A pancake-shaped vacuum that cleans an entire room without
guidance, sensing and avoiding drop-offs like stairs. An R2-D2
sized cart that navigates hospital hallways to deliver patient
samples to a lab. Airline computers that man 1-800 numbers to give
callers updated flight information.
Dr. Peter Stone in the Department of Computer Sciences
has won four league competitions at international robot soccer
challenges, and competes again this week.
Devices driven by artificial
intelligence (AI) are slowly becoming part of the fabric of life.
But these gadgets seem simple-minded
compared to the ones Peter Stone envisions. The assistant professor
of computer sciences at The University of Texas at Austin is laying
the groundwork for a not-too-distant future populated by adaptable
AI devices that not only can improve their interactions with objects
around them, but affect each other’s actions as well.
“It’s not out of the question that in 20 years we’ll
be seeing cars that drive themselves,” Stone says, noting
that the cars’ abilities to communicate with each other would
permit rapid, accident-free travel.
“Creating those cars is a multiagent systems problem,” he
says. “It’s about learning how to develop independently
controlled cars that will drive themselves by coordinating around
streets, which will be a limited resource. That’s at the
core of all the research problems I look at.”
His chosen path
to those advances may seem unusual at first. It is paved in forest
green, with the limited resource a bright orange
ball. Stone and students taking his computer science course at
the College of Natural Sciences are preparing a team of eight robots
to compete this week at an international robot soccer competition
in Padua, Italy.
An avid soccer player since childhood, Stone was
on the varsity team at the University of Chicago, and plays on
a top-notch amateur
team in Austin. Stone became interested in using the sport to bring
about advances in AI soon after beginning graduate school at Carnegie
Mellon University in Pittsburgh. He saw a one-on-one robot soccer
demonstration at an AI meeting in 1993 and was struck by the robots’ simplicity.
“I immediately thought, ‘Soccer is not about one robot
on each team—it’s about multiple robots cooperating
and working against the other team,’ ” Stone says.
With his adviser’s approval, Stone began the AI project that
Using miniature, Rubix cube-sized robots he helped
design, he won the small-sized robot league at the first international
in 1997. He also reached the semifinals that year in the computer
simulation league, in which on-screen players used an AI-based
computer program of previously developed strategies to respond
to patterns of soccer play.
That following year, Stone won both
these international RoboCup leagues. And in 1999, he posted a shut-out
during eight games in
the simulator competition, winning 110-0.
“Our team became one of the ones that people started building
off of, because we made a lot of our source code available,” he
Stone described how he created the 1999 simulation team in
his book “Layered Learning in Multiagent Systems: A Winning
Approach to Robotic Soccer.” He also helped edit a similar
book on the 2000 RoboCup competition, but notes that “the
competitions are a good motivating tool, but not the be-all and
end-all of it.”
his research bent, it’s not surprising that the
energetic scientist decided to go back to the AI drawing board
after coming to Austin and the university’s Department of
Computer Sciences in fall 2002. Stone’s effort focuses on
Aibo (I-Bo) robots from Sony that have well-defined hardware, permitting
him and his students to focus entirely on computer programming.
They reworked the dog-sized robots’ code from scratch in
preparation for the first American RoboCup competition, held in
May, and are fine-tuning it further for the Italian competition,
expected to draw tens of thousands of spectators from July 2-11.
With only 14 weeks to prepare for the warm-up competition, Stone
and 19 graduate students and an undergraduate worked doggedly to
improve the robots’ abilities by breaking into groups. One,
led by graduate student Mohan Sridharan, focused on teaching the
robots, known as the UT Austin Villa team, to distinguish the pattern
of numbers on their video screens representing the pastel colors
of the goals, the green of their 2.1-meter by 4.5-meter playing
field, the colorful outfits of teammates and other objects. [Play
video of robots in action, with Dr. Stone’s commentary. Download
free QuickTime Player. For high bandwidth use: 19MB; 1 min.,
58 sec. playing time]
group worked to improve the ‘bots’ jerky walks,
and another to guide them through the steps of scoring a goal:
find the ball, walk to the ball, find the goal, square up the ball
with the goal, re-check that everything’s aligned and then
shoot with a one-legged lunging kick or by popping front legs back
together. The reverse of the latter, front power kick was taught
to a “goalie” to block a ball.
In addition, the Aibos
were taught to play without human input based on instructions built
into their code before the games. For
example, the “Dibs” program allowed an Austin Villa
robot close to the ball to stake a claim to it, using an Ethernet
connection to communicate that decision to teammates.
legged robot work was in progress, graduate student Gregory Kuhlmann
and undergrads Justin Lallinger and Bharat Kejriwal applied
a similar approach to improve code for the computer simulation
competition. The three came in second in the American RoboCup coached
computer simulation league, in which an AI-based computer program
provided adaptable coaching advice to on-screen players. In the
legged league, the larger crew lost in their final game of three,
which was against the runner-up team from Georgia Institute of
Stone and a smaller student group have since been prepping
for this week’s international RoboCup challenge. The Austin
walk twice as fast as in May, thanks to programming tweaks provided
by graduate student Daniel Stronger, and can scan the field and
kick faster as well thanks to graduate student Peggy Fidelman and
undergraduate Ellie Lin. They’re also able to tuck a ball
under the chin as cargo, or do a head-butt to send it flying sideways.
“We’re now ready in this last stretch before Italy
to focus on completely new behaviors that can take advantage of
building blocks,” Stone says.
Mohan Sridharan, graduate student in Electrical and Computer
Engineering, works to refine the robots’ visual acuity.
For example, the robots are
being programmed to detect where they and their teammates are on
the playing field so that programs like “Dibs” are
easier to carry out. These programs, which create what Stone refers
to as “locker room” agreements that are made before
competition, will be a key part of the computer simulation teams
in Italy, and play a role in Stone’s other AI-based activities.
example is the mock travel agent competitions that he has been
a part of since 2000. The competitions involve on-screen agents
haggling to create the best vacation packages based on various
flight, hotel and entertainment options. Stone’s team won
the 2000 Trading Agent Competition and was one of two declared
winners the following year. He hopes his team fares as well in
this year’s competition in August.
Though Stone always considers
these more practical applications of autonomous machinery, he admits
to feeling a rush when considering
RoboCup creator Hirokai Kitano’s ultimate vision: creating
a team of autonomous humanoid robots by 2050 that duke it out and
win against a real soccer team.
“That’s akin to the dream of landing a man on the
says, noting that this dream will keep AI researchers challenged
Other AI dreams in their infancy include robots that
can perform search and rescue operations for military or other
a network of computers whose nodes can sense outages and autonomously
decide the quickest ways to re-route e-mail.
As the dreams become
reality, Stone notes that people will tend to forget the machine
learning behind them. Consider, for example,
the case of shoppers who think nothing about AI when they go to
a Web site and receive suggestions of other products to purchase,
or chess-playing programs that make moves without input. There
is an advantage to these fluid expectations, though.
“It means we have the freedom to keep looking 20 years into
the future,” Stone says.
And you can bet that he’ll
travel the journey of a thousand innovations needed to get there
a single, well-planned robot step
at a time.