Autonomous Agent Learning Lab
The central theme for research in this lab is autonomous agent learning.
The agents are robots, models of robots, and interactive video game players.
The learning is usually a form of evolutionary computation and almost always some type of computational intelligence.
The agents are autonomous in the respect that they can operate on their own and the learning either takes place before
operation or can happen during operation using learning systems that are offline from the agent.
Most of the learning is for control programs, although some is for the morphology of the robot or control/morphology
in combination. The following is a list of current or recent research topics.
Colony Robotics,
Gary Parker, Ozgur Izmirli, Art Potter, Richard Zbeda '06, William Tarimo '12, and Jim O’Connor '13.
An 8x8 foot area has been set aside in the
Robotics Lab for future colony robotics research. We are working
on developing a power supply for the robots, establishing communication
links from the learning system computer to the robot, and implementing
an overhead camera for colony observation.
Xpilot-AI,
Gary Parker, Sarah Penrose '12, Jim O’Connor '13, Evan Gray '13, and Jesse Newbold '13.
Xpilot is an online computer game that models space combatants in a 2D environment.
We have determined how to access the client program to be able to allow agents with artificial
intelligence to play the game
(Xpilot-AI).
With proper configuration, these agents look and act just like those controlled by human players. The long-term goal is to
have them continually learn as they join games in progress and compete against human players.
Using Cyclic Genetic Algorithms to Generate Gaits for
Hexapod Robots, Gary Parker and William Tarimo '12. In this
project we use CGAs (a form of evolutionary computation) to learn walking
patterns for six-legged robots. Learning takes place on a model of the
robot with the new control programs downloaded to the actual robot for
testing.
Learning Gaits with a Cyclic Genetic Algorithm and a Four-Legged
Robot, Gary Parker and William Tarimo '12.
This project involves the design of a four-legged robot with legs having
three degrees of freedom (the
hexapod robots we use have only two) and its controller. The CGA is being used
to learn the control programs needed to produce its gaits.
Punctuated Anytime Learning in Evolutionary Robotics,
Gary Parker, Basar Gulcu '08, William Tarimo '12, and Jim O’Connor '13. This research uses periodic tests
on the actual robot to test the control programs learned on the robot
model by evolutionary computation and improve the learning process by
altering the learning algorithm (Fitness Biasing) or changing the model
(Co-Evolving Model Parameters).
The Co-Evolution of a Team of Cooperative Autonomous Agents,
Gary Parker, Joseph Blumenthal '04, Max Rollins '12, Sarah Penrose '12, and Jim O'Connor '13.
In this research we are using evolutionary computation and
punctuated anytime learning to evolve the control programs for the
individuals in a team of robots or Xpilot-AI autonomous agents.
They have a common task
(such as pushing a box into the corner or catch a prey) and need to cooperate to
successfully accomplish it.
Genetic Algorithms,
Gary Parker.
This research involves studying the genetic algorithm methods of selection,
crossover, and mutation to improve results for categories of problem sets.
The Co-Evolution of Robot Control and Morphology,
Gary Parker and Pramod Nathan '06.
This research involves
concurrently evolving the body and the mind of a robot. We are using LEGO
Mindstorms for the evolution of full body robots and the ServoBot for the
evolution of sensor morphology.
Using Cyclic Genetic Algorithms to Evolve Multi-Loop
Control Programs, Gary Parker, Richard Zbeda '06, and Basar Gulcu '08.
In this research we are expanding the use of CGAs to evolve
multi-loop programs. Previous versions were limited to a single loop.
Mini-Robot Construction, Gary Parker.
This project involves the development of a six-legged robot that is 1/4 to 3/4
the size of the hexapod used in previous experimentation. This robot
will be of potential use in future colony robotics research.
Emergent Gaits Through the Co-Evolution of Leg Cycles,
Gary Parker. Incremental learning used a genetic
algorithm to coordinated previously learned leg cycles. In this research,
we attempt to
have gaits emerge during the co-evolution of the leg cycles.
Dynamic Neural Networks To Control Multi-Legged Robots,
Gary Parker. In this research we are designing a neural network
controller for
a simple multi-legged robot. A genetic algorithm is being used to evaluate
and improve the NN system in a computer simulation. This control system
is being tested on real robots (hexapod and/or octopod).
Neural Networks Implemented on PIC Chips,
Gary Parker.
We have are working on the implementation of a neural network on PIC chips.
Each chip will have a single
neuron and the program needed to learn using backpropagation.
Initially we will be using 3 chips to create a
neural network that can solve the AND, OR, and XOR.
In future work, we hope to use it for robot control.
Research Videos
Evolved ServoBot Gait
Evolved ServoBot Gait from Ground Level
Evolved Stiquito Gait (2x speed)
ServoBot with Capacitors Recharging
ServoBot Sensing Low Power and Navigating to Charger (2x)
Xpilot-AI Evolved Controller
Xpilot-AI Evolved Controller 2
ServoBot with Evolved Sensor Morphology/Control
ServoBot with Evolved Sensor Morphology/Control 2
General Areas of Research
Evolutionary Robotics
Adaptive Learning Systems for Autonomous Robot Control
Gait Generation for Multi-Legged Robots
Cyclic Genetic Algorithms
Punctuated Anytime Learning
Co-Evolving Cooperative Teams of Robots
RETURN
