innovation through science

innovation through science
Welcome
In 2003 Honda made the bold and foresighted move to
establish research institutes in Japan, the United States
and in Europe that should re-define corporate research.
The aim was to be conceptually close enough to academic
institutions to be able to make a lasting scientific impact
and at the same time to rapidly exploit new concepts and
principles for the next technological revolution.
At Honda we have a tradition to think in larger
perspectives. Companies can only prosper in prosperous
societies. Society’s challenges are our challenges and
together we have to find solutions. At HRI our products are
ideas. You cannot buy ideas and there is no stock
exchange for them, but they are the real currency of our
future. Ideas lead to innovations and innovations enable
positive change. In order for this change to be sustainable
and lasting, it has to grow on a solid fundament of
knowledge. In a complex, globally connected world like
ours, this is more important than ever. At HRI we have
understood that science and innovation are inseparable
twins. Science without innovation neglects opportunities
for the future of our society and innovation without
science remains shallow and superficial. For us innovation
through science is more than a motto, it is our philosophy
and our guiding principle that allows us to always see the
road ahead without getting lost in the daily traffic.
Honda Research Institute Europe
The Honda Research Institutes are three in one. We share
the same philosophy and our research complements each
other as does our cultural uniquenesses. Intelligence
Science and Nano-Science are the fields where HRI
contributes to Honda and to society as a whole. Intelligent
systems will shape our future in a variety of forms, ranging
from cognitive robotics to thinking networks and from
accident-free mobility to efficient use of resources. In
biology, intelligence increases flexibility, adaptivity and
cooperation. It seems that for us to overcome the
challenges ahead, we need to “biologize” our technology
in those capabilities. Nano-Science holds the key to many
fundamental problems that arise in diverse fields ranging
from material science to ecology. Only if we understand
more about how complex systems develop and operate at
the smallest scales, we will be able to control them and to
set the right constraints for a stable and directed selforganization process. Innovation most frequently occurs at
the intersection of traditional disciplines where true
synergy can be harvested and where we start to think out
of the box. Both intelligence science and nano-science are
interdisciplinary and they themselves overlap.
In Europe, we focus on the biology of intelligence and
complexity. It is evident that the brain (from simple
nervous systems for controlling reflexes to the human
brain with its storehouse - the neocortex - for learning
from interaction) is the only truly intelligent system that we
know of. Understanding the major principles behind the
processing of information in the neural structure of
biological brains will open a plethora of opportunities for
making technical systems smarter. Of course brains are
not isolated, neither temporally nor spatially. They have an
evolutionary history and they are an integral and
inseparable part of biological individuals. This
embodiment defines which knowledge is actively acquired
for generating adaptive behaviors. Any endeavor to
understand the brain without either of these constraints is
unlikely to succeed. Brains are truly complex systems,
however, in living organisms there are plenty more.
Metabolic networks, immune systems, genetic and protein
interaction networks and ecosystems as a whole are
hardly understood and we cannot but marvel at the degree
of precision, robustness and sustainability of their
operation. The biology of intelligence and complexity or
the intelligence and complexity of biology will lead us to a
new holistic view on “biological technology” and
engineering.
Welcome to the Honda Research Institute Europe!
Prof. Dr. Edgar Körner
Chairman
Yoshiaki Sakagami
President
Prof. Dr. Bernhard Sendhoff
Chief Technology Officer
HRI-EU shares the building with
Honda R&D Europe (Deutschland) GmbH
who focus on automotive and
motorcycle technology.
2
Welcome
Welcome
3
Brain-like Intelligence
Understanding how the brain works
Local and global models of brain function
The brain is organised in a highly distributed and
decentralised way, with control acting on a local as well as
on a global basis.
We investigate the brain at different levels of description,
ranging from very detailed networks of spiking neurons to
models comprising several specialised brain areas that
interact on a system level.
At the most detailed level, we investigate which properties
of nerve cells and neural dynamics are essential for brain
function. At a more functional level, we investigate the
active role of visual processes and memory required for
scene understanding. At a high level we study how the
brain learns visual objects and builds up sensory
representations in a self-organised way.
