澳洲dissertation网：Explicit Task Representation based on Gestu

Explicit Task Representation based on Gesture InteractionChristian Müller-Tomfelde and Cécile Paris
CSIRO - Information and Communication Technologies CentreLocked Bag 17, North Ryde, NSW 1670, [email protected] and [email protected]
Abstract
This paper describes the role and the use of an explicittask representation in applications where humans interactin non-traditional computer environments using gestures.The focus lies on training and 澳洲dissertation网assistance applications,where the objective of the training includes implicit
knowledge, e.g., motor-skills. On the one hand, theseapplications require a clear and transparent description ofwhat has to be done during the interaction, while, on theother hand, they are highly interactive and multimodal.Therefore, the human computer interaction becomesmodelled from the top down as a collaboration in whicheach participant pursues their individual goal that isstipulated by a task. In a bottom up processing, gesturerecognition determines the actions of the user by applyingprocessing on the continuous data streams from theenvironment. The resulting gesture or action is interpretedas the user’s intention and becomes evaluated during thecollaboration, allowing the system to reason about how tobest provide guidance at this point. A vertical prototypebased on the combination of a haptic virtual environmentand a knowledge-based reasoning system is discussed andthe evolvement of the task-based collaboration becomesdemonstrated.1
Keywords: task model, collaboration, gesture interaction,gesture recognition, virtual environment.
1 Introduction
The user interaction in virtual environments follows aspatial paradigm, while in traditional computer desktopenvironments the notion of Windows, Icon, Menus andPointing interaction (WIMP) is prevailing. http://www.ukthesis.org/Thesis_Writing/Computer_Science/Theinteraction in virtual environments provides an intuitiveaccess to simulated objects. Simulations have achievednowadays considerable realism or plausibility, and thefocus can be shifted towards more quality in the
interaction (Smith et al., 1999). On the one hand, researchhas been done to specify new software models andspecification languages to cope withnon- and post-WIMP user interfaces (Jacob, 1996 and Beaudouin-
with which traditional User Interface ManagementSystems (UIMS) are not designed to deal. On the other
Copyright © 2006, Australian Computer Society, Inc. Thispaper appeared at the NICTA-HCSNet Multimodal User
Interaction Workshop (MMUI2005), Sydney, Australia.Conferences in Research and Practice in InformationTechnology, Vol. 57. Fang Chen and Julien Epps, Eds.Reproduction for academic, not-for profit purposes permitted
provided this text is included.
hand, effort has been made to populate a virtualenvironment with additional information to enrich theenvironment. In an early work of Feiner et al. (1993) onaugmented reality systems, additional task-relevant#p#分页标题#e#
graphical information is displayed as an overlay to thereal environment. In contrast, in the approach of Bowmanet al. (1999), additional information becomes integratedat appropriate spatial locations for the user to recognise,memorise and with which to interact. In general, explicit
knowledge representation can be used to improve theinteraction, e.g., to direct and assist the user in virtual
environments (Aylett and Luck, 2000; Jung et al., 1998).In the approach of Bowman et al. (1999), additionalinformation appears when the user interacts withartefacts. We propose now, to not merely provideinformation about objects, but also make use of theexplicit knowledge about what the user wants or has to dowith the object. This knowledge about the task of the userallows the system to deliver to the user meaningful andrelevant information at the right time. This can beunderstood as supporting the user’s interaction at aconceptual level of interaction (Massink and Faconti,2002).
We developed a vertical prototype for training motorskillsusing a hand immersive haptic virtual environment.
