It's like trying to keep tabs on a lively group of 2-year-olds.
Lynne Parker, a computer scientist at Oak Ridge National Laboratory, chases her four charges around the room to prevent them from escaping. She has told them to play without setting up the playpen, and she runs to pull barriers in place before they get stuck behind desks or make a run for the door.
Ada, Alexandra, Edith and Grace chatter together in a language of beeps and monotone syllables. Their instructions from Parker are simple -- to keep from colliding with each other or anything else.
But those instructions come in the form of complex algorithms issued by a computer across the room. Ada, Alexandra, Edith and Grace are robots.
And Parker has developed the computer software that allows the robots to work together, performing tasks like passing a baton or tracking moving targets or simply avoiding each other. The program, called Alliance, has helped break new ground in a field known as cooperative robotics, which focuses on using groups, rather than single robots, to solve a variety of problems.
|
ORNL computer scientist wins national honor
by Larisa Brass
Oak Ridger staff
Lynne Parker, a computer scientist at Oak Ridge National Laboratory, has been selected as one of three scientists among the Department of Energy's federal laboratories to win a Presidential Early Career Award.
Parker won the award for her creation of a software architecture known as Alliance. The program is used to coordinate teams of robots to complete tasks together and compensate for each other's weaknesses. Robot groups could be used for everything from environmental cleanup to clearing land on Mars for manned missions.
She will travel to the White House to receive the award on April 12.
Parker has worked at ORNL as a research staff member in the Computer Science and Mathematics Division's Center for Engineering Science Advanced Research since 1994. She became a senior research staff member in 1998.
She received her doctorate from Massachusetts Institute of Technology in 1994.
|
Groups of robots could potentially tackle nasty cleanup projects for the Department of Energy, scout out suspicious buildings for the military, do reconnaissance work for the army or explore Mars. And Parker helps teach robots not only how to take orders but how to learn for themselves.
Her pioneering efforts have now received accolades all the way from Washington, D.C.
Parker will be one of three scientists from among DOE's national laboratories to be honored with a Presidential Early Career Award. She will receive the award during a ceremony at the White House on April 12. The president honors 60 government employees each year for achievement in their respective agencies.
Parker is one of the first researchers to develop software that coordinates robots with different capabilities to complete a job and teaches them to respond to the difficulties they encounter, according to Ronald Arkin from Georgia Institute of Technology's College of Computing. Arkin has collaborated with Parker, and he cites her work in a textbook he authored, "Behavior-Based Robotics."
"She's off to a great start in her career," says Arkin.
"It's just always something that's fascinated me," she says, "the whole idea of how you can make machines do what seem to us to be very smart things."
But it's a difficult problem, she says, because computers "think" in a completely different way than humans do.
"What we have discovered is that things that are easy for us as people to do are very hard for robots to do. And then things that are hard for people to do are easy for robots to do," says Parker.
|
Lynne Parker sits with her robot family in Oak Ridge National Laboratory's Center for Engineering Science Advanced Research. Parker recently won a Presidential Early Career Award for developing software that enables the robots to work cooperatively and learn on the job.
|
For instance, supercomputers have beat the reigning chess champion and can perform intricate calculations at mind-boggling speed, impossible feats for humans.
"But we can all walk down the hall, and we can all look out at traffic and avoid getting hit by a car as we cross the street," she says.
"And we can maybe state the rules for how we play a game of chess. But we can't really state the rules for how we recognize that a car's about to hit us or how we recognize that when we're holding an object and it's slipping out of our hand we need to grasp it better in order to keep it from falling. We don't really have a good handle on making explicit those kinds of rules."
Efforts to substitute robots for humans have led researchers to conclude that robot teams can more effectively attack problems than lone robots.
"We're looking at types of applications where the requirements are complex enough that you can't really imagine building one robot that can do everything," says Parker. "So we're looking at making a problem less complex by spreading the capabilities out among multiple robots."
This approach to NASA's recent Mars mission, for example, she says, might have prevented the Mars Polar Lander's failure to reach the planet's surface. Robot teams can operate even if some of the robots cop out along the way, she says. The remaining members pick up the slack to ensure the job gets done. That way, an entire mission doesn't rest on one highly trained robot's shoulders.
Parker is currently working on projects for DOE, NASA, the Defense Advanced Research Projects Agency and Caterpillar.
|
Lynne Parker
-- Photo Submitted
|
DOE wants to find better ways to attack its stores of hazardous and radioactive waste. NASA is investigating multi-robot missions to Mars to clear the way, and the land, for human visits. DARPA plans to use robot teams for urban operations, searching for evidence of terrorism in downtown buildings, for example. Caterpillar wants to use bulldozer robots to clear areas of land for surface mining in remote locations like northern Canada.
Parker is also beginning a project with the Army to develop robot teams to substitute for human scouting groups in missions such as the one to Kosovo last year.
Parker tests the Alliance system on four resident nomad robots -- named Ada, Alexandra, Edith and Grace for computer science's female pioneers -- in ORNL's Center for Engineering Science Advanced Research. ORNL has also gotten the first of a new shipment of robots more suited for outdoor endeavors. Picking up on the CESAR theme, they'll be named for the less notorious of Rome's emperors.
Rather than "hardwiring" the robots to complete certain tasks, Parker uses the software to gently prod them through their paces and teach them in the process.
The software helps the robots make decisions based on the information received through attached sensors and feedback received from other robots.
The software gives each robot the desire to complete the mission. "And incorporated within Alliance are motivations of behavior," Parker says. For example, if one robot learns that another robot is not completing its job, it will grows impatient to see the job finished.
Alliance then guides the malfunctioning robot to acquiesce, allowing the working robot to finish its job. The program gives robots different levels of impatience depending on their ability to do the work, she says.
"So that way you're able to have the robot that's best suited for a task to become impatient to perform that task before the ones that are not as well suited for it," Parker says. "But if the one that's best suited for it has a catastrophic failure and is not available, then the other robots that are left will still become impatient to perform that task."
In addition, Alliance teaches robots to adapt to their surroundings.
Take, for example, a group of robots assigned to mow a field, Parker says. If one robot has been assigned an area with patchy grass, it will learn to work differently than a robot mowing a lush section.
Robots can learn their work in two ways, says Parker. The "model-based" approach shows the robot what the solution to the problem should be. A "model-free" approach gives the robot no initial information but provides feedback throughout the process. "You say, 'You're doing great. You're hot. You're warm. You're cold,'" she says. "But we don't tell them in what way they could improve things. So they have to sort of figure out from experience ... ."
So far, the model-based approach achieves better results, Parker says. But researchers are hoping to develop programs that will set the robots free to find their own solutions.
"We would like for the robot team to figure out an approach that maybe we didn't think about," she says. "But the problem is that it's hard."
It will probably take at least a decade, she says, for robots to be able to learn complex tasks from the ground up. And computer scientists hope research on the human brain will shed light on ways to reproduce our ability to make quick observations and decisions in robots, she says.
It appears that the robots milling around the CESAR lab, like toddlers at playtime, have some growing up to do.
