Social Transmission of Information through Virtual Robotic Agents

Owais Hamid

1 a

, Shruti Chandra

, Kerstin Dautenhahn

2,1

and Chrystopher Nehaniv

1,2

Systems Design Engineering, University of Waterloo, Waterloo, ON, Canada

Electrical and Computer Engineering, University of Waterloo, Waterloo, ON, Canada

Keywords:

Social Learning, Game Environment, Observational Conditioning, Stimulus Enhancement, Response Facili-

tation.

Abstract:

Social learning includes simple or complex social mechanisms that allow us to understand cooperation and

communication in animals, giving them better chances to survive for longer and thrive as a society. Speciﬁ-

cally, certain types of social learning such as observational conditioning and stimulus enhancement have been

investigated in the context of social information spread between primates. However, not many studies have

utilized such social learning mechanisms to study social learning between humans and artiﬁcial agents. In the

work described here, we seek to understand if and how simple social learning mechanisms can inﬂuence hu-

man participants using an online game platform with an immersive ﬁrst person experience built through Unity.

Speciﬁcally, we designed a study inspired by experiments in behavioural sciences to investigate whether and

to what extent, a robotic agent can inﬂuence human’s actions. The study compared two conditions in which

the robot showed body-language based emotions in a positive or negative manner that could enhance certain

stimuli for the human participants and inﬂuence their decision making. From this, we wanted to understand

whether these effects are socially learned by humans. Objective (position of player in-game) and Subjec-

tive (questionnaires) measures were recorded, and markers using the objective data suggest successful social

transmission of information. We believe this approach can make a novel contribution to the ﬁeld of Human

Interaction with Artiﬁcial Agents.

1 INTRODUCTION

Social learning is the facilitation of learning by ob-

servation or interaction with another individual or its

products (Hoppitt and Laland, 2013). While these in-

teractions have typically been studied in humans and

other social animals, certain recent literature has fo-

cused on social interaction among generalized agents

that may be artiﬁcial, including robots (Steels and Ka-

plan, 2000) (Hamid et al., 2020).

Several mechanisms play a role in the learning

processes that we see every day. Some of them have

been termed cognitively more complex, requiring a

greater amount of cognition such as imitation. Others

are more simple and need not require complex per-

ceptual and cognitive processes, but may involve a

simple redirection of attention, local enhancement be-

ing one good example. In between these two levels of

complexity lie a plethora of social transmission mech-

anisms with varying levels of cognitive complexity

required. Mechanisms that were deﬁned earlier in

https://orcid.org/0000-0002-8942-4621

the literature and remain popular include Stimulus-

Stimulus learning which includes Observational Con-

ditioning, the mechanism identiﬁed behind Pavlov’s

famous experiments. Emulation is another mecha-

nism that was identiﬁed earlier in (Hoppitt and La-

land, 2013). However, while we know that humans

can socially learn from other humans, they may also

learn socially from non-human agents. Humans learn

a variety of skills and behaviours through interaction

with others, cf. language learning through interac-

tion with others (Tomasello, 2009). Social signals

can be non-verbal, or can involve vocal, non-speech

signals. For example, vervet monkeys (Cercopithe-

cus aethiops) use vocal calls for anti-predator alarm

signals, even going so far as to classify which type of

predator is hunting them (Seyfarth et al., 1980). How-

ever, such specialized signalling only works within a

certain species, not across species.

So far, we discussed examples where communi-

cation happens within a species of primates. How-

ever, there do exist several instances of humans ex-

tracting or delivering information socially from or to

Hamid, O., Chandra, S., Dautenhahn, K. and Nehaniv, C.

Social Transmission of Information through Virtual Robotic Agents.

DOI: 10.5220/0010823000003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 3, pages 361-372

ISBN: 978-989-758-547-0; ISSN: 2184-433X

361

animals. In (Byrne and Tanner, 2006), this type of

‘imitation’ by an animal (a western lowland gorilla,

Gorilla g. gorilla), who learns to reproduce observed

actions, is considered simple priming where the ani-

mal puts together actions it already knows, and then

reinforces them by individual learning. The gorilla

in this experiment was able to ’imitate’ gestures that

humans showed her to the extent that naive coders

who were asked to check if such actions resembled

the original gestures by humans gave scores that con-

ﬁrmed socially learned behaviors. This particular

form of learning, previously called gestural imitation,

was termed in this study as Response Facilitation.

Complex cognitive processes can be involved in

socially learning from others, in particular in the area

of imitation learning. Large bodies of work have been

developed on Imitation and Learning by Demonstra-

tion between robots and humans, where robots are

being taught by human teachers. However, crucial

perceptual and cognitive challenges remain to be re-

solved, and some of them are discussed in (Nehaniv

and Dautenhahn, 2001). This serves as the inspiration

for utilizing simpler methods of social learning so that

robots and artiﬁcial agents are able to learn socially in

human-agent interaction.

2 BACKGROUND

A number of efforts have been made to classify so-

cial mechanisms of information transfer in primates,

and the deﬁnitions have changed over time. Some of

the older efforts were made by Galef (2013), Heyes

(1994) and Zentall (2001). The ones utilized here are

more recent versions built on previous work (Hoppitt

and Laland, 2013).

Several researchers have studied learning behav-

ior among animals, where a demonstrator directs an

observer’s

attention to the speciﬁc location of a food

object using simple attention mechanisms such as tap-

ping the item with a body part or partially biting and

chewing a strategically important part that needs to

be exploited initially so that food acquisition by a

‘naive’ observer is made easier (Previde and Poli,

1996; Terkel, 1996). For example, in wild greylag

geese, attention is drawn by tapping a certain loca-

tion. This same effect can be had if a human taps a

certain object in a speciﬁc place (Fritz et al., 2000).

We use the words demonstrator and observer through-

out the paper to refer to the agent ‘demonstrating’ a certain

behavior in which it is already experienced in, and the naive

agent receiving social information as an ‘observer’ respec-

tively.

