On TAMEing Robots: A Framework for Affective Robotic Behavior

1
On TAMEing Robots A Framework for Affective
Robotic Behavior

• GVU Brown Bag Presentation
• Lilia Moshkina

Moshkina, L., and Arkin, R.C., On TAMEing Robots,
Proc. IEEE International Conference on Systems,
Man and Cybernetics, Oct. 2003
2
Preliminaries

• Supported in part by GVU seed
• Designing Human-Robot Collaborative Teams
• Create, demonstrate and evaluate methods to
  design human-robot teams that collaborate
  effectively
• Build a framework of affective robotic behavior
  to facilitate the collaboration
• Will focus on the second goal (the corresponding
  paper will be presented at IEEE Conference on
  Systems, Man and Cybernetics, Washington, DC,
  Oct. 8 2003

3
Overview

• Motivation and Related Research
• Architectural Framework
• Psychological Foundations
• Integration into AuRA (Autonomous Robot
  Architecture)
• Exploratory Experimental Study
• Scenario
• Study Design
• Evaluation Methods

4
TAMEing???

• Of course, its an acronym!
• TAME traits, attitudes, moods, emotions
• But theres more to it
• domesticating
• Quality of interaction ease and pleasantness
• Safety
• Bonding/attachment
• In short, TAMEing making robots suitable to
  live in our world
• As situated and embodied companions, as opposed
  to lifeless machines

5
Topic and Research Questions

• Build an Integrative Framework for Time-Varying
  Affective Robotic Behavior Cognitive Perspective
  and Implications for Human-Robot Interaction
• Can integration of various affective processes,
  such as moods, emotions, and attitudes help
  achieve synergy in generating affective behavior?
  
• What are the implications for HRI? Can complex
  affective robotic behavior help lead to more
  natural communication between humans and robots?
• Can the framework help make better predictions
  about human affective phenomena and potentially
  be a test-bed for cognitive/emotion
  psychologists?
• Can adding affective behaviors help robots
  achieve certain tasks more efficiently?

6
Why Affective?

• At least two reasons
• Because of us
• people treat computers as social actors (Clifford
  Nass), even if they dont realize it themselves
• anecdotal evidence military (!) personnel gets
  attached to simulated robots by the end of days
  training
• Challenging affective phenomena are often
  misunderstood
• Because of them
• Its all about adaptation
• affective phenomena are argued to have evolved to
  serve various adaptive purposes

7
Why Robotic?

• Thats where the future is!
• entertainment, edutainment, service robotics,
  elder-care, etc.
• Dynamic and Uncertain Environments
• where else is the need for fast-paced adaptation
  greater?
• Continuous vs. Discrete
• emotion and mood generation follows continuous
  paradigm robotics as test-bed
• Embodied Cognition
• does our embodiment define our cognitive and
  emotive processes?

8
Related Work

• Cynthia Breazeal robotic creature Kismet
• modeled after an infant
• capable of proto-social responses, providing an
  untrained user with natural and intuitive means
  of communication
• Juan Velasquez emotion-based approach to
  robotics
• extends the role of emotion from emotional
  expression for communication purposes to a
  determining factor in decision-making processes
• Ron Arkin Ethological and Emotional Basis for
  HRI
• Emotions and drives for Sony entertainment robot
  Aibo
• Affect-related work in animation and autonomous
  agents domains

9
General Approach

• Interactionist approach
• Combine a number of time-varying affective
  phenomena, such as personality traits,
  affect-based attitudes, moods and emotions in a
  unified framework, and explicitly define
  interactions between them and the robotic system

10
TAMETraits, Attitudes, Moods and Emotions
Environment
11
Behavior-Based Robotic Paradigm

• Robots control program consists of a collection
  of behaviors and coordination mechanisms
• Primitive behaviors have a set of defining
  parameters (e.g., obstacle avoidance sphere) and
  can be combined into assemblages, where each of
  the primitive behaviors is weighted
• Perceptual input causes transitions between
  behaviors, and the output of the behavior
  coordination potentially results in a motor
  action.

