Novelty Processing in Human Robot Interaction

fencinghuddleΤεχνίτη Νοημοσύνη και Ρομποτική

14 Νοε 2013 (πριν από 3 χρόνια και 1 μήνα)

51 εμφανίσεις

Novelty Processing in Human Robot Interaction
B.A.MacDonald and T.H.J.Collett
Department of Electrical and Computer Engineering
University of Auckland,New Zealand
fb.macdonald,t.collettg at auckland.ac.nz
April 2,2004
Abstract
Our interest is pragmatic,driven by our engineering research culture;we want
to break down the barriers to human-robot interaction,enabling people to commu-
nicate their wishes to subservient robots.It is notoriously dicult for humans to
get machines and computers to understand and carry out tasks on behalf of the
users.Normally many hours,much expertise and considerable funds are needed
to coax a computer or robot to do something new for us,or even to customise
an existing task.The diculty is the nature of human cognition;we understand
very little about how humans represent and process information during both the
teaching and the learning of new tasks.
In this paper and presentation we will analyse the nature of modern human-
robot communication,survey the modalities and techniques available for such com-
munication,discuss the role of novelty,and outline our thoughts on how to design
a suitable robot interaction system with novelty.
Novelty may be disruptive at low levels of robot interaction,while novelty is
desirable at the higher levels,such as social interaction.
1 Introduction
Novel events are important for any automated and/or intelligent system.Novelty rep-
resents new events and changes,which are likely to be a challenge to any programmed
intelligence.In robotics we see four kinds of novelty:
 In the environment itself,for example new objects.
 When the robot processes the environment the robot may have novel experiences,
for example it may travel a path it has not travelled before.
 When humans interact with the robot they may have novel experiences,for example
saying or doing new things.
 In the cognitive model of the robot itself.The robot's internal controller may reach
a novel state,for example the robot may count say 32 objects,where it has never
counted that many before.
1
There are some important questions that novelty raises,including:
 How can truly new events mean anything to a programmed system such as a robot?
Doesn't there have to be something recognisable so that the robot can respond based
on previous experience or programmed behaviour?
 Is a novel event novel to the robot,the human,or both?
 Does novelty have a useful,functional role in robotics?Or is it just something we
can talk about?
In this presentation we will summarise some of our interest and ideas in robotics and
how these relate to novelty.
We begin by introducing our goals,mention the importance of instructing robots,and
examining the dierent ways that novelty is important.We then discuss human-robot
interaction (HRI) in further detail.
2 Designing and building robots
As engineers our goals are about solving problems for people;about creating artifacts
that have some utility for a client.Our interest in novelty processing is pragmatic,driven
by our engineering research culture.
Our work will be successful if robots are useful assistants for people.We want to
break down the barriers to HRI,enabling people to communicate their wishes to sub-
servient robots.It is notoriously dicult for humans to get machines and computers
to understand and carry out tasks on behalf of the users.More generally,it is di-
cult for humans to tell any computer what to do;that is why software is so dicult,
time-consuming and expensive to build,yet still full of bugs.Describing complicated
behaviour is not easy,especially when you must specify all the details to a dumb ma-
chine.Our goals,robots and recent work can be seen on our robot group web site at:
http://www.ece.auckland.ac.nz/~robot.
The diculty on the human side is the nature of human cognition;we understand
very little about how humans represent and process information during both the teaching
and the learning of new tasks.
3 Instructing robots
To become useful assistants robots will need to take instructions from humans,not just
to fetch this and carry that,but to perform substantial tasks on behalf of the human,at
a later time.For example,humans may need to instruct robots to cook a meal,to clean
the house in a particular way,to photocopy a set of documents in a particular way,and
so on.
So this is one of our interests;how do you get a robot to internalize a description of
a task.For this to happen:
 there must be a way for the human to describe the task
2
 there must be ways for the robot to understand and generalise the task,so that it
can cope with variations
Some basic skills will be built in,but there will always be instructions to give.So
the interaction between the human and the robot is crucial;there must be enough of a
match that important information about the task can be passed across.Generally it is
recognised that this requires a signicant perceptual overlap between the human and the
robot;the perception of important events in the environment must be shared.Also,the
robot and human must perceive each other in an adequate fashion.This is a realistic
problem;robots use lasers and sonar sensors to measure the distance to objects around
them.Humans don't use those sensors.One of our interests is in augmented reality,
where the real world is augmented so that the human teacher does directly experience
the robot's sensations.For example laser ranging lines may be projected in a head
mounted display showing the real world.
