behavior. But if the robot is in an angry state, the robot refuses to display the “give me a pawâ€? behavior. Thus, the robot exhibits different responses even if the external stimuli are the same.     We use a state machine to generate behavior sequences. But we implemented the state machine with probabilistic transitions. Then varying behavior is generated even if the robot receives the same stimuli and the internal state.     The third mechanism is reinforcement learning. This is basically to modify the probabilities of the state machine. The modification is based on how the user interacts with AIBO. If AIBO receives a “reward,â€? it increases the probability of the transition. Therefore, it tends to exhibit the same behavior as it receives the reward.     The final method involves development. Because the behaviors are basically generated based using the probabilistic state machine described above, if the entire state machine is replaced, then the robot's behavior is a totally different one. We designed multiple state machines which are chosen depending on how a user interacts with the robot. For example, if the user often interacts with the robot and keeps the robot in a happy state as much as possible, then the state machine that exhibits “a good boyâ€? behaviors is chosen. If a user does not interact with a robot, then the state machine that exhibits “a bad boyâ€? behaviors is chosen. Thus, the robot has a development capability that enables it to develop a different robot persona for each user.     D.  Behavior Control Architecture    Â
Summarizing this section, Fig. 3 shows the behavior control architecture of AIBO. The architecture is based on the behavior-based architecture, where multiple behavior modules are implemented and selected according to situations. The behavior module is basically composed of probabilistic state machines, whose probabilities are modified through interactions so that a user can train the robot's behavioral tendencies. In addition, the state machine itself is replaced based on user interaction so that the robot's behaviors are affected by the interactions. The instincts and emotions are used for the regulation of the behaviors and realize different responses even when the same stimulus is presented. Finally, we implemented about 1000 motions, sounds, and blink patterns of a light-emitting diode (LED) to increase the complexity of behaviors.     All of these mechanisms contribute to increase the complexity of behaviors so that we can achieve the goal of a “lifelike� appearance.     Fig. 4 shows various postures and motions generated by AIBO. IV.  MARKET REACTIONS
    Fig. 5 shows ERS-111, ERS-210, ERS-311, ERS-312, and ERS-220.     We gathered information about AIBO owners from 2001 to 2002. These are owners of the ERS-110, 111, 210, 220, 311, and 312 models. The total number of owners for this statistics is about 4000. From these statistics it is possible to categorize the users into four groups, which are: 1) owners of ERS-110 or 111; 2) owners of ERS-210; 3) owners of ERS-311 or 312; and 4) owners of ERS-220. Since ERS-220 is not used as an interactive robot but remote-operated robot, in this paper we show the first three categories. Figs. 6–8 show some results. Precisely speaking, an owner of an AIBO may be different from the buyer of an AIBO. In this paper, we use “owner� as having the same meaning as “user.�     Fig. 6 shows the national ratio of AIBO owners. More than 80% of the owners are in Japan. This tendency is true for all of AIBO models. In general, it is said that the Japanese love an animal-like or human-like robot more than other any other country's people. At least for AIBO, the statistics confirm this as true. Many non-Japanese people think of AIBO as a toy, and they feel that it is expensive for a toy. Many non-Japanese people require some useful function for the robot if the cost is like that of AIBO.     Fig. 7 shows the gender ratio of AIBO owners in Japan. We could say the newer model is more often owned by females. It may be difficult to say by this data, however; from some observations of Web bulletin board systems (BBSs), females tend to prefer the model ERS-31X more than ERS-210. ERS-31X tends to be owned more by female users when compared to the other models.     Fig. 8 shows the age distribution of AIBO owners. This graph tells us that elderly users have become a far larger component when compared with the first model ERS-11X. The majority of owners of AIBO are between the ages of 30–40, and about half of the owners are between ages 30–50.     We have other statistical data about the attitudes of AIBO owners in 2002.     The major motivations to purchase AIBO include:
    How they play with AIBO is described as follows.
   Â
AIBO is quite a different product from ordinary home electronics products. The users feel attachment to AIBO so that they do not like to change the bodies, even if there is a damaged part. The statistics are as follows.
    Thus, most users are intensively involved in playing with AIBO in the beginning. The ratio goes down after awhile, but about one-third of the users continue to play with AIBO as part of their everyday life. These people may be attached to AIBO so that they feel AIBO is a pet. V.  SOME EXPERIMENTS USING AIBO
    This mental stability depends on how long the patient spends time with the animal. The relationship between the animal and the patient is not one-way but mutual, but unlike human–human relations, human–animal relationships do not cause “rival� competition.     Integrating all these facts, AAA/T calls forth healing effects, which can be categorized as:
    Compared with AAA/T, RAA/T has the following advantages.
    A.  Experiments in Pediatrics Hospital Ward     Yokoyama [3] made experiments of RAA/T using several robots in a hospital pediatrics ward. Their report can be summarized as follows.