The interaction environment is enriched with informationand makes uses of a comprehensive context model. Theprototype shows how a formal organisation of theinteraction in layers can be used to integrate differenthigh and low level components. The interaction of theuser can be characterised as parallel, continuous andmultimodal at the physical and perceptual level (Nigayand Coutaz, 1993), while, at a higher level, the actionsand goals are represented and processed in a serial,discrete and symbolic style. The user interaction withartefacts in the environment follows the metaphor of thecomputer as a tool or media, while on the task level, thecomputer can be understood as a dialogue partner
(Schomaker et al., 1995, Maybury and Wahlster, 1998).Finally, since the resulting system addresses issues of
training and assistance, the described approach can becompared to those of Intelligent Tutoring Systems (ITS),
where a computer is acting as a tutor, trainer or assistantand can collaborate with the user on tasks in simulated
environments (Rickel et al. 2000). The majority of
existing ITS are addressing the issue of teaching explicit
knowledge as, e.g., described in Core et al. (2000). In
contrast to that, the approach described here concentrateson training implicit or tactic knowledge, e.g., motorskills.In the following section, the paper describes a layeredinteraction model for information enriched interaction.
2
Then, in section 3, a detailed description of the used taskrepresentation is given. In section 4, a sensor network is
described that provides the system with symbolicinformation about the user interaction. Finally, thevertical prototype of the integration of a haptic virtualenvironment and a knowledge-based system is describedand the paper ends with a conclusion.
2 Information enriched interactionThe foundation for the support of user interaction in#p#分页标题#e#
virtual environments has been already established byintroducing interactive landmarks (Müller-Tomfelde etal., 2004). Interactive landmarks function as a metamodel in order to deal with the different models of therepresentation of an object in a three dimensional
interactive virtual environment and that in a knowledgebasereasoning system. This meta model avoids creating anew model with both representations. The landmark inthe virtual environment has the function of an annotationof the scene-graph, for example, the knob of a doorbecomes annotated with an interactive landmark tospecify a point of user interaction and to reason whetherthe user grabs the knob or not. The landmark enables themapping of objects in the scene with those in the domainmodel of the reasoning system. This approach overcomesthe structural barrier of virtual environments andknowledge-based reasoning systems and avoidsreinventing the wheel, by making use of existing
implementations.2.1 Layered interaction model
It is supposed that the virtual environment is a genericenvironment, in which the user can interact with any sorts
of artefacts. Each of these artefacts could host interactivelandmarks. How the interaction in the environment is
technically realised and implemented is not is the scopeof this paper. Instead, only the user’s actions and thespatial relation of the landmarks are relevant andconsidered for the interaction (Müller-Tomfelde et al.2004). A layered model for human computer interactionhas been chosen. It makes use of the reference frameworkfor continuous interaction, as proposed by Massink andFaconti (2002).Collaboration layer
Task layerPropositional layer
Perceptual layerPhysical layer
Figure 1 The layers of the interaction model based on
the reference framework for continuous interaction
(Massink and Faconti, 2002).
This framework for interaction addresses issues of
continuous and parallel interaction while, at the same
time, attempting to reduce the design complexity of
澳洲留学生计算机专业dissertation定制continuous interaction systems. Each layer has clear and
distinct functions and passes information up and down in
the model or framework, respectively. In the systemdescribed in this paper, issues of continuous and parallelinteraction are confined to the lower levels of thereference framework, where physical data streams areprocessed. Once these data streams have been
transformed and interpreted into a symbolicrepresentation, the interaction is characterised by serial
and discrete processing of events, actions or states. As itwill be discussed later, this boundary becomes less ridgedthan it may appear at this point.In the physical layer, the interaction between the
environment and, e.g., human sensors happens. Physicalsignals are perceived by the sensors and transformed toprovide means at the perceptual layer. In the followingthe physical layer is not discussed since this would#p#分页标题#e#
address the issues of psychophysics (like, e.g., describedand discussed in Schomaker et al. 1995) and would gobeyond the scope of this paper.
