JULY/AUGUST 2002 1094-7167/02/$17.00 © 2002 IEEE 17
Interactive Entertainment
Character-Based
Interactive Storytelling
Marc Cavazza, Fred Charles, and Steven J. Mead, University of Teesside, UK
I
nteractive storytelling promises to be an important evolution of computer entertain-
ment, introducing better narrative content into computer games and potentially sup-
porting the convergence of traditional and interactive media. Previous work has described
several paradigms for interactive storytelling,
1–3
each differing on various dimensions
such as user involvement and relations between the
character and plot. Our approach is character-based
and essentially follows Michael Young’s proposal
2
that autonomous actors, whose roles are imple-
mented using real-time planning systems, should
dynamically interact to generate the story.
Within the many possible implementations of
interactive storytelling, we target a specific kind of
application: letting users interfere, at any time, with
a predefined storyline’s progression. Furthermore,
rather than give instructions, users can alter the
environment by stealing an object or influence other
characters by offering advice. The consequences of
this intervention then affect the characters’ behav-
ior and alter the course of action, creating new dra-
matic situations and eventually leading to different
story endings.
System overview
We developed our prototype using the Unreal Tour-
nament game engine as a development environment
(see www.unrealtournament.com). The interactive
story appears as a real-time 3D interactive animation
with subtitles corresponding to the characters’ dia-
logue or important events. Users can physically inter-
act with the characters and navigate through their
environment using normal game controls, or they can
verbally interact with them using a speech recognition
system.
The test scenario we have been using is inspired
by the popular US television sitcom Friends (www.
nbc.com/Friends).
4
We chose a sitcom because, in
this genre, the story ending and intermediate situa-
tions are equally relevant, which provides a more
appropriate testbed for story generation. Further-
more, when developing the system, we defined var-
ious roles for each feature character and formalized
these roles as plans; when the system executes a plan,
it generates character behavior at runtime. Decom-
posing a plan into subgoals reflects an action’s dif-
ferent stages, while the lower layers of the plan
decomposition correspond to various ways to achieve
these goals. For example, if the character Ross wants
to ask out Rachel, then he must acquire information
about her, gain her friendship, find a way to talk to
her in private, and so forth. He faces several possi-
bilities at each stage—for example, to gain infor-
mation, he could steal her diary, talk to one of her
friends, or phone her mother. These various possi-
bilities correspond to subgoals in the description of
Ross’s plan, which can be further refined in the plan
representation until they can be described in terms
of terminal actions (that is, elementary actions car-
ried out by the characters). The system then plays
the actions in the virtual environment using standard
Unreal animation sequences or additional animations
that have been imported into the system.
One particularity of this character-based approach
is how it uses the same basic mechanisms to support
both story variability and interaction. Plan-based roles
for the various characters are dynamically combined
to generate multiple variants of an initial storyline.
Interactive storytelling
is a privileged
application of
intelligent virtual-
actors technology.
The authors introduce
their character-based
interactive storytelling
prototype that uses
Hierarchical Task
Network planning
techniques, which
support story
generation and anytime
user intervention.
Authorized licensed use limited to: Teesside University. Downloaded on June 24,2010 at 13:54:53 UTC from IEEE Xplore. Restrictions apply.
In the absence of any user intervention, this
mechanism will produce a variety of plot
instantiations. At the same time, user inter-
action can interfere with the characters’plans
(for example, causing action failure) and trig-
ger a replanning that varies the plot.
In our system prototype, we modeled the
graphic environment using the game’s level
editor and modeled additional objects using
3d studio max and textures from several
online resources. We imported the characters
from online repositories (Brian Collins cre-
ated the Ross character, “Austin” created
Rachel, and Roger Bacon created Phoebe
and Monica). We implemented the AI layer
in C++ and integrated it in Unreal as a set of
dynamic link libraries. UnrealScript defines
all the functions that interface with Unreal’s
events—that is, those functions dealing with
object interactions. We also fully integrated
communication into Unreal using a speech
recognition system (Babel Technologies’
Automatic Speech Recognition (ASR) soft-
ware development kit).
Planning techniques for
character performance
A wide range of AI techniques has been
proposed to support interactive storytelling
systems, including planning techniques
1,2,4,5
and techniques for augmented truth-mainte-
nance systems.
3
The technique used often
depends on the interactive storytelling para-
digm being implemented. However, there is
no direct correlation between a given AI tech-
nique and a storytelling paradigm. For
instance,Young has used planning to control
the narrative rather than just the behavior of
individual autonomous characters;
2
William
Swartout and his colleagues have used plan-
ning for autonomous characters, but they also
rely on causal narrative representations.
