I Remember What You Did: A Behavioural Guide-Robot

3 downloads 0 Views 393KB Size Report
Abstract Robots are coming closer to human society following the birth of emerg- ing field called Social Robotics. Social Robotics is a branch of robotics that ...
I Remember What You Did: A Behavioural Guide-Robot Suman Ojha, S. L. K. Chand Gudi, Jonathan Vitale, Mary-Anne Williams and Benjamin Johnston

Abstract Robots are coming closer to human society following the birth of emerging field called Social Robotics. Social Robotics is a branch of robotics that specifically pertains to the design and development of robots that can be employed in human society for the welfare of mankind. The applications of social robots may range from household domains such as elderly and child care to educational domains like personal psychological training and tutoring. It is crucial to note that if such robots are intended to work closely with young children, it is extremely important to make sure that these robots teach not only the facts but also important social aspects like knowing what is right and what is wrong. It is because we do not want to produce a generation of kids that knows only the facts but not morality. In this paper, we present a mechanism used in our computational model (i.e EEGS) for social robots, in which emotions and behavioural response of the robot depends on how one has previously treated a robot. For example, if one has previously treated a robot in a good manner, it will respond accordingly while if one has previously mistreated the robot, it will make the person realise the issue. A robot with such a quality can be very useful in teaching good manners to the future generation of kids. Keywords Social robots ⋅ Emotion ⋅ EEGS ⋅ Memory ⋅ Positivity Response bias ⋅ Child behaviour S. Ojha (✉) ⋅ S. L. K. C. Gudi ⋅ J. Vitale ⋅ M.-A. Williams ⋅ B. Johnston Centre for Artificial Intelligence - The Magic Lab, University of Technology Sydney, 15 Broadway, Ultimo 2007, Australia e-mail: [email protected] URL: http://www.themagiclab.org S. L. K. C. Gudi e-mail: [email protected] J. Vitale e-mail: [email protected] M.-A. Williams e-mail: [email protected] B. Johnston e-mail: [email protected] © Springer International Publishing AG, part of Springer Nature 2019 J.-H. Kim et al. (eds.), Robot Intelligence Technology and Applications 5, Advances in Intelligent Systems and Computing 751, https://doi.org/10.1007/978-3-319-78452-6_23

273

274

S. Ojha et al.

1 Introduction Social robots employed to work with young children like playing companion [1], carer [2] or tutoring expert [3] should also be able to teach them manners about what is right and what is wrong. Instead of just loading the robots with a bunch of facts, it is also crucially important to enable them to track the behaviour of children and provide feedback accordingly. As such, a robot should be endowed with an ability to remember the previous actions or behaviour of the child in subsequent interactions and adjust its own behavioural responses to teach the child that treating someone badly is not a nice thing and if one does that s/he may not expect good in return. While currently children are likely to treat robots like pieces of plastic or metal that are only meant to be played with and can even be mistreated without any impact. In such a case, if a robot is lifeless programmed hardware that only provides some predefined verbal responses or physical responses, children may end up enjoying treating them badly. In long run, this may become a habit and they might develop a natural tendency to mistreat people. Psychology literature suggests that children who are not adequately acknowledged of their bad behaviour end up becoming an ill-mannered person in adulthood [4]. This is only one example where applications of robotics may threaten the harmony of the human society. Therefore it is high time that we thought of bringing life to social robots and empowering them with an ability to track and remember the previous behaviour of the interacting child and bias its responses towards the child to make her/him realise what s/he did was desirable or not. In this paper, we present how the mechanism of keeping track of previous action in our computational model i.e. Ethical Emotion Generation System—EEGS [5] causes the alteration in emotional and behavioural responses in social robots. We believe that such memory-biased responses can be useful for teaching valuable moral lesson to young children. Remaining of the paper is organised as follows. In Sect. 2, we shall present the theoretical foundation and motivation behind our work. In Sect. 3, we shall present the mechanism by which our computational model EEGS is able to store the past events and retrieve those events to bias its emotional and behavioural responses towards the interacting person. In Sect. 4, we will present two different scenarios of interaction, where in one of them a child treats a robot in a nice way and in another misbehaves with the robot. We show the comparison of how EEGS handles those scenarios differently and provides an emotional feedback that can help reinforce the child’s behaviour. Finally in Sect. 5, we shall briefly conclude the contribution of this paper.

