Animals, even when domesticated, maintain a series of innate characteristics that humans perceive as wild. Human responses to animal behavior vary with many factors such as historical time and culture. Even after their domestication, humans perceive animals’ distinctive behavior patterns that are considered as being wild, even in pets. It is common understanding that these wild patterns are linked to the specific animal species but also that there is a common characteristic to wild behavior in general. In this paper, we present our work on coding such a quality, exemplified in an application to robot pets. We consider that an application of wild character in the behavior of the pet robot can improve its realism and make the human/robot social interaction more captivating. To this aim, we have designed and implemented a behavioral architecture to reproduce well-known behavioral patterns of domestic cats that might be considered wild and applied it to a commercial pet robot that then was used to assess human reaction to its behavior. We demonstrate the effectiveness of the proposed approach with experiments carried out with humans that found the robot cat “independent” and “unpredictable” to a significant degree, two distinctive features of the wild side of animals.
Hinweise
Esther Aguado, Claudio Rossi and Pablo Gil have contributed equally to this work.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
1 Introduction
In the novel by Philip K. Dick Do Androids Dream of Electric Sheep?, animals have almost completely disappeared from the world, and people have lost contact with wild life to the point that owning and caring for an animal is considered a symbol of high social status and a moral responsibility. Many people have to resort to artificial ones to show their social position.
We do not need to be in such a dark futuristic world to imagine artificial animals. For many years now, wild animals have disappeared from human settlements. City dwellers have lost any contact with the wild and enigmatic behavior of animals. This process has impoverished the human experience and has had an impact on its emotional, moral, and aesthetic development. The importance of the connection between humans and the natural environment is recognized by evidence from the fields of psychology and neuroscience.
Anzeige
According to Florian Mormann et al. [1], animal images mobilize brain resources to prepare them for entering our field of interaction. The amygdala detects that there is an animal that might affect us and organizes a response to it. Based on previous knowledge of the role of the right amygdala, Mormann considers that in our evolutionary past, animals had great behavioral relevance, the reason being their binary role as predators or prey. Both roles are fundamental for human existence and, therefore, have an important consequence in the responses that have been developed. This study is preceded by other studies in which it is shown that during the early evolution of vertebrates, “the right brain hemisphere has been specialized in dealing with unexpected and behaviorally relevant stimuli”.1 Considered from the point of view of human perception, it is possible therefore to consider that animals represent a form of behavior among others that can be termed wild, a term connected to Mormann’s conclusion that our brain has been shaped to respond to animals as prey or predators.
Also, the interest in reconnecting humans to the experience of nature is noticeable, as is manifested in the work of many designers, theoreticians, city planners, artists, architects, and engineers that in their respective fields have tried to design that connection throughout numerous theoretical studies and projects during the 19th, 20th and 21st centuries. It is less common the attempt by these disciplines to provide forms of re-connection with animals, although there are numerous studies about the health and psychological benefits of spending time with animals(see, e.g., [2‐4]).
The generation of human and animal-like movements for virtual characters and robots has been proven to create engaging human–machine interactions [5, 6]. Wildness has a great impact on humans’ response because they cannot fully predict the animals’ actions and reactions. It is important to distinguish between completely unpredictable behavior and the type of behavior that occurs when we consider that an animal is acting wildly. In wild behavior, we find that there must be a proportion of predictability and understandable behavior, a basis of animal character that is understood as logical and reasonable for us, and that creates a structure of what a specific animal is for us, the idea of a cat, a dog, or a panther. But then, in order to find wildness in an animal, there has to be sudden change of behavior into unpredictable and maybe unreasonable patterns of behavior, on a certain proportion that can vary from species to species. The different proportion of unpredictable behavior is connected with the degree of wildness that humans give to different animals. Humans are biased in their appreciation by many other circumstances, such as prior knowledge of that species, the morphology of the animal, its potential harm to humans, and other cultural values such as predisposition to feelings of disgust, fear, etc.
