Content uploaded by Fangkai Yang
Author content
All content in this area was uploaded by Fangkai Yang on Aug 18, 2020
Content may be subject to copyright.
Impact of Trajectory Generation Methods on Viewer Perception of
Robot Approaching Group Behaviors
Fangkai Yang1†, Wenjie Yin2†, M˚
arten Bj¨
orkman2, Christopher Peters1
Abstract— Mobile robots that approach free-standing con-
versational groups to join them should behave in a safe and
socially-acceptable way. Existing trajectory generation methods
focus on collision avoidance with pedestrians, and the models
that generate approach behaviors into groups are evaluated in
simulation. However, it is challenging to generate approach and
join trajectories that avoid collisions with group members while
also ensuring that they do not invoke feelings of discomfort.
In this paper, we conducted an experiment to examine the
impact of three trajectory generation methods for a mobile
robot to approach groups from multiple directions: a Wizard-
of-Oz (WoZ) method, a procedural social-aware navigation
model (PM) and a novel generative adversarial model imitating
human approach behaviors (IL). Measures also compared two
camera viewpoints and static versus quasi-dynamic groups. The
latter refers to a group whose members change orientation and
position throughout the approach task, even though the group
entity remains static in the environment. This represents a more
realistic but challenging scenario for the robot. We evaluate
three methods with objective measurements and subjective
measurements from viewer perception, and results show that
WoZ and IL have comparable performance, and both perform
better than PM under most conditions.
I. INTRODUCTION
As mobile robots continue to have increased autonomy in
human-robot interactions and are expected to work together
with humans in teams, the ability to robustly approach and
join groups, such as free-standing conversational groups
[1], is fundamental. When doing so, it is vital to generate
safe and socially-acceptable paths that avoid collisions with
group members and do not make them feel uncomfortable,
for example, by violating their personal space [2]. Recent
work [3], [4], [5], [6] has conducted experiments involving
generating safe and socially-acceptable paths. However, they
have shortcomings that limit their utility. Especially, most
research focuses on generating approach behaviors towards
groups whose members are assumed to be totally static
throughout the approach task. However, many situations in-
volve groups that while static in the environment, have quasi-
dynamic members, i.e., the positions and orientations of
group members change over time as they make adjustments
to account for a variety of factors, from slight weight-shifts
to changes of formation due to a change in the focus of
attention of the group or role of individuals within it [7].
This also relates to changes of position and orientation in
order to make space for newcomers to join the group [8].
1Department of Computational Science and Technology, KTH Royal
Institute of Technology, Stockholm, Sweden.
2Division of Robotics, Perception and Learning, KTH Royal Institute of
Technology, Stockholm, Sweden.
†{fangkai, yinw}@kth.se.
Another limitation of previous work is the lack of evaluation
with human participants and comparison with other trajectory
generation methods.
To study these limitations, we conduct an experiment
to evaluate robot approaching group trajectories that are
generated by three methods: a WoZ (Wizard-of-Oz) approach
[9], a procedural model [10] and an imitation learning based
model [11]. They are selected to represent manually con-
trolled models, computational models, and machine learning
models, due to their good performance (excluding WoZ) in
respective domains. Our experiment is conducted in a motion
capture lab to better capture the behaviors of group members
that are used as input to three methods (Figure 1). The major
contributions of the paper are summarized as follows:
•We conduct novel experiments to evaluate three tra-
jectory generation methods in human-robot interactions
from viewer perception in which robots approach to join
free-standing conversation groups.
•We consider robot approaching behaviors under various
experiment conditions, including group types, camera
viewpoints, and approaching directions.
Fig. 1: The setup for our experiment in a motion capture
lab where a Pepper robot approaches to join a free-standing
group with three group members.
II. RELATED WORK
In this section, we present a summary of research on robot
approaching group behaviors with related models to generate
such behaviors and perception of robot movements.
A. Approaching Group Behaviors
Many studies have been carried out that specifically con-
cern the approaching behaviors of robots into small free-
standing conversational groups. As proposed by Kendon
[12], the positions and orientations of individuals in such
small groups are defined as the F-formation system. The
central area within a group surrounded by group members,
called O-space by Kendon, is an exclusive space that pre-
vents robots from intruding into the group. For example,
robots outside a group want to approach to join it, and they
need to calculate a trajectory that does not intersect with the
group’s o-space.
Leveraging the F-formation system, Truong et al. [13]
proposed a framework to enable a robot to approach a
human group safely and socially. Claudio et al. [14] simulate
approaching behaviors for virtual characters towards small
groups. Samarakoon et al. [15] designed a rule-based method
to replicate the natural approaching behaviors of humans.
