The behavior system
 Kismet home
 Overview
 The robot
 THE FRAMEWORK
 Socializing
 Ongoing research
 Broader questions
 People
 Video
 Publications
 Media

 Facial expression
 Visual attention
 Ocular-motor control
 Low-level features
 Expressive speech
 Affective intent
 Regulation
 BEHAVIOR
 Emotions



Proto-social responses

Kismet's behavior system is configured to emulate those key action patterns observed in an infant's initial repertoire that allow him/her to interact socially with the caregiver. Because the infant's initial responses are often described in ethological terms, the architecture of the behavior system adopts several key concepts from ethology regarding the organization of behavior (esp. Lorenz and Tinbergen). To implement this organization, we have categorized a variety of infant proto-social responses into four categories. With respect to Kismet, the affective responses are important because they allow the caregiver to attribute feelings to the robot, which encourages the human to modify the interaction to bring Kismet into a positive emotional state. The exploratory responses are important because they allow the caregiver to attribute curiosity, interest, and desires to the robot. The human can use these responses to direct the interaction towards things and events in the world. The protective responses are important to keep the robot from damaging stimuli, but also to elicit concern and caring responses from the caregiver. The regulatory responses are important for pacing the interaction at a level that is suitable for both human and robot.

In addition, Kismet needs skills that allow it to engage the caregiver in tightly coupled dynamic interactions. Turn-taking is one such skill that is critical to this process. It enables the robot to respond to the human's attempts at communication in a tightly temporally correlated and contingent manner. If the communication modality is facial expression, then the interaction may take the form of an imitative game. If the modality is vocal, then proto-dialogs can be established. This dynamic is a cornerstone of the social learning process that transpires between infant and adult.

The Behavior System

In the current implementation of the behavior system there are three primary branches, each specialized for addressing a different need (as defined by the homeostatic regulation processes). Each branch is comprised of multiple levels, with three layers being the deepest. Each level of the hierarchy serves a different function, and addresses a different set of issues. As one moves down in depth, the behaviors serve to more finely tune the relation between the robot and its environment, and in particular, the relation between the robot and the human.

reject toy escape withdraw call to person greet person vocal play attentive regard play with toy orient to toy reject face escape withdraw avoid stim engage people seek people sleep quiet down seek toys engage toy avoid stim satiate social satiate fatigue satiate stimulation

Level Zero: Functional Level

Level 0 is is the functional level that establishes which need Kismet's behavior will be directed towards satiating. Specifically, whether the robot should engage people and satiate the social-drive, or engage toys and satiate the stimulation-drive, or rest and satiate the fatigue-drive. To make this decision, each behavior receives input from its affiliated drive. The larger the magnitude of the drive, the more urgently that need must be addressed, and the greater the contribution the drive makes to the activation of the behavior. Environmental factors (the presence or absense of the satiatory stimulus) also contribute to which behavior becomes active.

Level One: The Environment Regulation Level

The behaviors at this level are responsible for establishing a good intensity of interaction with the environment. As shown in the figure, satiate-social and satiate-stimulation pass activation to their behavior group below. At this level, the behavior group consists of three types of behaviors: searching behaviors set the current task to explore the environment and bring the robot into contact with the desired stimulus; avoidance behaviors set the task to move the robot away from stimuli that are too intense, undesirable, or threatening; and engagement behaviors set the task of interacting with desirable, good intensity stimuli.

Search Behaviors

Each search behavior establishes the goal of finding the desired stimuli. Thus, the goal of the seek-people behavior is to seek out skin-toned stimuli, and the goal of the seek-toys behavior is to seek out colorful stimuli. When active, each adjusts the gains of the attention system to facilitate these goals. Each search behavior receives contributions from releasers (signaling the current absence of the desired stimulus), or low arousal affective states (such as boredom, and sorrow) that signal a prolonged absence of the sought after stimulus.



Watch clip: Search behavior 
 (get viewer) 

Avoidance Behaviors

Each avoidance behavior, avoid-stimulus for both the social and stimulation hierarchies, establishes the goal of putting distance between the robot and the offending stimulus or event. The presence of an offensive stimulus or event contributes to the activation of an avoidance behavior through its releaser. At this level, an offending stimulus is either undesirable (not of the correct type), threatening (very close and moving fast), or annoying (too close or moving too fast to be visually tracked effectively). The behavioral response recruited to cope with the situation depends upon the nature of the offense. The coping strategy is defined within the behavior group one more level down.

Engagement Behaviors

The goal of the engagement behaviors, engage-people or engage-toys, is to orient and maintain the robot's attention on the desired stimulus. These are the consummatory behaviors of the level One group. With the desired stimulus found, and any offensive conditions removed, the robot can engage in play behaviors with the desired stimulus.

