Extreme Sensitive Robotic
A Context-aware Ubiquitous Learning
Nicolas Verstaevel
1,2
, Christine Régis
1
, Valérian Guivarch
1
,
Marie-Pierre Gleizes
1
and Fabrice Robert
2
1
IRIT, Université Paul Sabatier, Toulouse, France
2
Sogeti High Tech, Toulouse, France
Keywords: Ambient Intelligence, Self-organizing Systems, Machine Learning, Adaptive Multi-Agent Systems,
Robot and Multi-robot Systems.
Abstract: Our work focuses on Extreme Sensitive Robotic that is on multi-robot applications that are in strong
interaction with humans and their integration in a highly connected world. Because human-robots
interactions have to be as natural as possible, we propose an approach where robots Learn from
Demonstrations, memorize contexts of learning and self-organize their parts to adapt themselves to new
contexts. To deal with Extreme Sensitive Robotic, we propose to use both an Adaptive Multi-Agent System
(AMAS) approach and a Context-Learning pattern in order to build a multi-agent system ALEX (Adaptive
Learner by Experiments) for contextual learning from demonstrations.
1 INTRODUCTION
The drastic reduction in the cost of electronic
equipment allows populating our environment with a
multitude of devices and functions of rich interaction
capabilities. Information technologies, formerly
confined inside computers, are now distributed in
our homes, factories and companies. Those ambient
systems (also called ubiquitous systems) are
characterised by their dynamic and their complexity:
a huge number of heterogeneous devices evolves
autonomously and new devices can appear or
disappear at any time (Perera, 2014). One of the
desired properties for such systems is the ability to
self-adapt to the specific and changing needs of its
users. Nevertheless, such adaptation is complex
since we can make no a priori supposition on the
task to perform or on the entities composing the
system. Furthermore, users in these systems can
interact in several ways and this interaction brings
both new challenges and new solutions.
More and more works among the robotic
community focus on the design of physically
distributed applications where autonomous robots
interact with other systems to perform complex and
changing tasks in interaction with humans
(Brambilla, 2013). Factories of Future (FoF) are a
good illustration of this as they involve multiple
entities evolving in a complex and highly dynamical
environment with a need for sustainability and
adaptability to end-users (Siciliano, 2014). Hyper-
connectivity of FoF offers a new challenge to their
designers: how to handle the complexity and
dynamic brought by the affluence of data coming
from electronic systems and human activity.
In this paper, we propose an approach named
Extreme Sensitive Robotic to use the inherent
interactivity of ambient systems as the motor of self-
adaptation. The system learns the way users interact
with it. With this approach, the design is not guided
by the finality, but by the system’s functionalities.
Each functionality is then seen as an autonomous
system having the capacity to self-adapt its
behaviour to what it can perceive from its
environment (including human activity). Self-
observation capacities allow each functionality to
correlate its own activity to the observation it makes
from its environment.
First, we consider the use of Learning from
Demonstration, a paradigm to dynamically learn
new behaviours (Argall, 2009), mainly studied in the
robotic field. Next, we will dress our vision of
Extreme Sensitive Robotic and propose the use of the
Adaptive Multi-Agent Systems (AMAS) to achieve
this vision.
242
Verstaevel N., Régis C., Guivarch V., Gleizes M. and Robert F..
Extreme Sensitive Robotic - A Context-Aware Ubiquitous Learning.
DOI: 10.5220/0005282002420248
In Proceedings of the International Conference on Agents and Artificial Intelligence (ICAART-2015), pages 242-248
ISBN: 978-989-758-073-4
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
2 LEARNING FROM
DEMONSTRATIONS
Since we consider exploiting the interactivity of
ambient systems, we need a user-centred approach.
Observing the field of application made in robotic,
one approach has retained our attention. Learning
from Demonstrations (LfD, also called Imitation
Learning or Programming by Demonstrations) is a
paradigm to allow a system to autonomously learn
new behaviours from demonstrations (Argall, 2009).
