Intention Indication for Human Aware Robot Navigation

Oskar Palinko

, Eduardo R. Ramirez

, William K. Juel

, Norbert Kr

uger

and Leon Bodenhagen

SDU Robotics, Maersk Mc-Kinney Moller Institute, University of Southern Denmark,

Campusvej 55, 5230 Odense, Denmark

Keywords:

Human Aware Navigation, Mobile Robots, Intention.

Abstract:

Robots are gradually making their ways from factory ﬂoors to our everyday living environments. Mobile

robots are becoming more ubiquitous in many domains: logistics, entertainment, security, healthcare, etc. For

robots to enter the everyday human environment they need to understand us and make themselves understood.

In other words, they need to make their intentions clear to people. This is especially important regarding

intentions of movement: when robots are starting, stopping, turning left, right, etc. In this study we explore

three different ways for a wheeled mobile robot to communicate its intentions on which way it will go at

a hallway intersection: one analogous to automotive signaling, another based on movement gesture and as

a third option a novel light signal. We recorded videos of the robot approaching an intersection with the

given methods and asked subjects via a survey to predict the robot’s actions. The car analogy and turn gesture

performed adequately, while the novel light signal less so. In the following we describe the setup and outcomes

of this study, as well as give suggestions on how mobile robots should signal in indoor spaces based on our

ﬁndings.

1 INTRODUCTION

Mobile robots are having widespread success in con-

strained, industrial environments, executing various

logistic tasks and both freeing human resources as

well as providing additional ﬂexibility compared to

solutions based on conveyors. However, in other do-

mains such as healthcare, robots are still rarely found,

although they are considered to be one of the means to

mitigate the demographic challenge (Riek, 2017) and

a large variety of technical challenges can be solved

with existing technology already (Bodenhagen et al.,

2019). In healthcare, unlike the industrial domain,

robots are expected to encounter humans that are both

unfamiliar with the robot and vulnerable. However,

besides operating safely, which can be achieved by

utilizing adequate safety mechanism, it is also re-

quired to operate robustly – for mobile robots this im-

plies in particular to adapt the navigation strategies

with respect to humans that share the environment.

https://orcid.org/0000-0002-4513-5350

https://orcid.org/0000-0002-6575-7202

https://orcid.org/0000-0001-5046-8558

https://orcid.org/0000-0002-3931-116X

https://orcid.org/0000-0002-8083-0770

In this paper we will in particular investigate how in-

tention can be conveyed and aid humans in anticipat-

ing the actions of the robot without prior instruction.

Understanding the intention of the robot allows for

adjusting the own behaviour accordingly and thereby

to minimize interference with the robot. It should be

clariﬁed that robots of course do not have intentions

like humans, but as they are perceived as agents, we

should focus on displaying signals that will be per-

ceived as intentions by people. This understanding is

important for the acceptance of robots in our everyday

environments.

2 BACKGROUND

Mobile robots are entering into less constrained envi-

ronments such as public institutions and even private

homes. This shift is caused by technological progress

enabling robots to be reactive and responsive to hu-

mans in their environment (Svenstrup et al., 2009) and

to adjust the planned path online accordingly.

In the following, we will investigate prior work

relevant for the communication of intention since the

understanding of the robots’ intention by humans can

Palinko, O., Ramirez, E., Juel, W., Krüger, N. and Bodenhagen, L.

Intention Indication for Human Aware Robot Navigation.

DOI: 10.5220/0009167900640074

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 2: HUCAPP, pages

64-74

ISBN: 978-989-758-402-2; ISSN: 2184-4321

help avoiding conﬂicting situations which can impose

challenges to traditional navigation solutions and lead

to poor acceptance of robots (Hameed et al., 2016;

Beer et al., 2011). There are several approaches to

provide clues about the intention of robots including

the use of embodiment, physical attributes, or expres-

sive light (Juel et al., 2018). In the following, we

will summarize how humans anticipate the intention

of others (section 2.1) and use this to relate to rele-

vant research focusing on how robots communicate

their intentions (section 2.2).

2.1 Human Anticipation of Intention

Human-human interaction is often successful because

we can respond and react to the intention and action of

other humans. For this, we make use of the contextual

information within the nearby environment where ac-

tions are performed. If a teacup is on a kitchen table

with a tea-bag in it and a person reaches for a pot with

boiling water, we predict that he is probably going to

pour the boiling water into the cup. This type of con-

textual information surrounding an action allows us to

limit the possible outcomes of human intentions (Kil-

ner, 2011).

When the context in the environment cannot be

used or is not present, we use other techniques to pre-

dict intentions. Experiments by (Castiello, 2003) sug-

gest that the intentions of a human can be inferred

by monitoring gaze. If both contextual information

and gaze are absent it has been shown that from body

motion we can predict intentions of humans (Sciutti

et al., 2015) and that human actions translate into dif-

ferent kinematic patterns. (Ansuini et al., 2008) show

that depending on the end goal, we grasp the same ob-

ject differently which means we can use body motion

to anticipate intention.