5
We are mainly interested in the emergence
of story variants from the interaction of
autonomous actors, so our emphasis has been
on the actors’behavior rather than on explicit
plot representation or narrative control. Char-
acter-based systems provide a unified prin-
ciple for story generation and interactivity.
As such, they allow anytime interaction,
whereas plot-based systems tend to restrict
user intervention to selected key points in the
plot representation. However, we still needed
our planning formalism to accommodate the
authoring aspects of the baseline narrative.
These knowledge-representation require-
ments led us to investigate planning tech-
niques that we could use in knowledge-inten-
sive domains, and we eventually opted for
Hierarchical Task Networks planning.
6
We
picked HTN planning because it is generally
considered appropriate for knowledge-rich
domains, which can provide domain-specific
knowledge to assist the planning process.
7
It
also appeared that we could naturally repre-
sent the characters’ roles, which serve as a
basis for our narrative descriptions, as HTNs
in which the main characters’goals are decom-
posed into alternative actions.
Hierarchical Task Networks
A single HTN corresponds to several pos-
sible decompositions for the main task—in
other words, we can view HTNs as an implicit
representation for the set of possible solutions.
8
In the present context, each ordered decom-
position constitutes the basis for a character’s
plan, and each HTN associated with an artifi-
cial actor contains the set of all possible roles
for that character across story instantiations.
Although the set of all roles is sufficient,
the set of story instantiations is at least an
order of magnitude larger, because the story
is composed of situations that are the cross-
product of the actors’roles. This also provides
a principled fashion for authoring these story
variants, because that goal node in the net-
work can subsume several ways of solving a
narrative goal. For instance, if Ross needs to
talk to Rachel in private, he can isolate Rachel
from her friends by calling her aside, attract-
ing her attention, asking her friends to leave,
and so forth. This makes it easy to refine
potential variants by adding extra options at
authoring time. As representations, HTNs can
capture essential properties of a character’s
role through the actions the agent takes
toward its goals and the choices it faces.
There is a further need to categorize these
actions according to narrative criteria. These
categories should represent properties bear-
ing relevance for intercharacter relationships,
which we can match to the various actors’
personalities. For instance, actions targeting
other actors can be classified as “friendly,”
“rude,” and so forth. If, when faced with the
task of talking to Rachel in private, Ross inter-
rupts her previous conversation and sends her
friends away, we would tag the corresponding
option in the HTN as “rude.” In a similar fash-
ion, we can categorize single actors’occupa-
tions according to their degree of sociabil-
ity—for example, “lonely” or “sociable.”
To some extent, these categories are part of
an ontology of intercharacter relationships and
can help determine how other characters will
react to the actions taken. Intercharacter rela-
tionships, although obviously important in a
Friends context, are a generic problem in
interactive storytelling. The contents of the
HTN are determined by considering each
actors’ role in the baseline story in isolation.
These roles can be refined by providing addi-
tional options (this refine process is naturally
supported by the HTN formalism). The search
mechanisms associated with HTN planning
also makes them a useful tool for debugging.
Because HTNs are searched from the root
node, which is also the main goal, it is easier
to gain access to the corresponding state of the
world. One additional reason for selecting
HTNs as a formalism is that their graphic
nature seems more supportive of the authoring
phase than STRIPS-like planning formalisms.
However, we have not yet been able to test this
assumption with professional scriptwriters.
Figure 1 gives an overview of a typical
HTN for a character. Pre- and postconditions
for the various tasks (not explicitly repre-
sented in the figure) are associated with each
task node. Preconditions for the lowest-level
operators are constituted by the conjunction
of executability conditions for their associ-
ated terminal actions (those actually acted in
the 3D environment). For instance, if Ross
wants to read information from Rachel’s
diary, the diary should be at its initial loca-
tion, not in use by another agent or near any
witnesses. Some of these conditions are obvi-
ously subject to change in a dynamic envi-
ronment, so they become a main vehicle for
interaction. The system directly implements
postconditions through the effects of termi-
nal actions, which are rolled back to the high-
est-level task node subsuming these actions.
Furthermore, we can compare HTNs to
other forms of knowledge representation pro-
posed in interactive storytelling. In particu-
lar, there is a formal equivalence between
18 computer.org/intelligent IEEE INTELLIGENT SYSTEMS
Interactive Entertainment
Character-based systems provide
a unified principle for story
generation and interactivity.
As such, they allow anytime
interaction.