The hypothesis is that a system can learn its
behaviour from the observation of a human’s
activity. To adapt behaviour to specific needs, the
classical approach is to decompose the task to realise
into sub-tasks and to manually program the
resolution of each sub-task. At the opposite, the LfD
approach proposes that a new task can be derived
from the observation of human activity (Calinon,
2008). As it is a social learning that requires for the
user no expertise on the system (Dautenhahn, 2003),
LfD seems to be a good paradigm when the task to
perform is not a priori known.
The idea of LfD took its origins in the 80s, thus
there has been a resurgent work on the subject
(Argall, 2009) (Calinon, 2008). Applications of LfD
are various, like learning to fly (Sammut, 2002),
learning to react to an artefact (Knox, 2013) or
learning manipulation task in an industrial
environment (Tosello, 2014).
Despite its variety, LfD suffers from some
limitations in term of genericity. Indeed, in the
variety of LfD applications, we can distinguish two
families of algorithms: those using supervised
learning and those using reinforcement learning.
Both technics have their pros and cons.
Supervised learning algorithms use labelled data
to construct a decision model. Once this model is
constructed, the system is able to autonomously
follow a decision process. These algorithms include
the notion of demonstration, each labelled data
acting as a demonstration and seem relevant in the
LfD context. However, they are not robust to new
demonstrations. Indeed, the two-step operating mode
separating learning phase from behavioural phase
restricts the use of these algorithms for real-time
learning. A new demonstration implies to repeat the
complete learning phase.
Reinforcement learning algorithms use a
feedback function associating the performance of an
action to a utility value. A reinforcement learning
algorithm tries to establish an optimal control of its
environment in order to maximise this utility value.
The main advantage of these algorithms is that they
explore autonomously and in real-time the possible
states to find an optimal control policy.
Nevertheless, this advantage can also become a
default since to learn not to destroy itself, a robot
needs to experiment its destruction and receive a
negative feedback in response. Furthermore, those
algorithms suffer from a lack of genericity, since the
design of the feedback function requires knowledge
on the task to perform and the environment.
Some hybrid techniques try to use both
algorithms in the same approach. For example,
(Knox, 2013) infers a feedback function from
feedbacks coming from the user and then uses this
learned function to make reinforcement learning.
However, these approaches, which attempt to
remove limitations by combining reinforcement
learning and supervised learning, often simply
combine their limitations.
One of the current challenge in this domain is
then to propose a generic approach, independent to
the task to perform and able to self-adapt in real time
both to the specific needs of its users and to the
dynamic of its environment. However, we see that
LfD is an interesting paradigm for service robotics
as it is user-centred and allows day-to-day
interactions.
(Nehaniv, 2001) postulates that LfD has to answer
the following key questions:
- What to imitate?
- How to imitate?
- When
to imitate?
- Whom to imitate?
On the next section, we propose to answer to those
questions with an approach that is not guided by the
goal to achieve, but by the functionalities composing
the system.
3 EXTREME SENSITIVE
ROBOTIC
Since their beginning, robots become ever more
elaborated in terms of both hardware and software. It
may be considered that these systems will be truly
complex, complexity increased by their necessary
adaptation to the high dynamic of their environment
(including humans) and a dynamical coordination
with other robots or artificial ambient systems. This
is the presupposition made by Extreme Sensitive
Robotic where the system’s design should be made
bottom-up rather than top-down. We have at our
disposal libraries of various components realising
functions rather than objectives. Thus, a robot
ExtremeSensitiveRobotic-AContext-AwareUbiquitousLearning
243
consists of the aggregation of necessary functions to
satisfy user’s needs, but a group of robots has to be
considered exactly in the same way: a set of macro-
functions (each robot) working in coordination. An
Extreme Sensitive Robot is made of simpler Extreme
Sensitive Functions, each of these functions having
the ability to self-adapt to what it can perceive from
its environment. Thereby, the overall activity results
from local interactions between Extreme Sensitive
Functions and the environment. Such adaptation
requires both openness and self-observation
capacities, well known in the multi-agent field.
