These guys used a DENNING chasis for a school project.
The Little Robot That Could
A novice ISI team entered its first state-of-the-art robot competition just to gain experience. Their robot was the biggest, the oldest and the slowest. But Yoda surprised everyone.
by Eric Mankin
The Information Sciences Institute team called their robot Yoda, but perhaps it should have been Rocky. Yoda came into the competition just hoping to go the distance - but came out with a prize.
The competition, held this summer in Portland, Oregon, by the American Association for Artificial Intelligence, attracted state-of-the-art mobile automatons from around the world, programmed by the best specialists in the business.
Against this competition, the ISI's Yoda ("Young Observant Discovery Agent") didn't seem to measure up. Yoda looks nothing like its Star Wars namesake. Its chassis, roughly the size and shape of a heavy-duty cannister vacuum cleaner, is more than a decade old. Yoda's programming was done by a group of amateurs that included not a single specialist in robotics, a pick-up gang that worked after hours and on weekends. While the School of Engineering, of which ISI is a part, has an active robotics effort, ISI works in other areas.
"We were just hoping to complete the course," said Wei-Min Shen, the team leader, whose normal work involves "intelligent agents" of a much less tangible kind - specialized computer programs designed to autonomously carry out virtual tasks in cyberspace. "If no one laughed, I was going to be happy."
But despite inexperience and make-do equipment, Yoda came through on game day, even though a last-minute reprogramming turned out to be necessary.
Yoda's set task, "Calling a Meeting," seemed simple. The robot was given a map of a section of a typical office complex, essentially a maze, with doors opening off corridors. Starting in a "director's office," the robot first had to find which of two "conference rooms" was empty.
When the robot found an empty conference room, it then had to visit the offices of two professors, announcing to each a meeting in the empty room, and a meeting time. Then it had to return to its start point, the director's office. Ideally, it should have calculated times and distances accurately enough so that it would be able to say, on entering this office, "the conference begins in one minute." The rules deducted points for the robot bumping into walls, doorways - or human beings, and for not taking the shortest paths through the maze. It gave bonus points for speed.
As simple as this test seems, it involves tasks that challenge the current state of the robotic art. Just entering directions to tell the robot how to get through the maze is difficult. Even more difficult is creating a foolproof - the computer word is robust - procedure so that the robot can check its progress through the confusing real world.
"The hardest thing," said Shen, "is the unreliability of the sensory data."
The input information about the world that the robot must use is always imprecise and often confusing. Yoda perceives the world through 24 sonor sensors, like the ones used on Polaroid cameras for focusing. Each sensor gives a single reading - the distance to the nearest obstacle. To get around, Yoda had to be able to use this simple set of readings to recognize landmarks - such as doorways, turns in the corridor and rooms.
Unfortunately, the sensors are easily thrown off by objects that aren't at the right level, or by environmental noise, and cross-echoes. But when this happens, Yoda pauses to think things over. Yoda is also programmed to combine readings from a number of sensors, never relying on a single instrument, much less a single reading.
Yoda's wisdom was not gained easily. As a report in the ISI InSiDer newsletter notes, "one of the first things we learned... is that a robot of this size and mass can potentially be dangerous.... Bumps into the wall, doors, trash cans and people have proved to us that robustness is hard problem to solve, especially if your sensors are lying to you.... The bottom line is the same: in the real world, things don't work as they should.... Murphy's law is not the exception, it it is THE rule."
And it's easy to tell what Yoda is thinking, because Yoda is programmed - charmingly - to continually tell the world. The whole team, along with Shen's young sons, Ted and Lucas, recorded messages which monitor the program (running in a Macintosh powerbook, that rides on the robot's flat top).
For example, the lilting female voices of team members Jihie Kim and Sheila Tejada announce the start of the robot's mission: "Director, I'm going to find a conference room and schedule a meeting."
The high pitched voice of Shen's young son announces, "I'm looking for the door."
"The next landmark after I start moving - open space to the left" says Yoda in the voice of Bohan Cho.
"I'm looking for the landmark," says Behnam Salemi's voice.
"I passed the door!" exults the voice of Jafar Adibi.
Giving the robot a simple way to determine whether or not the conference rooms were occupied also required ingenuity. The solution for Yoda was an invitation: when it enters a conference room,Yoda invites humans to come take some M&M candies from a dish sitting atop its console. Movement of a human-sized object in the room produces a noticeable change in the distance sensor readings - and causes Yoda to conclude that the room is occupied.