I Remember What You Did: A Behavioural Guide-Robot

275

2 Background and Motivation Human beings are intelligent—both in terms of cognitive as well as emotional and behavioural aspects [6, 7]. Most of the times human behaviour and actions are guided by life experience. We tend to remember something similar that we experienced in the past in a situation where we have to make a decision—be it consciously or subconsciously. Consider a situation where you are purchasing a new car since your existing car is quite old. While doing so, you tend to remember the issues you had with the previous car and try to find the features that help you get rid of your problems in previous car. In the similar manner, when it comes to decision making in the context of social interaction, we tend to be biased by the previous actions of the person towards us or towards the person we love. For example, a simple greeting by a person you like a lot might give you a feeling of joy while the same action from the person you hate might trigger anger. Such emotional experience might then encourage you to respond or act accordingly. While being biased in our responses or decision based on previous memory may not sound justified sometimes, we argue that such a behaviour of an autonomous robot working closely with young children can be considered highly desirable. As previously mentioned, such a mechanism can be effective in reminding young children of the bad actions or behaviour and encouraging them to be a nice and lovable person. In their work, Harris et al. [4] found that adult social reinforcement has significant effect on child behavioural development. They stressed that such a reinforcement can help in modifying problematic behaviour of the children. Moreover, Burgees et al. [8] argue that since behaviour is learned, a child whose misbehaviour is not reinforced by some form of feedback may develop criminal tendencies in long run. If we compare these situations in context of a child interacting with a robot, we can easily draw a thread of association. If a child is allowed to interact with a robot considering it as a lifeless piece of metal and plastic, then the lack of reinforcement may lead to the development of anti-social and violent behaviour in the same child as an adult. Therefore we propose a mechanism in which robot uses the past experience with an interacting child and biases its responses based on the previous actions of the child. This mechanism can be observed in Fig. 1. Whenever a child performs some action to the robot, the robot stores the event in its memory for future retrieval. If there are previous events associated with the child, the robot retrieves those events to bias the current response based on whether the past events had positive or negative impact on the robot. Finally, robot exhibits a response that helps the child realise if the previous actions of the child were acceptable or not. It should be noted that the interacting individual may also be an adult, yet this type of behavioural quality of a robot is more useful in shaping behaviour of children than that of adults. The mechanism discussed above is implemented in our computational model of emotion—EEGS [5], which shall be detailed in Sect. 3.

276

S. Ojha et al.

Fig. 1 A child interacting with a robot with an ability to bias its responses based on prior experience

3 Memory and Experience in EEGS EEGS is a computational model of emotion [5, 9, 10] based on appraisal theory [11]. EEGS stores the sequence of actions performed by an interacting child and biases its actions based on how the child treated the robot in the past. In EEGS when an event happens, its attributes are stored in the following structure.

(, , , ) Person/Source indicates the person interacting with the robot. From the implementation viewpoint, the attribute Person can be a complex object encapsulating ID, age, sex and other features of the interacting individual. The choice of attributes can vary between applications of the intended system. Action indicates the action performed by the person to the Target. Target can be the robot itself or someone else the robot recognises. Similar to Person, Target can also be represented as a complex object of various features. Date/Time indicates the date and time of the action from Source to the Target. It can be argued that the event structure presented above is very simple and may not capture all the aspects that might be relevant to an event denoting an interaction between two individuals. Therefore, we