We are interested in two methods for communicating tasks:
 description in an expressive computer language that is very suitable for humans
 indirectly by demonstration (you showthe task and the robot watches and interprets
it;the robot creates its own description based on the interpretation)
One of the key factors in both is where the human and robot focus attention during
the interaction.When a robot watches a human there are many things happening;of
all the events and objects that can be seen,what is important for the task description?
When a human describes the task in some kind of computer language,it is important
that the human describes the things that are signicant.
4 Novelty in a robot
This is where novelty processing comes in.For us it is the novelty that answers the
questions above.When teaching a robot a task,it is the novel things that must be
focused on.
Studies of formal instruction from a teacher to a human student show this.Kurt van
Lehn [13] used the idea of\felicity conditions"to explain the way formal lessons work.
For example when children are taught subtraction,presented as a sequence of lessons.In
each lesson it is important that all steps are shown,that invisible objects and relations
are made visible,and each lesson introduces exactly one new part to the description.He
built a learning system based on these constraints.The felicity conditions are important
because the learner understands the rules,which help it to develop a better description
for each lesson.If there is exactly one new thing to add,and there is nothing hidden,
then it is easier to narrow down where the new part has to be added and what it is.
No need to consider adding two interacting things.For example,when you see a lesson
about borrowing for subtraction;the description has to have just the one thing added,
about borrowing.
It is the novel thing that must be focused on.The novel things a human teacher gives
to a robot must be recognised as novel.
3
For example,how do you teach a robot to clean up your living room?Well the robot
will come with some basic skills,for picking things up,for nding certain known objects,
for wiping surfaces,etc.But this will not be all that the robot needs to know.You will
need to tell it how to put these pieces of behaviour together to carry out the whole task.
A good way to do this is to show the robot what to do,either by controlling it yourself,
or by having it watch what you do.
What will be important in each lesson is what the novel things are.How do we
get a robot to focus on the right things?How do humans get other humans to focus
correctly?Well humans share the same visual system and are sometimes will focus on
similar things,but robots don't have this advantage.Humans use language (\Make sure
you check behind the sofa.") and gestures such as pointing to direct the attention of
fellow humans.How can we do this with robots?We need mechanisms for robots to nd
novelty in their experience,and to have it directed by humans they interact with.
For example,the rst time the robot encounters wine glasses on the oor while clean-
ing the living room,the important thing is that it notices this new occurrence,interprets
it,asks for help if necessary,and incorporates some new actions in its task description.
It must recognise the novel occurrence and respond appropriately.
4.1 John Andreae's novelty processing
In his book\Associative Learning for a robot intelligence"[1],John Andreae found
novelty to be important for additional reasons.
Firstly,what is novelty?Novelty in John's system works like this.The robot stores its
experience as associations.The rst time an association is stored it is marked as a novel
one.When that association is experienced again (either in fact or in a plan) the novelty
mark is removed.So seeking novelty is seeking associations that have been experienced
only once.
For example,this is what encourages a\baby"robot to explore its world,seeking out
new experiences because its action choices are biased in favour of repeating experiences
seen only once.This takes it to new places,to explore and gain further novelty,until
that experience is fully explored and renders no new novelty.
Novelty can't be something completely new,because it isn't until associations are
stored that they can aect the robot's behaviour.
Seeking novelty helps a robot explore an unfamiliar environment.Exploration is
important so that you can discover how to better meet your goals.
Seeking novelty gives a robot one way to set its own goals,to escape the goals the
designer builds in.Whether something is novel or not depends on the previous robot
experience,and in the sense that the robot is its experience,it sets its own goals.
John's robot system also exhibits boredomand frustration.As with Novelty,these are
given mechanistic interpretations in John's low level associative robot model.Boredom
prevents endless loops (which computers seems to get in to easily but humans do not
persevere with).Frustration prevents unnecessary loss of novelty goals.Encountering
novelty (or reward) resets the boredom and frustration parameters to zero.