A human blocking the way is also reliably seen by Yoda's sensors. The robot stops and loudly announces "I'm looking for a better direction," in the voice of Gal Kaminka, and then carefully bypasses the obstacle. Or if the passage is completely blocked, Yoda stops until the way is clear.
The team began working on the project in March, 1996. "Each of us came from a different project area," Shen said. "None of us had robotics background, and we did the project on our own, without input from a robotics expert."
Yoda's basic chassis is a 10-year-old 3-wheeled robot manufactured by Denning Mobile Robotics, Inc., and made available by Ramakant Nevatia of the department of computer science. One small electric motor runs the wheels, and another controls the steering.
For better or worse, the group decided on its own the best way to program Yoda, without preconceptions about how it was supposed to be done: the bibliography of the short paper in which the group reports its work cites only the manual of the Denning robot and a single 1987 paper on behavior-based robot navigation, plus a theoretical work written by Shen. "In some ways, our ignorance turned out to be a strength," one team member commented. "We didn't know what could and could not be done, we just did things."
This is not to say that the team was totally without background in the subject. Even though no team member had ever actually programmed a robot before, the group included a great deal of programming expertise in closely allied areas, such as machine learning, planning, control systems, and expert systems. In addition, some members had considerable experience in hardware systems.
The competition was divided up into three heats, with two preliminary trials and then a final. After the long drive up to Portland, the group was dismayed to see how stiff the competition looked, with group after group putting high-tech, sleek 'bots through their paces, while what seemed an enormous audience - more than 200 experts - looked on.
"Yoda was the biggest, the oldest and the slowest," said team member Gal Kaminka. "We were pretty intimidated at first."
Intimidation turned to horror when the exhausted team discovered, in the first trial, that the programming method it had used to allow Yoda to find and go through a door wasn't working at all. The maze being used had thin, freshly-painted cardboard walls and the sonar sensors weren't picking up a 'verging' effect programmers had depended on to guide Yoda through the center of the doors. It worked at home at ISI - but here, the doors were invisible.
The team threw fabric over the cardboard to make the doorways perceptible to the robot, took a deduction, and spent all night in a feverish reprogramming session, finding an alternate method that would work for cardboard.
The next day, in the second trials session, the patched software worked perfectly. Yoda saw the doors. In fact, Yoda went through the whole exercise flawlessly, if somewhat slowly.
Here the team made another smart engineering decision. "We resisted the temptation to try to improve it," Shen said proudly, even though it was clear other robots might be faster. The next day, in the finals, Yoda again worked flawlessly - slowly, but perfectly, arriving back in the director's office to announce, in team member Kim's voice: "Mr. Director, the meeting begins in one minute."
True, Yoda didn't win first prize - the gold went, instead, to a tag-team of three small robots from SRI, in Palo Alto, who divided up the task and radioed each other with instructions and updates.
But Yoda went the distance - and when bonuses and deductions were duly added and subtracted, it had tied in a solid second, finishing ahead of competitors from traditional robotics powerhouse institutions such as CMU. (More than 20 robots were entered in the competition.)
And the exuberant team came out with a clear message: it isn't the hardware that counts. It's the software, and the people behind the software.
"You can do something with a good idea," said one team members, "even if the equipment is not the best."
It's not the end for ISI robotics. The unexpected near-win brought attention, and even the possibility of financial support for more work. At least one member of the team is actually considering shifting his concentration to robotics.
And Shen wants to continue: "I now have confidence in what we can accomplish with our limited time and energy," he said, "I want to use this as a vehicle to launch a longer research program, to create robots that can learn autonomously from their environment and go places that humans may have difficulty enduring."
The Yoda team included: Wei-Min Shen, Jafar Adibi, Bonghan Cho, Gal Kaminka, Jihie Kim, Behnam Salemi, and Sheila Tejada.
Photo- IRENE FERTIK
"Team Yoda" clusters around the robot that is "roughly the size and shape of a heavy-duty cannister vacuum cleaner." In the front row: Gal Kaminka, Sheila Tajada, Jihie Kim and Behnam Salemi. In the back row: Jafar Adibi, team leader Wei-Min Shen and Bonghan Cho.
|
