|
|
|
|
|||
|
Home | People | Undergraduate Projects | Selected Publications by Topic| Selected Publication by Category | New Events ·
Call for papers - Special
session on “Bio-Inspired Self-Organizing Multi-Agent Systems”
at the 2010 IEEE Congress on Evolutionary
Computation ·
Call for papers - Workshop on
“Bio-Inspired Self-Organizing Robotic Systems” at the 2010 IEEE International Conference on
Robotics and Automation Research
Projects My research
projects are summarized here. For
more details, please refer to my online selected
publications by topic or selected
publications by category (journals, conferences, etc.) Morphogenetic Robotics Morphogenetic swarm robotics. For
self-organizing robotic systems, we want to construct the systems to be
robust, flexible, independent, autonomous, which can adapt dynamically to the
current conditions of their environments. In other words, they will exhibit some
life-like intelligence, which are characterized by self-organization,
self-configuration, self-adaptive, and self-repair. Distributed approaches
seem desirable. However, it is
well known that, for distributed approaches, it is difficult to predict the
emerging behaviors from local interactions of individual agents, neither is
it trivial to design rules for local interactions to generate a desired
global behavior. As we know,
biological systems, from macroscopic swarm systems of social insects to
microscopic cellular systems, can generate robust and complex emerging
behaviors through relatively simple local interactions subject to various
kinds of uncertainties. Embryonic
development of multi-cellular organisms is governed by gene regulatory
networks (GRNs), which are a collection of genes that interact with each
other and with other chemicals in the cell. Inspired by the major principles
of gene regulation and cellular interactions in multi-cellular development,
in this project, we aim to replicate biological robustness by organizing
complex behaviors from locally interacting individuals using morphogenetic
approaches, and eventually provide a self-organizing framework for
morphogenetic swarm robotic systems.
Some video
demos for morphogenetic swarm robotics:
Morphogenetic Modular Robotics.
Self-reconfigurable modular robots are autonomous robots with a variable
morphology, where they are able to deliberately change their own shape by
reorganizing the connectivity of their modules to adapt to new environments,
perform new tasks, or recover from damage. Each module is an independent unit
that is able to connect it to or disconnect it from other units to form
various structure/patterns dynamically. Compared with conventional robotic
systems, self-reconfigurable robots are potentially more robust and more
adaptive. In this project, we
mainly focus on developing a morphogenetic modular robot, CROSSCUBE, which is
inspired by the embryonic development of multi-cellular organisms. In
principle, CROSSCUBE is a two-layer modular robotic system. The first layer is a lattice-based locomotion
design, which defines the basic mechanics of the modules in CROSSCUBE and the
major constraints of the modules.
The second layer consists of a high-level controller using a
morphogenetic approach for CROSSCUBE to adaptively generate and form patterns. Some video demos for
morphogenetic modular robotics: ·
From a cube to a mobile
vehicle. Movie
(4.8MB) ·
Automatic change morphology
to traverse a narrow tunnel. Movie (5.4MB) ·
Self-reconfigure to adapt to
different environmental situations. Movie
(10MB) Social Cognitive Robotics Evolving Social Cognitive Agents System. In this project, we will develop a CIVS
(Civilization-Inspired Vying Societies) system with evolving social cognitive
agents, which is a novel evolutionary learning system loosely inspired by the
history of human civilization.
The system uses a bottom-up Alife (Artificial Life) approach with the
goal of generating intelligent emergent behavior in agents and agent groups.
CIVS proceeds from the idea that an agent will not develop complex,
intelligent behavior in isolation; it must be part of a culture and social
behaviors that support such development.
Thus, we propose a simulated world where agents exist as part of
societies, so that there are complex competition/cooperation relationships
acting at several different scales. Swarm
Intelligence based Multi-Robot Coordination. One of the main
challenges for multi-robot systems is to create intelligent agents that can
adapt their behaviors based on interaction with the environment and other
agents, become more proficient in their tasks over time, and adapt to new
situations as they occur. Such ability is crucial for developing robots
in human environments. Swarm robots are often observed to display many
of the attributes which are typical in collective intelligent system in
general including robustness, distributedness, adaptability, flexibility, and
self-organization. Typical problem domain for the study of swarm-based
robotic systems including foraging, box-pushing, aggregation and segregation,
formation forming, cooperative mapping, soccer tournaments, site preparation,
sorting, and collective construction. The main challenge of these control
architecture are scalability due to limited sensing and communication
capability of swarm robots. In this project, we are focusing on swarm
intelligence based control approaches, which is an innovative computational
and behavioral metaphor for solving distributed problems by taking its
inspiration from the behaviors of social insects swarming, flocking, herding,
and shoaling phenomena in vertebrates, where social insect colonies are able
to build sophisticated structures and regulate the activities of millions of
individuals by endowing each individual with simple rules based on local
perception.