In recent work, Yang et al. [10] proposed a social-aware
navigation method to navigate a robot to join a group in a
socially-acceptable manner.
With the advent of deep learning, machine learning meth-
ods are being used to generate safe and social approaching
group behaviors. Ram´
ırez et al. [16] adopted inverse rein-
forcement learning, involving several participants demon-
strating approaching behaviors for a robot to learn. Gao et
al. [17] proposed a deep reinforcement learning model to
generate robot approaching group behaviors. However, most
methods involving groups consider them to be totally static,
i.e., models assume that group members do not change body
orientations and positions over the course of the approach.
More recently, as pointed out in [7], individuals in groups
are quasi-dynamic: even in static free-standing groups, indi-
vidual members may routinely shift position and orientation.
An approaching newcomer may, therefore, need to update
their approach trajectory dynamically. By considering groups
with quasi-dynamic individuals, Yang et al. [5] proposed a
model based on Generative Adversarial Networks (GAN) to
generate safe and socially acceptable trajectories into free-
standing conversational groups that adapted to the position
and orientation adjustments of group members.
B. Perception of Robot Movements
Existing research showed that many factors in robot ap-
proaching behaviors influence human perception. Proxemics
is one vital factor that defines social distances between
humans and robots. When robots approach a human subject
or a group, they should not surprise the human subject and
should keep a safe distance, while obeying social norms
or being socially-acceptable [2]. Mead and Mataric [18]
presented an experiment on how human participants perceive
robots during interactions at various distances. Peters et
al. [19] proposed comfortable proxemics when a virtual
robot approaches a human user leveraging Hall’s model [2].
Kuzuoka et al. [20] demonstrated a positional model based
on proxemics that was able to reconfigure the F-formation.
Other works focus on the trajectories of the approach.
Satake et al. [21] found that simply approaching a human
subject by taking the shortest path is not enough to initiate
human-robot interactions. Correspondingly, an analysis of
approach behaviors towards groups by Yang et al. [22]
showed that we do not necessarily approach a group directly,
but take a comfortable path to the group.
Additionally, research has also focused on robot approach
speeds, timings, and directions. Lohse et al. [23] found
that inconstant robot speed during approaching behaviors
increases a robot’s legibility. Similarly, Henkel et al. [24]
found that nonlinear approaching speed makes robots more
acceptable. Huang et al. [25] found that robots with quick
responses to user requests are perceived as more polite.
Concerning approaching directions, studies show that people
prefer to be approached within their field-of-view (FOV)
rather than from directly behind [26]. Ball et al. [27] found
that seated people feel least comfortable when the approach-
ing robot cannot be seen. Moreover, there are works that
focus on eye gaze [28], [29], facial/vocal emotions [30] in
robot approaching behaviors and various human states, such
as sitting [3], walking [31], standing [32].
III. METHOD
Our primary objective is to find a method to generate
approaching group trajectories that are the most socially-
acceptable to the group. To achieve this goal, we proposed
three ways to generate the robot approaching behaviors.
A. WoZ
In the WoZ (Wizard-of-Oz) approach [9], the robot is
teleoperated by a human operator (a researcher who is a
trained operator). To better control the robot interacting with
groups, both the camera view from the robot forehead camera
from Choregraphe1and the reconstructed skeletons from
Motive2(see Figure 2) are used to help the wizard perceive
the environment and participants’ full-body behaviors and
reactions throughout the WoZ process.
Fig. 2: Real-time views that help operator to control the
robot. (Left) The camera view from the robot forehead cam-
era. (Right) The reconstructed scene from Motive including
three group members and the robot.
B. Procedural Model
We use the social-aware navigation method [10] as a
procedural model to generate approaching group trajectories
that are socially-acceptable and realistic. Additionally, Yang
et al. [7] showed the social-aware navigation method is
adopted to quasi-dynamic groups and outperforms other
1http://doc.aldebaran.com/2-4/software/choregraphe/index.html
2https://optitrack.com/products/motive/
state-of-the-art navigation methods, including A* [33] and
SF (Social Forces) [34]. The procedural model is built upon
a social-aware space [10] as shown in Figure 3, where a
darker value area in the social-aware space means the robot
walks in a lower speed and it will be harder to walk through.
A fast marching method [35] was used to navigate the robot.
It is an efficient method to track the motion of wavefronts.
Due to the wave expansion properties, the path following the
wavefront from the target point to the start point will be the
fastest path, a path that is unique and complete.
Fig. 3: The social-aware space of a standing agent facing
right (left) and a free-standing conversational group with
three group members (right) given a top-down view. The
robot starts from the red point and approaches the group
along the purple curve.