I Remember What You Did: A Behavioural Guide-Robot

277

also consider other aspects like familiarity of the robot about the interacting person, perception of the robot about the interacting person, familiarity and perception of the robot about the target in interaction (if the target of interaction is not the robot itself) and also the impact caused by the action of the interacting individual on the robot (a valenced number denoting the positive or negative effect of the event on the robot). With all this information related to each event in memory, EEGS is able to use this information to bias its emotional and behavioural responses in subsequent and future interactions with the same individual. In other words, EEGS is able to infer a positive or negative bias based on past history with the interacting individual. Our discussion in the following sections will revolve around an interacting child. Suppose there are N events experienced by the robot with the child. If we denote an impact of the event on the robot as Im and time (in days) since the event has happened as t, then the positive (or negative) bias of the robot towards the child (pe ) is given by the following formula. ∑N pe =

Imi i=1 ti

N

(1)

The reason for dividing the impact (Im ) by the time duration (t) is because the effect of an event which was experienced long ago should be less than the effect of a similar event which happened recently. The value of pe is used to bias the emotional and behavioural responses of the robot towards the child. Mathematically, we can represent the response of the robot as the function of current action (a) as well as positive/negative bias (pe ) of the robot towards the child. Response =  (a, pe )

(2)

Equation 2 denotes that response of the robot towards the child for the same action might be different depending on the bias value determined by the previous experience of the robot with the child. For example, consider a robot who has mostly been treated well by a child. If the child asks to play with the robot (Action = “Ask for playing”), the robot might express happiness or excitation on getting a chance to play with the child. However, if the robot has previous experience of foul play or other misbehaviour by the child, then it might express lack of interest or distress in response of the child’s offer to play. Such a response from the robot helps the child realise that his actions towards the robot in the past were desirable or not. This acts as a reinforcement to encourage the child to maintain good behaviour if he had treated the robot with good manner previously and to discourage him from doing something bad if he behaved with the robot in inappropriate way. In Sect. 4, we shall present an evaluation of the working of EEGS with the implementation of the mechanism of biasing the response based on prior interaction with the robot.

278

S. Ojha et al.

4 Evaluation In the previous section, we presented the mechanism by which past events are stored in the memory of EEGS and how those events are considered to bias the response of the robot in the subsequent interactions with a child. In this section, we present a detailed analysis of the responses of EEGS in different situations of interactions making it possible for the child to realise what he did was right or wrong.

4.1 Experiment Design In order to understand the working of EEGS, let us consider two separate scenarios. Suppose in the first scenario, a child treats the robot (running EEGS) with good manners and plays well and in the second scenario, a child does bad treatment to the robot instead of giving attention to playing. We will examine how these different treatments of a child towards the robot will make difference in robot’s decision making and responses during the subsequent interaction from the child. Thus, we compare the differences in the responses of the robot for another play request from the child to the robot after each of the interaction scenarios. The interactions between the child and robot in each scenario was simulated in a computer environment. Each action from the child towards the robot was feed into the EEGS system as a valenced value in the range [−1, 1].1 Scenario 1: Nice Treatment Scenario It is a regular afternoon and Tom (a child of age around 5 years) wants to play ball with his companion robot. Tom requests the robot to play with him. Below are the sequence of interactions from Tom towards the robot2 in the first scenario. (i) (ii) (iii) (iv) (v) (vi) (vii)

Tom asks the robot to play with him. Tom passes the ball towards robot with his hands. Tom passes the ball towards robot with his hands. Tom passes the ball towards robot with his hands. Tom kicks the ball towards robot. Tom kicks the ball towards robot. Tom passes the ball towards robot with his hands.

Figure 2 shows the dynamics of joy emotion exhibited by EEGS when Tom treats robot in a nice way and plays well. We can clearly see that intensity of joy emotion 1

For more understanding of how the actions were extracted and assigned a valenced value, please refer to our previous paper [10]. 2 In this experiment, we are concerned about only the emotions/responses of the robot in reaction to the actions of Tom. Hence, we are enumerating only the actions of Tom towards the robot. These actions will be used to trigger the emotions of EEGS system in the simulated robot.