This paper aims to start a discussion on the question: Could we use biological traits to arise a certain judgment of the characterization of a bioinspired robot? To address this question, we present three main contributions to the bioinspired social robot research: (1) the use of biological behavior to trigger a certain perception in robot users, (2) focus on behavior to increase engagement in human-robot interactions, and (3) a proof-of-concept application by evaluating with 30 people the perception of wild behavior in human-robot interaction.
Anzeige
The implementation of wild behavior in a social robot poses several requirements for the system under study. First, we need to approach ethological studies, i.e., behavioral studies of wild animals to emulate its conduct in our bioinspired robot. Then, this behavior needs to be modeled to be reproduced by the machine. This research is framed in bioinspired robots, particularly ethorobotics. Ethorobotics is a multidisciplinary field that targets the implementation of biological behavior in robots. Most etho-robots are inspired by dogs because these animals evolved from savage species to domestic ones, adapting to human behavior in this transition [7]. Hence, works on social robots are mainly aimed at making them look and behave in a friendly way.
Here, we focus on a particular trait in social interaction, the unpredictability characteristic of wild animals.
This paper is organized as follows. Section 2 discusses related research on bioinspired robots, triggering emotions on them, and behavior modeling on ethorobotics. Section 3 presents the ethological studies we have based to determine what humans understand by wild behavior. Section 4 focuses on the computational modeling and implementation of these traits, paying special attention to ensure unpredictability. Section 5 evaluates the modeling and implementation of animal behavior through a human-robot interaction and discusses the experiments carried out.
2 Related Work
2.1 Triggering Emotions Through Social Robots
Most animal-like robots are used in care and assistance tasks. Numerous studies show the effectiveness of pets in improving the elderly or patients with dementia [8‐10]. For this reason, most of the work in ethorobotics corresponds to this purpose [11].
The AIBO robot, is one of the most widely used when conducting these studies [12, 13]. AIBO is a robot with an attractive appearance and is ready for social interaction. Typical interactions are based on canine behaviors, such as walking and hitting a ball, and on actions associated with toys, such as dancing to music.
Another example of a robot with toy appearance is the commercial robot PARO, designed specifically for therapy to reduce stress and increase sociability. Studies have shown the effectiveness of this robot in helping dementia [14] and promoting social interaction [15‐17].
The studies in [18] an [19], focus in building artificial emotional creatures with a cat-like appearance. In [18] a robot is presented with tactile, auditory, and postural responses to humans and their environment. In addition, an emotion-based model is implemented. This study also reflects that cat owners expect the robot to behave like this animal, so its response is always evaluated by a comparison of what a feline would do. Similarly, in [19], a robotic cat is presented as a “friendly machine” through emotion modeling. Thanks to this system, the robot can adapt to changes in the environment and develop a behavior that is aligned with the particular stimuli.
In the present work, we target a behavioral model that orchestrates the actions of an artificial cat to resemble the response of the live animal, focusing on its “wildness” and “unpredictability” rather than “friendliness”.
2.2 Behavior Modeling in Etho-Robotics
Translating the generation of emotions into orders interpretable by the robot is not an easy task. According to [20], the affective loop is an iterative process in which the user expresses his emotions through some corporal mobility, such as gestures. The system answers by generating an action according to this expression, which triggers a response in the user. When the person presents a state that is detected by the robot, this should generate an emotional behavior appropriate to the read stimulus. The robot then synthesizes the emotion, adapts it to the system, and expresses an answer, which produces a new iteration in the interaction loop.
However, these emotions induced by the robot are often not intentional. By nature, humans attribute intelligence to objects and animals [21]. For example, the fact that a robot moving forward suddenly changes direction and goes back can be interpreted as “fear”. For this reason, many researchers seek to provoke sensations from behavioral patterns executed by the robot.
The morphology of the robot plays a crucial role in how the person evaluates the interaction, and the appearance of the robot has a great influence on triggering emotions. In [22] it is shown how the purely mechanical morphology makes people treat the robot like a servant, with substantially less interaction and low expectations. This mechanical morphology is a parameter the designer must take into account when developing an animal-like robot that should trigger a certain emotion in the user.
