Title: Adam Overholtzer presents
1Adam Overholtzer presents
2ROBOCOP
3Err
4RoboCup
5RoboCupThe Robot World Cup Initiative
- http//www.robocup.org
- Kitano, Asada, Kuniyoshi, Noda, Osawa
6RoboCups Ultimate Goooooooooooooooooal!!
- By mid-21st century, a team of fully autonomous
humanoid robot soccer players shall win a soccer
game against the most recent World Cup champions
while fully complying with the official rules of
the FIFA (Fédération Internationale de Football
Association).
7What is RoboCup?
- an international research and education
initiative - an attempt to foster AI and intelligent robotics
research by providing a standard problem where
wide range of technologies can be integrated and
examined, as well as being used for integrated
project-oriented education - for this purpose, a soccer game was chosen as the
primary domain - Why Soccer?
- RoboCup is a task for a team of multiple
fast-moving robots under a dynamic environment - in order for a robot team to actually play
soccer, various technologies must be
incorporated, such as autonomous agents,
multi-agent collaboration, strategy acquisition,
and real-time reasoning
8A New Standard Problem?
- a standard problem is a clearly-defined,
realistic and affordable challenge that can be
used to evaluate and compare various algorithms,
theories, and architectures - Turing proposed chess as a standard problem for
AI (because chess is hard) - the authors propose that RoboCup should be a new
standard problem for intelligent robotics
9A New Standard ProblemWhy RoboCup?
Well, Deep Blue already defeated human Grand
Master Garry Kasparov. The chess problem is
(virtually) solved and has become too simple to
spur innovation, whereas RoboCup is cutting-edge
10Standard ProblemsCriticisms and Responses
- standard problems are often abstract tasks and
thus they ignore essential difficulties of real
world problem solving - however, solving any real world problem must
involve domain-specific restraints - additionally, work on real world systems is far
too expensive for many groups
RoboCup is designed to meet the need of handling
real world complexity, though in a limited
environment, while maintaining an affordable
problem size and research cost.
11Research Issues Robot Design
- existing robots have been designed to perform
mostly single behavior actions, such as pushing,
kicking, juggling, etc.
- a RoboCup player would need to be able to perform
multiple subtasks such as shooting and passing,
while also avoiding opponents - such a multitasking robot could be built with
either (1) many combined single task components
or (2) a smaller number of components each able
to perform several subtasks - approach 2 seems preferable, since a player would
need to be compact and quick - ideally the robots would be humanoid, though this
is obviously a long term goal
12Research Issues Sensors
- sensors need to provide visual representation of
the world and information about the results of
the robots actions - Computer Vision researchers seek to convert 2-D
sensor data into complex 3-D geometry, but this
is slow and processor-intensive - RoboCup players need a simpler and faster vision
system so player can react in real time - other possible sensing devices for detecting
information not provided by sight include sonar,
touch, and force/torque sensors
13Research Issues Learning
- each player has to be able to perform one of
several behaviors depending on the current
situation - considering the uncertainties of RoboCups
dynamic and nondeterministic environment,
programming behaviors for every situation seems
unfeasible - thus, reinforcement learning seems promising
because the robots can evolve advanced behaviors
with little pre-programmed knowledge - so far, almost all of the existing applications
have been done with computer simulations, not
robots
14Research Issues Learning
- the first step would be to evolve one-on-one
competition, though this is quite complex because
the player must take the opponents motions into
account - to reduce complexity, task decomposition can be
used - some work so far
- Asada et al. proposed a system where the shooting
and avoiding behaviors are acquired independently
and coordinated through the learning, though this
method is slow due to the huge state space needed - Sahota proposed a reactive deliberation approach
architecture, though his method requires perfect
knowledge and RoboCop does now allow for it
15Research Issues Learning
- the ultimate goal would be a many-to-many
competition, i.e., a full game of soccer - this would require dynamic team interaction, and
defining all such collective behaviors seems
infeasible, so learning them would seem promising - unfortunately, it is difficult to find simple
methods for learning collective behaviors, so
instead you would need to use pattern finding
methods or a coordination by imitation system - other issues, such as task representation
environment modeling are also challenging, so,
uh, clearly there is lots of good work to be done!