The different techniques humans use to antici-

pate intention and action are naturally also used when

we want to understand the intention of robots. This

means that we ﬁrst and foremost will try to use con-

textual information to anticipate the intention or state

of the robot. If contextual information is absent it

is the robot’s task to provide us with clues about

its intentions. In the following section, we will de-

scribe how this has been done via anticipatory motion,

augmented reality, animation techniques, expressive

light, animated light, and biological inspired lights.

2.2 Communication of Robot’s

Intentions

(Gielniak and Thomaz, 2011) show that anticipatory

motion can be used to communicate motion intent

earlier than motion without anticipation. They ﬁnd

that when robots are displaying their anticipatory mo-

tion, humans have more time to respond in interactive

tasks. (Ferreira Duarte et al., 2018) show that when a

robot arm overemphasizes a motion, the intentions of

the robotic arms motion becomes predictable. In con-

trast, they also show that when adding gaze and reduc-

ing the motion of the robotic arm to normal the overall

readability of the intention increases. This suggests

that a combination of signaling modalities is stronger

than an overemphasized signal alone which could in-

crease the acceptability of the robot.

A number of HRI studies have employed anima-

tion techniques to improve robots’ intention legibil-

ity. These techniques are borrowed from animated

movies (cartoons). Some of these methods are: antic-

ipation (reaching back before throwing a ball), squash

and stretch (a falling character will squash at landing),

secondary action (a character puts on a jacket while

leaving a house), etc. For a detailed overview of the

ﬁeld, see (Schulz et al., 2019).

Augmented reality can also be used to display

a mobile robot’s intentions of movement. (Coovert

et al., 2014) looked at a robot projecting an arrow

in front of itself signaling its intended direction of

movement. Experiment participants correctly inter-

preted these projections and rated such a robot more

favorably. (Chadalavada et al., 2015) designed an

augmented reality signaling system where the robot

projected a line representing its exact intended path.

People interacting with this robot thought that it was

much more communicative, predictable and transpar-

ent compared to the same robot but without intention

projection.

ortner et al., 2018) hypothesized that colored

light is a suitable feedback mechanism for mobile

robots. They test which of six chosen colors repre-

sent three categories: Active, help needed from hu-

man, and error. They ﬁnd that a green signal should

be used to report active robot behavior, yellow/orange

signal for reporting that the robot needs help and red

signaling if an error occurs. Their results show that

humans interpret light on robots in very speciﬁc ways

and that it can be used to express the internal state of

robots.

Utilizing colored light to signal a robot’s state and

action is supported by (Baraka and Veloso, 2018) that

use light arrays to create both periodic and aperiodic

expressive light signals on a mobile robot. Their re-

sults suggest that the presence of lights on a mobile

robot can signiﬁcantly increase people’s understand-

ing of the robot’s intentions. They completed a user

study where they demonstrate that when using ex-

pressive light to show that the robot needs help more

Intention Indication for Human Aware Robot Navigation

humans would help it and understand that the robot

needed help. This shows that expressive light has an

impact on the behavior of humans around the robot.

Interestingly, their results also suggest that by using

expressive light the trust between humans and robots

increases.

ortner et al., 2018) and (Baraka and Veloso,

2018) suggest that expressive light and animated light

can be used to signal in what operating state the robot

is in while also compelling humans to partake in in-

teractions with a robot.

(Szaﬁr et al., 2015) investigate how expressive

light can be used to communicate directional ﬂight

intention in drones. They apply design constraints to

a drone’s ﬂight behaviors, using gaze, lighthouse bea-

con, blinkers, and airplane ﬂight as inspiration and

thereby design a set of signaling mechanisms to signal

directionality. In a user study, they asked participants

to predict the drone’s behavior. They ﬁnd that using

expressive light to signal directional movements sig-

niﬁcantly improves the understanding of the drone’s

directional ﬂight, where they found the gaze behavior

to be especially useful in communicating the intention

of the drone. (Hart et al., 2019) has made a user study

with a mobile robot that drives towards participants of

the study and the robot signals a lane switch. On the

mobile robot, they mount a virtual head that can turn

and have a gaze. They compare the performance of

the gaze from the virtual agent head against an LED

turn signal. They show that the gaze signal more often

prevents the human and robot from choosing conﬂict-

ing trajectories. This suggests that gaze has some an

advantage in being more explicit than expressive light

but with the essential problem of only being visual

when facing the robot which in turns are very limited

for its overall usability in public spaces because in

many scenarios humans would be interested in know-

ing the intention of the robot from both of its sides and

from behind which could be possible with e.g. expres-

sive lights. Another problem is the level of anthro-

pomorphism required for being able to convey gaze.

A very high level of anthropomorphism might not be

suitable for mobile robots performing logistic tasks

at e.g. hospitals. In general, the literature by (Fer-

reira Duarte et al., 2018), (Hart et al., 2019), (Szaﬁr

et al., 2015) suggests that using signaling mechanisms

that follow known and human-aware conventions and

biological signals (gaze) increased the understanding

of both drone’s and robot’s intentions.