Kaminka's previous work focused on the
necessity to roboticists and Multi-Agent System
community to work together as they share same
interests into autonomous decisions in complex
environment (Kaminka, 2012). As a matter of fact,
there has been an increasing work considering robots
as agents into multi-robot application for complex
tasks such as collective rescue (Lacouture, 2012)
(Couceiro, 2013), collective exploration or tasks
allocation (Navarro, 2012). A new challenge that
rises among robotic applications is the integration of
robots into smart environments where many
heterogeneous devices are in hyper-interactivity with
other systems and humans (Broxvall, 2006). In these
systems, robots have the challenging task to use this
hyper-connectivity to adapt their behaviour to
achieve complex and changing goals. The traditional
reductionist approach is not relevant for such
systems where no assumption can be made on goals
to achieve or the dynamic of the environment.
Extreme Sensitive Robotic proposes that interactivity
of such systems is more related to an autonomous
observation of the dynamic of the surrounding
environment (including the consequences of its own
mobility) than the explicit communication between
system entities. The absence of this explicit
communication reduces the need for a priori
knowledge on the system and allows each
functionality to be designed separately. Self-
observation capacities make the system extremely
sensitive to its environment allowing it to integrate
changes in its environment into its decision process.
In the remainder of this section, we focus on
important features that should be taken into account
within Extreme Sensitive Robotic.
The architecture of a robot is globally composed
of functions of perception, action and decision. A
robot that performs well has to permanently make
cooperate these three functions in association with a
loop-back correlating the consequences of its own
actions with the observation of changes on its
surrounding environment. This is what Brooks
(Brooks, 1990) expressed as the physical grounding
hypothesis. In opposition to the classical reductionist
approach, the physical grounding hypothesis
postulates that physical interactions with the
environment are the primary source of constraints
for the design of intelligent systems. Thus, there is
no need of symbolic representation of the
environment leading to a complex decision making.
On the contrary, the system's behaviour is a reaction
to a stimulus coming from its environment. Brooks’s
subsumption architecture (Brooks, 1987) is the
origin of behaviour-based robotics. In Brooks's
architecture, a robot controller is built layer by layer,
each layer responsible for one behaviour. The
subsumption architecture enables the robot to select
the most adequate layer in reaction to what it
perceives from its surrounding environment. The
postulate that can be made of Brooks's work is that
direct interactions with the environment have a
strong influence on the robot's decision process.
Making an Extreme Sensitive Robotic consists in
making it sensible to variations in the perception of
its surrounding environment and not to an internal
state representation. Extreme Sensitive Robotic is all
about to sense, not to model. Pioneering work of
Walter Grey (Walter, 1950) (Walter, 1951) in early
50s has shown that even without any form of
computational intelligence, a machine can produce a
behaviour that one can consider as a smart
behaviour, even showing some learning skills. In
Grey's robots, an active interaction between sensors
and actuators allows a strong interaction with the
surrounding environment and the emergence of a
behaviour. Braitenberg (Braitenberg, 1986)
proposed a set of vehicles where sensors are in direct
interaction with actuators. The sensors could have an
exciting or inhibitory influence on the actuators.
With an exciting influence, more the sensor is
excited more the actuator is excited. On the contrary,
with an inhibitory influence, more the sensor is
excited, less the actuator is excited. Depending of
the type of influence and how sensors and actuators
are connected, the same robot (same actuators and
same sensors) could perform radically different
behaviours. The only difference lies in how sensors
and actuators are connected. It results that the
robotic entity is not only influenced by its
surrounding environment but also by the nature of
the influence between sensors and actuators. Making
an Extreme Sensitive Robotic is then considering
what occurs both outside and inside the robot's body.
Pfeifer (Pfeifer, 2002) has named the relation
between an entity and its environment as the
embodiment relation. Pfeifer postulates that the
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
244
behaviour of an entity is highly influenced by the
environment in which it is immersed but also by its
own body. To illustrate this phenomenon, Pfeifer
proposed the following experiment: looking at the
trajectory of an ant walking on rocks, one could say
that the behaviour of the ant is smart. Indeed, as the
ant is avoiding obstacles, the trajectory appears to be
complex. However, if this ant were a thousand times
larger, the ant would not be blocked by stones
anymore and would then walk in a straight line. The
same observer would then say that the ant behaviour
is not smart any more. Whereas there has been no
change on the ant's mind or on the environment, the
observed behaviour differs. A change in the ant body
has changed the effect its body produces on the
environment. This philosophical experiment shows
that any changes in the body could radically change
the relation between an entity and its environment.