In a classic study in humans, Gerull and Rapee

(2002) have documented that affective responses to

novel objects can be learned socially. Speciﬁcally, if

humans observe something novel, they usually asso-

ciate certain emotions to the newly introduced object

of interest when they observe someone else who has

more experience react to the novel object or situa-

tion in a certain way. In (Gerull and Rapee, 2002),

toddlers, when ﬁrst being introduced to novel toy

creatures, learned to associate fear and avoidance be-

havior towards the objects when they witnessed their

mothers reacting to the objects with fear, horror or

disgust and avoidance. Toddlers reacted signiﬁcantly

differently when a mother’s reaction was positive, not

showing any avoidance to the toy physically. Fa-

cial expressions were also very different in the two

cases, the children in the negative condition showing

expressions of clear disgust or horror, while those in

the positive condition treated the toy normally. This

shows that the information provided by an experi-

enced demonstrator (e.g. mother) is crucial to the per-

ception of any object that the observer (e.g. toddler),

hence creating a Stimulus-Stimulus pairing. Gerull

and Rapee also found signiﬁcant differences between

the perception and avoidance behavior of male tod-

dlers and female toddlers to the mother expressing a

positive or negative emotion. Male toddlers were less

avoidant towards the toy creature in the negative con-

dition than female toddlers. Furthermore, the impres-

sion of these emotions lasted for at least up to 10 min-

utes, making this one of the ﬁrst studies to understand

potentially how long the effects might last.

This speciﬁc experiment is of great interest to us

for a variety of reasons. New environments pose a

challenging problem when humans who are naive to

the environment ﬁrst encounter challenges. An expe-

rienced agent that understands the subtleties and prob-

lems of an environment can pass on information to

newcomers, gained by previously exploring this envi-

ronment. We therefore envision situations when em-

bodied artiﬁcial agents, quite speciﬁcally robots, in-

troduce humans to a new environment efﬁciently. For

instance, teaching humans about features and situa-

tions in the environment that are dangerous, and oth-

ers that are harmless, even if they may appear danger-

ous. Our study therefore aims to draw inspiration the

above-mentioned social learning study (Gerull and

Rapee, 2002), in the new context of fostering Agent

Interaction and Communication that can be beneﬁcial

to learning about a new environment that the human

encounters. Note that the study by Gerull and Rapee

is itself an extension of Cook and Mineka’s series

of experiments (the original experiments were con-

ducted from 1984-1993, but a comprehensive review

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

362

can be found in (Mineka and Cook, 2013)) which

showed observational conditioning of fear of snakes

among rhesus monkeys, using toy snakes.

2.1 Types of Social Learning

The process of Social Learning is rarely random or

without reason. In fact, social transmission of infor-

mation has been termed strategic. A very complete

list of strategies for social learning can be found in

(Hoppitt and Laland, 2013). Some of the most com-

mon reasons are “when” strategies such as copy when

uncertain. For instance, naive rats copy other more

experienced rats for their choice of diet (Galef, 1996).

Historically, most sub-types of social learning have

focused on imitation, with earlier work considering

all forms of animal based social learning to be an at-

tempt to reproduce human-like imitation (Hoppitt and

Laland, 2013). However, other scientists have catego-

rized the same behaviours as different types of social

learning. Classiﬁcation schemes for social learning

mechanisms are broadly categorised according to the

following: (1) context speciﬁcity (i.e. whether depen-

dent on location or stimulus); (2) whether the out-

come is sensitive to the demonstrator’s actions; (3)

whether the information being transferred is action

speciﬁc and; (4) if the learned behaviour is a novel

action sequence (Hoppitt and Laland, 2013, p. 99).

In our study, we are interested in understanding

if social information can be spread by an emotion-

ally intelligent robotic agent to a human, similar to

how mothers’ emotional reactions were transmitted

to their infants (Gerull and Rapee, 2002). In doing

so, we wish to establish a precedent on the effects

that social behavior expressed by non-human artiﬁcial

agents can have on humans.

In the kinds of observation based learning that

were discussed above, certain speciﬁc types of social

learning stand out and are described and deﬁned as

follows:

Response Facilitation (RF): is deﬁned as the “pres-

ence of a demonstrator performing an act [that] in-

creases the probability of an animal which saw it do

the same” (Byrne, 1994, p. 237). As mentioned in

Section 1, a good example is that when a demonstra-

tor performs such actions, the observers have a “prim-

ing” effect. Note, this effect may be transient, i.e. the

probability of the observers performing the same ac-

tions might reduce later in time.

Observational Conditioning: is one of the older

mechanisms that was discovered through classical de-

signs of social learning studies. The earliest stud-

ies in this category include Pavlovian conditioning

where an Unconditioned Response (UR) to a stim-

ulus for the demonstrator acts as an Unconditioned

Stimulus (US), which then becomes a Conditioned

Stimulus (CS) for the observer, and the observer will

then respond the same way. Heyes’ deﬁnition is

a little broader in that stimuli need not be condi-

tioned (Heyes, 1994). It is deﬁned as “a subset of

Stimulus-Stimulus learning where an observer ob-

serving a demonstrator is exposed to a relationship

between stimuli at t1 and the observer’s behavior

changes in a detectable manner at a later time t2”.

Finally, Stimulus Enhancement: as deﬁned by

Heyes (ibid.) occurs “when an observer observing

a stimulus at time t1 from a demonstrator and this

stimulus causes an observable change in the observer

at time t2”. Stimulus enhancement requires memory

and learning, since the observer has to retain informa-

tion about the object and how it was used, in order to

use the same object in a different context, or a differ-

ent object in a similar context(Giraldeau, 1997).

It should be noted that there exists plenty of

overlap between the above-mentioned types of social

transmission mechanisms. This is due to the reason

that a combination of factors might be at play when

we see certain types of imitation-like behaviors in an-

imals.