Inspired by the design methodology by (Szaﬁr

et al., 2015) we investigate in this paper the use of

signaling schemes from the automobile domain. The

beneﬁt of employing such signals for mobile robots is,

as suggested by (P

ortner et al., 2018) and (Baraka and

Figure 1: Robot test platform used for the experiments

(left); top view of the robot with standard car light pattern

(right).

Veloso, 2018), that animated lights occur (blinking,

etc.) and they follow well-known conventions. We

also incorporate anticipatory motion by using the em-

bodiment of the robot to signal directional movement

(turn cue) and we combined the usage of expressive

light and anticipatory motion.

3 METHODOLOGY

This section covers all details of the research ap-

proach that was taken in this paper. First we will

describe the mobile robot platform with the signal-

ing unit (Section 3.1) and the implementation of the

signals (Section 3.2). Finally we will detail the exper-

imental design including the recording of the videos,

online survey and data collection in Section 3.3.

3.1 Experimental Platform - The Robot

In order to perform the experiments, we built a test

platform, using a MiR100 autonomous mobile robot

as the base due to its stability, payload capacity and

off the shelf implementation. We made electrical

modiﬁcations to control the LEDs around the robot.

Figure 2: Robot system architecture.

HUCAPP 2020 - 4th International Conference on Human Computer Interaction Theory and Applications

(a) (b) (c) (d) (e)

Figure 3: Signaling behaviors. Each image shows a robot top view where the lines in the perimeter represent the bottom

LEDs, and the circles inside the rectangle represent the top LED ring arrays. The arrow in the top indicates the front side of

the robot. (a) Standard car lights (b) Left blinking signal (c) Right blinking signal (d) Left rotating signal (yellow LEDs move

in the arrow direction) e) Right rotating signal (yellow LEDs move in the arrow direction).

On top of the mobile platform, a chest of drawers was

placed to emulate a logistic use case and design of

the Health-CAT robot. Atop the drawers we installed

four LED ring arrays. Both, the bottom and top sets

of LEDs, could be addressed individually by a con-

troller, giving the possibility of deﬁning the custom

patterns needed for the experiments (see Figure 1).

The system architecture is outlined in Figure 2.

We used an Intel NUC as the master computer. It had

ROS (Robot Operating System) framework installed,

and it provided an interface between the mobile plat-

form and the microcontroller. In order to trigger the

robot behaviors needed for the experiment, we used

a Logitech Wireless Gamepad F710. This remote

controller was constantly sending an array ﬁlled with

0 and 1 (buttons array), representing which buttons

from the gamepad were pressed. The array was then

read by the master computer, which, based on the but-

tons pressed, ran the corresponding robot behavior.

A robot behavior consisted of both, sending veloc-

ity commands (velocity cmd) to the robot platform to

move and pattern commands (pattern cmd) with the

light pattern ID to the microcontroller in charge of

the light control. The communication between these

two devices followed the MQTT messaging protocol,

with the master computer being the MQTT server.

The light control was implemented on a ESP8285

microcontroller, which provides not only a digital

I/O interface, but also a wireless communication with

other devices. Following the master computer re-

quirements, the microcontroller was programmed as

a MQTT client. For the light control, we pro-

grammed the patterns, ﬁrst, in a way they emulated

car lights: white for the headlights, red for the tail

lights and yellow for the signal lights (Figure 3a).

Along with the car signals (Figure 3b), an extra pat-

tern was programmed, showing a LED rotation mo-

tion in each of the rings and the base strip in the

direction of the turn (Figure 3d). Further explana-

tion about the robot signals can be found in Section

3.2. Based on the reading from the pattern cmd, it

was selected which pattern to run, which was later

decomposed into light commands that were sent to

both, the LED strip (led strip cmd) and the LED

rings (led ring cmd). The commands transmission

followed a NZR communication protocol managed

by open source libraries speciﬁc for each LED type

(Fast LED for the strip and Adafruit NeoPixel for the

rings). The structure of those commands consisted of

the LED position and the color to light on.

3.2 Experimental Conditions - Robot

Signals

It is important for a mobile robot intended for hospital

use like ours, to communicate its intentions clearly.

In the following sections, three potential solutions are

outlined: blinking lights, rotating lights, and a turning

gesture.

3.2.1 Blinking Lights

This solution for signaling was designed as an anal-

ogy with standard signaling on cars for making turns.

In this, the outer halves of the top lights were blinking

with yellow light at a frequency of about 1Hz. The

blink frequency was designed to comply with auto-

motive industry standards. The inner halves of the

top front lights were kept constant white, while the

See a video of the implemented conditions here https:

//youtu.be/J6jtDH6ZSuw

Intention Indication for Human Aware Robot Navigation

inner parts of the top back lights were constantly red

(Figures 3b, 3c and 4b). These patterns were emulat-

ing the front and back lights on cars. Using this setup

also adds information about which end is the front of

the robot (white inner lights) and which the back (red

inner lights), again in analogy with cars. The bot-

tom LED strip was also blinking at the corner position

with the same color LEDs.