12
How TAME fits in

• Composed of four interrelated components
  Personality Traits, Attitudes, Moods, and
  Emotions
• Emotions and moods constitute dynamically
  changing robots affective state
• Traits and attitudes are more or less
  time-invariant, and define general dispositions
• The module continuously (in parallel) scans the
  environment for relevant cues
• Instead of directly defining behavioral
  transitions, the module rather modifies
  behavioral parameters, which affect currently
  active behaviors

13
Psychological Foundations

• Each component serves a distinct adaptive
  function
• traits serve as an adaptation mechanism to
  specialized tasks and environments
• emotions mobilize the organism to provide a fast
  response to significant environmental stimuli
• moods bias behavior according to
  favorable/unfavorable environmental conditions
• attitudes facilitate decision-making process by
  reducing decision space

14
Position in Time/Specificity Space
15
Personality Traits

• Five-Factor Model (FFM) of Personality developed
  by McCrae and Costa serves as a basis for the
  trait component
• Offers a comprehensive taxonomy, consistent over
  time, age and cultural differences
• Personality traits, are mainly inherited or
  imprinted by early experience
• Influence a wider range of behavior than
  emotions, as they are not limited to
  emotionally-charged situations

16
Dimensions of FFM

• Neuroticism (N) contrasts adjustment or
  emotional stability with maladjustment or
  neuroticism the general tendency to experience
  negative affects such as fear, sadness, etc. is
  the core of the N domain
• Extraversion (E) refers to liking people and
  preferring large groups and gatherings, being
  sociable, assertive, active and talkative

17
Dimensions of FFM, cont.

• Openness (O) the elements of O are active
  imagination, preference for variety, intellectual
  curiosity, and independence of judgment
• Agreeableness (A) a dimension of interpersonal
  tendencies, refers to being altruistic,
  sympathetic to others, cooperative and eager to
  help
• Conscientiousness (C) concerns the control of
  impulses high scores mean purposeful,
  strong-willed, achievement oriented
  individuals.

18
Emotions

• Emotion is an organized reaction to an event that
  is relevant to the needs, goals, or survival of
  the organism (Watson)
• Short in duration and noncyclical
• Characterized by a high activation state and
  significant energy and bodily resources
  expenditure
• Core set of emotions modeled are joy, interest,
  surprise, fear, anger, sadness and disgust

19
Functions of Select Emotions

• Interest
• Motivates exploration and learning guarantees
  persons engagement in the environment serves as
  a mechanism of selective attention
• Joy/happiness
• Contributes to affiliative behavior and
  strengthens social bonds has recuperative powers
  and serves as antidote to stress
• Anger
• Mobilizes and sustains energy at high levels
• Fear
• Motivates escape from dangerous situations
  organizes and directs perceptual and cognitive
  processes, focusing attention on the source of
  threat

20
Moods

• Mood is a continuous variable affective state, or
  stream of affect (Watson)
• Represents low activation state and is less
  intense than emotion
• Expends less energy and bodily resources than
  emotion
• Two dimensions positive affect and negative
  affect, which are fairly independent of each
  other
• Correlated with FFM e.g., Neuroticism is
  strongly correlated with negative affect,
  Extroversion -with positive Conscientiousness is
  moderately correlated with positive affect and
  Agreeableness is negatively correlated with
  negative affect

21
Affect-Based Attitudes

• Attitude component forms attitudes towards
  particular objects or situations, and reduces
  decision state space by automatically rejecting
  outcomes connected to undesirable entities/events
  (e.g., dislike or hatred), or providing incentive
  for choosing those connected to desirable
  entities/events
• Functions of attitudes
• Adaptive (guides behavior towards desirable
  goals)
• Knowledge (e.g., stereotypes and prejudices)
• Expressive (expressing personalities and values)
• Ego-defensive (protecting from self-threatening
  thoughts)

22
Integration with AuRA

• MissionLab as a version of Autonomous Robot
  Architecture (Arkin, Balch)
• Overall hybrid architecture is schema-based
  reactive system at a low-level combined with a
  high-level deliberative system
• The reactive component is composed of primitive
  behaviors (schemas) grouped into behavioral
  assemblages
• Each individual primitive behavior produces a
  motor response based on relevant perceptual
  input, and behavior fusion is done through
  cooperation, by summing up vector responses from
  each of the active schemas, where the normalized
  weighted sum of vectors forms an overall motor
  output