2.2 Collaboration and task layer
The top level of the interaction model is the collaborationlayer. The collaboration management organises the
activity of the participants in the collaboration. In a firstand simple realisation of this organisation, a token ispassed periodically to each participant and enables themto ‘speak’, while all others are ‘listening’. This timeperiod of ‘speaking’ in the collaboration is defined to bethe turn of the participant. It has to be noticed here thatthe organisation of the collaboration at the collaborationlayer is independent from the content of the task of eachparticipant.The task is to be understood as a structured sequence of
primitive tasks or actions, that have to be performed inorder to achieve the overall goal of the task. For example,the task to open a door becomes decomposed by thesequence: approach the door, grab the door knob, thenturn the knob and push the door to open. Once aparticipant has the token, she pursues the task. The nextstep to do or the next goals to achieve in the situation isdetermined. It is conceivable that multiple executions of
the same task result in different sequences of taskprimitives, but the overall goal of the task remains thesame. This is due to the fact that a primitive task can befollowed by a choice of other primitive tasks. Which next
primitive task becomes selected for further processingdepends on conditions and the actual state of the context
of the task. In other words, multiple ways can lead to thesame task goal, depending on the level of detail of the
task description.The task representation is declarative and does not
contain any element that supports processing in a
collaboration. Therefore a structure called task
environment enables the management of the task
processing and the storage of relevant information about
the task during its execution. It contains, for example, the
task name, the actual task primitive, the task processing
mode, etc. Each participant of the collaboration owns a
task environment and hence has a task to operate on
3
during the turns in the collaboration. The task
environment also provides means for the collaboration
management, e.g., to signal a ‘claim’ of a participant to
‘speak’ in the collaboration.
2.3 Propositional and perceptual layers
Once the next task primitive has been selected, the
subgoal of this primitive becomes further planned top
down in a content and presentation layer (equivalent to
the propositional and perceptual layer, see in Figure 1).
The execution of this plan then delivers information on
the physical layer, e.g., to a visual display device or a
text-to-speech generator. If the goal is a communicative#p#分页标题#e#
goal, then the planning and the execution of the plan
leads to an ‘utterance’ and can be used, e.g., to create
text, as described in Moore and Paris (1993).
In case the computer functions as a trainer or assistant,
the task of the trainer becomes executed by the computer,
comparable to the execution of a computer program. But
in contrast to a programming language, the task
description is a specific language that supports the
efficient description and authoring of tasks (more about
this in section 3.1). The computer, in the role of a trainer,
has the task to provide appropriate instructions and
feedback to the user performing the task. This can be
understood as a communicative goal which becomes
processed by the content and presentation layer. This top
down processing from the task layer to the perceptual
layer enriches the environment at the physical layer by
delivering tailored information to the user, based on the
current state of the task. In a bottom up processing, from
the perceptual layers to the propositional, the user’s
action becomes detected by refining the data streams
from the environment and obtaining a reasonable
representation of the current goal of the user (see section
4.1). This user goal gets evaluated to determine whether
the goal matches with the stipulated task primitive or not.
3 Hierarchical task description
It is assumed that the user’s mental model of A) the
environment, the virtual environment in which the user is
physically interacting and B) the training situation, in
which the user is in collaboration with a trainer, are well
established (Norman 1986). In other words, the user
knows in what and with whom she is interacting. Now,
we focus on the model of the topic of the training or
assistance: the task of the user or the domain task. The
domain task is a task that refers to a certain domain and is
not further specified for the following considerations.
Both the author, who creates the domain task and the
http://www.ukthesis.org/Thesis_Writing/Computer_Science/user/trainee who performs or wants to learn this task,
must have an unequivocal understanding about the
represented task. This is considered to be a fundamental
requirement of successful applications for training and
assistance. Therefore, a hierarchical task representation
based on a formalism that supports task annotations and
procedural relationships (Tarby and Barthet, 1996) was
selected for the task description. Advantages are efficient
authoring and learning, as well as easy conceptualisation
(Lu et al., 2000). The use of a task model has additional
advantages at various stages in the Software
Development Life Cycle, analysing the task of users, but#p#分页标题#e#
also describing user interaction tasks in virtual
environments (as for instance in Murray and Fernando,
1999). The domain task is graph-based represented and
can become decomposed in a tree-like structure of
primitive and composite tasks. The latter become further
decomposed in their parts (see an example task
representation in Figure 2). Once the model has been
created, annotations and mapping functions enable further
processing of the task model.