The same idea has to be applied to robotic because
some part of a robot could disappear (for example, a
sensor failure) or functionalities added during robot's
activity. Even two robots with the same architecture
can have electronic differences such as a motor
rotating faster than the other does. These
modifications of robot's body could have a strong
impact on consequences of robot actions on the
environment. Making an Extreme Sensitive Robotic
consists in making it sensitive to the effects that their
actions have on the environment.
Thus, building an extremely sensitive robot is then
to make it sensitive and adaptive to:
- How its environment evolves,
- How its functionalities interact,
- Appearance or disappearance of functionalities
and their effects on the environment.
Unlike traditional robotic approach, which consists
in building robust controllers for robotic platforms,
the Extreme Sensitive Robotic approach deals with
functionalities. Each functionality is designed to be
self-adaptive, as self-adaptation is driven by a local
observation of the environment. Thus, a robot is a
set of functionalities interacting through the
environment.
As each functionality has to correlate its own
activity to both users and the observation of the
environment, the use of LfD seems here relevant.
Thus, answering (Nehaniv, 2001) questions would
be:
- What: how users use system functionalities.
- How: by correlating the performance of an
action to the observation of the environment.
- When: each time a user uses a functionality.
- Who: whoever has to act on a functionality.
To enable these functionalities to self-adapt, we
propose to use the Adaptive Multi-Agent System
theory, which is presented now.
4 AMAS THEORY
4.1 Landscape
The Adaptive Multi-Agent System theory (Capera,
2003) addresses the problematic of complex systems
with a bottom-up approach where the concept of
cooperation is the core of self-organisation. The
theorem of functional adequacy (Camps, 1998)
states that for all functionally adequate systems,
there is at least one system with a cooperative
internal state that realizes the same function in the
same environment. A general definition of
cooperation could be the golden mean between
altruism and selfishness (Picard, 2005). Three
mechanisms allow repairing an uncooperative state
(Capera, 2003):
- Tuning: the agent adjusts its internal state to
modify its behaviour,
- Reorganisation: the agent modifies the way it
interacts with its neighbourhood,
- Evolution: the agent can create other agents or
self-suppress when there is no other agent to
produce a functionality or when a functionality is
useless.
The system will then self-organise to stay in a
cooperative state. From cooperative interactions
between the system's entities emerges a global
function that is more than the sum of the parts. This
theory is applied in Extreme Sensitive Robotic with
the Context-Learning pattern.
4.2 Context-learning Pattern
The term context refers to all information external to
the activity of an entity that affects its activity. This
set of information describes the environment as the
entity sees it (Guivarch, 2012). Context-Learning is
based on the idea that the activity of an entity is
correlated with the observation made by the entity of
its own context. Thus, when an action is performed
on (or by) an entity, this entity can make a
correlation between the performance of this action in
the current context and the effects of this action and
then, learns the relevance of this action in this
particular context. So the entity becomes able to
correlate what it feels to what it does and to reuse its
knowledge when it is confronted to an already
known context. The entity is then called context-
aware, which means that it is able to perceive,
ExtremeSensitiveRobotic-AContext-AwareUbiquitousLearning
245
interpret and use the information from its current
context in order to dynamically adapt its
functionality.
The Context-Learning pattern has been applied
to the control of bioprocess (Videau, 2011), the
control of engine and energy (Boes, 2013) and the
observation of users activity (Guivarch, 2012). In the
next section, we will present general principles of
the Context-Learning approach, based on an
Adaptive Multi-Agent System (AMAS).