4
Brain-like Intelligence
The human brain is the most powerful information
processing system we know of. Its universality, flexibility
and robustness are unrivalled by any technical system,
despite the long-lasting efforts and advances in
information science and computer engineering.
At HRI-EU we aim to identify and understand the principles
of brain-like information processing for building intelligent
systems. Our research concentrates on the functions of the
brain, inspired and guided by findings from neuroscience.
Thus, if we want to transfer the power of brain-like
processing to technical systems, we must understand how
the brain works. We have to investigate how the brain
learns, how it represents knowledge, and how it adapts
and organises behaviour according to what it knows and
perceives. These capabilities enable intelligent systems to
operate autonomously in complex and changing
environments.
Brain-like Intelligence
5
Discovering the neuronal mechanisms
of brain function
Detailed neural modelling and simulation
Intelligence is the collective work of billions of nerve cells
that communicate through a dense web of connections.
Already 1 cubic millimetre of cortex contains more than
100,000 neurons with over a billion connections.
To understand how neural circuits and brain areas work
and how they implement brain functions, we create
models that we simulate on a computer. In simulations, we
can then compare different models or subject one model
to different stimuli or parameters. Thus, we can test and
extend our understanding of how the brain works.
The software tools for this research must optimally exploit
the performance of modern computers and computer
clusters. Together with our partners in universities in
Germany, Norway, and Japan, we develop NEST, one of
the leading simulation tools for large nervous systems.
Which properties of nerve cells and neural circuits are
essential for brain function? To answer this question, we
look in detail at the neuronal structures and processes in
the brain and how they contribute to intelligence. With a
focus on visual processing in mammals, we combine
results from brain sciences, mathematical analysis and
computer models to investigate and answer the following
questions:
• Which neuronal circuits are responsible for perception,
behaviour and learning?
• How is processing organized and controlled?
• How does the brain maintain its stable and robust
function?
Our research is guided by a few basic assumptions:
• Columns in the cortex
The cortex is composed of neuronal circuits, called
columns. They are the smallest functional unit of the
cortex.
• Spike-latency code
Neurons produce brief electric pulses, called spikes.
Information is carried by the latency of spikes relative
to stimuli or other spikes.
• Bottom-up and top-down processing
The brain has two main processing modes. In bottomup mode, signals flow asynchronously from the
sensors to the central areas. In top-down mode, large
parts of the cortex synchronously focus their resources
on a common target, e.g. when we attend to an object.
• Rhythmic control
Rhythmic neuronal activity is abundant in the brain. It
is an important control signal that switches the cortex
from bottom-up to top-down processing.
6
Brain-like Intelligence
Brain-like Intelligence
7
Exploring and understanding visual scenes
Hierarchies for motion estimation
To resolve motion uncertainties and ambiguities during
visual perception, we use a motion estimation approach
analogous to biological V1-MT processing which
incorporates probabilistic motion descriptions. The motion
estimations are integrated over space, time and spatial
scales using Bayesian prediction and refinement methods.
We research hierarchical motion estimation models from
which robust local velocity and confidence estimates for
both slow and high speeds can be computed. From there,
we extract and estimate large-scale motion patterns that
serve as a basis for perceptual object pop-out, detection
and segmentation.
Such a motion estimating system is able to overcome local
ambiguities and exhibits response time-courses as found
in physiological measurements of motion selective cells. In
addition, motion-based object segmentation is a key
capability for the decomposition and interpretation of
dynamically changing visual scenes.
After decades of research on visual processing, the puzzle
of visual cognition is still far from being resolved. What
happens within the short time of a few 100 ms when
observing a dynamic visual scene like the one on the right?
How does the brain select the necessary information and
neglect the irrelevant parts?
To investigate this, we build models of visual cognition
that focus on the active and task-dependent parts of visual
processing. It is our goal to
• understand the visual mechanisms that allow selective
processing for dynamically varying tasks, and
• model visual processing as a task-dependent control
process that consistently links sensory modules to an
internal scene representation.
It would have to detect and isolate objects in the scene and
estimate its dynamic parameters. Additionally, it would
have to be able to lock attention on parts of the visual
input, mentally following them over time. Furthermore, it
would have to decide on serialisation strategies for visual
inspection and to couple single visual processes in an
active, intentional vision architecture. Here, a key issue is
bi-directional processing, with topdown prediction shaping
the afferent information flow.