3.1 Domain task and tutorial task
A training situation is considered to be a collaboration in
which a trainee is directed or taught by a trainer to do a
domain task. Since participants of the training session are
represented at the collaboration layer in the same manner,
not only the trainee has a domain task, but also the trainer
or assistant has a task, the tutorial task. This tutorial task
is represented in the same way as the domain task and is
explicitly available for processing by the computer. The
goal of the tutorial task is to train the user performing this
domain task.
Figure 2 An example for a simple domain task
This reveals that both tasks need to be processed
differently: On the one hand, the domain task is a
description of what the user wants or has to do and
becomes referred by the tutorial task. While, on the other
hand, the tutorial task becomes executed on the computer.
The separation between the tasks is a consequence, when
considering the interaction in the environment as a
collaboration of multiple participants. Both tasks have to
be created by authors with expertise in the specific
domain and in tutoring. Furthermore, to increase the
impact of the explicit tutorial task it must be independent
form the domain task to achieve, e.g., a high reusability.
The explicitness of both tasks allow advanced processing,
e.g., for documentation and transformation into other
representations.
The tutorial task is implicitly defining the structure of the
collaboration as a training session. A possible training
session has a three-tier structure and consists of :
• Introduction: the trainer introduces the subject of
training to the trainee. This could be, e.g., a list of
subtasks that have to be fulfilled to complete the
task.
4
• Practice: the trainer and trainee are working
cooperatively on their individual tasks. The
collaboration guarantees that each participant in
the collaboration is able to ‘speak’ and operate on
their individual task.
• Summary: the trainer informs the trainee about
his or her overall performance and about the
assessment for the session.
In order to enable the assessment of task performance the
task description not only has to describe explicitly what#p#分页标题#e#
has to be done in a task, but also provide means to decide
whether the goal of a primitive task has been
accomplished or not. An annotation on a task element
called ‘post condition’ is used to reason about the goal
achievement of the primitive task. This allows the trainer
to evaluate the trainee’s performance by examining the
associate post condition of the task element the user is
performing in the domain task. This link or connection
between the domain task and tutorial task allows a sort of
synchronisation of the domain task execution in the
environment and its representation in the computer on the
task layer of the interaction model (see Figure 1). This
interaction can be understood as a conceptual interaction
in contrast to the apparent interaction on the physical
layer.
3.2 Tutorial strategy
During the tutorial practice, the strategy of the tutor how
to teach can change. If, for instance, the candidate fails
performing a sub task, the tutor can give further details
about the actual task or otherwise reduce the amount of
feedback if the user is performing well. The latter case
helps avoiding overload of a good performing candidate
with redundant and irrelevant information. Currently the
tutorial task handles three levels of strategies, which can
be changed by the tutorial ‘reflection’ during the
interaction in the session practice:
• Step-by-step: During the collaboration, each step
in the domain task becomes explicitly announced
to the trainee and comprehensive feedback is
provided.
• Guide: The trainee receives guidance for the task
that has to be performed. Instructions and
feedback are given whenever the trainee seems to
require it. Situations where the trainee is not
continuing in doing at least something, can be, for
example, interpreted as an uncertainty on the
trainee. Appropriate information at that point
might help.
• Rehearse: This tutorial strategy mimics the
situation where the trainee has to perform the
domain task on his or her own. Nevertheless, the
trainer monitors and tracks each step of the trainee
in the domain task to give a comprehensive
summary and assessment at the end of the training
session.
In the course of the training session, the tutorial strategy
influences the way the current instruction or feedback of
the trainer becomes realised. The ‘trainer’ exhibits basic
characteristics of individualised training by adapting of
the tutorial strategy. This could lead to an optimised
development of the learning rate over time, due to the fact
that, on the one hand, minor errors of the trainee do not
block the training progress, while on the other hand, a
performing trainee is not forced to read instructions she#p#分页标题#e#
not requiring. In a first approach, the adaptation process
is designed asymmetrically. The step back to more
verbose strategies, like, e.g., from guidance to step-bystep
strategy requires less errors than it needs right
actions stepping up again.
