Transcript RoboticsLabProjects - Computer Science & Engineering

Enhanced Pattern Recognition
of target Chemicals and Bacteria
with Cantilever Sensor Arrays
Asya Nikitina
Robotics Research Laboratory
Computer Science Department
University of Nevada, Reno
http://www.cse.unr.edu/~nikitina
The Problem
 Goal:
 Develop and test a reliable system for detection and
differentiation of selected chemicals and bacteria in the
unknown background mixture of gases using a
microcantilever-based sensor array
 Motivation:
 Due to the high concern over chemical and biological
weapons and significant interest from governmental,
environmental and health-related agencies, there is a
high need for miniaturized, portable, accurate,
inexpensive, low power use chemical sensors capable
of rapid, in situ detection of chemical signatures of
bacteria and of chemical agents in the air
2/4
Approach
 The ability to detect extremely small displacements
makes the cantilever an ideal device for detection of
extremely small forces and stresses
 Adsorption of molecules on the surface of a cantilever or
absorption of molecules by a coating material changes
the total mass and, consequently, the resonance
frequency of the cantilever
 The resonance frequency shifts
and bending of a cantilever can be
measured with very high precision
using different readout techniques
and used as model’s features for
the future classification
3/4
Results
 During the current stage of our project, we are
performing experiments in which the cantilever array is
exposed to different concentration of some specified
analytes (stage I - data collection)
 Next stage will be training the cantilever array using
different mixtures of the same analytes (stage II –
training the cantilever sensor array)
 The final stage of our project will be developing a kernelbased pattern recognition algorithm, which will be able
not only to recognize the specified analytes in the
gaseous mixture, but also to determine the relative
concentrations of the analytes presented in the mixture
(stage III – developing and implementing a kernel-based
algorithm for pattern classification)
4/4
Intention prediction using Hidden Markov Model
Jigar Patel
Robotics Research Laboratory
Computer Science Department
University of Nevada, Reno
http://www.jspatel.com
The Problem
 Goal:
 The main goal of my research is to predict robot’s intention in
single or multi agent environment using trained Hidden Markov
Model.
 Make robot to predict other’s intention based on his own
experience with environment like humans do.
 Motivation:
 When one person is going towards door another person can
easily say he/she is going out of the door. This looks simple and
obvious to us. But this is not true for robots.
 Won’t be it interesting to see robot predicting other robot’s
intention/behavior.
6/4
Approach




As required in my goal I need a
robot having experience with
environment before it starts
predicting.
I am using Hidden Markov Model
(HMM) to represent robot’s
experience.
To train (or gain experience) I am
letting robot achieve certain goal
and train HMM for that behaviors
related to it. Using the same
technique train different HMM for
more behaviors.
Now try to predict intention in
single or multi agent environment
based on training (or gained
experience).
C4 C5 C6
C13
C1 C2 C3
C4 C5 C6
C13
C7
APPROACH
ONSET
INIT
Stage Distance
FINAL
Angle
//C1
++
++
//C2
++
--
//C3
++
Static
//C4
--
++
//C5
--
--
//C6
–
Static
//C13
END
Unknown UnKnown
AWAY
Example HMM for
object Seeking
Behavior
C4 C5 C6
C13
7/4
Results


Active / Deactive
HMM1
Action Performing
Robot set
Active / Deactive
Environ
ment
State
HMM2
Action Performing
Robot set
Active / Deactive
HMM3
Action Performing
Robot set
Possible Active
behavior in
multi agent
environment
1

5
So far I have trained HMM for
interfere, follow, meet, seek
behaviors.
In multi agent environment run
more than one behaviors. Feed
current status of environment to
all HMM in parallel and get
results.
I am also trying to derive
behaviors that are not explicitly
defined like Group Meet. This
behavior may occur when more
than one pair of robots try to meet.
I am getting overall 96% true
prediction for simulated
environment using Player/Stage
as tool.
4
3

2
8/4