Our long-term target is to create a system that
continuously and autonomously analyses a visual scene in
a context and task-dependent way, driven by sensory
input, scene memory and world knowledge.
We concentrate on dynamic scenes and the role of
temporal and predictive processes for brain-like visual
analysis. A visual system with scene understanding
capabilities would have to be able to accomplish a whole
series of visual subtasks.
Offenbach am Main
8
Brain-like Intelligence
Brain-like Intelligence
9
Learning of sensory representations
Hierarchies for the representation of visual objects
We develop models of object representation that focus on principles of hierarchical
feature detection in the „what“-pathway of cortical visual processing. Along the
hierarchy, the visual input is decomposed into more and more specific features and
parts. Using this decomposition, objects can be rapidly learned by capturing their
appearance variations under rotation. The results are view-tuned neurons, which
are also found in the primate visual cortex. We apply these models to the
interactive training of individual objects and more general object categories. A
particular focus is the intuitive presentation of objects to the learning system
without strong constraints on the environment.
10
Brain-like Intelligence
Learning and recognition of visual objects is a task so easy
for humans that we rarely notice its importance in our
daily activities. There is broad neurobiological evidence
that this ability is achieved by hierarchical processing of
increasingly complex visual feature representations in the
brain. Therefore, we develop functional models of this
pathway for the recognition of objects and visual
categories.
We focus on learning algorithms to obtain such a
hierarchical feature decomposition, from simple low-level
detectors to analytic parts-based features that are directly
linked to meaningful object categories. While the lower
levels are general and capable of representing arbitrary
objects, the higher levels are required to be more specific
to certain visually related tasks.
Our target is to obtain a self-referential visual knowledge
representation about objects that is incrementally acquired
by learning. Similar to well-known concepts in human
learning, we use a functional separation into working,
short-term and long-term memory. This enables new
approaches to fast learning of objects in interaction with a
human teacher. Another important aspect is the
embedding of the visual knowledge acquisition process
into a task-driven behaviour. This ensures that learning is
goal-directed and driven by direct feedback from the
environment.
Brain-like Intelligence
11
Evolutionary and
Learning Technology
The basic evolutionary loop
The basic evolutionary loop consists of the following six
elements:
1. select the mating partner;
2. recombine the genetic material;
3. mutate the genetic material;
4. evaluate the individuals;
5. select the best individuals and
6. adapt the individuals, i.e., life-time learning.
These elements and operations are not strictly sequential.
They operate on different spaces and time scales:
variations are introduced in the genotype space - on the
DNA/RNA level (blue in the image); evaluation and
selection act on the phenotype space - the space of
individual characteristics (green in the image). The
phenotype is a result of the decoding of the genetic
information and the environmental influences during
development.
Biological evolution demonstrates that sustained
innovation and improvement of complex systems ranging
from the ecosystem in the rainforest to information
processing in the human brain are possible. It is
remarkable that such innovations are achieved in perfect
harmony with the respective environment.
Evolutionary design is situated design,
i.e. issues like robustness,
evolvability and environmental
impact are as much inherent in the
process as are the system‘s level
holistic operation and the
constructive use of non-linearities
and non-equilibrium states.
However, evolution has gone one
step further by inventing learning to
allow adaptation to take place on
arbitrary time scales. Individual and
cultural learning enables systems to cope
with environmental diversity and with rapid
environmental changes.
Furthermore, we want to translate this knowledge into
unique technology that enables us to understand and
design complex systems:
• Understanding evolution
understanding the organisation of the genetic
representation as a process that controls cellular
growth of the phenotype under environmental
influence during development
• Evolution to create
evolutionary system optimization of the
structure of complex and adaptive systems
and its application to multi-disciplinary
technical problems
• Evolution to understand
research the evolution and the selforganization of early nervous systems
embedded in the cellular organization of
artificial organisms
It is the mission of HRI-EU to gain a deeper
understanding of both processes and their
interaction: evolution and learning.
12
Evolutionary and Learning Technology
Evolutionary and Learning Technology
13
Understanding evolution
A different perspective:
Vector Field Embryogeny
In biology, regulation during development is generally
modelled by gene regulatory networks (GRN) whose
structure captures the complex dynamics of the process.