Boredom increases gradually if the expectation of novelty decreases (so that endless
loops,which cannot continually provide novelty,cause boredom to increase).If boredom
increases too much,the robot changes it tendencies for selecting actions;it starts selecting
4
User
Robot
Interface(s)
User Roles
- Developer Interaction
- Operator Interaction
- Social Interaction
- Cybernetic Interaction
Autonomy Mode
- Fully Autonomous
- Director-Agent
- Human-Robot Cooperation
- Direct Control
Figure 1:HRI Summary
alternative actions rather than the\best"ones,so that it can break out of the loop.
Frustration rises when plans do not reach the expected novelty.If frustration is too
high (meaning the robot can't reach the novelty is it chasing),it starts making random
actions.This prevents the novelty marks being removed when the robot cannot reach
novelty;well some are removed but not all.
The next section examines the interaction between humans and robots in more detail.
5 Human Robot Interaction
HRI can be divided into three key components;the human users,the robots autonomy
mode and the interface between the two,see Figure 1.Here the term interfaces describes
the medium that the two sides use to communicate,such as speech,gesture or visual
cues.
5.1 User roles
In order to promote successful integration all three components need to be considered.In
order to select the interfaces to be used and how much and what intelligence to program
into the robot an appropriate user model needs to be chosen.In general the type of
interaction system is driven by the role that the developer takes during the interaction.
In general the dierent interactions can be comprised of four basic user roles;developer
interaction,operator interaction,social interaction and cybernetic interaction.A user
may take on one or more of these roles while interacting with a robot,however at any
one time one role will generally be dominant.The important feature of this division is
that the needs of each role can be treated distinctly,simplifying the design of the system.
Table 1 gives a summary of the roles,typical tasks that would use the role,and the
interface priorities.
Developer Interaction which is the role used during the programming,testing and
debugging of the robot.The priorities for this interface are ecient ways of presenting
5
Role
Typical Tasks
Interface Priorities
Developer
Interaction
Robot design,testing and de-
bugging
Large output bandwidth,e-
ciency and precision
Operator In-
teraction
Industrial automation,task
programming
Precision and Eciency
Cybernetic
Interaction
Prosthetic limb
Seamless interface with body
Social Inter-
action
Reception robot,Guide robot
Natural and intuitive interface
Table 1:User Roles
large amounts of information about robot and world states while allowing precise control
of the robot's operation.
Operator Interaction which is based around industrial automation,where humans
set up and modify robot tasks.It is also applicable to robot assistants for medical work,
for home assistant robots,and other task level programming.In this role the interaction
should be ecient and precise.
Cybernetic Interaction where the robot is an extension of the user's natural body,for
example a prosthetic limb or exoskeleton.The interaction must be completely seamless
with control being exactly as if the user was making the movement with their own body.
Social Interaction where the interfaces between the user and the robot should be
natural and intuitive,similar to communicating with another human.The tasks to be
undertaken are likely to be relatively simple from a human point of view,for example
guiding humans around an exhibit,or managing a reception desk.
For the rst three types of interaction (Developer,Operator and Cybernetic),the
interface needs to behave predictably,any novelty in the system becomes a hindrance to
its use.The cybernetic role takes this to the extreme,an arm that moves to the right
when you want it to move left would certainly be novel,but also completely dysfunctional.
The social interaction user role has two con icting requirements,the system must be
recognisable so that the user can relate to it,often through projection of either animal or
human characteristics,and at the same time the behaviour of the robot must have a degree
of novelty so that its behaviour is socially engaging.This concept of social engagement
is particularly important for entertainment robots where predictable behaviour leads to
disinterest very quickly.
5.2 Robot Autonomy Mode
The selection of the robot's autonomy mode depends mainly in the application the sys-
tem is being designed for.There is a continuum from direct remote control of the robot
6
Interfaces
Traditional
Remote
Gesture
Speech
Gaze
VR & AR
Emotion & Personality
Facial Features
Body Language & Gesture
Speech (Intonation)
Demonstration (Implicit)
Haptic
Imitation
Direct Interfaces
Indirect Interfaces
Figure 2:Interface Methods
through to complete autonomy.For tasks such as bomb disposal direct control is of-
ten appropriate.For a robot designed solely to interact with passing users.complete
autonomy is ideal.
