Socially Acceptable Behaviour for Robots Approaching Humans using an

Adaptable Personal Space

Kasper Camillus Jeppesen, Leon Bodenhagen and Norbert Kr

uger

Maersk Mc-Kinney Moller Institute, University of Southern Denmark, Campusvej 55, 5230 Odense, Denmark

Keywords:

Social Robotics, Service Robotics, Human Robot Interaction, Personal Space, Acceptability.

Abstract:

In this paper a new adaptable model of the personal space is proposed. This model takes into account the

position of a persons hands aiming to facilitate interactions with the human while maintaining an appropriate

social distance during the approach. This personal space model has been used as a cost function in the path

planning algorithm Transition-based Rapidly exploring Random Trees. It allows users to inﬂuence the robot’s

generated approach at the planning stage by varying their body and hand positions. Results from an online sur-

vey, where participants were shown different simulated approach behaviours, indicate that the model performs

well when it comes to distance regulation and how close the robot comes during the approach. An interesting

discovery from the survey is that the maintenance of eye contact, i.e. the robot keeping oriented towards the

person during the approach, was positively associated with both the closeness of the robot during movement

and judging the intentions of the robot for unaware users.

1 INTRODUCTION

As personal care robots become more commonplace

and people start sharing their environment with such

robots, certain expectations for the robot’s behaviour

arise. For example, the robot should try to follow sim-

ilar social conventions to the ones that humans are

conforming to. Few path planners create paths that

resemble human-like motion behaviour or are subject

to the same rules and constraints. Therefore, adapting

navigation strategies to conform to social conventions

and be as non-obtrusive as possible is important for

improving the perception and acceptance of personal

care robots.

This paper investigates how social conventions,

tendencies and expectations can be modelled and used

as cost functions for a path planning algorithm. In

particular, it proposes a new adaptable model of the

personal space, which describes distance regulations

between individuals, and uses this model for plan-

ning approach trajectories. It builds on the concept

of proxemics, see Figure 1, a term and ﬁeld of study

coined by Hall (1966) and in particular the idea of

the personal space. Each space or region emanating

from the person relates to the comfortable distance

that a certain type of social interaction should take

place (Hall, 1966). In this paper the deﬁnition of the

personal space model is extended to depend on the

Figure 1: Spaces associated with a person.

position of a persons hands, such that it allows for

intrusions into the personal space based on hand po-

sition. The hypothesis is that an adaptable personal

space will allow a path planning algorithm to create

approach strategies that can be tailored to the indi-

vidual’s preferences. The robot used for this research

is the Care-O-Bot (Graf et al., 2009), a service robot

with an omnidirectional mobile base with laser range

scanners, a 7-DOF manipulator with gripper, a tray

and a sensor head with a stereo camera and a Kinect-

like depth sensor. Figure 2 shows the robot interacting

with a person using its tray and manipulator.

116

Jeppesen K., Bodenhagen L. and KrÃijger N.

Socially Acceptable Behaviour for Robots Approaching Humans using an Adaptable Personal Space.

DOI: 10.5220/0006193801160121

In Proceedings of the 12th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2017), pages 116-121

ISBN: 978-989-758-229-5

Figure 2: The Care-O-Bot, a personal care robot.

2 RELATED WORK

The ﬁeld of robotics research that focuses on human-

robot interaction and collaboration is one of the more

recent and interdisciplinary areas.

Rios-Martinez et al. (2015) discuss a domain

called Social Signal Processing. They estimate that

over 60% of communication between two people

come from nonverbal communication, i.e. body lan-

guage and social cues. It discusses the connection be-

tween social cues (e.g. hand gestures and posture)

and social signals (e.g. emotion, personality, status)

and the importance of this connection in future re-

search of human-robot interactions. The survey also

looks at proxemics and various spaces such as com-

mon models of the personal space, which are mostly

static and non-adaptable. Interaction spaces, related

to groups of people, affordance or activity and spaces

related to objects are discussed and it is apparent that

modelling these spaces efﬁciently is important for so-

cially acceptable robot navigation. The possibility of

a dynamic personal space model is brieﬂy mentioned,

and in our paper the concept of an adaptable personal

space is explored.