The resulting connectivity patterns allow us to find basic
units so called motifs that occur in different networks and
organisms.
At the same time, it is notoriously difficult to evolve gene
regulatory networks in silico experiments.
We look at the evolution of GRN dynamics from a novel
perspective: instead of concentrating on the network
structure and its changes, we focus on the dynamical
behavior of a system alone.
During development, genes, the basic information units of
the genome, are expressed and used to determine the
sequence of amino acids of proteins which are the most
basic structuring units of every organism. The expression
pattern of genes is regulated in multiple ways, e.g. by
proteins and microRNA. The role of this regulation as a
driving force behind biological evolution has long been
underestimated. Today, it is clear that much of biological
innovation is a result of regulatory changes: evolution is
teaching old genes new tricks.
We have built a model of genetic regulation and combined
it with the simulation of physical interaction between cells.
Gene regulatory networks influence cell function, like
division and apoptosis. By studying the evolution of cell
growth based on the manipulation of regulatory
information, we are able to get a better understanding
about how regulation must be structured to enable
innovation.
Biological regulatory systems often seem to be overly
complex and redundant. By analysing measures of
robustness, evolvability and redundancy, we can shed
some light on issues like modularization and duplication of
regulatory structures. This insight can be used to better
understand biological evolution as well as for the cellular
composition of novel materials and structures.
In the framework of Vector Field Embryogeny, we
superimpose basic field elements to describe the system
dynamics.
This enables evolution to manipulate the dynamics
directly, which makes the optimization process more
efficient and helps us to analyse evolutionary innovation.
14
Evolutionary and Learning Technology
Evolutionary and Learning Technology
15
Evolution to understand
When learning hinders evolution
– lazy for the benefit of all
Things that are good for the individual do not have to be
good for the group. This intuitively correct statement can
also be observed when studying the interaction between
learning and evolution.
Fast evolution needs large differences among different
individuals. If individuals can catch up through learning,
the individual differences are hidden - we speak of the
hiding effect - and evolution slows down.
Using a theoretical model for a simple evolutionary
system, we can show under which circumstances learning
is beneficial for evolution and when it hinders evolution.
The structure of the brain reflects its current functionality
as much as its evolutionary history. In order to understand
the organisation of information processing in the brain, we
must analyse the major evolutionary transitions in the
structural realisation of brain-like information processing.
In simple simulation models, we study the onset and the
constraints of early transitions of the nervous system
using organisms like hydra or the flatworm as biological
examples.
There is a great similarity between the old and new
structures in the brain, such as the hippocampus and the
neocortex. This suggests that evolution of the brain is
constrained, and system-level changes made to the cortex
are limited. This can be attributed to the unique genetic
representation of the brain. Studies on gene expression
indicate that the complex temporal and spatial interactions
of functional and regulatory genes enable such
constrained evolution.
The relation between functional and regulatory
subsystems is a fundamental aspect of biological
information processing.
We study and analyse regulation and its influence on
dynamically stable growth processes during the evolution
of the cellular development of worm-like artificial
organisms. By observing the emergence of feedback loops
and the role of maternal gradients for initial symmetry
breaking during development, we can improve the
artificial evolutionary processes and get a better
understanding of biological evolution. The analysis of
neural development must be embedded into a
morphological framework. In our models, the evolution of
morphology and control are simulated together, which
enables us to study their mutual influence.
Different neural connectivity patterns dependent on morphological constraints
16
Evolutionary and Learning Technology
Evolutionary and Learning Technology
17
Evolution to create
Evolutionary design: duplication and
specialisation
Free Style Deformation is a unique way to manipulate
shapes and structures. It introduces a modular concept for
design modifications with Free Form Deformation (FFD). In
this context, a module is defined as the geometry which is
situated in an FFD sub-control volume. The combination of
common FFD techniques and module exchanges is an
efficient and highly flexible representation for complex
structures. A design is finally defined by a code specifying
which sub-control volumes have to be combined, and by
the deformations of each control point.
Evolutionary engineering is a powerful design method that
is neither restricted by the need for analytical models, nor
by designer prejudice. It is capable of exploiting directed
random changes adaptively by re-using and re-organizing
modular structures.