3.2.2 Rotating Lights

We implemented a rotation of the top and bottom

lights as a novel signaling method. In this condition

only two LEDs of each top light ring were on at one

time. The lit LEDs kept changing to create an ef-

fect of rotating lights. All four lights were moving in

the same direction. This direction was correspondent

with the future turn of the robot: the lights were turn-

ing clockwise when the robot wanted turn right and

anti-clockwise when it intended to turn left. The bot-

tom lights were displaying a ”running” pattern around

the base of the robot, corresponding to the future turn

direction of the robot Figures 3d, 3e and 4c.

3.2.3 Turn Gesture

A ﬁnal signaling option was implemented in the form

of movement: when the robot reached the intersec-

tion, it made a turning gesture of about 30 degrees

towards the side it wanted to turn to (Figure 4a). We

considered this the strongest signal indicating the in-

tention of the robot: there was no reason why a turn

gesture to one side would be interpreted as an inten-

tion to move in the other direction. In the experiment

itself we used this signal in two ways: 1) by itself

without additional indicators and 2) in combination

with the above mentioned two other signaling meth-

ods.

3.2.4 Combination of Conditions

The main independent variable of our study was the

signaling method with conditions: blinking light, ro-

tating light and turn gesture. The last of the condi-

tions could be administered either by itself or in com-

bination with the ﬁrst two methods. We also wanted

to show movements of the robot using a standard car

light pattern without any turn signals (none), to estab-

lish a baseline behavior. A secondary necessary in-

dependent variable we considered was turn direction

with levels: left and right. All these combinations of

conditions are represented in Table 1.

For the combination of no-turn and no-light the

left and right conditions are the same, i.e. the robot

just approaches the intersection without any turning

Table 1: Combination of conditions.

without turn gesture with turn gesture

blink rotate none blink rotate none

left • • • • •

right • •

•

• • •

(a) (b) (c)

Figure 4: Signaling behaviors on real robot. (a) Right turn

gesture with standard car lights (b) Blinking signal to left

lights or gestures. Therefore these two cases are con-

joined into one condition. This give us a total of 11

combinations of conditions.

3.3 Experimental Design

We ran a human subject study to investigate which

signaling approach would be the most appropriate for

our robot. In order to ensure repeatability and efﬁ-

cient gathering of human data, we opted for recording

videos of the robot’s signaling movements and show-

ing them to human participants via online surveys. As

it was essential for all subjects to see exactly the same

robot behaviors, video recordings were the best op-

tion.

3.3.1 Video Recordings

One video was recorded for each of the 11 combina-

tions of conditions mentioned above. The videos were

shot with a Motorola moto G7 Plus mobile phone’s

primary camera in 4K resolution with 30 frames per

second. The recording was done from an initial dis-

tance to the robot of 3,5 m. At the end of the video

the robot approached the intersection and was at a dis-

tance of 2,3 m to the camera, see Figure 5. The mo-

bile phone was mounted on a ﬁxed tripod at a height

of 1,7 m, thus simulating the point of view of a per-

son. The location of the video recording was exactly

the same in all clips. We made a careful selection of a

location that represents a symmetrical intersection of

hallways at the university. The robot always started

from the same position of 1,25 m from the intersec-

HUCAPP 2020 - 4th International Conference on Human Computer Interaction Theory and Applications

tion. It always ended its movement just at the bor-

derline of the intersection box. It was always in the

center of the corridor allowing the same amount of

space on both of its sides. The width of the hallway

at the entrance to the intersection was 2,22 m. The

width of the robot is 0,58 m. This allowed a space of

0,82 m on both sides of the robot for passing around

it. In the conditions where the robot also performed

a turning gesture, this space somewhat decreased, be-

cause of its rectangular geometry. The videos ended

at the point when the robot stops at the intersection

and has performed the turn gesture and light actions.

We intentionally cut it off before the robot performed

the actual turn at the intersection, because we wanted

to ask our participants to tell us what their prediction

would be on what will happen next, thus giving us

insight on the effectiveness of our signaling methods

in conveying information about the robot’s future ac-

tions.

3.3.2 Online Survey

An online survey was created to test the designed ex-

perimental conditions with human subjects. As a sur-

vey platform, we selected the site soscisurvey.de be-

cause of its high customizability. Among other things,

most importantly it allows many options for random-

izing the order of presentation of the videos. The sur-

vey started with a quick explanation of the experiment

without giving away its scientiﬁc purpose. After the

initial slide, we presented the 11 videos, each on a

separate page, in pseudo-random order. It was de-

signed to be counter-balanced, but not all ﬁelds of the

procedure were covered because there were 120 com-

binations of order and 30 subjects. The videos were

Figure 5: Video recordings setup layout. Dimensions are in

mm.

Figure 6: The last frame of the video, representing the route

options in the multiple choice questions.

divided in two groups:

1. With turn gesture,

2. Without turn gesture.

The randomized ﬁrst group was always shown before

the randomized second group. The videos contain-

ing the turn gesture were selected to be shown later,

because when they were in combination with either

of the light signals, they could inﬂuence the subjects’

subsequent decision on light signals appearing with-

out the turn gesture. This could happen as we ex-

pected the turn gesture to be the strongest indicator for

the turning intention of the robot. The questions we

asked the participants right after showing each video

are shown in Table 2.