Dautenhahn et al. (2006) carried out human-robot

interaction trials to determine the preferred approach

direction and other deﬁning characteristics of a robot

approach. Results based on a live trial showed that

the most preferred direction was to the right or left

hand side of the person and the least preferred direc-

tion was from the front. The approaches were also

rated in terms of practicality (seen in relation to the

trial environment) and comfort, and the frontal ap-

proach was again the least preferred or lowest rated.

In our work we found the same trend of people prefer-

ring non-frontal or direct approaches when it comes to

judging the closeness of the robot. The paper also dis-

cusses the idea of combining safety, visibility and hid-

den zone criteria that together seek to model the cost

map of an environment. The cost is modelled based

on e.g. whether the person can get the robot into his

ﬁeld of view by moving just his eyes or if he needs to

turn his head. The cost function also tries to penalize

the robot for making surprising appearances, such as

when coming from a hidden zone that the person can-

not see and into the persons ﬁeld of view. Instead, the

robot should seek to enter the persons ﬁeld of view at

a comfortable distance so there is enough time for the

person to react and be aware of the robot. In our work

we use a cost function which is based on the position

of a persons body and hands. We also vary the ori-

entation of the robot so the robot either looks in the

direction of travel or tries to maintain a sort of eye-

contact by looking at the person.

Kirby et al. (2009) implement a navigation frame-

work where human social conventions such as per-

sonal space and tending to one side of hallways are

represented as constraints on a robot’s navigation.

The following constraints were identiﬁed as impor-

tant for social behaviour in hallway situations; Min-

imize travelled distance, obstacle avoidance, person

avoidance with personal space and passing on the

right hand side, default velocity where the robot tries

to keep a constant velocity and inertia where the robot

should try to keep moving straight as much as pos-

sible. Each of these constraints are weighted and

combined linearly and used as the objective function

in a modiﬁed A

∗

planner. The method was success-

ful in generating paths that resemble human-like be-

haviour when moving down a hallway; passing on-

coming trafﬁc on the right and cutting across the hall-

way based on how far oncoming trafﬁc is. In our work

we utilize a sampling based method for path planning.

Our cost function is also based on a personal space

model, but our model is adaptable instead of constant.

Woods et al. (2006) investigate differences, be-

tween live and video based trials, in responses and

preferences for different characteristics of robot ap-

proaches. Participants, both in the live and video

based trials, were questioned about their preferred

approach direction, stopping distance and approach

speed. There were a high levels of correspondence

(85% resp. 87%) for the least preferred approach di-

rection (frontal) and the ratings of the robot’s speed.

Moderate to high (60-80%) agreement was found be-

tween video and live trials for most preferred ap-

proach direction and the robot’s stopping distance

from the subject. Almost all subjects (93%) preferred

the live trials over the videos. These results are inter-

esting, since in our paper we utilize videos of simu-

lated approaches and online questionnaires for evalu-

ation. This evaluation method is a lot faster, allows

for more experimentation and does not have the same

Socially Acceptable Behaviour for Robots Approaching Humans using an Adaptable Personal Space

117

)(x

)

Figure 3: Left: Non-adaptable model. Right: Proposed adaptable model. (x

, y

) and (x

, y

) are coordinates for the body

center and left hand respectively. White corresponds to high values, purple to low values.

safety requirements as live trials, which are time con-

suming to set up and carry out.