However, such a method requires a representation of the
design or structure and its variations, which allows
efficient exploration.
We have established a framework for evolutionary system
optimization, which allows the adaptation of the
representation to the local topology of the search space.
Furthermore, the changes to a system design are encoded
instead of the design itself which allows the framework to
hierarchically optimise complex shapes and structures.
Important system level properties like robustness and
efficiency are inherent in a natural design process like
evolution. Placing the population of designs in a changing
environment furthers the desired combination of stability
and flexibility. Additional biological mechanisms like cell
growth allow a structure-focused three-dimensional
design in a multidisciplinary and multi-objective quality
space.
In nature, evolution is an asynchronous, massively parallel
process. In artificial system design, we can simulate such
an environment using large-scale computing clusters and
grids in order to flexibly assign computing power across
projects and continents. Currently, we employ Linux
clusters with over 600 compute nodes.
We employ this framework for the optimisation of engine
components, e.g. turbine blades, and of vehicle parts.
duplication of a modular
wing structure
18
Evolutionary and Learning Technology
Evolutionary and Learning Technology
19
Embodied Brain-like Intelligence
Self-development of practical intelligence
• The developmental process
Research an intelligent system that develops individual cognitive abilities based
on its practical experience in interacting with the external world.
• The initial conditions and architecture
Research a system that evolves itself from initially relatively few innate abilities
towards an autonomous and socially compliant / integrated partner.
• The constraints
Validate the researched systems interactively in real-time in the real world for
demonstrating the achieved performance.
20
Embodied Brain-like Intelligence
The concept of Embodied Brain-like Intelligence regards
intelligence as a system property, i.e. the result of the
well-orchestrated interplay of several elements including
vision, speech, thinking and acting.
Starting from an architecture inspired from developmental
biology and psychology, we equip systems with the ability
to autonomously learn and develop in interaction with the
natural environment, growing own cognitive abilities
based on experience.
One goal is to replace handcrafted design of complex
information processing systems by autonomous
adaptation and learning for solving specific tasks. The
feasibility of the scientific results is shown in the areas of
cognitive humanoid robots and intelligent assistant
systems, providing practical intelligence for a wide
spectrum of tasks.
In the area of concept and semantic acquisition the
concurrent learning of basic visual categories, words and
their associations during interacting with a tutor has been
successfully shown in an integrated system with the
HONDA humanoid robot. New motions can be acquired in
interaction and employed immediately within an
elaborated framework for flexible motion planning and
control.
Several systems for the investigation of human-like
sensing have been researched and evaluated in the
automotive domain.
They represent approaches towards context based
dynamic scene analysis for advanced driver assistant
functions.
Embodied Brain-like Intelligence
21
System environment and architecture
Complex information processing and
behaviour controlling systems
In order to provide the optimal framework for the system
research, we have created a software suite providing tools
for design, implementation, testing, distributed real-time
execution, maintenance and documentation of complete
large-scale systems and system elements. The framework
is already being used for creating research prototypes
comprising vision, speech and behaviour generation in
real-time. The complexity of the current systems is of the
order of several hundred modules, which is manageable
only with appropriate tools.
22
Embodied Brain-like Intelligence
The brain is the most complex structure known to man.
Researching brain-like systems raises implicitly the
question of how to deal with complex architectures
processing information inside of an artifact‘s control
system. The solutions shown by nature are developmental
and learning processes within a suitable control
framework. Following this example, we have to
understand what the necessary control framework is, what
the principles behind those developmental and learning
processes are and how to create and maintain the
resulting systems based on current or necessary future
computing technology.
For the targeted systems the research and creation of this
technological environment is not a trivial task, since it has
to cope with several contradicting requirements. It should
allow for all the necessary freedom of research, but it
should minimise the step from exploratory research
questions towards well-integrated reliable systems that
can be validated on a robot or a car.
Functional requirements may also be contradictory like
low latency and high bandwidth communication within
one system.
Within the above mentioned framework we are
researching global system properties on the architectural
level, like the internal control processes governing the
generation of interactive behaviour and the necessary
needs & drives behind continuous self-development.