Table 2: Questions about the videos.

QA1

Which way would you go around the robot to get

to the end of the hallway ahead?

QA2

Which way will the robot turn after the end of the

video?

For the ﬁrst question we offered multiple choice

radio buttons with the labels: a,b,c,d,e,f,g,h,i,j. These

labels represented ten possible routes of movement

for people to take and corresponded to arrow repre-

sentations of these paths at the end of the video, which

were added in a video editing application, see Fig-

ure 6. The second question’s answers were three radio

buttons with the labels m, n and o. These were also

represented by arrows at the last frame of the video.

The videos only played once (of which the subjects

were informed at the initial page) and stopped at the

last frame with the arrow representations.

After the survey pages with the videos the subjects

Intention Indication for Human Aware Robot Navigation

had two more pages with questions to ﬁll out: one

about the experiment and another about demographic

information. The experimental page started with two

questions for validating if subjects paid attention to

the videos.

Table 3: Validation questions about the experiment.

QB1 What was the color of the robot?

QB2 What was the general shape of the robot?

The ﬁrst of these questions had an open text ﬁeld,

so people could enter the name of the color they per-

ceived. The second question was a multiple choice

one with the following options: cylindrical, box-

shaped, ball-shaped, snake-shaped, humanoid, other.

The correct choice was ’box-shaped’. The follow-

ing 9 questions were in the form of statements with

7-point Likert-scaled answers ranging from ’strongly

disagree’ to ’strongly agree’, see Table 4.

With these statements we expected to learn more

about people’s preference for the conditions we were

suggesting. This page ended with a comments sec-

tion. We asked subjects in a large textual ﬁeld to let

us know about their thoughts, observations, and sug-

gestions concerning the experiment. The last input

page asked for demographic information, see Table 5.

3.3.3 Data Collection

We opted for collecting data using Mechanical Turk.

Thirty-one participants were recruited with MTurk

Master qualiﬁcation and at least 90 percent job quality

approval rating. They were all located int the United

States. We opted for this country as our questionnaire

was in English and the USA is the largest English

speaking country (as their native language), thus we

could get quick high quality responses. One of the

subjects showed irregular behavior according to our

survey collection system: she spent very little time

on each slide (around 11,7 seconds per slide, while

the average was 25,7 seconds), i.e. she didn’t pay at-

tention to the videos, thus we elimiated her from the

results. Out of the leftover 30, 10 were female and

20 male. The average age was 41,9. Two people were

left-handed and 28 right-handed. They all had driver’s

licenses except one. Twelve subjects never interacted

with robots before, 16 a few times, while two experi-

enced robots a number of times.

4 RESULTS

Results can be categorized into a number of groups.

First we’ll discuss subjective responses, then partici-

pants’ movements as reactions to the robot and ﬁnally

people’s understanding of robot’s intentions.

4.1 Subjective Results

This section will discuss the Likert-scaled agreements

of participants with the statements found in Table 4.

The ﬁrst analysis focuses on opinions about the three

basic signaling methods: blink, rotate and turn. These

refer to statements QC1, QC2 and QC3. There is

some disagreement in the literature on how Likert-

scaled values should be analyzed. The more con-

servative approaches suggest non-parametric statisti-

cal methods as the Friedman test and the Wilcoxon

signed-rank test, which we will use here.

Figure 7 shows the averages of responses to state-

ments QC1, QC2 and QC3 about the clearness of sig-

nals when using rotating lights, blinking lights and

turn gesture. We applied a Friedman test to assess

the difference between these. The results show a sta-

tistically signiﬁcant difference between the signaling

methods, χ

(2) = 44.1, p < 0.001. Post-hoc analy-

sis using Wilcoxon signed-rank test was conducted

with a Bonferroni correction. This has shown signif-

icant differences between all pairs of conditions: ro-

tate compared to blink (Z = -3.781, p < 0.001), blink

compared to turn (Z = -3.13, p = 0.002) and rotate

compared to turn (Z = -4.475, p < 0.001). This means

that participants found the rotating lights the least in-

formative, and the blinking more informative than ro-

tation but less than the turning gesture. Therefore, the

turning gesture was the strongest signal, followed by

blinking, followed by rotating lights.

We intended to explore participants’ preference

of the LED rings on the top versus the LED strips

near the bottom of the robot, statements QC6 and

QC7 in Table 4. Figure 8 shows the averages and

standard deviations for this comparison. Adapting a

Wilcoxon signed-rank test we did not ﬁnd statistically

Figure 7: Averages of opinions on clearness of signaling

with rotating lights, blinking lights and turning. +/- 1SD.

HUCAPP 2020 - 4th International Conference on Human Computer Interaction Theory and Applications

Table 4: Likert-scaled questions about the experiment.

QC1 The blinking lights on the robot in some of the videos made it very clear which way it would want to go.

QC2 The rotating lights on the robot in some of the videos made it very clear which way it would want to go.

QC3 The turning of the robot at the end of some videos made it very clear which way it would want to go.