Distributed
Reinforcement Learning for Multi-Agent Systems. Most available
distributed multi-robot paradigms require expert behavior development with
extensive tuning for each specific application, and lack of capability to
generate new behaviors to adapt to dynamic environments. Therefore, it is desirable that robots
can learn by themselves to accomplish the assigned tasks, instead of
following some predefined rules or behaviors which may not be suitable for
unknown environments. More and more researchers realized that automatic
learning approaches have more potential for cooperative multi-robot system
under unknown dynamic environments. However, multi-robot learning is a very
challenging problem due to the dynamic interactions among the robots. More specifically, the very
large state spaces may make it infeasible for a system to define the state. Due to the dynamic interaction among
robots and with environment, it may bring some uncertainty in the reward
assignment after each action of an agent. Usually in the real-world scenarios,
sensor noise and uncertainty are unpreventable, which leads to more challenge
in learning performance since these sensing information are the only feedback
that an learning agent can rely on. Furthermore, it is very
hard to merge information from different robot experiences due to its
distributed characteristics.
The goal of this project is to develop a decentralized cooperative
multi-Q reinforcement learning algorithm, where each robot learns to make
optimal decisions cooperatively with other robots based on its own Q value
and the correlated Q values of other robots in a team.
Object
Detection and Tracking.
Visual
interpretation of people and their behaviors is an important issue in many
applications, such as service robots, surveillance systems, public security
systems, virtual reality interfaces, and other government, commercial, and
entertainment applications. The ability to recognize and track people
is therefore an important visual problem. When the vision system is installed
on a mobile platform, such as a mobile robot, the problem becomes more
challenging since the background environments are dynamic and the people are
not always visually isolated but are partially occluded by other
objects. To tackle this issue, in this project, we are focusing on
developing some robust adaptive-window-tracking algorithms with real-time
performance. More specifically, we propose a new object tracking
algorithm that embeds swarming particles into generic particle filter
framework (PSO-PF) to achieve more robustness and flexibility. Firstly a
group of particles associated with potential solutions are initialized in a
high-dimensional space. Then particle swarm optimization (PSO) is used to
drive particles flying. The object is tracked when the particles reach
convergence. This PSO-based algorithm contains resample, similarity measure,
and integration together such that the degeneracy problem of particle filter
can be avoided. Furthermore, a multiple feature model is proposed for object
description to enhance the tracking accuracy and efficiency. The proposed
algorithm is independent with specific objects and can be used for any free-selected
object tracking. Some experimental results demonstrate efficiency and
robustness of the algorithm. Now we are focusing on developing a bio-inspired hierarchical neural
network (HNN) model for complex pattern recognition and learning. Behavior
Learning and Recognition. Recognizing
human behaviors in a video stream is critical in many applications, such as
video surveillance, video indexing, video annotation, and video
summarization. One major objective of automatic visual surveillance systems
is to recognize normal patterns and detect abnormal behaviors. Automatic behavior detection is a
challenging problem due to its large amount of video data to be analyzed and
its real-time requirements. Furthermore, it is difficult to represent the behavior
patterns based on video signal data, which is not always a one-to-one
mapping. The goal of automatic
behavior recognition is to learn a model that is capable of detecting unseen
abnormal behavior patterns while recognizing expected normal behavior
patterns. The
primary source of difficulty in behavior recognition is that behaviors may
vary dynamically in both shape and duration. Although it is unrealistic to
obtain a large training data set for abnormal behaviors, it is possible to do
so for normal behaviors, allowing the creation of well-estimated models for
normal behaviors. A better approach for abnormal behaviors detection is to
learn the model of normal behaviors, and then detect an abnormality based on
its dissimilarity from normal behaviors.
Therefore, in this project, we propose a self-organizing hidden Markov
models (SO-HMMs) based framework to tackle this issue, which is an online
learning method that can dynamically generate models for normal behavior
recognition and abnormal behavior detection in an unsupervised manner. This
SO-HMMs method is able to learn from the current data and generate new
corresponding HMMs if a new data sequence cannot fit into any of the existing
trained HMMs. This self-growing
and self-organizing capability can provide more robustness and flexibility
for dynamic scenarios where no prior knowledge about abnormal behaviors is
given. Real-Time Embedded Systems SMARbot. The goal of this project is to
construct a simple, easy to use, embedded system framework that can generally
meet or exceed the specifications of various robots, while maintaining a
great deal of flexibility. For the purposes of architecture design, details
concerning sensors and the chassis are intentionally ambiguous. Instead of
viewing the problem from the perspective of what the end system should do, we
take an approach that is similar to the design of a development board.