The length of time where novelty remains in a system is related to the level of au-
tonomy in the system.A system that is completely under direct control has very little
long term novelty as the responses of the system will always be the same to a repeated
sequence of commands.For a system with complete autonomy the response will depend
on the internal state of the robot,leading to possibilities for rich interactions.
5.3 Interfaces
Robot interfaces can be divided into two groups,direct interaction,which results in a
command or explicit piece of information being transferred,and indirect interaction where
the information transferred is not explicit,such as emphasis,mood and context.Indirect
interfaces are generally used in conjunction with direct interfaces to make the interaction
more natural,or convey deeper meaning,such as emotion.Figure 2 shows the range of
interfaces currently used in HRI.
The selection of the interfaces for a particular systemdepends on the application,both
to determine which interfaces are appropriate to the users and also to the environment
that the robot will be working in,for example loud audio would not be appropriate for a
museumguide robot and a complex text based command systemwould not be appropriate
for a reception robot.
5.3.1 Direct Interfaces
Traditional Interfaces consist of keyboards,mice and joysticks for input and a com-
puter monitor for output.One of the drawbacks of traditional interfaces is that many
robots are mobile and thus require the ability to interact when there are no computer
terminals available.Wearable computing and PDA's provide possible solutions,however
this still restricts who can interact with the robot and so a range of alternative interfaces
are being explored that provide mobile,hardware free interaction;for example gesture
and speech.
7
Other traditional interfaces include audio alarms and warning lights,these provide a
small amount of information and so their use is limited in complex systems.
Remote Interfaces typically consist of a client system with a traditional computer
interface linked to the robot via a network link such as a LAN or the internet.Remote
interface methods may be used by teleoperated robots,semi{autonomous task based
systems,and to provide feedback or conguration with fully autonomous systems,for
example streaming video and map data from an exploration robot.
Acommon application of remote interfaces is remote exploration of museums,libraries
[12] and art galleries by avatars [7].They are also used for search and rescue scenarios,
bomb disposal and other dangerous environments.
The main challenges of remote interfaces relate to the time delays inherent in any
communication link.In the extreme case of remote command of space robots (such as
the Mars rovers) this delay can be hours.In these cases the concept of\hand to eye"
coordination is redundant and low level control must be delegated to software agents
running on the robot itself.
Gesture and hand posture can be used as a direct control method where each gesture is
mapped to a command.The challenge with hand gesture systems is to accurately capture
the hand position.Image based techniques require signicant processing power and have
diculty with changing light conditions and hand becoming obscured,either partially by
their own form or by other obstacles in the environment.Instrumented gloves overcome
these problems but are often cumbersome and expensive and have limited freedom of
movement if position measurement is necessary.
Speech is a fundamental communication method between humans and so is a powerful
interface between humans and robots,particularly for users unfamiliar with the robot
system.Current research focuses on both the interpretation and the generation of speech;
both the meaning and context of natural language are important.For example,a street
map navigation robot [3] accepts speech commands fromusers such as\take the 3rd left."
Gaze direction has been used to directly control robots.This is particularly useful for
users with limited mobility,such as the elderly or disabled.Yoo et al use IR LED's to
estimate gaze direction for robot control by elderly people without accurate arm control
[14] and Law et al use electrodes to measure gaze direction for the control of an electric
wheelchair [4].
Virtual and Augmented Reality provide an articial 3D world for the users to
interact with.With virtual reality (VR) the user is immersed in the virtual world and is
able to move around freely.The capabilities of the system and abilities given to the user
vary between dierent implementations,with some systems including accurate physics
and force feedback to the user and others allowing users to y through the world with no
restrictions.Augmented reality (AR) uses similar concepts but here the users real world
is augmented by adding virtual objects to the users vision.Movement within the world
is restricted to what the user is physically able to do.
8
Both VR and AR have a range of hardware setups but typically feature as a minimum
a head mounted display,and an inertial head tracker.Additional components include
position trackers,data gloves and a variety of hand props that allow the user to feel as if
they are holding whatever tool they are using in the virtual world.