4 Action and gesture detection
The action and gesture detection is based on sensors,
which transform incoming data streams into more
meaning full information. For example, a stylus functions
as the user representation at the physical level and its
orientation, orientation and the applied force are fed into
the sensors network to detect specific features. The state
of one sensor becomes activated when its detection
condition is fulfilled and all sensors become processed
and their inherent states updated in real-time at a
frequency of 60 Hz.
In this paper the term gesture refers to the user action that
becomes recognised by processing and transforming the
continuous data streams into a discrete and symbolic
representation of the action. The sensor processing is
located in the physical and perceptual layer and leads to a
representation of the user actions, as an intention or goal
at the propositional layer (see Figure 1). In other words,
during the processing of the gesture recognition, a
transition is made from a parallel continuous interaction,
to a serial and discrete symbolic representation in the
upper layer of the interaction model.
4.1 Sensor network
To determine a gesture, it is not useful just looking at the
states of the sensors and pass them on for further
reasoning. Ambiguous situations could occur, if the
relation between the sensor results are not considered.
Therefore, we organise the sensors in a hierarchical
network to establish relationships between them (see
Figure 3).
Figure 3 The sensor network of the action and gesture
recognition. Each node represents a sensor for a
specific action.
In the gesture recognition described in Latoschik (2001),
nodes of a network are connected by a data routing
mechanism. Instead, the connections in the network of
sensors in the approach described in this paper refer to the
conceptual relation between the sensors. On the basis of
the current states of the sensors, a simple tree algorithm
determines a resulting gesture or action that is passed on.
5
The algorithm of the search for a currently intended
action by the user starts at the root of the tree and can be
described as follows: Traverse the tree and return the
activated sensor with no activated subsensors. The
priority amongst multiple activated subsensors at one
level is increasing from the right to the left. Repetitive
detections are ignored. The resulting behaviour
guarantees that only one gesture or action becomes#p#分页标题#e#
passed on per detection cycle. This tree search algorithm
also operates at real-time at 60 Hz, after the processing of
all sensor states have been finished.
4.2 Temporal organisation
The relationships in the sensor network rely on spatiotemporal
aspects, as well as on haptic aspects of the user
interaction. The latter aspect addresses issues of the force
applied by the user during the interaction, like push, pull,
etc. In the simple example case, the user holds the stylus
steady for more then 500 ms and applies a force in the
direction of the orientation of the stylus, then the dwell
and push sensors become activated, but only the push
gesture is passed on. This behaviour is realised by the
above proposed search algorithm and can be understood
as early fusion of the hand position and orientation with
the tactile modality (Oviatt et al., 2000). In another
example the user released the stylus and no movement
occurs anymore, the sequence of detected ‘actions’ is the
following: after 500 ms the dwell action is detected, since
in the next 3 s nothing changes, the stop sensor becomes
activated. The tree algorithm determines stop as the next
‘action’ assuming that no other sensor at that level in the
tree is activated. Finally, after another 15s the halt sensor
gets activated. The used terminology is unfortunately
inexact, since dwelling, stopping, etc. can not be
considered as a user action; it is rather an absence of
action. A more appropriate terminology is under study,
following, e.g., the taxonomy of Bobick (1997).
4.3 Complex actions and gestures
So far, only basic actions or gesture like approach, point,
pull, push and slide have been implemented. As depicted
in Figure 3, an alternative to the dwell sensor is a move
sensor. This branch in the sensor tree could serve to
detect more motion oriented aspects of the user
interaction like, translation, a linear trajectory in space,
or rotation, a circular or curved trajectory in space. Since
these sensors require more development, their realisation
is postponed until application scenarios demand them. To
illustrate further possibilities based on these sensors, we
now describe the detection of a more complex gesture:
assumed that the user’s tool is a knife, and the user moves
the tool forward in a linear manner, while at the same
time applying force perpendicular to the blade of the
knife. The user’s action could then be detected in the
sensor tree as a slice action. Furthermore, multiple
consecutive slice actions with alternating translation
direction could be grouped and constitute a cutting action.