QC4 The combination of blinking lights and turning made it very clear which way the robot would go

QC5 The combination of rotating lights and turning made it very clear which way the robot would go.

QC6 The lights on the top of the robot were very useful in understanding where it would go.

QC7 The lights near the bottom of the robot were very useful in understanding where it would go.

QC8 I did not notice any difference between the videos.

QC9 The different signals were insufﬁcient for one to understand which way the robot would go.

Table 5: Demographic information questions.

QD1 What is your gender?

QD2 What is your age?

QD3 Are you left-handed or right-handed?

QD4

How many times have you interacted with robots

before?

QD5 Do you have a driver’s license?

QD6

In your country of residence, which side of the

road do cars drive on?

QD7 Have you participated in this experiment before?

Figure 8: Averages of opinions on usefulness of top LED

lights vs. bottom LED lights. +/- 1SD.

signiﬁcant difference between the two conditions (Z

= −1.668, p = 0.095). Even though this results is not

signiﬁcant, we report it because our expectation was

that the top LEDs would be preferred, which was not

the case at all.

4.2 Participant Movement Intentions

This section will present the results derived from par-

ticipants’ answer to question QA1 in Table 2: Which

way would you go around the robot to get to the end

of the hallway ahead? This question was asked once

for each of the 11 videos, showing the 11 conditions.

The answers were in the form of multiple choices

(a,b,...,j). These letters refer to paths proposed in the

last frame of the videos, see Figure 6. Paths a,b,c,d,e

led around the robot from the left side, while paths

f,g,h,i,j went around from the right. Figure 9 con-

Figure 9: Histogram of participants’ bimodal choices on

which way to go around the robot that doesn’t show any

signals.

tains the histogram of answers to this question while

showing the video of the robot with no signals. This

was the baseline condition administered to investigate

people’s default preference of sides.

Participants were fairly evenly divided between

left and right sides. On the left they tended to stay

further away from the robot while on the right, they

approached somewhat closer. There were still more

people (16) circumventing from the right, as opposed

to left (14). This makes intuitive sense, if the robot

Figure 10: Histogram on participants’ binomial choices on

which side to take to go around the robot while displaying

rotation, blinking lights and turning gesture to the left.All

direction labels (left, right) are from the subject’s point of

view.

Intention Indication for Human Aware Robot Navigation

is perceived as a vehicle and right-hand-side driving

road rules apply. It was noticed that these histograms

have an irregular bimodal distribution. Therefore we

opted to reduce the complexity by clustering all left

side paths (a,b,c,d,e) together and all right side paths

(f,g,h,i,j) together, thus simplifying the bimodal to a

binomial (left, right) distribution. This allowed us to

apply simpler statistical methods, while keeping the

most important part of the data. With this new ap-

proach, we compared which way people want to cir-

cumvent the robot when it is signaling to the left and

right. Figure 10 shows the binomial distributions of

side selection when the robot wants to turn to its own

left side. It can be noticed that for blinking and turn-

ing, people follow the expected route, to their left,

because they are correctly perceiving the robot’s in-

tention to turn to its own left and want to avoid col-

lision. However, for the rotating light, we don’t see

the same ’keep left’ distribution. Rather, it is split be-

tween going left and right. This might be because the

rotating light was not perceived as an intention to turn

to that side. To check for differences between these

three binomial distributions we ran a 3x2 Chi-squared

test and found signiﬁcant difference between the three

conditions χ

(2,N=30)=21.67, p<0.001. Post-hoc

analysis of adjusted residuals with Bonferroni correc-

tion revealed (p<0.001) that it was the rotating light

condition that was signiﬁcantly different compared to

the other conditions.

We also investigated the analogous situation, but

when the robot is signaling for a right turn, Figure 11.

Here we do not see a deviation from the expected

choice even for the rotating light. Most people chose

to go to the right side from their point of view, to

avoid the robot which was signaling a turn towards

its own right. Indeed, a Chi-squared test also did not

report a signiﬁcant difference between conditions for

this case, χ

(2,N=30)=5.21, p<0.074.

4.3 Understanding of Robot’s Intention

In this section we report on question QA2 from Ta-

ble 2: Which way will the robot turn after the end of

the video? This question had three possible answers:

paths m,n and o (see Figure 6). It was asked to see if

the subject can make correct robot movement predic-

tions based on the signaling methods. Similarly as in

the previous section we see that blinking and turning

are adequately interpreted, but the rotating light signal

is not. Half of the participants thought that the robot

will continue going straight even though it was rotat-

ing to signal a right turn. We conducted a Chi-squared

analysis on this dataset too and found signiﬁcant dif-

ferences between levels χ

(4,N=30)=27.6, p<0.001,

Figure 11: Histogram on participants’ binomial choices on

which side to take to go around the robot while displaying

rotation, blinking lights and turning gesture to the right.All

direction labels (left, right) are from the subject’s point of

view.

Figure 12: Histogram on participants’ opinions on which

way the robot will go for different signaling methods.

