Investigating spatial relationships in human-robot interaction

fencinghuddleΤεχνίτη Νοημοσύνη και Ρομποτική

14 Νοε 2013 (πριν από 3 χρόνια και 4 μήνες)

78 εμφανίσεις

Inves
tigating

spatial r
elationships in

human
-
robot i
nteraction

HELGE HÜTTENRAUCH

K
ERSTIN
S
EVERINSON
E
KLUNDH

A
NDERS
G
REEN


E
LIN
A
T
OPP

Human computer interaction (HCI)

Computer science and communication (CSC)

Royal institute of technology (KTH)


HCI
-
31

In:
Proceedings
of the IEEE/RSJ international c
onference on

i
ntelligent
r
obots and systems (IROS 2006)
,

Oct.

9

15, 2006, Beijing, Chin
a

E
-
mail:
{
hehu
, kse,
green,

top
p
}
@csc.kth.se

Human computer interaction (HCI)

Computer science and communication (CSC)

Royal institute of technology (KTH)

S
-
100 44 Stockholm, Sweden

URL: www.csc.kth.se

Investigating Spatial Relationships
in Human-Robot Interaction


Helge Huettenrauch, Kerstin Severinson Eklundh, Anders Green, Elin A.Topp
School of Computer Science and Communication (CSC)
Royal Institute of Technology (KTH)
100 44 Stockholm, Sweden
{hehu, kse, green, topp}@csc.kth.se


Abstract - Co-presence and embodied interaction are two
fundamental characteristics of the command and control
situation for service robots. This paper presents a study of
spatial distances and orientation of a robot with respect to a
human user in an experimental setting. Relevant concepts of
spatiality from social interaction studies are introduced and
related to Human-Robot Interaction (HRI). A Wizard-of-Oz
study quantifies the observed spatial distances and spatial
formations encountered. However, it is claimed that a
simplistic parameterization and measurement of spatial inter-
action misses the dynamic character and might be counter-
productive in the design of socially appropriate robots.


Index Terms – spatiality in human-robot interaction.


I.

I
NTRODUCTION

Humans engaged in physical activities deal with spatial
relationships. The physical mass and degrees of freedom of
body, head, and limbs need to be orchestrated for
movements or manipulation based upon sensory perception
and cognitive abilities. The necessary understanding of
spatiality is claimed to have its origin in evolutionary traits
that shaped not only perception, but influenced human
usage of linguistic metaphors in daily usage [17].
A service robot that operates in the co-presence of a
human user might become engaged in activities that are
determined by the human’s and the robot’s co-presence,
mobility, multimodal communication, and embodied in-
teraction [6].
Trained by daily experience humans are in general
skilled in dealing with other people in managing space and
in handling objects. The signaling, whether through
nonverbal or verbal expressions is well understood,
building upon the ability to notice other people’s body
movements, gaze exchanges, gestures or mimic expressions
[4]. Furthermore signaling in and through the environment
is possible and anchored within the sociocultural context
and practice [2], e.g., a closed door can signify a “please do
not disturb me right now” if this convention is well estab-
lished and adhered to.
Interactive mobile robots are machines that test many
human assumptions about interactive artefacts by pushing
the borderline of our understanding, differentiation, and

Fig. 1 User teaching the robot objects
reactions towards what is alive or inanimate [16], [20]. The
robots’ self-locomotion and the attribution of “body”-
movement as expression of own intentions are contributing
factors. As humans and robots interact, this attribution of
character towards robots might influence humans’ spatial
behaviour in the presence of such devices.
This paper investigates the spatial management in a
Human-Robot-Interaction scenario as illustrated in figure
1: A user guides her new robot around with a “follow-me”
behaviour and shows it the operation area. In this way the
user is teaching the robot places and objects that will allow
the robot to perform service missions afterwards. At the
intended locations, the user and the robot need to position
themselves so that objects or locations can be shown and
named to the robot by speech dialogue.
Our research questions related to the scenario can be
phrased like this:
• How do the spatial distances and orientations of a
user in relation to a robot vary throughout a
cooperative task performed in a home-like environ-
ment?
• Can patterns of spatial HRI behaviours be identified to
guide the design of robots’ spatial conduct?