4.3 Context-learning Principle
The Context-Learning process is the result of two
kinds of agents, each one being responsible for a
particular activity in the system:
- A Context Agent associates a low-level
context description (See section 4.4) to an action
proposal. It receives signals from the environment
and uses them to characterize the context. When the
current context belongs to the context description of
a Context Agent, this agent considers itself as valid,
which means the current context is relevant to make
an action proposal to its associated Controller Agent.
The action proposition is composed of the action
description itself, and information about the
relevance level to perform this action. After each
proposal, it receives a positive feedback from its
associated controller if the action is selected or a
negative one if the action is not selected. The role of
the Context Agent is then to self-adapt to feedbacks
from its associated Controller Agent by modifying
its situation description or by adjusting the
information of its action proposition.
- A Controller Agent is associated with each
controllable variable of the environment and
controls the modification of this variable to produce
an adequate behaviour. In order to do this, it receives
action proposals from Context Agents. Then, it
selects the best action proposal, performs this action
and observes the impact of this action (this
observation is domain dependant) to send feedbacks
to Context Agents. It is also responsible of Context
Agents creation when there is no relevant Context
Agent. (See section 4.4).
Learning is then the result of a self-organisation
process inside Context Agents as each Context
Agent dynamically adjusts its validity domain in
reaction to feedbacks from the Controller Agent.
4.4 Context-learning Formalism
A Context Agent receives signals from its
environment and uses these signals to describe the
current context.
Definition 1: Let ∈ where is a set of
signals and ⊂
,
a signal such as
,
∈ .
A low-level context description of a Context Agent
is made with validity ranges.
Definition 2: A validity range
associated to a
value ∈ is a range
,
where
,
⊂
,
.
A validity range allows the Context Agent to
compare the current signal value to its associated
validity range and to decide if it is relevant to send
an action proposal.
Definition 3: A validity range
is if and
only if ∈
,
.
For each received signal, a Context Agent creates an
associated validity range. The set of validity ranges
is then called a validity domain.
Definition 4: Let ∈ a Context Agent. Let
a validity domain associated to a Context
Agent.
is a set of validity range such
as∀ ∈ , ∃
.
To determine its validity state, a Context Agent
verifies each validity range of its validity domain.
Definition 5: ∈ is if and only if
∀
,
is , otherwise.
The role of the Context Agent is then to dynamically
adjust its validity domain to feedbacks from its
Controller Agent in order to be valid when its
associated action is relevant. The main advantage of
this context description is that it uses no semantic:
only variations of signal value are observed and no a
priori is made on signal value meaning.
4.5 Context-learning Properties
The Context-Learning pattern presents interesting
properties:
- On-line learning: the process of self-
organisation (especially the creation and tuning
of the Context Agents) is performed at runtime
on the fly without necessity to stop it or reboot it;
- Openness: new variables can easily be added in
the system thanks to the low-level description of
situations in the Context Agents. Moreover, new
functionalities can also be easily added because
of the independence between the learning
processes of each controllable variable;
- Generic: the Context-Learning pattern uses no
semantic on the signals perceived neither on the
controlled system, making it highly generic.
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
246
4.6 Context-learning Modes
The Controller Agent in charge of the Context
Agents creation can apply either a supervised
strategy, either a reinforcement strategy to explore
the possible states.
- Supervised strategy: the Controller Agent bases
its Context Agents creation on the observation of
the actions performed by an entity (for example,
a user). When this entity performs an action
whereas there was no Context Agent that
proposed this action, the Controller Agent
observes it and creates a Context Agent in order
to represent this action associated with the
current situation. With this strategy, the system
only performs actions previously observed;
- Reinforcement strategy: the Controller Agent
generates Context Agents by itself for each
situation where the current set of valid Context
Agents is empty or composed of unsatisfying
Context Agents. In this case, the Controller
Agent applies different strategies in order to
evaluate what seems to be the correct action to
perform (for example, the same action as that of
the previously selected Context Agent if it was a
satisfying action, or else the opposite action).
In both case, self-observation is the engine of
learning. These two strategies differ in the way they
explore the possibilities space. In supervised
learning, exploration is guided by an external entity
whereas in reinforcement learning, exploration is
guided by a trial/error process.