Right, straight, left is from the robot’s point of view. Ro-

tate left, blink left and turn left are from the subject’s point

of view.

indicating that the rotation signal is improperly inter-

preted.

Regarding the signals in combination with turn,

they were very correctly interpreted by almost all

participants. Figure 13 shows three conditions ”ro-

tate&turn”, ”blink&turn” and ”turn”. It can be no-

ticed that the turning signal already provides all in-

formation for making a right decision. The addi-

tion of the two light signals doesn’t change the out-

come. As expected, a Chi-squared test did not show

and signiﬁcant difference between these conditions

(2,N=30)=0.36, p=1. The results in this section

tell us that all subsequent robot movements were cor-

rectly predicted by the subjects, except the rotating

light.

5 DISCUSSION

Section 4.1 demonstrated that people found the ro-

tating lights the least informative, the blinking lights

HUCAPP 2020 - 4th International Conference on Human Computer Interaction Theory and Applications

more and the turn gesture the most informative. These

differences were also statistically very signiﬁcant. We

did expect that the turn gesture will be the most

prominent indicator for the intention of the robot, but

we did not expect the rotating lights to be evaluated

so poorly. It was thought, that a rotational signal

would clearly communicate the robot’s intention to

turn. However this was not reﬂected in the data. It

might be the case that people associate rotating lights

with heavy machinery or oversized vehicles on the

road, which utilize similar signals, but only for at-

tracting people’s attention and not for signaling di-

rection. Even the running lights on the bottom LED

strips did not help in clarifying the intention. The

blinking light performed quite well, as expected. As

we were aiming for analogy with regular vehicles in

terms of lights, we expected that people will easily

make the connection between the robot and a car on

the road. Even the turning gesture can be associ-

ated with vehicular trafﬁc: in many situations at in-

tersections cars will turn towards their intended ﬁnal

goal, especially for right turns in right-hand-side traf-

ﬁc countries. We emphasize that the turn gesture was

not the turn itself. After the gesture, the robot made

a full stop before performing the actual turning ac-

tion. The actual turn was not part of our videos, as we

wanted subjects to predict this.

Next, we looked at the preference of top or bottom

lights. We expected that the top lights would be pre-

ferred as there were more of them and they were more

in the line of sight of passers by. The ring LEDs were

also more expressive, thanks to their numbers and

customizability. However, this expectation was not

supported by the results. Figure 8 shows participants’

opinions on this matter: the average opinion was even

better for the bottom lights, than top, even though this

difference was not statistically signiﬁcant. One of the

possible explanations for this kind of outcome might

Figure 13: Histogram on participants’ opinions on which

way the robot will go for combined signals. Right, straight,

left is from the robot’s point of view. Rotate&turn left,

blink&turn left and turn left are from the subject’s point of

view.

be the similar position of the bottom lights to lights

on regular cars. Although, we designed the top lights

to emulate car signals in one condition, maybe their

location on the top was not appropriate for making a

closer connection with automotive lights.

Regarding the analysis of participants’ intention to

go around the robot from the right or left, we looked at

their answers to question QA1, see Table 2, Figure 10

and Figure 11. There is an inconsistency between the

left and right turns. The leftward rotating condition

produced an even outcome between taking either the

right or left routes. This tells us that the rotating lights

were not a good indicator of turning intention. On the

other hand, for the rightward turn, rotation behaved

as expected. This might be caused by the fact that

more people tend to choose passing on the right (see

Figure 9) in analogy with passing an oncoming car

from the right. Figure 12 might give some insight on

this: it shows people’s prediction of the robot’s turn-

ing intention under the basic conditions. It can be no-

ticed that the leftward rotation signal is interpreted as

an intention to go straight mostly and going left sec-

ondly. Thus, people who think that the robot will con-

tinue straight can go around the robot from their own

right side, even though in the designer’s intention this

would lead them to collision with the robot. Finally,

Figure 13 showed us that the turn gesture overpowers

any additional light signal that might occur contem-

poraneously.

Regarding the number of experiment participants

(30), we acknowledge that their number could have

be higher for even more convincing results, but on the

other hand most of our results showed very high levels

of signiﬁcance (p=0.001), thus adding more subject

most likely wouldn’t have changed these outcomes.

6 CONCLUSION AND FUTURE

WORK

Signaling movement intention is essential for mobile

robots in public environments like hospitals, univer-

sities, airports, etc. Finding an effective signal for

indicating direction at a hallway intersection is cru-

cial for the acceptance of such robots. In this paper,

we discussed three such signals and their combina-

tions. One was a turning gesture which was expected

and proven to be the strongest indicator. Next there

were the blinking signals in analogy with automo-

tive signaling, which performed quite well too. Fi-

nally there was the newly designed rotational signal

which proved to be the least efﬁcient for the purpose,

as some participants misinterpreted it as a general at-

tention signal.