To investigate these questions we performed a study
with 22 subjects where we observed and recorded the
movement and positioning during this interaction to
1-4244-0259-X/06/$20.00 ©2006 IEEE
understand how robot motion and interaction behaviours
can be designed to be perceived as socially appropriate.
We are interested in this spatial management behaviour
as it requires the active monitoring and dynamic reaction to
each others’ movement and position changes. We also want
to determine how the robot should select appropriate
movement behaviours when interacting with a human user
in a spatial context. Understanding posture and positioning
changes in HRI are prerequisites to reading one another’s
signalling through joint spatial management. It is assumed
that it is used in parallel to other communication modalities
like spoken utterances. To find the relevant features of such
spatial interaction between a robot and a user we let a robot
interact with users and analysed the interaction for
variations in distance and spatial orientation.
The remaining paper is organized as follows: The
background to relevant concepts from social interaction
studies and related research in robotics is given in the next
part. In Section III the user study conducted is presented.
Finally, in section IV we discuss the findings of the study.

II.

B
ACKGROUND

Many disciplines contribute to our understanding of
spatial (inter-) action in co-presence of people and (inter-
active) artefacts. Below relevant concepts such as Hall’s
Proxemics and Kendon’s F-formation system are intro-
duced and discussed for their possible significance in HRI.
A. Hall’s Proxemics
Hall studied interpersonal distances and coined the
term Proxemics [10], i.e., “the interrelated observations and
theories of man’s use of space as a specialized elaboration
of culture” [ibid, p.1]. In the human-robot-interaction
context of posture and positioning, mainly three findings
are of importance: The classification of interpersonal
distances into 4 different classes, the realization of cultural
differences in the spatial behaviour, and last but not least
man’s perception of space. From his observations in the
USA, Hall concluded that social interaction is based upon
and governed by four interpersonal distances: intimate (0-
0.46 meter), personal (0.46-1.22 meters), social (1.22-3.66
meters), and public (>3.66 meters). The combination of
measurable spatial parameters, human ergonomic and
kinetic capabilities, different social roles and interaction as
well as typical characteristics and interaction situations
make Hall’s interpersonal distances interesting for HRI. It
might be hypothesized that the most co-present HRI ex-
changes and reciprocal adaptations between a human and a
robot will happen in the social and the personal distances.
The public distance is of interest as this seems like an
appropriate distance to perhaps try to signal that an ex-
change can or is about to happen. The social and the
personal distance seem appropriate in theory to facilitate
both the communication and the exchange of goods (for
example the manipulation with a robotic arm). The intimate
distance seems to be better suited for exchanges with, e.g.,
so called “mental commit robots” like the seal-robot Paro
[18], where touch is an intended interaction modality.
B. Kendon’s F-formation system
Kendon’s F-Formation system [12] is based upon the
observations that people often group themselves in a spatial
formation, e.g., in clusters, lines, circles, or other patterns.
The term formation is used to express the dynamic aspect
of this spatial arrangement, i.e., the need to actively sustain
it during interaction. This can be observed as small, well
synchronized movements of the participating interactors.
An F-Formation arises when two or more people form a
shared space between them to which they have equal and
direct access due to their sustained spatial and orientational
configuration. The necessary behavioural organization and
movement patterns which are used to sustain this F-Forma-
tion is called an F-Formation system. The F-Formation sys-
tem can be applied directly to an interactive encounter be-
tween a robot and a human: Between the two a so called
transactional segment or o-space is established (marked
with ellipses in figure 2), i.e., a space that both participants
are able to look and speak into, and in which they can
handle objects of shared interest.
Fig. 2: Kendon’s F-formation arrangements
Kendon showed that joint activities and spatial interac-
tions are supported by certain F-Formation system
arrangements, and thus often are encountered in prototypi-
cal situations. In the Vis-à-vis arrangement (figure 2, left)
two participants normally face one another directly; an L-
Shape arrangement (see fig. 2, middle) usually indicates a
joint system in which something is shared in the o-space.
The Side-by-side configuration (fig.2, right) allows two
participants to stand closely together and to face the same
direction. This arrangement often occurs in situations were
both interactors are facing an outer edge given externally
by the environment, e.g., in the form of a table or a wall.
For HRI it is important to notice that all F-formation
arrangements support a triadic relationship between the two
interactors and one or more objects of shared interest, e.g.,
objects that a robot should learn.
C. Spatiality in HRI
Several systems have been designed or studied to en-
able the robots to actively manage spatiality in interaction
with humans.
Yoda and Shiota [22] take the need for safety in pass-
ing a human in a hallway as motivation to develop control
strategies for the robot. Three types of encounters were
anticipated as test cases for their control algorithm, includ-
ing a standing, a walking, and a running person.
Nakauchi and Simmons [13] present another approach
by first collecting empirical data on how people stand in
line. They use these data to model a set of behaviours for a
robot that needs to get into a queue, wait and advance in
the queue for being serviced along with other people there.
Butler and Agah [3] varied a robot’s movement behaviours
and performed a user study to evaluate how different robot
speeds and distances were perceived by users. However, no
interactive task was performed by the robot or user during
this experiment.
A study reported by Althaus et al. [1] used a complex,
room-based sensor array to track the fine movements and
spatial adaptations of a group of people and a robot during
its initial appearance, its “joining of the group”, and finally,
the robot’s departure. The authors concluded that the spa-
tial adaptation observed for the humans could be matched
by the robot’s reacting (in turn) with a dynamic adaptation
in its positioning.
Prassler et al. [15] introduced a robot wheelchair con-
trol system that allowed the system to stay close to an
accompanying person in a crowed subway station, i.e., the
robot movement (with a person) in a highly dynamic con-
text could be demonstrated. Other people (besides the
accompanying person) in this public space were treated as
“dynamic obstacles” that needed to be avoided.
In [19] Topp and Christensen also addressed the dy-
namic, joint movement of a robot and its user. However,
their robot operation setting is confined to an indoor office
space. The interaction is focused on providing a robot
navigation component that can follow users with its laser-
based tracking system during a so called Human Aug-
mented Mapping mission.
Using Hall’s interpersonal distances as parameters in a
robotic system, Pacchierotti et al. [14] recently devised an
algorithm that allows robots to pass people in hallways.

