Context-awareness in wearable and ubiquitous computing
Gregory D. Abowd, Anind K. Dey, Gregory Abowd, Robert Orr & Jason Brotherton
Graphics, Visualization & Usability Center
Georgia Institute of Technology
Atlanta, GA 30332-0280 USA
+1-404-894-7512
{abowd, anind, rjo, brothert}@cc.gatech.edu
GVU Technical Report GIT-GVU-97-11
Submitted to the 1st International Symposium on Wearable Computers, 1997
Abreviated version to be presented as a poster
Table of Contents
A common focus shared by researchers in mobile, ubiquitous and wearable
computing is the attempt to break away from the traditional desktop
computing paradigm. Computational services need to become as mobile
as their users. Whether that service mobility is achieved by
equipping the user with computational power or by instrumenting the
environment, all services need to be extended to take advantage of the
constantly changing context in which they are accessed. This paper
will report on work done in the Future Computing Environments Group at
Georgia Tech to provide infrastructure for context-aware computing.
We will describe some of the fundamental issues involved in
context-aware computing, solutions we have generated to provide a
flexible infrastructure and several applications that take advantage
of context-awareness to allow freedom from traditional desktop computing.
Context-aware computing, ubiquitous computing, consumer
applications, personal information management, tourism, voice-only
interaction, positioning systems.
Wearable computing. Mobile computing. Ubiquitous computing. Perhaps
these terms are synonymous in the mind of the reader. Or perhaps they
evoke some religious zeal concerning fundamental distinctions in
future trends for computing. Regardless, researchers in these
fields all share one common belief, namely that it is time to shift
our research focus away from the traditional paradigm of desktop
computing. Rather than force the user to search out and find
the computer's interface, our new aim is to provide an interface that
can take on the responsibility of locating and serving the
user. In this paper we present the case in support of a research
agenda on context-aware computing. We will provide some general
mechanisms and architectures to support context-awareness and justify
their effectiveness through a number of case studies on applications
that benefit from context-aware services.
The work reported here is a summary of research within the Future
Computing Environments (FCE) Group at Georgia Tech. The FCE Group is
dedicated to the invention of novel applications of computing
technology to assist everyday activities. We are trying to create an
array of computational services that permeate everyday life without
becoming too much of a physical, cognitive or social burden. Some of
the principles we have used to direct our research method are
- Maintain an applications focus, as opposed to a technology
infrastructure focus.
- item Assume a mode of ``fail fast'' research, otherwise known as
rapid prototyping.
- Do not spend too much time predicting the future; aim to invent
it.
After enough experience has been gained through rapidly prototyping
novel applications, general themes and mechanisms can be generated.
In our work in FCE, we have generated three general themes, of which
context-aware computing is one (the others are the capture,
integration and access problem [2] and
ubiquitous software services [18, 1].
The general mechanism for context-aware computing is summarized in the
following steps:
- Collect information on the user's physical, informational
or emotional state.
- Analyze the information, either by treating it as an independent
variable or by combining it with other information collected in the
past or present.
- Perform some action based on the analysis.
- Repeat from Step 1, with some adaptation based on previous
iterations.
After providing a brief overview of related research on context-aware
computing, we will give an overview of four different attempts to
provide context-aware computing. Each attempt is a project that has at
least one (and usually more) running prototypes that we have
developed. Greater details of each project can be found at our Web
site (http://www.cc.gatech.edu/fce) or through various cited
publications. Each project will be presented in terms of a motivating
application, discussion of how context is important in that
application, some general mechanisms or architectural solutions that
were used in our prototyping efforts and a discussion of some issues
that have arisen with respect to context-aware computing.