C. Imitation Learning Model
Imitation learning aims to learn a policy from expert
observations. We propose to generate approaching group
trajectories use a Generative Adversarial Imitation Learning
(GAIL) [11] framework with a Group Behavior Recognition
framework [22]. In our method, the policy generator πθ
generates state-action pairs (S × A) while the discriminator
Dωis trained to distinguish a trajectory generated by the
policy generator πθfrom the expert observations πE. The
objective function is Eπθ[Dω(s,a)] −EπE[Dω(s,a)].
In each training step, expert trajectories τEsampled from
the CongreG8 dataset [22] are fed into the discriminator
with mini-batch updates, together with sampled trajectories
τθgenerated by the policy generator πθ. The discriminator
is updated with the objective function, and its parameters
ωiare then clipped between (-0.01, 0.01). After updating
the discriminator, the policy parameters θare updated with
Proximal Policy Optimization (PPO) [36].
We adopt a Group Behavior Recognition framework sim-
ilar to the AG-GCN [22] in the policy generator and dis-
criminator as the state encoder. The full-body markers of the
three players are connected as skeleton graphs and fed into
a Spatial Graph Convolutional Neural Network (S-GCN),
which encodes the marker’s spatial relationships into feature
vectors. Past generated or expert trajectories are fed into
an LSTM network, which encodes the past trajectories into
feature vectors. On the group level, the feature vectors of
the three players and the past trajectory are fed to the Group
Graph Convolutional Neural Network (G-GCN). The steps
involved in the imitation learning algorithm are summarized
in Algorithm 6.
Algorithm 1 GAIL Algorithm for Robot Approaching Group
Trajectory Generation
Input:
Input: player markers and expert trajectories
1: Initialize policy generator πθand discriminator Dω
2: for i=0,1,2, ... do
3: Sample expert τEand generated τθtrajectories
4: Update discriminator parameters from ωito ωi+1
with gradient ascent on mini-batches with objective
function: Eπθ[Dω(s,a)] −EπE[Dω(s,a)]
5: Update policy parameters from θito θi+1using the
PPO rule.
6: end for
Fig. 4: The architecture of the state encoder
IV. EXP ER IME NT
A. Experiment Design
To explore our research questions, we designed an ex-
ploratory scenario for human-robot interactions. In order to
keep consistent with the dataset [8] used in imitation learning
(Section III-C), a similar game Who’s the Spy was used as
the scenario (See Figure 5). This game involves three players
in a group. In every game round, each player is given a card
with a word on it. Among the three cards, only one card has
a different word. When the game starts, the players take turns
to describe the material properties of the word they have in
hand. The robot acts as an adjudicator to identify the player
who holds the different word cards, i.e., the spy. Once the
robot established the identity of the spy, it approaches to join
the group to inform group members of the outcome.
In this paper, we use video footage rather than a full live
interaction. Live interaction trials are particularly challenging
to develop when they involve complex human behaviors that
need to be reliable and replicable for statistical analysis.
For example, it is challenging to keep the group member
behaviors to be consistent in various trails when the robot is
controlled by the three methods (Section III). As suggested
in [37], subjective ratings in live and remote (via video)
trials are similar. Shinozawa et al. [38] found that human
decision making depends more on interaction environment
Fig. 5: The Who’s the Spy scenario. Each of the three players
has a card with a word on it, and only one word card is
different. One player only knows the card on hand but does
not know others. They take turns to describe the word, e.g.,
A: It is a fruit. B: We add its juice to salmon. C: It’s sweet if
ripe. The conversation goes on until the robot approaches to
identify the spy, i.e., player B above, and all the players show
the card on hand to confirm if the identification is correct.
and consistency. Additionally, Woods et al. [39] found that
there are no significant findings for the subjective rating of
the practicality for the robot approach direction task, and
comfort levels between the live and video trial. We thus use
video footage as stimuli in this paper as a way to reach a
large number of participants. We record videos under various
experimental conditions (see Section IV-B) performed by
three players. The collected videos are shown to participants
online for subjective assessment of approaching trajectories
from viewer perception (see details in Section IV-E).
B. Experimental Conditions
1) Group Type: The experiment aims to explore the
robot approaching group behaviors under three robot control
methods. Most previous work [13], [17], [26], [27] focused
on static groups where group members keep body positions
and orientations unchanged in the human-robot interaction
process. However, as proposed in [7], conversational groups
are not always static but quasi-dynamic while the group as
a whole is not moving, the group members may change
position and orientation, and the robot needs to update its
approaching trajectory. Yang et al. [8] confirmed the quasi-
dynamic nature in conversational groups via data collection.
In this paper, we thus explore robot approaching behaviors
in static and quasi-dynamic groups.