An algorithm evaluation system is presented in [11],which uses robots in a VR world
to test the performance of avoidance algorithms when human users are involved.The
main advantages are that it uses a real human rather than a scripted one,thus giving
more accurate (less predictable) behaviour while at the same time using virtual robots
so there are no safety issues.
Imitation is an important source of information for humans,likewise there has been a
lot of interest in transferring this learning technique to robots.When using imitation for
training a robot one of the key problems is selecting what to imitate,and the solution to
this is to focus on the novel aspects of the'performance'or on removing the novelty.
Pollard et al [9] present a system for full body imitation that attempts to maintain
the individuality of the performance within the limited movement space of the robot,this
is particularly relevant for programming performances for entertainment robots.In the
system movements that are beyond the physical abilities of the robot cause the entire
performance to be scaled so that the relative movements are still true to the original,
maintaining its individuality.
Billard et al take the opposite approach [2],specically the system examines repeated
manipulations and selects the invariant aspects,which it then reproduces.
5.4 Indirect Interfaces
Emotion,Personality,Facial Features,Body Language and Intonation play
an important role in making a robot socially acceptable,particularly in long term in-
teractions.If a robot's behaviour is predictable then social interaction quickly becomes
boring.A variety of research robots have been created that attempt to reproduce human-
like emotions and personality,generally focusing on facial features.
Emotion or aective state is also a useful input from the user,allowing the robot to
respond to the users emotional state appropriately,for example when the user is fatigued
or stressed.
Demonstration (Implicit Communication) is useful when the robot does not have
the capacity for direct communication.Nicolescu and Mataric [8] implemented a mobile
robot that requests help by demonstrating failed task attempts in the presence of a user.
Haptic interfaces are generally used to provide a feedback path to the human opera-
tor in teleoperated environments.They are particularly important for tasks that involve
precision movement and grasping.Haptic interfaces have huge potential in medical ap-
plications,where subtle changes in texture can be very important to the surgeon during
diagnosis and surgery,for example when removing tumours [10].
Haptic interfaces are useful for remote operation as they transmit information about
the physical environment to the remote operator.Lo et al [6] describe a system for
9
capturing the tactile contact of a robot nger and displaying it on a tactile display so
that a remote user can sense what the robot is touching.
Lin et al describes the use of haptic interfaces to enhance the use of computers in
`creative processes'[5] by sensing and displaying attributes such as stroke pressure.
5.4.1 Multi Modal Communication
In order to achieve eective and ecient interaction between humans and robots several
or all of the above interfaces must be used in a single robot,as humans use many com-
munication means simultaneously.The use of several overlapping interfaces also allows
for more robust communication as when one interface breaks down another can be used.
5.5 Novelty in Interfaces
In general the direct interfaces are designed to convey a predictable ow of information
between the robot and thus novelty is unwanted and great lengths are taken to avoid
it.Indirect interfaces have a lot more use for novelty,this is particularly the case in
robots that are designed to interact in social situations.A robot which uses xed rules
to emulate personality and emotion quickly becomes articial and boring,in order to
remain socially engaging the robot must be capable of novel responses to its interaction
partner and the environment.
The need for novelty in the environment is common in all areas of human existence.
Humans have evolved to respond to changes in the environment and not to a lack of
change.Attention is seldom attracted by a stationary object,yet we can pick up even
small movements far out in our peripheral vision.Similarly air conditioning manufac-
turers have found that people are far more comfortable if there are small but random
variations in the temperature of a room.
Novelty is also particularly important for speech as an interface.In processing speech
a robot must search for the invariant aspects of the sound stream,extracting the known
words despite variations in pitch,volume and pronunciation.The opposite is true when
synthesising speech,here attempts are made to reproduce the personal variations that
exist between one human and another,adding novelty to the raw speech stream.
An important,yet simple way for a robot to deal with novelty in its environment is
through using implicit communication.If an unexpected obstacle appears in the robots
environment that it is unable to overcome it can simple demonstrate failed attempts to
surmount the obstacle and wait for a passing user to notice and remove the obstacle.
This is similar to the way a pet will stare under the couch after a lost toy until their
owner fetches it out for them.
