Collaboration with Agents in VR Environments

1
Collaboration with Agents in VR Environments

• Ján Cíger
• jan.ciger_at_epfl.ch
• Virtual Reality Laboratory, EPFL

2
Outline

• Introduction
• Motivation
• Objectives
• Related work
• Research results
• Symbolic representation of the VRE
• Collaboration model
• Collaborative problem solving
• Multi-agent simulation framework

3
Outline

• Case studies
• Box world, Eye tracking
• Virtual guide
• Natural language interface
• Interface for problem solving
• Declarative story specification for VRET
• Riot in the city
• Conclusions Future work

4
Motivation

• Problems of current VR applications
• Single purpose, hardwired
• User unable to choose own approach/pace
• Difficult to extend and reuse components
• No plugplay for new objects
• Lack high-level metadata
• Impossible or very difficult to reason about

5
Motivation

• User's control over VR applications with virtual
  characters
• Strong
• Everything scripted, predefined
• Common in games
• Very complex, laborious to develop
• Weak
• Autonomous characters, agents
• Non-deterministic, problematic for therapy,
  training, storytelling

6
Motivation

• User's control over VR applications with virtual
  characters
• Intermediate ?
• Characters smart enough to be able to do the
  boring parts alone
• Characters controllable enough to keep the user
  in charge
• Neither scripting and neither fully autonomous
  agents are suitable

7
Objectives

• Unified semantic representation of the VRE
• Enable reasoning
• Enable reusability of components
• Different modes of control
• Let the user choose his pace, method
• Let the user retain control

8
Objectives

• Collaboration problem solving
• Let the machine do the hard/boring work
• Machine should be smart enough to be able to
  find the solution by itself

9
Related work

• Semantic information for animation/reasoning
• Object-specific reasoning (Levison 1996)
• Informed environment (Farenc 1999)
• Parametrized action representation (Badler 2000)
• Task definition language (Vosinakis 2003)
• Smart objects (Kallmann 2001, Abaci 2005)

10
Related work

• Human-agent
• STEVE (Rickel 1997)
• Jacob (Evers 2000)
• Intelligent guide (Doyle 1997)

11
Related work

• Collaborative systems
• Agent-agent or mixed
• STEAM, TEAMCORE, RoboCup (Tambe 1997-2000)
• NASA training systems (Miller 2000)

12
Related work

• Theories techniques - collaboration
• Partial Global Planning (Durfee 1991)
• Joint intentions (Jennings 1992-2000)
• Shared plans (Grosz 1999, Rich 1997)
• Task delegation
• Theory - Ioerger 2001
• Open Agent Architecture - Cohen 1994, Martin 1999

13
Related work

• Action (non-)planning
• GOLOG (Lespérance 1999)
• Cognitive modelling (Funge 1999)
• Hierarchical planners (Baxter 2000, Alonso 2000)
• Action planning (iterative)
• BDI (Bratman 1987)
• Rule-based (Soar Newell, 80-ties)
• Procedural Reasoning System (Georgeff 1992)
• ItPlanS (Badler 1994)

14
Related work

• Action planning (propositional)
• STRIPS (Fikes 1971)
• UCPOP (Penberthy 1992)
• Prodigy (Veloso 1995)
• Graphplan (Blum 1997)
• Sensory Graphplan (Weld 1998)
• Fast Forward (Hoffmann 2001, 2003)

15
Outline

• Introduction
• Motivation
• Objectives
• Related work
• Research results
• Symbolic representation of the VRE
• Collaboration model
• Collaborative problem solving
• Multi-agent simulation framework

16
Symbolic representation of the VRE

• Environment representation
• State - Gino is ready
• Qualitative relationships - A is adjacent to B
• Quantitative properties - Agent A is 30
  thirsty
• General rules - Agent cannot be in two places at
  the same time
• Partial vs. full observability

17
Symbolic representation of the VRE

• State, qualitative and quantitative properties
• Predicates, PDDL notation
• (blue sky)
• Numeric fluents functions of arity N
• (distance ?x ?y)
• Closed world assumption

