The research activity of Francesco Amigoni
sits at the intersection of the fields of autonomous robotics and artificial
intelligence and, specifically, deals with multiagent systems. It addresses the
study, development, and experimental assessment of models and algorithms for
decision-making of systems composed of autonomous agents operating in the real
world.
The research activities reported below have been conducted
in cooperation with students
and researchers, prominently
including: Davide Azzalini, Jacopo Banfi, Nicola
Basilico, Vincenzo Caglioti, Andrea Castelletti,
Giulio Fontana, Nicola Gatti, Maria Gini, Matteo Giuliani, Michèle Lavagna,
Matteo Luperto, Alberto Quattrini Li, Ioannis Rekleitis,
Alessandro Riva, Viola Schiaffonati, and Marco Somalvico.
The current research topics of Francesco Amigoni are articulated in the following areas ([] refers
to the main publications in the Publication List of the Curriculum Vitae).
Navigation
Strategies
The research in this topic aims
at defining efficient navigation strategies for autonomous mobile robots.
Navigation strategies are methods used by mobile robots, either operating
individually or organized in multirobot systems, to autonomously decide where
to go when performing a task [B11] (in contrast to classical path planning
methods that look for how to reach given locations, see below).
Francesco Amigoni has significantly contributed to
the theoretical and practical development of navigation strategies, especially
for the exploration and patrolling tasks.
When exploring initially unknown
environments in order to discover their geometrical
features, exploration strategies determine the next observation locations
mobile robots should reach in a partially known environment. Francesco Amigoni has initially proposed an information-based
criterion for defining an exploration strategy for a mobile robot equipped with
a laser range finder [A19]. Moreover, Francesco Amigoni
has proposed decision-theoretical exploration strategies that exploit
multi-objective optimization [C26] and multi-criteria decision making [A22]
[C49]. An exploration strategy that, differently of the mainstream approaches,
exploits not only metric information but also semantic information to determine
the next observation locations has been proposed [A27], while an exploration
strategy that exploits prior knowledge that may be available on the
environments is described in [A41]. An exploration strategy that chooses the
locations to reach also according to the predicted observations obtainable from
there is proposed in [C103].
Exploration strategies have an
intrinsic online nature and their performance can be
compared with that of optimal offline exploration paths. An approach to find
the optimal exploration paths for covering arbitrary environments is presented
in [C60], while suboptimal covering paths can be found with the method of [C79]
for generic environments and with that of [C100] for modular environments. Some
bounds on performance of exploration strategies are reported in [C70].
Experimental comparisons of different exploration strategies in different
settings are reported in [C39] [C51] [C56] [C63]. The relation between
exploration strategies (where to go) and coordination methods (who goes where)
in multirobot exploration has been analyzed in [C58].
Multirobot exploration becomes
more challenging when robots cannot always communicate with each other [A30].
Different strategies for communication-constrained exploration are compared in
[C75], while in [A34] a multirobot exploration strategy is proposed that deals
with recurrent connectivity constraints. The methods proposed in [A40] build
communication maps that can be used to deploy robots in
communication-constrained environments, even during the exploration of the same
environments [C92]. [A36] presents an approach to efficiently reconnect robots
if they move out of each other’s communication ranges. The more abstract
problem of collecting joint measurements from locations between which
communication is possible in addressed in [C89]. Switching between different
communication modalities to efficient explore environments is investigated in
[C83].
Exploration of environments with
mobile robots to find gas sources is discussed in [C82].
In patrolling, mobile robots move
in an environment to prevent intrusions. An efficient strategy for a mobile
robot that is tasked to patrol an environment is presented in [C42]. The
strategy has been developed considering patrolling as a strategic game played
by the mobile robot and the intruder. Following the same approach, a more
mature model for finding optimal randomized patrolling strategies is reported
in [C44] [C45] and extended in [C46]. A deterministic patrolling strategy is
developed in [C47], while a deterministic capture strategy is described in
[C90]. A step toward the application of the approach to real mobile robots is
described in [C50], an application to active patrolling cameras is reported in
[C53], and an application to the problem of pursuit evasion is illustrated in
[C57]. Most of the above results are summarized and applied to realistically-sized environments in [A23].