III.

U
SER STUDY

To investigate the spatial distances and orientations of a
user interacting with a robot we designed a study based
upon the idea of a “Home Tour” [5], where a user shows a
robot around and teaches it places and objects in a office or
home-like environment.
A. Scenario and setup
In our trial scenario a user has received a robot and is
ready to use it for the first time. To introduce the robot to
the environment it needs to be shown around to learn rele-
vant places and objects. Once the robot has learned these
the user is encouraged to test the robot. Users could send

Fig. 3 “Living-room” experiment area
the robot on a “search-” or a “find” mission to verify that it
could find locations or previously encountered objects. The
task embedded in the HRI scenario was thus for invited
trial users to (a) get familiar with the robot and navigate it
by letting it follow him or her, (b) teach it places and ob-
jects, (c) validate already taught places and objects, and
(d) handle interaction practically with the robot, including
an initial opening and a closing.
The robot used in this study is an ActivMedia Perform-
ance PeopleBot
1
. It comes equipped with an on-board pan-
tilt-zoom camera. Trial users were told that this camera was
employed by the robot for object and place recognition.
They were also informed that the microphones placed upon
the robot were used by the interactive speech system ena-
bling the commanding of the robot by speech.
The trial was conducted in a room approximately five
by five meter in size. It is furnished with IKEA living room
furniture, including different tables, a bookshelf, and two
sofas (see figure 3). Indicated with numbers are  the
entrance,  the bookshelf,  the Wizard of Oz control
station (with a video camera),  a small table with a
telephone,  a low coffee table upon which different
objects like a remote control and magazines were placed,
 two sofas,  a TV and a VCR combination placed on a
small table, and finally,  a small dining table with a fruit
bowl, a coffee cup etc.
The trial subjects were recruited within the Royal Insti-
tute of Technology, i.e., young technical students of both
genders. Requirements for selection were that they did not
work with or performed research in robotics or computer
vision, as this was judged to be the requirement of a robot