Ideally, a maximal amount of interfaces and commonly used components should
be provided, with a minimal number of constraints that would prevent the
addition of other components. In this project, we propose a generic modular
embedded system architecture called SMARbot (Stevens Modular Autonomous
Robot), which consists of a set of hardware and software modules that can be
configured to construct various types of robot systems. These modules
include a high performance microprocessor, a reconfigurable hardware
component, wireless communication, and diverse sensor and actuator
interfaces. The proposed
modular architecture of SMARbot consists of the following design features: · A modular hardware
and software architecture, where individual modules should be easily
modifiable and reusable. · A high performance
microcontroller in a common and well supported architecture. Open
source compilers and debugging tools are available. · Reconfigurable
hardware, well supported with software tools and documentation. · Power-efficient
wireless communication, where inexpensive and off the shelf module are ideal. · A large number of
interfaces for sensors. Sensors such as infrared proximity sensors, cameras,
sonars, and bumper switches need to be connected while leaving room for
future expansion. · Low power. Mobile
robots generally run off of small batteries. The system should
not consume more than 1000mW. · Low cost. Using
high end microprocessors or FPGAs is out of the question. PCB design should
fit the design rules of low cost, fast turn prototypes. Total cost
should be under $600USD each in single quantity. · Small Size. The
complete system should be approximately 3” cubed. · Low risk
design. Proven designs and off the shelf hardware will be used whenever
available. · Ease of
fabrication. A reasonably skilled user should be able to assemble the
necessary components into a working system.
The prototype of the SMARbot. Reconfigurable
System-on-Chip Architecture Design for Real-Time Systems. Conventionally, all
the tasks of a complex real-time system are implemented on a general
processor in software only due to its great flexibility. A powerful
processor is needed to guarantee that hard deadlines always are met, even for
very rare conjunctions of events. Sometimes the sensors of a real time
system are supposed to work in parallel and independently, it would be
difficult to implement the parallel sensing in software. Furthermore, with
multi-tasks, signals, events, it would be a challenging job for a software engineer
to implement all of the tasks under real-time constraints especially with
some computation-intensive information, say video images. On the other hand,
the specific hardware can respond to external inputs more efficiently since
the multiple hardware units can all operate in parallel, concurrent, and
asynchronous, so that individual response times are much less variable and
easier to guarantee, even as the number of tasks increases. Parallel
hardware is not so affected by issues such as task swapping, scheduling,
interrupt service, and critical sections, which complicate real-time software
solutions. The expensive ASIC can fulfill the speed criteria. However,
it is a complicated and expensive procedure if a slight change occurs.
The newest FPGA technology allows a designer to use a single reconfigurable
platform to instantiate both the processors and the required logic units,
which directly leads to feasibility of the reconfigurable
system-on-chip (rSoC) architecture. In this project, a unified agent-based
hardware/software representation is proposed to simplify the
hardware/software codesign procedure. Currently, we are developing
dynamic task allocation and scheduling algorithms for this rSoC
platform. This rSoC platform is desirable for those real-time systems
which must be highly responsive to the dynamic environments. Bioinformatics Bioinformatics is
the science of managing, mining, and interpreting information from biological
sequences and structures. Previous applications of data mining and machine
learning to bioinformatics include gene finding, protein function domain
detection, function motif detection, disease diagnosis, disease prognosis,
disease treatment optimization, protein and gene interaction network
reconstruction, and data cleansing. Some clustering algorithms have
been proposed for these applications, such as K-means algorithms, Genetic
Algorithm (GA), Self-Organizing Maps (SOM), Support Vector Machine (SVM), artificial
Neural networks, and decision trees. Even though most available
classification methods are powerful in the information they produce, they may
not be efficient or applicable to predict the outcomes of other clinic
occasions. Furthermore, the choice of classification algorithms can
give different classifiers, which may or may not be biologically
significant. Therefore, in this project, we plan to develop a swarm
intelligence based neural network technique for a generic analytical and
prediction system, which can analyze and interpret the dataset, predict
the results in a consistent way with known biological information about the
clinic case at hand. Eventually, this system will automatically provide
an optimal treatment strategy for specific clinic occasion. These research projects are partially
sponsored by NSF, DOD, Honda Research Institute Europe, Armament Research,
Development and Engineering Center (ARDEC), and Stevens Faculty Startup
Funding.
|
|||||
|
Please
send comments to: yan.meng@stevens.edu |