Authorized licensed use limited to: Teesside University. Downloaded on June 24,2010 at 13:54:53 UTC from IEEE Xplore. Restrictions apply.
subtasks of the HTN and narrative functions
described in narratology that stand for key
narrative actions seen from a given charac-
ter’s perspective. The difference lies in the
fact that the agentive (or predicative) struc-
ture for the equivalent narrative functions lies
outside the corresponding portion of the
HTN, in the interaction with narrative objects
and other characters filling up the roles for
that narrative function. For instance, when
seeking information about Rachel, Ross
could talk to her friend Phoebe. If he talks to
Phoebe, she will complement the agentive
role of the corresponding narrative function.
Also, whenever multiple characters interact,
they potentially instantiate narrative func-
tions “bottom-up” through the conjunction
of activities from their respective HTNs.
HTN planning
Interactive storytelling requires interleav-
ing planning and execution.
2
We have thus
devised a search algorithm to produce a suit-
able plan from the HTN. Exploiting our total
ordering assumption and subtask indepen-
dence, the algorithm searches the HTN
depth-first and left-to-right and executes any
primitive action it encounters in the process.
It allows backtracking when primitive
actions fail (such as following competition
for action resources by other agents, or user
intervention). In addition, it attaches heuris-
tic values to the various subtasks, so forward
search can use these values to select a sub-
task decomposition (this is similar to the use
of heuristics that Peter Weyhrauch described
to “bias” a story instantiation
9
).
An essential aspect of HTN planning is that
it is based on forward search while being goal-
directed at the same time, because the top-
level task is the main goal. (Other recent for-
ward-search planning systems, such as the
Heuristic Search Planner
10
or MinMin,
11
search forward from the initial state to the
goal.) Consequently, because the system is
planning forward from the initial state and
expands the subtasks left-to-right, the current
state of the world is always known (in this
case, the current state reached by the plot).
When initially describing the roles, we
chose to adopt total ordering of subtasks.
Total-order HTN planning precludes the pos-
sibility of interleaving subtasks from differ-
ent primitive tasks, thus eliminating task inter-
action to a large extent.
6
In the case of story-
telling, the subtasks are largely independent
because they represent the story’s stages.
Decomposability of the problem space derives
from the inherent decomposition of the story
into various stages or scenes—a classical rep-
resentation for stories. Our use of HTN is cur-
rently associated with substantial simplifica-
tions of the associated planning problems,
such as subgoal independence, empty delete
lists, and total ordering of subtasks at AND
nodes. However, this approach to planning
seems consistent with the knowledge-inten-
sive nature of interactive storytelling and some
of its inherent properties, such as the tempo-
ral ordering of various scenes. Other planning
techniques—ones more oriented toward a
problem-solving approach, for example—
could be used, such as one that manages
resources and orders actions (see, for instance,
D. Weld’s “dinner date” example, which
describes planning in a domain similar to our
sitcom example
12
). However, it is still unclear
under which conditions a more generic
approach will benefit interactive storytelling.
In addition to their top-down plans, char-
acters also react to specific events. For exam-
JULY/AUGUST 2002 computer.org/intelligent 19
Ring
Go to
Rachel
Give
gift
Go to
Rachel
Give
gift
Give
gift
Select
gift
Be
friendly
Go to
Rachel
Say
nice things
to her
Send
message
Go to
friends
Befriend
her
friends
Send
message
Offer
gift
Gain
affection
Go to
diary
Pick up
diary
Read
diary
Go to
phone
Dial phone
diary
Send
message
Borrow
her diary
Acqiure
infor-
mation
Phone
her mom
Go to
friend
Ask her
friend
Send
message
Ask her
Take
her out
Send
message
Ask
someone
else
Get
reply
Send
message
Ask
yourself
Get
reply
Attract her
attention
Go to
place
Sing her
favorite
song
Send
message
Isolate
her
Go to
worst enemy
Talk to
her worst
enemy
Send
message
Ask
them
Go to
others
Ask
them
Send
message
Take her
aside
Go to
telephone
Phone
Send
message
Turn
towards her
Shout
Send
message
Go to
diary
Pick up
diary
Read
diary
Go to
phone
Dial phone
number
Send
message
Borrow
her diary
Acqiure
infor-
mation
Phone
her mom
Go to
friend
:Friend_Free :Friend_Listen :Diary_Free :Hands_Empty :Phone_Nearby
:Phone_Free
:Phone# :Mom_Listen
Ask her
friend
Send
message
3415
2
5
1
1
32
1
5
31
1
23
55
15
2
1
32
1
5
Figure 1. A Hierarchical Task Network for the main character, Ross.