In addition to morphology, behavior is crucial for triggering a certain feeling during social interaction. Behavior modeling is based on providing an emotional architecture to the robot. An example of emotion-centered behavior is [23], where four basic emotions are considered: happiness, sadness, fear, and anger. These emotions are modulated through reinforcement-learning techniques.
Examples of modeling and programming behavior using ethological rules can be found in [24, 25] and [26]. In these works, a robot with a canine appearance uses behavioral rules to adapt to particular situations. These rules drive the decision-making about which behavior should be applied or how the fusion between several behaviors occurs. They use an emotional model to trigger the best behavior according to the situation. For example, when the robot does not see his owner, its anxiety level increases, so the exploration task is triggered. This task selection is a linear combination of factors that determines the response.
As mentioned above, in this work, our aim is that the user identifies the robot behavior as “wild” after interaction. Therefore, the triggering events must be complex enough to resemble the proportion of predictable and unpredictable behavior in the real animal, and the movements must be organic enough to create this realistic illusion. To provide such complexity, we use a composition of state machines that select the response depending on previous actions, sensory perception, a characterization of the robot “personality”, -i.e., the probability to do a certain type of actions- and some degree of randomness.
We also encode a set of routines and responses that emulate a wild feline and define an architecture to orchestrate the responses depending on previous actions, sensory perception, and a grade of unpredictability.
3 Characterization of Wild Behavior
The concept of wildness is directly linked to an unfettered behavioral repertoire, extremely influenced by environmental factors. Further details on what wild behavior can be found in [27, 28] and [29]. [30] uses components -such as anxious, insecure, tense, stable, friendly, calm, fearful of people, curious, dominant, impulsive, independent, aggressive, etc.- to characterize the personality of five wild felines.
The behavioral characterization of animals requires a set of terms and adjectives that can be perceived to different degrees by each ethologist [31]. Tinbergen [32] presents four questions to explain animal behavior according to the environment in which it is developed: (i) Function, how behavior improves survival, (ii) Evolution, how behavior has evolved with the development of the species, (iii) Causality, which factors cause a particular behavior to be exhibited at a specific moment, (iv) Development, how behavior changes from youth to adulthood.
To develop our bioinspired wild robot, we focus on the third question, why a certain action is activated at a certain moment.
In [30] five wild felines are studied to make a taxonomy of their personality. Each species was found to have three personality factors: dominance, impulsiveness, and neuroticism.
Furthermore, in [33] the personality of the domestic cat is complete with extroversion and agreeableness to conform to the domestic aspects of the feline. All these characteristics together conform the Five Factors Model (FFM).
Nowadays, the FFM model is widely used in animal personality characterization (not only in cats) because it allows using standard terminology and coefficient modification for an accurate taxonomy. As the aim of this research is to resemble unpredictability and wildness, the factors addressed in our robot behavior are neuroticism, impulsiveness, and extroversion.
4 Modeling Wild Behavior
We use the concept of FFM to set parameters and encode unpredictable and wild behavior to reproduce wildness through a degree of unpredictability in the human-robot interaction.
4.1 Behavioral Actions and Interactions
The first step in reproducing animal behavior is to emulate its actions. We have replicated six well-known behavior patterns in felines [34]: exploration, hunting, playing, grooming, drawing attention, and knead, then experiment with the parameters so that the characteristics of the wilderness are perceived by the human participants in the study. Every pattern is divided into several actions. Figure 1 shows the four actions for the hunting pattern.
Fig. 1
Movement emulation of wild animals by our test robot corresponding with the hunting sequence in the robot. The hunt state executes the actions (i) approach, (ii) wait with a low body, (iii) rise up, and (iv) extend the front leg
×
We use a state machine to make a hierarchical structure of the robot response according to human interaction, adopting the six behavioral patterns mentioned above as the main states. Each state acts as a sub-state machine that regulates the different actions that comprise a pattern. For example, the “hunting” pattern is a state that acts as a state machine for the four actions shown in Fig. 1.
The main states are linked through transitions at three levels: a trigger, a guard condition, and transition actions. These elements will be discussed in Sect. 4.2. The sub-states (concrete actions) also constitute a state machine, but with more simple transitions; e.g, time or sensor readings.