1
www.activrobots.com

encounter with inexperienced users. We conducted 22 trials
(after 4 initial pilot trials for trial-adjustments) with 9
women and 13 men. Participants of the study were
rewarded a cinema ticket for their time and effort.
Upon arrival participants received an introduction to
the robot and the task, both in written form and in the form
of a short demonstration by one of the experiment leaders.
They were then asked to use the robot to teach it new
places and objects and validate these. A time-limit was set
for the interaction, i.e., after 15 minutes a sound indicating
empty batteries for the robot was played to end the trial.
Upon completion of the experiment users were asked to fill
in a questionnaire before being debriefed and told that the
robot’s behaviours during the experiment were simulated.
The robot behaviours were controlled by two experi-
ment leaders who used a wireless robot navigation and on-
board-camera control and a speech synthesizer to produce
spoken dialogue in a Wizard-of-Oz setup [8].
B. Data collection
Multiple sources were used for data collection during
the trial: An external video camera taped the trial in audio
and video from the experiment leaders’ position and per-
spective. Placed in the room’s corners, four webcams run-
ning at a frame-rate of about 1 Hz recorded the interaction.
The images taken with the webcams ensured that the user
and robot movements, postures, and gestures would be
captured from different angles to avoid possible occlusions.
Data from a laser range finder on the robot were stored
and analysed with the help of a person tracking system
[19]. This data represents information about the spatial
distance and positioning of the user under the condition
that the user is in a 180° degree half-circle in front of the
robot.
A system log stored all commands that were sent to the
robot. The different systems mentioned were synchronized
against a local Network Time Protocol (NTP) server. To-
gether with the timing information the robot trials can thus
be run in a simulator at a later point of time. Finally, a
digital recorder was used to record the spoken commands
on the robot itself for detailed speech dialog analysis and
future speech recognition training.
C. Data analysis
To find the relevant spatial interaction patterns and
ways to categorize them, we first carefully examined the
data of a few trials. After this first round of finetuning we
settled for our analysis on a process as follows.
As starting point for the analysis the timeline of the
external video was taken to synchronize the interaction
transcriptions. Based upon these synchronized transcrip-
tions the interaction was then categorized into three
interaction episodes termed “FOLLOW” (user guiding the
robot around), “SHOW” (user teaching the robot places
and objects), and “VALIDATE” (user testing the taught

