New Robot, Old Tricks


January 8, 2012

four students stand behind robot in classroom

By Laura Donnelly-Smith

GW senior Sam Zapolsky is walking a robot. On a leash. Or at least that’s what he’s attempting to do. The robot isn’t being entirely cooperative. The problem, Mr. Zapolsky explained, is that the robot—a 400-pound, anthropomorphic fellow who rolls around fluidly—senses the leash as an obstacle. The robot is programmed for self-protection, and he won’t advance toward something he thinks he’ll hit.

“Even if I hold the robot’s hand to lead him, he sees me as an obstacle,” Mr. Zapolsky said. So that means it’s time to go back to his computer simulation of the robot program, reinvestigate the code he’s written, and try to make some tweaks that will allow the robot to ignore the leash or the hand that’s leading him.

Mr. Zapolsky wasn’t frustrated that the robot hadn’t performed exactly as expected. It’s all part of the learning process, he said. “The robot is so capable, you just need to know how to ask it the right questions.”

Mr. Zapolsky, an economics major and computer science minor, was one of several undergraduate students in Autonomous Robotics, a dual-listed graduate/undergraduate course taught last semester by Evan Drumwright, assistant professor of computer science in the School of Engineering and Applied Science.

One of Dr. Drumwright’s main research interests is developing robots that can perform occupational tasks—things humans do during manual labor—autonomously and effectively. But before we can develop better robots, we need better computer robotic simulations, Dr. Drumwright said. So in his course, students spent much of their class time learning about programming languages and the mathematics of robotics and control theory. Then they worked on their own to write simulation programs for specific robot-performed tasks.

“The students proposed tasks that they would tackle, first in simulation and then in the real world with the robot,” Dr. Drumwright explained. Most of students chose pet-related tasks—things a robot would do in the course of taking care of a pet. Mr. Zapolsky proposed teaching the robot to walk a dog. Ph.D. student James Taylor worked on a “fetch” game, where the robot would throw a ball for a pet to retrieve. And Ph.D. student Roxana Leontie developed a program that would allow the robot to carry items without stable centers of gravity—say, bottles of water or half-filled bags of pet kibble.

The class material demands a lot from the students, Dr. Drumwright said. “It involves electrical and computer engineering, mechanical engineering and computer science. And the projects the students proposed were hard on several different levels. For example, what does it mean to be successful at walking a dog? It brings up a whole slew of research questions. But there’s no failure as long as the students learn.”

Mr. Zapolsky said the course was appealing because he had taken a lot of advanced mathematics courses and was interested in applying what he’d learned. One of his biggest challenges during the semester was learning several programming languages as a first step, before he could even begin working on simulations for the robot. And his dog-walking problem was one of the most complex of the semester.

“Walking a dog is a difficult concept, since dogs are variable creatures,” he said. “You’re not just dragging a dog—they’re in front of you but still following you. You realize how smart dogs are when you try to get a robot to do it. So it’s turned into a project about walking a robot.”

In developing his fetch scenario, Mr. Taylor first had to write a program in which the robot would track a ball with its eyes, which are fitted with a sensor. The ball—lime green and ping-pong-ball sized—could then be thrown or dropped. Once the robot was able to successfully track it, the next step was to write a program allowing it to toss another ball to land near the first one.

“It’s like the idea of bocce,” Mr. Taylor said, describing the lawn game where players try to toss a ball as close as possible to a target. “Rather than having an arbitrary target, we’re having a human throw a ball, and then the robot throws the ball as close as possible to that target.”

Mr. Taylor and the other students in the class all spent a lot of time working on computer simulations of their problems before moving to tests with the robot. Computer simulation is a vital first step to success in autonomous robotics, Dr. Drumwright said, because it allows the researcher to try various scenarios in a safe environment, without risking damage to the robot—or anyone else.

“In a virtual environment, you can send commands to the robot that might break it or cause it to hurt someone, and it’s not such a problem. Once students show proficiency in simulation, they come into the lab and work with the actual robot.”

Ms. Leontie’s project posed a special challenge because it is difficult to simulate objects that aren’t stable, such as bottles half-filled with liquids. In the simulation, she had to simplify her research question and make more assumptions. Then she had to spend a lot of time working with the actual robot, tweaking her program slightly with every try.

“For testing purposes, the robot holds a box with a ball in it that rolls around. It needs to correct the balance when the ball rolls,” she said. “I like working on problems that can apply in real life.”

And ultimately, real-life application is what autonomous robotics is all about, Dr. Drumwright said. While the students were focused on pet-related problems, similar applications could be developed to help care for the elderly, for example, he said. An autonomous robot might be able to help an elderly person perform certain routine tasks safely, so that the person could remain in his or her own home rather than moving into an assisted-living facility.

Even though the fall semester is over, many of the Autonomous Robotics students plan to continue working on their research problems. For Ph.D. student Mr. Taylor, the fetch project fits nicely into his larger body of research, which focuses on improving computer simulations.

“The real lesson is that our models aren’t good enough to take a simulation right into the real world,” he said. “Simulation doesn’t yet approximate the real world well enough, and my research is about how to make it better.”

For Mr. Zapolsky, there are most likely many more computer simulations—and robots—in his future. He recently applied to a number of Ph.D. programs in robotics, including GW’s.

“This class was important to me in terms of seeing where the research is heading,” he said. “It’s really given me a good view of what I might be doing.”