23
Integration with AuRA, cont.

• Each schema is weighted by a gain value, thus
  determining the relative importance of each
  primitive behavior
• A finite-state automation defines the high-level
  plan of a robots mission, where each state
  corresponds to a behavioral assemblage, and
  perceptual inputs (triggers) cause transitions
  between the states

24
Integration with AuRA, cont.

• Each TAME component is implemented as a set of
  primitive behaviors
• These dont output motor vectors, but rather
  change parameters for relevant motor schemas
• Each component runs as a separate thread
  continuously throughout the execution (except for
  Traits, which are defined once per execution).

25
Example Obstacle Avoidance

• Obstacle avoidance is a primitive behavior, and
  the magnitude of the motor vector it outputs is
  defined as follows
• In case of trait influence, we would modify the
  parameters of obstacle gain and sphere of
  influence to obtain the vector magnitude

where S is the default sphere of influence, R is
the radius of the obstacle, G is the default
avoidance gain, and d is the distance of robot to
center of obstacle.
26
Trait Influence on Parameters

• User defines a personality configuration
• New behavioral parameters are calculated once at
  the beginning of execution based on user-defined
  traits
• More than one trait can influence a parameter,
  and their combination produces an overall
  personality value affecting a particular
  parameter
• The new trait-based values modify the default
  ones

27
Emotion Influence on Parameters

• Emotions are dynamically generated throughout the
  execution
• based on the presence and strength of
  environmental stimuli
• Can have no, direct or inverse influence, and
  multiple emotions can be combined
• Similarly to traits, emotions modify behavioral
  parameters

28
Exploratory Experimental Study

• Overall goal to identify aspects of effective
  human-robot collaboration
• In particular, specific affective phenomena to
  include into the framework
• Robot as a companion
• IRB-approved
• Study
• Longitudinal study 4 thirty-minute sessions to
  allow the subjects to bond with the robot
• A new task for the participant to perform will be
  introduced at each session, and the last session
  will be cumulative.

29
Longitudinal Study Scenario

• Robot as a pet and protector scenario
• Aibo ERS 210A Sony entertainment robot
• 20 degrees of freedom
• Expressive features mouth, ears, tail, LCD
  display
• Participants will be asked to interact with the
  robot over a number of short sessions, with the
  goal to assess the ease and pleasantness of
  interaction, as well as the extent to which they
  treat the robot as a companion (i.e., bond with
  it)

30
Study Design

• Independent variable presence of affect
• No Affect vs. Affect Conditions
• Affect condition high on openness,
  agreeableness, and extraversion
• Dependent variables
• Pleasantness of interaction
• Amount of praise, petting, results of
  questionnaires
• Ease of interaction
• Time taken to make the dog perform a task number
  of time succeeded
• Level of bonding
• The dog was given a nickname, total time of play

31
Study Design, cont.

• Hypotheses
• Null hypothesis there will be no difference in
  participants attitude towards the robot
• Alternative Pleasantness and ease of interaction
  will be greater in Affect condition than in No
  Affect condition
• Bonding will be the greater in Affect condition.

32
Tasks

• Follow the ball kick the ball
• Participant asks the dog to follow a ball
• And/or to kick it when close
• Come here
• Participant asks the dog to move towards him/her
  and stop when close
• Follow me
• The dog follows the person
• Sick em!
• The dog will move to intercept the intruder
  (Amigobot)

33
Evaluation Methods

• Introspection
• Negative/positive emotionality questionnaire
• Goldbergs Unipolar Big-Five Markers
  Questionnaire
• Post-study questionnaire
• Observation
• Average distance to the robot
• Number of times the robot was petted/touched
• Ratio of praise to scolding to neutral utterances
• Analysis of interactive behaviors (e.g., talking
  to the robot, petting it, etc.)

34
What Next?

• Perform the study
• Extract the information to inform the framework
• Explore time-varying aspects of affective
  phenomena
• Extend affect-behavior mapping
• Explore the interactions between affective
  phenomena

35
Rhetorical Question Do we really need TAME
robots?

• Rhetorical Answer
• Just think of the alternative!
• Thank you!