2.2 Social Learning with Artiﬁcial

Agents

Little research has been conducted on how artiﬁcial

agents interacting with humans might shape or in-

ﬂuence their experiences. This is speciﬁcally the

case for non-verbal social information spread not in-

cluding cognitively complex forms of imitation. A

certain amount of literature review on social learn-

ing in robotics, speciﬁcally multi-agent robotics was

conducted in (Hamid et al., 2020). Cakmak et al.’s

work is one of the few examples that can be found

of human-robot interaction with social mechanisms

simpler than imitation (Cakmak et al., 2010). It in-

vestigates how different social learning mechanisms

such as emulation, mimicking and stimulus enhance-

ment (three other simpler forms of social transmis-

sion) can be used to teach robots certain simple tasks.

Similarly, social learning was also utilized for ba-

sic language learning using the Aibo robot (Cakmak

et al., 2010). While no speciﬁc type of social trans-

mission mechanism was identiﬁed, social learning for

language acquisition through a human mediator is re-

alized through observational learning. Additionally,

supervised and unsupervised machine learning algo-

rithms were used for this task.

Social Transmission of Information through Virtual Robotic Agents

363

Figure 1: A decision tree describing which types of social transmission are at play depending on the effect of the stimulus and

the demonstrator’s actions, adapted from (Hoppitt and Laland, 2013). The social learning mechanisms are classiﬁed according

to different levels: if the effect of the stimulus is Context Speciﬁc, if it is dependent on the demonstrator’s actions, if the action

itself plays a large role in the effect, and if the effect depends on an action sequence that is novel. PI = Product Imitation, CI

= Contextual Imitation, ORSL = Observational R-S Learning, OC = Observational Conditioning, LE = Local Enhancement,

RF = Response Facilitation, SE = Stimulus Enhancement.

In our work, we investigated simple forms of so-

cial learning mechanisms that could be transmitted by

the artiﬁcial agents such as robots to humans. Addi-

tionally, we wanted to understand if robots are just as

effective as experienced demonstrators in conveying

social information to humans.

2.3 Using Virtual Environments as a

Stand-in for Real World

Experiments

Since the platform used for conducting this experi-

ment is entirely virtual, using a game environment, it

is worth asking whether this study can be emulated in

real world conditions, and how close the study gets

to evoking actual emotional responses from partici-

pants. The pertinent practical question being, can we

compare emotional arousal in games to experiencing

emotions in the real world?

While there are different opinions about the extent

of the human reaction to events in digital media (in-

cluding games and movies), (Reeves and Nass, 1996)

pointed out that the responses to media events are re-

markably similar to real world responses. While there

is some agreement that the immediate emotional re-

sponse is very similar, the longer term effects might

be more blunted for virtual events. (Cantor, 2009) as-

serts that previous experiences do play a strong role

in evoking emotional responses, especially from the

real world, however in the virtual world these are less

intense than the same happening in the real world.

(Cantor, 2009) and (Lynch and Martins, 2015) argue

that previous experiences may allow participants to

cope with fear based scenarios more effectively and

allow a more balanced reaction towards these events.

Given the current trend of artiﬁcial agents play-

ing increasingly bigger roles in human society, under-

standing how these agents may inﬂuence humans is an

important question to consider. We address this ques-

tion by designing a virtual scenario where a robotic

agent needs to guide a human who is new to the en-

vironment, assuming that these might be some of the

roles given to intelligent agents in the future.

3 RESEARCH QUESTIONS &

HYPOTHESES

Despite drawing inspiration from the above men-

tioned experiment that involves a mother transmitting

social information to a child about a novel stimulus

in the environment (Gerull and Rapee, 2002), several

factors have been changed to better suite the study

currently conducted. First, we focus on adults rather

than infants because our use case, i.e. familiarizing

individuals to new environments, suites adults. In the

case of grown adults, they do not necessarily react to

surprising stimuli in a new environment in a manner

similar to how toddlers or infants react. For example,

adults may have inherent biases from previous experi-

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

364

ences that might make them more or less susceptible

to either positive or negative actions. Second, adults

are already used to a large variety of different objects

and experiences, so it can be difﬁcult to create truly

novel stimuli that they previously have never encoun-

tered, making our task of replicating the experiment

difﬁcult. To counter this, we attempted to make the

animal novel and thus chose an alien animal that is

not common in general, does not exist in reality, and

resembles a combination of two or three animals. Ta-

ble 1 outlines the similarities and differences in our

approach as compared to (Gerull and Rapee, 2002).

In this study we are interested in understanding if

social information (in this case affective behaviour by

a virtual robot) can spread to humans who share the

environment (designed similar to a computer game)

with the virtual robot. We study the role of simple so-

cial transmission mechanisms such as Observational

Conditioning, Stimulus Enhancement and Response

Facilitation. Figure 1 portrays the decision making

process that allows us to narrow down these mecha-

nisms.

The methodology of the experiment is described

in detail in Section 4.1, however a functional descrip-

tion of the experiment is given here. The experimen-

tal design requires a novel stimulus in an environment

unfamiliar to the participants, with a virtual robotic

agent acting as a demonstrator. The game was de-

signed such that a novel stimulus would startle the

demonstrator (robotic agent) in a programmed fash-

ion, with the observer constrained to watch the ex-

change in a virtual forest environment. The stimu-

lus chosen had to be novel to some extent, therefore

an “Alien Animal” was chosen for this task. The

demonstrator reacts either positively

or negatively

to the animal, i.e. acting in either a friendly manner

or running away in terror, respectively. All the while,

the participant/observer watches this reaction of the

demonstrator towards the novel stimulus without be-

ing able to move. The study used a between-subjects

design, i.e. participants saw either the demonstrator’s

positive or negative reaction.