3 ADAPTABLE PERSONAL

SPACE MODEL

While a person only physically occupies a portion of

the intimate space, see Figure 1, there is a sense of

ownership and occupation associated with the per-

sonal space. If someone else enters ones personal

space and it was neither expected nor allowed, that

person will be negatively judged because of the in-

trusion. Of course, it is not always possible to keep

ones personal space free of other people, e.g. in pub-

lic transport situations. However, a person willingly

lets go of personal space privileges when participat-

ing in certain activities or entering certain locations.

In approach scenarios the robot should try to obey

the personal space of the person it is approaching and

should only enter when allowed.

In this paper, a new variant of the personal space

model, which is meant to be used for approach-based

human-robot interaction, is proposed. The main dif-

ference between the proposed model and previous

variants of the personal space is that it is dependent

upon the placement of the hands of the person. Each

hand has its own small space with the opposite sign to

the space radiating out from the center of the person.

This can be viewed as a sort of intrusion space where

a cavity is created in the personal space to allow the

robot to intrude. Figure 3 shows the difference be-

tween a common static model (left) and our proposed

adaptable model (right) with a cavity to the persons

left hand side. In the proposed model the space cen-

tered around the left hand allows for intrusion because

of the low value of the personal space in that area.

Each of the spaces that make up the personal space

are modeled by (1).

The center of the Gaussian coincides with the cur-

rent position of hand p

∈ R

resp. the body center,

∈ R

. The shape of the Gaussian is controlled by

the amplitude M

and the covariance matrix C

which

depend on which body part is used – for the body

a isotropic Gaussian has been used, while an elon-

gated Gaussian, hence having elliptical isocontours,

was used for the hands. The cost associated to a posi-

tion p in the personal space is determined by the sum

of spaces for all body parts, each being modeled by:

(p) = M

exp



− p)C

(p − p

)



(1)

with i ∈

{

h, bc

}

indicating the body part. Such that

the total cost is:

cost

(p) = PS

(p) + PS

(p) (2)

4 PATH PLANNING WITH THE

PERSONAL SPACE MODEL

The path planning algorithm Transition-based

Rapidly exploring Random Trees (Jaillet et al., 2008)

(TRRT) works on the same principles as the well

known Rapid Random Trees (LaValle, 2006), but

changes the way new conﬁgurations are accepted.

The transition test, which gives the algorithm its

name, accepts a new node based on the costs and

distance between the two nodes in the extend step.

Negative cost slops are accepted, nodes with costs

above a max threshold are rejected and nodes that

lead to positive cost slopes are accepted with a low

(self-tuning) probability.

The cost of a robot conﬁguration is determined

by the value of the personal space at that conﬁgura-

tion’s coordinates, given by (2) with the proper inputs.

Therefore, the negatively valued parts of the personal

space, determined by the persons hands, will contain

HUCAPP 2017 - International Conference on Human Computer Interaction Theory and Applications

118

Figure 4: TRRT exploration in an artiﬁcial environment

with many high cost areas.

the goal conﬁguration or goal region for the path plan-

ning and allow for easy exploration due to the associ-

ated low cost.

Figure 4 shows the TRRT exploring and ﬁnding

a solution in a cost environment with many high cost

areas. The cost of the environment in which the sim-

ulated approaches take place in is considered obsta-

cle free, and therefore only inﬂuenced by the personal

space model.

5 EXPERIMENTS

Recordings of the body and hand position of a real

person, using the Microsoft Kinect V2 sensor, were

used to obtain the input for the personal space model

for the simulation environment. The recorded person

was sitting in a chair and stretched his left hand out to

indicate where he wants the robot to end up. The sim-

ulated person is static throughout the robot approach

and is merely there for visualization purposes.

The TRRT algorithm generated a path ofﬂine and

post-processing steps for smoothing the path were ap-

plied. The approach was visualized in the Gazebo

simulator using the Care-O-Bot ROS/Gazebo pack-

ages. A series of videos was recorded, where the

robot approached a sitting man using different trajec-

tories and different orientations along the trajectory.

An online survey gathered responses (between 36 and

39 for each question) about agreement with the state-

ments S1-S4 about the robot approaches.