Several architectural design principles for large-scale
distributed systems have been researched and formulated
paving the way towards more comprehensive artifacts
with a new performance quality.
Embodied Brain-like Intelligence
23
Towards interactive cognitive robots
Embodied systems
Researching intelligent systems is a long
term endeavor that has to be pursued in an
incremental fashion.
Since concepts will build hierarchically on
each other it is important to ensure that all
novel fundamental research results scale
properly towards real world requirements.
An embodiment evaluates this kind of
scaling and provides a proper grounding for
abstractions and symbols within interacting
systems.
Humanoid robots fit well into our anthropocentric
environment and, based on their similarities, enable us to
develop and test methods for intelligence, interaction and
social behaviour.
As prerequisites, a general and rich environmental
perception, as well as basic movement and behaviour
capabilities have to be established before cognitive
abilities can be realized.
Among the key questions is how to learn, represent and
use not only purely sensory information about our
environment - such as object appearances, but how to
include functional information and semantic knowledge:
e.g. what can an object be used for - its so-called
affordance.
Methods to realize this include the imitation of human
movements and tasks via imitation learning as well as
exploration and organizing reward for successful actions.
Those abilities are needed to design truly helpful and
socially compliant robots blending seamlessly into our
natural environment.
Intelligent driver assistance
For supporting the driver and improving the safety of all
traffic participants, we research driver assistance systems
exhibiting human-like robustness and flexibility. Towards
building an integrated system capable of orchestrating all
the different functionalities needed for such an artificial
Co-Pilot, we aim at a flexible control architecture with
bottom-up (sensor-driven) and top-down (predictiondriven) information flow for generating safe driving
behavior in complex traffic scenes.
Focusing first on the visual processing stages analysing
the scene, we have developed a vision system that is
inspired by human visual perception. Using a tunable
attention system and state-of-the-art perception
algorithms, the prototype is capable of analysing the
scenery for task-relevant information.
A key element for future driver assistance systems and
intelligent systems in general are brain-like learning
methods that allow to acquire multiple situationdependent prediction models.
Especially for the task of driving, we investigate spatial
representations with associated prediction models that
support the analysis of the current scene as well as the
prediction of possible future changes that may pose a
danger.
24
Embodied Brain-like Intelligence
Embodied Brain-like Intelligence
25
Acoustic scene decomposition for
developmental language and semantics learning
Language acquisition
The acquisition of language observed in 1-2 year old
children represents only a tiny portion of emotional,
perceptual and behaviour skills developed continuously
even before birth.
Infant developmental psychology begins to shed some
light on the developmental steps, the causal influence
linking these steps, the interactions between sensorymotor modalities and the determining role of the
environmental social relations that support each learning
step. In this context, we consider language, speech and
auditory perception as part of a multimodal, socially
interacting system going beyond input/output interfaces.
Speech is a natural communication channel as well as a
crucial element in linking perceptions, actions and
behaviours to a specific meaning (semantics). Our
research focuses on incrementally acquiring the internal
representations for language and semantics following a
developmental path. One of the first steps along this path
is the decomposition of an acoustic scene into its basic
behaviorally relevant elements.
Our approach is based on several features including:
• binaural hearing, using a biomimetic ear for the
improved 3D localisation of sound sources,
• spectro-temporal feature extraction for the robust
recognition of speech units in natural echoic and noisy
environment,
• robust pitch and formant extraction and tracking.
The next step is the further decomposition of streams into
syllables and words.
Here we concentrate our research on algorithms able to
create, in interaction with the tutor, on-line categories of
speech units that can optimally explain the perceived
audio stream. A further step consists of linking the stream
of recognized speech building blocks to the perception of
other multi-modal cues in order to create semantic
perceptual categories.
The particular challenge here consists of letting the system
discover new feature spaces that optimally describe the
semantic task using only a limited set of tutoring steps.
The latest results concern the learning of perceptual or
motor categories in an interactive scenario between a tutor
and a humanoid robot. Furthermore, we investigate how
speech production can develop in an imitative context.
These features are the basis for the creation of acoustic
proto-objects allowing the representation of audio scene
elements.