For more information on Context-Learning
pattern, the reader can refer to previous works of
(Guivarch, 2012) and (Boes, 2013).
5 ALEX: AN ADAPTIVE MULTI-
AGENT SYSTEM FOR
CONTEXTUAL-LEARNING
FROM DEMONSTRATION
ALEX (Adaptive Learner by EXperiments) is an
AMAS based on the Context-Learning pattern in
respect with the Extreme Sensitive Robotic vision. It
allows a real-time learning from the observation of
user’s activity in distributed applications. ALEX is
able to control a device or a functionality by
correlating actions performed by a tutor to the
observations it makes of its own environment.
The main hypothesis made by ALEX is that
when a user has to act on it, it is because the ongoing
behaviour is not satisfying the user anymore. Thus,
the system has to self-organise to reach a
functionally adequate behaviour. The actions
performed by the user will be relevant under the
same context. Then, ALEX tries to learn all
contextual actions performed by a tutor and to
reproduce them.
An ALEX is an Extreme Sensitive Function. Its
functionality could be the control of a high-level
function (such as "Go back", "Turn left") or a low-
level control on an effector (such as the rotating
speed of a motor or its angular position). It receives
signals from its surrounding environment that could
come from sensors, other Extreme Sensitive
Functions, or even humans. A signal is composed of
a unique identifier and a value. The identifier has no
specific significance and by consequence, it has no
semantic. These signals are used to determine
contexts.
When a user acts on it, the ALEX system
analyses all signals to discover the current context.
Then, it determines what action should have been
performed if the user did not act and adapts the
behaviour in response. Every time an ALEX
performs an action, it communicates its new state as
a new signal. Thus, each Extreme Sensitive Function
can sense the activity of other within a
communication range. An ALEX is then an
autonomous controller trying to correlate the user
activity to the observation of its own environment,
including other ALEX.
ALEX is currently used and tested on multi-
robot and multi-user ambient applications. Some
examples of experiments can be viewed on one of
the author’s website (www.irit.fr/
~Nicolas.Verstaevel/ALEX).
6 CONCLUSIONS
This paper presents Extreme Sensitive Robotic, an
approach where the design is not guided by the goal
to achieve but by the functionalities composing the
system. The overall activity results from local
interactions between each functionality.
It also presents ongoing work on Learning by
Demonstration. More precisely, it illustrates how we
are combining the Extreme Sensitive Robotic
approach and the AMAS approach. We propose the
use of the Context-Learning pattern to enable self-
observation capacities in each system’s
functionality. It results in ALEX, a multi-agent
system for contextual learning by demonstration that
brings Context-Awareness in ambient systems.
Functional prototypes have been developed and we
ExtremeSensitiveRobotic-AContext-AwareUbiquitousLearning
247
now consider the application of our approach on
concrete problems coming from industry.
REFERENCES
Argall B. D., Chernova S., Veloso M., & Browning B.
(2009). A survey of robot learning from
demonstration. Robotics and autonomous systems,
57(5), 469-483.
Braitenberg, V. (1986). Vehicles: Experiments in synthetic
psychology. MIT press.
Brambilla M., Ferrante E., Birattari M., & Dorigo M.
(2013). Swarm robotics: a review from the swarm
engineering perspective. Swarm Intelligence, 7(1), 1-
41.
Boes, J., Gatto, F., Glize, P., & Migeon, F. (2013).
Controlling Complex Systems Dynamics without Prior
Model. In ICAART (1) (pp. 389-392).
Brooks, R. A. (1987). A hardware retargetable distributed
layered architecture for mobile robot control. In
Robotics and Automation. Proceedings. 1987 IEEE
International Conference on (Vol. 4, pp. 106-110).
Brooks, R. A. (1990). Elephants don't play chess.
Robotics and autonomous systems, 6(1), 3-15.
Broxvall, M., Gritti, M., Saffiotti, A., Seo, B. S., & Cho,
Y. J. (2006). PEIS ecology: Integrating robots into
smart environments. In Robotics and Automation,
2006. ICRA 2006. Proceedings 2006 IEEE
International Conference on (pp. 212-218).