Such a complex gesture recognition would be the result
of a late fusion, where a sequence of actions over a#p#分页标题#e#
couple of seconds become grouped and classified as one
action or gesture (Oviatt et al., 2000). The different
complementary information of the different sensors over
time becomes integrated at a sematic level into a single
action or intention.
4.4 Queries to the sensors
The processing of the sensors network can be understood
as a bottom-up oriented processing, which pushes new
detected actions and gestures into higher layers in the
interaction model. These actions become interpreted as
the user’s goal and drive the collaboration. During the
planning of the information that is needed to be delivered
to the user, it can be the case that information about the
state of a sensor is required in order to optimally plan the
delivery. In these situations, function calls enable to
directly access the required information from the sensor
in a pull-like or query-style manner. During planning, a
top-down request is issued towards the sensors in the
perceptual layer, and a response is fed back into the
planning process. Therefore, managing the progress of
the domain task and planning the information delivery
can be influenced by or dependant on the states of the
sensors. This pull-like access to sensor data blurs the
boundary between the continuous and discrete layers in
the interaction model, when, e.g., the success of a
primitive task depends on a significant value of a physical
sensor parameter.
5 Architecture of the vertical prototype
We developed vertical prototype for an informationenriched
virtual environment to demonstrate the interplay
of explicit task representations and gesture interaction.
The system architecture of the prototype is the integration
of a haptic virtual environment or Haptic Workbench
(HWB, Stevenson et al., 1999) and the Myriad platform
for information delivery (Paris et al., 2004) (Figure 4).
Figure 4. The basic components of the vertical
prototype: a hand-immersive haptic virtual
environment and the Myriad platform for
information delivery.
These systems are heterogeneous in the structure of their
representation of objects and information. We connected
them through the concept of interactive landmarks. The
user interacts in a hand-immersive virtual environment
and receives force feedback corresponding to the
dynamic data model of the virtual environment.
Instructions and feedback generated by the Myriad
6
platform become visually displayed in the virtual
environment, but can also be delivered via text-to-speech
(TTS) or non-speech audio (as depicted in Figure 4)
(Müller-Tomfelde, 2004). The communication between
the HWB, the Myriad Platform and other applications
make uses of standard computer network communication
facilities.#p#分页标题#e#
The context model depicted in Figure 4 is accessed
directly by the Myriad platform and provides the planning
with all relevant information that is required for
managing the collaboration and for executing the tutorial
task. The context model specifies a domain knowledge
model, a user model and a discourse history and the task
models, amongst others.
5.1 The functional parts of the interaction
model
The interaction model we described is realised partially in
the haptic virtual environment and partially using the
Myriad platform. While the continuous processing of the
sensors detection is part of the rendering loop of the
virtual environment, the Myriad platform basically
handles with the symbolic output of this processing and
sends queries directly to the sensors. The architecture of
the of the prototype follows the schema of Figure 1,
except that this representation (Figure 5) is rotated
clockwise by 90 degrees.
Figure 5. Details of the realisation of the interaction
model in our prototype.
On the left hand side, the user interacts with the system
on the physical level, i.e., the user moves a stylus in the
haptic virtual environment. This user activity is processed
by sensors, and the review of the states of all sensors
provide actions and gestures after the fusion and
discrimination (see section 4). Finally, the action or
gesture becomes interpreted as the user’s intention and
represents the user’s turn in the collaboration. At that
stage, further development happens in the Myriad
platform. The collaboration gets managed, and the turn is
up to the trainer to act corresponding the tutorial task.
The appropriate tactics are selected based on the
prevailing tutorial strategy (see section 3.2). In the next
layer the instruction and feedback becomes planned and
finally delivered to the user in the virtual environment. In
the Myriad platform, the embedded engine builds plans
for all layers of the interaction model based on
declarative plan operators, which become decomposed
into subgoals (Paris et al., 2004). This chain of processing
happens every time the sensor network emits an action
based on the user’s activity. Since the model of
interaction is not hard coded, all the described processing
include massive planning operations in real time. The
dynamics of the collaboration and the task execution
emerge through this planning with respect to an explicit
domain model and the explicit tutorial and domain tasks.