In this context, and with reference to the decision

tree shown in Figure 1, the effect is stimulus speciﬁc,

sensitive to the demonstrator’s actions, action spe-

ciﬁc, but not a novel action sequence

. Thus, we are

The actions taken by the virtual robotic agent for the

positive condition can be seen by clicking HERE

The actions taken by the virtual robotic agent for the

negative condition can be seen by clicking HERE . It is to be

noted that the alien animal’s actions were exactly the same

in both conditions. Only the demonstrator’s actions were

changed. Further, care was taken to show the same actions

by the animal in both conditions

We assume that adult participants are fully capable of

left with three possibilities to classify the social trans-

mission that might occur in this study, namely, ORSL

(Observational Response-Reinforcer Learning), RF

(Response Facilitation) + OC (Observational Condi-

tioning) or RF + SE (Stimulus Enhancement) which

is the red box, second from left in Figure 1. In or-

der to differentiate between these 4 types of social

transmission mechanisms, two factors are important;

whether the observer observes the demonstrator get

rewarded for its behavior, and whether this learn-

ing is S-S (Stimulus- Stimulus) or R-S (Response-

Reinforcer). From our earlier discussion

in Section

2, SE, OC and RF do not typically involve rewards

ORSL can therefore be ruled out. This narrows down

the possible social transmission mechanisms to either

RF + OC or RF + SE. Further information is provided

once the results are explained in Section 6.

We designed a between-subject study that aims to

investigate the effect of a robotic demonstrator on an

observer

in an unknown terrain, i.e. a virtual forest.

These characteristics include two affective behaviours

(positive and negative emotions expressed through

body movements) portrayed by a robotic agent when

encountering a novel creature. Note, in order to pro-

vide context for the demonstrator agent’s skills, it is

introduced as having prior knowledge of the environ-

ment, which is new to the observer. We formulated

our research questions as follows:

• RQ1: Does the perceived reaction (positive, neg-

ative) of the demonstrator (robot) affect the ob-

server’s (human participant’s) response? If so, to

what extent?

• RQ2: Does the participants’ gaming experience

or gender affect their responses towards the crea-

ture? If so, to what extent?

With respect to these research questions, we pro-

pose the following hypotheses:

• H1 for RQ1: Distance of the participant to the an-

imal in the positive condition should be lower than

in the negative condition. We expect participants

in the positive condition to have a more positive

attitude towards the alien animal. In (Gerull and

Rapee, 2002), children observe and infer the na-

ture of the relationship between their carer and the

recognizing the demonstrator’s actions. Hence, this effect

is not based on a novel action sequence.

A detailed version of these subtleties is discussed

in (Hoppitt and Laland, 2013), Section 4.2

There is some debate regarding usage of reward for

OC, see (Palameta and Lefebvre, 1985). However, OC com-

bined with RF may not necessarily involve rewards

We use the word observer and participant interchange-

ably since in this experiment, the observing agents are the

human study participants.

Social Transmission of Information through Virtual Robotic Agents

365

Table 1: Description of similarities and differences of the present study, as compared to (Gerull and Rapee, 2002).

Gerull and Rapee Current Study

Mean Age 17 months 30.6 years

Type of Social Transmission Observational Learning Observational Conditioning

Number of Participants

30 44

Gender of Observers

F-15, M-15 F-17, M-27

Demonstrator Mother (Human) Robotic Agent

Stimulus Snake/Spider toy Virtual Alien Animal

Measurement Subjective - Likert Scale Subjective and Objective

Number of times observer

is exposed to stimulus

3 times

1st at 1 min,

2nd at 2 min,

3rd at 10 min

2 times

1st at 1 min,

2nd at 7 min

novel creature by recognizing disgust or fear, and

replicating the same avoidance behavior towards

these creatures.

• H2 for RQ1: Participants in the positive condi-

tion should have a more positive and less violent

perception of the creature as compared to the neg-

ative condition.

• We do not present any hypotheses regarding RQ2

since the question for the effect of gender was

only included because Gerull and Rapee found an

effect, although this was for toddlers. We include

the effects of gaming in the RQ since we would

like to know if previous experience has a signiﬁ-

cantly different effect on a participant’s behavior

in-game.

4 METHODOLOGY

We used Amazon Mechanical Turk to carry out this

remote study. The virtual environment is set in an

unknown terrain (a forest). The game scene includes

a virtual robotic agent capable of two types of emo-

tional body expressions (positive, negative) portrayed

by the robotic-like agent when subjected to a novel

creature, with the robotic agent being depicted as

someone with experience in the environment through

the game narrative. For the participants, the goal of

the game is to collect spheres scattered around the

game world. The spheres were distributed around the

environment in such a way that it would not bias par-

ticipants towards either moving towards or away from

the alien animal. The entire participant experience

was divided into three parts:

• The pre-game Questionnaires: This includes

some basic demographic information, consent and

general information regarding the game. This was

administered using Qualtrics

(2021 Qualtrics).

• The game experience included three levels:

– Level 1: Familiarization phase where partici-

pants get used to the game interface using their

keyboard and mouse on a PC/Laptop

– Level 2: Participants explore the area, look-

ing for in-game rewards. The event with social

transmission where the robot reacts either neg-

atively or positively to the alien animal happens

then. Next, the animal is removed from the

game (in accordance with the methodology fol-

lowed in (Gerull and Rapee, 2002)). This level

then continues for another 6 minutes, to make

sure there is no immediate emotional carry-over

to the next phase.

– Level 3: The participant starts the game at a

position where they observe the alien animal

directly. The creature remains at the same po-

sition and does not move in space, it is sim-

ply looking around. The participants can now

explore the area in the vicinity of the animal,

as well as moving further away, picking up re-

sources. This level lasts 7 minutes before the

game is bought to a close.

• The post-game Questionnaires: A set of custom-

made questions ask about participants’ experi-

ences during the game, notably their perception

of the alien animal, and including a few questions

as attention-checks. This questionnaire was also

hosted on Qualtrics.

This study was approved by the Ofﬁce of Research

Ethics at University of Waterloo.

4.1 Game Development

The game was designed with the Unity 2019 editor.

Participants were asked to enter their ID, and their

in-game positions were recorded once every second.

The game is hosted through the Web Graphics Library

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

366

(WebGL) format on Github pages. This allows the

game to be hosted remotely and to be playable on a

web browser without the user having to install any

plugins, and multiple participants can play simultane-

ously. In-game data is stored on an external online

DataBase based on a MongoDB server on the cloud.