Intention Indication for Human Aware Robot Navigation

All in all, as expected we found that using car sig-

nals on vehicle-shaped mobile robots is a good idea

and a design recommendation for the robotics com-

munity. What we did not expect was that the less vis-

ible bottom lights were at least as important for sig-

naling as the top LED rings, which we expected to

be more informative. We also noticed a tendency of

people passing by an oncoming robot from the right

side, as in vehicular trafﬁc. We might have expected

this to be more signiﬁcant, which it was not, because

many participants chose to go on the left side of the

robot too. As future work it might be interesting to

compare this effect with populations where the driv-

ing is on the left-hand-side of streets. We will also

consider redesigning our robots lights and enforcing

the ones near the bottom of the robot, according to our

ﬁndings. Finally, based on the conclusions from this

controlled study on the most efﬁcient ways to signal

turning intent, we plan to drive the robot on the uni-

versity hallways and collect data in this uncontrolled

environment and analyze how people ﬂow around it

depending on turn signaling.

ACKNOWLEDGEMENTS

This work was supported by the project Health-CAT,

funded by the European Regional Development Fund.

REFERENCES

Ansuini, C., Giosa, L., Turella, L., Alto

e, G., and Castiello,

U. (2008). An object for an action, the same object for

other actions: effects on hand shaping. Experimental

Brain Research, 185(1):111–119.

Baraka, K. and Veloso, M. M. (2018). Mobile service robot

state revealing through expressive lights: Formalism,

design, and evaluation. International Journal of So-

cial Robotics, 10(1):65–92.

Beer, J. M., Prakash, A., Mitzner, T. L., and Rogers, W. A.

(2011). Understanding robot acceptance. Technical

report, Georgia Institute of Technology.

Bodenhagen, L., Suvei, S.-D., Juel, W. K., Brander, E.,

and Kr

uger, N. (2019). Robot technology for future

welfare: meeting upcoming societal challenges – an

outlook with offset in the development in scandinavia.

Health and Technology, 9(3):197–218.

Castiello, U. (2003). Understanding other people’s ac-

tions: Intention and attention. Journal of experimen-

tal psychology. Human perception and performance,

29:416–30.

Chadalavada, R. T., Andreasson, H., Krug, R., and Lilien-

thal, A. J. (2015). That’s on my mind! robot to human

intention communication through on-board projection

on shared ﬂoor space. In 2015 European Conference

on Mobile Robots (ECMR), pages 1–6. IEEE.

Coovert, M. D., Lee, T., Shindev, I., and Sun, Y. (2014).

Spatial augmented reality as a method for a mobile

robot to communicate intended movement. Comput-

ers in Human Behavior, 34:241–248.

Ferreira Duarte, N., Tasevski, J., Coco, M., Rakovi

c, M.,

and Santos-Victor, J. (2018). Action anticipation:

Reading the intentions of humans and robots. IEEE

Robotics and Automation Letters, PP.

Gielniak, M. J. and Thomaz, A. L. (2011). Generating an-

ticipation in robot motion. In 2011 RO-MAN, pages

449–454.

Hameed, I., Tan, Z.-H., Thomsen, N., and Duan, X. (2016).

User acceptance of social robots. In The Ninth In-

ternational Conference on Advances in Computer-

Human Interactions.

Hart, J., Mirsky, R., Tejeda, S., Mahajan, B., Goo, J., Bal-

dauf, K., Owen, S., and Stone, P. (2019). Unclogging

our arteries: Using human-inspired signals to disam-

biguate navigational intentions.

Juel, W., Kr

uger, N., and Bodenhagen, L. (2018). Robots

for elderly care institutions: How they may affect

elderly care. In Coeckelbergh, M., Loh, J., Funk,

M., Seibt, J., and Nørskov, M., editors, Envisioning

Robots in Society – Power, Politics, and Public Space,

pages 221–230. IOS Press.

Kilner, J. (2011). More than one pathway to action under-

standing. Trends in cognitive sciences, 15:352–7.

ortner, A., Schroder, L., Rasch, R., Sprute, D., Hoffmann,

M., and Koenig, M. (2018). The power of color: A

study on the effective use of colored light in human-

robot interaction. In 2018 IEEE/RSJ International

Conference on Intelligent Robots and Systems (IROS),

pages 3395–3402.

Riek, L. D. (2017). Healthcare robotics. Commun. ACM,

60(11):68–78.

Schulz, T., Torresen, J., and Herstad, J. (2019). Anima-

tion techniques in human-robot interaction user stud-

ies: A systematic literature review. ACM Transactions

on Human-Robot Interaction (THRI), 8(2):12.

Sciutti, A., Ansuini, C., Becchio, C., and Sandini, G.

(2015). Investigating the ability to read others’ inten-

tions using humanoid robots. Frontiers in Psychology,

Svenstrup, M., Tranberg, S., Andersen, H. J., and Bak, T.

(2009). Pose estimation and adaptive robot behaviour

for human-robot interaction. In 2009 IEEE Interna-

tional Conference on Robotics and Automation, pages

3571–3576.

Szaﬁr, D., Mutlu, B., and Fong, T. (2015). Communicating

directionality in ﬂying robots. 2015 10th ACM/IEEE

International Conference on Human-Robot Interac-

tion (HRI), pages 19–26.

HUCAPP 2020 - 4th International Conference on Human Computer Interaction Theory and Applications