by Alexandra Kirsch
Abstract:
Robots performing complex tasks in changing, everyday environments and required to improve with experience must continually monitor the way they execute their routines and revise them if necessary. Existing approaches, which use either monolithic or isolated, nonrecurring learning processes, cannot sufficiently focus their learning processes to satisfy these requirements. To meet this challenge we propose to make learning an integral part of the control program by providing a control language that includes constructs for specifying and executing learning problems. Our Robot Learning Language (RoLL) makes learning tasks executable within the control program. It allows for the specification of complete learning processes including the acquisition of experience, the execution of learning algorithms and the integration of learning results into the program. RoLL is built upon the concept of experience, which is a learning task specific symbolic summary of a problem solving episode. This means that experiences do not only record the observed data, but also include the robot's intentions and the perceived execution context. The experience acquisition in RoLL is designed in a way that experiences can be defined outside the primary control program, using hybrid automata as a tool for declaratively specifying experience and anchoring it to the program. The rich experience concept enables convenient abstraction and an economic use of experiences. RoLL's design allows the inclusion of arbitrary experience-based learning algorithms. Upon the completion of the learning process RoLL automatically integrates the learned function into the control program without interrupting Program Execution. RoLL enables the plug-and-play addition of new learning problems and keeps the control program modular and transparent. RoLL's control structures make learning an integral part of the control program and can serve as a powerful implementational platform for comprehensive learning approaches such as developmental, life-long and imitation learning.
Reference:
Alexandra Kirsch, "Integration of Programming and Learning in a Control Language for Autonomous Robots Performing Everyday Activities", PhD thesis, Technische Universität München, 2008.
Bibtex Entry:
@PhdThesis{kirsch08integration,
author = {Alexandra Kirsch},
title = {Integration of Programming and Learning in a Control Language for Autonomous Robots Performing Everyday Activities},
school = {Technische Universit\"at M\"unchen},
year = {2008},
bib2html_pubtype = {PhD Thesis},
bib2html_rescat = {Learning,Models,Representation},
bib2html_groups = {Cogito,PARA},
bib2html_keywords = {},
pdf = {http://mediatum2.ub.tum.de/doc/625553/625553.pdf},
url = {http://mediatum2.ub.tum.de/node?id=625553},
abstract = {Robots performing complex tasks in changing, everyday environments and required to improve
with experience must continually monitor the way they execute their routines and revise them
if necessary. Existing approaches, which use either monolithic or isolated, nonrecurring
learning processes, cannot sufficiently focus their learning processes to satisfy these
requirements. To meet this challenge we propose to make learning an integral part of the
control program by providing a control language that includes constructs for specifying
and executing learning problems.
Our Robot Learning Language (RoLL) makes learning tasks executable within the control program.
It allows for the specification of complete learning processes including the acquisition of
experience, the execution of learning algorithms and the integration of learning results into
the program. RoLL is built upon the concept of experience, which is a learning task specific
symbolic summary of a problem solving episode. This means that experiences do not only record
the observed data, but also include the robot's intentions and the perceived execution context.
The experience acquisition in RoLL is designed in a way that experiences can be defined outside
the primary control program, using hybrid automata as a tool for declaratively specifying
experience and anchoring it to the program. The rich experience concept enables convenient
abstraction and an economic use of experiences. RoLL's design allows the inclusion of arbitrary
experience-based learning algorithms. Upon the completion of the learning process RoLL
automatically integrates the learned function into the control program without interrupting
Program Execution.
RoLL enables the plug-and-play addition of new learning problems and keeps the control program
modular and transparent. RoLL's control structures make learning an integral part of the
control program and can serve as a powerful implementational platform for comprehensive
learning approaches such as developmental, life-long and imitation learning.}
}