4.2 Questionnaires

Two standardized and two custom-made question-

naires were included in the study and administered

through the University of Waterloo’s Qualtrics sys-

tem, along with the accompanying consent form at

the beginning, an explanation of the tasks, instruc-

tions to play the game and a web page explaining the

exact purpose of the game at the end of the experi-

ment. Custom-made pre-game and post-game ques-

tionnaires were completed by participants. Attention

check questions were used to address some of the is-

sues regarding Amazon Mechanical Turk studies such

as lack of attention or misrepresentation (MacInnis

et al., 2020) (e.g. Have you encountered the creature?

How does the creature look like?). Other questions

concerned participants’ reaction to and perception to-

wards the alien animal in relation to what they ob-

served from the virtual robot demonstrator (e.g. How

did you perceive the behaviour of the robot?).

4.3 Participants

A total of 49 successful participants were recruited

on Amazon Mechanical Turk. Data integrity was im-

portant since there was some packet data loss through

the three-step process with respect to the participants’

position. In order to enhance the quality of the data

collection, participants were recruited with respect to

speciﬁc metrics (>96% completion rate for at least 50

previous tasks completed).

The participants were equally distributed accord-

ing to the two experimental conditions. However, due

to some participants failing the attention checks, the

study ﬁnally had 27 participants in the negative con-

dition and 22 in the positive condition. Of the 27 in

the negative condition, 3 were excluded due to packet

data loss regarding their location in-game. A further

2 were excluded due to low quality of submissions.

A low quality submission was deﬁned as the partici-

pant either standing at the same place for a long time

(over 3 minutes of the total 6 minutes in the ﬁrst level,

hence showing inactivity), or answering the question-

naire by clicking on the same options for multiple

questions. This left us with 22 participants for each

condition, in total 44 participants, 28 males and 16

females, 19 to 55 years of age, with the mean age

being 30.64, S.D. of 8.9 years, 32 gamers and 12 non-

gamers.

4.4 Statistical Analyses

To understand the effect of social learning mecha-

nisms on the participants, we collected the data in the

form of objective and subjective measures. The ob-

jective data collected during the study in the second

phase includes mean and absolute distances

from the

participant to the alien creature, the frequency of re-

turns of the participant (moving back towards) to the

creature, and the time duration that the participants

took to their ﬁrst return to the creature. Since there is

no way to directly measure (objectively) how a partic-

ipant feels regarding the animal, we believe distance

of the participant to the animal in-game represents a

faithful marker regarding how wary they feel towards

the animal. A ‘return’ is deﬁned as the participant

walking to a position where they can clearly observe

the animal from a distance of 40 game distance units.

40 is chosen as the cut-off since the landscape con-

sists of hills near the position of the creature where

a participant might be behind one of them and there-

fore not directly have an animal in their line of sight

despite close proximity. These distances are calcu-

lated for each participant in relation to the creature at

different time points: 15 sec; 30 sec; 45 sec; 60 sec

and 100 sec. Because the alien animal was not a real

threat to the participants’ in-game character and al-

ways stood in one place, participants became curious

about the animal after about 100s and started wan-

dering closer to the animal. Therefore, the important

period was the ﬁrst 100s of the participants observing

the animal and deciding whether to move away or ex-

plore around the animal. We explored different time

stamps from 15 to 100, but from 60s up until 100s,

no difference in results is found. Table 2 describes

the normality tests that were performed to determine

whether a non-parametric or parametric test was re-

quired to test for signiﬁcance differences between the

positive and negative conditions. Shapiro-Wilk tests

were performed to understand normality of the objec-

tive data. This test checks whether a distribution is

non-normal, i.e. if p<0.05, the data is not normal. If

the data is indeed non-normal, the Mann-Whitney U

test was performed, and the independent-samples t-

test on the data that were normally distributed.

The subjective data includes a question regard-

ing the participant’s perception of the behaviour of

For mean distance, we averaged the distance of a par-

ticipant to the animal up to a certain time interval, for ex-

ample, 30 seconds. The absolute distance is the distance to

the animal at that point in time, for instance, at 30 seconds.

Social Transmission of Information through Virtual Robotic Agents

367

(a)

(b)

(c)

Figure 2: a A screenshot of the game showing the robot introducing itself to the participant. A game world map in a circle can

be seen on the top right. bThe Alien Animal that was present in the game world. c An overview of the world. The dirt paths

can be observed crisscrossing the landscape, while towards the bottom left, an overhead view of the alien animal can be seen.

the creature. The question that the participants were

asked was: How did you perceive the behaviour of

the creature?. Two example responses were: ”I per-

ceived the behaviour as curiosity” and ”It seemed vi-

olent”. The responses of all the participants were

classiﬁed into three categories of perceived threat:

Mild, High and None. Two independent coders not

involved in the study rated the responses of the par-

ticipants according to the three categories. Cohen’s κ

was run to determine if there was agreement between

the rater’s judgement on whether the participants per-

ceived high, mild or no threat from the alien animal.

There was good agreement between the two judge-

ments, κ = .726 (95% CI, .556 to .896), p ¡ .001.

One other pertinent question asked in the post-

game questionnaire was about the participants’ reac-

tions when they ﬁrst saw the creature. This is in ref-

erence to Level 2 (see Section 4) when the partici-

pants ﬁrst saw the creature and the robot’s reaction to

it. The responses were one of [Run Away, Approach,

Neither, Other] and this data was classiﬁed as Nomi-

nal, and therefore the Chi-Square test was used with

z-tests for post-hoc analysis.

5 RESULTS

Mean Distance between Participants and Alien

Animal: A Mann-Whitney U test was run to deter-

mine if there were differences in mean distance score

between positive and negative conditions. Mean dis-

tance score at time point 30 sec, was statistically sig-

niﬁcantly higher in the negative condition (Mdn =

57.58) than in the positive condition (Mdn = 48.40),

U = 152, z = -2.113, p = .035, using an exact sam-

pling distribution for U (2-tailed). Similarly, at time

points (for Mean distance) 45 sec and 60 sec, a Welch

t-test was run to determine if there were differences in

distance score between the positive and negative con-

dition. The mean distance for the participants from

the creature in the negative condition was higher (at

45 sec: M = 76.07, SD = 36.11; at 60 sec: M = 88.91,

SD = 42.43 ) than the participants in the positive con-

dition (at 45 sec: M = 54.49, SD = 19.00; at 60 sec:

M = 62.64, SD = 21.12), a statistically signiﬁcant dif-

ference was found for both time points 45 sec, M =

21.58, 95% CI [-39.31, -3.85], t(31.79) = -2.481, p =

.019, d= -.52 (Moderate effect) and 60 sec M = 26.27,

95% CI [-46.89, -5.65], t(30.80) = -2.6, p = .014, d=

-.55 (Moderate effect).