S1: The robot came too close during its movement

S2: The robot stopped too close

S3: The robot’s movements were predictable

S4: The robot’s intention was clear

The participants rated their agreement on a 5-point

Likert-scale between 1 (strongly disagree) and 5

(strongly agree). Participants were not privy to any

information about what they were going to be asked

Figure 5: Simulated environment, personal space model and

TRRT exploration tree.

or the content of the videos and the order of the videos

was randomized.

In the simulation scenario, the robot starts near the

kitchen area and moves to the persons left hand side

to deliver a beverage. The start and goal location are

therefore the same in all the videos, but the robot path

and orientation differs. Figure 5 shows the different

steps in simulating the approach scenario. Pictured

is the kitchen environment, the exploration of possi-

ble robot conﬁgurations by the TRRT algorithm and a

plot of the personal space model (cost function) with

a cavity in front of the person to his left hand side.

More details about the scenario and simulations can

be found in Jeppesen (2015).

The four different approach strategies are visual-

ized in a time-lapse manner in Figure 6(a-d) and are

accessible online

6 FINDINGS

The paths that were planned using the personal space

model (video 1 and 2) tended to stay farther from

the person during the approach and only come close

once there was a cavity in the personal space (see

Figure 5). Approaches planned without the per-

sonal space model (video 3 and 4) would take almost

straight line paths to the goal, violating the personal

space boundaries.

Participants rated their agreement with statements

in a questionnaire on a 5-point Likert-scale. For each

statement an ANOVA was conducted to investigate if

the usage of the proposed model had a positive ef-

fect. The results indicate a positive effect for the

https://www.youtube.com/playlist?list=PLKvb2SEOLj

c9Wi ts3GwhRE 7aV7UR 2z

Socially Acceptable Behaviour for Robots Approaching Humans using an Adaptable Personal Space

119

(a) (b)

Figure 6: Illustration of the videos. Top row (a, b): using the adaptable model; bottom row (c, d): using straight motions; left

column (a, c): facing the in direction of motion; right column (b, d): facing the subject.

closure, based on the statement “the robot came too

close during its movement”, (p=0.061). Interestingly,

the intention of the robot was rated to be less clear

(p=0.049) when the proposed model is used. This can

indicate that the participants might expect additional

cues, revealing the robots intention, to be provided by

the robot.

The paths planned with the adaptable personal

space model (video 1 and 2) rated more positively for

the closeness of the robot during the movement, i.e.

the participants had stronger disagreement with the

statement. The results clearly indicate that video 2

shows the best rated approach. Keep in mind that the

difference between video 1 and 2 is that in video 2 the

robot keeps looking at the person during the approach,

i.e. the simulated eye-contact. The straight-line paths

(video 3 and 4) rated higher in terms of predictability

of motion, which was expected since it is a simpler

motion.

An interesting ﬁnding was that similar paths that

utilized different robot orientation strategies rated dif-

ferently. The paths planned with the adaptable per-

sonal space model were tested with the robot looking

in the direction of travel and with the robot trying to

maintain ‘eye-contact’ by facing the user the entire

time. The approach with eye-contact (video 2) had

signiﬁcantly better ratings in terms of closeness of the

robot compared to video 1 without eye-contact. The

difference in rating between video 3 (no eye-contact)

and video 4 (eye-contact) is smaller, but supports the

trend. The down-side to the strategy of maintaining

eye-contact is that it can make the robot’s movement

feel less predictable, especially when the robot is not

moving in a straight line, because it is easier to guess

where the robot is going when the robot faces in the

direction of movement.