26
Embodied Brain-like Intelligence
Embodied Brain-like Intelligence
27
HRI European Graduate Network
Partnership in science projects
HRI European Graduate Network
Numbers in 2010
• 16 partnership in science projects in 6 countries
• one guest scientist
• 22 PhD students
• 19 MSc and Diploma students
• 7 internships
• we provide two lecture series:
- Technical University Darmstadt
- University of Applied Science Frankfurt
The HRI European Graduate Network is the home of all
graduate students that are supported by the Honda
Research Institute Europe and of all of our associates who
want to pursue a Doctorate degree in addition to their
work at our institute.
The HRI European Graduate Network shall foster the spirit
of togetherness among all of „our“ students, further
strengthen our advisory role and create a network which
shall last beyond the time the students spend at our
institute.
Nearly all of the students of the HRI European Graduate
Network are supported by Partnership in Science projects.
Partnership in Science projects are conducted together
with universities and research institutes and facilitate open
discussions on the research challenges of the 21st century
which cannot be addressed by a single institution.
• HRI-EU joins its scientific expertise with academic
partners to advance our understanding of intelligent
systems for the benefit of our society.
• HRI-EU contributes to scientific programmes and open
source software development.
• HRI-EU participates in the education of the next
generation of researchers in the local and global
community (joint PhD, Diploma and MSc
programmes).
HRI European Graduate Network Symposium 2008
28
HRI European Graduate Network
HRI European Graduate Network
29
Irene Ayllón Clemente
”HRI-EU supports me in discovering
new ways to extend my imagination
and to harvest my potential.“
Chen Zhang
Project Title:
Project Title:
Representations for Incremental Speech Acquisition
Prediction Strategies for Dynamic Objects in a Visual
Scene
Partner:
Research Institute for Cognition and Robotics
Bielefeld University
HRI European Graduate Network
Technical University Darmstadt
PhD Advisory team:
Dr. Britta Wrede (Bielefeld University)
Prof. Dr. Gerhard Sagerer (Bielefeld University)
Dr. Martin Heckmann (HRI-EU)
Dr. Heiko Wersing (HRI-EU)
Prof. Dr. Jürgen Adamy (TU Darmstadt)
Dr. Julian Eggert (HRI-EU)
Dr. Martin Heckmann (HRI-EU)
I. A. Clemente, M. Heckmann, G. Sagerer, and F. Joublin.
Multiple sequence alignment based bootstrapping for
improved incremental word learning. 35th International
Conference on Acoustics. Speech and Signal Processing
(ICASSP), 2010.
30
Partner:
PhD Advisory team:
Recent Publication:
”The fusion of scientific and industrial
competence at HRI-EU inspires me
to tackle challenging problems that
impact the real world.“
Recent Publication:
C. Zhang, J. Eggert, and N. Einecke. Robust tracking by
means of template adaptation with drift correction.
Computer Vision Systems, 7th International Conference
(ICVS), Springer Verlag, 2009.
HRI European Graduate Network
31
Heiko Lex
Lisa Schramm
Project Title:
Project Title:
Cognitive Planning and Motor Adaptation in Manual
Action
Simulation of the cellular growth of the morphology and
the nervous system of artificial organisms
Partner:
Partner:
Research Institute for Cognition and Robotics
Bielefeld University
Technical University Darmstadt
PhD Advisory team:
”Experiencing both a university and
an industrial research environment
provides me with a wider perspective
than most other PhD students have.“
PhD Advisory team:
Prof. Dr. Thomas Schack (Bielefeld University)
Dr. Yaochu Jin (HRI-EU)
Prof. Dr. Jürgen Adamy (TU Darmstadt)
Prof. Dr. Bernhard Sendhoff (HRI-EU)
Dr. Marc-Oliver Gewaltig (HRI-EU)
Recent Publication:
Recent Publication:
H. Lex, M. Weigelt, and T. Schack. The relationship
between sensorimotor adaptation and cognitive
structures. Journal of Sport & Exercise Psychology, 31,
2009.
L. Schramm, Y. Jin, and B. Sendhoff. Emerged Coupling of
Motor Control and Morphological Development in
Evolution of Multi-Cellular Animats. 10th European
Conference on Artificial Life, Springer Verlag, 2009.