In the static group, three players stand as a circle on three
equally distributed positions (Figure 6). The group radius is
0.8 meters (the average value from the CongreG8 dataset
[8]). The players stand on the initial location during the
robot approaching procedure. On the other hand, the quasi-
dynamic group has the same group radius.
Unlike the static group that has three possible entering
points for the approaching robot, two players in the quasi-
dynamic group stand closely at the start, which makes only
two entering points available. As shown in Figure 7, while
the robot approaches the group by following an initially
planned path, player C notices the approaching robot and
makes space for it. The robot thus changes the planned
trajectory to join in the new entering point. In order to keep
consistency in various quasi-dynamic trials, all the players
stand on the same initial positions, i.e., two players stand
closely initially to the opposite of the third player. After 2
seconds, one of the two closely standing players walks 0.7
meters aside and looks towards the robot to make space for
it, showing the awareness of its presence.
Fig. 6: A static group with three group members. The robot
approaches from 9 directions including 6 direct approaching
directions (1-6) and 3 indirect approaching directions (7-9).
2) Approaching Direction: Previous work [26], [27] eval-
uated the robot approaching directions towards either an indi-
vidual person or two sitting persons. However, the approach-
ing direction of a humanoid robot towards a conversational
group remains unexplored. In the static group, the robot
approaches from 9 directions, including direct and indirect
ones (see Figure 6), starting 2.15 meters from the group
center. On the other hand, the robot approaches and joins
in the newly available entering point in the quasi-dynamic
group (Figure 6). Note that a mirrored trial is performed
where the robot starts from the back of player A, and at
T=T1player A moves aside to make space. In this case,
the robot initially plans to approach from the left and then
changes to join in the newly opened position. We thus have
two robot approaching directions in the quasi-dynamic group.
3) Camera Viewpoint: As suggested in [40], [41], [42],
camera viewpoint influences human perception on agent
behaviors that they are perceived as more salient in the
egocentric view but with a clearer vision in the perspective
view. We thus collected video stimuli from the egocentric
view and the perspective view (see Figure 8).
C. Experiment Stimuli
Considering that the two group types have different robot
approaching directions, we collected a different number
of video clips in each group type scenario. In the static
Fig. 7: A quasi-dynamic group with three group members.
(Top) The robot started from the back of player C and
planned a path (dash line) to approach and join the group.
However, player C noticed the approaching robot and moved
aside to make space for it at T=T1, and the planned path
changed. (Bottom) A mirrored case where the robot initially
approached to the left and player A moved aside to make
space.
Fig. 8: Collected videos in the egocentric view (left) and in
the perspective view (right).
group, we collected 9 (approaching directions) ×2 (camera
viewpoint) ×3 (robot control methods), i.e., 54 clips for the
static group experiment scenario (27 clips in each camera
viewpoint). On the other hand, in the quasi-dynamic group,
we collected 2 (approaching directions) ×3 (robot control
methods), i.e., 6 clips in the perspective view. However,
since the quasi-dynamic group is asymmetric, we collected
videos in the egocentric view of each player, and it results
in 2 (approaching directions) ×3 (group members) ×3
(robot control methods), i.e., 18 egocentric clips. In total, we
collected 78 clips that represent all condition combinations.
D. Experiment Apparatus
1) Motion Capture: As previously mentioned in Sec-
tion III, both Imitation Learning Model and Procedural
Model need real-time behavioral information of group mem-
bers as inputs. We thus performed the experiment in a motion
capture lab. The room has an approximate 5m ×5m ×3m
active capture volume, which is equipped with a NaturalPoint
Optitrack3system with 16 Prime 41 cameras. Each camera
has a 4 mega-pixel resolutions with a frame rate of 120
fps. The motion of each group member was captured with a
Motion Capture suit with 37 markers placed at respective
anatomical locations of the body (see Figure 9 middle),
and then the captured information is passed to our python
program using NatNet SDK4.
2) Robot: We used a physical Pepper5robot as the
approaching group robot. Additionally, in order to enable a
human operator to monitor the robot without actually seeing
it (Section III-A), four motion capture markers were attached
on its base and chest to track the motion, including its
position and orientation via Motive (see Figure 9 left).
3) Video Recordings: We recorded robot approaching
videos from both the egocentric view and the perspective
view. The egocentric videos were recorded by attaching a
GoPro Hero 5 camera to the upper-chest of one player (see
Figure 9 middle), and the perspective videos were recorded
with a Canon EOS 60D camera equipped with a Sigma
35mm f/1.4 camera lens on a tripod (see Figure 9 right).
Fig. 9: The experiment apparatus: (Left) Pepper robot with
four motion capture markers, (Middle) a player in a motion
capture suit with 37 markers and a GoPro camera to take
egocentric videos, (Right) Camera to take perspective videos.