Calinon, S. (2008). Robot programming by demonstration.
In Springer handbook of robotics (pp. 1371-1394).
Springer Berlin Heidelberg.
Camps V, (1998). Application of a self-organizing method
based on cooperation to information retrieval. In
Young researcher paper to European Conference on
Artificial Intelligence (ECAI-98), Brighton, England.
Capera, D., Georgé, J., Gleizes, M. P., & Glize, P. (2003).
The AMAS theory for complex problem solving based
on self-organizing cooperative agents. In Enabling
Technologies: Infrastructure for Collaborative
Enterprises, 2003. WET ICE 2003. Proceedings.
Twelfth IEEE International Workshops on (pp. 383-
388).
Couceiro, M. S., Portugal, D., & Rocha, R. P. (2013). A
collective robotic architecture in search and rescue
scenarios. In Proceedings of the 28th Annual ACM
Symposium on Applied Computing (pp. 64-69).
Dautenhahn, K., Nehaniv, C. L., & Alissandrakis, (2003).
Learning by Experience from Others—Social Learning
and Imitation in Animals and Robots. In Adaptivity
and Learning (pp. 217-241). Springer Berlin
Heidelberg.
Guivarch, V., Camps, V., & Péninou, A. (2012). Context
awareness in ambient systems by an adaptive multi-
agent approach. In Ambient Intelligence (pp. 129-
144). Springer Berlin Heidelberg.
Kaminka, G. A. (2012). Autonomous Agents Research in.
Robotics: A Report from the Trenches. In AAAI Spring
Symposium: Designing Intelligent Robots.
Knox W. B., Peter Stone, and Cynthia Breazeal, (2013).
Training a Robot via Human Feedback: A Case Study.
In International Conference on Social Robotics.
Lacouture, J., Gascueña, J. M., Gleizes, M. P., Glize, P.,
Garijo, F. J., & Fernández-Caballero, A. (2012).
Rosace: Agent-based systems for dynamic task
allocation in crisis management. In Advances on
Practical Applications of Agents and Multi-Agent
Systems (pp. 255-259). Springer Berlin Heidelberg.
Navarro, I., & Matía, F. (2012). An introduction to swarm
robotics. ISRN Robotics, 2013.
Nehaniv, C. L., & Dautenhahn, K. (2001). Like me?-
measures of correspondence and imitation.
Cybernetics & Systems, 32(1-2), 11-51.
Perera, C., Zaslavsky, A., Christen, P., & Georgakopoulos,
D. (2014). Context aware computing for the internet of
things: A survey. Communications Surveys &
Tutorials, IEEE, 16(1), 414-454.
Picard G and Glize P, (2005). Model and Experiments of.
Local Decision Based on Cooperative Self-
Organization. In second International Indian
Conference on Artificial Intelligence (IICAI’05).
Pfeifer, R. (2002). On the role of embodiment in the
emergence of cognition: Grey Walter’s turtles and
beyond. In Proc. of the Workshop “The Legacy of
Grey Walter”.
Sammut, C., Hurst, S., Kedzier, D. et al (2002). Learning
to Fly. Imitation in animals and artifacts, p. 171.
Siciliano Bruno & al., (2014). EuRoC - The Challenge
Initiative for European Robotics ISR/Robotik 2014.
Proceedings of 41st International Symposium on
Robotics, vol., no., pp.1,7, 2-3.
Tosello, E., Michieletto, S., Bisson, A., Pagello, E., &
Menegatti, E. (2014). A learning from demonstration
framework for manipulation tasks. In ISR/Robotik
2014; 41st International Symposium on Robotics;
Proceedings of (pp. 1-7).
Videau, S., Bernon, C., Glize, P., & Uribelarrea, J. L.
(2011). Controlling bioprocesses using cooperative
self-organizing agents. In Advances on Practical
Applications of Agents and Multi-agent Systems (pp.
141-150). Springer Berlin Heidelberg.
Walter, W. G. (1950). An imitation of Life.
Walter, W. G. (1951). A machine that learns. Scientific
American, 185(2), 60-63.
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
248