Two strategies for multiple
robots patrolling environments with constraints on communication are reported
in [C72] and in [C78], while a multirobot system that tracks several targets is
described in [A37].
Multirobot
Path Planning
In the classical problem of path
planning, robots autonomously plan how to reach given goal locations. Francesco
Amigoni has contributed to show that this problem
turns out to be computationally difficult for multirobot systems [A29] [C86]
and for single robots [C96] [A39] in presence of constraints on communication
and to propose solving algorithms. There are many real-life applications where
several autonomous agents must coordinate to reach their goals without
collisions while optimizing some performance measure. Solutions to this
Multi-Agent Path Finding (MAPF) problem usually assume environments to be
static and known in advance. [C99] introduces C-MAPF, a novel variant of the
MAPF problem in which the environment is configurable and its structure and
topology can be modified within some given constraints. Other variants address
the case in which more teams of robots operate in the same environment [C106] and
exploit the knowledge about the appearance of future tasks [C107].
Semantic Maps and Map Prediction
Semantic maps associate
high-level information (e.g., labels like ‘kitchen’) to locations of
environments. Francesco Amigoni has addressed
different aspects involved in building semantic maps of indoor environments
considering a priori knowledge about the type of the building in which robots
are operating in [C62] [C68] [C74] [C80]. In [A38] a method able to predict the
structure of buildings of a given type is described, while the method in [A45] reconstructs
the geometry of parts of buildings that have not been observed and the method
in [A46] predicts the geometry of rooms behind closed doors. The identification
of the structure of buildings starting from observations made by robots is
addressed in [A44].
Anomaly Detection for Autonomous Robots
Detection of anomalies and faults
is a key element for long-term robot autonomy, because, together with
subsequent diagnosis and recovery, allows the robots to reach the required
levels of robustness and persistency. In [C98] and [A42], Francesco Amigoni has proposed approaches based on HMMs (Hidden
Markov Models) and on neural networks, respectively, for detecting anomalous
behaviors in autonomous robots starting from data collected during their
routine operations.
Experimental Methodologies and Standards
The research in this topic aims
at improving experimentation in autonomous mobile robotics, which has not yet
reached a level of maturity comparable to that reached in other scientific and
engineering disciplines [B10]. Francesco Amigoni has
been among the firsts to contribute to the definition of good experimental
methodologies for robotic mapping [C37] and to propose a general theoretical
framework in which these methodologies are inserted [A18] [B7] [B9] [B12], with
also attention to generalization of experimental results [A28] and ethical
issues [A32]. A comprehensive summary of the main results is reported in [B15].
In this context, competitions [B8] represent an interesting approach to the
experimental analysis of whole robotic systems [C65] [C66] [B16] [A26]. From an
operative point of view, [C91] presents a system that supports repeatability of
some experiments on robots.
Francesco Amigoni
has also substantially contributed (with leadership roles) the definition of an
IEEE standard for representing the maps employed by robots for navigation [C69]
[A33] [C101].
Multiagent Decision-making, Planning, and Scheduling
Francesco Amigoni
has studied multiagent decision-making, planning, scheduling, and anomaly
detection in several application scenarios: ambient intelligence and energy and
comfort management, space systems, water resources systems management, and
electrical systems.
In ambient intelligence, several
distributed devices, considered as agents, collectively operate to support the
activities of the user and the needs of energy and comfort management.
Distributed decision-making, planning, and scheduling are needed to implement
goal-oriented behaviors. In this research topic, Francesco Amigoni
has proposed a multiagent planner that can adapt its performance to the agents,
and thus to the devices, currently present in the environment [A8]. The
planner, called D-HTN, is based on the Hierarchical Task Network (HTN)
approach. Some technological issues of the implementation of the planner in
JADE are reported in [C41].