16Research Issues Collaboration
- a soccer game can be viewed as a real time
multi-agent environment great for distributed
AI and multi-agent research (like that Ant
Algorithm stuff) - each team has a team-wide common goal (to win)
and these goals are incompatible (i.e., both
teams cant win) - a team might have global (team-wide) strategies
to fulfill this common goal, as well as both
local and global tactics to achieve subgoals
(scoring points, blocking an enemy, etc.) - the opposing team could be seen as a dynamic and
obstructive environment that can prevent our team
from achieving its goal
17Research Issues Collaboration
- there are many challenges to overcome
- the environment, i.e. the movements of both
teams, is highly dynamic - each players perception is locally limited (no
global perception or knowledge) - each player may have a different role
- communication between teammates is limited and so
a player must be able to act autonomously - in summary, a soccer team can be viewed as a
cooper-ative distributed real-time planning
scheme, embedded in a highly dynamic environment
and it still aint as bad as real life!
18The Rules of RoboCup
- RoboCup consists of three competitions the real
robot competition, the software (or simulated)
robot competition, and the special skills
competition - for the real robot competition, the real
worldness of RoboCup arises from the vast
complexity of the overall situation due to
uncertainty and uncontrollability in the
structures and functions of the real robots along
with real physical phenomena - therefore the authors have aimed for minimal
regulations and the simplest possible environment
in the real robot competitions
19Rules of RoboCup Real Robots
- for both the small and medium sized real robot
leagues, the size and structure of the
environment, as well as the number of players, is
predetermined - the colorings of the ball, field and robots are
also predetermined so robots can be designed
accordingly - fouls are called when a robot intentionally
attacks another robot, or when a robot holds the
ball or crowds the defense zone
The field for small robots.
20Rules of RoboCup Real Robots
- the Simulation Track for the field of play is
based on a server-client system where the server
provides the virtual field and simulates all
movements - the simulated environment is extremely
straightforward, with a 2-D field and players
represented as circles with simple stepwise
motion and collision - a player is only aware of its position relative
to nearby landmarks, not its absolute position
on the field - each player is controlled by a client program
connected to the Soccer Server and every action,
including player-to-player communication, is
handled through the server
21Rules of RoboCup Simulated
- Soccer Server is also used to simulate games
using client programs as virtual robots, just as
discussed on the previous slide
- these virtual games are especially useful for
comparing multi-agent systems without having to
deal with all the physical limitations of robot
design - thus complex player coordination can be developed
while others develop workable robots
22Rules of RoboCup Special Skills
- in order to fine-tune various advanced physical
skills, these humanoid robots are designed to
compete at only one or two key abilities - this is the newest completion at RoboCup and very
few groups can afford to compete - such a robot is needed to accomplish RoboCups
ultimate goal, but theyve got a long way to go
see this kicking robot in action!
23In Summary
- RoboCup should be a new standard AI problem
- RoboCup provides rich research issues for a wide
range of AI and robotics studies - building a robot soccer team would require
integration of a broad range of technologies and
fundamental breakthroughs on some of the key
issues in agent design - the authors hope that the RoboCup initiative will
play a key role in promoting the state-of-the-art
in AI and robotics research - someday, robots will win the World Cup and the
fans will riot in the streets Domo arigato, Mr.
Roboto!
24Web Sites of Interest
- Main RoboCup Site
- RoboCup Soccer Simulator
- Special Interest Developer Groups
- Multi-agent Learning
- Multi-agent Modeling
- Vision
- Simulation Tools for Real Robots
- Configurable and Modular Robotics
- Lots of Movies by the Carnegie Mellon Team