Student Name: Igor Labutov Major: Mechanical Engineering Grade: Freshman Immigration Status: US Citizen Student Organization: Active member of CCNY Robotics Club Address: 6608 Grand Central Parkway, Apt. 1B, Forest Hills, NY 11375 Telephone: 718-275-7234

glibdoadingAI and Robotics

Oct 20, 2013 (3 years and 5 months ago)

76 views

Student Name: Igor Labutov

Major: Mechanical Engineering

Grade: Freshman

Immigration Status: US Citizen

Student Organization: Active member of CCNY Robotics Club


Address: 6608 Grand Central Parkway, Apt. 1B, Forest Hills, NY 11375

Telephone: 718
-
275
-
72
34

Cell: 646
-
239
-
9079

E
-
mail: azimuthny@yahoo.com


Project Title: Neural Networks in Robotics

Faculty Mentor: Dr. Jizhong Xiao




NEURAL NETWORKS IN ROBOTICS

Igor Labutov

The City College of the City University of New York

February 16, 2006



Backgro
und:
Modern embedded algorithms for autonomous control consist of high
speed processors utilizing RISC and CISC architectures. Their high speed serves as a
vehicle for intense calculations needed when performing complex tasks that are very
comfortably perf
ormed by humans. Tasks such as throwing and catching a ball,
synchronizing complex muscle movements and walking are all inherently learned human
behaviors. Traditional model of the brain however indicates that internal brain signaling
is a thousand fold sl
ower than even modest microcontrollers


the robotic “brains”.
Advanced progress in modern robots that imitate aforementioned human behavior
approach the problem in a fundamentally different way. The advantage in the speed of
the traditional processor is u
sed to perform intense mathematical calculations needed for
actions such as catching, throwing and walking. Highly advanced programs must be
written by a programmer and only tend to perform under carefully controlled conditions.
Theoretical research in Neu
ral Networks has been conducted for decades to simulate
brain processing. A Neural Network models the brain and needs minimum programming
to realize complex behavior. Robustness of this method is provided not through intense
computation, but through memory

and invariant associations during learning. A Neural
Network is imitated in a standard processor through a neuron based object oriented
programming. High clock speed of the microcontroller compensates for the myriad of
parallel connections through fast se
rial communication.


Abstract
Neural Networks have been simulated extensively as a theoretical tool to
study its properties. The purpose of this project however is to apply the fundamentally
radical processing system to a common engineering problem


a mot
or
-
visual system.
The design will consist of a mechanical arm capable of a wide range of movement
controlled solely be a Neural Network emulator developed to coordinate the arm to
catch a flying object. There will be no need to program the details specific

to this
task, as the Neural Network is fundamentally based on a principal of learning. The
associative memory is not centralized in any dedicated compartment (such as RAM in
a traditional computing system), but instead spread out through the network in a
form
of synaptic intra
-
neural connections. The advantage of the non
-
centralized system,
such as the brain is advantageous. Associative memory is stored in a hierarchical
fashion, where low level inputs are generalized to invariant representations of real
w
orld objects (such as the flying ball, or the thrower’s hand). The Neural Network
then modifies its own connections (its memory) to adapt to a general, rather than a
specific environment.