In space systems, autonomy in
managing onboard activities is one of the most important requirements. In this
scenario, Francesco Amigoni has developed multiagent
systems to manage the activities on Earth and onboard space systems. Each
device of the system is associated to an agent that is in
charge of planning, scheduling, and executing its specific activities.
The description of a system for onboard activities is reported in [A20]. A
study on the applicability of agent technologies to space systems is reported
in [C52], while a multiagent system to detect anomalies on space systems is
described in [C88].
The management of water resources
systems requires the accurate modeling of the entities involved and of their
interactions. Distributed approaches based on multiagent systems can be
fruitfully adopted to this end. A multiagent system that can help in deciding
and planning the use of water resources from the perspective of a regulation
body is presented in [A25], while a more evolved multi-objective approach is
reported in [C71]. A multiagent system that models preference evolution in the
management of water resources systems is presented in [A35].
A
system to detect anomalies in operations of electrical circuit breakers is
presented in [C95].
Two
systems for classifying road accidents from textual descriptions are presented
in [C102] and [C104].
The
research in this area, in cooperation with Viola Schiaffonati,
addresses the study of the relationships and of the reciprocal influences
between philosophy and artificial intelligence and robotics. In
particular, we have addressed some foundational issues about the nature
of interactions between computers and robots, on the one side, and the world
[A1] and humans [A5] [C12], on the other side, and about the ethical[C4] [C33]
and the design [B17] implications of these interactions. We have also addressed
topics related to creativity, using the multiagent paradigm as a metaphor to
represent the results of creative processes [A3] [B3] and to partially
represent the creative processes themselves, proposing an operational approach
to creativity [C16]. Finally, the two roles multiagent systems can play in
scientific discovery, as support to scientists and as representation of results
[A15] [B6], have been analyzed with respect to specific examples in [A4], [C2],
and [C8]. The main results of the investigation of the roles of multiagent
systems in scientific discovery are summarized in [A13].
The past research activity of Francesco Amigoni has been articulated in the following areas ([]
refers to the main publications in the Publication List of the Curriculum Vitae).
Development of Cooperative Multiagent Systems
Francesco Amigoni
proposed some formalisms for describing the properties of a generic system
composed of cooperative agents, both software and robotic [A2]. Such a system
has been called agency to emphasize its unitary nature. Moreover, Francesco Amigoni proposed the dynamic agency architecture for
developing cooperative multiagent systems. In this architecture, each agent is
composed of a pair of semiagents (modules): the
operative one, that offers specialized functions to operate in an environment,
and the cooperative one, that integrates the agents in a uniform cooperation
framework [C1]. The dynamic agency approach enables the development of a
cooperative multiagent system for a given task according to a sequence of
steps. Initially, the most suitable operative semiagents
for the task are selected (this recruitment process has been studied
theoretically in [A24] and more algorithmically in [C3] and [B1]). Then, the
cooperative semiagents are installed exploiting the
software technique of mobile code systems (a software framework for this
purpose has been proposed in [C7]). The flexibility offered by the dynamic
agency methodology in the management of agents allows their easy reuse and the
dynamic variation of the system composition. In this scenario, the use of
ontologies for describing the capabilities of agents is fundamental, especially
in the case of robotic agents [C32] and of Internet of Things technologies
[B14]. The main results obtained in this topic are collected in [A16].
Cooperative Negotiation
Francesco Amigoni
has proposed the use of cooperative negotiation techniques to model complex
phenomena. In particular, when a phenomenon (like some
physiological processes) is described by a set of models that only partially
capture the phenomenon, it is possible to associate a software agent to each
partial model and let a global more complete model emerge as the agreement of
the cooperative negotiation between these agents [A15] [B4]. In this research
topic, a cooperative negotiation model that allows the agents to reach an
agreement has been formulated and its stability studied [C21], defining the
conditions that guarantee stability regardless the number of agents
participating in the negotiation [A14]. This approach has been applied to the
development of systems for modeling and regulating the glucose-insulin
metabolism [A6] and the heart rate [A9] [A10], for modeling network normality
in anomaly-based intrusion detection [C43], and for modeling operations in
management of water resources systems [B13].