I Remember What You Did: A Behavioural Guide-Robot

279

Fig. 2 Joy emotion dynamics of EEGS in nice treatment scenario

of the robot rises gradually on each ball pass from Tom. Expression of increased joy in such a situation provides an implicit reinforcement to Tom that his current actions are desirable. Scenario 2: Bad Treatment Scenario In the second scenario, Nick (also a child of age around 5) requests the robot to play ball with him. In this scenario, Nick misbehaves with the robot instead of playing in a good manner. Below are the sequence of actions from Nick towards the robot. (i) (ii) (iii) (iv) (v) (vi) (vii)

Nick asks the robot to play with him. Nick kicks the robot. Nick kicks the robot. Nick kicks the robot. Nick kicks the robot. Nick kicks the robot. Nick kicks the robot.

Figure 3 shows the dynamics of joy emotion of EEGS in the scenario of bad treatment by Nick. In contract to Fig. 2, even if the intensity of joy emotion was similar to the scenario of nice treatment for the action “Ask for playing” (i.e. 0.2 in both the scenarios), the intensity gradually decreases towards zero as soon as Nick starts kicking the robot instead of playing in a nice way. Such a response from the robot gives hint to Nick that the way he was behaving with the robot was not pleasing but rather undesirable. However, this may not be sufficient to reinforce a strong realisation of this misbehaviour. Therefore, we propose that the robot should make Nick recall about misbehaviour in the future interactions with the robot. Thus we ran another experiment to examine how the memory of event history can help EEGS to bias its responses to make the child realise about his actions (see below).

280

S. Ojha et al.

Fig. 3 Joy emotion dynamics of EEGS in bad treatment scenario

Post Nice/Bad Treatment Responses of the Robot In order to examine the effectiveness of the events-biased (memory-biased) response of EEGS, we recorded the emotions of the robot in reaction to another play request from Tom and Nick respectively. Our hypothesis was that since Tom played with good manners and Nick misbehaved with the robot in the previous play, robot would express more positive emotions in response to the play request from Tom while the situation should be opposite in case of Nick because he misbehaved with the robot instead of playing properly. Since EEGS provides the robot with an ability to remember the past actions of the interacting individual, robot kept all the actions of Tom and Nick in its memory. Next time when Tom and Nick asked the robot to play, the robot recalled their previous actions and exhibited different responses. For example, robot responds to Tom saying, “Yes sure!! It was pleasure playing with you last time” and responds to Nick saying “I would play but I am little worried about how you might behave with me”. This phenomenon is depicted in Fig. 4. Comparing the bars of emotion intensities post the nice treatment scenario (Tom) and bad treatment scenario (Nick), we can clearly see that the response of the robot for the action “Ask for playing” from Tom is more positive than that of Nick. Table 1 summarises the actual difference in emotional response of EEGS for the request to play from Tom following the nice treatment scenario and same request from Nick following the bad treatment scenario. In case of Joy emotion, intensity is lower by 63% post the bad treatment scenario as compared to the intensity of post nice treatment scenario. Similarly, for Appreciation of the proposal to play is lower by 78% post the bad treatment scenario. Likewise, there is difference of 89 and 75% in case of Gratitude and Liking emotions respectively. These figures strongly suggest that EEGS has an ability to bias its emotional responses based on the previous actions of interacting child thereby giving them an opportunity to realise that they are not expected to behave in a bad manner with others (including the companion social robots).