Figure 4: Visual inspection tool
places and objects by sending the robot on missions to find
them again). Another category of interaction was termed
“BREAKDOWN”, i.e., scenes where miscommunication
and /or task-level incidents led to interruptions in interac-
tion. Often this was accompanied by repair attempts
through speech dialogue, adaptations of position towards
the interaction partner, speech-command repetitions, or a
change of interaction strategy altogether (see [7] for de-
tails).
For each of the identified interaction episodes the ini-
tial posture and positioning, i.e., the distance, orientation,
body posture, gesture(s), utterances, and dynamic position-
ing changes within the episode itself were annotated. The
spatial formation of the user and the robot was analysed
with help of the laser range finder data and a visual inspec-
tion tool (see figure 4). As the laser range finder data is
only available when the user is standing in a 180º degree
half-circle in front of the robot, the visual inspection tool
was applied in situations in which the user was standing
“behind” the robot or laser data was unavailable.
The visual inspection tool displays different webcam
images simultaneously and supports the annotation of the
posture and positioning by pressing pre-defined keys on a
keyboard. Before loaded into this visual annotation tool,
still images are first overlaid and fused with a calibration
image so that virtual dots on the images mark a grid to
calculate distances and positions with. With this aid, marks
in the trial environment that could possibly bias users to
align themselves with were avoided. Image sequences can
be played back and forth and give the possibility to quickly
annotate movements, positioning, and postures.
The outcome of the analysis has been termed a “thick
description” giving the literally frame-by-frame com-
mented observations from the trials. These thick descrip-
tions are accompanied with numerical, quantitative interac-
tion episode descriptions (including still image-sequences
for illustration) for each of the observed Follow-, Show-,
and Validate episodes. This analysis has been conducted
for 11 trials so far, i.e., only half of the available data have
been subjected to this in-depth spatial interaction analysis.
Focusing on the questions posed initially with respect
to the spatial distance and formations of the robot and the
human, the following section will focus on the results of
the spatial management during the Follow, Show, and Vali-
date episodes as analysed from eleven trial sessions based
upon a total of N=321 HRI initiations.
A. Findings
Tables 1-3 give the summarized findings for the HRI
episodes Follow, Show, and Validate as introduced above
for eleven trial subjects. Column 1 (numbering from left to
right) gives the trial-subject’s “identity”, column 2 holds
the number of episodes encountered. Episodes themselves
were then categorized according to Hall’s interpersonal dis-
tances of “Intimate”, “Personal”, and “Social” depicted in
columns 3-5 by checking the metric distance between the
robot and the user and classifying it according to the
appropriate Hall distance.

TABLE

I


F
OLLOW
-E
PISODES
A
NALYSIS


TABLE

II
S
HOW
-E
PISODES
A
NALYSIS


TABLE

III
V
ALIDATE
-E
PISODES
A
NALYSIS


Finally a categorization according to Kendon’s F-
formation arrangements was made: Columns 6-8 give the
number of events recorded as “Vis-à-vis”, “L-Shape”, or
“Side-by-side” F-formations.
Note that the subtotals do not necessarily have to add
up to the absolute number of episodes. The reason is that
subjects also initiated missions while not being in one of
the Kendon F-formations analyzed.
Subjects were free to decide for themselves how to
conduct the trial in detail. Some choose to first iterate
FOLLOW and SHOW missions to teach places and objects
to the robot before trying VALIDATE-missions with a few
selected places and objects. As an alternative strategy
subjects could keep a strict sequential order of FOLLOW,
SHOW, and VALIDATE after one another. The preference
to iterate multiple Follow- and Show-missions first as well
as the observation that Validate missions are taking longer
in duration than Follow- or Show-missions explain why
only 93 Validate missions were observed.

For the Hall’s interpersonal distances it is striking how
predominant the “Personal zone” is, i.e., independent upon
mission-type subjects preferred to position themselves in
the range of 1 to 4 feet (0.45 to 1.2 meters). Interesting is
that the number of subjects who command a Follow-,
Show-, or Validate-mission from the intimate zone is much
smaller than, for example the robot approach distance
reported in Walters et al. [21]. The authors requested
subjects to “move toward the robot” as far as they felt
comfortable and reported that up to 40% of their subjects
came closer than 0.45 meters. Our figures on users entering
into an intimate distance towards the robot are much
smaller as given in Table 1-3 above, e.g., for Follow = 5
(4.6%); Show = 7 (5.8%), and Validate = 12 ( 12.9%) of
users ordered the robot to perform a mission while being in
the intimate Hall distance. Although both experiments used
an ActivMedia PeopleBot
2
the high number of people
coming very close to the robot was not encountered in our
experiment.

Looking at the Kendon F-formations a dominance of
the Vis-à-vis (or face-to-face) positioning of the user to-
wards the robot can be noted, independent upon interaction
episode. The “L-Shape” F-formation arrangement is in
comparison less often observed. Especially in the Follow-
episodes the L-Shape formations are rarely encountered.
The Validate and the Show episodes seem more
appropriate to be handled in an L-Shape formation – as can
be seen from the more frequent occurrences. Especially for
the Show episodes, used to present and label objects and
places in the environment for the robot, the formation of
the L-Shape seems to be more natural.