Future computing environments promise to free the user from the
constraints of stationary desktop computing, so researchers should
focus on what applications maximally benefit from mobility. Highly
portable devices, such as personal digital assistants (PDAs), pagers, and
cellular telephones are starting to proliferate. The wearable
computing community is rapidly providing even more power to support a
mobile user. Too many of the applications provided on these portable
computing devices, however, are simple duplications of what we have on
our desktops, or simple messaging devices that work too much like
traditional ones. None of these devices take into account the one
thing that changes most when a user is mobile ---location. Building
applications that are customized to the user's context can be of
benefit in stationary modes of interaction as well. We define
context-aware computing as any attempt to use knowledge of a user's
physical, social, informational and even emotional state as input to
adapt the behavior of one or more computational services.
The majority of context-aware computing to date has been restricted to
location-aware computing for mobile applications, and our own work
started out that way as well. In thinking about and developing our
own location-aware application, we were greatly influenced by work
such as the PARCTab at Xerox PARC [16], the InfoPad
project at Berkeley [12], the Olivetti Active
Badge system [16] and the Personal Shopping Assistant
proposed at AT&T [5]. A more general programming
framework for describing location-aware objects was the subject of
Schilit's thesis and reflected a lot of the work done at PARC
[15].
There has been some interesting work recently directly related to
context-aware computing. Essa and Pentland have used computational
perception techniques in an attempt to match actual facial expressions
with some prescribed expressions indicating the state of the human
(e.g., smiling, frowning, surprised, etc.) [[8],
[9]].
Though this work does not claim to be a way to predict human emotions,
there is a clear suggestion of how this and related perception
research can improve the quality of contextual information that can be
gathered. Picard's work on affective computing [14]
suggests a similar objective, only through the use
bio-electric signals, coupled with theories on emotion and cognition.
Informational context, defined below in the section on CyberDesk,
has been the subject of work done at Apple (Data Detectors
[4] and Intel (Pandit and Kalbag's Selection
Recognition Agent [13]).
There are a number of PDAs available today, yet there are not many
people who are devoted to using them, for a wide variety of
reasons. One way to thwart this lack of market success is to invent
new ways to use a PDA. The size of current PDAs is
similar in size to a guidebook that a tourist might take with them on
vacation. The book lists places to visit and
provides important practical information, such as locations of interesting
sights, or categorization of hotels and restaurants. The one thing a
book does not know, however, is where the tourist is located when they want
information. Position information augmenting an electronic
guidebook could address that problem. We initiated the Cyberguide
project to experiment with location as a context cue
[[10],[11],[3]].
We have developed a
number of Cyberguide prototypes that support both
indoor and outdoor tours.
Figure 1 shows a version of the outdoor Cyberguide
used for touring local establishments in Atlanta.
Figure 1: The CyBARguide interface. The left shows the interactive map
indicating the user's location (the triangle) and the location of
establishments previously visited (the beer mugs). The user
modifiable database shown on the right supports the
long-term development of touring information for a location.
In general, the purpose of context-awareness in a touring application
is to predict what the user is attending to and provide information on
that entity. As the user moves around, her location (the arrowhead in
Figure 1) is updated on the map. The user
can either explicitly query the map for information in the local
surroundings, or that information can be automatically provided using
some proximity algorithm. Getting close to something is an indication
that information on that object is to be requested.
We currently have a very limited notion of
context in Cyberguide ---physical location and crude orientation.
Tracking where a single individual is currently located provides
adequate information for very simple context-awareness.
We have experimented with capturing historical context (what sights
have been visited already), but there are a number of other aspects of
the tourist's context that can be useful. For instance, knowing where
everyone else is located might suggest places of potential interest.
Knowing the tourist's reaction to some exhibits would help in
suggesting other related places of interest. Being aware of time of
day or day of week may aid in more intelligent user queries of
interesting attractions or activities.
The strategy in the Cyberguide project was to build a number of
prototypes that varied on certain critical features. This was our
first experimentation with context-aware computing, so we chose to
work with the most easily obtained physical context ---the position of
a mobile user. We developed prototypes using both indoor and outdoor
positioning systems. Our high-level design of the family of
Cyberguide prototypes separated the physical positioning
system from the rest of the system, in the hope that a system built
for outdoor use could be easily adapted for indoor use, and vice
versa.
