A Framework for Learning by Demonstration in Multi-teacher Multi-robot Scenarios
Author(s)
Fernandes Martins, Murilo
Type
Thesis
Abstract
As robots become more accessible to humans, more intuitive and human-friendly ways of
programming them with interactive and group-aware behaviours are needed.
This thesis addresses the gap between Learning by Demonstration and Multi-robot
systems. In particular, this thesis tackles the fundamental problem of learning multi-robot
cooperative behaviour from concurrent multi-teacher demonstrations, problem which had
not been addressed prior to this work.
The core contribution of this thesis is the design and implementation of a novel, multi-
layered framework for multi-robot learning from simultaneous demonstrations, capable
of deriving control policies at two different levels of abstraction. The lower level learns
models of joint-actions at trajectory level, adapting such models to new scenarios via
feature mapping. The higher level extracts the structure of cooperative tasks at symbolic
level, generating a sequence of robot actions composing multi-robot plans. To the best of
the author's knowledge, the proposed framework is the first Learning by Demonstration
system to enable multiple human demonstrators to simultaneously teach group behaviour
to multiple robots learners.
A series of experimental tests were conducted using real robots in a real human workspace
environment. The results obtained from a comprehensive comparison confirm the appli-
cability of the joint-action model adaptation method utilised. What is more, the results
of several trials provide evidence that the proposed framework effectively extracts rea-
sonable multi-robot plans from demonstrations. In addition, a case study of the impact
of human communication when using the proposed framework was conducted, suggesting
no evidence that communication affects the time to completion of a task, but may have
a positive effect on the extraction multi-robot plans. Furthermore, a multifaceted user
study was conducted to analyse the aspects of user workload and focus of attention, as
well as to evaluate the usability of the teleoperation system, highlighting which parts were
necessary to be improved.
programming them with interactive and group-aware behaviours are needed.
This thesis addresses the gap between Learning by Demonstration and Multi-robot
systems. In particular, this thesis tackles the fundamental problem of learning multi-robot
cooperative behaviour from concurrent multi-teacher demonstrations, problem which had
not been addressed prior to this work.
The core contribution of this thesis is the design and implementation of a novel, multi-
layered framework for multi-robot learning from simultaneous demonstrations, capable
of deriving control policies at two different levels of abstraction. The lower level learns
models of joint-actions at trajectory level, adapting such models to new scenarios via
feature mapping. The higher level extracts the structure of cooperative tasks at symbolic
level, generating a sequence of robot actions composing multi-robot plans. To the best of
the author's knowledge, the proposed framework is the first Learning by Demonstration
system to enable multiple human demonstrators to simultaneously teach group behaviour
to multiple robots learners.
A series of experimental tests were conducted using real robots in a real human workspace
environment. The results obtained from a comprehensive comparison confirm the appli-
cability of the joint-action model adaptation method utilised. What is more, the results
of several trials provide evidence that the proposed framework effectively extracts rea-
sonable multi-robot plans from demonstrations. In addition, a case study of the impact
of human communication when using the proposed framework was conducted, suggesting
no evidence that communication affects the time to completion of a task, but may have
a positive effect on the extraction multi-robot plans. Furthermore, a multifaceted user
study was conducted to analyse the aspects of user workload and focus of attention, as
well as to evaluate the usability of the teleoperation system, highlighting which parts were
necessary to be improved.
Date Issued
2012-01
Date Awarded
2012-02
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Demiris, Yiannis
Sponsor
CAPES Brazil
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)