From High to Low Level and Vice-Versa: A New Language for the - - PowerPoint PPT Presentation

From High to Low Level and Vice-Versa:

A New Language for the Translation between Abstraction Levels in Robot Control Architectures José Carlos González, Javier García, Raquel Fuentetaja, Ángel García-Olaya and Fernando Fernández

5 October 2018 Computer Science Department

Planning and Learning Group

2/10

Autonomous robots

• Deliberation
• Required: Decision making
• Not required: Executing a sequence of actions
• Automated Planning
• Searches for a valid action plan to achieve goals
• Symbolic representation
• Coded in PDDL (Planning Domain Definition Language)
• Abstraction layers of knowledge
• Deliberate with high-level about the sensory data
• Remove low-level sensory data from deliberation
• Avoid wasting time planning sequential arrays of actions

?

Introduction

Contribution

From High to Low Level and Vice-Versa: A New Language for the Translation between Abstraction Levels in Robot Control Architectures

3/10

Abstraction layers

• High-level
• High states to deliberate with (PDDL)
• High actions to define behaviors
• Low-level
• Low states with data from the sensors
• Low actions are instructions for the robot

Low actions

Approach Move arm Grip handle Turn hand Push door

High actions

Open door . . .

Low state

Image raw data Depth raw data . . .

High state

Door closed Handle at right

H H

Deliberation

From High to Low Level and Vice-Versa: A New Language for the Translation between Abstraction Levels in Robot Control Architectures Introduction

Contribution

LTH HTL LTH States HTL Actions

4/10

Contribution

HighToLow Robot LowToHigh High Predicates Low Predicates LTH

Description

High Actions Low Actions

Catalogues

Monitoring Executive Decision Support HTL

Description

• Declarative HTL and LTH
• Specified in a new language
• Easier to change behaviors
• No need to recompile code
• Based on the PELEA architecture
• Mature software
• Planning
• Monitoring
• Execution
• Replanning

Introduction

Contribution

Language

From High to Low Level and Vice-Versa: A New Language for the Translation between Abstraction Levels in Robot Control Architectures

5/10

Declarative LTH and HTL

• New language to describe both translations
• Suitable for Automated Planning
• Less code to define behaviors
• Could be generated automatically, e. g. with a GUI
• Allows to extract useful analysis (behavior trees)

High actions Low act. sets Low actions Selector Sequential Parallel

High to Low (actions) Low to High (states)

Sensor data A Sensor data B Sensor data C Sensor data D

Low state

Add predicate A Delete predicate B Increase predicate C

PDDL changes High state . . . . . . . . . . . .

Introduction

Contribution

Language Predicate A Predicate C

?

→ From High to Low Level and Vice-Versa: A New Language for the Translation between Abstraction Levels in Robot Control Architectures

6/10

Low-to-high description

• Low state is checked to change the high state accordingly

If: True add(checkPoseResult $checkPoseResult) add(exerciseTotalPoses $exerciseTotalPoses) If: $checkPoseResult=="OK" increase(performedPoses 1) If: $person==false delete(detected_event person) delete(can_continue) Low-level predicate PDDL predicate

H

LTH States

Contribution

Language

Conclusions

From High to Low Level and Vice-Versa: A New Language for the Translation between Abstraction Levels in Robot Control Architectures

7/10

High-to-low description

• Chosen decomposition is the first that fulfills conditions
• Low-level actions of each set are executed in parallel

High: say(speech,behavior), $behavior==show_video Lows: say(speech) show_video(speech) High: say(speech,behavior), $behavior==insert_subtitle Lows: say(speech) insert_subtitle(speech) High: say(speech,behavior) Lows: say(speech) Label PDDL action

H

HTL Actions

Contribution

Language

Conclusions

From High to Low Level and Vice-Versa: A New Language for the Translation between Abstraction Levels in Robot Control Architectures

8/10

High-to-low description

• Several high-level actions can share the same decomposition
• The array of low-level sets is executed sequentially

High: finish-exercise-mirror() High: finish-exercise-simon(mode), $mode==classic_finish Lows: print("FINISH-EXERCISE") allowAutonomousMovements(true) executeAnimation(blinking) executeAnimation(initFull) say(speech_success) $pauseatend(200) Lows: say("We have finished this exercise.") Lows: executeAnimation(blinking) Lows: say(speech_encouragement) $pauseatend(1000) Internal instruction

H

HTL

Contribution

Language

Conclusions

Actions

From High to Low Level and Vice-Versa: A New Language for the Translation between Abstraction Levels in Robot Control Architectures

9/10

Conclusions

• New declarative language
• State translation from low to high level
• Action translation from high to low level
• Suitable for Automated Planning
• Insight of several benefits
• Much easier to change behavior definitions
• No need to modify the software, only descriptions
• Could be generated automatically, e. g. with a GUI
• Currently working in two of our robots

Language

Conclusions From High to Low Level and Vice-Versa: A New Language for the Translation between Abstraction Levels in Robot Control Architectures

10/10

Our (very near) future work

Robot High to Low Low to High Executive Monitoring Decision Support Low Actions Set Partial Low State 1 2 3 4 5 6 7 8 9 10 Domain Problem Session state PDDL state Planner Declarative file Generic module External application Internal state of the world Decomposition Durative conditions Extrapolation

Language

Conclusions From High to Low Level and Vice-Versa: A New Language for the Translation between Abstraction Levels in Robot Control Architectures

From High to Low Level and Vice-Versa:

A New Language for the Translation between Abstraction Levels in Robot Control Architectures José Carlos González, Javier García, Raquel Fuentetaja, Ángel García-Olaya and Fernando Fernández

5 October 2018 Computer Science Department

Planning and Learning Group

Thank you for your attention

josgonza@inf.uc3m.es