”I enjoy having the opportunity to work
in a fruitful and inspiring environment
and to pursue my research interests in
the area of movement organization .“
32
HRI European Graduate Network
HRI European Graduate Network
33
Selected Publications
E. Nordlie, M.-O. Gewaltig and H. Plesser. Towards
reproducible descriptions of neuronal network models.
PLoS Computational Biology, 5(8), 2009.
S. Schrader, M.-O. Gewaltig, U. Körner, and E. Körner.
Cortext: a columnar model of bottom-up and top-down
processing in the neocortex. Neural Networks, 22(8),
1055-70, 2009.
V. Willert and J. Eggert. Dynamic Visual Motion
Estimation. Book chapter in the IN-TECH Machine Learning
series, 2009.
V. Willert and J. Eggert. A stochastic dynamical system for
optical flow. International Conference on Computer Vision
(ICCV) 2009, Kyoto. Workshop on Dynamical Vision. Best
Paper Award.
L. Graening, S. Menzel, M. Hasenjäger, T. Bihrer, M.
Olhofer, and B. Sendhoff. Knowledge extraction from
aerodynamic design data and its application to 3d turbine
blade geometries. Journal of Mathematical Modelling and
Algorithms, 7(4), 329-350, 2008.
I. Paenke, B. Sendhoff, and T. Kawecki. Influence of
plasticity and learning on evolution under directional
selection. American Naturalist, 170(2), 1-12, 2007.
T. Steiner, Y. Jin, and B. Sendhoff. Vector field
embryogeny. PLoS ONE, 4(12), e8177, 2009.
H.-G. Beyer and B. Sendhoff. Robust optimization: A
comprehensive survey. Computer Methods in Applied
Mechanics and Engineering, 196(33-34), 3190-3218, 2007.
B. Sendhoff, E. Körner, O. Sporns, H. Ritter, and K. Doya,
editors. Creating Brain-like Intelligence. Number 5436 in
LNCS. Springer Verlag, 2009.
M. Gienger, M. Toussaint, and C. Goerick. Whole body
motion planning: Building blocks for intelligent systems,
Motion Planning for Humanoid Robots, Springer Verlag,
2010.
T. Michalke, J. Fritsch, and C. Goerick. A biologicallyinspired vision architecture for resource-constrained
intelligent vehicles. Computer Vision and Image
Understanding, 2010.
A. Ceravola, M. Stein, and C. Goerick. Researching and
developing a real-time infrastructure for intelligent
systems - Evolution of an integrated approach. Robotics
and Autonomous Systems, 56(1), 14-28, 2008.
C. Goerick, et al. Interactive online multimodal association
for internal concept building in humanoids, IEEE-RAS
International Conference on Humanoids, 2009.
C. Gläser, M. Heckmann, F. Joublin, and C. Goerick.
Combining auditory preprocessing and bayesian
estimation for robust formant tracking, IEEE Transactions
on Audio, Speech, and Language Processing, 18(2),
224-236, 2010.
S. Menzel and B. Sendhoff. Representing the change: free
S. Hasler, H. Wersing, and E. Körner. Combining
reconstruction and discrimination with class-specific
sparse coding. Neural Computation, 19(7), 1897-1918,
2007.
S. Kirstein, H. Wersing, and E. Körner. A biologically
motivated visual memory architecture for online learning
of objects. Neural Networks, 21, 65-77, 2008.
34
Selected Publications
form deformation for evolutionary design optimization. In
T. Yu, et al., editors, Evolutionary Computation in Practice,
Springer Verlag, 63-86, 2007.
H. Guo, Y. Meng, and Y. Jin. A cellular mechanism for
multi-robot construction via evolutionary multi-objective
optimization of a gene regulatory network. BioSystems,
98(3), 193-203, 2009.
Selected Publications
35
Honda Research Institute
Europe GmbH
Phone: +49 (0) 69-8 90 11–750
Fax:
+49 (0) 69-8 90 11–749
Carl-Legien-Str. 30
63073 Offenbach / Main
Germany
Email:
Web:
[email protected]
http://www.honda-ri.de
HRI-EU offers internship places for
students in higher semesters from
European and international universities.
Within our HRI European Graduate
Network programme it is possible for
PhD, MSc and Diploma students to
spend time at HRI-EU.
innovation through science