6 Conclusion
Our goal is to break down the barriers to human-robot interaction,enabling people to
communicate their wishes to subservient robots.In this paper we have discussed the
nature of modern human-robot communication,and outlined some thoughts on how to
design a suitable interaction system for robots.
10
Novel events are important for any automated and/or intelligent system.Novelty
may appear in the environment,the robot's experiences,the human's experiences,or
internally in the robot controller.
Novelty is an important aspect of robot design,because it represents the signicant
new aspects in new lessons and interactions between humans and robots.A robot's
attention should focus on novelty for example.
John Andreae's work shows how novelty may be used to allow a robot to set its own
goals and to encourage it to explore its environment,without causing the robot to enter
an innite behaviour loop.
For direct control of robots novelty is an unwanted complication.For social interaction
and indirect communication it is an essential consideration.In order to carry out tasks
in the real world a robot must be aware of novel events and make decisions on their
relevance to the task at hand.
References
[1] John Andreae.Associative learning for a robot intelligence.Imperial College,1998.
[2] Aude Billard.Robota:Clever toy and educational tool.Robotics and Autonomous
Systems,42(3-4):259{269,2003.
[3] T.Kyriacou,G.Bugmann,and S.Lauria.Vision-based urban navigation procedures
for verbally instructed robots.In Proc.IEEE/RSJ International Conference on
Intelligent Robots and System (IROS 02),volume 2,pages 1326{1331,2002.
[4] C.K.H.Law,M.Y.Y.Leung,Y.Xu,and S.K.Tso.A cap as interface for wheelchair
control.In Proc.IEEE/RSJ International Conference on Intelligent Robots and
System (IROS 02),volume 2,pages 1439{1444,2002.
[5] M.C.Lin,W.Baxter,M.Foskey,M.A.Otaduy,and V.Scheib.Haptic interaction
for creative processes with simulated media.In Proc.IEEE International Conference
on Robotics and Automation (ICRA 02),volume 1,pages 598{604,2002.
[6] Wang-Tai Lo,Yantao Shen,and Yun-Hui Liu.An integrated tactile feedback system
for multingered robot hands.In Proc.IEEE/RSJ International Conference on
Intelligent Robots and System (IROS 01),volume 2,pages 680{685,2001.
[7] S.Maeyama,S.Yuta,and A.Harada.Remote viewing on the web using multiple
mobile robotic avatars.In Proc.IEEE/RSJ International Conference on Intelligent
Robots and System (IROS 01),volume 2,pages 637{642,2001.
[8] M.N.Nicolescu and M.J.Mataric.Learning and interacting in human-robot domains.
31(5):419{430,2001.
[9] N.S.Pollard,J.K.Hodgins,M.J.Riley,and C.G.Atkeson.Adapting human motion
for the control of a humanoid robot.In Proc.IEEE International Conference on
Robotics and Automation (ICRA 02),volume 2,pages 1390{1397,2002.
11
[10] A.E.Quaid and R.A.Abovitz.Haptic information displays for computer-assisted
surgery.In Proc.IEEE International Conference on Robotics and Automation (ICRA
02),volume 2,pages 2092{2097,2002.
[11] R.Shikata,T.Goto,H.Noborio,and H.Ishiguro.Wearable-based evaluation of
human-robot interactions in robot path-planning.In Proc.IEEE International Con-
ference on Robotics and Automation (ICRA 03),volume 2,pages 1946{1953,2003.
[12] T.Tomizawa,A.Ohya,and S.Yuta.Book browsing system using an autonomous
mobile robot teleoperated via the internet.In Proc.IEEE/RSJ International Con-
ference on Intelligent Robots and System (IROS 02),volume 2,pages 1284{1289,
2002.
[13] Kurt VanLehn.Learning one subprocedure per lesson.Articial Intelligence,31:1{
40,1987.
[14] Dong Hyun Yoo,Jae Heon Kim,Do Hyung Kim,and Myung Jin Chung.A human-
robot interface using vision-based eye gaze estimation system.In Proc.IEEE/RSJ
International Conference on Intelligent Robots and System (IROS 02),volume 2,
pages 1196{1201,2002.
12