Level Two: The Protective Behaviors

There are three types of protective behaviors that co-exist within the protective behavior group. Each represents a different coping strategy that is responsible for handling a particular kind of offense. Each coping strategy receives contributions from its affiliated releaser as well as from its affiliated emotion process.

Escape Behavior

When active, the goal set by the escape behavior is to flee from the offending stimulus. This behavior sends a request to the motor system to perform the fleeing response where the robot closes its eyes, grimaces, and turns its head away from a threatening stimulus. It doesn't matter whether this stimulus is skin-toned or colorful -- if anything is very close and moving fast, then it is interpreted as a threat by the low-level visual perception system.

Withdraw Behavior

The withdraw behavior is active when the robot finds itself in an unpleasant, but not threatening situation. Often this corresponds to a situation where the robot's visual processing abilities are over challenged. For instance, if a person is too close to the robot, the eye-detector has difficulty locating the person's eyes. Alternatively, if a person is waving a toy too fast to be tracked effectively, the excessive amount of motion is classified as ``annoying'' by the low level visual processes. The primary function of this response is to send a social cue to the human that they are offending the robot and thereby encourage the person to modify their behavior.

Reject Behavior

The reject behavior is active when the robot is being offered an undesirable stimulus. The affiliated emotion process is disgust. It is similar to the situation where an infant will not accept the food it is offered. It has nothing to do with the offered stimulus being noxious, it is simply not what the robot is after.



Watch clip: Withdrawal and rejection 
 (get viewer) 

Level Two: The Play Behaviors

Kismet exhibits different play patterns when engaging toys versus people. Kismet will readily track and occasionally vocalize while its attention is drawn to a colorful toy, but it will not evoke its repertoire of envelope displays that characterize vocal play. These proto-dialog behaviors are reserved for interactions with people. These social cues are not exhibited when playing with toys. The difference in the manner Kismet interacts with people versus toys provides observable evidence that these two categories of stimuli are distinguished by Kismet.

In this section we focus our discussion on those four behaviors within the social-play behavior group. This behavior group encapsulates Kismet's engagement strategies for establishing proto-dialogs during face-to-face exchanges. They finely tune the relation between the robot and the human to support interactive games at a level where both partners perform well.

Calling Behavior

The first engagement task is the call-to-person behavior. This behavior is relevant when a person is in view of the robot but too far for face-to-face exchange. The goal of the behavior is to lure the person into face-to-face interaction range (ideally, about three feet from the robot). To accomplish this, Kismet sends a social cue, the calling display, directed to the person within calling range. The display is designed to attract a person's attention.

Greeting Behavior

The goal of the behavior is to socially acknowledge the human and to initiate a close interaction. When active, it makes a request of the motor system to perform the greeting display. This behavior is relevant when the person has just entered into face-to-face interaction range. It is also relevant if the social-play behavior group has just become active and a person is already within face-to-face range. The display involves making eye contact with the person and smiling at them while waving the ears gently. It often immediately follows the success of the call-to-person behavior. It is a transient response, only issued once, as its completion signals the success of this behavior.



Watch clip: Drawing people in 
 (get viewer) 

Attentive Regard Behavior

The third task is attentive-regard. This behavior is active when the person has already established a good face-to-face interaction distance with the robot but remains silent. The goal of the behavior is to visually attend to the person and to appear open to interaction. To accomplish this, it sends a request to the motor system to hold gaze on the person, ideally looking into the person's eyes if the eye detector can locate them. The robot watches the person intently and vocalizes occasionally. If the person does speak, this behavior loses the competition to the vocal-play behavior.

Turn-taking Behavior

The forth task is vocal-play. The goal of this behavior is to carry out a proto-dialog with the person. It is relevant when the person is within face-to-face interaction distance and has spoken. To perform this task successfully, the vocal-play behavior must closely regulate turn-taking with the human. This involves a close interaction with the perceptual system to perceive the relevant turn-taking cues from the person (i.e., that a person is present and whether or not there is speech occurring), and with the motor system to send the relevant turn-taking cues back to the person. There are four turn-taking phases this behavior must recognize and respond to: 1) Relinquish speaking turn; 2) Attend to human's speech; 3) Reacquire speaking turn; and 4) Deliver speech. Each state is recognized using distinct perceptual cues, and each phase involves making specific display requests of the motor system.



Watch clip: Proto conversation 
 (get viewer) 



Watch clip: Conversation with Rich 
 (get viewer) 



Watch clip: Conversation with Adrian 
 (get viewer) 



Watch clip: Conversation with Hannah 
 (get viewer) 



Watch clip: Three way conversation 
 (get viewer) 

Other topics
Kismet's hardware
Facial expression
Visual attention
Ocular-motor control
Low-level features
Expressive speech
Affective intent in speech
Homeostatic regulation mechanisms
Emotions



         

    contact information: cynthia@ai.mit.edu