The
IEEE/RSJ IROS – MOR 2008 WORKSHOP
on
Multi-Competence
Optimization and Adaptation
in Robotics and A-life
September 22,
2008
Nice, France
In
conjunction with
IEEE/RSJ Int. Conf. on Intelligent Robots and Systems
IROS 2008
Workshop
Organizers:
Amiram (Ami)
Moshaiov, Tel Aviv University, Israel
Hussein Abbass, University of New South
Wales, Australia
Contact: For further information please
contact [moshaiov AT eng.tau.ac.il]
-------------------------------------------------------------------------------------------------------------------
Workshop
Theme:
Multi-Competence
Optimization and Adaptation in Robotics
and A-life
NOTE - We use the term multi-competence
interchangeably with multi-objective to encompass both natural and artificial
adaptation processes.
-------------------------------------------------------------------------------------------------------------------
Workshop
Main Goals:
The main goals of the IROS-MOR workshop series are: 1. To
serve as a meeting place for researchers dealing with various multi-objective
optimization problems in robotics, 2. To inform researchers in robotics about
the state-of-the-art in multi-objective optimization methods, 3. To explore the
similarities and dissimilarities between bio-inspired and bio-plausible
multi-competence adaptation, as related to robotics and A-life.
------------------------------------------------------------------------------------------------------------------
Background:
The 1st Multi-Objective Robotics (MOR) workshop
took place at IROS-2006. It focused on the 1st and 2nd goals of the
IROS-MOR workshop series, namely to deal with Multi-Objective Problems (MOPs)
in robotics and the state-of-the-art methods to achieve it. As detailed at http://www.eng.tau.ac.il/events/MOR-Workshop.html
, the 2006 meeting aimed at contributions in Multi-objective
Design (MOD) of Robot Mechanics, MOD of Robot Control, MOD of Mixed
(Simultaneous MOD of) Mechanics and Control of Robots, Multi-objective
Evolutionary Robotics, Multi-objective Path Planning, Multi-objective Obstacle
Avoidance, Multi-objective Sensing (Attention), Multi-objective Manipulation
and Grasping, Interactive MOR, Robust MOR, Multi-objective Learning and
Adaptation, Multi-objective Agents (MOA), Concept-based MOP in Robotics, MOP in
Application Areas of Robotics, and other related areas.
In 2008, we plan to extend the scope of IROS-MOR to also include A-life
related research. The background for this 3rd goal is given below.
Comparing nature and the artificial has been proven to be
fruitful for bio-inspired technical systems and methods, as well as for
scientific studies of nature. Well known examples of the former are bio-inspired
robotics and soft computing methods, whereas the use of bio-morphs to study
"the blind watchmaker" is a renowned example of the latter. Here, the
focus of comparison is on tradeoffs. Despite
of the notions of 'fitness' and 'performance' often being considered closely
related, the idea of 'Pareto-front,' which helps engineers to investigate tradeoffs
among design objectives, appear strange or irrelevant to most biologists. Yet,
tradeoffs are neither new to biologists, nor to cognitive scientists and are often
being considered in a weighted sum approach in such related studies.
Bio-inspired and bio-plausible design may be particularly
useful in attempts to understand the nature of evolution tradeoffs and the
degree to which evolution involves a "balance" between selection for
multiple "objectives." Or, in more general terms, bio-inspired and
bio-plausible design, sensing, learning, and decision making, could support the
study of multi-competence adaptation in natural and artificial systems. This,
in turns, may help developing robust robots that are adaptable to changing
conditions. Evolution, learning, and development influence adaptation in
different time scales and forms. Hence we seek to understand the role of
tradeoffs in such individual modes of adaptation as well as in their combined adaptation
effects (as in ENN), in both robotics and A-life. We further want to explore,
which are the most fundamental tradeoffs in natural adaptation, and in
engineering design that could strongly influence the design of multi-competence
cognitive robots.
------------------------------------------------------------------------------------------------------------------
Topics:
·
Tradeoffs and multi-objective optimization in robotics
including: tradeoffs in robot mechanical design, sensing, control, behavior,
planning, learning, and development (see also detailed list at the first
paragraph of the background above).