18
Symbolic representation of the VRE

• Rules of the game - axioms/invariants
• If the agent is near object he has to be at the
  same place as the object

(implies (and (near ?agent ?object)
(at ?object ?place) (not (at ?agent
?place))) (not (near ?agent ?object)))
19
Symbolic representation of the VRE

• Task representation
• Imperative
• Specify both goal and solution method
• Declarative
• Specify only goal, solution is up to the recipient

(move A X Y)
(at A Y)
20
Symbolic representation of the VRE

• Action semantics
• STRIPS derived (PDDL 2.1 subset), operators
• Preconditions (when is the action allowed)
• Effects (what are the results of the action)

(action preparepush params (?X ?Y) precond
(and (at ?X ?P) (at ?Y ?P)
(agent ?X) (place ?P))
effect (pushing ?X ?Y))
21
Symbolic representation of the VRE

• Challenges
• De-coupling between animation and semantics
• Actions can have side-effects
• Human-agent communication
• Common language?
• Translators

22
Collaboration model

• Facilitator
• delegated computing by Cheyer Martin
• Simplified facilitator
• Simpler
• Faster
• - No automatic taskdecomposition

23
Collaboration model

• Facilitator
• Global store of the world state
• Data vs. procedural solvables
• Request matching against available capabilities
• Request matching
• Unification
• Simple compound requests

24
Collaboration model

• Teamwork
• No notion of team in OAA-like systems
• By design nobody knows his collaborators
• Works well for service agents
• Impractical for VR, human-like characters
• Solution - explicit dynamic team creation
• No need to predefine/hardwire the team
• Agent commits to the team

25
Collaboration model

• Team forming
• Modification of Contract Net protocol

26
Collaboration model

• Team roles
• Teammate
• Leader
• Coordinates activities
• Fits better the human-agent case
• Most common hierarchy in human society
• Manager vs. employee
• Military chain of command

27
Collaboration model

• Position of the human user
• Direct interaction
• I will do it myself
• Teamwork with one or multiple collaborators
• Help me with this, please
• High level teamwork
• I want this to be true ...

28
Collaborative problem solving

• Planning

29
Collaborative problem solving

• Planning
• Delegated actions
• Speculative planning
• somebody will be able to do it
• Delegate the responsibility for the task
• Protoagents
• Dummy agents used for planning
• Enable planning before team is available
• Substituted for real agents after the team is
  formed
• Plan using protoagents generic plan

30
Collaborative problem solving

• Protoagents
• self denotes the agent requesting planning
• Recruiting teammate maps protoagent to real agent
• Disbanding team undefines the mapping
• Delegated action operator can be instantiated
  only for already recruited agent
• Forces team forming during planning

31
Collaborative problem solving

• Multi-stage planning
• Leader receives task
• Leader plans and creates team-plan using
  delegated actions
• Team member can plan further to elaborate the
  delegated action

32
Collaborative problem solving

• Object-specific planning
• Agents can learn new actions by exploration
• Extension of smartobjects
• Semantic informationis added
• Properties
• Possible actions(operator animation)

33
Multi-agent simulation framework

• Implementation of the presented concepts
• Visualization and animation engine
• VHD or Delta3D
• Facilitator
• Planning agents
• Auxiliary agents
• Application/domain specific agents

34
Multi-agent simulation framework

• Implementation of the presented concepts

35
Multi-agent simulation framework

• Ghosts puppets
• Puppet is an abstraction layer
• Ghost is an agent (autonomous or interface)
• Puppet is body, ghost is the brain
• Ghost can assert control over puppet(s) possess
  it

36
Multi-agent simulation framework

• Ghosts puppets
• Controllability
• User/agent can take over errant character
• Additional stimulation by trainer/therapist
• Control sharing exchange
• Assisted interaction
• Information access
• Access low level information in a generic way