Social interaction for emotion triggering in humans is obtained through sensors. Our robot is equipped with two VMA309 microphones to respond to audible stimuli (one on each side of the head), eight tactile points (two on the head, three on the upper back, and three on the lower back) connected to a CAP1188 sensor, and an HC-SR04 ultrasound sensor in the eyes to detect the presence of humans or objects.
The robot response to sensory inputs differs according to its state. This is controlled by the transitions on the state machine. Transitions’ syntax follows UML (Unified Modeling Language) style, and is composed of three parts: event[guard]/action. To execute a transition, the associated event must occur and the guard condition must be true. When the transition is activated, the associated action(s) are performed. Guard conditions encode the cat’s personality (see below) and determine which transition is actually taken with a probability threshold. Actions are postural movements that show the general mood (e.g., happy, aggressive, etc.).
In the state transitions, the triggering event allows us to encode the social response, whereas the guard condition preserves the individual personality giving unpredictability and wild appearance.
Figure 2 shows a portion of the state machine that we have used. For simplicity, we only represent the relationships from/to the “explore” state. However, every state has at least three incoming and three outgoing transitions, just as in reality, where the same sequence of actions can be triggered by different events.
For example, a trigger event from the explore state is a touch on the spine sensor. The transition associated to this event to the hunt state determines the cat’s personality, i.e. the guard condition.
Fig. 2
Details of the state machine related to the “explore” state and associated transitions. From the explore state, we can reach the play, hunt, wait, and attention states, and the explore state can be reached from the start, groom, and wait states. The syntax of the transitions is composed of three parts: event[guard]/action
×
4.2 Personality Encoding
According to the FFM presented in Sect. 3, felines are expressed through five dimensions: dominance, impulsiveness and neuroticism, extroversion, and agreeability. We use the term personality to encode these factors in the robot response.
Our robot personality is defined by four parameters: suspicion, related to neuroticism; sociability, as an extroversion trait; aggressiveness, related to dominance and fear, linked to low agreeableness. The fifth factor, impulsiveness, is encoded through a random variable in the behavior selector. These parameters are based on the personality characterization of wild felines from [33].
We normalize these qualitative factors through a numerical component of 0 to 1. Therefore, our robotic cat personality is a combination of different traits at different levels.
As mentioned above, the personality of the robot is expressed through the guard conditions associated with the transitions. If the triggering event occurs but the guard condition threshold is not exceeded, the new state cannot be reached. Therefore, the guard condition thresholds encode the personality of each particular individual. Table 1 shows an example of the probability of a change of the state matrix. Therefore, each personality trait has a probability of a state change associated according to the source and destination state. We combine each trait probability to get a probability matrix. The combination of probability traits is done by the law of total probability, as shown below.
Probability of action according to robot “personality”
Source/Dest
Start
Explore
Hunt
Play
Wait
Groom
Attention
Knead
Start
0.23
0.1
0.2
0.1
0.16
0.1
0.03
0.02
Explore
0.05
0.15
0.1
0.15
0.03
0.02
0.26
0.3
Hunt
0.02
0.08
0.3
0.05
0.23
0.28
0.03
0.02
Play
0.07
0.3
0.07
0.33
0.15
0.1
0.3
0.07
Wait
0.1
0.02
0.1
0.02
0.08
0.27
0.02
0.1
Groom
0.15
0.07
0.15
0.07
0.3
0.07
0.07
0.15
Attention
0.08
0.23
0.08
0.23
0.02
0.1
0.2
0.26
Knead
0.3
0.05
0
0.05
0.03
0.06
0.09
0.08
When a transition is triggered by sensory input, the probability of a change of state is compared to the change of state threshold. If the value is exceeded, the change of state occurs. If not, the robot remains in the previous state, but personality changes, varying the probability values shown in Table 1. Therefore, if the robot stays in the same state for several triggering events, the guard condition checked for reaching other states may increase or decrease, simulating how animal behavior changes during interactions with humans.