Outdoor positioning was a simple matter, as it is now easy to purchase
inexpensive off-the-shelf GPS systems with a serial interface. The
high price of the infrastructure to support GPS (satellites) has been
paid for by military applications and this cost is not passed on to
the consumer. Indoor positioning is a more difficult proposition, as
we discuss below.
It is easy to get caught up in the particulars of positioning
information. A lesson we have learned is that it is less important to
focus on more and more precise positioning for the purpose of
information delivery. This is not the case for a system which would
require the registration of information overlaid on top of a physical
object, as is suggested by augmented reality research. In Cyberguide,
the key is to determine what the user is attending to in order to
predict what information they are likely to request. It would be much
more useful to augment crude positioning information with orientation
information.
There are a number of indoor positioning systems currently on the
market, and others that are simple enough to assemble. The systems
vary dramatically in price and resolution. There is no one system
that can meet every user need. Most systems excel in a few dimensions
and are forced to compromise in others. The dimensions can be broken
into two categories: quantitative and qualitative. The quantitative
dimensions are: cost, number of users supported, area of coverage,
data rate, latency, resolution, and accuracy. Scaling to support a
large number of users and a large indoor area are the two hardest
dimensions to satisfy simultaneously.
The qualitative dimensions are: obtrusiveness to the user,
obtrusiveness to the environment, and location of information (either
with the user or in the environment). For the CyberGuide
application, these dimensions are very important. The positioning
system used should be unobtrusive to the user and the environment.
Users do not want to be constrained in any way by the technology,
ruling out all systems that require the user to be tethered
(mechanical and magnetic systems). Whether the user or the
environment should hold position information depends on the
application mode. If all information is local and CyberGuide is
working in single user mode, the user benefits by having the position
information. If information is distributed and CyberGuide is working
in a collaborative mode, users benefit by having the system know their
positions. This enables the system to dynamically provide pertinent
information based on position and to notify users of each others'
positions.
In the design of our prototypes, we modularized communications
capabilities and positioning to enable us to experiment with
different realizations of each. Ultimately, we are aiming for a positioning
service that works both indoors and outdoors.
Communication is important for several reasons. It is required in
order to keep track of the location of all mobile units and to
broadcast information to them. Our view of the mobile unit
was not one of a mass storage device, so it will soon become necessary
for information (a map of a location or a description of an
interesting sight) to be delivered on demand to the unit, rather than
stored locally. This is even more important for dynamic information,
such as traffic reports.
There is an interesting relationship between the positioning and
communication systems. In one version of Cyberguide we provided
indoor cell-based IR positioning using communication beacons with a
limited range. Communication and positioning requirements, therefore,
could be satisfied by one system. However, with the Sharp IR units, we can
couple positioning and communication, but the range of the IR link is
so limited (3 feet) that communication is cumbersome. However, it can
be impossible or undesirable to couple positioning and communication
together. For example, if position is coming from GPS, then a
separate means of communication must be used. It makes sense in some
cases to use a short-range IR positioning system that doesn't cover
all space but where it is available provides more exact positioning
(and orientation) information. Communication, however, needs to be
uniform throughout some space, so a short-range communication system
would require many beacons to cover some larger area.
The previous example of context-aware computing emphasized use of a
user's physical context. In this second example, the CyberDesk
project [[18],[6]], we use
informational context to aid in the integration of desktop and network
services. Informational context refers to any artifact (e.g., words
on a screen or a picture at a museum) that the user is attending to.
