A Task Definition Language for Virtual Agents

1
A Task Definition Language for Virtual Agents

• WSCG03
• Spyros Vosinakis, Themis Panayiotopoulos
• Informatics Dept., University of Piraeus
• Presenter Sie-kyung Jang, VR Lab., KAIST

2
Outline

• Introduction
• Approach
• Related Work
• 3D Environment with Virtual Agent
• The Task Definition Language
• Actions
• Arguments
• Defining a Task
• Example
• Conclusion

3
Introduction (1/2)

• Virtual Environment as a user interface can be
  important for certain type
• Virtual Environment is more attractive when they
  are populated by virtual agents
• Virtual Agent
• Autonomous entity in a virtual environment
• Look like and behave as a living organism
• To enhance virtual agents autonomy model the
  agents functionality and behavior so as to
  resemble the real one

4
Introduction (2/2)

• Strong dependence between the action definition
  and context
• Lack of general-purpose tools for designing and
  implementing intelligent virtual environment
• No tool to describe action combinations and
  sequences needed to achieve specific tasks

5
Approach

• Lack of standard architectures, methodologies and
  general-purpose tools
• ? Context-independent task definition
• No tool to define action combination and
  sequences needed to achieve specific tasks
• ? Use of a procedural language
• Context-independent definition tasks using a
  high-level language

6
Related Work

• Task execution of agents is an important issue
• Three Important Approaches
• Parameterized Action Representation Bad00
• Smart Object approach Kal02
• Improv system Per96

7
3D Environment with Virtual Agent

• SimHuman
• Tool for creation of 3D Environments with virtual
  agents
• Consists of a programming library
• Embedded characteristics Inverse Kinematics,
  Physically Based Modeling, Collision Detection,
  Response, Vision
• Define and animate 3D scenes with an arbitrary
  number of objects and virtual agents
• Consists of two utilities
• Designing the environments
• Designing the agents animation sequences

8
3D Environment with Virtual Agent

• SimHuman
• Entities
• Type
• Agent Perform actions and perceive the current
  state
• Have an autonomous operation
• Object World,
• Tree-structures hierarchy
• Geometry
• A set of attributes
• ltname, type, valuegt

9
The Task Definition Language

• Fill the gap between higher-level decision
  processes and the agents motion and interaction
  in the environment
• Combine numerous built-in functions and commands
  to describe complex task
• Advantages
• Easier for the user to specify action
  combinations and scenarios for virtual agents
• Easily reuse task with different agents ,
  environments

10
Task Structure

• Structure

lttask Definitiongt Variables lt variable
declarationgt Body ltblock of commandsgt
11
Actions

• Process that causes changes to the world ,
  entities
• Has duration (Fixed or variant)
• Executed in continuous timeframes
• Primitive action
• Set of commands that an agent can perform in one
  timeframe
• SimHuman
• Changing the geometric properties of entities
• Adding / removing entities to / from the world
• Sending messages to other agents
• Agent can execute more than one primitive actions
  in one timeframes

12
Actions Representation

• Implemented as a sequence of primitive action
  sets
• Predefined
• a1, a2, ,an with duration n ?t
• ai set of primitive actions performed in
  timeframe I
• Goal-oriented
• ai f(ai-1, I, G)
• I Set of info. About the object properties
  the agent receives
• G Goal

13
Actions

• Keyframing
• Ability to execute predefined animation sequences
• Selecting the agents body parts and adjusting
  their rotation
• Animation library
• Set of user defined animation sequences
• Anim ltnamegt

14
Actions

• Locomotion
• Ability to walk inside the environment
• Use state machine
• Ensure that the movement and rotation of the body
  takes place in a correct and believable manner
• Inverse Kinematics
• IK ltchaingt ltpositiongt
• Continuous correction sequence
• Tests at every step the best rotation for each
  joint to achieve the target
• Works with moving targets, avoid collision

15
Arguments Variable

• Variable type
• boolean, integer, float, string, entity,
• list of entities, vector, list of vectors,
  relation
• - relation composition type
• ex) person(John, 28, 1.80, single)
• Variable set Types
• Agents attributes
• Task arguments
• Tasks internal variables
• Other entity's attributes
• ltentity namegt ltvariable namegt

16
Arguments function

• Use of function as arguments allow actions to be
  called with values that are adapted to different
  environments
• One can define tasks that may be executed in
  dynamic world
• Track the current state of the world, possible
  relations between entities.

17
Arguments function

• Some functions detect spatial relations
• Conditional execution of actions
• Managing the agents beliefs about the world
• Using current geometric properties (position,
  size, orientation)
• Evaluated by current geometric properties
• (position, size, orientation)
• Near, on, front_of, behind, left_of, right_of,
  above, below
• Some function deal with logical properties
• Useful for defining complex conditions
• and, or, not

18
Task Structure

• Structure

lttask Definitiongt Variables lt variable
declarationgt Body ltblock of commandsgt
- TASK name (type1 arg1,.., typen argn)
- Defines local variables

• c1 c2 ci ci is task commands
• Possible Commands
• ltactiongt
• PAR(ltblock b1gt, ltblock b2gt)
• DO(ltblock bgt) UNTIL c
• IF ltbool cgt THEN (block b1gt)
• ELSE (ltblock b2gt)

19
Interacting with Object
Example

• Catch an object using hand
• lt Doing without grasping gt
• End-effecter on the surface of the hand
• Spot on the surface of the object
• Use inverse kinematics action
• Suitable for large environments that need
  simplified agent models.

20
Interacting with Object
Example

• lt Doing with grasping gt
• Need complex agent models and skeleton
• Define many spots
• Use IK motions using PAR command
• Constantly rotate all fingers
• Use DO-UNTIL command
• Avoids the definition of spots
• ? More general
• Must use of constant collision detection checks
• ? increase the computational cost

21
Observing the Environment
Example

• Observing is needed when agent do not know which
  object to interact with

22
A complete Scenario

• Agent visiting a bar
• walking around until there is a free table
• Sitting on a chair
• Calling a waiter
• Ordering drink
• Drinking from a bottle
• Decrease every time the agent brings the bottle
  to the mouth
• Communication between agent and waiter

23
Conclusion

• Context-independent definition tasks
• using a high-level language
• Future Work
• Addition of more complex object interactions
• ( Facial animation for expressing the agents
  emotions )
• Improve action execution and the world
  functionality
• (Able to add more bio-mechanical characteristics
  to the agent)