The comments from the participants provide addi-

tional clues on this aspect. One of them (participant

31) argues:

“There is a scenario where the robot makes a

wide arc whilst keeping the person in its focus

– this seems predatory and is alarming.”

on the other hand participant 38 states:

“The ﬁrst one where it curved around you to

your side whilst facing you gave a much more

waiterly impression, I liked that one.”

which indicates that neither of the strategies might

fully comply with the behavioral norms that are ap-

plied unconsciously. This might be one underlying

reason for the results for the different conditions not

to differ as much as expected. Furthermore, this leads

to the conclusion that a more complex model for the

approach behaviour, with the gaze as an active el-

ement and taking characteristics of the person, e.g.

gender, into account, will improve the robots abilities

to optimize the social distance (Mumm and Mutlu,

2011).

HUCAPP 2017 - International Conference on Human Computer Interaction Theory and Applications

120

7 CONCLUSION

An adaptable personal space model was used as a cost

function for a path planning algorithm, TRRT, to plan

simulated approaches toward a sitting person. The ap-

proaches using the proposed adaptable personal space

model were rated more positively than the direct ap-

proaches, according to results from an online survey

asking if the robot came too close during its move-

ment. Furthermore, approaches where the robot ori-

ents itself to look at the person received more positive

ratings than approaches where it was oriented in the

direction of travel.

Due to the simplistic nature of the simulation envi-

ronment, the adaptability of the personal space model

could not be tested fully. Live trials with the real robot

would be the best way to verify the trends that the

online survey has revealed, concerning eye-contact

and the distance regulation due to the personal space

model. The different approaches will likely be rated

slightly differently in live trials due to the embodi-

ment and physical presence of the robot, which are

an important aspects of human-robot trials. How-

ever, our results give good indications that our adapt-

able personal space model has relevant properties and

that gaze contact is rated positively in approach be-

haviours.

ACKNOWLEDGEMENTS

This work has been supported by the project Pa-

tient@Home which is funded as a strategic platform

for innovation and research by the Danish Innovation

Fond.

REFERENCES

Dautenhahn, K., Walters, M., Woods, S., Koay, K. L., and

Nehaniv, C. L. (2006). How may I serve you? A robot

companion approaching a seated person in a help-

ing context. In 1st ACM SIGCHI/SIGART Conf. on

Human-robot Interaction.

Graf, B., Reiser, U., H

agele, M., Mauz, K., and Klein, P.

(2009). Robotic home assistant Care-O-bot 3 - prod-

uct vision and innovation platform. In IEEE Workshop

on Advanced Robotics and its Social Impacts, pages

139–144.

Hall, E. T. (1966). The Hidden Dimension. Garden City,

N.Y., Doubleday.

Jaillet, L., Cort

es, J., and Sim

eon, T. (2008). Transition-

based RRT for Path Planning in Continuous Cost

Spaces. In Intelligent Robots and Systems, pages

2145–2150.

Jeppesen, K. C. (2015). Socially acceptable behavior for

robots approaching humans. Master’s thesis, Univer-

sity of Southern Denmark.

Kirby, R., Simmons, R., and Forlizzi, J. (2009). Com-

panion: A constraint optimizing method for person-

acceptable navigation. In IEEE Int. Symp.on Robot

and Human Interactive Communication.

LaValle, S. M. (2006). Planning Algorithms. Cambridge

University Press, Cambridge, U.K.

Mumm, J. and Mutlu, B. (2011). Human-robot proxemics:

physical and psychological distancing in human-robot

interaction. In 6th Int. Conf. on Human-robot Interac-

tion.

Rios-Martinez, J., Spalanzani, A., and Laugier, C. (2015).

From proxemics theory to socially-aware navigation:

A survey. Int. Journal of Social Robotics, 7(2):137–

153.

Woods, S., Walters, M., Koay, K. L., and Dautenhahn, K.

(2006). Comparing human robot interaction scenarios

using live and video based methods: towards a novel

methodological approach. In 9th IEEE Int. Work. on

Advanced Motion Control, pages 750–755.

Socially Acceptable Behaviour for Robots Approaching Humans using an Adaptable Personal Space

121