E. Experiment Procedure
The experiment contains a video collection phase and a
online survey phase.
1) Video Collection: First, an experimenter gave a brief
introduction about the game to the players and informed them
that they could stop their participation at any time. Prior to
the start of the session, the players were asked to fill in
consent forms. The experiments helped each player to select
a Motion Capture suit with an appropriate size and to wear
it. Then the videos were collected in the following phases:
Static Group Phase: The players were asked to keep
standing on the fixed positions during the video recording
phase, but they could talk and perform upper-body behaviors.
The GoPro camera is fixed on the forehead of one player in
this phase to collect egocentric video clips. Then the game
3https://optitrack.com/
4https://optitrack.com/products/natnet-sdk/
5https://www.softbankrobotics.com/emea/en/pepper
started, and the players took turns to describe word proper-
ties, and after 15 seconds, the robot approaches the group to
identify the spy. One trial is thus collected via two cameras
that result in an egocentric video and a perspective video.
Then the next game started with new word cards, and the
robot approached with another combination of approaching
directions and robot control methods.
Quasi-dynamic Group Phase: The setup has discussed in
Section IV-B. The player who moves aside to make space
was asked to perform the same behaviors for other trials
to keep consistency. Note that the GoPro camera was not
attached to one player in this phase, but instead was switched
among all players to capture asymmetric egocentric videos.
After the video collection, we found the egocentric camera
could barely capture the robot approaching behaviors, and
the player (who had the camera under the neck) realized
the robot until the robot already joined the group. We thus
remove clips that robots approach from back left and right,
i.e., direction 4, 5, 8, 9 in Figure 6 if the egocentric camera
is attached to Player B.
2) Online Survey: The online survey phase was divided
into two sessions corresponding to the egocentric and per-
spective viewpoints, as shown in Figure 10. Within each
session, the video clips were divided into two blocks con-
cerning group types, and the ordering of video clips within
each block was counter-balanced. Before each session, the
participant received a training trial where a clip example
with a questionnaire was presented. The data of the training
session was not included in the analysis.
Fig. 10: The online survey procedure. Within each group
block (red cube), the clips are counter-balanced. When it
came to the next participant, the ordering of camera view-
point sessions was switched as well as group blocks.
Twenty-seven participants (18M:9F) aged between 23-
43 years old (M=28.1, SE=4.7) were recruited from the
university locale to take the online survey. Most participants
were not very familiar with robots. They were asked to
watch the videos and answer three questions for each video.
Specifically, participants were asked to rate how much they
thought the robot approaching behaviors were polite,human-
like, and safe, using a 1-7 Likert scale, where 1 means ”not
at all” and 7 means ”very”. These questions are designed to
evaluate the robot approaching behaviors in three dimensions
of social appropriateness, i.e., polite, social, and safe, as
in [17], [43]. In the end, participants were asked to give
feedback. Note that both players in the video collection and
participants in the online survey were told that the robot
was autonomous and lead them to believe that the robot was
actively involved in the game during the group conversation.
V. RESU LTS
We present the results in both objective and subjective
aspects for generated trajectories from WoZ, Procedural
Model (PM), and Imitation Learning Model (IL) under
various experimental conditions.
A. Objective Measurements
Two measurements, collision index (CI), and interaction
index (II), are used to evaluate the social acceptableness
of generated trajectories (see [13] for detailed definitions).
CI is used to measure the physical safety, and II is used
to evaluate the social interactions between the robot and
groups. Note that WoZ and IL generate different trajectories
even though the environment is the same. We thus randomly
sampled one trajectory from WoZ and IL. Figure 11 shows
the sampled trajectories that approach a static group (left)
and a quasi-dynamic group (right) with related CI and II
values. Trajectories generated by IL and WoZ have more
deviated from the group so that they have lower risks of
collision. PM makes abrupt turns when approaching a quasi-
dynamic group, due to its costly computation that results in
reaction delay. IL, however, adopts the quasi-dynamics by
taking an early turn while keeping lower collision risks and
higher group interactions.
Fig. 11: Sampled trajectories (top) with related CI and II
values (bottom) in a static group (left) and a quasi-dynamic
group (right) where a group member moves aside from the
initial position (yellow dot). The robot starts at the red dot
and approaches groups (pink dots).
B. Subjective Measurements
The subjective measurement is done with the online sur-
vey analysis. The mean responses obtained in the different
experimental settings were compared through a repeated
measures Analysis of Variance (ANOVA) F-test. As the input
factors are different in two viewpoints, we perform ANOVA
tests for both egocentric session and perspective session.