Examples of network and desktop services are an e-mail browser, a
Web-based map service, or a contact manager. Interacting with one of
these services might suggest some action to be invoked on another
service. For example, as shown in Figure 2, the
user can be reading an e-mail message which has information on a new
book written by a favorite author. The e-mail contains a Web site
address and an e-mail address for the author. The user highlights the
e-mail address (a), thus explicitly announcing the information
context, and the system gives him some suggestions (b) on what he can
do: search for more information on the author, put the author's
contact information in the contact manager, call the author, or send
an e-mail to the author. He chooses the first two options (c and d),
and saves the e-mail.
Figure 2: An annotated scenario of usage for the CyberDesk system.
Click on the figure to get an image with better resolution.
The behavior we are trying to provide in
CyberDesk is automatic integration of separate services. Current
software suites suffer from problems due to poor integration of their
individual tools. They require the designer to think of all possible
integrating behaviors and leave little flexibility to the
user. CyberDesk is a component software framework that
automatically integrates desktop and network services, requiring no
integrating decisions to be made by the tool designers and giving
total control to the user.
When integrating the behavior of separate applications, the programmer
or the user needs to transfer data from one application to another.
The typical scenario, as shown above, takes part of the output from
one application (the name in the e-mail browser in
Figure 2) and uses it as input to another
application. In CyberDesk, after the user selects some string in one
application, all possible type interpretations of that string are
generated and announced (e.g., name, date, friend, etc.). Any
resident service which can accept the announced data then becomes
available to the user via a simple menu button. The context-awareness
provides the ability to change the set of resident services and frees
the user from having to remember any integration procedures.
The CyberDesk system has a simple but innovative architecture. It is
based on an event-driven model, where components act as event sources
and/or event sinks. Events, in this current version, are generated
from explicit user interaction with the system. The system consists of
five core components: the Locator, the IntelliButton, the ActOn
Button Bar, the desktop and network services, and the type
converters. The Locator maintains the registry of event sources and
sinks. This allows the IntelliButton to automatically find matches
between event sources and event sinks based on a given input event, a
task normally required of the system or service designer. The
IntelliButton displays the matches in the form of suggestions to the
user, via the ActOn Button Bar. It is through the ActOn Button Bar
that the user accesses the integrating functionality of CyberDesk.
The services are the event sources and sinks themselves, and are the
tools the user ultimately wants to use. The type converters provide
more powerful integrating behavior by converting given events into
other events, allowing for a greater number of matches. These five
components are discussed in greater detail elsewhere
[6].
The CyberDesk framework was designed to be easily extensible, and we
have so far been able to incorporate over 50 different desktop and
network services, including ones that involve mobile PDAs connected
over a wireless network. Simple extensions to CyberDesk include
adding additional types, type converters, desktop services and network
services. The real advantages with CyberDesk can be seen with more
complex extensions that include adapting the behavior of CyberDesk to
individual use and creating more interesting integrating behavior.
Two examples we have implemented are chaining, in which the output of
a service also acts as a type converter to announce higher level
information to the system, and combining, using multiple pieces of
information to invoke better services. For example, in chaining, a
name selection might suggest an e-mail address lookup service. By
executing this service automatically, the system can also suggest
e-mail address-related services, such as searching newsgroups or
sending an e-mail. In combining, using time with a name can allow the
automatic search of a friend's calendar.
The CyberDesk system relies on an explicit indication by the user of
the informational context. For example, the user must select a region
of text to activate any suggestions of integrating
behavior. While this may not seem so onerous in the example cited
above, when we start to use physical context information, such as
position, it will be less desirable for the user to explicitly
announce information to the system. Moving around (considered
implicit with respect to CyberDesk) should be the only action
required by the user.
Perhaps the biggest limitation of the system is the user interface
implemented by the ActOn Button Bar. It consists of a window that
displays a list of suggested user actions. It is clear that the
number of possible suggestions could quickly become overwhelming to
the user. We are currently looking at different ways to adapt the
interface to initially show actions that the user is likely to take,
but provide a way for the user to see other possible actions as well.