Absolute Distance between Participants and Ani-

mal: A Mann-Whitney U test was run to determine if

there were differences in absolute distance score be-

tween positive and negative conditions. The absolute

distance scores at time point 15 sec were statistically

signiﬁcantly higher in the negative condition (Mdn =

56.36) than in the positive condition (Mdn = 46.75),

U = 153.50, z = -2.078, p = .037, using an exact

sampling distribution for U (2-tailed). The absolute

distance scores at time point 30 sec were statistically

signiﬁcantly higher in the negative condition (Mdn =

70.08) than in the positive condition (Mdn = 51.06),

U = 175=4.00, z = -1.596, p = .013, using an exact

sampling distribution for U (2-tailed).

Perceptions of Participants towards the Behavior

of the Animal: With respect to the ﬁrst impression of

participants towards the creature, a chi-square test of

homogeneity was run, with an adequate sample size

established according to Cochran (Cochran, 1954).

The two multinomial probability distributions were

equal in the population, χ

(3) = 8.581, p = .035. Par-

ticipants in the negative condition were more likely

to respond with “Run Away” (n = 19, 86.4% versus

n = 12, 54.5%). Post hoc analysis involved pairwise

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

368

Table 2: Mean and Absolute distances of participants to the animal at 15, 30, 45, 60 and 100s. For Absolute distance at 60s,

where the positive was normal and the negative non-normal, the Mann-Whitney was used, since both distributions need to be

normal for the T-test. Here N- refers to Non-Normal, and N+ to Normal.

15s 30s 45s 60s 100s

Mean

Positive p<0.05, N- p<0.05, N- p>0.05, N+ p>0.05, N+ p>0.05, N+

Negative p<0.05, N- p<0.05, N- p>0.05, N+ p>0.05, N+ p>0.05, N+

Absolute

Positive p<0.05, N- p<0.05, N- p>0.05, N+ p>0.05, N+ p>0.05, N+

Negative p<0.05, N- p<0.05, N- p>0.05, N+ p<0.05, N- p>0.05, N+

(a) (b)

Figure 3: (a) The mean distance of participants to the alien animal up until 30s, 45s and 60s, respectively, showed a signiﬁcant

difference between positive and negative conditions. (b) The absolute distance of participants to the animal up until 15s and

30s showed a signiﬁcant difference between positive and negative conditions.

comparisons using multiple z-tests of two proportions

with a Bonferroni correction. Statistical signiﬁcance

was accepted at p < .0125. With the Bonferroni Cor-

rections, there were no statistically signiﬁcant differ-

ences in any of the cases. While the ‘Run Away’ re-

sponse for Positive vs Negative Condition is statis-

tically signiﬁcant without the Bonferroni correction

(χ

(1)=.021), with the correction this is still greater

than .0125.

A Mann-Whitney U test was run to determine if

there were differences in subjective rating scores be-

tween positive and negative conditions for the ques-

tion “How did you perceive the behaviour of the crea-

ture?”. No signiﬁcant results were found for either of

the raters’ scores between Positive and Negative con-

ditions.

Gaming Experience: Mann-Whitney U tests were

run to determine if there were differences in the dis-

tances between gamers in the positive and negative

conditions. No statistically signiﬁcant differences

were found for either Mean or Absolute distances.

This was extended to Mann Whitney U-tests between

gamers and non-gamers without regard for the con-

dition they came from. No signiﬁcant results were

found for this test either.

A Mann-Whitney U test was also run to determine

if there were differences in subjective ratings score

between Gamers in the two conditions. No signiﬁ-

cant results were found for either of the raters’ scores

between the two conditions. Further, subjective rat-

ing scores between gamers and non-gamers was also

non-signiﬁcant.

Effect of Gender: The Mann-Whitney U tests were

run to determine if there were differences in the dis-

tances according to gender. None of these data points

for either Mean or Absolute distance were found to be

statistically signiﬁcantly.

Further, the Mann-Whitney U test to determine if

there were differences in subjective rating scores be-

tween the self-reported Male and Female genders (no

other genders were reported) found no signiﬁcant re-

sults for either of the raters’ scores between Male and

Female participants.

Frequency of Return to Animal: An analysis was

conducted to understand whether participants en-

gaged in searching the area around the animal for re-

wards and returning to it. As before, the deﬁnition of

‘return’ was given as approach of participant to within

40 distance units, i.e. within observable distance.

A Mann-Whitney U test was run to determine if

there were differences in the number of times partici-

pants return to the animal in the Positive and Negative

Conditions. The statistic for frequency of returns was

statistically signiﬁcantly higher in the negative condi-

tion (Mdn = 1, Mean Rank=25.73) than in the positive

condition (Mdn = 2, Mean Rank=19.27), U = 171, z =

-1.722, p = .045, using an exact sampling distribution

for U (1-tailed).

Social Transmission of Information through Virtual Robotic Agents

369

6 ANALYSIS & DISCUSSION

It is essential to interpret the results detailed in Sec-

tion 5 within the context of the deﬁnitions provided

in Section 2 and understand if the results support our

stated hypotheses within the deﬁnitions laid out ear-

lier.

Objective Measures for Positive & Negative Con-

ditions: A closer look at the histograms in Fig-

ure 4 provides an interesting trend where we observe

the distance of participants in positive condition (blue

bars) to be less spread out and closer to the animal

(higher towards the left side of the histograms), which

represents a lower distance to the alien animal. This

is visible more clearly for 30s, 45s and 60s in Row

1. The differences in the mean distances can be seen

more clearly in Figure 3a that shows signiﬁcant dif-

ferences between the two conditions.