·
Handling tradeoffs in A-life research including: tradeoffs in
sensing, control, behavior, planning, learning, and development
·
Tradeoffs in co-evolution and in game theory with
applications to robotics and A-life
·
Tradeoffs in related application areas such as: Engineering
Design, Control, Mechatronics, Informatics, Multi-Criteria Decision Making,
Complexity Science, Biological Systems, Biomechanics, Environmental Planning,
Economics, Social Sciences, Cognitive and Behavioral Sciences (see note below)
·
Tradeoffs in computational methods such as: Soft Computing,
Organic Computing, Bio-inspired Hybrid Metaheuristics, and their applications
to robotics and A-life (see note below)
·
Metaphors and analogies in evolutionary computations and bio-inspired
hybrid-metaheuristics (as related to Multi-Objective Optimization, and to
applications in robotics and A-life)
·
Similarities and dissimilarities between natural and
artificial processes of adaptation and their "handling" of tradeoffs
·
The adaptation/optimization debate and its relation to
teleology and the notion of objectives
·
Tradeoffs in natural systems and their comparison with
tradeoffs in engineering design and in multi-criteria decision making
·
Tradeoffs in conceptual design and in natural "concepts"
such as species and neural ensembles
-------------------------------------------------------------------------------------------------------------------
Workshop Schedule:
9:00 -9:15
Introduction to the IROS-MOR workshop on Multi-Competence Adaptation
in Robotics and A-life
Amiram Moshaiov
9:15- 9:40
Integrated visual information processing on a humanoid robot with
foveated vision
Ales Ude
9:40-10:05
Incremental evolution using multi-objective evolutionary algorithms and
behavior sharing
Jean-Baptiste Mouret and St´ephane Doncieux
10:05-10:30
Exploration strategy based on multi-criteria decision making for an
autonomous mobile robot
Nicola Basilico and Francesco Amigoni
10:30 -10:50
Coffee break
10:50 - 11:15
Multi-objective multi-robot surveillance
Canu, Delle Fave, Iocchi, Ziparo
11:15-11:40
Evolutionary search of conceptual path plans using a
relaxed-Pareto-optimality approach
Elad Denenberg and Amiram Moshaiov
11:40- 12:30
Discussion
-------------------------------------------------------------------------------------------------------------------
Workshop Abstracts:
Integrated visual information processing on a humanoid robot
with foveated vision
Ales Ude
To
successfully act in real-world environments, a robot must simultaneously
perform a number of nontrivial visual tasks such as: 1. detecting objects and
activities of interest, 2. observing them by actively controlling its eyes
and body, and 3. analyzing the acquired images to extract new information about
objects and events in the scene. Humanoid vision systems that replicate
the foveated structure of the human eye are well suited for these tasks because
they allow simultaneous processing of visual information at different
resolutions. Robots with foveated vision can monitor and explore their
surroundings using wide-angle images of low resolution, thereby increasing the
efficiency of the search process, while simultaneously extracting additional
information – once the area of interest is determined – from narrow-angle but
high resolution foveal images that contain more detail. In this talk I shall
present our work on integrated processing of information acquired by peripheral
and foveal vision and the associated motor control algorithms. The developed
algorithms can be applied to solve multiple objectives at the same time, where
the problems are as different as smooth pursuit and tracking on the one hand,
and object recognition on the other hand.
Incremental evolution using
multi-objective evolutionary algorithms and behavior sharing
Jean-Baptiste Mouret and St´ephane Doncieux
While many robot controllers
have been successfully generated using evolutionary algorithms, current methods
do not scale well to complex robotics tasks. The bootstrap problem is often
recognized as one of the main causes of this difficulty: if all individuals
from the first randomly generated population perform equally poorly, the
evolutionary process won’t generate any interesting solution. In this paper, we
introduce two approaches to overcome this problem, both of them based on
multi-objective evolution. In the first method, called multi-subgoal evolution,
the user defines a set of subgoals which could act as potential stepping stones
to the final goal. We show how a multi-objective evolutionary algorithm can
then be used to obtain working controllers. In the second proposed method, a
distance between behaviors has to be designed. This distance is used to apply a
selective pressure towards the most original solutions. These two processes are
fully automatic: once the experiments have been defined, no further human
intervention is required. These approaches have been successfully tested and
compared on the evolution of a neuro-controller for a complex light-seeking
mobile robot.
Exploration strategy based on
multi-criteria decision making for an
autonomous mobile robot
Nicola Basilico and Francesco Amigoni
In robot mapping, exploration
strategies are used to determine locations visited by an autonomous mobile
robot in building a spatial representation of an environment. Usually, a next-best-view
approach is used. The robot iteratively chooses the best location where to go
among a set of candidates, evaluating each one of them on the basis of
different criteria. For every candidate, utilities associated to single
criteria are aggregated by means of ad hoc functions. The maximization of the
aggregated utility greedily drives the robot’s choices. In this paper we propose
an exploration strategy based on a multi-criteria decision making technique in
which evaluation of candidate locations takes into account different criteria
together with their mutual interactions. Our approach is more general than
those proposed in literature and it’s independent of the nature and the number of
criteria considered. Some preliminary experimental results are presented in
order to assess the feasibility of our approach.
Multi-Objective
Multi-Robot Surveillance
Canu, Delle Fave,
Iocchi, Ziparo
In
many surveillance applications, there are different properties of the
environment to check. For example, in the case of robots surveillance at an
industrial depot, one could be interested in verifying for fire alarms,
intrusion alarms or bio-hazards. Given these different goals, it is very hard
to characterize the solution of the multi-objective problem by defining a
unique utility function. Indeed, this would require to define
a measure of the trade-off among objectives, which are, by definition, incommensurable
quantities. In this paper, we present an approach to address such issues. In
particular, we define the multi-robot multi-objective surveillance problem
and show how this can be solved with a multi-objective heuristic search. The
approach has been experimented by using an off-line planner and actual
implementation on simulated and real multi-robot systems.
Evolutionary Search of Conceptual
Path Plans
Using a Relaxed-Pareto-optimality
Approach
Elad Denenberg and Amiram Moshaiov
This
paper presents an evolutionary concept-based multi-objective optimization
approach. A relaxed Pareto-optimality criterion is used to allow for conceptual
plans, with performances close to those of the Pareto-optimal set, to survive
the evolutionary search process. The proposed approach is significant to the
final selection of conceptual and particular path plans. Such a selection may
involve, in addition to the modeled objectives, post-run considerations such as
robustness. Here, the relaxed problem is defined using a Pareto-epsilon
measure. Furthermore, this paper provides a description of a novel
multi-objective evolutionary algorithm that is tailored for the described
problem. Finally, the algorithm is demonstrated using both an academic test
function and a path planning problem.
-------------------------------------------------------------------------------------------------------------------
Special Issue:
We plan to negotiate a related special issue in
a leading journal to which a selection of the workshop accepted submissions
will be considered.