It is likely that some other contextual data (physical, emotional)
could assist in determining the most desirable actions to offer. We are
also looking at different presentation methods for the suggestions,
including pop-up hierarchical menus and document lenses.
One of the problems we have found with chaining is the
potential for multiple services to generate a data type: a Name
object, for example. Since the services are running independently,
the Name objects that they generate could be different. If a
suggested action to the user is to put this name in the Contact
Manager, which Name object should be used? We need to investigate
methods for determining relevancy and confidence of suggested actions
and results, in order to rank suggestions for the user.
In the Savoir (Somewhat assisted voice-only interaction research)
project, we are exploring a number of issues related to context
awareness and voice-only computer interactions [7]. As
computing and
communication devices become smaller, lighter, and more powerful,
voice will become a primary medium of interaction. Our work examines
voice-only interactions and aims to develop a prototype application
that demonstrates useful voice-only functions and well designed
voice-only interfaces.
We have created a prototype application that allows a user to retrieve
information from the Internet using a standard wired or cellular
telephone. The user has access to e-mail, US and world news,
financial reports, weather forecasts, restaurant listings, and sports
scores. We do not currently have access to a voice recognition system
robust enough to handle natural spoken interaction, so we have instead
developed a Wizard of Oz infrastructure to support our investigation
until better speech technology becomes available. We are currently
developing a version of Cyberguide using the Savoir infrastructure,
adding an interactive tour guide facility to Savoir that utilizes
outdoor GPS data. In addition to interpreting a user's voice
commands/query, the Savoir version of Cyberguide will receive
positioning information to help predict what requests for information
a user might make. Figure 3 shows a tourist equipped
to use the Savoir Cyberguide interface.
Figure 3: A tourist equipped to use the Savoir interface to Cyberguide.
Our aim is to enable a user to travel from place to place and to
request and receive relevant information based on location and other
contextual information (such as time-of-day or user history). Current
voice recognition technologies are not adequate to support full
vocabulary natural language interactions. In Savoir, we use
informational context (noting that requests from a user are likely to
be based on information that they have recently heard) to provide the
recognition engine with a relevant grammar, thus improving recognition
of continuous speech. In the Savoir Cyberguide prototype, we will be
able to use physical context to help select a recognition grammar as
well. For example, in our campus tour-guide application, if the user
is standing in front of the Financial Aid building, we would load a
grammar that contains words such as "aid" and "tuition", in addition
to the standard grammar. We could also pre-fetch information on the
Financial Aid office for the user to access next. In this way, we use
physical location to establish context within the application and
improve the accuracy of the speech recognizer.
The position information in Savoir is provided by a portable GPS receiver
that is connected to a portable computer worn by the user. The portable
computer parses the output of the GPS and keeps track of the position
state of the user. When the position changes by a large enough distance,
the portable computer encodes the current position in a sequence of DTMF
tones and sends these tones over the cellular telephone carried by the
user. This cellular telephone forms the communication link between mobile
user and immobile base station.
Once the positioning data has been sent over the cellular link, the base
unit decodes the DTMF tones and makes the position data available to the
application. In our initial prototype, the Map Agent, the component that
provides tour-guide data to the user, can then locate the nearest feature
of interest (be it a building, a park, transportation, etc.) and provide a
description to the user via the application's text-to-speech module.
This conversion from position information to a feature of interest is
exactly the kind of conversion supported by the CyberDesk framework.
We have already described how context in Savoir is used to improve the
performance of speech recognition based on grammar loading. Since it
is unlikely that the grammar loaded will be able to match all possible
questions a user will ask, it becomes important to remember the
utterances that have not been recognized in the past and even record
the place where they occurred. Maintaining this context history may
improve the grammar to be loaded the next time a location is visited
by the same person.