In summary, robot behavior is not hardcoded. There is a set of predefined actions structured in sequences (see Fig. 1); each sequence constitutes a state of the state machine. These states correspond to some well-known behavioral traits in felines. The exhibition of each sequence depends on the human-robot interaction, i.e., the sensori input, and the “personality” of the cat, i.e., the probability matrix that can evolve in time.
5 Experiments: Social Interaction Evaluation
In this section, we describe the robotic platform we have used and define the experiments carried out and their results.
5.1 Robotic Platform
For our experimental setup, we have used the commercial cat robot Nybble. This is an open source robot with a three-month kitten appearance from the Petoi project [35]. Nybble comes with a board to encode different moves of the robot. We have extended its computing capabilities with a Raspberry Pi 3 B+. The default sensory input in Nybble is ultrasound, which has been extended with eight tactile sensors and two microphones to allow for a richer interaction. Figure 3 shows the robot.
Fig. 3
Nybble robot with modifications for improved human interaction
×
5.2 Social Interaction Tests
The evaluation of social interaction must take into account the different types of user. We follow the guidelines for ethical studies and use a standard nomenclature according to [30]. Here, each adjective is directly related to an FFM factor. In addition, we use an open and closed questions questionnaire to reduce the influence on the answers. In open questions, the participant can use any adjective(s) to describe the robot, while in closed questions, they shall select from a list of predefined adjectives. The questionnaire used during the social interaction tests can be found in the Appendix A.
The tests described below constitute a first proof of concept on the effectiveness of the proposed approach.
We have performed 30 in-person tests with an interaction of 10 min per user. The experiments were carried out individually to avoid influence among the participants. Participants are of different groups of ages between 20 and 60 years, with an average of 31 years and a standard deviation of 11.4 years. 79% of the participants were used to interaction with animals, but only 10% of them with cats. We did not find significant differences among responses depending on interaction with cats or other animals. 82% of the participants saw the robot for the first time when they performed the test. The objective of these tests was to determine to what extent the behavior encoded in the cat robot can be classified as “wild”.
Robot categorization is done by selecting from a list of adjectives corresponding to the taxonomy presented in [31]. For comparison purposes, participants were also asked to evaluate the behavior and perception of robots from videos of BigDog, a robot from Boston Dynamics [36]. This robot has been chosen for its fierce appearance while being a loading robot. The objective was to evaluate to what extent the appearance of the embodiment is more important than the actions of the robot when triggering human judgments about its behavior.
In the first block of questions to define the robot, the most repeated are independent (33%), curious (25%) and affectionate (18%). In the closed answer, the most selected is sociable (70%), independent (40%), and friendly (15%). These factors are related to the extroversion factor of FFM according to the list of adjectives analyzed in [31]. Moreover, an important characterization of the robot is the term “independent”. This is related to the unpredictability we aimed to achieve in our wild behavior implementation, as users pretend to “domesticate” the behavior of the cat or at least to establish a pattern, but rather characterize its behavior as not doing what the subject expects. In Fig. 4a, these answers are represented. The size of the word represents repetition, so the larger adjectives are the most repeated. On a final note, we highlight the appearance of the adjective “free”, which is also associated with wildness.
Fig. 4
Word cloud of adjectives used to describe our interaction with our pet robot (a), and the perception of BigDog (b). The size is proportional to the repetition of the word
×
Compared to the characterization of our robot, the BigDog robot is perceived as robust and strong (66%). In the closed question, the adjectives most commonly used are aggressive (56%) and loyal (38%), as shown in Fig. 4b with these words in larger font. This suggests the importance of appearance in the perception of human behavior. Because our cat robot looks like a kitten, the perception of wild and fierce is difficult; however, we can resemble unpredictability but not fall into randomness, missing the feeling that we are interacting with a feline. When we directly ask if the users perceive the cat as unpredictable, 75% of the users agree (note also the relative sizes of the term “aggressive” in the two-word cloud) (Table 2).