6 First Experiences and future work
The vertical prototype as described in the prior section
provides all the required feature and means to build an
information-enriched virtual environment. Furthermore,
the collaboration and task layer provides information#p#分页标题#e#
about the task that is performed by the user. First
experiences with simple training sessions reveal that the
response time of the vertical prototype is reasonable using
standard computer hardware. The sensor processing as
well as the goal planner operate and communicate fast
enough and provides the information to the user nearly
instantaneously. More tasks have to be tested with the
prototype to investigate the influence of, e.g., the
complexity of the task on the system performance.
Currently the turns of the participants in the collaboration
is strictly alternating so when the ‘speaker’ remains silent
the collaboration could be blocked. Although not a
problem now, future demonstrators and applications have
to show whether this approach has to be revisited or not.
Finally, evaluations and user studies have to be planned
and conducted to undermine the advantages of the
proposed system and bring out the benefits for
applications, like training or assistance.
7 Conclusion
This paper describes an approach for linking explicit task
representations and gesture interaction to create an
information–enriched virtual environment. The approach
focuses on applications where the computer acts as a
trainer or assistant to train implicit knowledge within a
virtual environment. Another aspect of this approach is
that it suits well for applications that require an explicit
task representation, e.g., to document or assess the
performance of user. The goal of this prototype is not
only to enrich objects in the virtual environment, but also
to support users with useful information to proceed in
their tasks. Our approach makes use of a reference
framework for continuous interaction to integrate the
organisation of the collaboration and the tasks of the
participants with the more continuous data processing for
the gesture recognition. The explicit domain task
representation allows the computer to mimic a trainer or
assistant and to deliver appropriate information to the
user in context. The tutorial task becomes executed on the
computer to deliver the right instruction at the right time
to the trainee during the course of the training session.
First experiences with the vertical prototype using simple
domain tasks are promising, and possible application
scenarios considering other roles of the computer as a
participant in a collaboration are investigated.
7
8 References
Aylett, R. and Luck, M. (2000): Applying artificial
intelligence to virtual reality: Intelligent virtual
environments. Applied Artificial Intelligence, 14(1):3-
32.
Beaudouin-Lafon, M. (2000): Instrumental Interaction:
An Interaction Model for Designing Post-WIMP User
Interfaces. In Proceedings of the CHI2000 Conference,#p#分页标题#e#
ACM, CHI Letters 2(1): 446-453.
Bobick, A. (1997): Movement, Activity, and Action: The
Role of Knowledge in the Perception of Motion. Phil.
Trans. Royal Society London B, 352, pp.1257-1265.
Bowman, D., Wineman, J., Hodges, L. and Allison, D.
(1999): The Educational Value of an Information-Rich
Virtual Environment. Presence 8(3): 317-331.
Core, M., Moore, J. and Zinn, C. (2000): Supporting
constructive learning with a feedback planner. In
Papers from the 2000 AAAI Fall Symposium, Menlo
Park, CA: AAAI Press, pp, 1-9.
Feiner, S., MacIntyre, B., and Seligmann, D. (1993):
Knowledge-Based Augmented Reality,
Communications of the ACM 36(7): 53–62.
Jacob., R. (1996): A Visual Language for Non-WIMP
User Interfaces. IEEE Symposium on Visual
Languages, pp. 231-238.
Jung, B., Latoschik, M. and Wachsmuth, I. (1998):
Knowledge-Based Assembly Simulation for Virtual
Prototype Modeling. In Proceedings of the 24th Annual
Conference of the IEEE Industrial Electronics Society,
Vol. 4, IEEE, pp. 2152-2157.
Latoschik, M. (2001): A gesture processing framework
for multimodal interaction in virtual reality. In
Proceedings of the 1st international Conference on
Computer Graphics, Virtual Reality and Visualisation,
AFRIGRAPH '01. ACM Press, New York, NY, pp. 95-
100.
Lu, S., Paris, C. and Vander Linden, K. (2002): Tamot:
Towards a Flexible Task Modeling Tool. In
Proceedings of Human Factors, Melbourne, Australia,
pp. 25-27.
