designing libraries of transparent, general and restartable activity plans for a simulated...
DESCRIPTION
This presentation was given on the occasion of the defense of my M.Sc. thesis in computer science on February 1, 2006, at the Technical University of Munich (TUM).TRANSCRIPT
Designing Libraries of Transparent, Generaland Restartable Activity Plans for a Simulated
Autonomous Service Robot
Andreas J. Häusler
Intelligent Autonomous Systems GroupInstitut für Informatik
Technische Universität München
February 1, 2006
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Outline
1 Motivation
2 Simulation Environment
3 Plan Library
4 Evaluation
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Outline
1 Motivation
2 Simulation Environment
3 Plan Library
4 Evaluation
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Robots Today
current stageacting most accurately
highly specialized
do unpleasant, dangerous or often recurring work
drawbacksfocused on single tasks
hardware depends on application
acting in static environments
need supervision in most cases
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Robots Today
current stageacting most accurately
highly specialized
do unpleasant, dangerous or often recurring work
drawbacksfocused on single tasks
hardware depends on application
acting in static environments
need supervision in most cases
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Challenges to Future Robots
robots of the future will have to be. . .
adaptable to dynamically changing environments
robust to failures
able to act on their own
able to accomplish a wide range of different tasks
much cheaper than today
⇒ change from industrial to every-day household robots
failure recovery
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Challenges to Future Robots
robots of the future will have to be. . .
adaptable to dynamically changing environments
robust to failures
able to act on their own
able to accomplish a wide range of different tasks
much cheaper than today
⇒ change from industrial to every-day household robots
task diversity
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Challenges to Future Robots
robots of the future will have to be. . .
adaptable to dynamically changing environments
robust to failures
able to act on their own
able to accomplish a wide range of different tasks
much cheaper than today
⇒ change from industrial to every-day household robots
failure recovery
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Challenges to Future Robots
robots of the future will have to be. . .
adaptable to dynamically changing environments
robust to failures
able to act on their own
able to accomplish a wide range of different tasks
much cheaper than today
⇒ change from industrial to every-day household robots
failure recovery
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Requirements
The crucial point lies in how the robot’s behaviour isimplemented. It has to be. . .
transparent
general
restartable
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Requirements
The crucial point lies in how the robot’s behaviour isimplemented. It has to be. . .
transparent
general
restartable
transparencyinferences about actions
meaningful code structures
code abstraction
improved understandability
efficient learning
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Requirements
The crucial point lies in how the robot’s behaviour isimplemented. It has to be. . .
transparent
general
restartable
generalityadaptability
task reusability
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Requirements
The crucial point lies in how the robot’s behaviour isimplemented. It has to be. . .
transparent
general
restartable
restartabilityfailure recovery
task interruptability
repeatable results of task execution
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Outline
1 Motivation
2 Simulation Environment
3 Plan Library
4 Evaluation
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The Environment
dynamically changing
oriented at the real world
allows manipulation
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The B21 Robot
environmental sensors
internal sensors
actors
knowledge bases
required abilitiesperforming differenttasksfailure awareness
Video
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The B21 Robot
environmental sensorscamerasonar sensorslaser sensor belt
internal sensors
actors
knowledge bases
required abilitiesperforming differenttasksfailure awareness
Video
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The B21 Robot
environmental sensorsinternal sensors
touch sensorsrotation sensors
actors
knowledge bases
required abilitiesperforming differenttasksfailure awareness
Video
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The B21 Robot
environmental sensors
internal sensorsactors
robot motion actorshand movement actorsarm movement actors
knowledge bases
required abilitiesperforming differenttasksfailure awareness
Video
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The B21 Robot
environmental sensors
internal sensors
actorsknowledge bases
environmental beliefentity-specificknowledgeheuristics
required abilitiesperforming differenttasksfailure awareness
Video
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The B21 Robot
environmental sensors
internal sensors
actors
knowledge bases
required abilitiesperforming differenttasksfailure awareness
Video
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Outline
1 Motivation
2 Simulation Environment
3 Plan Library
4 Evaluation
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The Concept Behind Activity Plans
Goals
Goal Routines
Low Level Routines
High Level PlansLow Level PlansBasic Control CommandsLisp ProcessesExternal Processes
compositionactivity plans
goals
goal routines
low-level routines
collections of goal achievements
define what task has to befulfilled instead of how
allow the robot to know aboutthe task’s purpose
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The Concept Behind Activity Plans
Goals
Goal Routines
Low Level Routines
High Level PlansLow Level PlansBasic Control CommandsLisp ProcessesExternal Processes
compositionactivity plans
goals
goal routines
low-level routines
directives for robotic actions
slots constitute “predicates” forfulfillment tests
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The Concept Behind Activity Plans
Goals
Goal Routines
Low Level Routines
High Level PlansLow Level PlansBasic Control CommandsLisp ProcessesExternal Processes
compositionactivity plans
goals
goal routines
low-level routines
specify actions for goal achieves
high-level plans contain failuremonitoring
low-level plans contain“computations”
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
The Concept Behind Activity Plans
Goals
Goal Routines
Low Level Routines
High Level PlansLow Level PlansBasic Control CommandsLisp ProcessesExternal Processes
compositionactivity plans
goals
goal routines
low-level routines
directly controlling the robot’sactors
contain no further goal achieves
act on the process level of therobot’s architecture
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Designing the Plans
roadmap1 thinking about the
requirements2 introducing goal
inheritance3 specifying the
execution order4 construction
always keep in mind thedemanded behaviour
rembember to think aboutpossible failure situations
also consider recoveryapproaches if applicable
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Designing the Plans
roadmap1 thinking about the
requirements2 introducing goal
inheritance3 specifying the
execution order4 construction
simplifies the implementationprocess
enables “semantical”structuring of the plans
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Designing the Plans
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Designing the Plans
roadmap1 thinking about the
requirements2 introducing goal
inheritance3 specifying the
execution order4 construction