A bargaining protocol for
obtaining Pareto optimal agreements is described in [C31].
Virtual Museums
The
research activity on this topic, under the name Minerva
project, was oriented to enhance the use of advanced artificial intelligence
techniques to support some human activities related to the museum organization.
The project involved heterogeneous contributions from different sources, both
academic and not. The system that has been developed, called Minerva,
automatically organizes a virtual museum starting from the works of art and the
environments in which they should be displayed. Minerva is implemented as a
system composed of agents that communicate and cooperate. A first version of
Minerva, oriented to archeological museums, has been presented in [C6], while
an updated version, oriented to archeological and design museums, has been
presented in [C16]. In [A17], the main features of these versions of Minerva
are summarized. Moreover, a further version of Minerva for archeological
objects found on the Isola Comacina (Como, Italy) has
been presented in [C35].
Supporting Research in Bio-Medical Fields
Autonomous
agents can be employed to make the execution of distributed scientific
experiments that span different organizations more flexible [C38]. Moreover,
multiagent systems have been employed to provide simulations of biological
processes, as critically analyzed in [B6].
A
survey on distributed learning systems for medical applications is reported in
[A43].
Line
Segment Maps
The
most used techniques for building two-dimensional maps of environments (which
represent obstacles and free space) with mobile robots work incrementally by
integrating a sequence of partial maps acquired by the sensors of the robots.
Usually, this integration is based on the use of other localization sensors
that give information on robots’ positions in the already built map. The
research on this topic aims at representing obstacles with line segments and at
exploiting geometrical features of partial maps to improve the quality of their
integration. Francesco Amigoni has proposed important
contributions in this field, including a method for integrating two partial
maps [C24] and a method for integrating a sequence of partial maps [C19]
without using any information about robots’ position. The obtained results are
collected in [A11].
A
line segment-based method for matching partial maps (or scans) that exploits
geometrical features to improve the estimate provided by localization sensors
is presented in [C67].
Francesco
Amigoni has also proposed some methods for reducing
the size of line segment-based maps by merging redundant line segments [C34]
[C48] and has experimentally compared them with other methods for reducing line
segments in maps [C40] [A31].
Robotic Systems for Environmental Monitoring and
Search and Rescue
The research in this topic aimed at developing and
implementing a system, called perceptive agency, composed of mobile robots
equipped with different sensors for monitoring indoor environments. In this
sense, a perceptive agency is a particular sensor network in which the nodes
are implemented as perceptive mobile robots. Francesco Amigoni
defined the general aspects and requirements of perceptive agencies in [A7],
[C5], [C13], and [C14]. An implementation of a
perceptive agency oriented to monitor electromagnetic fields, developed within the research project described in
[A12], is presented in [C20] (with simulated sensors) and [C28] (with real
magnetic field sensors), while [B5] summarizes the main features of this
perceptive agency providing further experimental results.
The interface between human operators and autonomous
robots is fundamental to obtain effective systems in search and rescue
applications. A proposal relative to this topic is reported in [C61].
Flying Robot Formations
The research in this topic aimed
at studying some parameters of formations of flying robots [C18].
The
research in this topic aims at reassembling two-dimensional images starting
from their fragments and without any knowledge of the final images. Francesco Amigoni has proposed a method for reassembling fragments
only on the basis of their shape [C17].
The goal of the research in this
topic was to study a minimal model of differential equations for describing the
dynamics of production in creative professions. Francesco Amigoni,
in cooperation with Sergio Rinaldi, has proposed to consider, as a fundamental
variable of the model, the satisfaction of a creative person, represented as
the composition of internal selfmotivation and
external judgment given to production [B2].