    The experiments of RAA/T using AIBO were done for approximately one year. The first reactions of the children were to avoid AIBO's eyes, but to keep watching AIBO's behaviors. They all laugh, are confused and surprised at AIBO's reactions, and eventually get close to it. Typically, the first contact with AIBO is to pat its head, which makes AIBO blink its eyes and generate soft sounds and motions. The children recognize that their actions produce AIBO's reactions. Then, they start further interacting with AIBO, by calling its name, patting its back, and so on.     Some differences in reactions by children depending on their ages were found. The infant children tend to have no interest in AIBO. Younger children tend to regard AIBO as an animal. The older children tend to be interested in the mechanism of AIBO. However, it tends to be difficult for the older children to keep their interest in AIBO. They found a difference in the reactions between a girl and a boy. Compared with boys, girls tend to take good care of AIBO as if it is a real living thing,     In addition, they found that there were mutual interactions among the children involving eye contact and some conversations when AIBO was present in a group of the children.     Yokoyama also reports a precious case. A patient, age 10, had a problem in her left leg. There was no communication with others and there was no smile in her face. There was no physical cause of the symptom, so it was considered that the cause was mental. They treated her using RAA/T with AIBO. They took her to the room where the children were absorbed in playing with AIBO. She was also interested in AIBO and interacted with AIBO by patting its head and so on. Then, she started to interact with a girl in the room. In the next day, she spent the most of her time with the girl, with smiles and conversations. In parallel, her left leg problem disappeared, and finally she left the hospital.     Furthermore, Yokoyama made similar experiments with other robots. In summary, a robot producing too-simple motions or when aiming at only language communication without motion did not show good results. There are a couple of other robots with good results, which are NeCoRo and Robovie.     NeCoRo is a cat-type robot with soft hair like that of a real cat and smooth motion, but not walking. Children and elderly people like NeCoRo as much as AIBO. They hug or keep patting NeCoRo, which is not observed in the interaction with AIBO. NeCoRo gives more “stability� than “stimuli,� which is the reverse effect ratio of AIBO.     Robovie is a human-sized wheel-based robot with a human upper body. Unlike AIBO and NeCoRo, Robovie speaks a natural language (Japanese). Children start to make contact with the robot with much interest but tend to hesitate when actually contacting it. On the other hand, elderly people actively try to make contact with Robovie. Thus, the ability of verbal communication causes different interaction phenomena.     B.  Effect on Immune System     Suga et al. report on their experiments using AIBO for elderly people [4]. In these experiments, they evaluate the effect on the immune system of the elderly. They measure some secretions in their saliva. The total participants were 23 elderly people, whose average age was 79.87. Among them 13 elderly people, whose average age was 72.0, participated in three trials of AIBO interactions for 30 min every other day. Some applicants continued the experiments continuously for six months.     The levels of dementia of the patients are distributed. However, the all patients can stay at home. The patients first tend to hesitate interacting with AIBO. But after led by an experimenter, they start to pat the head of AIBO and call its name. Eventually, they spontaneously interact with AIBO by showing a pink ball, calling its name, and so on.     s-IgA is a useful secretion to measure immune system activity [14]. The s-IgA density significantly increased (from 100 to 175 ![[$\mu$]](http://mathfigs.ieeexplore.ieee.org/iel5/5/29676/1347460/1032415.gif){kind=small} g/ml) after
three trials. In addition, among the patients, the subjects who had participated
in the activity at least once and continued their interactions with AIBO for
two months improved their density of s-IgA from 170 to 300 g/ml.    In addition, they measured density of homovanillic acid (HVA), which is usually low for Alzheimer dementia patients. They did the experiment with 15 patients, where five are in the Alzheimer dementia group, five are in the senile dementia group, and five are in the normal group. The density of HVA after the first trial of Alzheimer dementia patients increased to the same level of the ordinary elderly.     From the experiments, they concluded that interaction with AIBO is effective to improve the immune system of the elderly. The effectiveness can especially be said to arise from activating their emotions rather than healing their emotions, so that AIBO can be used for rehabilitation for Alzheimer elderly patients.     C.  Online Discussion Analysis     There are many Web sites about AIBO. Most members of the AIBO Web sites own their AIBO. They discuss and exchange their information on what happens with their AIBO. Through the discussion, we can observe what they consider AIBOs to be. Friedman et al. [5] analyzed the discussion data to extract five overarching categories in members' postings about AIBO. These are:
    The analyzed data consisted of 3119 postings by 182 members over a three-month period. The results showed that many members affirmed (negated) that AIBO had:
    According to these results, although the members consider AIBO as an artifact (from 75% technological), they observe mental states in AIBO (from 60% mental states), communicate with AIBO (59% social rapport; in this main category, the subcategory of “communication� is 45%, emotional connection is 28%, and companionship is 26%), and about half of them consider AIBO as something lifelike. We think these ratios are not valid for general people, because these statistics are based solely on the AIBO owners' Web.     Note that the members tend not to attribute moral standing to AIBO (from 12% moral standing). This means that even if the owners consider that AIBO is lifelike, AIBO has mental states, and AIBO provides company with reciprocal communication, they seldom attribute moral standing to AIBO.     A group of researchers has also been investigating the psychological effects of robotic pets in the lives of children [6] and the elderly [7]. Unfortunately, there are no results available as of the writing of this paper. VI.  DISCUSSION VII.  CUSTOMIZATION AND OPEN-ENDED SYSTEMS VIII.  SUMMARY AND FUTURE WORK     The author would like to thank Dr. Yokoyama at National Defense Medical
Collage, Japan, who kindly showed the author some experimental results. The
author would also like to thank Dr. Suga at Aichi Prefectural College of Nursing
and Health, Japan, who allowed the author to describe their experiments on
the immune system. The author would also like to thank Dr. P. Kahn, University
of Washington, who also kindly introduced their paper describing elderly people's
relationship with AIBO. The author would also like to thank Mr. Amagai, President
of ERC, Sony, and Mr. Kawanami, a Manager of ERC, Sony, for their disclosing
marketing data.
| |||||||||||||||||||||||||||||||||||||||||||||||||||
Masahiro Fujita (Member, IEEE) was born in Kanagawa,
Japan, in 1959. He received the B.S. degree from Waseda University, Tokyo,
Japan, in 1981, and the M.S. degree from the University of California, Irvine,
in 1989.