A similar trend emerges when we compare abso-

lute distances between the two conditions. Histogram

bars in Row 2 of Figure 4 show that participants in

the positive condition are in higher numbers at shorter

distances from the animal, i.e. there are more partic-

ipants closer to the animal in the positive condition

for at least 15s and 30s. This trend disappears as time

stretches further. This means they tend to be closer to

the animal during the earlier stages than the partici-

pants in the negative condition. We hypothesize that

this is because participants start exploring the area

more randomly in search of rewards as time goes on.

The trend of signiﬁcant differences between partici-

pants in the two conditions is shown in Figure 3b.

One other point of interest is the result from the

frequency of number of returns, as presented in Sec-

tion 4.4. This frequency was calculated to provide an

understanding of how comfortable participants felt in

searching for rewards closer to the alien animal. As-

suming participants from the positive condition return

to the animal more often due to having been exposed

to a prior positive relationship, the 1-tailed test is sig-

niﬁcant. This conﬁrms Hypothesis 1 for Research

Question 1, i.e. the distance of the participant to the

animal in general is signiﬁcantly lower in the positive

condition than the negative.

Subjective Measures: The subjective data that was

collected with respect to independent coders’ ratings

of the threat level that participants expressed towards

the animal was non-signiﬁcant. While the results

suggest that there is an effect of the demonstrator’s

(robot) reaction on the participant, these results are

not signiﬁcant. We hypothesize that this is due to hes-

itance among participants to admit fear in general, and

they may not agree that they felt frightened even if

their in-game behavior shows this. This means that

Hypothesis 2 for Research Question 1 cannot be ver-

iﬁed. We consider subjective data to be supporting

data, and secondary to the objective data.

Gender and Gaming Experience: A close look at

the relevant Sections regarding previous gaming ex-

perience and effect of Gender in Section 4.4 show

us that neither gender nor prior gaming experience

made a signiﬁcant difference in participants’ attitudes

towards the alien animal. Since no differences can

be found between gamers and non-gamers for either

the subjective or objective data, our data suggests that

gamers and non-gamers have a similar experience in

our study. Furthermore, gamers in the positive and

negative conditions also do not seem to have had a

very different experience. This is suggested by nei-

ther the objective nor subjective data for the positive

or negative conditions being signiﬁcantly different.

General: Taking the three types of objective data

measures together suggests that participants in the

positive condition tend to stay closer to the animal,

and the only difference between the games played by

participants in the two conditions is the type of reac-

tion to the demonstrator.

With this in mind, referring back to the litera-

ture in social transmission among humans and other

animals, Heyes’ deﬁnition of Stimulus Enhancement

(Heyes, 1994), where an observer observing a stimu-

lus at time t1 causes an observable change in the ob-

server’s behavior at time t2, is satisﬁed. Further, it

can be argued that the observer participants are ex-

posed to a stimulus relationship (fear or positive re-

actions towards a stimulus, i.e. an alien animal) at t1,

which then leads to a similar manner being adopted

by the observers at t2, thus ﬁtting the deﬁnition of

Observational Conditioning as well. To provide sup-

port that Response Facilitation has happened, we have

to show that the probability of an observer doing the

exact same thing that it saw the demonstrator doing

must increase. Two problems can arise while trying

to conform to this deﬁnition. First, we cannot cal-

culate the probability of adoption of the same action

because we must ﬁrst deﬁne what the ‘same’ action

is. Second, we must demonstrate such a probability

increases, which we cannot do since the repertoire of

actions that the participant can take through the game

is very limited. Hence, it is difﬁcult to provide sup-

port for RF through our online experiment, and so this

theoretical issue remains inconclusive and has to be

investigated in future research.

Of particular importance to theories of fear ac-

quisition, negative stimuli pairing created a negative

impression that lasted over 6 minutes, which was the

time given between Level 2 and Level 3. This means

that the effect lasts at least 6 minutes and does not

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

370

Figure 4: Row 1: The mean distance of participants to the alien animal up until 30s, 45s and 60s, respectively, showed a

signiﬁcant difference between positive and negative conditions. Row 2: The absolute distance of participants to the animal up

until 15s and 30s showed a signiﬁcant difference between positive and negative conditions.

simply disappear after the initial appearance of the

alien animal.

As a comparison to the earlier study by (Gerull

and Rapee, 2002), this study suggests the effect of

Observational Conditioning and Stimulus Enhance-

ment to last at least 6 minutes instead of 10 minutes

in the earlier paper. The sample size for this study

includes adults instead of children, is more numerous

(44 versus 30), and is virtual in nature with an artiﬁ-

cial agent as a demonstrator.

7 LIMITATIONS AND FUTURE

WORK

The study is limited in nature for two reasons. First,

it is very difﬁcult to ﬁnd other studies in human-

human social transmission that can be replicated for

Human Agent Interaction where the agent is artiﬁ-

cial. Therefore, the scope is limited to suggesting

that such types of social transmission are detectable

and quantiﬁable. Second, while original plans to con-

duct this study were in person, the complete halt of in-

person activities during COVID-19 forced this study

to be conducted online. We would like to note that

as is presented in Section 2.3, since virtual media and

real world experiences are to some extent comparable,

there is a good likelihood that the results produced in

this study can be reproduced using real world scenar-

ios. Therefore, these two limitations become scope

for future work, i.e. the goal would be to conduct in-

person experiments with real robots for human-robot

interaction in order to verify the results gained in this

remote study, and also to expand the studies to further

incorporate other types of social transmission mecha-

nisms. Certain other limitations in this study include

unequal sample sizes regarding gender and gaming

experience, which again was due to limitations of us-

ing crowd-sourcing methods, however, taken care of

by the chosen statistical tests.

