Outline History Motivation Comparison with real soccer
Artificial Intelligence in Robotic Soccer Simulation of Robotic
Soccer
Slide 3
What is Soccer Robot? A soccer robot is a specialized
autonomous robot and mobile robot that is used to play variants of
soccer -Wikipedia
Slide 4
Robots Playing Soccer The idea of robots playing soccer was
first mentioned by Professor Alan Mackworth (University of British
Columbia, Canada) in a paper entitled On Seeing Robots presented at
VI-92, 1992.
Slide 5
History Independently, a group of Japanese researchers
organized a Workshop on Grand Challenges in Artificial Intelligence
in October, 1992 in Tokyo, discussing possible grand challenge
problems. This workshop led to a serious discussions of using the
game of soccer for promoting science and technology.
Slide 6
Cont.. Factors Considered in the discussion: Technology
feasibility. Social impact assessment. Financial feasibility.
Slide 7
Result As a result of these studies, it has been concluded that
the project is feasible and desirable. In June 1993, a group of
researchers, including Minoru Asada, Yasuo Kuniyoshi, and Hiroaki
Kitano, decided to launch a robotic competition, tentatively named
as Robot J-League Within a month, on receiving overwhelming
reactions from researchers outside of Japan, renamed the project as
the Robot World Cup Initiative, RoboCup for short.
Slide 8
The Year 1997 In the history of artificial intelligence and
robotics, the year 1997 will be remembered as a turning point. In
May 1997, IBM Deep Blue defeated the human world champion in chess.
Forty years of challenge in the AI community came to a successful
conclusion. On July 4, 1997, NASAs pathfinder mission made a
successful landing and the first autonomous robotics system,
Sojourner, was deployed on the surface of Mars.
Slide 9
Together with these accomplishments, RoboCup made its first
step towards the development of robotic soccer with an aim of
beating human World Cup champion team by the year 2050. RoboCup and
its aim
Slide 10
Motivation Why all this focus letting robots play games of
soccer? Dealing with the real world properties. The element of
competition is not only an interesting research question, it also
serves as a way to put the research to the test. Element of
Competition.
Slide 11
Motivation Cont. Standard problems. Research on games, chess in
particular, has generated significant advances in the theory of
search algorithms and search control, as well as motivating
cognitive studies into the ways that humans approach the same
problems
Lessons from real Soccer The problem of creating a team
strategy from individual behavior Performance is probably very bad
on skills like passing and shooting. Aritificial Intelligence.
Slide 14
Solutions for good team work in human soccer Game system
Positioning Offensive Positioning Defensive Positioning The
golie
Slide 15
Goalie
Slide 16
Others Freezing the game Positioning at special kick
situations
Slide 17
Tactics Simplicity Let the ball work Only Stop when one cannot
advance immediately Be receptible for a pass Make passes playable
for the receiving robot.
Slide 18
Tactics Cont.. Dont be to obvious in your intentions.
Camouflage your weaknesses. Move towards a passed ball. Shield the
ball of the opponent. Do not slide unless absolutely necessary. Be
a good team player.
Slide 19
AI in Robotic Soccer Three divisions Machine Vision Camera
Snapshots + Image Processing etc. Artificial Intelligence Logic of
players movements. Communication Signal Transmission etc.
Slide 20
AI in Robotic Soccer Some Basic Definitions Roles Modularity
Layered Operations Decision Matrix
Slide 21
A decision matrix is a list of values in rows and columns that
allows an analyst to systematically identify, analyze, and rate the
performance of relationships between sets of values and
information.
Slide 22
AI in Robotic Soccer Artificial Intelligence Layers Strategy
layer Path planning layer Fuzzy reactive layer Motor control
layer
Slide 23
Strategy Layer Strategy layer Operations Task identification
Role prioritization Role assignment Robot destination
assignment
Slide 24
Role Assignment Using weighted importance of roles Role
allocation Heuristic
Slide 25
Task Identification & Role prioritization Inputs Ball speed
towards own goal Ball distance from own goal Large,
Multi-dimensional decision matrix Outputs Decision Matrix Heuristic
weights of importance of the goalie, attacker role, defender
role.
Slide 26
Robot destination assignment Based on role stereotype
Slide 27
Robot destination assignment Based on role stereotype Goalie
Movement between the goal and the position of the ball but within
the goal area. Ball moving toward the goal action Zonal division of
goal area
Slide 28
Robot destination assignment
Slide 29
Based on role stereotype Defender Common Role Interchangeable
role Similarity with Goalies target destination
Slide 30
Robot destination assignment
Slide 31
Based on role stereotype Attacker Aggressive Behavior Evasive
Behavior Similarity with Goalies target destination
Slide 32
Merits & Demerits Operations satisfactory in simulation
Inflexibility of fixed arrays to future additions Lack of sensible
interpolation policy for unknowns Constructing, storing and
accessing in decision matrix Inflexibility of this approach
Fuzzification Fuzzy set pair (U,m) where U is a set and sets
whose elements have degrees of membership. Classical set and Fuzzy
set Input variables are mapped into fuzzy sets Process of
converting a crisp input value to a fuzzy value
Fuzzy Output Sets Zero, Small Positive, Medium Positive, Large
Positive, Very Large Positive
Slide 39
Fuzzy Output Sets one Fuzzy Output Value representing each of
the three robot roles Output = { Attacker Fuzzy Output Set,
Defender Fuzzy Output Set, Goalie Fuzzy Output Set, }
Slide 40
Fuzzy Rules & Association memory matrix Fuzzy Rules are
very natural and can be expressed in natural Language. Example: If
ball speed is negative and ball position is away then attacker
priority is very large and defender priority is zero and goalie
priority is small. If ball speed is zero and ball position is away
then attacker priority is large and defender priority is small and
goalie priority is medium.