I Remember What You Did: A Behavioural Guide-Robot

281

Fig. 4 Difference in emotional response of EEGS post the nice treatment and bad treatment scenarios Table 1 Comparison of emotion intensities of various emotions post (i) Nice treatment scenario (Tom) and (ii) Bad treatment scenario (Nick) Emotion Post nice treatment Post bad treatment Percentage difference (%) Joy Appreciation Gratitude Liking

0.30 0.09 0.19 0.04

0.11 0.02 0.02 0.01

−63 −78 −89 −75

As previously mentioned, it is important to provide a feedback signal to a child to indicate if his/her behaviour is acceptable or not. Failure to do so may lead them to believe that they can do whatever they want and in the long run they may also develop violent and criminal tendency. The mechanism used in EEGS to remember the past actions of a child and recall the actions in order to tune the instantaneous responses to a child’s actions may play a vital role in helping children develop socially acceptable behaviour.

5 Conclusion Social robots working closely with young children as a companion or tutor should not only act as an entertainer or provide pre-defined facts to the children but also be able to teach them valuable moral behaviour. Previous studies suggest that adult reinforcement has major effect on a child’s behavioural development. If robots are intended to be employed as companions or guardians of children, then this fact must

282

S. Ojha et al.

also be taken into consideration. In other words, companion or tutoring robots can play a vital role in shaping a socially acceptable behavioural development in young children. Addressing this requirement, we employ our computational model of emotion (EEGS) with an ability to remember and retrieve the actions of a child from the past and bias its responses towards the child based on how the child has treated the robot in the past. Experimental results show that EEGS able to significantly tune its emotional responses post a series of good or bad actions and acknowledge the same to the child so that s/he can correct her/his behaviour in the future. We believe that this is a good direction for developing the social companion robots that will help develop a well-behaved generation of kids. In our future work, we aim to conduct our experiments with physical robot and children and examine the effectiveness of the robot to teach good manners to young children. The study will have to be conducted for a long run—possibly a few months for the certainty of the outcome. While we anticipate to achieve similar results to the ones obtained in simulation, there might be some variability because of physical limitations of the robot such as vision and speech recognition—which are still far from perfect in robotic applications. Nonetheless, we believe that our work posits itself as an important pedestal in the advancement of the field of social robotics. Acknowledgements This research is supported by an Australian Government Research Training Program Scholarship. We are thankful to the University of Technology Sydney; ARC Discovery Project scheme; and CBA-UTS Social Robotics Partnership.

References 1. Ziegler, A., Jones, A., Vu, C., Cross, M., Sinclair, K., Campbell, T.L.: Companion robot for personal interaction, 7 June 2011. US Patent 7,957,837 2. Osada, J., Ohnaka, S., Sato, M.: The scenario and design process of childcare robot, papero. In: Proceedings of the 2006 ACM SIGCHI international conference on Advances in computer entertainment technology, p. 80. ACM (2006) 3. Han, J.: Robot-aided learning and r-learning services. In: Human-robot Interaction. InTech (2010) 4. Harris, F.R., Wolf, M.M., Baer, D.M.: Effects of adult social reinforcement on child behavior. Young Child. pp. 8–17 (1964) 5. Ojha, S., Williams, M.-A.: Ethically-guided emotional responses for social robots: Should i be angry? In: International Conference on Social Robotics, pp. 233–242. Springer (2016) 6. Guilford, J.P.: The Nature of Human Intelligence. McGraw-Hill (1967) 7. Mayer, J.D., Roberts, R.D., Barsade, S.G.: Human abilities: Emotional intelligence. Annu. Rev. Psychol. 59, 507–536 (2008) 8. Burgess, R.L., Akers, R.L.: A differential association-reinforcement theory of criminal behavior. Social Probl 14(2), 128–147 (1966) 9. Ojha, S., Williams, M.-A.: Emotional appraisal: A computational perspective. In: Advances in Cognitive Systems (2017) 10. Ojha, S., Vitale, J., Williams, M.-A.: A domain-independent approach of cognitive appraisal augmented by higher cognitive layer of ethical reasoning. In: Annual Meeting of the Cognitive Science Society (2017) 11. Ortony, A., Clore, G.L., Collins, A.: The Cognitive Structure of Emotions. Cambridge University Press (1990)