37
Multi-agent simulation framework

• Ghost's activity loop
• Sensing
• Autonomous activity
• Fail-safe execution
• Repeat
• If nothing to do
• Wait for delegated task
• Perform idle activity

38
Multi-agent simulation framework

• Technology
• Python - agents
• CORBA - communications
• C/C - planner, visualization
• Qt toolkit, PyGame UI agents, HW interfaces

39
Outline

• Case studies
• Box world, Eye tracking
• Virtual guide
• Natural language interface
• Interface for problem solving
• Declarative story specification for VRET
• Riot in the city
• Conclusions Future work

40
Case studies

• Box World
• Simple testbed
• Validate the facilitator concept
• Move boxes around
• Door needs to be opened by partner (agent or
  human)

41
Case studies

• Eye tracking ghost
• Two modes
• Follow my gaze!
• Go there!
• Evaluate usability of gaze tracking

42
Case studies

• Eye tracking ghost

43
Case studies

• Virtual guide
• Let the user explore virtual museum
• Multimodal interface
• Two modes
• Direct exploration
• Follow the guide
• Let the user choose his own way

44
Case studies

• Virtual guide

45
Case studies

• Natural language interface
• Experimental translator agent
• Translate written English to task specification
• Chat-like interface

46
Case studies

• Interface for problem solving
• Attempt to make generic interface to the
  framework
• Suports
• Introspection
• Action scheduling
• Action pre-validation
• Task delegation
• Grants full observability to the user (god mode)

47
Case studies

• Interface for problem solving

48
Case studies

• Declarative story specification for VR exposure
  therapy
• Requirements Believable, Reproducible,
  Controllable
• Non-expert user
• I want this to happen, arrange it!
• I want this amount of stimuli
• Action scheduler

49
Case studies

• Declarative story specification for VR exposure
  therapy
• Virtual bar (social phobia treatment)
• Variable amount of background activity

50
Case studies

• Riot in the city
• Mockup of a training system for crowd management
• Two modes of control
• Direct user is the policeman at the scene
• Indirect user is the commander coordinating the
  forces
• Units are smart able to solve high-level tasks
• Block this street!
• Protect this building!

51
Case studies

• Riot in the city

52
Case studies

• Riot in the city

53
Case studies

• Riot in the city
• Large scale scenario
• 90 agents
• 1000 people in the crowd
• 20 policemen, 20 cars
• Parallel execution of certain actions (movement)
• Breaks total ordering, potential problem
• Multiple planners (performance)

54
Outline

• Case studies
• Box world, Eye tracking
• Virtual guide
• Natural language interface
• Interface for problem solving
• Declarative story specification VRET
• Riot in the city
• Conclusions Future work

55
Conclusions Future work

• Summary of the research
• Different control modes
• Ghosts puppets enable multiple modes selectable
  at runtime
• User is given choice
• It is possible to intervene manually, if needed
• Task delegation
• Difficult or boring tasks can be given to the
  machine

56
Conclusions Future work

• Summary of the research
• Automatic sub-task solving
• Agents capable to solve sub-problems
  autonomously
• Teamwork
• Collaboration with others helps solve tasks
  unsolvable otherwise
• Delegated actions
• Team planning

57
Conclusions Future work

• Contributions
• The ghost puppets framework
• Knowledge and semantic information representation
  in the virtual environment
• A multi-agent collaborative simulation framework
  based on task delegation, facilitation and
  planning
• Delegated actions in standard STRIPS-like
  planners
• Object-speci?c planning

58
Conclusions Future work

• Future work
• Problem of state consistency
• Side effects
• Disconnect between reality and computed state
• Sub-teams
• Agent leading a smaller team focused on a
  subproblem
• Agent autonomy vs. controlability
• Agent's desires and goals of the user may be
  contradictory

59
Conclusions Future work

• Future work
• Facilitator improvements
• Better data storage backend (faster unification)
• Multiple facilitators with replication
• Data partitioning, deeply recursive queries

60

• Thank you for your attention ...