The egocentric session has factors including methods (Woz,
PM or IL), approaching directions (left, front, right) and
group type (static or quasi-dynamic), and the perspective
session has methods (Woz, PM or IL), direction (directly
or indirectly), and group type (static or quasi-dynamic) as
input factors.
1) Egocentric Session: As shown in Figure 12 in a static
group, people consider the approaching behaviors in the front
direction as the least polite (F(2,52) = 10.56,p<0.01) and
the least sociable (F(2,52) = 8.02,p<0.01). However, we
found no significant difference in safety. One possible reason
from the feedback is that people are fearful that the robot will
collide even if the speed is imitated from human behaviors.
IL has comparable scores as WoZ in politeness, sociality,
and safety (p0s>0.01), excluding approaching in the front,
and both are rated higher than the procedural model (PM).
Fig. 12: Comparison of three methods in different approach-
ing directions in the static group from the egocentric view.
In the quasi-dynamic group, as shown in Figure 13, there is
no significant difference between IL and WoZ in politeness,
sociality, and safety with p0s>0.01, and they both perform
better than PM significantly.
Fig. 13: Comparison of three methods in the quasi-dynamic
group from the egocentric view.
2) Perspective Session: As shown in Figure 14, there is a
significant effect of three methods in politeness (F(2,52) =
5.14,p<0.01) and safety (F(2,52) = 7.96,p<0.01), but not
in sociality (F(2,52) = 1.62,p>0.01). WoZ is rated as the
highest in politeness and sociality significantly (p0s<0.01)
if the robot approaches directly towards the group, but IL
is rated as the highest in politeness and sociality when the
robot approaches indirectly (p0s<0.01). Both WoZ and IL
perform similarly in the safety dimension.
In the quasi-dynamic group (Figure 15), there is also a
significant effect of methods (F(2,52) = 7.81,p<0.01).
Similar to the conclusion in the egocentric view, IL and WoZ
are rated higher than PM in all three social appropriateness
dimensions, and IL is comparable with WoZ.
Fig. 14: Comparison of three methods in the static group
from the perspective view.
Fig. 15: Comparison of three methods in the quasi-dynamic
group from the perspective view.
VI. CONCLUSIONS
In this paper, we conduct an experiment to evaluate three
methods that generate robots approaching group trajecto-
ries under various experimental conditions. The imitation
learning model has comparable performance with WoZ, and
they both outperform the procedural model in objective
measurements (with a lower risk of collision and higher
group interaction) and subjective measurements from viewer
perception. This offers a way to increase autonomy in mobile
robots interacting with groups. It also raise a question to find
a better solution to control mobile robot than imitating human
behaviors. In the future, we will perform a similar IL model
to output full-body behaviors and transfer imitated behaviors
to mobile robots. We will experiment with groups in different
formations to improve the adaptability of trajectory genera-
tion methods. Moreover, we will extract full-body behaviors
from videos as input to adopt the imitation learning model
in a general scenario without a motion capture setup.
ACK NOW LED GE MEN TS
This research has received funding from the European
Union‘s Horizon 2020 research and innovation program
under grant agreement n. 824160 (EnTimeMent).
REFERENCES
[1] X. Alameda-Pineda, Y. Yan, E. Ricci, O. Lanz, and N. Sebe, “Ana-
lyzing free-standing conversational groups: A multimodal approach,”
in Proceedings of the 23rd ACM International Conference on Multi-
media, ser. MM ’15. New York, NY, USA: ACM, 2015, pp. 5–14.
[2] E. T. Hall, The hidden dimension. Garden City, NY: Doubleday,
1910, vol. 609.
[3] K. L. Koay, D. S. Syrdal, M. Ashgari-Oskoei, M. L. Walters, and
K. Dautenhahn, “Social roles and baseline proxemic preferences for
a domestic service robot,” International Journal of Social Robotics,
vol. 6, no. 4, pp. 469–488, 2014.
[4] R. Triebel, K. Arras, R. Alami, L. Beyer, S. Breuers, R. Chatila,
M. Chetouani, D. Cremers, V. Evers, M. Fiore, et al., “Spencer: A
socially aware service robot for passenger guidance and help in busy
airports,” in Field and service robotics. Springer, 2016, pp. 607–622.
[5] F. Yang and C. Peters, “Appgan: Generative adversarial networks for
generating robot approach behaviors into small groups of people,”
in 2019 28th IEEE International Conference on Robot and Human
Interactive Communication (RO-MAN). IEEE, 2019, pp. 1–8.
[6] T. Kruse, A. K. Pandey, R. Alami, and A. Kirsch, “Human-aware robot
navigation: A survey,” Robotics and Autonomous Systems, vol. 61,
no. 12, pp. 1726–1743, 2013.