It is also desirable to maintain parts of the informational context
over short periods of time. For instance, a request to get ``Dave's''
e-mail address could result in the name ``Dave'' being placed in a
short-term context area in order to disambiguate pronouns in later
utterances (e.g., ``Send these comments to him'').
Voice-only interfaces have more of an obligation to achieve a
flexibility with respect to user expertise than is the case for
graphical interfaces. Analyzing a voice pattern and matching it
against features indicative of important user states (e.g., interest
or confusion) can help to automatically adjust the form of voice
interaction that the system supports, accommodating those that are
comfortable with the system and those that are not.
Our final example application involves the use of ubiquitous computing
in education. We are actively engaged in a project, Classroom 2000,
which is attempting to augment both teacher and learner in a
lecture-room environment [2]. Much of our lives
are spent engaged in active discussions with other people. How many
times have we been stuck trying to recall the one important message
from a meeting only to have that memory elude us? A similar situation
arises in some forms of classroom education as well. The purpose of
Classroom 2000 is to use automated tools to capture different streams
of classroom activities, such as prepared lecture materials, audio,
video, and handwritten notes on an electronic surface. The captured
material is then integrated together and made accessible via the Web
to provide a facsimile of the actual classroom experience.
The most obvious context is the time in which different events occur.
We provide a lot of synchronization of class materials based on common
times (e.g., linking handwritten notes on an electronic whiteboard to
an audio track for the lecture). When attending a lecture, it is
possible that a student will become lost in the train of thought
of the lecturer. By detecting and logging this, the system can
provide a pointer into the lecture during the access phase that will
bring up the various streams of class activity at the point when
confusion began.
As another example, if the majority of the class is confused at some
point, then that piece of group context may be useful for the
instructor. Students might take advantage of such a facility if
anonymity were guaranteed. Frequently, questions arise during a
lecture and the lecturer is unwilling to address the question. If a
student were able to submit a question at the time in which it occurs
to her, the lecturer could return to the question, and the context in
which it arose, at a later time. Finally, effective use of context
can allow relationships between material across lectures that are
related by topic, not time. We are experimenting with speech
technology to feed into an audio indexing and concept indexing scheme
to support search across notes.
The primary concern of Classroom 2000 has been to use time as a
synchronizing quantity, so that lecture notes written by an instructor
or student can be automatically connected to the audio or video record
of the class at that time. Our challenge so far has been to reduce
all contextual actions into equivalent time information. The
knowledge of when significant activities occur is useful in providing
hooks into other streams of information.
Context-awareness should be done behind the scenes. In early
prototypes of Classroom 2000 we used a graphical interface to allow
students to provide anonymous feedback on their reactions to class
pace and content. We found that most students did not use these
explicit context indicator features because they provided too much of
a distraction from the activity of the lecture. We need to find more
implicit mechanisms to capture group context.
Future computing environments promise to break the paradigm of
desktop computing. To do this, computational services
need to take advantage of the changing context of the user. The
context-awareness we have promoted considers physical, social,
informational and even emotional information. Beginning with an
applications perspective on future computing environments, we have
been able to identify some of the common mechanisms and architectures
that best support context-aware computing. We have described
applications in the domain of tourism, education and personal
information management, and in each case, there is a strong argument
in favor of context-aware computing. We have also raised some issues
to direct further research in context-aware computing.
Our experimental research method is very heavily biased in favor of an
applications focus. We argue, along with Weiser [17],
that the primary purpose for mobile, ubiquitous and wearable computing
is the applications that are enabled. Though research on
infrastructure is also essential, the current balance of research is
weighted too much in favor of infrastructure at the expense of
applications. We hope that our success at generating a productive
research group is enough to encourage other researchers to focus more
on applications-level research.
The authors would like to acknowledge the active support of all
members of the FCE Group for their comments that have lead to the
formulation of these general notions of context-aware computing.
- Abowd, G. D.
Ubiquitous computing: Research themes and open issues from an applications perspective. Technical Report GIT-GVU 96-24, GVU Center, Georgia Institute of Technology, October 1996.