While two experimental conditions were studied

here, it would also be desirable to compare these two

conditions with a neutral condition. While we drew

inspiration from (Gerull and Rapee, 2002) who com-

pared only the positive and negative conditions, others

(Mineka and Cook, 2013) utilize the approach with

a neutral condition as well. Finally, variations of

the virtual environment, tasks, and the nature of the

demonstrator (e.g. whether a robotic or human-like

or animal-like agent) could be studied further. All of

these factors make for exciting possibilities to con-

duct future work.

Social Transmission of Information through Virtual Robotic Agents

371

8 CONCLUSION

An online study was designed to understand if sim-

pler forms of social learning (i.e. simpler than imita-

tion) can be observed between artiﬁcial agents and hu-

man participants being present in a virtual, game-like

experimental environment. Results from objective

and subjective data collected during the online game,

which was carefully designed to perform stimulus

pairing, point towards successful social transmission

between a robotic virtual agent and human partici-

pants of information utilizing a mixture of methods,

as identiﬁed in the literature as Observational Con-

ditioning, Stimulus Enhancement and possibly Re-

sponse Facilitation. The study closely emulates work

done previously in the form of mother-child interac-

tion, (Gerull and Rapee, 2002), and to some extent,

human-animal interaction (Mineka and Cook, 2013).

Neither gender nor previous gaming experience seem

to play any signiﬁcant role in the efﬁcacy of social

transmission of information in our study.

The study, to the best of our knowledge, is a novel

approach in the ﬁeld of Human Interaction with Arti-

ﬁcial Agents, inspired by experiments in behavioural

sciences. Further studies with in-person participation

and real robots would be beneﬁcial, once such re-

search is possible, to verify and extend the results.

REFERENCES

Byrne, R. W. (1994). The evolution of intelligence. In

Behavior and Evolution, pages 223–265. Cambridge

University Press, New York, NY, US.

Byrne, R. W. and Tanner, J. E. (2006). Gestural imitation

by a gorilla: evidence and nature of the capacity. In-

ternational Journal of Psychology and Psychological

Therapy, 6(2):215–231.

Cakmak, M., DePalma, N., Arriaga, R. I., and Thomaz,

A. L. (2010). Exploiting social partners in robot learn-

ing. Autonomous Robots, 29(3):309–329.

Cantor, J. (2009). Fright reactions to mass media. In Media

Effects, pages 303–319. Routledge.

Cochran, W. G. (1954). Some methods for strengthening

the common χ

tests. Biometrics, 10(4):417–451.

Fritz, J., Bisenberger, A., and Kotrschal, K. (2000). Stim-

ulus enhancement in greylag geese: socially medi-

ated learning of an operant task. Animal Behaviour,

59(6):1119–1125.

Galef, Jr., B. G. (1996). Social enhancement of food prefer-

ences in norway rats: a brief review. In Heyes, C. M.

and B. G. Galef, J., editors, Social Learning in Ani-

mals: The Roots of Culture, pages 49–64. Academic

Press.

Galef, Jr., B. G. (2013). Imitation in animals: History, def-

inition, and interpretation of data from the psycholog-

ical laboratory. In Social learning, pages 15–40. Psy-

chology Press.

Gerull, F. C. and Rapee, R. M. (2002). Mother knows best:

effects of maternal modelling on the acquisition of

fear and avoidance behaviour in toddlers. Behaviour

Research and Therapy, 40(3):279–287.

Giraldeau, L.-A. (1997). The ecology of information use. In

Krebs, J. R. and Davies, N. B., editors, Behavioural

Ecology: An Evolutionary Approach, pages 42–68.

Blackwell Oxford, UK, 4 edition.

Hamid, O., Dautenhahn, K., and Nehaniv, C. L. (2020). En-

gineering social learning mechanisms for minimalistic

multi-agent robots. In 2020 3rd International Con-

ference on Control and Robots (ICCR), pages 90–99.

IEEE.

Heyes, C. M. (1994). Social learning in animals: categories

and mechanisms. Biological Reviews, 69(2):207–231.

Hoppitt, W. and Laland, K. N. (2013). Social learning:

an introduction to mechanisms, methods, and models.

Princeton University Press.

Lynch, T. and Martins, N. (2015). Nothing to fear? an anal-

ysis of college students’ fear experiences with video

games. Journal of Broadcasting & Electronic Media,

59(2):298–317.

MacInnis, C. C., Boss, H. C., and Bourdage, J. S. (2020).

More evidence of participant misrepresentation on

mturk and investigating who misrepresents. Person-

ality and Individual Differences, 152:109603.

Mineka, S. and Cook, M. (2013). Social learning and the

acquisition of snake fear in monkeys. In Social learn-

ing, pages 63–86. Psychology Press.

Nehaniv, C. L. and Dautenhahn, K. (2001). Like me?-

measures of correspondence and imitation. Cybernet-

ics & Systems, 32(1-2):11–51.

Palameta, B. and Lefebvre, L. (1985). The social transmis-

sion of a food-ﬁnding technique in pigeons: what is

learned? Animal Behaviour, 33(3):892–896.

Previde, E. P. and Poli, M. D. (1996). Social learning in

the golden hamster (mesocricetus auratus). Journal

of Comparative Psychology, 110(2):203.

Reeves, B. and Nass, C. (1996). The media equation: How

people treat computers, television, and new media like

real people. Cambridge University Press.

Seyfarth, R. M., Cheney, D. L., and Marler, P. (1980). Mon-

key responses to three different alarm calls: evidence

of predator classiﬁcation and semantic communica-

tion. Science, 210(4471):801–803.

Steels, L. and Kaplan, F. (2000). Aibo’s ﬁrst words: The

social learning of language and meaning. Evolution of

Communication, 4(1):3–32.

Terkel, J. (1996). Cultural transmission of feeding behavior

in the black rat (rattus rattus). In Heyes, C. M. and

B. G. Galef, J., editors, Social Learning in Animals:

The Roots of Culture, pages 17–47. Academic Press.

Tomasello, M. (2009). The Cultural Origins of Human Cog-

nition. Harvard University Press.

Zentall, T. R. (2001). Imitation in animals: evidence, func-

tion, and mechanisms. Cybernetics & Systems, 32(1-

2):53–96.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

372