allows bottom-upimplementation and testing
enables “semantical”structuring of the control flow
allows further distinctionbetween high-level andlow-level plans
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Designing the Plans
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Designing the Plans
roadmap1 thinking about the
requirements2 introducing goal
inheritance3 specifying the
execution order4 construction
building a collection ofheuristical functionsalways have in mind that
calls to heuristics andtrajectory computationsreside in low-level plansfailure monitoring andrecovery always arecontained in high-level plans
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Capability of the Plans
the demandstransparency
generality
restartability
others
achieving transparency throughproper implementation
splitting up the activity plans
externalizing the heuristicalfunctions
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Capability of the Plans
the demandstransparency
generality
restartability
others
thoughts about the achievement ofgenerality
coping with “new” situations
adapting to changed demands
splitting up the activity plans
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Capability of the Plans
the demandstransparency
generality
restartability
others
restartability through failure recovery
implementing sophisticatedfailure monitoring
using the knowledge base
interruptable activity plans
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Capability of the Plans
the demandstransparency
generality
restartability
others
robustness of the plans
support for future learningmechanisms
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Outline
1 Motivation
2 Simulation Environment
3 Plan Library
4 Evaluation
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
for ensuring high levels of generality and restartability
however, these cannot be tested separately⇒ scenarios for the overall performance are needed
evaluation of transparency only through “reading the code”
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
for ensuring high levels of generality and restartability
however, these cannot be tested separately⇒ scenarios for the overall performance are needed
evaluation of transparency only through “reading the code”
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
for ensuring high levels of generality and restartability
however, these cannot be tested separately⇒ scenarios for the overall performance are needed
evaluation of transparency only through “reading the code”
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
for ensuring high levels of generality and restartability
however, these cannot be tested separately⇒ scenarios for the overall performance are needed
evaluation of transparency only through “reading the code”
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
prerequisites to proper testingmain focus on failure monitoring and recovery⇒ think about all failures the robot is aware of, and
the situations in which they will be invokedhow the robot has to react
low-level plan recovery only as consecutive failures
tests compare “normal” to “non-recovery” behaviour
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
prerequisites to proper testingmain focus on failure monitoring and recovery⇒ think about all failures the robot is aware of, and
the situations in which they will be invokedhow the robot has to react
low-level plan recovery only as consecutive failures
tests compare “normal” to “non-recovery” behaviour
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
prerequisites to proper testingmain focus on failure monitoring and recovery⇒ think about all failures the robot is aware of, and
the situations in which they will be invokedhow the robot has to react
low-level plan recovery only as consecutive failures
tests compare “normal” to “non-recovery” behaviour
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
prerequisites to proper testingmain focus on failure monitoring and recovery⇒ think about all failures the robot is aware of, and
the situations in which they will be invokedhow the robot has to react
low-level plan recovery only as consecutive failures
tests compare “normal” to “non-recovery” behaviour
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
prerequisites to proper testingmain focus on failure monitoring and recovery⇒ think about all failures the robot is aware of, and
the situations in which they will be invokedhow the robot has to react
low-level plan recovery only as consecutive failures
tests compare “normal” to “non-recovery” behaviour
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
the “pick-up”-scenariofor testing the stability of low-level plans
randomized robot- and object-positions
in each pass losing the object had been provoked
success if the robot managed to lift the object again
the “move-entity”-scenariotesting the robot’s general performance
randomized starting positions
success if the object was placed on the table
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Evaluation Scenarios
the “pick-up”-scenariofor testing the stability of low-level plans
randomized robot- and object-positions
in each pass losing the object had been provoked
success if the robot managed to lift the object again
the “move-entity”-scenariotesting the robot’s general performance
randomized starting positions
success if the object was placed on the table
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Results of the Evaluation
performance of pick-up-cup actions; entity-lost failure provokedIntelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Results of the Evaluation
performance of put-down-cup actionsIntelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Results of the Evaluation
performance of the robot’s arms when picking up a cupIntelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Summary
reflecting the requirementstransparency through introducing a goal hierarchy
generality through heuristical functions
restartability through failure recovery
results of the workuniversal approach for implementing activity plans
improved performance through failure recovery
appearance of (rudimentary) emergent behaviour
the library provides a reliable base for further work
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Summary
reflecting the requirementstransparency through introducing a goal hierarchy
generality through heuristical functions
restartability through failure recovery
results of the workuniversal approach for implementing activity plans
improved performance through failure recovery
appearance of (rudimentary) emergent behaviour
the library provides a reliable base for further work
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Summary
perspectives for the futureimproving the present plans
adding new plans and new kinds of behaviour
inserting new technologies
In our vision, the robot shall sometime be like a “human butler”to private households.
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Summary
perspectives for the futureimproving the present plans
adding new plans and new kinds of behaviour
inserting new technologies
In our vision, the robot shall sometime be like a “human butler”to private households.
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Summary
perspectives for the futureimproving the present plans
adding new plans and new kinds of behaviour
inserting new technologies
In our vision, the robot shall sometime be like a “human butler”to private households.
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler
Motivation Simulation Environment Plan Library Evaluation Summary
Summary
perspectives for the futureimproving the present plans
adding new plans and new kinds of behaviour
inserting new technologies
In our vision, the robot shall sometime be like a “human butler”to private households.
Intelligent Autonomous SystemsDepartment of InformaticsTechnische Universität München
Designing Activity Plan LibrariesAndreas J. Häusler