- Abowd, G. D., Atkeson, C. G., Feinstein, A., Hmelo, C., Kooper, R., Long, S., Sawhney, N., and Tani, M.
Teaching and learning as multimedia authoring: The classroom 2000 project. In Proceedings of the ACM Conference on Multimedia --- Multimedia '96, 1996.
- Abowd, G. D., Atkeson, C. G., Hong, J., Kooper,
R., Long, S., and Pinkerton, M.
Cyberguide: A mobile context-aware tour guide. To appear in ACM Wireless Networks, 1997.
- Apple Research Labs.
Apple Data Detectors homepage. Available at
http://www.research.apple.com/research/tech/AppleDataDetectors/.
- Asthana, A., Cravatts, M., and Krzyzanouski, P.
An indoor wireless system for personalized shopping assistance. In L.F. Cabrera and M. Sattyanarayanan, editors, Workshop on Mobile Computing Systems and Applications, pages 69-74. Computer Society Press, December 1994.
- Dey, A. K., Abowd, G. D., Pinkerton, M. and Wood, A. CyberDesk: A framework for providing self-integrating ubiquitous software services. Paper submitted to UIST '97, April 1997.
- Dey, A. K., Catledge, L. D., Abowd, G. D., and Potts, C. Developing voice-only applications in the absence of speech recognition technology. Technical Report GIT-GVU-97-06, GVU Center, Georgia Institute of Technology, March 1997. Submitted to DIS '97.
- Essa, I. and Pentland, A. A vision system for observing and extracting facial action parameters. In Proceedings of the Computer Vision and Patter Recognition Conference, pages 76-83. IEEE Computer Society, 1994.
- Essa, I. and Pentland, A. Facial expression recognition using a dynamic model and motion energy. In Proceedings of the International Conference on Computer Vision, pages 360-367. IEEE Computer Society, Cambridge, MA, 1995.
- Long, S., Aust, D., Abowd, G. D., and Atkeson, C. G. Cyberguide: Prototyping context-aware mobile applications.Rapid Prototyping of mobile context-aware applications: The cyberguide case study. In Proceedings of the 1996 conference on Human Factors in Computing Systems, Chi '96, 1996. Short paper.
- Long, S., Kooper, R., Abowd, G. D., and Atkeson, C. G. Rapid Prototyping of mobile context-aware applications: The cyberguide case study. In Proceedings of the 2nd Annual International Conference on Mobile computing and Networking, November, 1996.
- Long, Jr., A. C., Narayanaswamy, S., Burstein, A., Han, R., Lutz, K., Richards, B., Sheng, S., Broderson, R. W., and Rabaey, J. A prototype user interface for a multimeida terminal. In Proceedings of the 1995 conferecne on Human Factors in Computing Systems, CHI '95, 1995. Interactive experience demonstration.
- Pandit, M. and Kalbag, S. The selection recognition agent: Instant accessto relevant information and operations. In Proceedings of Intelligent User Interfaces '97. ACM Press, 1997.
- Picard, R. Affective computing. Technical Report 321, MIT Media Lab, Perceptual Computing, November 1995. Available as MIT Media Lab Perceptual Computing Techreport 362
- Schilit, W. N. System architecture for context-aware mobile computing. PhD thesis, Colubmia University, 1995.
- Want, R., Hopper, A., Falcao, V., Gibbons, J. The active bage location system. ACM Transactions on Information Systes, 10(1), pages 91-102, January, 1992.
- Weiser, M. Some computer science issues in ubiquitous computing. Communications of the ACM, 36(7), pages 75-84, July 1993.
- Wood, A., Dey, A. K., and Abowd, G. D. CyberDesk: Automated integration of desktop and network services. In Proceedings of the 1997 conference on Human Factors in Computing Systems, CHI '97. pages 552-3. 1997. Technical Note.