Massink, M. and Faconti, G. (2002): A reference
framework for continuous interaction. In International
Journal Universal Access in the Information Society,
Electronic journal Vol. 1 n. 4, pp. 237-251.
Maybury, M. and Wahlster, W. (eds.) (1998): Readings in
Intelligent User Interfaces. San Francisco: Morgan
Kaufmann.
Müller-Tomfelde, C., Paris, C. and Stevenson, D. (2004):
Interactive Landmarks: Linking Virtual Environments
with Knowledge-Based Systems. In Proceedings of the
OZCHI, Wollongong, Australia.
Müller-Tomfelde, C. (2004): Interaction sound feedback
in a haptic virtual environment to improve motor skill
acquisition. In Proceedings of the International
Conference on Auditory Display (ICAD 2004).
Moore, J. and Paris, C (1993): Planning text for advisory
dialogues: Capturing intentional and rhetorical
information. In Computational Linguistics, 19(4), 651-
694.
Murray, N. and Fernando, T. (1999): A Task Manager for
Virtual Environments. In Proceedings of the Workshop
on User Centered Design and Implementation of
Virtual Environments, University of York.
Nigay, L. and Coutaz, J. (1993): A Design Space for
Multimodal Systems Concurrent Processing and Data
Fusion. In Proceedings of the INTERCHI’93
Conference, ACM, Amsterdam, pp. 172- 178.#p#分页标题#e#
Norman, D. A. (1986): Cognitive Engineering, Ch. 3 in
Draper S.W. and Norman, D.A. (Editors) UCSD, User-
Centered System Design, Lawrence Erlbaum Ass.
Oviatt, S.L., Cohen, P., Wu, L., Vergo, J., Duncan, L.,
Suhm, B., Bers, J., Holzman, T., Winograd, T.,
Landay, J., Larson, J. and Ferro, D. (2000): Designing
the User Interface for Multimodal Speech and Pen-
Based Gesture Applications: State-of-the-Art Systems
and Future Research Directions. Human-Computer
Interaction, 15, pp. 263-322.
Paris, C., Wu, M., Vander Linden, K., Post, M. and Lu, S.
(2004): Myriad: An Architecture for Contextualized
Information Retrieval and Delivery. In AH2004:
International Conference on Adaptive Hypermedia and
Adaptive Web-based Systems, The Netherlands, pp.
205-214.
Rickel, J., Ganeshan, R., Rich, C., Sidner, C. and Lesh,
N. (2000): Task-Oriented Tutorial Dialogue: Issues and
澳洲dissertation网Agents. In AAAI Fall Symposium on Building Dialogue
Systems for Tutorial Applications, North Falmouth,
MA.
Schomaker, L., Nijtmans, J., Camurri, A., Lavagetto, F.,
Morasso, P., Benoît, C., Guiard-Marigny, T., Le Goff,
B., Robert-Ribes, J., Adjoudani, A., Defée, I., Münch,
S., Hartung, K. and Blauert, J. (1995): A Taxonomy of
Multimodal Interaction in the Human Information
Processing System. A Report of the ESPRIT Project
8579 MIAMI.
Smith, S., Duke, D. and Massink, M. (1999): The Hybrid
World of Virtual Environments. In P. Brunet and R.
Scopigno (eds), Computer Graphics Forum, 18(3), pp.
C297-C307.
Stevenson, D., Smith, K., Veldkamp, P., McLaughlin J.,
Gunn, C. and Dixon, M. (1999): Haptic Workbench: A
Multisensory Virtual Environment. The Engineering
Reality of Virtual Reality, Electronic Imaging ’99, San
Jose.
Tarby, J.-C. and Barthet, M.-F. (1996): The Diane+
method. In Computer-Aided Design of UserInterfaces,
Proceedings of the Second International Workshop on
Computer -Aided Design of User Interfaces
(CADUI'96). Namur, Belgium. J. Vanderdonckt (eds),
Presses Universitaires de Namur, Namur, pp. 95-119.