[7] F. Yang and C. Peters, “App-lstm: Data-driven generation of socially
acceptable trajectories for approaching small groups of agents,” in
Proceedings of the 7th International Conference on Human-Agent
Interaction, 2019, pp. 144–152.
[8] F. Yang, W. Yin, T. Inamura, M. Bj ¨
orkman, and C. Peters, “Group
behavior recognition using attention- and graph-based neural net-
works,” in Proceedings of the 24th European Conference on Artificial
Intelligence, 2020.
[9] L. D. Riek, “Wizard of oz studies in hri: a systematic review and
new reporting guidelines,” Journal of Human-Robot Interaction, vol. 1,
no. 1, pp. 119–136, 2012.
[10] F. Yang and C. Peters, “Social-aware navigation in crowds with
static and dynamic groups,” in 2019 11th International Conference
on Virtual Worlds and Games for Serious Applications (VS-Games).
IEEE, 2019, pp. 1–4.
[11] J. Ho and S. Ermon, “Generative adversarial imitation learning,” in
Advances in neural information processing systems, 2016, pp. 4565–
4573.
[12] A. Kendon, Conducting interaction: Patterns of behavior in focused
encounters. CUP Archive, 1990, vol. 7.
[13] X.-T. Truong and T.-D. Ngo, “Dynamic social zone based mobile
robot navigation for human comfortable safety in social environments,”
International Journal of Social Robotics, vol. 8, no. 5, pp. 663–684,
2016.
[14] C. Pedica and H. H. Vilhj´
almsson, “Study of nine people in a hallway:
Some simulation challenges,” in Proceedings of the 18th International
Conference on Intelligent Virtual Agents, IVA 2018, Sydney, NSW,
Australia, November 05-08, 2018, 2018, pp. 185–190.
[15] S. B. P. Samarakoon, M. V. J. Muthugala, and A. B. P. Jayasekara,
“Replicating natural approaching behavior of humans for improving
robot’s approach toward two persons during a conversation,” in 2018
27th IEEE International Symposium on Robot and Human Interactive
Communication (RO-MAN). IEEE, 2018, pp. 552–558.
[16] O. A. I. Ram´
ırez, H. Khambhaita, R. Chatila, M. Chetouani, and
R. Alami, “Robots learning how and where to approach people,” in
Robot and Human Interactive Communication (RO-MAN), 2016 25th
IEEE International Symposium on. IEEE, 2016, pp. 347–353.
[17] Y. Gao, F. Yang, M. Frisk, D. Hemandez, C. Peters, and G. Castellano,
“Learning socially appropriate robot approaching behavior toward
groups using deep reinforcement learning,” in 2019 28th IEEE Inter-
national Conference on Robot and Human Interactive Communication
(RO-MAN). IEEE, 2019, pp. 1–8.
[18] R. Mead and M. J. Matari´
c, “Proxemics and performance: Subjective
human evaluations of autonomous sociable robot distance and social
signal understanding,” in 2015 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS). IEEE, 2015, pp. 5984–5991.
[19] C. Peters, F. Yang, H. Saikia, C. Li, and G. Skantze, “Towards the
use of mixed reality for hri design via virtual robots,” in Proceedings
of the 1st International Workshop on Virtual, Augmented, and Mixed
Reality for HRI (VAM-HRI), 2018.
[20] H. Kuzuoka, Y. Suzuki, J. Yamashita, and K. Yamazaki, “Reconfigur-
ing spatial formation arrangement by robot body orientation,” in 2010
5th ACM/IEEE International Conference on Human-Robot Interaction
(HRI). IEEE, 2010, pp. 285–292.
[21] S. Satake, T. Kanda, D. F. Glas, M. Imai, H. Ishiguro, and N. Hagita,
“How to approach humans? strategies for social robots to initiate inter-
action,” in Proceedings of the 4th ACM/IEEE international conference
on Human robot interaction, 2009, pp. 109–116.
[22] F. Yang, W. Yin, T. Inamura, M. Bj ¨
orkman, and C. Peters, “Group be-
havior recognition using attention-and graph-based neural networks,”
2020.
[23] M. Lohse, N. van Berkel, E. M. van Dijk, M. P. Joosse, D. E.
Karreman, and V. Evers, “The influence of approach speed and
functional noise on users’ perception of a robot,” in 2013 IEEE/RSJ
International Conference on Intelligent Robots and Systems. IEEE,
2013, pp. 1670–1675.
[24] Z. Henkel, C. L. Bethel, R. R. Murphy, and V. Srinivasan, “Evaluation
of proxemic scaling functions for social robotics,” IEEE Transactions
on Human-Machine Systems, vol. 44, no. 3, pp. 374–385, 2014.