2
Note that Walters et al [21] had modified their robot: The on-board
camera position was different – additionally, a “lifting arm” was put on the
robot.
Side-by-side F-formation arrangements were rarely en-
countered; most often they occurred in the Follow episode.
This spatial formation – facing an outer edge together – is
likely very dependent upon the environment in which the
human-robot interaction is conducted. The setup in the
“living room”, e.g., in furniture, might, beside the
bookshelf, simply not provide the situation of this
formation to appear very often.
An important limitation to tables 1-3 above should be
explicitly mentioned: Each occurrence in the table is based
upon a clearly identifiable, often speech-dialog initiated,
interaction episode of Follow, Show, or Validate. It is thus
the starting point that was taken as marker of the spatial
relationship between the robot and a subject. This limits the
categorization to a static perspective, i.e., the dynamics of
change over (even short) time periods is not covered.
The fact that this missing dynamic aspect might
however have deeper implications can be seen in figure 5.
It shows the laser tracking plot of a subject’s distance
3
from
the robot centric perspective. The user is approaching the
robot (coming into the view of the laser range finder) and
starts the “Follow-1” episode (depicted through boxes
below the graph) after a short while standing still in front
of the robot at a distance of about 1.2 meters. After spoken
dialogue initiation the subject takes a step from the robot
and waits for the robot’s initial movement as feedback
(visible as an increase of the distance, then again a
decrease). When the robot starts its movement this
feedback signal is taken up by the subject who starts going


3
as a graphical reduction, orientation data of the subject was removed;

towards a corner of the room, rapidly increasing the
distance towards the robot (peaking at about 2.3 meters).
Arriving at a goal-position the subject stops and turns
around waiting for the robot. The robot’s approach towards
the non-moving subject gives a sharp falling flank at the
end of “Follow-1”.
Once the robot has reached the subject’s position the
trial participant makes an observable position and
orientation switch that mark the beginning of the following
two “Show”-episodes. These are then initiated after one
another without noticeable changes in position from the
subject. This is shown through the almost horizontal
(distance-) line of “Show-1” and “Show-2”.

Note however the small position changes in distance
just before and at the end of the “Show”-episodes (pointed
out by arrows in the graph). Almost none-noticeable in the
video-data, these small alignment movements can be found
in the data to often signify transitions from one interaction
episode into another. We find these micro-spatial
adaptations interesting as they might in the future provide
a possibility to try sensor-perception-based triggers
indicating that new interaction tasks or episodes are
prepared for.
The subject’s mission depicted in figure 5 is continued
with multiple “Follow”-episodes; the illustration example
finally ends with another “Show”. While somewhat
disturbed by laser-sensor jitter, even the “Show-4” episode
is characterized by a straight horizontal line.
From the data we have analyzed so far we saw that
different HRI-interaction episodes will also produce
Fig. 5 Robot centric laser data plot showing distance between robot and subject during different interaction episodes
different spatial patterns in the sensor readings that monitor
the (subject) user’s movements and positioning.
Summarizing, we describe the observed dimensions and
differences of the interaction episodes of Follow, Show,
and Validate by their characteristics: “Follow” is best
typified by a paired-dynamic and user-initiative driven
joint activity which, e.g., can be seen from the dynamics of
distance/orientation measurements and high spatial-change
frequencies. “Show” instead has a paired-static, joint
interactivity attribution. Movements are confined to small
adaptive and co-operative engagements and each of the
interactors can be acting or reacting in shaping the
interaction progress. Finally and as in our scenario tried,
“Validate” is neither paired, nor tightly coupled. Once
initiated from the user the robot is acting autonomously
while the user becomes a supervisor monitoring the
progress at best, or possibly, starting a side activity
altogether. What becomes more important with this type of
interaction episode is thus, how both the robot and the user
come together again and continue with their joint track of
interaction once the Validate-mission has been finished.

IV.