Slide 41
Fuzzy Rules & Association memory matrix Fuzzy Association
Memory Matrix (FAMM) is a complete mapping of all the fuzzy rules
in our controller. Example
Slide 42
Defuzzification Necessity of defuzzifying our Fuzzy Output
Values. Aggregation procedure v i is the centre value for the rule
that corresponds to weight w i.
Slide 43
Path Planning Layer Inputs are outputs from Strategy Layer. A*
implementation Fuzzy logic and A* navigation challenge Remedy:
Create a supplementary algorithm to generate a map of robot soccer
environment. Extensive refinement to ensure compatibility of A*
with Fuzzy logic system.
Slide 44
Path Planning Layer Red bot is expected to move along the
dotted path. Supplementary Algorithm generates the Map. A*
Algorithm generates the points.
Slide 45
Fuzzy Reactive layer Inputs are outputs from Path Planning
Layer Calculates angular velocity and forward speed for the
bot
Slide 46
Motor control layer Inputs are outputs from Fuzzy reactive
Layer Calculates the left motor velocity and right motor velocity
Informs the communication module to transmit the required data to
the correct bot.
Slide 47
Soccer Simulation Evaluation of various multi-agent systems and
cooperative algorithms. Components :- Soccer Server Client
Slide 48
Soccer Server o Provides Virtual soccer field o Simulates the
movements of players and the ball. o client-server
communication
Selection of Play-plans Situation: Two attackers attempting to
score a goal against a single opponent.
Slide 52
1) The offensive player shoot the ball at the goal. 2) Pass the
ball to his teammate. Here its a binary choice. but in general,
there are a very large number of possible play situations,
including many where both play plans may appear equally good.
Possibilities
Slide 53
Experiment The three players are placed in the penalty area
randomly. Ball placed on a line parallel to touchlines at a dist 1
m from the main attacker. Main Attackers Behavior Collect
information about the positions of all objects. Input the
information to a Selector module. Choose either a pass or a shoot
according to the output of the selector module.
Slide 54
Defender is programmed to move toward and, if possible, to
kick, the ball. The attacker s teammate is programmed to wait for a
pass and, if the pass comes, to shoot at the goal.
Slide 55
Types of selectors 1) Neural Network trained on a sample set of
data using back-propagation. 2) Apply the Decision tree learning
procedure to the same data set.
Slide 56
Neural Network Eight inputs, which corresponded to the eight
values of the distance and the angle from the main attacking player
to his teammate, the defender, the center of the goal, the ball. 30
hidden units. 2 output units, Opass and Oshoot. All fully
connected.
Slide 57
Selection The pass plan is chosen with probability Opass
/(Opass + Oshoot) The shoot plan is chosen with probability
Oshoot/(Opass + Oshoot).
Slide 58
Training Create a random database of 1000 situations. For each
situation, we then chose, at random, whether to carry out a pass
plan or a shoot plan and add the success or failure of a single
simulation of this plan to the data item Apply Back-Propagation
method to train the weights of the network.
Slide 59
Result
Slide 60
Output vs Direction of Opposite Player
Slide 61
On also varying the distance of opposite player
Slide 62
Decision Tree Method Decision Attributes Relative positions of
the objects. Two classes Success and failure.
Slide 63
Steps 1) Data set 1 with positions of the objects and action =
shoot. 2) Data set 2 with positions of the objects and action =
pass. 3) Predict a class, success or failure, for each of the two
data sets. 4) If only one plan is success, do the corresponding
action. 5) If both classes are success do the action whose
certainty factor is higher. 6) If both classes are failure pass or
shoot randomly
Slide 64
Result
Slide 65
Conclusion Soccer Server, a simulator of the game of soccer in
which players, controlled by individual client programs, can play a
soccer match Demonstrated the potential of Soccer Server by
reporting experiments that used the system to investigate how the
selection of play plans could be learned Research has been going in
domain of Robotic Soccer with an aim of defeating World Champion
Human team by the year 2050.
Slide 66
References http://en.wikipedia.org/wiki/Soccer_robot
http://en.wikipedia.org/wiki/Soccer_robot
http://www.robocup.org/about-robocup/a-brief- history-of-robocup/
http://www.robocup.org/about-robocup/a-brief- history-of-robocup/
Artificial Intelligence in a multi agent robot soccer domain by
Remco Anthony Seesink Applied Artificial Intelligence Journal:
Soccer server: A tool for research on multiagent systems by Itsuki
Noda, Hitoshi Matsubara, Kazuo Hiraki & Ian Frank Artificial
Intelligence in Robot Soccer by A. P. Gerdelan supervised by Dr. N.
H. Reyes