[25] C.-M. Huang, T. Iio, S. Satake, and T. Kanda, “Modeling and
controlling friendliness for an interactive museum robot.” in Robotics:
Science and Systems, 2014, pp. 12–16.
[26] M. L. Walters, K. Dautenhahn, S. N. Woods, and K. L. Koay, “Robotic
etiquette: results from user studies involving a fetch and carry task,”
in 2007 2nd ACM/IEEE International Conference on Human-Robot
Interaction (HRI). IEEE, 2007, pp. 317–324.
[27] A. K. Ball, D. C. Rye, D. Silvera-Tawil, and M. Velonaki, “How should
a robot approach two people?” Journal of Human-Robot Interaction,
vol. 6, no. 3, pp. 71–91, 2017.
[28] L. Takayama and C. Pantofaru, “Influences on proxemic behaviors in
human-robot interaction,” in 2009 IEEE/RSJ International Conference
on Intelligent Robots and Systems. IEEE, 2009, pp. 5495–5502.
[29] K. Fischer, L. C. Jensen, S.-D. Suvei, and L. Bodenhagen, “Between
legibility and contact: The role of gaze in robot approach,” in 2016
25th IEEE International Symposium on Robot and Human Interactive
Communication (RO-MAN). IEEE, 2016, pp. 646–651.
[30] S. Bhagya, P. Samarakoon, M. Viraj, J. Muthugala, A. Buddhika,
P. Jayasekara, and M. R. Elara, “An exploratory study on proxemics
preferences of humans in accordance with attributes of service robots,”
in 2019 28th IEEE International Conference on Robot and Human
Interactive Communication (RO-MAN). IEEE, 2019, pp. 1–7.
[31] E. Torta, R. H. Cuijpers, and J. F. Juola, “Design of a parametric
model of personal space for robotic social navigation,” International
Journal of Social Robotics, vol. 5, no. 3, pp. 357–365, 2013.
[32] D. Carton, A. Turnwald, D. Wollherr, and M. Buss, “Proactively
approaching pedestrians with an autonomous mobile robot in urban
environments,” in Experimental Robotics. Springer, 2013.
[33] P. E. Hart, N. J. Nilsson, and B. Raphael, “A formal basis for the
heuristic determination of minimum cost paths,” IEEE transactions on
Systems Science and Cybernetics, vol. 4, no. 2, pp. 100–107, 1968.
[34] C. Pedica and H. Vilhj´
almsson, “Social perception and steering
for online avatars,” in International Workshop on Intelligent Virtual
Agents. Springer, 2008, pp. 104–116.
[35] S. Osher and J. A. Sethian, “Fronts propagating with curvature-
dependent speed: algorithms based on hamilton-jacobi formulations,”
Journal of computational physics, vol. 79, no. 1, pp. 12–49, 1988.
[36] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov,
“Proximal policy optimization algorithms,” arXiv preprint
arXiv:1707.06347, 2017.
[37] C. D. Kidd, “Sociable robots: The role of presence and task in
human-robot interaction,” Ph.D. dissertation, Massachusetts Institute
of Technology, 2003.
[38] K. Shinozawa, F. Naya, J. Yamato, and K. Kogure, “Differences in
effect of robot and screen agent recommendations on human decision-
making,” International journal of human-computer studies, vol. 62,
no. 2, pp. 267–279, 2005.
[39] S. Woods, M. Walters, K. L. Koay, and K. Dautenhahn, “Comparing
human robot interaction scenarios using live and video based methods:
towards a novel methodological approach,” in 9th IEEE International
Workshop on Advanced Motion Control, 2006. IEEE, 2006, pp. 750–
755.
[40] C. Ennis and C. O’Sullivan, “Perceptually plausible formations for
virtual conversers,” Computer Animation and Virtual Worlds, vol. 23,
no. 3-4, pp. 321–329, 2012.
[41] F. Martinez-Gil, M. Lozano, I. Garc´
ıa-Fern´
andez, and F. Fern´
andez,
“Modeling, evaluation, and scale on artificial pedestrians: a literature
review,” ACM Computing Surveys (CSUR), vol. 50, no. 5, pp. 1–35,
2017.
[42] F. Yang, J. Shabo, A. Qureshi, and C. Peters, “Do you see groups? the
impact of crowd density and viewpoint on the perception of groups,”
in Proceedings of the 18th International Conference on Intelligent
Virtual Agents, 2018, pp. 313–318.
[43] B. Okal and K. O. Arras, “Learning socially normative robot nav-
igation behaviors with bayesian inverse reinforcement learning,” in
Robotics and Automation (ICRA), 2016 IEEE International Conference
on. IEEE, 2016, pp. 2889–2895.