D
ISCUSSION

A descriptive analysis of static measurements showed
that Hall’s personal distance, i.e., a distance between robot
and user in the range of 0.46 to 1.22 meters was preferred
in 73% of the observed Follow-, in 85% of the observed
Show- and in 78% of the observed Validate-mission
initiations. Furthermore, Kendon’s “Vis-à-vis” F-formation
arrangement was found to be prevailing among the spatial
configurations tested for. A note of caution was raised to
the applicability of the terms of both Hall and Kendon
however: The dynamic changes and transitions from one
interaction episode state into another one are difficult to
express in terms of Hall’s interpersonal distances and
Kendon’s F-formations arrangements. Kendon’s F-
formations arrangements are dynamically sustained by
small position changes, but the lead-in and lead-out into
these formations, e.g., from a human and a robot need to be
carefully studied. A simplistic parameterization of the
preferred Hall distances and Kendon F-formation
configurations alone therefore seem unsuited to achieve a
socially appropriate robot behavior. A more successful
alternative might reside in the attempt to make other robot
interaction components aware both of the communicative
as well as the coordinating requirements of spatial
interaction. Examples would be to allow the spoken
dialogue model to trigger spatial behavior-signaling or pre-
emptive robot movements as spatial prompts in HRI [9].
A. Design Implications
Findings from this trial might be applied to test the
following robot design enhancements and behavior
strategies to improve spatial management in HRI:
• Testing an interactive robot in its targeted usage
scenario in an early design phase will reveal spatial
management challenges that can be used to improve
the robot’s performance and HRI; detailed findings
might differ according to the context studied.
• Established user preferences in interaction distance
(personal to social) and formation (Vis-à-vis or L-
Shape), are dependent upon specific interaction-
episodes. The transitions between different interaction
episodes should be carefully evaluated and designed
for with appropriate HRI spatial management
behaviors of the robot.
• Looking at the available internal-states and perceptual
data from the robot’s sensory system, e.g., the laser-
range finder (figure 5), it seems possible to evaluate
the current interaction state and spatial management in
HRI by looking for reoccurring and characterizing
patterns such as “user-quickly-leaving”, “robot-
approaching-standing-user”, “user-standing-still-and-
showing-object”, small alignment movements, etc.
This context-interpretation could be used as input in
the design of an interaction planner.
• The robot’s perceptual capabilities should allow for an
extended view on the operation environment and
human movements within it; the used 180
°
degrees in
front of the robot (based upon the capacity of the laser
range finder) appear too limiting in this regard. A
perceptual extension could enable the tracking of user
positions in a 360º degree around the robot. This
would, for example, give the robot the chance to detect
users approaching the robot from behind.

B. Future Work
The study reported was constrained in several aspects
to keep complexity to a level that allowed us to experiment
and investigate aspects of the spatial interaction with a
robot. Potential directions to extend our work include
gradually phasing out the Wizard-of-Oz control elements
with working robot system components as a first step.
Especially the teleoperation control of the robot’s
locomotion (and orientation) will be substituted to validate
our findings and to compare them with experimental HRI
data on spatial positioning that is only governed by imple-
mented robot behaviors.
We are also interested in extending our trial setup to
multiple rooms, possibly making it necessary to traverse
narrow passages together to examine further how elements
of the physical environment shape the spatial cooperation
between a human and a robot.
V.

A
CKNOWLEDGEMENT

The work described in this paper was conducted within
the EU Integrated Project COGNIRON (’The Cognitive
Robot Companion’, www.cogniron.org) and was funded by
the European Commission Division FP6-IST Future and
Emerging Technologies under Contract FP6-002020.
002020.