Table 2
Most repeated answer for pet robot and BigDog
Adjective selection
Petoi cat (%)
Adjective selection
Big dog (%)
Independent
40
Aggressive
56
Sociable
70
Loyal
38
Friendly
15
Independent
17
Table 3
Summary of other interaction results
Sensor response
US response
Microphone response
Touch response
77%
50%
97%
Participants profile
Any pet interaction
Cat interaction
First time seeing this robot
79%
10%
82%
Improvements
Faster
More unpredictable
More fierce
37%
27%
35%
Once the wild behavior has been characterized, we asked about the quality of interaction to improve our testbed platform. According to the interaction block, 77% of the users found a correct response to the ultrasound sensor associated with the feeling that the robot knows when the human is near. 50% of the users have classified the microphone response as adequate, probably because the microphone sensibility (48-66 dB) sometimes cannot filter the sound emitted by the motors located near these devices. The best way of interaction is touch, which was found to be satisfactory by 97% of the users.
We also asked what they would improve to make the robot resemble more wildness. A 37% said faster and stronger movements, a 27% make it even more unpredictable, and 35% would change its appearance to a fierce big animal (Table 3).
6 Conclusions and Future Work
Current works on social robots are mainly aimed at making these look friendly. In our current research, we aim to expand social robots’ capabilities to better emulate nature, reproducing unpredictable behavior and wildness.
In this work, we focused on structuring the behavior and patterns in order to trigger a perception in the user. The tests performed constitute a proof-of-concept on the viability of reproducing wild behavior. Even if the number of tests we carried out is relatively small, we believe that the experiments carried out confirm that wildness can be artificially provided.
However, we observe some limitations that can be explained mainly by the appearance of the robot and the motion possibilities the platform adopted allows. The study is limited to body movement through posture and responses to manifest behavior. Other elements such as facial expression, spine flexibility, or even claw extension are crucial to feline behavior, particularly when the animal shows its most wild behavior. We think that not having those features had an impact on the perceived degree of wildness. As far as programming behavior is concerned, it will be important to study the possibility to adapt the robot’s personality through social interaction, i.e., to train it to behave according to the user’s preferences, just as it occurs with a pet that develops its conduct during socialization. In this case, the objective would be to increase the level of perceived wildness according to a specific user.
Future work will also be devoted to determining the relevance of the morphology of the robot to the appreciation of wildness. Also, the relevance of the parameters linked to the actions of the robot, such as speed, acceleration, trajectory, in its parts, relative to each other, and generally, in the appreciation of wildness.
Additionally, future work should extend the validation with more experimental data to a larger evaluation group. We also plan to extend this evaluation from an exploratory survey to a more formal and reliable study, using an ethics board, and measure the evaluation with elements such as quality and performance indicators from the European Social Survey program [37].
Even if this robot does not intend to substitute real animals, it is designed to generate experiences of wildness that are less present in urban societies. These human-robot interactions can provide compelling experiences to the public without exposing them to the risk of being close to a wild animal.
More importantly, this study is a first step into the codification of wild behavior to be used by robots, which will perform a mode of intelligence- animal inspired intelligence, that diverges from human inspired intelligence so present in current day applications. The application of such a kind of intelligence, embodied in robots, could have many applications, from the development of more realistic and engaging video game characters, toys, and in general artificial entities for entertainment, e.g., in thematic parks, defense applications in which a degree of unpredictability could have beneficial implications, and biomimetic robots that interact with other animals or humans in diverse circumstances, avoiding risks of animal suffering.
Acknowledgements
This research has received funding from the RoboCity2030-DIH-CM, Madrid Robotics Digital Innovation Hub, S2018/NMT-4331, funded by Programas de Actividades I+D en la Comunidad de Madrid and co-funded by Structural Funds of the EU.
Declarations
Ethical Approval
The manuscript has not been submitted to other journals for simultaneous consideration. The submitted work is original and has not been published elsewhere. No data, text or theories from others are presented as if they were the authors’ own. The authors have permission for the use of software and hardware used in this study. This research does not pose a threat to public health or national security.
Open AccessThis article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Die im Laufe eines Jahres in der „adhäsion“ veröffentlichten Marktübersichten helfen Anwendern verschiedenster Branchen, sich einen gezielten Überblick über Lieferantenangebote zu verschaffen.