Authorized licensed use limited to: Teesside University. Downloaded on June 24,2010 at 13:54:53 UTC from IEEE Xplore. Restrictions apply.
ple, Rachel might become jealous whenever
she sees Ross talking alone to another female
character, or she might be upset if he is rude
to one of her friends. These reactions dynam-
ically update “mood” values that affect the
other characters’plans. There is thus more to
authoring than just describing the various sub-
tasks for each actor’s role in an HTN. It is also
necessary to describe the character’s reactions
to various generic situations, mostly arising
from the conjunction of actions from the char-
acters’ respective plans.
Interactive story generation
One main challenge in generating a story
using a character-based approach is achiev-
ing story variability while preserving a well-
defined story genre. In other words, in the
course of various plot instantiations, differ-
ent situations occur that generate different
endings. However, these situations should
generally fall in line with the sitcom genre.
Having a consistent genre helps the user
understand the course of events and decide
whether to intervene and in what fashion.
Story generation results from dynamic
interaction between the main characters’
plans,
4
which correspond to a top-down
approach, because characters’ behavior is
generated from their predefined HTNs. How-
ever, in the course of the action, situations
might emerge that do not form part of the ini-
tial plans. The interaction between charac-
ters’plans results in random onstage encoun-
ters between agents that have the potential to
create situations of narrative relevance. These
interactions constitute a bottom-up approach
(because plan-based behaviors don’t account
for these situations) and thus create a need
for two specific mechanisms: situated rea-
soning and action repair.
Situated reasoning in plan-based actors’
behaviors
13
originates from the discrepancy
between an agent’s expectations and action
preconditions. One defining aspect of situated
reasoning is that it is oriented toward obtain-
ing a specific resulting state in a given situa-
tion.
13
Situated reasoning should include
avoiding an undesirable result. One such
example in interactive storytelling consists of
reacting to situations that emerge from the spa-
tial interactions of artificial actors. The sys-
tem randomly positions the characters on the
set before the story begins. Consequently,
although characters will try to follow their
independent plans, they might find themselves
in situations that are not (and cannot be)
explicitly represented as part of their plan—
and the system can’t ignore these situations.
One example is Ross meeting Rachel by
accident while he is still at the early phase of
his plan (see Figure 2). He can choose to talk
to her or hide from her, but he can’t, from a
narrative perspective, walk past her without
any interaction. One option that situated rea-
soning offers is to hide from her, and a user
can implement this action by interrupting
Ross’current action. Ross could also resume
his initial plan: If his current action is to meet
Phoebe, he can return to her after Rachel
passes (not noticing him). In this specific
case, hiding from Rachel does not impair
subplan continuation.
Consider a similar case, where Ross wants
to talk to Phoebe without Rachel knowing
because he’s afraid Rachel might get jealous
(a feature actually implemented in the sys-
tem). He might wait, but unlike the diary,
Phoebe can in the meantime move to another
location or engage in other activities, caus-
ing the initial intended action to fail. The
interruption caused by situated reasoning can
thus have an irreversible impact on the ini-
tial plan whenever time and duration or loca-
tion constraints appear. However, even in this
case, situated reasoning (hiding from Rachel)
preserves the plot’s relevance and coherence,
because it is properly dramatized and con-
stitutes a part of the story.
One of the main causes for action failure
is not satisfying executability conditions.
Consider the case where other agent behav-
iors affect the executability conditions.
11
One
example is Ross needing to access Rachel’s
diary early in the story. This action can fail in
several cases (corresponding to different con-
texts): the user hides the diary, Rachel is writ-
ing in it, Ross’ sister Monica is in the same
room so he cannot steal it, and so forth. The
first case imposes replanning, because action
repair cannot be applied to the user’s nonde-
terministic behavior (for example, the user
likely won’t return the diary). The second sit-
uation can be a target for action repair,
because Ross could simply wait until Rachel
has finished her task. More interestingly, the
latter case offers the widest range of options.
Ross can choose another source of informa-
tion about Rachel, wait for Monica to leave
the room and resume his initial plan, or try
to influence Monica so that he can still carry
on his original action.
There is sometimes a fine line between
action repair and situated reasoning. Strictly
speaking, action repair should be dedicated
to recovering from action failure. However,
in our storytelling context, action failure is
most often due to not satisfying executabil-
ity conditions due to external factors. For
instance, Ross cannot read Rachel’s diary
because it is missing, Rachel is using it, or
Monica is in the same room. In other words,
action repair is dedicated to restoring exe-
cutability conditions or reaching the same
final state as the original action, whereas sit-
uated reasoning essentially consists of inter-
20 computer.org/intelligent IEEE INTELLIGENT SYSTEMS
Interactive Entertainment
Figure 2. The emergence of situations: Ross meets Rachel by accident while still in the
early phase of his plan.