R
EFERENCES

[1] Althaus, P., Ishiguro, H., Kanda, T., Miyashita T., and Chris-
tensen, H. I., Navigation for human-robot interaction tasks,.
in Proc. of the IEEE International Conference on Robotics
and Automation (ICRA), vol. 2, pp. 1894.1900, April 2004.
[2] Argyle, M. Social Interaction. London: Tavistock Publi-
cations Ltd., 1973.
[3] Butler, J. T. and Agah, A., Psychological effects of behavior
patterns of a mobile personal robot, Autonomous Robots,
Vol. 10, pp. 185-202, March 2001.
[4] Clark, H.H., Krych, M.A., Speaking while monitoring
addressees for understanding, in Journal of Memory and
Language, Vol.50, pp. 62-81.Elsevier, 2004.
[5] COGNIRON. Annex 1 “Description of Work”, EU Project
document, Contract number FP6-IST-002020, 2003.
[6] Dourish, P. Where The Action Is: The Foundations of
Embodied Interaction. MIT Press. 2001.
[7] Green, A., Severinson Eklundh, K. Wrede, B., Li, S.
Integrating Miscommunication Analysis in Natural
Language Interface Design for a Service Robot, submitted.
[8] Green, A, Hüttenrauch, H., Severinson Eklundh, K. Apply-
ing the Wizard-of-Oz Framework to Cooperative Service
Discovery and Configuration. In. Proc. of the 13
th
IEEE
International Workshop on Robot and Human Interactive
Communication, Kurashiki, Japan, 2004.
[9] Green, A, Hüttenrauch, H., Severinson Eklundh, K. Making
a Case for Spatial Prompting in Human-Robot Communica-
tion., submitted.
[10] Hall, E.T., The Hidden Dimension: Man's Use of Space in
Public and Private. The Bodley Head Ltd, London, UK,
1966.
[11] Kanda, Takayuki, Ishiguro, Hiroshi, Imai, Michita, and Te-
tsuo Ono, Body Movement Analysis of Human-Robot
Interaction, in International Joint Conference on Artificial
Intelligence (IJCAI 2003), pp. 177-182, 2003.
[12] Kendon, A., Conducting Interaction – Patterns of Behavior
in Focused Encounters. Cambridge University Press, 1990.
[13] Nakauchi, Y. and Simmons, R., A social robot that stands in
line, in Proc. of the IEEE/RSJ Intern. Conference on Intelli-
gent Robots and Systems (IROS), pp. 357-364, 2000.
[14] Pacchierotti, E., Christensen, H.I., Jensfelt, P. Human-Robot
Embodied Interaction in Hallway Settings: A Pilot Study. In
IEEE International Workshop on Robots and Human
Interactive Communication (Ro-man), pp. 164-171, 2005.
[15] Prassler, E., Bank, D., and Kluge, B., Motion Coordination
between a Human and a Mobile Robot, in Proc. of the 2002
IEEE/RSJ Intl. Conference on Intelligent Robots and Sys-
tems, Lausanne, Switzerland, October 2002
[16] Reeves, B. & Nass, C. The media equation: How people
treat computers, television, and the new media like real peo-
ple and places. NY: Cambridge University Press, 1996.
[17] Shepard, R. N., Ecological constraints on internal representa-
tion: Resonant kinematics of perceiving, imagining, thinking,
and dreaming, Psychological Review, Vol. 91, pp. 417-447,
1984
[18] Shibata, T.; Wada, K.; Tanie, K., Subjective evaluation of a
seal robot in the National Museum of Science and Technol-
ogy in Stockholm, in Robot and Human Interactive
Communication (RO-MAN 2003), pp. 397-402, 2003.
[19] Topp, E.A., Christensen, H.I. Tracking for Following and
Passing Persons. In Proc. of the IEEE/RSJ Intern. Confer-
ence on Intelligent Robots and Systems (IROS), pp. 70-76,
2005
[20] Turkel, S. Life on the Screen. Simon and Schuster. New
York, New York. 1995.
[21] Walters et al. The Influence of Subjects’ Personality Traits
on Personal Spatial Zones in a Human-Robot Interaction Ex-
periment. In: 2005 IEEE International Workshop on Robots
and Human Interactive Communication. pp. 347-352, 2005.
[22] Yoda, M. and Shiota, Y., .The mobile robot which passes a
man, in Proc. of the IEEE International Workshop on Robot
and Human Interactive Communication (ROMAN), pp. 112-
117, September 1997.