Authorized licensed use limited to: Teesside University. Downloaded on June 24,2010 at 13:54:53 UTC from IEEE Xplore. Restrictions apply.
rupting the current plan and dealing with a
specific situation. It hence does so more from
the dramatization perspective than from the
planning perspective.
Although the basic elements of actors’
behaviors are deterministic, several factors
contribute to make the action nonpredictable
from the user’s perspective:
• The actors’ initial positions on stage
• The interaction between actors’ plans—
the various characters essentially compet-
ing for resources for action (whether nar-
rative objects or other characters)
• The random output of some terminal
actions
• The characters’ mood status
• User intervention
For instance, the initial positions on stage
strongly influence the emerging situations.
Depending on their positions and activities,
Ross might not be able to acquire informa-
tion from Phoebe before she leaves the apart-
ment to go shopping. Consequently, similar
conditions or user interventions might not
always produce the same results.
User intervention and plot
variation
The user watches the story as a spectator.
He or she can follow the story from any char-
acter’s perspective or navigate the virtual set
while the action is in progress. Then, depend-
ing on the situation, the user can choose
whether to interfere with the characters’
goals. Characters’ actions are dramatized
through the timing of appropriate animations.
Because the actors are playing a role rather
than improvising, their actions are always
narratively meaningful. Hence, if a charac-
ter moves toward a given object, it likely
bears significance on the story and can be a
target for user intervention (for instance, if
the user sees Ross moving toward Rachel’s
diary, he or she can steal or hide the diary).
Users can intervene any time—they don’t
need to wait for key situations or for the sys-
tem to prompt them. However, it is impor-
tant that they understand the story. Thus,
users should be aware from the onset of the
overall dramatic situation—namely, Ross’
interest in Rachel. The system can best con-
vey this using an opening full-motion video
sequence, generated with the game engine.
A user can intervene by either acting on
physical objects onstage that bear narrative
relevance or by advising the characters using
speech recognition. The possibility for phys-
ical intervention is based on the notion of
narrative objects. These objects act as dis-
patchers—that is, they bear narrative signif-
icance because they are the compulsory
objects of key narrative functions. Dispatch-
ers naturally arise from the current course of
action: when Ross seeks a gift for Rachel,
objects such as flowers, chocolates, or jew-
elry become explicit potential targets for user
interaction. These objects, now resources for
actions, can force the character into replan-
ning or action repair, thus creating a new
course for the plot. The user simply uses the
Unreal Tournament’s ordinary “player” fea-
tures to navigate in the virtual set to steal or
hide narrative objects (the user, however, is
not embodied through a character and thus
maintains spectator status).
In Figure 3, a user steals the chocolate box,
so Ross must offer Rachel roses (which hap-
pens to be a favorable gift). This situation can
correspond to various sorts of user interven-
tions, depending on the user’s understanding
of the plot. The user could have realized that
Phoebe lied about Rachel’s preferences and
tried to help Ross. Or, the initial intention
might have been to interfere with Ross’ plan,
in which case the user involuntarily helped
him. Dispatchers crystallize choices both
from the characters’perspective and from the
user standpoint, the latter having to decide
whether to interfere. We do not resort to the
traditional notion of affordance nor to its
implementation in current computer games,
where potentially reactive objects are often
signaled as such. Rather, we intend to use the
same kind of narrative cues as traditional
media, such as camera close-ups in films.
The other mode of interaction consists of
influencing actors using speech recognition.
Speech intervention is the most natural way
of influencing the characters and is ideally
suited to the interactive storytelling paradigm
of user-as-spectator. Several interactive sto-
rytelling systems have reported the use of lin-
guistic interaction,
1,5
essentially in the form
of user–agent dialogue. The rationale being
that, in these systems, the user is a member
of the cast and acts by engaging in conver-
JULY/AUGUST 2002 computer.org/intelligent 21
Figure 3. User intervention: (a) Ross goes to get a box of chocolates. (b) The user sees this and steals the chocolates. (c) Ross can’t
find them, so he (d) replans and gets roses instead.
(a) (b) (c) (d)
Authorized licensed use limited to: Teesside University. Downloaded on June 24,2010 at 13:54:53 UTC from IEEE Xplore. Restrictions apply.
